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 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
86 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
88 return get_resync_pages(bio)->raid_bio;
91 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
96 /* allocate a r10bio with room for raid_disks entries in the
98 return kzalloc(size, gfp_flags);
101 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102 /* amount of memory to reserve for resync requests */
103 #define RESYNC_WINDOW (1024*1024)
104 /* maximum number of concurrent requests, memory permitting */
105 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
116 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
133 nalloc = 2; /* recovery */
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
142 goto out_free_r10bio;
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
151 r10_bio->devs[j].bio = bio;
152 if (!conf->have_replacement)
154 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
157 r10_bio->devs[j].repl_bio = bio;
160 * Allocate RESYNC_PAGES data pages and attach them
163 for (j = 0; j < nalloc; j++) {
164 struct bio *rbio = r10_bio->devs[j].repl_bio;
165 struct resync_pages *rp, *rp_repl;
169 rp_repl = &rps[nalloc + j];
171 bio = r10_bio->devs[j].bio;
173 if (!j || test_bit(MD_RECOVERY_SYNC,
174 &conf->mddev->recovery)) {
175 if (resync_alloc_pages(rp, gfp_flags))
178 memcpy(rp, &rps[0], sizeof(*rp));
179 resync_get_all_pages(rp);
182 rp->raid_bio = r10_bio;
183 bio->bi_private = rp;
185 memcpy(rp_repl, rp, sizeof(*rp));
186 rbio->bi_private = rp_repl;
194 resync_free_pages(&rps[j]);
198 for ( ; j < nalloc; j++) {
199 if (r10_bio->devs[j].bio)
200 bio_put(r10_bio->devs[j].bio);
201 if (r10_bio->devs[j].repl_bio)
202 bio_put(r10_bio->devs[j].repl_bio);
206 rbio_pool_free(r10_bio, conf);
210 static void r10buf_pool_free(void *__r10_bio, void *data)
212 struct r10conf *conf = data;
213 struct r10bio *r10bio = __r10_bio;
215 struct resync_pages *rp = NULL;
217 for (j = conf->copies; j--; ) {
218 struct bio *bio = r10bio->devs[j].bio;
221 rp = get_resync_pages(bio);
222 resync_free_pages(rp);
226 bio = r10bio->devs[j].repl_bio;
231 /* resync pages array stored in the 1st bio's .bi_private */
234 rbio_pool_free(r10bio, conf);
237 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241 for (i = 0; i < conf->geo.raid_disks; i++) {
242 struct bio **bio = & r10_bio->devs[i].bio;
243 if (!BIO_SPECIAL(*bio))
246 bio = &r10_bio->devs[i].repl_bio;
247 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
253 static void free_r10bio(struct r10bio *r10_bio)
255 struct r10conf *conf = r10_bio->mddev->private;
257 put_all_bios(conf, r10_bio);
258 mempool_free(r10_bio, &conf->r10bio_pool);
261 static void put_buf(struct r10bio *r10_bio)
263 struct r10conf *conf = r10_bio->mddev->private;
265 mempool_free(r10_bio, &conf->r10buf_pool);
270 static void reschedule_retry(struct r10bio *r10_bio)
273 struct mddev *mddev = r10_bio->mddev;
274 struct r10conf *conf = mddev->private;
276 spin_lock_irqsave(&conf->device_lock, flags);
277 list_add(&r10_bio->retry_list, &conf->retry_list);
279 spin_unlock_irqrestore(&conf->device_lock, flags);
281 /* wake up frozen array... */
282 wake_up(&conf->wait_barrier);
284 md_wakeup_thread(mddev->thread);
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
292 static void raid_end_bio_io(struct r10bio *r10_bio)
294 struct bio *bio = r10_bio->master_bio;
295 struct r10conf *conf = r10_bio->mddev->private;
297 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
298 bio->bi_status = BLK_STS_IOERR;
300 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
301 bio_end_io_acct(bio, r10_bio->start_time);
304 * Wake up any possible resync thread that waits for the device
309 free_r10bio(r10_bio);
313 * Update disk head position estimator based on IRQ completion info.
