1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid10.c : Multiple Devices driver for Linux
5 * Copyright (C) 2000-2004 Neil Brown
7 * RAID-10 support for md.
9 * Base on code in raid1.c. See raid1.c for further copyright information.
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
24 #include "md-bitmap.h"
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69 static int _enough(struct r10conf *conf, int previous, int ignore);
70 static int enough(struct r10conf *conf, int ignore);
71 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
73 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74 static void end_reshape_write(struct bio *bio);
75 static void end_reshape(struct r10conf *conf);
77 #define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
83 #define cmd_before(conf, cmd) \
85 write_sequnlock_irq(&(conf)->resync_lock); \
88 #define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
90 #define wait_event_barrier_cmd(conf, cond, cmd) \
91 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
94 #define wait_event_barrier(conf, cond) \
95 wait_event_barrier_cmd(conf, cond, NULL_CMD)
98 * for resync bio, r10bio pointer can be retrieved from the per-bio
99 * 'struct resync_pages'.
101 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
103 return get_resync_pages(bio)->raid_bio;
106 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
108 struct r10conf *conf = data;
109 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
111 /* allocate a r10bio with room for raid_disks entries in the
113 return kzalloc(size, gfp_flags);
116 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
117 /* amount of memory to reserve for resync requests */
118 #define RESYNC_WINDOW (1024*1024)
119 /* maximum number of concurrent requests, memory permitting */
120 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
121 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
122 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
125 * When performing a resync, we need to read and compare, so
126 * we need as many pages are there are copies.
127 * When performing a recovery, we need 2 bios, one for read,
128 * one for write (we recover only one drive per r10buf)
131 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
133 struct r10conf *conf = data;
134 struct r10bio *r10_bio;
137 int nalloc, nalloc_rp;
138 struct resync_pages *rps;
140 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
144 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
145 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
146 nalloc = conf->copies; /* resync */
148 nalloc = 2; /* recovery */
150 /* allocate once for all bios */
151 if (!conf->have_replacement)
154 nalloc_rp = nalloc * 2;
155 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
157 goto out_free_r10bio;
162 for (j = nalloc ; j-- ; ) {
163 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
166 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
167 r10_bio->devs[j].bio = bio;
168 if (!conf->have_replacement)
170 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
173 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
174 r10_bio->devs[j].repl_bio = bio;
177 * Allocate RESYNC_PAGES data pages and attach them
180 for (j = 0; j < nalloc; j++) {
181 struct bio *rbio = r10_bio->devs[j].repl_bio;
182 struct resync_pages *rp, *rp_repl;
186 rp_repl = &rps[nalloc + j];
188 bio = r10_bio->devs[j].bio;
190 if (!j || test_bit(MD_RECOVERY_SYNC,
191 &conf->mddev->recovery)) {
192 if (resync_alloc_pages(rp, gfp_flags))
195 memcpy(rp, &rps[0], sizeof(*rp));
196 resync_get_all_pages(rp);
199 rp->raid_bio = r10_bio;
200 bio->bi_private = rp;
202 memcpy(rp_repl, rp, sizeof(*rp));
203 rbio->bi_private = rp_repl;
211 resync_free_pages(&rps[j]);
215 for ( ; j < nalloc; j++) {
216 if (r10_bio->devs[j].bio)
217 bio_uninit(r10_bio->devs[j].bio);
218 kfree(r10_bio->devs[j].bio);
219 if (r10_bio->devs[j].repl_bio)
220 bio_uninit(r10_bio->devs[j].repl_bio);
221 kfree(r10_bio->devs[j].repl_bio);
225 rbio_pool_free(r10_bio, conf);
229 static void r10buf_pool_free(void *__r10_bio, void *data)
231 struct r10conf *conf = data;
232 struct r10bio *r10bio = __r10_bio;
234 struct resync_pages *rp = NULL;
236 for (j = conf->copies; j--; ) {
237 struct bio *bio = r10bio->devs[j].bio;
240 rp = get_resync_pages(bio);
241 resync_free_pages(rp);
246 bio = r10bio->devs[j].repl_bio;
253 /* resync pages array stored in the 1st bio's .bi_private */
256 rbio_pool_free(r10bio, conf);
259 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
263 for (i = 0; i < conf->geo.raid_disks; i++) {
264 struct bio **bio = & r10_bio->devs[i].bio;
265 if (!BIO_SPECIAL(*bio))
268 bio = &r10_bio->devs[i].repl_bio;
269 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
275 static void free_r10bio(struct r10bio *r10_bio)
277 struct r10conf *conf = r10_bio->mddev->private;
279 put_all_bios(conf, r10_bio);
280 mempool_free(r10_bio, &conf->r10bio_pool);
283 static void put_buf(struct r10bio *r10_bio)
285 struct r10conf *conf = r10_bio->mddev->private;
287 mempool_free(r10_bio, &conf->r10buf_pool);
292 static void wake_up_barrier(struct r10conf *conf)
294 if (wq_has_sleeper(&conf->wait_barrier))
295 wake_up(&conf->wait_barrier);
298 static void reschedule_retry(struct r10bio *r10_bio)
301 struct mddev *mddev = r10_bio->mddev;
302 struct r10conf *conf = mddev->private;
304 spin_lock_irqsave(&conf->device_lock, flags);
305 list_add(&r10_bio->retry_list, &conf->retry_list);
307 spin_unlock_irqrestore(&conf->device_lock, flags);
309 /* wake up frozen array... */
310 wake_up(&conf->wait_barrier);
312 md_wakeup_thread(mddev->thread);
316 * raid_end_bio_io() is called when we have finished servicing a mirrored
317 * operation and are ready to return a success/failure code to the buffer
320 static void raid_end_bio_io(struct r10bio *r10_bio)
322 struct bio *bio = r10_bio->master_bio;
323 struct r10conf *conf = r10_bio->mddev->private;
325 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
326 bio->bi_status = BLK_STS_IOERR;
330 * Wake up any possible resync thread that waits for the device
335 free_r10bio(r10_bio);
339 * Update disk head position estimator based on IRQ completion info.
341 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
343 struct r10conf *conf = r10_bio->mddev->private;
345 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
346 r10_bio->devs[slot].addr + (r10_bio->sectors);
350 * Find the disk number which triggered given bio
352 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
353 struct bio *bio, int *slotp, int *replp)
358 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
359 if (r10_bio->devs[slot].bio == bio)
361 if (r10_bio->devs[slot].repl_bio == bio) {
367 update_head_pos(slot, r10_bio);
373 return r10_bio->devs[slot].devnum;
376 static void raid10_end_read_request(struct bio *bio)
378 int uptodate = !bio->bi_status;
379 struct r10bio *r10_bio = bio->bi_private;
381 struct md_rdev *rdev;
382 struct r10conf *conf = r10_bio->mddev->private;
384 slot = r10_bio->read_slot;
385 rdev = r10_bio->devs[slot].rdev;
387 * this branch is our 'one mirror IO has finished' event handler:
389 update_head_pos(slot, r10_bio);
393 * Set R10BIO_Uptodate in our master bio, so that
394 * we will return a good error code to the higher
395 * levels even if IO on some other mirrored buffer fails.
397 * The 'master' represents the composite IO operation to
398 * user-side. So if something waits for IO, then it will
399 * wait for the 'master' bio.
401 set_bit(R10BIO_Uptodate, &r10_bio->state);
403 /* If all other devices that store this block have
404 * failed, we want to return the error upwards rather
405 * than fail the last device. Here we redefine
406 * "uptodate" to mean "Don't want to retry"
408 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
413 raid_end_bio_io(r10_bio);
414 rdev_dec_pending(rdev, conf->mddev);
417 * oops, read error - keep the refcount on the rdev
419 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
422 (unsigned long long)r10_bio->sector);
423 set_bit(R10BIO_ReadError, &r10_bio->state);
424 reschedule_retry(r10_bio);
428 static void close_write(struct r10bio *r10_bio)
430 /* clear the bitmap if all writes complete successfully */
431 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
433 !test_bit(R10BIO_Degraded, &r10_bio->state),
435 md_write_end(r10_bio->mddev);
438 static void one_write_done(struct r10bio *r10_bio)
440 if (atomic_dec_and_test(&r10_bio->remaining)) {
441 if (test_bit(R10BIO_WriteError, &r10_bio->state))
442 reschedule_retry(r10_bio);
444 close_write(r10_bio);
445 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
446 reschedule_retry(r10_bio);
448 raid_end_bio_io(r10_bio);
453 static void raid10_end_write_request(struct bio *bio)
455 struct r10bio *r10_bio = bio->bi_private;
458 struct r10conf *conf = r10_bio->mddev->private;
460 struct md_rdev *rdev = NULL;
461 struct bio *to_put = NULL;
464 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
466 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
469 rdev = conf->mirrors[dev].replacement;
473 rdev = conf->mirrors[dev].rdev;
476 * this branch is our 'one mirror IO has finished' event handler:
478 if (bio->bi_status && !discard_error) {
480 /* Never record new bad blocks to replacement,
483 md_error(rdev->mddev, rdev);
485 set_bit(WriteErrorSeen, &rdev->flags);
486 if (!test_and_set_bit(WantReplacement, &rdev->flags))
487 set_bit(MD_RECOVERY_NEEDED,
488 &rdev->mddev->recovery);
491 if (test_bit(FailFast, &rdev->flags) &&
492 (bio->bi_opf & MD_FAILFAST)) {
493 md_error(rdev->mddev, rdev);
497 * When the device is faulty, it is not necessary to
498 * handle write error.
500 if (!test_bit(Faulty, &rdev->flags))
501 set_bit(R10BIO_WriteError, &r10_bio->state);
503 /* Fail the request */
504 set_bit(R10BIO_Degraded, &r10_bio->state);
505 r10_bio->devs[slot].bio = NULL;
512 * Set R10BIO_Uptodate in our master bio, so that
513 * we will return a good error code for to the higher
514 * levels even if IO on some other mirrored buffer fails.
516 * The 'master' represents the composite IO operation to
517 * user-side. So if something waits for IO, then it will
518 * wait for the 'master' bio.
524 * Do not set R10BIO_Uptodate if the current device is
525 * rebuilding or Faulty. This is because we cannot use
526 * such device for properly reading the data back (we could
527 * potentially use it, if the current write would have felt
528 * before rdev->recovery_offset, but for simplicity we don't
531 if (test_bit(In_sync, &rdev->flags) &&
532 !test_bit(Faulty, &rdev->flags))
533 set_bit(R10BIO_Uptodate, &r10_bio->state);
535 /* Maybe we can clear some bad blocks. */
536 if (is_badblock(rdev,
537 r10_bio->devs[slot].addr,
539 &first_bad, &bad_sectors) && !discard_error) {
542 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
544 r10_bio->devs[slot].bio = IO_MADE_GOOD;
546 set_bit(R10BIO_MadeGood, &r10_bio->state);
552 * Let's see if all mirrored write operations have finished
555 one_write_done(r10_bio);
557 rdev_dec_pending(rdev, conf->mddev);
563 * RAID10 layout manager
564 * As well as the chunksize and raid_disks count, there are two
565 * parameters: near_copies and far_copies.
566 * near_copies * far_copies must be <= raid_disks.
567 * Normally one of these will be 1.
568 * If both are 1, we get raid0.
569 * If near_copies == raid_disks, we get raid1.
571 * Chunks are laid out in raid0 style with near_copies copies of the
572 * first chunk, followed by near_copies copies of the next chunk and
574 * If far_copies > 1, then after 1/far_copies of the array has been assigned
575 * as described above, we start again with a device offset of near_copies.
576 * So we effectively have another copy of the whole array further down all
577 * the drives, but with blocks on different drives.
578 * With this layout, and block is never stored twice on the one device.
580 * raid10_find_phys finds the sector offset of a given virtual sector
581 * on each device that it is on.
583 * raid10_find_virt does the reverse mapping, from a device and a
584 * sector offset to a virtual address
587 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
595 int last_far_set_start, last_far_set_size;
597 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
598 last_far_set_start *= geo->far_set_size;
600 last_far_set_size = geo->far_set_size;
601 last_far_set_size += (geo->raid_disks % geo->far_set_size);
603 /* now calculate first sector/dev */
604 chunk = r10bio->sector >> geo->chunk_shift;
605 sector = r10bio->sector & geo->chunk_mask;
607 chunk *= geo->near_copies;
609 dev = sector_div(stripe, geo->raid_disks);
611 stripe *= geo->far_copies;
613 sector += stripe << geo->chunk_shift;
615 /* and calculate all the others */
616 for (n = 0; n < geo->near_copies; n++) {
620 r10bio->devs[slot].devnum = d;
621 r10bio->devs[slot].addr = s;
624 for (f = 1; f < geo->far_copies; f++) {
625 set = d / geo->far_set_size;
626 d += geo->near_copies;
628 if ((geo->raid_disks % geo->far_set_size) &&
629 (d > last_far_set_start)) {
630 d -= last_far_set_start;
631 d %= last_far_set_size;
632 d += last_far_set_start;
634 d %= geo->far_set_size;
635 d += geo->far_set_size * set;
638 r10bio->devs[slot].devnum = d;
639 r10bio->devs[slot].addr = s;
643 if (dev >= geo->raid_disks) {
645 sector += (geo->chunk_mask + 1);
650 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
652 struct geom *geo = &conf->geo;
654 if (conf->reshape_progress != MaxSector &&
655 ((r10bio->sector >= conf->reshape_progress) !=
656 conf->mddev->reshape_backwards)) {
657 set_bit(R10BIO_Previous, &r10bio->state);
660 clear_bit(R10BIO_Previous, &r10bio->state);
662 __raid10_find_phys(geo, r10bio);
665 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
667 sector_t offset, chunk, vchunk;
668 /* Never use conf->prev as this is only called during resync
669 * or recovery, so reshape isn't happening
671 struct geom *geo = &conf->geo;
672 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
673 int far_set_size = geo->far_set_size;
674 int last_far_set_start;
676 if (geo->raid_disks % geo->far_set_size) {
677 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
678 last_far_set_start *= geo->far_set_size;
680 if (dev >= last_far_set_start) {
681 far_set_size = geo->far_set_size;
682 far_set_size += (geo->raid_disks % geo->far_set_size);
683 far_set_start = last_far_set_start;
687 offset = sector & geo->chunk_mask;
688 if (geo->far_offset) {
690 chunk = sector >> geo->chunk_shift;
691 fc = sector_div(chunk, geo->far_copies);
692 dev -= fc * geo->near_copies;
693 if (dev < far_set_start)
696 while (sector >= geo->stride) {
697 sector -= geo->stride;
698 if (dev < (geo->near_copies + far_set_start))
699 dev += far_set_size - geo->near_copies;
701 dev -= geo->near_copies;
703 chunk = sector >> geo->chunk_shift;
705 vchunk = chunk * geo->raid_disks + dev;
706 sector_div(vchunk, geo->near_copies);
707 return (vchunk << geo->chunk_shift) + offset;
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
722 * The rdev for the device selected will have nr_pending incremented.
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
729 static struct md_rdev *read_balance(struct r10conf *conf,
730 struct r10bio *r10_bio,
733 const sector_t this_sector = r10_bio->sector;
735 int sectors = r10_bio->sectors;
736 int best_good_sectors;
737 sector_t new_distance, best_dist;
738 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
740 int best_dist_slot, best_pending_slot;
741 bool has_nonrot_disk = false;
742 unsigned int min_pending;
743 struct geom *geo = &conf->geo;
745 raid10_find_phys(conf, r10_bio);
748 min_pending = UINT_MAX;
749 best_dist_rdev = NULL;
750 best_pending_rdev = NULL;
751 best_dist = MaxSector;
752 best_good_sectors = 0;
754 clear_bit(R10BIO_FailFast, &r10_bio->state);
756 * Check if we can balance. We can balance on the whole
757 * device if no resync is going on (recovery is ok), or below
758 * the resync window. We take the first readable disk when
759 * above the resync window.
761 if ((conf->mddev->recovery_cp < MaxSector
762 && (this_sector + sectors >= conf->next_resync)) ||
763 (mddev_is_clustered(conf->mddev) &&
764 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
765 this_sector + sectors)))
768 for (slot = 0; slot < conf->copies ; slot++) {
772 unsigned int pending;
775 if (r10_bio->devs[slot].bio == IO_BLOCKED)
777 disk = r10_bio->devs[slot].devnum;
778 rdev = rcu_dereference(conf->mirrors[disk].replacement);
779 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
780 r10_bio->devs[slot].addr + sectors >
781 rdev->recovery_offset) {
783 * Read replacement first to prevent reading both rdev
784 * and replacement as NULL during replacement replace
788 rdev = rcu_dereference(conf->mirrors[disk].rdev);
791 test_bit(Faulty, &rdev->flags))
793 if (!test_bit(In_sync, &rdev->flags) &&
794 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
797 dev_sector = r10_bio->devs[slot].addr;
798 if (is_badblock(rdev, dev_sector, sectors,
799 &first_bad, &bad_sectors)) {
800 if (best_dist < MaxSector)
801 /* Already have a better slot */
803 if (first_bad <= dev_sector) {
804 /* Cannot read here. If this is the
805 * 'primary' device, then we must not read
806 * beyond 'bad_sectors' from another device.
808 bad_sectors -= (dev_sector - first_bad);
809 if (!do_balance && sectors > bad_sectors)
810 sectors = bad_sectors;
811 if (best_good_sectors > sectors)
812 best_good_sectors = sectors;
814 sector_t good_sectors =
815 first_bad - dev_sector;
816 if (good_sectors > best_good_sectors) {
817 best_good_sectors = good_sectors;
818 best_dist_slot = slot;
819 best_dist_rdev = rdev;
822 /* Must read from here */
827 best_good_sectors = sectors;
832 nonrot = bdev_nonrot(rdev->bdev);
833 has_nonrot_disk |= nonrot;
834 pending = atomic_read(&rdev->nr_pending);
835 if (min_pending > pending && nonrot) {
836 min_pending = pending;
837 best_pending_slot = slot;
838 best_pending_rdev = rdev;
841 if (best_dist_slot >= 0)
842 /* At least 2 disks to choose from so failfast is OK */
843 set_bit(R10BIO_FailFast, &r10_bio->state);
844 /* This optimisation is debatable, and completely destroys
845 * sequential read speed for 'far copies' arrays. So only
846 * keep it for 'near' arrays, and review those later.
848 if (geo->near_copies > 1 && !pending)
851 /* for far > 1 always use the lowest address */
852 else if (geo->far_copies > 1)
853 new_distance = r10_bio->devs[slot].addr;
855 new_distance = abs(r10_bio->devs[slot].addr -
856 conf->mirrors[disk].head_position);
858 if (new_distance < best_dist) {
859 best_dist = new_distance;
860 best_dist_slot = slot;
861 best_dist_rdev = rdev;
864 if (slot >= conf->copies) {
865 if (has_nonrot_disk) {
866 slot = best_pending_slot;
867 rdev = best_pending_rdev;
869 slot = best_dist_slot;
870 rdev = best_dist_rdev;
875 atomic_inc(&rdev->nr_pending);
876 r10_bio->read_slot = slot;
880 *max_sectors = best_good_sectors;
885 static void flush_pending_writes(struct r10conf *conf)
887 /* Any writes that have been queued but are awaiting
888 * bitmap updates get flushed here.
890 spin_lock_irq(&conf->device_lock);
892 if (conf->pending_bio_list.head) {
893 struct blk_plug plug;
896 bio = bio_list_get(&conf->pending_bio_list);
897 spin_unlock_irq(&conf->device_lock);
900 * As this is called in a wait_event() loop (see freeze_array),
901 * current->state might be TASK_UNINTERRUPTIBLE which will
902 * cause a warning when we prepare to wait again. As it is
903 * rare that this path is taken, it is perfectly safe to force
904 * us to go around the wait_event() loop again, so the warning
905 * is a false-positive. Silence the warning by resetting
908 __set_current_state(TASK_RUNNING);
910 blk_start_plug(&plug);
911 raid1_prepare_flush_writes(conf->mddev->bitmap);
912 wake_up(&conf->wait_barrier);
914 while (bio) { /* submit pending writes */
915 struct bio *next = bio->bi_next;
917 raid1_submit_write(bio);
921 blk_finish_plug(&plug);
923 spin_unlock_irq(&conf->device_lock);
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
948 static void raise_barrier(struct r10conf *conf, int force)
950 write_seqlock_irq(&conf->resync_lock);
952 if (WARN_ON_ONCE(force && !conf->barrier))
955 /* Wait until no block IO is waiting (unless 'force') */
956 wait_event_barrier(conf, force || !conf->nr_waiting);
958 /* block any new IO from starting */
959 WRITE_ONCE(conf->barrier, conf->barrier + 1);
961 /* Now wait for all pending IO to complete */
962 wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
963 conf->barrier < RESYNC_DEPTH);
965 write_sequnlock_irq(&conf->resync_lock);
968 static void lower_barrier(struct r10conf *conf)
972 write_seqlock_irqsave(&conf->resync_lock, flags);
973 WRITE_ONCE(conf->barrier, conf->barrier - 1);
974 write_sequnlock_irqrestore(&conf->resync_lock, flags);
975 wake_up(&conf->wait_barrier);
978 static bool stop_waiting_barrier(struct r10conf *conf)
980 struct bio_list *bio_list = current->bio_list;
981 struct md_thread *thread;
983 /* barrier is dropped */
988 * If there are already pending requests (preventing the barrier from
989 * rising completely), and the pre-process bio queue isn't empty, then
990 * don't wait, as we need to empty that queue to get the nr_pending
993 if (atomic_read(&conf->nr_pending) && bio_list &&
994 (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
997 /* daemon thread must exist while handling io */
998 thread = rcu_dereference_protected(conf->mddev->thread, true);
1000 * move on if io is issued from raid10d(), nr_pending is not released
1001 * from original io(see handle_read_error()). All raise barrier is
1002 * blocked until this io is done.
1004 if (thread->tsk == current) {
1005 WARN_ON_ONCE(atomic_read(&conf->nr_pending) == 0);
1012 static bool wait_barrier_nolock(struct r10conf *conf)
1014 unsigned int seq = read_seqbegin(&conf->resync_lock);
1016 if (READ_ONCE(conf->barrier))
1019 atomic_inc(&conf->nr_pending);
1020 if (!read_seqretry(&conf->resync_lock, seq))
1023 if (atomic_dec_and_test(&conf->nr_pending))
1024 wake_up_barrier(conf);
1029 static bool wait_barrier(struct r10conf *conf, bool nowait)
1033 if (wait_barrier_nolock(conf))
1036 write_seqlock_irq(&conf->resync_lock);
1037 if (conf->barrier) {
1038 /* Return false when nowait flag is set */
1043 raid10_log(conf->mddev, "wait barrier");
1044 wait_event_barrier(conf, stop_waiting_barrier(conf));
1047 if (!conf->nr_waiting)
1048 wake_up(&conf->wait_barrier);
1050 /* Only increment nr_pending when we wait */
1052 atomic_inc(&conf->nr_pending);
1053 write_sequnlock_irq(&conf->resync_lock);
1057 static void allow_barrier(struct r10conf *conf)
1059 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1060 (conf->array_freeze_pending))
1061 wake_up_barrier(conf);
1064 static void freeze_array(struct r10conf *conf, int extra)
1066 /* stop syncio and normal IO and wait for everything to
1068 * We increment barrier and nr_waiting, and then
1069 * wait until nr_pending match nr_queued+extra
1070 * This is called in the context of one normal IO request
1071 * that has failed. Thus any sync request that might be pending
1072 * will be blocked by nr_pending, and we need to wait for
1073 * pending IO requests to complete or be queued for re-try.
1074 * Thus the number queued (nr_queued) plus this request (extra)
1075 * must match the number of pending IOs (nr_pending) before
1078 write_seqlock_irq(&conf->resync_lock);
1079 conf->array_freeze_pending++;
1080 WRITE_ONCE(conf->barrier, conf->barrier + 1);
1082 wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1083 conf->nr_queued + extra, flush_pending_writes(conf));
1084 conf->array_freeze_pending--;
1085 write_sequnlock_irq(&conf->resync_lock);
1088 static void unfreeze_array(struct r10conf *conf)
1090 /* reverse the effect of the freeze */
1091 write_seqlock_irq(&conf->resync_lock);
1092 WRITE_ONCE(conf->barrier, conf->barrier - 1);
1094 wake_up(&conf->wait_barrier);
1095 write_sequnlock_irq(&conf->resync_lock);
1098 static sector_t choose_data_offset(struct r10bio *r10_bio,
1099 struct md_rdev *rdev)
1101 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1102 test_bit(R10BIO_Previous, &r10_bio->state))
1103 return rdev->data_offset;
1105 return rdev->new_data_offset;
1108 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1110 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1111 struct mddev *mddev = plug->cb.data;
1112 struct r10conf *conf = mddev->private;
1115 if (from_schedule) {
1116 spin_lock_irq(&conf->device_lock);
1117 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1118 spin_unlock_irq(&conf->device_lock);
1119 wake_up_barrier(conf);
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 raid1_prepare_flush_writes(mddev->bitmap);
1128 wake_up_barrier(conf);
1130 while (bio) { /* submit pending writes */
1131 struct bio *next = bio->bi_next;
1133 raid1_submit_write(bio);
1141 * 1. Register the new request and wait if the reconstruction thread has put
1142 * up a bar for new requests. Continue immediately if no resync is active
1144 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1146 static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1147 struct bio *bio, sector_t sectors)
1149 /* Bail out if REQ_NOWAIT is set for the bio */
1150 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1151 bio_wouldblock_error(bio);
1154 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1155 bio->bi_iter.bi_sector < conf->reshape_progress &&
1156 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1157 allow_barrier(conf);
1158 if (bio->bi_opf & REQ_NOWAIT) {
1159 bio_wouldblock_error(bio);
1162 raid10_log(conf->mddev, "wait reshape");
1163 wait_event(conf->wait_barrier,
1164 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1165 conf->reshape_progress >= bio->bi_iter.bi_sector +
1167 wait_barrier(conf, false);
1172 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1173 struct r10bio *r10_bio, bool io_accounting)
1175 struct r10conf *conf = mddev->private;
1176 struct bio *read_bio;
1177 const enum req_op op = bio_op(bio);
1178 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1180 struct md_rdev *rdev;
1181 char b[BDEVNAME_SIZE];
1182 int slot = r10_bio->read_slot;
1183 struct md_rdev *err_rdev = NULL;
1184 gfp_t gfp = GFP_NOIO;
1186 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1188 * This is an error retry, but we cannot
1189 * safely dereference the rdev in the r10_bio,
1190 * we must use the one in conf.
1191 * If it has already been disconnected (unlikely)
1192 * we lose the device name in error messages.
1196 * As we are blocking raid10, it is a little safer to
1199 gfp = GFP_NOIO | __GFP_HIGH;
1202 disk = r10_bio->devs[slot].devnum;
1203 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1205 snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1208 /* This never gets dereferenced */
1209 err_rdev = r10_bio->devs[slot].rdev;
1214 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1216 rdev = read_balance(conf, r10_bio, &max_sectors);
1219 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1221 (unsigned long long)r10_bio->sector);
1223 raid_end_bio_io(r10_bio);
1227 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1230 (unsigned long long)r10_bio->sector);
1231 if (max_sectors < bio_sectors(bio)) {
1232 struct bio *split = bio_split(bio, max_sectors,
1233 gfp, &conf->bio_split);
1234 bio_chain(split, bio);
1235 allow_barrier(conf);
1236 submit_bio_noacct(bio);
1237 wait_barrier(conf, false);
1239 r10_bio->master_bio = bio;
1240 r10_bio->sectors = max_sectors;
1242 slot = r10_bio->read_slot;
1244 if (io_accounting) {
1245 md_account_bio(mddev, &bio);
1246 r10_bio->master_bio = bio;
1248 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1250 r10_bio->devs[slot].bio = read_bio;
1251 r10_bio->devs[slot].rdev = rdev;
1253 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1254 choose_data_offset(r10_bio, rdev);
1255 read_bio->bi_end_io = raid10_end_read_request;
1256 read_bio->bi_opf = op | do_sync;
1257 if (test_bit(FailFast, &rdev->flags) &&
1258 test_bit(R10BIO_FailFast, &r10_bio->state))
1259 read_bio->bi_opf |= MD_FAILFAST;
1260 read_bio->bi_private = r10_bio;
1263 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1265 submit_bio_noacct(read_bio);
1269 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1270 struct bio *bio, bool replacement,
1273 const enum req_op op = bio_op(bio);
1274 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1275 const blk_opf_t do_fua = bio->bi_opf & REQ_FUA;
1276 unsigned long flags;
1277 struct r10conf *conf = mddev->private;
1278 struct md_rdev *rdev;
1279 int devnum = r10_bio->devs[n_copy].devnum;
1283 rdev = conf->mirrors[devnum].replacement;
1285 /* Replacement just got moved to main 'rdev' */
1287 rdev = conf->mirrors[devnum].rdev;
1290 rdev = conf->mirrors[devnum].rdev;
1292 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1294 r10_bio->devs[n_copy].repl_bio = mbio;
1296 r10_bio->devs[n_copy].bio = mbio;
1298 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1299 choose_data_offset(r10_bio, rdev));
1300 mbio->bi_end_io = raid10_end_write_request;
1301 mbio->bi_opf = op | do_sync | do_fua;
1302 if (!replacement && test_bit(FailFast,
1303 &conf->mirrors[devnum].rdev->flags)
1304 && enough(conf, devnum))
1305 mbio->bi_opf |= MD_FAILFAST;
1306 mbio->bi_private = r10_bio;
1308 if (conf->mddev->gendisk)
1309 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1311 /* flush_pending_writes() needs access to the rdev so...*/
1312 mbio->bi_bdev = (void *)rdev;
1314 atomic_inc(&r10_bio->remaining);
1316 if (!raid1_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) {
1317 spin_lock_irqsave(&conf->device_lock, flags);
1318 bio_list_add(&conf->pending_bio_list, mbio);
1319 spin_unlock_irqrestore(&conf->device_lock, flags);
1320 md_wakeup_thread(mddev->thread);
1324 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1327 struct r10conf *conf = mddev->private;
1328 struct md_rdev *blocked_rdev;
1331 blocked_rdev = NULL;
1333 for (i = 0; i < conf->copies; i++) {
1334 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1335 struct md_rdev *rrdev = rcu_dereference(
1336 conf->mirrors[i].replacement);
1339 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1340 atomic_inc(&rdev->nr_pending);
1341 blocked_rdev = rdev;
1344 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1345 atomic_inc(&rrdev->nr_pending);
1346 blocked_rdev = rrdev;
1350 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1352 sector_t dev_sector = r10_bio->devs[i].addr;
1357 * Discard request doesn't care the write result
1358 * so it doesn't need to wait blocked disk here.
1360 if (!r10_bio->sectors)
1363 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1364 &first_bad, &bad_sectors);
1367 * Mustn't write here until the bad block
1370 atomic_inc(&rdev->nr_pending);
1371 set_bit(BlockedBadBlocks, &rdev->flags);
1372 blocked_rdev = rdev;
1379 if (unlikely(blocked_rdev)) {
1380 /* Have to wait for this device to get unblocked, then retry */
1381 allow_barrier(conf);
1382 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1383 __func__, blocked_rdev->raid_disk);
1384 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1385 wait_barrier(conf, false);
1390 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1391 struct r10bio *r10_bio)
1393 struct r10conf *conf = mddev->private;
1398 if ((mddev_is_clustered(mddev) &&
1399 md_cluster_ops->area_resyncing(mddev, WRITE,
1400 bio->bi_iter.bi_sector,
1401 bio_end_sector(bio)))) {
1403 /* Bail out if REQ_NOWAIT is set for the bio */
1404 if (bio->bi_opf & REQ_NOWAIT) {
1405 bio_wouldblock_error(bio);
1409 prepare_to_wait(&conf->wait_barrier,
1411 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1412 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1416 finish_wait(&conf->wait_barrier, &w);
1419 sectors = r10_bio->sectors;
1420 if (!regular_request_wait(mddev, conf, bio, sectors))
1422 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1423 (mddev->reshape_backwards
1424 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1425 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1426 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1427 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1428 /* Need to update reshape_position in metadata */
1429 mddev->reshape_position = conf->reshape_progress;
1430 set_mask_bits(&mddev->sb_flags, 0,
1431 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1432 md_wakeup_thread(mddev->thread);
1433 if (bio->bi_opf & REQ_NOWAIT) {
1434 allow_barrier(conf);
1435 bio_wouldblock_error(bio);
1438 raid10_log(conf->mddev, "wait reshape metadata");
1439 wait_event(mddev->sb_wait,
1440 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1442 conf->reshape_safe = mddev->reshape_position;
1445 /* first select target devices under rcu_lock and
1446 * inc refcount on their rdev. Record them by setting
1448 * If there are known/acknowledged bad blocks on any device
1449 * on which we have seen a write error, we want to avoid
1450 * writing to those blocks. This potentially requires several
1451 * writes to write around the bad blocks. Each set of writes
1452 * gets its own r10_bio with a set of bios attached.