315 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
317 struct r10conf *conf = r10_bio->mddev->private;
319 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
320 r10_bio->devs[slot].addr + (r10_bio->sectors);
324 * Find the disk number which triggered given bio
326 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
327 struct bio *bio, int *slotp, int *replp)
332 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
333 if (r10_bio->devs[slot].bio == bio)
335 if (r10_bio->devs[slot].repl_bio == bio) {
341 update_head_pos(slot, r10_bio);
347 return r10_bio->devs[slot].devnum;
350 static void raid10_end_read_request(struct bio *bio)
352 int uptodate = !bio->bi_status;
353 struct r10bio *r10_bio = bio->bi_private;
355 struct md_rdev *rdev;
356 struct r10conf *conf = r10_bio->mddev->private;
358 slot = r10_bio->read_slot;
359 rdev = r10_bio->devs[slot].rdev;
361 * this branch is our 'one mirror IO has finished' event handler:
363 update_head_pos(slot, r10_bio);
367 * Set R10BIO_Uptodate in our master bio, so that
368 * we will return a good error code to the higher
369 * levels even if IO on some other mirrored buffer fails.
371 * The 'master' represents the composite IO operation to
372 * user-side. So if something waits for IO, then it will
373 * wait for the 'master' bio.
375 set_bit(R10BIO_Uptodate, &r10_bio->state);
377 /* If all other devices that store this block have
378 * failed, we want to return the error upwards rather
379 * than fail the last device. Here we redefine
380 * "uptodate" to mean "Don't want to retry"
382 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
387 raid_end_bio_io(r10_bio);
388 rdev_dec_pending(rdev, conf->mddev);
391 * oops, read error - keep the refcount on the rdev
393 char b[BDEVNAME_SIZE];
394 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
396 bdevname(rdev->bdev, b),
397 (unsigned long long)r10_bio->sector);
398 set_bit(R10BIO_ReadError, &r10_bio->state);
399 reschedule_retry(r10_bio);
403 static void close_write(struct r10bio *r10_bio)
405 /* clear the bitmap if all writes complete successfully */
406 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
408 !test_bit(R10BIO_Degraded, &r10_bio->state),
410 md_write_end(r10_bio->mddev);
413 static void one_write_done(struct r10bio *r10_bio)
415 if (atomic_dec_and_test(&r10_bio->remaining)) {
416 if (test_bit(R10BIO_WriteError, &r10_bio->state))
417 reschedule_retry(r10_bio);
419 close_write(r10_bio);
420 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
421 reschedule_retry(r10_bio);
423 raid_end_bio_io(r10_bio);
428 static void raid10_end_write_request(struct bio *bio)
430 struct r10bio *r10_bio = bio->bi_private;
433 struct r10conf *conf = r10_bio->mddev->private;
435 struct md_rdev *rdev = NULL;
436 struct bio *to_put = NULL;
439 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
441 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
444 rdev = conf->mirrors[dev].replacement;
448 rdev = conf->mirrors[dev].rdev;
451 * this branch is our 'one mirror IO has finished' event handler:
453 if (bio->bi_status && !discard_error) {
455 /* Never record new bad blocks to replacement,
458 md_error(rdev->mddev, rdev);
460 set_bit(WriteErrorSeen, &rdev->flags);
461 if (!test_and_set_bit(WantReplacement, &rdev->flags))
462 set_bit(MD_RECOVERY_NEEDED,
463 &rdev->mddev->recovery);
466 if (test_bit(FailFast, &rdev->flags) &&
467 (bio->bi_opf & MD_FAILFAST)) {
468 md_error(rdev->mddev, rdev);
472 * When the device is faulty, it is not necessary to
473 * handle write error.
475 if (!test_bit(Faulty, &rdev->flags))
476 set_bit(R10BIO_WriteError, &r10_bio->state);
478 /* Fail the request */
479 set_bit(R10BIO_Degraded, &r10_bio->state);
480 r10_bio->devs[slot].bio = NULL;
487 * Set R10BIO_Uptodate in our master bio, so that
488 * we will return a good error code for to the higher
489 * levels even if IO on some other mirrored buffer fails.
491 * The 'master' represents the composite IO operation to
492 * user-side. So if something waits for IO, then it will
493 * wait for the 'master' bio.
499 * Do not set R10BIO_Uptodate if the current device is
500 * rebuilding or Faulty. This is because we cannot use
501 * such device for properly reading the data back (we could
502 * potentially use it, if the current write would have felt
503 * before rdev->recovery_offset, but for simplicity we don't
506 if (test_bit(In_sync, &rdev->flags) &&
507 !test_bit(Faulty, &rdev->flags))
508 set_bit(R10BIO_Uptodate, &r10_bio->state);
510 /* Maybe we can clear some bad blocks. */
511 if (is_badblock(rdev,
512 r10_bio->devs[slot].addr,
514 &first_bad, &bad_sectors) && !discard_error) {
517 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
519 r10_bio->devs[slot].bio = IO_MADE_GOOD;
521 set_bit(R10BIO_MadeGood, &r10_bio->state);
527 * Let's see if all mirrored write operations have finished
530 one_write_done(r10_bio);
532 rdev_dec_pending(rdev, conf->mddev);
538 * RAID10 layout manager
539 * As well as the chunksize and raid_disks count, there are two
540 * parameters: near_copies and far_copies.