1455 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1456 raid10_find_phys(conf, r10_bio);
1458 wait_blocked_dev(mddev, r10_bio);
1461 max_sectors = r10_bio->sectors;
1463 for (i = 0; i < conf->copies; i++) {
1464 int d = r10_bio->devs[i].devnum;
1465 struct md_rdev *rdev, *rrdev;
1467 rrdev = rcu_dereference(conf->mirrors[d].replacement);
1469 * Read replacement first to prevent reading both rdev and
1470 * replacement as NULL during replacement replace rdev.
1473 rdev = rcu_dereference(conf->mirrors[d].rdev);
1476 if (rdev && (test_bit(Faulty, &rdev->flags)))
1478 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1481 r10_bio->devs[i].bio = NULL;
1482 r10_bio->devs[i].repl_bio = NULL;
1484 if (!rdev && !rrdev) {
1485 set_bit(R10BIO_Degraded, &r10_bio->state);
1488 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1490 sector_t dev_sector = r10_bio->devs[i].addr;
1494 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1495 &first_bad, &bad_sectors);
1496 if (is_bad && first_bad <= dev_sector) {
1497 /* Cannot write here at all */
1498 bad_sectors -= (dev_sector - first_bad);
1499 if (bad_sectors < max_sectors)
1500 /* Mustn't write more than bad_sectors
1501 * to other devices yet
1503 max_sectors = bad_sectors;
1504 /* We don't set R10BIO_Degraded as that
1505 * only applies if the disk is missing,
1506 * so it might be re-added, and we want to
1507 * know to recover this chunk.
1508 * In this case the device is here, and the
1509 * fact that this chunk is not in-sync is
1510 * recorded in the bad block log.
1515 int good_sectors = first_bad - dev_sector;
1516 if (good_sectors < max_sectors)
1517 max_sectors = good_sectors;
1521 r10_bio->devs[i].bio = bio;
1522 atomic_inc(&rdev->nr_pending);
1525 r10_bio->devs[i].repl_bio = bio;
1526 atomic_inc(&rrdev->nr_pending);
1531 if (max_sectors < r10_bio->sectors)
1532 r10_bio->sectors = max_sectors;
1534 if (r10_bio->sectors < bio_sectors(bio)) {
1535 struct bio *split = bio_split(bio, r10_bio->sectors,
1536 GFP_NOIO, &conf->bio_split);
1537 bio_chain(split, bio);
1538 allow_barrier(conf);
1539 submit_bio_noacct(bio);
1540 wait_barrier(conf, false);
1542 r10_bio->master_bio = bio;
1545 md_account_bio(mddev, &bio);
1546 r10_bio->master_bio = bio;
1547 atomic_set(&r10_bio->remaining, 1);
1548 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1550 for (i = 0; i < conf->copies; i++) {
1551 if (r10_bio->devs[i].bio)
1552 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1553 if (r10_bio->devs[i].repl_bio)
1554 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1556 one_write_done(r10_bio);
1559 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1561 struct r10conf *conf = mddev->private;
1562 struct r10bio *r10_bio;
1564 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1566 r10_bio->master_bio = bio;
1567 r10_bio->sectors = sectors;
1569 r10_bio->mddev = mddev;
1570 r10_bio->sector = bio->bi_iter.bi_sector;
1572 r10_bio->read_slot = -1;
1573 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1574 conf->geo.raid_disks);
1576 if (bio_data_dir(bio) == READ)
1577 raid10_read_request(mddev, bio, r10_bio, true);
1579 raid10_write_request(mddev, bio, r10_bio);
1582 static void raid_end_discard_bio(struct r10bio *r10bio)
1584 struct r10conf *conf = r10bio->mddev->private;
1585 struct r10bio *first_r10bio;
1587 while (atomic_dec_and_test(&r10bio->remaining)) {
1589 allow_barrier(conf);
1591 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1592 first_r10bio = (struct r10bio *)r10bio->master_bio;
1593 free_r10bio(r10bio);
1594 r10bio = first_r10bio;
1596 md_write_end(r10bio->mddev);
1597 bio_endio(r10bio->master_bio);
1598 free_r10bio(r10bio);
1604 static void raid10_end_discard_request(struct bio *bio)
1606 struct r10bio *r10_bio = bio->bi_private;
1607 struct r10conf *conf = r10_bio->mddev->private;
1608 struct md_rdev *rdev = NULL;
1613 * We don't care the return value of discard bio
1615 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1616 set_bit(R10BIO_Uptodate, &r10_bio->state);
1618 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1620 rdev = conf->mirrors[dev].replacement;
1623 * raid10_remove_disk uses smp_mb to make sure rdev is set to
1624 * replacement before setting replacement to NULL. It can read
1625 * rdev first without barrier protect even replacement is NULL
1628 rdev = conf->mirrors[dev].rdev;
1631 raid_end_discard_bio(r10_bio);
1632 rdev_dec_pending(rdev, conf->mddev);
1636 * There are some limitations to handle discard bio
1637 * 1st, the discard size is bigger than stripe_size*2.
1638 * 2st, if the discard bio spans reshape progress, we use the old way to
1639 * handle discard bio
1641 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1643 struct r10conf *conf = mddev->private;
1644 struct geom *geo = &conf->geo;
1645 int far_copies = geo->far_copies;
1646 bool first_copy = true;
1647 struct r10bio *r10_bio, *first_r10bio;
1651 unsigned int stripe_size;
1652 unsigned int stripe_data_disks;
1653 sector_t split_size;
1654 sector_t bio_start, bio_end;
1655 sector_t first_stripe_index, last_stripe_index;
1656 sector_t start_disk_offset;
1657 unsigned int start_disk_index;
1658 sector_t end_disk_offset;
1659 unsigned int end_disk_index;
1660 unsigned int remainder;
1662 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1665 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1666 bio_wouldblock_error(bio);
1669 wait_barrier(conf, false);
1672 * Check reshape again to avoid reshape happens after checking
1673 * MD_RECOVERY_RESHAPE and before wait_barrier
1675 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1678 if (geo->near_copies)
1679 stripe_data_disks = geo->raid_disks / geo->near_copies +
1680 geo->raid_disks % geo->near_copies;
1682 stripe_data_disks = geo->raid_disks;
1684 stripe_size = stripe_data_disks << geo->chunk_shift;
1686 bio_start = bio->bi_iter.bi_sector;
1687 bio_end = bio_end_sector(bio);
1690 * Maybe one discard bio is smaller than strip size or across one
1691 * stripe and discard region is larger than one stripe size. For far
1692 * offset layout, if the discard region is not aligned with stripe
1693 * size, there is hole when we submit discard bio to member disk.
1694 * For simplicity, we only handle discard bio which discard region
1695 * is bigger than stripe_size * 2
1697 if (bio_sectors(bio) < stripe_size*2)
1701 * Keep bio aligned with strip size.
1703 div_u64_rem(bio_start, stripe_size, &remainder);
1705 split_size = stripe_size - remainder;
1706 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1707 bio_chain(split, bio);
1708 allow_barrier(conf);
1709 /* Resend the fist split part */
1710 submit_bio_noacct(split);
1711 wait_barrier(conf, false);
1713 div_u64_rem(bio_end, stripe_size, &remainder);
1715 split_size = bio_sectors(bio) - remainder;
1716 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1717 bio_chain(split, bio);
1718 allow_barrier(conf);
1719 /* Resend the second split part */
1720 submit_bio_noacct(bio);
1722 wait_barrier(conf, false);
1725 bio_start = bio->bi_iter.bi_sector;
1726 bio_end = bio_end_sector(bio);
1729 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1730 * One stripe contains the chunks from all member disk (one chunk from
1731 * one disk at the same HBA address). For layout detail, see 'man md 4'
1733 chunk = bio_start >> geo->chunk_shift;
1734 chunk *= geo->near_copies;
1735 first_stripe_index = chunk;
1736 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1737 if (geo->far_offset)
1738 first_stripe_index *= geo->far_copies;
1739 start_disk_offset = (bio_start & geo->chunk_mask) +
1740 (first_stripe_index << geo->chunk_shift);
1742 chunk = bio_end >> geo->chunk_shift;
1743 chunk *= geo->near_copies;
1744 last_stripe_index = chunk;
1745 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1746 if (geo->far_offset)
1747 last_stripe_index *= geo->far_copies;
1748 end_disk_offset = (bio_end & geo->chunk_mask) +
1749 (last_stripe_index << geo->chunk_shift);
1752 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1753 r10_bio->mddev = mddev;
1755 r10_bio->sectors = 0;
1756 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1757 wait_blocked_dev(mddev, r10_bio);
1760 * For far layout it needs more than one r10bio to cover all regions.
1761 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1762 * to record the discard bio. Other r10bio->master_bio record the first
1763 * r10bio. The first r10bio only release after all other r10bios finish.
1764 * The discard bio returns only first r10bio finishes
1767 r10_bio->master_bio = bio;
1768 set_bit(R10BIO_Discard, &r10_bio->state);
1770 first_r10bio = r10_bio;
1772 r10_bio->master_bio = (struct bio *)first_r10bio;
1775 * first select target devices under rcu_lock and
1776 * inc refcount on their rdev. Record them by setting
1780 for (disk = 0; disk < geo->raid_disks; disk++) {
1781 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1782 struct md_rdev *rrdev = rcu_dereference(
1783 conf->mirrors[disk].replacement);
1785 r10_bio->devs[disk].bio = NULL;
1786 r10_bio->devs[disk].repl_bio = NULL;
1788 if (rdev && (test_bit(Faulty, &rdev->flags)))
1790 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1792 if (!rdev && !rrdev)
1796 r10_bio->devs[disk].bio = bio;
1797 atomic_inc(&rdev->nr_pending);
1800 r10_bio->devs[disk].repl_bio = bio;
1801 atomic_inc(&rrdev->nr_pending);
1806 atomic_set(&r10_bio->remaining, 1);
1807 for (disk = 0; disk < geo->raid_disks; disk++) {
1808 sector_t dev_start, dev_end;
1809 struct bio *mbio, *rbio = NULL;
1812 * Now start to calculate the start and end address for each disk.
1813 * The space between dev_start and dev_end is the discard region.
1815 * For dev_start, it needs to consider three conditions:
1816 * 1st, the disk is before start_disk, you can imagine the disk in
1817 * the next stripe. So the dev_start is the start address of next
1819 * 2st, the disk is after start_disk, it means the disk is at the
1820 * same stripe of first disk
1821 * 3st, the first disk itself, we can use start_disk_offset directly
1823 if (disk < start_disk_index)
1824 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1825 else if (disk > start_disk_index)
1826 dev_start = first_stripe_index * mddev->chunk_sectors;
1828 dev_start = start_disk_offset;
1830 if (disk < end_disk_index)
1831 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1832 else if (disk > end_disk_index)
1833 dev_end = last_stripe_index * mddev->chunk_sectors;
1835 dev_end = end_disk_offset;
1838 * It only handles discard bio which size is >= stripe size, so
1839 * dev_end > dev_start all the time.
1840 * It doesn't need to use rcu lock to get rdev here. We already
1841 * add rdev->nr_pending in the first loop.
1843 if (r10_bio->devs[disk].bio) {
1844 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1845 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1847 mbio->bi_end_io = raid10_end_discard_request;
1848 mbio->bi_private = r10_bio;
1849 r10_bio->devs[disk].bio = mbio;
1850 r10_bio->devs[disk].devnum = disk;
1851 atomic_inc(&r10_bio->remaining);
1852 md_submit_discard_bio(mddev, rdev, mbio,
1853 dev_start + choose_data_offset(r10_bio, rdev),
1854 dev_end - dev_start);
1857 if (r10_bio->devs[disk].repl_bio) {
1858 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1859 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1861 rbio->bi_end_io = raid10_end_discard_request;
1862 rbio->bi_private = r10_bio;
1863 r10_bio->devs[disk].repl_bio = rbio;
1864 r10_bio->devs[disk].devnum = disk;
1865 atomic_inc(&r10_bio->remaining);
1866 md_submit_discard_bio(mddev, rrdev, rbio,
1867 dev_start + choose_data_offset(r10_bio, rrdev),
1868 dev_end - dev_start);
1873 if (!geo->far_offset && --far_copies) {
1874 first_stripe_index += geo->stride >> geo->chunk_shift;
1875 start_disk_offset += geo->stride;
1876 last_stripe_index += geo->stride >> geo->chunk_shift;
1877 end_disk_offset += geo->stride;
1878 atomic_inc(&first_r10bio->remaining);
1879 raid_end_discard_bio(r10_bio);
1880 wait_barrier(conf, false);
1884 raid_end_discard_bio(r10_bio);
1888 allow_barrier(conf);
1892 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1894 struct r10conf *conf = mddev->private;
1895 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1896 int chunk_sects = chunk_mask + 1;
1897 int sectors = bio_sectors(bio);
1899 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1900 && md_flush_request(mddev, bio))
1903 if (!md_write_start(mddev, bio))
1906 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1907 if (!raid10_handle_discard(mddev, bio))
1911 * If this request crosses a chunk boundary, we need to split
1914 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1915 sectors > chunk_sects
1916 && (conf->geo.near_copies < conf->geo.raid_disks
1917 || conf->prev.near_copies <
1918 conf->prev.raid_disks)))
1919 sectors = chunk_sects -
1920 (bio->bi_iter.bi_sector &
1922 __make_request(mddev, bio, sectors);
1924 /* In case raid10d snuck in to freeze_array */
1925 wake_up_barrier(conf);
1929 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1931 struct r10conf *conf = mddev->private;
1934 if (conf->geo.near_copies < conf->geo.raid_disks)
1935 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1936 if (conf->geo.near_copies > 1)
1937 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1938 if (conf->geo.far_copies > 1) {
1939 if (conf->geo.far_offset)
1940 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1942 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1943 if (conf->geo.far_set_size != conf->geo.raid_disks)
1944 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1946 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1947 conf->geo.raid_disks - mddev->degraded);
1949 for (i = 0; i < conf->geo.raid_disks; i++) {
1950 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1951 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1954 seq_printf(seq, "]");
1957 /* check if there are enough drives for
1958 * every block to appear on atleast one.
1959 * Don't consider the device numbered 'ignore'
1960 * as we might be about to remove it.
1962 static int _enough(struct r10conf *conf, int previous, int ignore)
1968 disks = conf->prev.raid_disks;
1969 ncopies = conf->prev.near_copies;
1971 disks = conf->geo.raid_disks;
1972 ncopies = conf->geo.near_copies;
1977 int n = conf->copies;
1981 struct md_rdev *rdev;
1982 if (this != ignore &&
1983 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1984 test_bit(In_sync, &rdev->flags))
1986 this = (this+1) % disks;
1990 first = (first + ncopies) % disks;
1991 } while (first != 0);
1998 static int enough(struct r10conf *conf, int ignore)
2000 /* when calling 'enough', both 'prev' and 'geo' must
2002 * This is ensured if ->reconfig_mutex or ->device_lock
2005 return _enough(conf, 0, ignore) &&
2006 _enough(conf, 1, ignore);
2010 * raid10_error() - RAID10 error handler.
2011 * @mddev: affected md device.
2012 * @rdev: member device to fail.
2014 * The routine acknowledges &rdev failure and determines new @mddev state.
2015 * If it failed, then:
2016 * - &MD_BROKEN flag is set in &mddev->flags.
2017 * Otherwise, it must be degraded:
2018 * - recovery is interrupted.
2019 * - &mddev->degraded is bumped.
2021 * @rdev is marked as &Faulty excluding case when array is failed and
2022 * &mddev->fail_last_dev is off.
2024 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
2026 struct r10conf *conf = mddev->private;
2027 unsigned long flags;
2029 spin_lock_irqsave(&conf->device_lock, flags);
2031 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
2032 set_bit(MD_BROKEN, &mddev->flags);
2034 if (!mddev->fail_last_dev) {
2035 spin_unlock_irqrestore(&conf->device_lock, flags);
2039 if (test_and_clear_bit(In_sync, &rdev->flags))
2042 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2043 set_bit(Blocked, &rdev->flags);
2044 set_bit(Faulty, &rdev->flags);
2045 set_mask_bits(&mddev->sb_flags, 0,
2046 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2047 spin_unlock_irqrestore(&conf->device_lock, flags);
2048 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2049 "md/raid10:%s: Operation continuing on %d devices.\n",
2050 mdname(mddev), rdev->bdev,
2051 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2054 static void print_conf(struct r10conf *conf)
2057 struct md_rdev *rdev;
2059 pr_debug("RAID10 conf printout:\n");
2061 pr_debug("(!conf)\n");
2064 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2065 conf->geo.raid_disks);
2067 /* This is only called with ->reconfix_mutex held, so
2068 * rcu protection of rdev is not needed */
2069 for (i = 0; i < conf->geo.raid_disks; i++) {
2070 rdev = conf->mirrors[i].rdev;
2072 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2073 i, !test_bit(In_sync, &rdev->flags),
2074 !test_bit(Faulty, &rdev->flags),
2079 static void close_sync(struct r10conf *conf)
2081 wait_barrier(conf, false);
2082 allow_barrier(conf);
2084 mempool_exit(&conf->r10buf_pool);
2087 static int raid10_spare_active(struct mddev *mddev)
2090 struct r10conf *conf = mddev->private;
2091 struct raid10_info *tmp;
2093 unsigned long flags;
2096 * Find all non-in_sync disks within the RAID10 configuration
2097 * and mark them in_sync
2099 for (i = 0; i < conf->geo.raid_disks; i++) {
2100 tmp = conf->mirrors + i;
2101 if (tmp->replacement
2102 && tmp->replacement->recovery_offset == MaxSector
2103 && !test_bit(Faulty, &tmp->replacement->flags)
2104 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2105 /* Replacement has just become active */
2107 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2110 /* Replaced device not technically faulty,
2111 * but we need to be sure it gets removed
2112 * and never re-added.
2114 set_bit(Faulty, &tmp->rdev->flags);
2115 sysfs_notify_dirent_safe(
2116 tmp->rdev->sysfs_state);
2118 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2119 } else if (tmp->rdev
2120 && tmp->rdev->recovery_offset == MaxSector
2121 && !test_bit(Faulty, &tmp->rdev->flags)
2122 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2124 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2127 spin_lock_irqsave(&conf->device_lock, flags);
2128 mddev->degraded -= count;
2129 spin_unlock_irqrestore(&conf->device_lock, flags);
2135 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2137 struct r10conf *conf = mddev->private;
2139 int mirror, repl_slot = -1;
2141 int last = conf->geo.raid_disks - 1;
2142 struct raid10_info *p;
2144 if (mddev->recovery_cp < MaxSector)
2145 /* only hot-add to in-sync arrays, as recovery is
2146 * very different from resync
2149 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2152 if (md_integrity_add_rdev(rdev, mddev))
2155 if (rdev->raid_disk >= 0)
2156 first = last = rdev->raid_disk;
2158 if (rdev->saved_raid_disk >= first &&
2159 rdev->saved_raid_disk < conf->geo.raid_disks &&
2160 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2161 mirror = rdev->saved_raid_disk;
2164 for ( ; mirror <= last ; mirror++) {
2165 p = &conf->mirrors[mirror];
2166 if (p->recovery_disabled == mddev->recovery_disabled)
2169 if (test_bit(WantReplacement, &p->rdev->flags) &&
2170 p->replacement == NULL && repl_slot < 0)
2176 disk_stack_limits(mddev->gendisk, rdev->bdev,
2177 rdev->data_offset << 9);
2179 p->head_position = 0;
2180 p->recovery_disabled = mddev->recovery_disabled - 1;
2181 rdev->raid_disk = mirror;
2183 if (rdev->saved_raid_disk != mirror)
2185 rcu_assign_pointer(p->rdev, rdev);
2189 if (err && repl_slot >= 0) {
2190 p = &conf->mirrors[repl_slot];
2191 clear_bit(In_sync, &rdev->flags);
2192 set_bit(Replacement, &rdev->flags);
2193 rdev->raid_disk = repl_slot;
2196 disk_stack_limits(mddev->gendisk, rdev->bdev,
2197 rdev->data_offset << 9);
2199 rcu_assign_pointer(p->replacement, rdev);
2206 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2208 struct r10conf *conf = mddev->private;
2210 int number = rdev->raid_disk;
2211 struct md_rdev **rdevp;
2212 struct raid10_info *p;
2215 if (unlikely(number >= mddev->raid_disks))
2217 p = conf->mirrors + number;
2218 if (rdev == p->rdev)
2220 else if (rdev == p->replacement)
2221 rdevp = &p->replacement;
2225 if (test_bit(In_sync, &rdev->flags) ||
2226 atomic_read(&rdev->nr_pending)) {
2230 /* Only remove non-faulty devices if recovery
2233 if (!test_bit(Faulty, &rdev->flags) &&
2234 mddev->recovery_disabled != p->recovery_disabled &&
2235 (!p->replacement || p->replacement == rdev) &&
2236 number < conf->geo.raid_disks &&
2242 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
2244 if (atomic_read(&rdev->nr_pending)) {
2245 /* lost the race, try later */
2251 if (p->replacement) {
2252 /* We must have just cleared 'rdev' */
2253 p->rdev = p->replacement;
2254 clear_bit(Replacement, &p->replacement->flags);
2255 smp_mb(); /* Make sure other CPUs may see both as identical
2256 * but will never see neither -- if they are careful.
2258 p->replacement = NULL;
2261 clear_bit(WantReplacement, &rdev->flags);
2262 err = md_integrity_register(mddev);
2270 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2272 struct r10conf *conf = r10_bio->mddev->private;
2274 if (!bio->bi_status)
2275 set_bit(R10BIO_Uptodate, &r10_bio->state);
2277 /* The write handler will notice the lack of
2278 * R10BIO_Uptodate and record any errors etc
2280 atomic_add(r10_bio->sectors,
2281 &conf->mirrors[d].rdev->corrected_errors);
2283 /* for reconstruct, we always reschedule after a read.
2284 * for resync, only after all reads
2286 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2287 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2288 atomic_dec_and_test(&r10_bio->remaining)) {
2289 /* we have read all the blocks,
2290 * do the comparison in process context in raid10d
2292 reschedule_retry(r10_bio);
2296 static void end_sync_read(struct bio *bio)
2298 struct r10bio *r10_bio = get_resync_r10bio(bio);
2299 struct r10conf *conf = r10_bio->mddev->private;
2300 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2302 __end_sync_read(r10_bio, bio, d);
2305 static void end_reshape_read(struct bio *bio)
2307 /* reshape read bio isn't allocated from r10buf_pool */
2308 struct r10bio *r10_bio = bio->bi_private;
2310 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2313 static void end_sync_request(struct r10bio *r10_bio)
2315 struct mddev *mddev = r10_bio->mddev;
2317 while (atomic_dec_and_test(&r10_bio->remaining)) {
2318 if (r10_bio->master_bio == NULL) {
2319 /* the primary of several recovery bios */
2320 sector_t s = r10_bio->sectors;
2321 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2322 test_bit(R10BIO_WriteError, &r10_bio->state))
2323 reschedule_retry(r10_bio);
2326 md_done_sync(mddev, s, 1);
2329 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2330 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2331 test_bit(R10BIO_WriteError, &r10_bio->state))
2332 reschedule_retry(r10_bio);
2340 static void end_sync_write(struct bio *bio)
2342 struct r10bio *r10_bio = get_resync_r10bio(bio);
2343 struct mddev *mddev = r10_bio->mddev;
2344 struct r10conf *conf = mddev->private;
2350 struct md_rdev *rdev = NULL;
2352 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2354 rdev = conf->mirrors[d].replacement;
2356 rdev = conf->mirrors[d].rdev;
2358 if (bio->bi_status) {
2360 md_error(mddev, rdev);
2362 set_bit(WriteErrorSeen, &rdev->flags);
2363 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2364 set_bit(MD_RECOVERY_NEEDED,
2365 &rdev->mddev->recovery);
2366 set_bit(R10BIO_WriteError, &r10_bio->state);
2368 } else if (is_badblock(rdev,
2369 r10_bio->devs[slot].addr,
2371 &first_bad, &bad_sectors))
2372 set_bit(R10BIO_MadeGood, &r10_bio->state);
2374 rdev_dec_pending(rdev, mddev);
2376 end_sync_request(r10_bio);
2380 * Note: sync and recover and handled very differently for raid10
2381 * This code is for resync.
2382 * For resync, we read through virtual addresses and read all blocks.
2383 * If there is any error, we schedule a write. The lowest numbered
2384 * drive is authoritative.
2385 * However requests come for physical address, so we need to map.
2386 * For every physical address there are raid_disks/copies virtual addresses,
2387 * which is always are least one, but is not necessarly an integer.
2388 * This means that a physical address can span multiple chunks, so we may
2389 * have to submit multiple io requests for a single sync request.
2392 * We check if all blocks are in-sync and only write to blocks that
2395 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2397 struct r10conf *conf = mddev->private;
2399 struct bio *tbio, *fbio;
2401 struct page **tpages, **fpages;
2403 atomic_set(&r10_bio->remaining, 1);
2405 /* find the first device with a block */
2406 for (i=0; i<conf->copies; i++)
2407 if (!r10_bio->devs[i].bio->bi_status)
2410 if (i == conf->copies)
2414 fbio = r10_bio->devs[i].bio;
2415 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2416 fbio->bi_iter.bi_idx = 0;
2417 fpages = get_resync_pages(fbio)->pages;
2419 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2420 /* now find blocks with errors */
2421 for (i=0 ; i < conf->copies ; i++) {
2423 struct md_rdev *rdev;
2424 struct resync_pages *rp;
2426 tbio = r10_bio->devs[i].bio;
2428 if (tbio->bi_end_io != end_sync_read)
2433 tpages = get_resync_pages(tbio)->pages;
2434 d = r10_bio->devs[i].devnum;
2435 rdev = conf->mirrors[d].rdev;
2436 if (!r10_bio->devs[i].bio->bi_status) {
2437 /* We know that the bi_io_vec layout is the same for
2438 * both 'first' and 'i', so we just compare them.
2439 * All vec entries are PAGE_SIZE;
2441 int sectors = r10_bio->sectors;
2442 for (j = 0; j < vcnt; j++) {
2443 int len = PAGE_SIZE;
2444 if (sectors < (len / 512))
2445 len = sectors * 512;
2446 if (memcmp(page_address(fpages[j]),
2447 page_address(tpages[j]),
2454 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2455 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2456 /* Don't fix anything. */
2458 } else if (test_bit(FailFast, &rdev->flags)) {
2459 /* Just give up on this device */
2460 md_error(rdev->mddev, rdev);
2463 /* Ok, we need to write this bio, either to correct an
2464 * inconsistency or to correct an unreadable block.
2465 * First we need to fixup bv_offset, bv_len and
2466 * bi_vecs, as the read request might have corrupted these
2468 rp = get_resync_pages(tbio);
2469 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2471 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2473 rp->raid_bio = r10_bio;
2474 tbio->bi_private = rp;
2475 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2476 tbio->bi_end_io = end_sync_write;
2478 bio_copy_data(tbio, fbio);
2480 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2481 atomic_inc(&r10_bio->remaining);
2482 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2484 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2485 tbio->bi_opf |= MD_FAILFAST;
2486 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2487 submit_bio_noacct(tbio);
2490 /* Now write out to any replacement devices
2493 for (i = 0; i < conf->copies; i++) {
2496 tbio = r10_bio->devs[i].repl_bio;
2497 if (!tbio || !tbio->bi_end_io)
2499 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2500 && r10_bio->devs[i].bio != fbio)
2501 bio_copy_data(tbio, fbio);
2502 d = r10_bio->devs[i].devnum;
2503 atomic_inc(&r10_bio->remaining);
2504 md_sync_acct(conf->mirrors[d].replacement->bdev,
2506 submit_bio_noacct(tbio);
2510 if (atomic_dec_and_test(&r10_bio->remaining)) {
2511 md_done_sync(mddev, r10_bio->sectors, 1);
2517 * Now for the recovery code.
2518 * Recovery happens across physical sectors.
2519 * We recover all non-is_sync drives by finding the virtual address of
2520 * each, and then choose a working drive that also has that virt address.
2521 * There is a separate r10_bio for each non-in_sync drive.
2522 * Only the first two slots are in use. The first for reading,
2523 * The second for writing.
2526 static void fix_recovery_read_error(struct r10bio *r10_bio)
2528 /* We got a read error during recovery.
2529 * We repeat the read in smaller page-sized sections.
2530 * If a read succeeds, write it to the new device or record
2531 * a bad block if we cannot.
2532 * If a read fails, record a bad block on both old and
2535 struct mddev *mddev = r10_bio->mddev;
2536 struct r10conf *conf = mddev->private;
2537 struct bio *bio = r10_bio->devs[0].bio;
2539 int sectors = r10_bio->sectors;
2541 int dr = r10_bio->devs[0].devnum;
2542 int dw = r10_bio->devs[1].devnum;
2543 struct page **pages = get_resync_pages(bio)->pages;
2547 struct md_rdev *rdev;
2551 if (s > (PAGE_SIZE>>9))
2554 rdev = conf->mirrors[dr].rdev;
2555 addr = r10_bio->devs[0].addr + sect,
2556 ok = sync_page_io(rdev,
2560 REQ_OP_READ, false);
2562 rdev = conf->mirrors[dw].rdev;
2563 addr = r10_bio->devs[1].addr + sect;
2564 ok = sync_page_io(rdev,
2568 REQ_OP_WRITE, false);
2570 set_bit(WriteErrorSeen, &rdev->flags);
2571 if (!test_and_set_bit(WantReplacement,
2573 set_bit(MD_RECOVERY_NEEDED,
2574 &rdev->mddev->recovery);
2578 /* We don't worry if we cannot set a bad block -
2579 * it really is bad so there is no loss in not
2582 rdev_set_badblocks(rdev, addr, s, 0);
2584 if (rdev != conf->mirrors[dw].rdev) {
2585 /* need bad block on destination too */
2586 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2587 addr = r10_bio->devs[1].addr + sect;
2588 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2590 /* just abort the recovery */
2591 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2594 conf->mirrors[dw].recovery_disabled
2595 = mddev->recovery_disabled;
2596 set_bit(MD_RECOVERY_INTR,
2609 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2611 struct r10conf *conf = mddev->private;
2613 struct bio *wbio = r10_bio->devs[1].bio;
2614 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2616 /* Need to test wbio2->bi_end_io before we call
2617 * submit_bio_noacct as if the former is NULL,
2618 * the latter is free to free wbio2.
2620 if (wbio2 && !wbio2->bi_end_io)
2623 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2624 fix_recovery_read_error(r10_bio);
2625 if (wbio->bi_end_io)
2626 end_sync_request(r10_bio);
2628 end_sync_request(r10_bio);
2633 * share the pages with the first bio
2634 * and submit the write request
2636 d = r10_bio->devs[1].devnum;
2637 if (wbio->bi_end_io) {
2638 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2639 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2640 submit_bio_noacct(wbio);
2643 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2644 md_sync_acct(conf->mirrors[d].replacement->bdev,
2645 bio_sectors(wbio2));
2646 submit_bio_noacct(wbio2);
2651 * Used by fix_read_error() to decay the per rdev read_errors.
2652 * We halve the read error count for every hour that has elapsed
2653 * since the last recorded read error.
2656 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2659 unsigned long hours_since_last;
2660 unsigned int read_errors = atomic_read(&rdev->read_errors);
2662 cur_time_mon = ktime_get_seconds();
2664 if (rdev->last_read_error == 0) {
2665 /* first time we've seen a read error */
2666 rdev->last_read_error = cur_time_mon;
2670 hours_since_last = (long)(cur_time_mon -
2671 rdev->last_read_error) / 3600;
2673 rdev->last_read_error = cur_time_mon;
2676 * if hours_since_last is > the number of bits in read_errors
2677 * just set read errors to 0. We do this to avoid
2678 * overflowing the shift of read_errors by hours_since_last.
2680 if (hours_since_last >= 8 * sizeof(read_errors))
2681 atomic_set(&rdev->read_errors, 0);
2683 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2686 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2687 int sectors, struct page *page, enum req_op op)
2692 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2693 && (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2695 if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2698 if (op == REQ_OP_WRITE) {
2699 set_bit(WriteErrorSeen, &rdev->flags);
2700 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2701 set_bit(MD_RECOVERY_NEEDED,
2702 &rdev->mddev->recovery);
2704 /* need to record an error - either for the block or the device */
2705 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2706 md_error(rdev->mddev, rdev);
2711 * This is a kernel thread which:
2713 * 1. Retries failed read operations on working mirrors.
2714 * 2. Updates the raid superblock when problems encounter.
2715 * 3. Performs writes following reads for array synchronising.
2718 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2720 int sect = 0; /* Offset from r10_bio->sector */
2721 int sectors = r10_bio->sectors;
2722 struct md_rdev *rdev;
2723 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2724 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2726 /* still own a reference to this rdev, so it cannot
2727 * have been cleared recently.