541 * near_copies * far_copies must be <= raid_disks.
542 * Normally one of these will be 1.
543 * If both are 1, we get raid0.
544 * If near_copies == raid_disks, we get raid1.
546 * Chunks are laid out in raid0 style with near_copies copies of the
547 * first chunk, followed by near_copies copies of the next chunk and
549 * If far_copies > 1, then after 1/far_copies of the array has been assigned
550 * as described above, we start again with a device offset of near_copies.
551 * So we effectively have another copy of the whole array further down all
552 * the drives, but with blocks on different drives.
553 * With this layout, and block is never stored twice on the one device.
555 * raid10_find_phys finds the sector offset of a given virtual sector
556 * on each device that it is on.
558 * raid10_find_virt does the reverse mapping, from a device and a
559 * sector offset to a virtual address
562 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
570 int last_far_set_start, last_far_set_size;
572 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
573 last_far_set_start *= geo->far_set_size;
575 last_far_set_size = geo->far_set_size;
576 last_far_set_size += (geo->raid_disks % geo->far_set_size);
578 /* now calculate first sector/dev */
579 chunk = r10bio->sector >> geo->chunk_shift;
580 sector = r10bio->sector & geo->chunk_mask;
582 chunk *= geo->near_copies;
584 dev = sector_div(stripe, geo->raid_disks);
586 stripe *= geo->far_copies;
588 sector += stripe << geo->chunk_shift;
590 /* and calculate all the others */
591 for (n = 0; n < geo->near_copies; n++) {
595 r10bio->devs[slot].devnum = d;
596 r10bio->devs[slot].addr = s;
599 for (f = 1; f < geo->far_copies; f++) {
600 set = d / geo->far_set_size;
601 d += geo->near_copies;
603 if ((geo->raid_disks % geo->far_set_size) &&
604 (d > last_far_set_start)) {
605 d -= last_far_set_start;
606 d %= last_far_set_size;
607 d += last_far_set_start;
609 d %= geo->far_set_size;
610 d += geo->far_set_size * set;
613 r10bio->devs[slot].devnum = d;
614 r10bio->devs[slot].addr = s;
618 if (dev >= geo->raid_disks) {
620 sector += (geo->chunk_mask + 1);
625 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
627 struct geom *geo = &conf->geo;
629 if (conf->reshape_progress != MaxSector &&
630 ((r10bio->sector >= conf->reshape_progress) !=
631 conf->mddev->reshape_backwards)) {
632 set_bit(R10BIO_Previous, &r10bio->state);
635 clear_bit(R10BIO_Previous, &r10bio->state);
637 __raid10_find_phys(geo, r10bio);
640 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
642 sector_t offset, chunk, vchunk;
643 /* Never use conf->prev as this is only called during resync
644 * or recovery, so reshape isn't happening
646 struct geom *geo = &conf->geo;
647 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
648 int far_set_size = geo->far_set_size;
649 int last_far_set_start;
651 if (geo->raid_disks % geo->far_set_size) {
652 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
653 last_far_set_start *= geo->far_set_size;
655 if (dev >= last_far_set_start) {
656 far_set_size = geo->far_set_size;
657 far_set_size += (geo->raid_disks % geo->far_set_size);
658 far_set_start = last_far_set_start;
662 offset = sector & geo->chunk_mask;
663 if (geo->far_offset) {
665 chunk = sector >> geo->chunk_shift;
666 fc = sector_div(chunk, geo->far_copies);
667 dev -= fc * geo->near_copies;
668 if (dev < far_set_start)
671 while (sector >= geo->stride) {
672 sector -= geo->stride;
673 if (dev < (geo->near_copies + far_set_start))
674 dev += far_set_size - geo->near_copies;
676 dev -= geo->near_copies;
678 chunk = sector >> geo->chunk_shift;
680 vchunk = chunk * geo->raid_disks + dev;
681 sector_div(vchunk, geo->near_copies);
682 return (vchunk << geo->chunk_shift) + offset;
686 * This routine returns the disk from which the requested read should
687 * be done. There is a per-array 'next expected sequential IO' sector
688 * number - if this matches on the next IO then we use the last disk.