2729 rdev = conf->mirrors[d].rdev;
2731 if (test_bit(Faulty, &rdev->flags))
2732 /* drive has already been failed, just ignore any
2733 more fix_read_error() attempts */
2736 check_decay_read_errors(mddev, rdev);
2737 atomic_inc(&rdev->read_errors);
2738 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2739 pr_notice("md/raid10:%s: %pg: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2740 mdname(mddev), rdev->bdev,
2741 atomic_read(&rdev->read_errors), max_read_errors);
2742 pr_notice("md/raid10:%s: %pg: Failing raid device\n",
2743 mdname(mddev), rdev->bdev);
2744 md_error(mddev, rdev);
2745 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2751 int sl = r10_bio->read_slot;
2755 if (s > (PAGE_SIZE>>9))
2763 d = r10_bio->devs[sl].devnum;
2764 rdev = rcu_dereference(conf->mirrors[d].rdev);
2766 test_bit(In_sync, &rdev->flags) &&
2767 !test_bit(Faulty, &rdev->flags) &&
2768 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2769 &first_bad, &bad_sectors) == 0) {
2770 atomic_inc(&rdev->nr_pending);
2772 success = sync_page_io(rdev,
2773 r10_bio->devs[sl].addr +
2777 REQ_OP_READ, false);
2778 rdev_dec_pending(rdev, mddev);
2784 if (sl == conf->copies)
2786 } while (!success && sl != r10_bio->read_slot);
2790 /* Cannot read from anywhere, just mark the block
2791 * as bad on the first device to discourage future
2794 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2795 rdev = conf->mirrors[dn].rdev;
2797 if (!rdev_set_badblocks(
2799 r10_bio->devs[r10_bio->read_slot].addr
2802 md_error(mddev, rdev);
2803 r10_bio->devs[r10_bio->read_slot].bio
2810 /* write it back and re-read */
2812 while (sl != r10_bio->read_slot) {
2816 d = r10_bio->devs[sl].devnum;
2817 rdev = rcu_dereference(conf->mirrors[d].rdev);
2819 test_bit(Faulty, &rdev->flags) ||
2820 !test_bit(In_sync, &rdev->flags))
2823 atomic_inc(&rdev->nr_pending);
2825 if (r10_sync_page_io(rdev,
2826 r10_bio->devs[sl].addr +
2828 s, conf->tmppage, REQ_OP_WRITE)
2830 /* Well, this device is dead */
2831 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2833 (unsigned long long)(
2835 choose_data_offset(r10_bio,
2838 pr_notice("md/raid10:%s: %pg: failing drive\n",
2842 rdev_dec_pending(rdev, mddev);
2846 while (sl != r10_bio->read_slot) {
2850 d = r10_bio->devs[sl].devnum;
2851 rdev = rcu_dereference(conf->mirrors[d].rdev);
2853 test_bit(Faulty, &rdev->flags) ||
2854 !test_bit(In_sync, &rdev->flags))
2857 atomic_inc(&rdev->nr_pending);
2859 switch (r10_sync_page_io(rdev,
2860 r10_bio->devs[sl].addr +
2862 s, conf->tmppage, REQ_OP_READ)) {
2864 /* Well, this device is dead */
2865 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2867 (unsigned long long)(
2869 choose_data_offset(r10_bio, rdev)),
2871 pr_notice("md/raid10:%s: %pg: failing drive\n",
2876 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2878 (unsigned long long)(
2880 choose_data_offset(r10_bio, rdev)),
2882 atomic_add(s, &rdev->corrected_errors);
2885 rdev_dec_pending(rdev, mddev);
2895 static int narrow_write_error(struct r10bio *r10_bio, int i)
2897 struct bio *bio = r10_bio->master_bio;
2898 struct mddev *mddev = r10_bio->mddev;
2899 struct r10conf *conf = mddev->private;
2900 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2901 /* bio has the data to be written to slot 'i' where
2902 * we just recently had a write error.
2903 * We repeatedly clone the bio and trim down to one block,
2904 * then try the write. Where the write fails we record
2906 * It is conceivable that the bio doesn't exactly align with
2907 * blocks. We must handle this.
2909 * We currently own a reference to the rdev.
2915 int sect_to_write = r10_bio->sectors;
2918 if (rdev->badblocks.shift < 0)
2921 block_sectors = roundup(1 << rdev->badblocks.shift,
2922 bdev_logical_block_size(rdev->bdev) >> 9);
2923 sector = r10_bio->sector;
2924 sectors = ((r10_bio->sector + block_sectors)
2925 & ~(sector_t)(block_sectors - 1))
2928 while (sect_to_write) {
2931 if (sectors > sect_to_write)
2932 sectors = sect_to_write;
2933 /* Write at 'sector' for 'sectors' */
2934 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2936 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2937 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2938 wbio->bi_iter.bi_sector = wsector +
2939 choose_data_offset(r10_bio, rdev);
2940 wbio->bi_opf = REQ_OP_WRITE;
2942 if (submit_bio_wait(wbio) < 0)
2944 ok = rdev_set_badblocks(rdev, wsector,
2949 sect_to_write -= sectors;
2951 sectors = block_sectors;
2956 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2958 int slot = r10_bio->read_slot;
2960 struct r10conf *conf = mddev->private;
2961 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2963 /* we got a read error. Maybe the drive is bad. Maybe just
2964 * the block and we can fix it.
2965 * We freeze all other IO, and try reading the block from
2966 * other devices. When we find one, we re-write
2967 * and check it that fixes the read error.
2968 * This is all done synchronously while the array is
2971 bio = r10_bio->devs[slot].bio;
2973 r10_bio->devs[slot].bio = NULL;
2976 r10_bio->devs[slot].bio = IO_BLOCKED;
2977 else if (!test_bit(FailFast, &rdev->flags)) {
2978 freeze_array(conf, 1);
2979 fix_read_error(conf, mddev, r10_bio);
2980 unfreeze_array(conf);
2982 md_error(mddev, rdev);
2984 rdev_dec_pending(rdev, mddev);
2986 raid10_read_request(mddev, r10_bio->master_bio, r10_bio, false);
2988 * allow_barrier after re-submit to ensure no sync io
2989 * can be issued while regular io pending.
2991 allow_barrier(conf);
2994 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2996 /* Some sort of write request has finished and it
2997 * succeeded in writing where we thought there was a
2998 * bad block. So forget the bad block.
2999 * Or possibly if failed and we need to record
3003 struct md_rdev *rdev;
3005 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
3006 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
3007 for (m = 0; m < conf->copies; m++) {
3008 int dev = r10_bio->devs[m].devnum;
3009 rdev = conf->mirrors[dev].rdev;
3010 if (r10_bio->devs[m].bio == NULL ||
3011 r10_bio->devs[m].bio->bi_end_io == NULL)
3013 if (!r10_bio->devs[m].bio->bi_status) {
3014 rdev_clear_badblocks(
3016 r10_bio->devs[m].addr,
3017 r10_bio->sectors, 0);
3019 if (!rdev_set_badblocks(
3021 r10_bio->devs[m].addr,
3022 r10_bio->sectors, 0))
3023 md_error(conf->mddev, rdev);
3025 rdev = conf->mirrors[dev].replacement;
3026 if (r10_bio->devs[m].repl_bio == NULL ||
3027 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
3030 if (!r10_bio->devs[m].repl_bio->bi_status) {
3031 rdev_clear_badblocks(
3033 r10_bio->devs[m].addr,
3034 r10_bio->sectors, 0);
3036 if (!rdev_set_badblocks(
3038 r10_bio->devs[m].addr,
3039 r10_bio->sectors, 0))
3040 md_error(conf->mddev, rdev);
3046 for (m = 0; m < conf->copies; m++) {
3047 int dev = r10_bio->devs[m].devnum;
3048 struct bio *bio = r10_bio->devs[m].bio;
3049 rdev = conf->mirrors[dev].rdev;
3050 if (bio == IO_MADE_GOOD) {
3051 rdev_clear_badblocks(
3053 r10_bio->devs[m].addr,
3054 r10_bio->sectors, 0);
3055 rdev_dec_pending(rdev, conf->mddev);
3056 } else if (bio != NULL && bio->bi_status) {
3058 if (!narrow_write_error(r10_bio, m)) {
3059 md_error(conf->mddev, rdev);
3060 set_bit(R10BIO_Degraded,
3063 rdev_dec_pending(rdev, conf->mddev);
3065 bio = r10_bio->devs[m].repl_bio;
3066 rdev = conf->mirrors[dev].replacement;
3067 if (rdev && bio == IO_MADE_GOOD) {
3068 rdev_clear_badblocks(
3070 r10_bio->devs[m].addr,
3071 r10_bio->sectors, 0);
3072 rdev_dec_pending(rdev, conf->mddev);
3076 spin_lock_irq(&conf->device_lock);
3077 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
3079 spin_unlock_irq(&conf->device_lock);
3081 * In case freeze_array() is waiting for condition
3082 * nr_pending == nr_queued + extra to be true.
3084 wake_up(&conf->wait_barrier);
3085 md_wakeup_thread(conf->mddev->thread);
3087 if (test_bit(R10BIO_WriteError,
3089 close_write(r10_bio);
3090 raid_end_bio_io(r10_bio);
3095 static void raid10d(struct md_thread *thread)
3097 struct mddev *mddev = thread->mddev;
3098 struct r10bio *r10_bio;
3099 unsigned long flags;
3100 struct r10conf *conf = mddev->private;
3101 struct list_head *head = &conf->retry_list;
3102 struct blk_plug plug;
3104 md_check_recovery(mddev);
3106 if (!list_empty_careful(&conf->bio_end_io_list) &&
3107 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3109 spin_lock_irqsave(&conf->device_lock, flags);
3110 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3111 while (!list_empty(&conf->bio_end_io_list)) {
3112 list_move(conf->bio_end_io_list.prev, &tmp);
3116 spin_unlock_irqrestore(&conf->device_lock, flags);
3117 while (!list_empty(&tmp)) {
3118 r10_bio = list_first_entry(&tmp, struct r10bio,
3120 list_del(&r10_bio->retry_list);
3121 if (mddev->degraded)
3122 set_bit(R10BIO_Degraded, &r10_bio->state);
3124 if (test_bit(R10BIO_WriteError,
3126 close_write(r10_bio);
3127 raid_end_bio_io(r10_bio);
3131 blk_start_plug(&plug);
3134 flush_pending_writes(conf);
3136 spin_lock_irqsave(&conf->device_lock, flags);
3137 if (list_empty(head)) {
3138 spin_unlock_irqrestore(&conf->device_lock, flags);
3141 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3142 list_del(head->prev);
3144 spin_unlock_irqrestore(&conf->device_lock, flags);
3146 mddev = r10_bio->mddev;
3147 conf = mddev->private;
3148 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3149 test_bit(R10BIO_WriteError, &r10_bio->state))
3150 handle_write_completed(conf, r10_bio);
3151 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3152 reshape_request_write(mddev, r10_bio);
3153 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3154 sync_request_write(mddev, r10_bio);
3155 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3156 recovery_request_write(mddev, r10_bio);
3157 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3158 handle_read_error(mddev, r10_bio);
3163 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3164 md_check_recovery(mddev);
3166 blk_finish_plug(&plug);
3169 static int init_resync(struct r10conf *conf)
3173 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3174 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3175 conf->have_replacement = 0;
3176 for (i = 0; i < conf->geo.raid_disks; i++)
3177 if (conf->mirrors[i].replacement)
3178 conf->have_replacement = 1;
3179 ret = mempool_init(&conf->r10buf_pool, buffs,
3180 r10buf_pool_alloc, r10buf_pool_free, conf);
3183 conf->next_resync = 0;
3187 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3189 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3190 struct rsync_pages *rp;
3195 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3196 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3197 nalloc = conf->copies; /* resync */
3199 nalloc = 2; /* recovery */
3201 for (i = 0; i < nalloc; i++) {
3202 bio = r10bio->devs[i].bio;
3203 rp = bio->bi_private;
3204 bio_reset(bio, NULL, 0);
3205 bio->bi_private = rp;
3206 bio = r10bio->devs[i].repl_bio;
3208 rp = bio->bi_private;
3209 bio_reset(bio, NULL, 0);
3210 bio->bi_private = rp;
3217 * Set cluster_sync_high since we need other nodes to add the
3218 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3220 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3222 sector_t window_size;
3223 int extra_chunk, chunks;
3226 * First, here we define "stripe" as a unit which across
3227 * all member devices one time, so we get chunks by use
3228 * raid_disks / near_copies. Otherwise, if near_copies is
3229 * close to raid_disks, then resync window could increases
3230 * linearly with the increase of raid_disks, which means
3231 * we will suspend a really large IO window while it is not
3232 * necessary. If raid_disks is not divisible by near_copies,
3233 * an extra chunk is needed to ensure the whole "stripe" is
3237 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3238 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3242 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3245 * At least use a 32M window to align with raid1's resync window
3247 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3248 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3250 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3254 * perform a "sync" on one "block"
3256 * We need to make sure that no normal I/O request - particularly write
3257 * requests - conflict with active sync requests.
3259 * This is achieved by tracking pending requests and a 'barrier' concept
3260 * that can be installed to exclude normal IO requests.
3262 * Resync and recovery are handled very differently.
3263 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3265 * For resync, we iterate over virtual addresses, read all copies,
3266 * and update if there are differences. If only one copy is live,
3268 * For recovery, we iterate over physical addresses, read a good
3269 * value for each non-in_sync drive, and over-write.
3271 * So, for recovery we may have several outstanding complex requests for a
3272 * given address, one for each out-of-sync device. We model this by allocating
3273 * a number of r10_bio structures, one for each out-of-sync device.
3274 * As we setup these structures, we collect all bio's together into a list
3275 * which we then process collectively to add pages, and then process again
3276 * to pass to submit_bio_noacct.
3278 * The r10_bio structures are linked using a borrowed master_bio pointer.
3279 * This link is counted in ->remaining. When the r10_bio that points to NULL
3280 * has its remaining count decremented to 0, the whole complex operation
3285 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3288 struct r10conf *conf = mddev->private;
3289 struct r10bio *r10_bio;
3290 struct bio *biolist = NULL, *bio;
3291 sector_t max_sector, nr_sectors;
3294 sector_t sync_blocks;
3295 sector_t sectors_skipped = 0;
3296 int chunks_skipped = 0;
3297 sector_t chunk_mask = conf->geo.chunk_mask;
3299 int error_disk = -1;
3302 * Allow skipping a full rebuild for incremental assembly
3303 * of a clean array, like RAID1 does.
3305 if (mddev->bitmap == NULL &&
3306 mddev->recovery_cp == MaxSector &&
3307 mddev->reshape_position == MaxSector &&
3308 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3309 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3310 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3311 conf->fullsync == 0) {
3313 return mddev->dev_sectors - sector_nr;
3316 if (!mempool_initialized(&conf->r10buf_pool))
3317 if (init_resync(conf))
3321 max_sector = mddev->dev_sectors;
3322 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3323 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3324 max_sector = mddev->resync_max_sectors;
3325 if (sector_nr >= max_sector) {
3326 conf->cluster_sync_low = 0;
3327 conf->cluster_sync_high = 0;
3329 /* If we aborted, we need to abort the
3330 * sync on the 'current' bitmap chucks (there can
3331 * be several when recovering multiple devices).
3332 * as we may have started syncing it but not finished.
3333 * We can find the current address in
3334 * mddev->curr_resync, but for recovery,
3335 * we need to convert that to several
3336 * virtual addresses.
3338 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3344 if (mddev->curr_resync < max_sector) { /* aborted */
3345 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3346 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3348 else for (i = 0; i < conf->geo.raid_disks; i++) {
3350 raid10_find_virt(conf, mddev->curr_resync, i);
3351 md_bitmap_end_sync(mddev->bitmap, sect,
3355 /* completed sync */
3356 if ((!mddev->bitmap || conf->fullsync)
3357 && conf->have_replacement
3358 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3359 /* Completed a full sync so the replacements
3360 * are now fully recovered.
3363 for (i = 0; i < conf->geo.raid_disks; i++) {
3364 struct md_rdev *rdev =
3365 rcu_dereference(conf->mirrors[i].replacement);
3367 rdev->recovery_offset = MaxSector;
3373 md_bitmap_close_sync(mddev->bitmap);
3376 return sectors_skipped;
3379 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3380 return reshape_request(mddev, sector_nr, skipped);
3382 if (chunks_skipped >= conf->geo.raid_disks) {
3383 pr_err("md/raid10:%s: %s fails\n", mdname(mddev),
3384 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
3385 if (error_disk >= 0 &&
3386 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3388 * recovery fails, set mirrors.recovery_disabled,
3389 * device shouldn't be added to there.