689 * There is also a per-disk 'last know head position' sector that is
690 * maintained from IRQ contexts, both the normal and the resync IO
691 * completion handlers update this position correctly. If there is no
692 * perfect sequential match then we pick the disk whose head is closest.
694 * If there are 2 mirrors in the same 2 devices, performance degrades
695 * because position is mirror, not device based.
697 * The rdev for the device selected will have nr_pending incremented.
701 * FIXME: possibly should rethink readbalancing and do it differently
702 * depending on near_copies / far_copies geometry.
704 static struct md_rdev *read_balance(struct r10conf *conf,
705 struct r10bio *r10_bio,
708 const sector_t this_sector = r10_bio->sector;
710 int sectors = r10_bio->sectors;
711 int best_good_sectors;
712 sector_t new_distance, best_dist;
713 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
715 int best_dist_slot, best_pending_slot;
716 bool has_nonrot_disk = false;
717 unsigned int min_pending;
718 struct geom *geo = &conf->geo;
720 raid10_find_phys(conf, r10_bio);
723 min_pending = UINT_MAX;
724 best_dist_rdev = NULL;
725 best_pending_rdev = NULL;
726 best_dist = MaxSector;
727 best_good_sectors = 0;
729 clear_bit(R10BIO_FailFast, &r10_bio->state);
731 * Check if we can balance. We can balance on the whole
732 * device if no resync is going on (recovery is ok), or below
733 * the resync window. We take the first readable disk when
734 * above the resync window.
736 if ((conf->mddev->recovery_cp < MaxSector
737 && (this_sector + sectors >= conf->next_resync)) ||
738 (mddev_is_clustered(conf->mddev) &&
739 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
740 this_sector + sectors)))
743 for (slot = 0; slot < conf->copies ; slot++) {
747 unsigned int pending;
750 if (r10_bio->devs[slot].bio == IO_BLOCKED)
752 disk = r10_bio->devs[slot].devnum;
753 rdev = rcu_dereference(conf->mirrors[disk].replacement);
754 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
755 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
756 rdev = rcu_dereference(conf->mirrors[disk].rdev);
758 test_bit(Faulty, &rdev->flags))
760 if (!test_bit(In_sync, &rdev->flags) &&
761 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
764 dev_sector = r10_bio->devs[slot].addr;
765 if (is_badblock(rdev, dev_sector, sectors,
766 &first_bad, &bad_sectors)) {
767 if (best_dist < MaxSector)
768 /* Already have a better slot */
770 if (first_bad <= dev_sector) {
771 /* Cannot read here. If this is the
772 * 'primary' device, then we must not read
773 * beyond 'bad_sectors' from another device.
775 bad_sectors -= (dev_sector - first_bad);
776 if (!do_balance && sectors > bad_sectors)
777 sectors = bad_sectors;
778 if (best_good_sectors > sectors)
779 best_good_sectors = sectors;
781 sector_t good_sectors =
782 first_bad - dev_sector;
783 if (good_sectors > best_good_sectors) {
784 best_good_sectors = good_sectors;
785 best_dist_slot = slot;
786 best_dist_rdev = rdev;
789 /* Must read from here */
794 best_good_sectors = sectors;
799 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
800 has_nonrot_disk |= nonrot;
801 pending = atomic_read(&rdev->nr_pending);
802 if (min_pending > pending && nonrot) {
803 min_pending = pending;
804 best_pending_slot = slot;
805 best_pending_rdev = rdev;
808 if (best_dist_slot >= 0)
809 /* At least 2 disks to choose from so failfast is OK */
810 set_bit(R10BIO_FailFast, &r10_bio->state);
811 /* This optimisation is debatable, and completely destroys
812 * sequential read speed for 'far copies' arrays. So only
813 * keep it for 'near' arrays, and review those later.
815 if (geo->near_copies > 1 && !pending)
818 /* for far > 1 always use the lowest address */
819 else if (geo->far_copies > 1)
820 new_distance = r10_bio->devs[slot].addr;
822 new_distance = abs(r10_bio->devs[slot].addr -
823 conf->mirrors[disk].head_position);
825 if (new_distance < best_dist) {
826 best_dist = new_distance;
827 best_dist_slot = slot;
828 best_dist_rdev = rdev;
831 if (slot >= conf->copies) {
832 if (has_nonrot_disk) {
833 slot = best_pending_slot;
834 rdev = best_pending_rdev;
836 slot = best_dist_slot;
837 rdev = best_dist_rdev;
842 atomic_inc(&rdev->nr_pending);
843 r10_bio->read_slot = slot;
847 *max_sectors = best_good_sectors;
852 static void flush_pending_writes(struct r10conf *conf)
854 /* Any writes that have been queued but are awaiting
855 * bitmap updates get flushed here.