3391 conf->mirrors[error_disk].recovery_disabled =
3392 mddev->recovery_disabled;
3396 * if there has been nothing to do on any drive,
3397 * then there is nothing to do at all.
3400 return (max_sector - sector_nr) + sectors_skipped;
3403 if (max_sector > mddev->resync_max)
3404 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3406 /* make sure whole request will fit in a chunk - if chunks
3409 if (conf->geo.near_copies < conf->geo.raid_disks &&
3410 max_sector > (sector_nr | chunk_mask))
3411 max_sector = (sector_nr | chunk_mask) + 1;
3414 * If there is non-resync activity waiting for a turn, then let it
3415 * though before starting on this new sync request.
3417 if (conf->nr_waiting)
3418 schedule_timeout_uninterruptible(1);
3420 /* Again, very different code for resync and recovery.
3421 * Both must result in an r10bio with a list of bios that
3422 * have bi_end_io, bi_sector, bi_bdev set,
3423 * and bi_private set to the r10bio.
3424 * For recovery, we may actually create several r10bios
3425 * with 2 bios in each, that correspond to the bios in the main one.
3426 * In this case, the subordinate r10bios link back through a
3427 * borrowed master_bio pointer, and the counter in the master
3428 * includes a ref from each subordinate.
3430 /* First, we decide what to do and set ->bi_end_io
3431 * To end_sync_read if we want to read, and
3432 * end_sync_write if we will want to write.
3435 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3436 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3437 /* recovery... the complicated one */
3441 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3447 struct raid10_info *mirror = &conf->mirrors[i];
3448 struct md_rdev *mrdev, *mreplace;
3451 mrdev = rcu_dereference(mirror->rdev);
3452 mreplace = rcu_dereference(mirror->replacement);
3454 if (mrdev && (test_bit(Faulty, &mrdev->flags) ||
3455 test_bit(In_sync, &mrdev->flags)))
3457 if (mreplace && test_bit(Faulty, &mreplace->flags))
3460 if (!mrdev && !mreplace) {
3466 /* want to reconstruct this device */
3468 sect = raid10_find_virt(conf, sector_nr, i);
3469 if (sect >= mddev->resync_max_sectors) {
3470 /* last stripe is not complete - don't
3471 * try to recover this sector.
3476 /* Unless we are doing a full sync, or a replacement
3477 * we only need to recover the block if it is set in
3480 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3482 if (sync_blocks < max_sync)
3483 max_sync = sync_blocks;
3487 /* yep, skip the sync_blocks here, but don't assume
3488 * that there will never be anything to do here
3490 chunks_skipped = -1;
3495 atomic_inc(&mrdev->nr_pending);
3497 atomic_inc(&mreplace->nr_pending);
3500 r10_bio = raid10_alloc_init_r10buf(conf);
3502 raise_barrier(conf, rb2 != NULL);
3503 atomic_set(&r10_bio->remaining, 0);
3505 r10_bio->master_bio = (struct bio*)rb2;
3507 atomic_inc(&rb2->remaining);
3508 r10_bio->mddev = mddev;
3509 set_bit(R10BIO_IsRecover, &r10_bio->state);
3510 r10_bio->sector = sect;
3512 raid10_find_phys(conf, r10_bio);
3514 /* Need to check if the array will still be
3518 for (j = 0; j < conf->geo.raid_disks; j++) {
3519 struct md_rdev *rdev = rcu_dereference(
3520 conf->mirrors[j].rdev);
3521 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3527 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3528 &sync_blocks, still_degraded);
3531 for (j=0; j<conf->copies;j++) {
3533 int d = r10_bio->devs[j].devnum;
3534 sector_t from_addr, to_addr;
3535 struct md_rdev *rdev =
3536 rcu_dereference(conf->mirrors[d].rdev);
3537 sector_t sector, first_bad;
3540 !test_bit(In_sync, &rdev->flags))
3542 /* This is where we read from */
3544 sector = r10_bio->devs[j].addr;
3546 if (is_badblock(rdev, sector, max_sync,
3547 &first_bad, &bad_sectors)) {
3548 if (first_bad > sector)
3549 max_sync = first_bad - sector;
3551 bad_sectors -= (sector
3553 if (max_sync > bad_sectors)
3554 max_sync = bad_sectors;
3558 bio = r10_bio->devs[0].bio;
3559 bio->bi_next = biolist;
3561 bio->bi_end_io = end_sync_read;
3562 bio->bi_opf = REQ_OP_READ;
3563 if (test_bit(FailFast, &rdev->flags))
3564 bio->bi_opf |= MD_FAILFAST;
3565 from_addr = r10_bio->devs[j].addr;
3566 bio->bi_iter.bi_sector = from_addr +
3568 bio_set_dev(bio, rdev->bdev);
3569 atomic_inc(&rdev->nr_pending);
3570 /* and we write to 'i' (if not in_sync) */
3572 for (k=0; k<conf->copies; k++)
3573 if (r10_bio->devs[k].devnum == i)
3575 BUG_ON(k == conf->copies);
3576 to_addr = r10_bio->devs[k].addr;
3577 r10_bio->devs[0].devnum = d;
3578 r10_bio->devs[0].addr = from_addr;
3579 r10_bio->devs[1].devnum = i;
3580 r10_bio->devs[1].addr = to_addr;
3583 bio = r10_bio->devs[1].bio;
3584 bio->bi_next = biolist;
3586 bio->bi_end_io = end_sync_write;
3587 bio->bi_opf = REQ_OP_WRITE;
3588 bio->bi_iter.bi_sector = to_addr
3589 + mrdev->data_offset;
3590 bio_set_dev(bio, mrdev->bdev);
3591 atomic_inc(&r10_bio->remaining);
3593 r10_bio->devs[1].bio->bi_end_io = NULL;
3595 /* and maybe write to replacement */
3596 bio = r10_bio->devs[1].repl_bio;
3598 bio->bi_end_io = NULL;
3599 /* Note: if replace is not NULL, then bio
3600 * cannot be NULL as r10buf_pool_alloc will
3601 * have allocated it.
3605 bio->bi_next = biolist;
3607 bio->bi_end_io = end_sync_write;
3608 bio->bi_opf = REQ_OP_WRITE;
3609 bio->bi_iter.bi_sector = to_addr +
3610 mreplace->data_offset;
3611 bio_set_dev(bio, mreplace->bdev);
3612 atomic_inc(&r10_bio->remaining);
3616 if (j == conf->copies) {
3617 /* Cannot recover, so abort the recovery or
3618 * record a bad block */
3620 /* problem is that there are bad blocks
3621 * on other device(s)
3624 for (k = 0; k < conf->copies; k++)
3625 if (r10_bio->devs[k].devnum == i)
3627 if (mrdev && !test_bit(In_sync,
3629 && !rdev_set_badblocks(
3631 r10_bio->devs[k].addr,
3635 !rdev_set_badblocks(
3637 r10_bio->devs[k].addr,
3642 if (!test_and_set_bit(MD_RECOVERY_INTR,
3644 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3646 mirror->recovery_disabled
3647 = mddev->recovery_disabled;
3653 atomic_dec(&rb2->remaining);
3656 rdev_dec_pending(mrdev, mddev);
3658 rdev_dec_pending(mreplace, mddev);
3662 rdev_dec_pending(mrdev, mddev);
3664 rdev_dec_pending(mreplace, mddev);
3665 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3666 /* Only want this if there is elsewhere to
3667 * read from. 'j' is currently the first
3671 for (; j < conf->copies; j++) {
3672 int d = r10_bio->devs[j].devnum;
3673 if (conf->mirrors[d].rdev &&
3675 &conf->mirrors[d].rdev->flags))
3679 r10_bio->devs[0].bio->bi_opf
3683 if (biolist == NULL) {
3685 struct r10bio *rb2 = r10_bio;
3686 r10_bio = (struct r10bio*) rb2->master_bio;
3687 rb2->master_bio = NULL;
3693 /* resync. Schedule a read for every block at this virt offset */
3697 * Since curr_resync_completed could probably not update in
3698 * time, and we will set cluster_sync_low based on it.
3699 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3700 * safety reason, which ensures curr_resync_completed is
3701 * updated in bitmap_cond_end_sync.
3703 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3704 mddev_is_clustered(mddev) &&
3705 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3707 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3708 &sync_blocks, mddev->degraded) &&
3709 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3710 &mddev->recovery)) {
3711 /* We can skip this block */
3713 return sync_blocks + sectors_skipped;
3715 if (sync_blocks < max_sync)
3716 max_sync = sync_blocks;
3717 r10_bio = raid10_alloc_init_r10buf(conf);
3720 r10_bio->mddev = mddev;
3721 atomic_set(&r10_bio->remaining, 0);
3722 raise_barrier(conf, 0);
3723 conf->next_resync = sector_nr;
3725 r10_bio->master_bio = NULL;
3726 r10_bio->sector = sector_nr;
3727 set_bit(R10BIO_IsSync, &r10_bio->state);
3728 raid10_find_phys(conf, r10_bio);
3729 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3731 for (i = 0; i < conf->copies; i++) {
3732 int d = r10_bio->devs[i].devnum;
3733 sector_t first_bad, sector;
3735 struct md_rdev *rdev;
3737 if (r10_bio->devs[i].repl_bio)
3738 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3740 bio = r10_bio->devs[i].bio;
3741 bio->bi_status = BLK_STS_IOERR;
3743 rdev = rcu_dereference(conf->mirrors[d].rdev);
3744 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3748 sector = r10_bio->devs[i].addr;
3749 if (is_badblock(rdev, sector, max_sync,
3750 &first_bad, &bad_sectors)) {
3751 if (first_bad > sector)
3752 max_sync = first_bad - sector;
3754 bad_sectors -= (sector - first_bad);
3755 if (max_sync > bad_sectors)
3756 max_sync = bad_sectors;
3761 atomic_inc(&rdev->nr_pending);
3762 atomic_inc(&r10_bio->remaining);
3763 bio->bi_next = biolist;
3765 bio->bi_end_io = end_sync_read;
3766 bio->bi_opf = REQ_OP_READ;
3767 if (test_bit(FailFast, &rdev->flags))
3768 bio->bi_opf |= MD_FAILFAST;
3769 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3770 bio_set_dev(bio, rdev->bdev);
3773 rdev = rcu_dereference(conf->mirrors[d].replacement);
3774 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3778 atomic_inc(&rdev->nr_pending);
3780 /* Need to set up for writing to the replacement */
3781 bio = r10_bio->devs[i].repl_bio;
3782 bio->bi_status = BLK_STS_IOERR;
3784 sector = r10_bio->devs[i].addr;
3785 bio->bi_next = biolist;
3787 bio->bi_end_io = end_sync_write;
3788 bio->bi_opf = REQ_OP_WRITE;
3789 if (test_bit(FailFast, &rdev->flags))
3790 bio->bi_opf |= MD_FAILFAST;
3791 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3792 bio_set_dev(bio, rdev->bdev);
3798 for (i=0; i<conf->copies; i++) {
3799 int d = r10_bio->devs[i].devnum;
3800 if (r10_bio->devs[i].bio->bi_end_io)
3801 rdev_dec_pending(conf->mirrors[d].rdev,
3803 if (r10_bio->devs[i].repl_bio &&
3804 r10_bio->devs[i].repl_bio->bi_end_io)
3806 conf->mirrors[d].replacement,
3816 if (sector_nr + max_sync < max_sector)
3817 max_sector = sector_nr + max_sync;
3820 int len = PAGE_SIZE;
3821 if (sector_nr + (len>>9) > max_sector)
3822 len = (max_sector - sector_nr) << 9;
3825 for (bio= biolist ; bio ; bio=bio->bi_next) {
3826 struct resync_pages *rp = get_resync_pages(bio);
3827 page = resync_fetch_page(rp, page_idx);
3828 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
3829 bio->bi_status = BLK_STS_RESOURCE;
3834 nr_sectors += len>>9;
3835 sector_nr += len>>9;
3836 } while (++page_idx < RESYNC_PAGES);
3837 r10_bio->sectors = nr_sectors;
3839 if (mddev_is_clustered(mddev) &&
3840 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3841 /* It is resync not recovery */
3842 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3843 conf->cluster_sync_low = mddev->curr_resync_completed;
3844 raid10_set_cluster_sync_high(conf);
3845 /* Send resync message */
3846 md_cluster_ops->resync_info_update(mddev,
3847 conf->cluster_sync_low,
3848 conf->cluster_sync_high);
3850 } else if (mddev_is_clustered(mddev)) {
3851 /* This is recovery not resync */
3852 sector_t sect_va1, sect_va2;
3853 bool broadcast_msg = false;
3855 for (i = 0; i < conf->geo.raid_disks; i++) {
3857 * sector_nr is a device address for recovery, so we
3858 * need translate it to array address before compare
3859 * with cluster_sync_high.
3861 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3863 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3864 broadcast_msg = true;
3866 * curr_resync_completed is similar as
3867 * sector_nr, so make the translation too.
3869 sect_va2 = raid10_find_virt(conf,
3870 mddev->curr_resync_completed, i);
3872 if (conf->cluster_sync_low == 0 ||
3873 conf->cluster_sync_low > sect_va2)
3874 conf->cluster_sync_low = sect_va2;
3877 if (broadcast_msg) {
3878 raid10_set_cluster_sync_high(conf);
3879 md_cluster_ops->resync_info_update(mddev,
3880 conf->cluster_sync_low,
3881 conf->cluster_sync_high);
3887 biolist = biolist->bi_next;
3889 bio->bi_next = NULL;
3890 r10_bio = get_resync_r10bio(bio);
3891 r10_bio->sectors = nr_sectors;
3893 if (bio->bi_end_io == end_sync_read) {
3894 md_sync_acct_bio(bio, nr_sectors);
3896 submit_bio_noacct(bio);
3900 if (sectors_skipped)
3901 /* pretend they weren't skipped, it makes
3902 * no important difference in this case
3904 md_done_sync(mddev, sectors_skipped, 1);
3906 return sectors_skipped + nr_sectors;
3908 /* There is nowhere to write, so all non-sync
3909 * drives must be failed or in resync, all drives
3910 * have a bad block, so try the next chunk...
3912 if (sector_nr + max_sync < max_sector)
3913 max_sector = sector_nr + max_sync;
3915 sectors_skipped += (max_sector - sector_nr);
3917 sector_nr = max_sector;
3922 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3925 struct r10conf *conf = mddev->private;
3928 raid_disks = min(conf->geo.raid_disks,
3929 conf->prev.raid_disks);
3931 sectors = conf->dev_sectors;
3933 size = sectors >> conf->geo.chunk_shift;
3934 sector_div(size, conf->geo.far_copies);
3935 size = size * raid_disks;
3936 sector_div(size, conf->geo.near_copies);
3938 return size << conf->geo.chunk_shift;
3941 static void calc_sectors(struct r10conf *conf, sector_t size)
3943 /* Calculate the number of sectors-per-device that will
3944 * actually be used, and set conf->dev_sectors and
3948 size = size >> conf->geo.chunk_shift;
3949 sector_div(size, conf->geo.far_copies);
3950 size = size * conf->geo.raid_disks;
3951 sector_div(size, conf->geo.near_copies);
3952 /* 'size' is now the number of chunks in the array */
3953 /* calculate "used chunks per device" */
3954 size = size * conf->copies;
3956 /* We need to round up when dividing by raid_disks to
3957 * get the stride size.