857 spin_lock_irq(&conf->device_lock);
859 if (conf->pending_bio_list.head) {
860 struct blk_plug plug;
863 bio = bio_list_get(&conf->pending_bio_list);
864 conf->pending_count = 0;
865 spin_unlock_irq(&conf->device_lock);
868 * As this is called in a wait_event() loop (see freeze_array),
869 * current->state might be TASK_UNINTERRUPTIBLE which will
870 * cause a warning when we prepare to wait again. As it is
871 * rare that this path is taken, it is perfectly safe to force
872 * us to go around the wait_event() loop again, so the warning
873 * is a false-positive. Silence the warning by resetting
876 __set_current_state(TASK_RUNNING);
878 blk_start_plug(&plug);
879 /* flush any pending bitmap writes to disk
880 * before proceeding w/ I/O */
881 md_bitmap_unplug(conf->mddev->bitmap);
882 wake_up(&conf->wait_barrier);
884 while (bio) { /* submit pending writes */
885 struct bio *next = bio->bi_next;
886 struct md_rdev *rdev = (void*)bio->bi_bdev;
888 bio_set_dev(bio, rdev->bdev);
889 if (test_bit(Faulty, &rdev->flags)) {
891 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
892 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
896 submit_bio_noacct(bio);
899 blk_finish_plug(&plug);
901 spin_unlock_irq(&conf->device_lock);
905 * Sometimes we need to suspend IO while we do something else,
906 * either some resync/recovery, or reconfigure the array.
907 * To do this we raise a 'barrier'.
908 * The 'barrier' is a counter that can be raised multiple times
909 * to count how many activities are happening which preclude
911 * We can only raise the barrier if there is no pending IO.
912 * i.e. if nr_pending == 0.
913 * We choose only to raise the barrier if no-one is waiting for the
914 * barrier to go down. This means that as soon as an IO request
915 * is ready, no other operations which require a barrier will start
916 * until the IO request has had a chance.
918 * So: regular IO calls 'wait_barrier'. When that returns there
919 * is no backgroup IO happening, It must arrange to call
920 * allow_barrier when it has finished its IO.
921 * backgroup IO calls must call raise_barrier. Once that returns
922 * there is no normal IO happeing. It must arrange to call
923 * lower_barrier when the particular background IO completes.
926 static void raise_barrier(struct r10conf *conf, int force)
928 BUG_ON(force && !conf->barrier);
929 spin_lock_irq(&conf->resync_lock);
931 /* Wait until no block IO is waiting (unless 'force') */
932 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
935 /* block any new IO from starting */
938 /* Now wait for all pending IO to complete */
939 wait_event_lock_irq(conf->wait_barrier,
940 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
943 spin_unlock_irq(&conf->resync_lock);
946 static void lower_barrier(struct r10conf *conf)
949 spin_lock_irqsave(&conf->resync_lock, flags);
951 spin_unlock_irqrestore(&conf->resync_lock, flags);
952 wake_up(&conf->wait_barrier);
955 static void wait_barrier(struct r10conf *conf)
957 spin_lock_irq(&conf->resync_lock);
959 struct bio_list *bio_list = current->bio_list;
961 /* Wait for the barrier to drop.
962 * However if there are already pending
963 * requests (preventing the barrier from
964 * rising completely), and the
965 * pre-process bio queue isn't empty,
966 * then don't wait, as we need to empty
967 * that queue to get the nr_pending
970 raid10_log(conf->mddev, "wait barrier");
971 wait_event_lock_irq(conf->wait_barrier,
973 (atomic_read(&conf->nr_pending) &&
975 (!bio_list_empty(&bio_list[0]) ||
976 !bio_list_empty(&bio_list[1]))) ||
977 /* move on if recovery thread is
980 (conf->mddev->thread->tsk == current &&
981 test_bit(MD_RECOVERY_RUNNING,
982 &conf->mddev->recovery) &&
983 conf->nr_queued > 0),
986 if (!conf->nr_waiting)
987 wake_up(&conf->wait_barrier);
989 atomic_inc(&conf->nr_pending);
990 spin_unlock_irq(&conf->resync_lock);
993 static void allow_barrier(struct r10conf *conf)
995 if ((atomic_dec_and_test(&conf->nr_pending)) ||
996 (conf->array_freeze_pending))
997 wake_up(&conf->wait_barrier);
1000 static void freeze_array(struct r10conf *conf, int extra)
1002 /* stop syncio and normal IO and wait for everything to
1004 * We increment barrier and nr_waiting, and then
1005 * wait until nr_pending match nr_queued+extra
1006 * This is called in the context of one normal IO request
1007 * that has failed. Thus any sync request that might be pending
1008 * will be blocked by nr_pending, and we need to wait for
1009 * pending IO requests to complete or be queued for re-try.