3959 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3961 conf->dev_sectors = size << conf->geo.chunk_shift;
3963 if (conf->geo.far_offset)
3964 conf->geo.stride = 1 << conf->geo.chunk_shift;
3966 sector_div(size, conf->geo.far_copies);
3967 conf->geo.stride = size << conf->geo.chunk_shift;
3971 enum geo_type {geo_new, geo_old, geo_start};
3972 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3975 int layout, chunk, disks;
3978 layout = mddev->layout;
3979 chunk = mddev->chunk_sectors;
3980 disks = mddev->raid_disks - mddev->delta_disks;
3983 layout = mddev->new_layout;
3984 chunk = mddev->new_chunk_sectors;
3985 disks = mddev->raid_disks;
3987 default: /* avoid 'may be unused' warnings */
3988 case geo_start: /* new when starting reshape - raid_disks not
3990 layout = mddev->new_layout;
3991 chunk = mddev->new_chunk_sectors;
3992 disks = mddev->raid_disks + mddev->delta_disks;
3997 if (chunk < (PAGE_SIZE >> 9) ||
3998 !is_power_of_2(chunk))
4001 fc = (layout >> 8) & 255;
4002 fo = layout & (1<<16);
4003 geo->raid_disks = disks;
4004 geo->near_copies = nc;
4005 geo->far_copies = fc;
4006 geo->far_offset = fo;
4007 switch (layout >> 17) {
4008 case 0: /* original layout. simple but not always optimal */
4009 geo->far_set_size = disks;
4011 case 1: /* "improved" layout which was buggy. Hopefully no-one is
4012 * actually using this, but leave code here just in case.*/
4013 geo->far_set_size = disks/fc;
4014 WARN(geo->far_set_size < fc,
4015 "This RAID10 layout does not provide data safety - please backup and create new array\n");
4017 case 2: /* "improved" layout fixed to match documentation */
4018 geo->far_set_size = fc * nc;
4020 default: /* Not a valid layout */
4023 geo->chunk_mask = chunk - 1;
4024 geo->chunk_shift = ffz(~chunk);
4028 static void raid10_free_conf(struct r10conf *conf)
4033 mempool_exit(&conf->r10bio_pool);
4034 kfree(conf->mirrors);
4035 kfree(conf->mirrors_old);
4036 kfree(conf->mirrors_new);
4037 safe_put_page(conf->tmppage);
4038 bioset_exit(&conf->bio_split);
4042 static struct r10conf *setup_conf(struct mddev *mddev)
4044 struct r10conf *conf = NULL;
4049 copies = setup_geo(&geo, mddev, geo_new);
4052 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
4053 mdname(mddev), PAGE_SIZE);
4057 if (copies < 2 || copies > mddev->raid_disks) {
4058 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
4059 mdname(mddev), mddev->new_layout);
4064 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
4068 /* FIXME calc properly */
4069 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
4070 sizeof(struct raid10_info),
4075 conf->tmppage = alloc_page(GFP_KERNEL);
4080 conf->copies = copies;
4081 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
4082 rbio_pool_free, conf);
4086 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
4090 calc_sectors(conf, mddev->dev_sectors);
4091 if (mddev->reshape_position == MaxSector) {
4092 conf->prev = conf->geo;
4093 conf->reshape_progress = MaxSector;
4095 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
4099 conf->reshape_progress = mddev->reshape_position;
4100 if (conf->prev.far_offset)
4101 conf->prev.stride = 1 << conf->prev.chunk_shift;
4103 /* far_copies must be 1 */
4104 conf->prev.stride = conf->dev_sectors;
4106 conf->reshape_safe = conf->reshape_progress;
4107 spin_lock_init(&conf->device_lock);
4108 INIT_LIST_HEAD(&conf->retry_list);
4109 INIT_LIST_HEAD(&conf->bio_end_io_list);
4111 seqlock_init(&conf->resync_lock);
4112 init_waitqueue_head(&conf->wait_barrier);
4113 atomic_set(&conf->nr_pending, 0);
4116 rcu_assign_pointer(conf->thread,
4117 md_register_thread(raid10d, mddev, "raid10"));
4121 conf->mddev = mddev;
4125 raid10_free_conf(conf);
4126 return ERR_PTR(err);
4129 static void raid10_set_io_opt(struct r10conf *conf)
4131 int raid_disks = conf->geo.raid_disks;
4133 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4134 raid_disks /= conf->geo.near_copies;
4135 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4139 static int raid10_run(struct mddev *mddev)
4141 struct r10conf *conf;
4143 struct raid10_info *disk;
4144 struct md_rdev *rdev;
4146 sector_t min_offset_diff = 0;
4149 if (mddev_init_writes_pending(mddev) < 0)
4152 if (mddev->private == NULL) {
4153 conf = setup_conf(mddev);
4155 return PTR_ERR(conf);
4156 mddev->private = conf;
4158 conf = mddev->private;
4162 rcu_assign_pointer(mddev->thread, conf->thread);
4163 rcu_assign_pointer(conf->thread, NULL);
4165 if (mddev_is_clustered(conf->mddev)) {
4168 fc = (mddev->layout >> 8) & 255;
4169 fo = mddev->layout & (1<<16);
4170 if (fc > 1 || fo > 0) {
4171 pr_err("only near layout is supported by clustered"
4178 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4179 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4180 raid10_set_io_opt(conf);
4183 rdev_for_each(rdev, mddev) {
4186 disk_idx = rdev->raid_disk;
4189 if (disk_idx >= conf->geo.raid_disks &&
4190 disk_idx >= conf->prev.raid_disks)
4192 disk = conf->mirrors + disk_idx;
4194 if (test_bit(Replacement, &rdev->flags)) {
4195 if (disk->replacement)
4197 disk->replacement = rdev;
4203 diff = (rdev->new_data_offset - rdev->data_offset);
4204 if (!mddev->reshape_backwards)
4208 if (first || diff < min_offset_diff)
4209 min_offset_diff = diff;
4212 disk_stack_limits(mddev->gendisk, rdev->bdev,
4213 rdev->data_offset << 9);
4215 disk->head_position = 0;
4219 /* need to check that every block has at least one working mirror */
4220 if (!enough(conf, -1)) {
4221 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4226 if (conf->reshape_progress != MaxSector) {
4227 /* must ensure that shape change is supported */
4228 if (conf->geo.far_copies != 1 &&
4229 conf->geo.far_offset == 0)
4231 if (conf->prev.far_copies != 1 &&
4232 conf->prev.far_offset == 0)
4236 mddev->degraded = 0;
4238 i < conf->geo.raid_disks
4239 || i < conf->prev.raid_disks;
4242 disk = conf->mirrors + i;
4244 if (!disk->rdev && disk->replacement) {
4245 /* The replacement is all we have - use it */
4246 disk->rdev = disk->replacement;
4247 disk->replacement = NULL;
4248 clear_bit(Replacement, &disk->rdev->flags);
4252 !test_bit(In_sync, &disk->rdev->flags)) {
4253 disk->head_position = 0;
4256 disk->rdev->saved_raid_disk < 0)
4260 if (disk->replacement &&
4261 !test_bit(In_sync, &disk->replacement->flags) &&
4262 disk->replacement->saved_raid_disk < 0) {
4266 disk->recovery_disabled = mddev->recovery_disabled - 1;
4269 if (mddev->recovery_cp != MaxSector)
4270 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4272 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4273 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4274 conf->geo.raid_disks);
4276 * Ok, everything is just fine now
4278 mddev->dev_sectors = conf->dev_sectors;
4279 size = raid10_size(mddev, 0, 0);
4280 md_set_array_sectors(mddev, size);
4281 mddev->resync_max_sectors = size;
4282 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4284 if (md_integrity_register(mddev))
4287 if (conf->reshape_progress != MaxSector) {
4288 unsigned long before_length, after_length;
4290 before_length = ((1 << conf->prev.chunk_shift) *
4291 conf->prev.far_copies);
4292 after_length = ((1 << conf->geo.chunk_shift) *
4293 conf->geo.far_copies);
4295 if (max(before_length, after_length) > min_offset_diff) {
4296 /* This cannot work */
4297 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4300 conf->offset_diff = min_offset_diff;
4302 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4303 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4304 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4305 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4306 rcu_assign_pointer(mddev->sync_thread,
4307 md_register_thread(md_do_sync, mddev, "reshape"));
4308 if (!mddev->sync_thread)
4315 md_unregister_thread(&mddev->thread);
4316 raid10_free_conf(conf);
4317 mddev->private = NULL;
4322 static void raid10_free(struct mddev *mddev, void *priv)
4324 raid10_free_conf(priv);
4327 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4329 struct r10conf *conf = mddev->private;
4332 raise_barrier(conf, 0);
4334 lower_barrier(conf);
4337 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4339 /* Resize of 'far' arrays is not supported.
4340 * For 'near' and 'offset' arrays we can set the
4341 * number of sectors used to be an appropriate multiple
4342 * of the chunk size.
4343 * For 'offset', this is far_copies*chunksize.
4344 * For 'near' the multiplier is the LCM of
4345 * near_copies and raid_disks.
4346 * So if far_copies > 1 && !far_offset, fail.
4347 * Else find LCM(raid_disks, near_copy)*far_copies and
4348 * multiply by chunk_size. Then round to this number.
4349 * This is mostly done by raid10_size()
4351 struct r10conf *conf = mddev->private;
4352 sector_t oldsize, size;
4354 if (mddev->reshape_position != MaxSector)
4357 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4360 oldsize = raid10_size(mddev, 0, 0);
4361 size = raid10_size(mddev, sectors, 0);
4362 if (mddev->external_size &&
4363 mddev->array_sectors > size)
4365 if (mddev->bitmap) {
4366 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4370 md_set_array_sectors(mddev, size);
4371 if (sectors > mddev->dev_sectors &&
4372 mddev->recovery_cp > oldsize) {
4373 mddev->recovery_cp = oldsize;
4374 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4376 calc_sectors(conf, sectors);
4377 mddev->dev_sectors = conf->dev_sectors;
4378 mddev->resync_max_sectors = size;
4382 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4384 struct md_rdev *rdev;
4385 struct r10conf *conf;
4387 if (mddev->degraded > 0) {
4388 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4390 return ERR_PTR(-EINVAL);
4392 sector_div(size, devs);
4394 /* Set new parameters */
4395 mddev->new_level = 10;
4396 /* new layout: far_copies = 1, near_copies = 2 */
4397 mddev->new_layout = (1<<8) + 2;
4398 mddev->new_chunk_sectors = mddev->chunk_sectors;
4399 mddev->delta_disks = mddev->raid_disks;
4400 mddev->raid_disks *= 2;
4401 /* make sure it will be not marked as dirty */
4402 mddev->recovery_cp = MaxSector;
4403 mddev->dev_sectors = size;
4405 conf = setup_conf(mddev);
4406 if (!IS_ERR(conf)) {
4407 rdev_for_each(rdev, mddev)
4408 if (rdev->raid_disk >= 0) {
4409 rdev->new_raid_disk = rdev->raid_disk * 2;
4410 rdev->sectors = size;
4412 WRITE_ONCE(conf->barrier, 1);
4418 static void *raid10_takeover(struct mddev *mddev)
4420 struct r0conf *raid0_conf;
4422 /* raid10 can take over:
4423 * raid0 - providing it has only two drives
4425 if (mddev->level == 0) {
4426 /* for raid0 takeover only one zone is supported */
4427 raid0_conf = mddev->private;
4428 if (raid0_conf->nr_strip_zones > 1) {
4429 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4431 return ERR_PTR(-EINVAL);
4433 return raid10_takeover_raid0(mddev,
4434 raid0_conf->strip_zone->zone_end,
4435 raid0_conf->strip_zone->nb_dev);
4437 return ERR_PTR(-EINVAL);
4440 static int raid10_check_reshape(struct mddev *mddev)
4442 /* Called when there is a request to change
4443 * - layout (to ->new_layout)
4444 * - chunk size (to ->new_chunk_sectors)
4445 * - raid_disks (by delta_disks)
4446 * or when trying to restart a reshape that was ongoing.
4448 * We need to validate the request and possibly allocate
4449 * space if that might be an issue later.
4451 * Currently we reject any reshape of a 'far' mode array,
4452 * allow chunk size to change if new is generally acceptable,
4453 * allow raid_disks to increase, and allow
4454 * a switch between 'near' mode and 'offset' mode.
4456 struct r10conf *conf = mddev->private;
4459 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4462 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4463 /* mustn't change number of copies */
4465 if (geo.far_copies > 1 && !geo.far_offset)
4466 /* Cannot switch to 'far' mode */
4469 if (mddev->array_sectors & geo.chunk_mask)
4470 /* not factor of array size */
4473 if (!enough(conf, -1))
4476 kfree(conf->mirrors_new);
4477 conf->mirrors_new = NULL;
4478 if (mddev->delta_disks > 0) {
4479 /* allocate new 'mirrors' list */
4481 kcalloc(mddev->raid_disks + mddev->delta_disks,
4482 sizeof(struct raid10_info),
4484 if (!conf->mirrors_new)
4491 * Need to check if array has failed when deciding whether to:
4493 * - remove non-faulty devices
4496 * This determination is simple when no reshape is happening.
4497 * However if there is a reshape, we need to carefully check
4498 * both the before and after sections.
4499 * This is because some failed devices may only affect one
4500 * of the two sections, and some non-in_sync devices may
4501 * be insync in the section most affected by failed devices.
4503 static int calc_degraded(struct r10conf *conf)
4505 int degraded, degraded2;
4510 /* 'prev' section first */
4511 for (i = 0; i < conf->prev.raid_disks; i++) {
4512 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4513 if (!rdev || test_bit(Faulty, &rdev->flags))
4515 else if (!test_bit(In_sync, &rdev->flags))
4516 /* When we can reduce the number of devices in
4517 * an array, this might not contribute to
4518 * 'degraded'. It does now.
4523 if (conf->geo.raid_disks == conf->prev.raid_disks)
4527 for (i = 0; i < conf->geo.raid_disks; i++) {
4528 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4529 if (!rdev || test_bit(Faulty, &rdev->flags))
4531 else if (!test_bit(In_sync, &rdev->flags)) {
4532 /* If reshape is increasing the number of devices,
4533 * this section has already been recovered, so
4534 * it doesn't contribute to degraded.
4537 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4542 if (degraded2 > degraded)
4547 static int raid10_start_reshape(struct mddev *mddev)
4549 /* A 'reshape' has been requested. This commits
4550 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4551 * This also checks if there are enough spares and adds them
4553 * We currently require enough spares to make the final
4554 * array non-degraded. We also require that the difference
4555 * between old and new data_offset - on each device - is
4556 * enough that we never risk over-writing.
4559 unsigned long before_length, after_length;
4560 sector_t min_offset_diff = 0;
4563 struct r10conf *conf = mddev->private;
4564 struct md_rdev *rdev;
4568 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4571 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4574 before_length = ((1 << conf->prev.chunk_shift) *
4575 conf->prev.far_copies);
4576 after_length = ((1 << conf->geo.chunk_shift) *
4577 conf->geo.far_copies);
4579 rdev_for_each(rdev, mddev) {
4580 if (!test_bit(In_sync, &rdev->flags)
4581 && !test_bit(Faulty, &rdev->flags))
4583 if (rdev->raid_disk >= 0) {
4584 long long diff = (rdev->new_data_offset
4585 - rdev->data_offset);
4586 if (!mddev->reshape_backwards)
4590 if (first || diff < min_offset_diff)
4591 min_offset_diff = diff;
4596 if (max(before_length, after_length) > min_offset_diff)
4599 if (spares < mddev->delta_disks)
4602 conf->offset_diff = min_offset_diff;
4603 spin_lock_irq(&conf->device_lock);
4604 if (conf->mirrors_new) {
4605 memcpy(conf->mirrors_new, conf->mirrors,
4606 sizeof(struct raid10_info)*conf->prev.raid_disks);
4608 kfree(conf->mirrors_old);
4609 conf->mirrors_old = conf->mirrors;
4610 conf->mirrors = conf->mirrors_new;
4611 conf->mirrors_new = NULL;
4613 setup_geo(&conf->geo, mddev, geo_start);
4615 if (mddev->reshape_backwards) {
4616 sector_t size = raid10_size(mddev, 0, 0);
4617 if (size < mddev->array_sectors) {
4618 spin_unlock_irq(&conf->device_lock);
4619 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4623 mddev->resync_max_sectors = size;
4624 conf->reshape_progress = size;
4626 conf->reshape_progress = 0;
4627 conf->reshape_safe = conf->reshape_progress;
4628 spin_unlock_irq(&conf->device_lock);
4630 if (mddev->delta_disks && mddev->bitmap) {
4631 struct mdp_superblock_1 *sb = NULL;
4632 sector_t oldsize, newsize;
4634 oldsize = raid10_size(mddev, 0, 0);
4635 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4637 if (!mddev_is_clustered(mddev)) {
4638 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4645 rdev_for_each(rdev, mddev) {
4646 if (rdev->raid_disk > -1 &&
4647 !test_bit(Faulty, &rdev->flags))
4648 sb = page_address(rdev->sb_page);
4652 * some node is already performing reshape, and no need to
4653 * call md_bitmap_resize again since it should be called when
4654 * receiving BITMAP_RESIZE msg
4656 if ((sb && (le32_to_cpu(sb->feature_map) &
4657 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4660 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4664 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4666 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4671 if (mddev->delta_disks > 0) {
4672 rdev_for_each(rdev, mddev)
4673 if (rdev->raid_disk < 0 &&
4674 !test_bit(Faulty, &rdev->flags)) {
4675 if (raid10_add_disk(mddev, rdev) == 0) {
4676 if (rdev->raid_disk >=
4677 conf->prev.raid_disks)
4678 set_bit(In_sync, &rdev->flags);
4680 rdev->recovery_offset = 0;
4682 /* Failure here is OK */
4683 sysfs_link_rdev(mddev, rdev);
4685 } else if (rdev->raid_disk >= conf->prev.raid_disks
4686 && !test_bit(Faulty, &rdev->flags)) {
4687 /* This is a spare that was manually added */
4688 set_bit(In_sync, &rdev->flags);
4691 /* When a reshape changes the number of devices,
4692 * ->degraded is measured against the larger of the
4693 * pre and post numbers.
4695 spin_lock_irq(&conf->device_lock);
4696 mddev->degraded = calc_degraded(conf);
4697 spin_unlock_irq(&conf->device_lock);
4698 mddev->raid_disks = conf->geo.raid_disks;
4699 mddev->reshape_position = conf->reshape_progress;
4700 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4702 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4703 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4704 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4705 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4706 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4708 rcu_assign_pointer(mddev->sync_thread,
4709 md_register_thread(md_do_sync, mddev, "reshape"));
4710 if (!mddev->sync_thread) {
4714 conf->reshape_checkpoint = jiffies;
4715 md_wakeup_thread(mddev->sync_thread);
4720 mddev->recovery = 0;
4721 spin_lock_irq(&conf->device_lock);
4722 conf->geo = conf->prev;
4723 mddev->raid_disks = conf->geo.raid_disks;
4724 rdev_for_each(rdev, mddev)
4725 rdev->new_data_offset = rdev->data_offset;
4727 conf->reshape_progress = MaxSector;
4728 conf->reshape_safe = MaxSector;
4729 mddev->reshape_position = MaxSector;
4730 spin_unlock_irq(&conf->device_lock);
4734 /* Calculate the last device-address that could contain
4735 * any block from the chunk that includes the array-address 's'
4736 * and report the next address.
4737 * i.e. the address returned will be chunk-aligned and after
4738 * any data that is in the chunk containing 's'.
4740 static sector_t last_dev_address(sector_t s, struct geom *geo)
4742 s = (s | geo->chunk_mask) + 1;
4743 s >>= geo->chunk_shift;
4744 s *= geo->near_copies;
4745 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4746 s *= geo->far_copies;
4747 s <<= geo->chunk_shift;
4751 /* Calculate the first device-address that could contain
4752 * any block from the chunk that includes the array-address 's'.
4753 * This too will be the start of a chunk
4755 static sector_t first_dev_address(sector_t s, struct geom *geo)
4757 s >>= geo->chunk_shift;
4758 s *= geo->near_copies;
4759 sector_div(s, geo->raid_disks);
4760 s *= geo->far_copies;
4761 s <<= geo->chunk_shift;
4765 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4768 /* We simply copy at most one chunk (smallest of old and new)
4769 * at a time, possibly less if that exceeds RESYNC_PAGES,
4770 * or we hit a bad block or something.
4771 * This might mean we pause for normal IO in the middle of
4772 * a chunk, but that is not a problem as mddev->reshape_position
4773 * can record any location.
4775 * If we will want to write to a location that isn't
4776 * yet recorded as 'safe' (i.e. in metadata on disk) then
4777 * we need to flush all reshape requests and update the metadata.
4779 * When reshaping forwards (e.g. to more devices), we interpret
4780 * 'safe' as the earliest block which might not have been copied
4781 * down yet. We divide this by previous stripe size and multiply
4782 * by previous stripe length to get lowest device offset that we
4783 * cannot write to yet.
4784 * We interpret 'sector_nr' as an address that we want to write to.
4785 * From this we use last_device_address() to find where we might
4786 * write to, and first_device_address on the 'safe' position.
4787 * If this 'next' write position is after the 'safe' position,
4788 * we must update the metadata to increase the 'safe' position.
4790 * When reshaping backwards, we round in the opposite direction
4791 * and perform the reverse test: next write position must not be
4792 * less than current safe position.
4794 * In all this the minimum difference in data offsets
4795 * (conf->offset_diff - always positive) allows a bit of slack,
4796 * so next can be after 'safe', but not by more than offset_diff
4798 * We need to prepare all the bios here before we start any IO
4799 * to ensure the size we choose is acceptable to all devices.
4800 * The means one for each copy for write-out and an extra one for
4802 * We store the read-in bio in ->master_bio and the others in
4803 * ->devs[x].bio and ->devs[x].repl_bio.
4805 struct r10conf *conf = mddev->private;
4806 struct r10bio *r10_bio;
4807 sector_t next, safe, last;
4811 struct md_rdev *rdev;
4814 struct bio *bio, *read_bio;
4815 int sectors_done = 0;
4816 struct page **pages;
4818 if (sector_nr == 0) {
4819 /* If restarting in the middle, skip the initial sectors */
4820 if (mddev->reshape_backwards &&
4821 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4822 sector_nr = (raid10_size(mddev, 0, 0)
4823 - conf->reshape_progress);
4824 } else if (!mddev->reshape_backwards &&
4825 conf->reshape_progress > 0)
4826 sector_nr = conf->reshape_progress;
4828 mddev->curr_resync_completed = sector_nr;
4829 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4835 /* We don't use sector_nr to track where we are up to
4836 * as that doesn't work well for ->reshape_backwards.
4837 * So just use ->reshape_progress.
4839 if (mddev->reshape_backwards) {
4840 /* 'next' is the earliest device address that we might
4841 * write to for this chunk in the new layout
4843 next = first_dev_address(conf->reshape_progress - 1,
4846 /* 'safe' is the last device address that we might read from
4847 * in the old layout after a restart
4849 safe = last_dev_address(conf->reshape_safe - 1,
4852 if (next + conf->offset_diff < safe)
4855 last = conf->reshape_progress - 1;
4856 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4857 & conf->prev.chunk_mask);
4858 if (sector_nr + RESYNC_SECTORS < last)
4859 sector_nr = last + 1 - RESYNC_SECTORS;
4861 /* 'next' is after the last device address that we
4862 * might write to for this chunk in the new layout
4864 next = last_dev_address(conf->reshape_progress, &conf->geo);
4866 /* 'safe' is the earliest device address that we might
4867 * read from in the old layout after a restart
4869 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4871 /* Need to update metadata if 'next' might be beyond 'safe'
4872 * as that would possibly corrupt data
4874 if (next > safe + conf->offset_diff)
4877 sector_nr = conf->reshape_progress;
4878 last = sector_nr | (conf->geo.chunk_mask
4879 & conf->prev.chunk_mask);
4881 if (sector_nr + RESYNC_SECTORS <= last)
4882 last = sector_nr + RESYNC_SECTORS - 1;
4886 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4887 /* Need to update reshape_position in metadata */
4888 wait_barrier(conf, false);
4889 mddev->reshape_position = conf->reshape_progress;
4890 if (mddev->reshape_backwards)
4891 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4892 - conf->reshape_progress;
4894 mddev->curr_resync_completed = conf->reshape_progress;
4895 conf->reshape_checkpoint = jiffies;
4896 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4897 md_wakeup_thread(mddev->thread);
4898 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4899 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4900 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4901 allow_barrier(conf);
4902 return sectors_done;
4904 conf->reshape_safe = mddev->reshape_position;
4905 allow_barrier(conf);
4908 raise_barrier(conf, 0);
4910 /* Now schedule reads for blocks from sector_nr to last */
4911 r10_bio = raid10_alloc_init_r10buf(conf);
4913 raise_barrier(conf, 1);
4914 atomic_set(&r10_bio->remaining, 0);
4915 r10_bio->mddev = mddev;
4916 r10_bio->sector = sector_nr;
4917 set_bit(R10BIO_IsReshape, &r10_bio->state);
4918 r10_bio->sectors = last - sector_nr + 1;
4919 rdev = read_balance(conf, r10_bio, &max_sectors);
4920 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4923 /* Cannot read from here, so need to record bad blocks
4924 * on all the target devices.
4927 mempool_free(r10_bio, &conf->r10buf_pool);
4928 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4929 return sectors_done;
4932 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4933 GFP_KERNEL, &mddev->bio_set);
4934 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4935 + rdev->data_offset);
4936 read_bio->bi_private = r10_bio;
4937 read_bio->bi_end_io = end_reshape_read;
4938 r10_bio->master_bio = read_bio;
4939 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4942 * Broadcast RESYNC message to other nodes, so all nodes would not
4943 * write to the region to avoid conflict.
4945 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4946 struct mdp_superblock_1 *sb = NULL;
4947 int sb_reshape_pos = 0;
4949 conf->cluster_sync_low = sector_nr;
4950 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4951 sb = page_address(rdev->sb_page);
4953 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4955 * Set cluster_sync_low again if next address for array
4956 * reshape is less than cluster_sync_low. Since we can't
4957 * update cluster_sync_low until it has finished reshape.
4959 if (sb_reshape_pos < conf->cluster_sync_low)
4960 conf->cluster_sync_low = sb_reshape_pos;
4963 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4964 conf->cluster_sync_high);
4967 /* Now find the locations in the new layout */
4968 __raid10_find_phys(&conf->geo, r10_bio);
4971 read_bio->bi_next = NULL;
4974 for (s = 0; s < conf->copies*2; s++) {
4976 int d = r10_bio->devs[s/2].devnum;
4977 struct md_rdev *rdev2;
4979 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4980 b = r10_bio->devs[s/2].repl_bio;
4982 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4983 b = r10_bio->devs[s/2].bio;
4985 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4988 bio_set_dev(b, rdev2->bdev);
4989 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4990 rdev2->new_data_offset;
4991 b->bi_end_io = end_reshape_write;
4992 b->bi_opf = REQ_OP_WRITE;
4997 /* Now add as many pages as possible to all of these bios. */
5000 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5001 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
5002 struct page *page = pages[s / (PAGE_SIZE >> 9)];
5003 int len = (max_sectors - s) << 9;
5004 if (len > PAGE_SIZE)
5006 for (bio = blist; bio ; bio = bio->bi_next) {
5007 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
5008 bio->bi_status = BLK_STS_RESOURCE;
5010 return sectors_done;
5013 sector_nr += len >> 9;
5014 nr_sectors += len >> 9;
5017 r10_bio->sectors = nr_sectors;
5019 /* Now submit the read */
5020 md_sync_acct_bio(read_bio, r10_bio->sectors);
5021 atomic_inc(&r10_bio->remaining);
5022 read_bio->bi_next = NULL;
5023 submit_bio_noacct(read_bio);
5024 sectors_done += nr_sectors;
5025 if (sector_nr <= last)
5028 lower_barrier(conf);
5030 /* Now that we have done the whole section we can
5031 * update reshape_progress
5033 if (mddev->reshape_backwards)
5034 conf->reshape_progress -= sectors_done;
5036 conf->reshape_progress += sectors_done;
5038 return sectors_done;
5041 static void end_reshape_request(struct r10bio *r10_bio);
5042 static int handle_reshape_read_error(struct mddev *mddev,
5043 struct r10bio *r10_bio);
5044 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
5046 /* Reshape read completed. Hopefully we have a block
5048 * If we got a read error then we do sync 1-page reads from
5049 * elsewhere until we find the data - or give up.
5051 struct r10conf *conf = mddev->private;
5054 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
5055 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
5056 /* Reshape has been aborted */
5057 md_done_sync(mddev, r10_bio->sectors, 0);
5061 /* We definitely have the data in the pages, schedule the
5064 atomic_set(&r10_bio->remaining, 1);
5065 for (s = 0; s < conf->copies*2; s++) {
5067 int d = r10_bio->devs[s/2].devnum;
5068 struct md_rdev *rdev;
5071 rdev = rcu_dereference(conf->mirrors[d].replacement);
5072 b = r10_bio->devs[s/2].repl_bio;
5074 rdev = rcu_dereference(conf->mirrors[d].rdev);
5075 b = r10_bio->devs[s/2].bio;
5077 if (!rdev || test_bit(Faulty, &rdev->flags)) {
5081 atomic_inc(&rdev->nr_pending);
5083 md_sync_acct_bio(b, r10_bio->sectors);
5084 atomic_inc(&r10_bio->remaining);
5086 submit_bio_noacct(b);
5088 end_reshape_request(r10_bio);
5091 static void end_reshape(struct r10conf *conf)
5093 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
5096 spin_lock_irq(&conf->device_lock);
5097 conf->prev = conf->geo;
5098 md_finish_reshape(conf->mddev);
5100 conf->reshape_progress = MaxSector;
5101 conf->reshape_safe = MaxSector;
5102 spin_unlock_irq(&conf->device_lock);
5104 if (conf->mddev->queue)
5105 raid10_set_io_opt(conf);
5109 static void raid10_update_reshape_pos(struct mddev *mddev)
5111 struct r10conf *conf = mddev->private;
5114 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
5115 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
5116 || mddev->reshape_position == MaxSector)
5117 conf->reshape_progress = mddev->reshape_position;
5122 static int handle_reshape_read_error(struct mddev *mddev,
5123 struct r10bio *r10_bio)
5125 /* Use sync reads to get the blocks from somewhere else */
5126 int sectors = r10_bio->sectors;
5127 struct r10conf *conf = mddev->private;
5128 struct r10bio *r10b;
5131 struct page **pages;
5133 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5135 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5139 /* reshape IOs share pages from .devs[0].bio */
5140 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5142 r10b->sector = r10_bio->sector;
5143 __raid10_find_phys(&conf->prev, r10b);
5148 int first_slot = slot;
5150 if (s > (PAGE_SIZE >> 9))
5155 int d = r10b->devs[slot].devnum;
5156 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5159 test_bit(Faulty, &rdev->flags) ||
5160 !test_bit(In_sync, &rdev->flags))
5163 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5164 atomic_inc(&rdev->nr_pending);
5166 success = sync_page_io(rdev,
5170 REQ_OP_READ, false);
5171 rdev_dec_pending(rdev, mddev);
5177 if (slot >= conf->copies)
5179 if (slot == first_slot)
5184 /* couldn't read this block, must give up */
5185 set_bit(MD_RECOVERY_INTR,
5197 static void end_reshape_write(struct bio *bio)
5199 struct r10bio *r10_bio = get_resync_r10bio(bio);
5200 struct mddev *mddev = r10_bio->mddev;
5201 struct r10conf *conf = mddev->private;
5205 struct md_rdev *rdev = NULL;
5207 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5209 rdev = conf->mirrors[d].replacement;
5212 rdev = conf->mirrors[d].rdev;
5215 if (bio->bi_status) {
5216 /* FIXME should record badblock */
5217 md_error(mddev, rdev);
5220 rdev_dec_pending(rdev, mddev);
5221 end_reshape_request(r10_bio);
5224 static void end_reshape_request(struct r10bio *r10_bio)
5226 if (!atomic_dec_and_test(&r10_bio->remaining))
5228 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5229 bio_put(r10_bio->master_bio);
5233 static void raid10_finish_reshape(struct mddev *mddev)
5235 struct r10conf *conf = mddev->private;
5237 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5240 if (mddev->delta_disks > 0) {
5241 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5242 mddev->recovery_cp = mddev->resync_max_sectors;
5243 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5245 mddev->resync_max_sectors = mddev->array_sectors;
5249 for (d = conf->geo.raid_disks ;
5250 d < conf->geo.raid_disks - mddev->delta_disks;
5252 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5254 clear_bit(In_sync, &rdev->flags);
5255 rdev = rcu_dereference(conf->mirrors[d].replacement);
5257 clear_bit(In_sync, &rdev->flags);
5261 mddev->layout = mddev->new_layout;
5262 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5263 mddev->reshape_position = MaxSector;
5264 mddev->delta_disks = 0;
5265 mddev->reshape_backwards = 0;
5268 static struct md_personality raid10_personality =
5272 .owner = THIS_MODULE,
5273 .make_request = raid10_make_request,
5275 .free = raid10_free,
5276 .status = raid10_status,
5277 .error_handler = raid10_error,
5278 .hot_add_disk = raid10_add_disk,
5279 .hot_remove_disk= raid10_remove_disk,
5280 .spare_active = raid10_spare_active,
5281 .sync_request = raid10_sync_request,
5282 .quiesce = raid10_quiesce,
5283 .size = raid10_size,
5284 .resize = raid10_resize,
5285 .takeover = raid10_takeover,
5286 .check_reshape = raid10_check_reshape,
5287 .start_reshape = raid10_start_reshape,
5288 .finish_reshape = raid10_finish_reshape,
5289 .update_reshape_pos = raid10_update_reshape_pos,
5292 static int __init raid_init(void)
5294 return register_md_personality(&raid10_personality);
5297 static void raid_exit(void)
5299 unregister_md_personality(&raid10_personality);
5302 module_init(raid_init);
5303 module_exit(raid_exit);
5304 MODULE_LICENSE("GPL");
5305 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5306 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5307 MODULE_ALIAS("md-raid10");
5308 MODULE_ALIAS("md-level-10");