1010 * Thus the number queued (nr_queued) plus this request (extra)
1011 * must match the number of pending IOs (nr_pending) before
1014 spin_lock_irq(&conf->resync_lock);
1015 conf->array_freeze_pending++;
1018 wait_event_lock_irq_cmd(conf->wait_barrier,
1019 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1021 flush_pending_writes(conf));
1023 conf->array_freeze_pending--;
1024 spin_unlock_irq(&conf->resync_lock);
1027 static void unfreeze_array(struct r10conf *conf)
1029 /* reverse the effect of the freeze */
1030 spin_lock_irq(&conf->resync_lock);
1033 wake_up(&conf->wait_barrier);
1034 spin_unlock_irq(&conf->resync_lock);
1037 static sector_t choose_data_offset(struct r10bio *r10_bio,
1038 struct md_rdev *rdev)
1040 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1041 test_bit(R10BIO_Previous, &r10_bio->state))
1042 return rdev->data_offset;
1044 return rdev->new_data_offset;
1047 struct raid10_plug_cb {
1048 struct blk_plug_cb cb;
1049 struct bio_list pending;
1053 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1055 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1057 struct mddev *mddev = plug->cb.data;
1058 struct r10conf *conf = mddev->private;
1061 if (from_schedule || current->bio_list) {
1062 spin_lock_irq(&conf->device_lock);
1063 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1064 conf->pending_count += plug->pending_cnt;
1065 spin_unlock_irq(&conf->device_lock);
1066 wake_up(&conf->wait_barrier);
1067 md_wakeup_thread(mddev->thread);
1072 /* we aren't scheduling, so we can do the write-out directly. */
1073 bio = bio_list_get(&plug->pending);
1074 md_bitmap_unplug(mddev->bitmap);
1075 wake_up(&conf->wait_barrier);
1077 while (bio) { /* submit pending writes */
1078 struct bio *next = bio->bi_next;
1079 struct md_rdev *rdev = (void*)bio->bi_bdev;
1080 bio->bi_next = NULL;
1081 bio_set_dev(bio, rdev->bdev);
1082 if (test_bit(Faulty, &rdev->flags)) {
1084 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1085 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
1086 /* Just ignore it */
1089 submit_bio_noacct(bio);
1096 * 1. Register the new request and wait if the reconstruction thread has put
1097 * up a bar for new requests. Continue immediately if no resync is active
1099 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1101 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1102 struct bio *bio, sector_t sectors)
1105 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1106 bio->bi_iter.bi_sector < conf->reshape_progress &&
1107 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1108 raid10_log(conf->mddev, "wait reshape");
1109 allow_barrier(conf);
1110 wait_event(conf->wait_barrier,
1111 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1112 conf->reshape_progress >= bio->bi_iter.bi_sector +
1118 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1119 struct r10bio *r10_bio)
1121 struct r10conf *conf = mddev->private;
1122 struct bio *read_bio;
1123 const int op = bio_op(bio);
1124 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1126 struct md_rdev *rdev;
1127 char b[BDEVNAME_SIZE];
1128 int slot = r10_bio->read_slot;
1129 struct md_rdev *err_rdev = NULL;
1130 gfp_t gfp = GFP_NOIO;
1132 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1134 * This is an error retry, but we cannot
1135 * safely dereference the rdev in the r10_bio,
1136 * we must use the one in conf.
1137 * If it has already been disconnected (unlikely)
1138 * we lose the device name in error messages.
1142 * As we are blocking raid10, it is a little safer to
1145 gfp = GFP_NOIO | __GFP_HIGH;
1148 disk = r10_bio->devs[slot].devnum;
1149 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1151 bdevname(err_rdev->bdev, b);
1154 /* This never gets dereferenced */
1155 err_rdev = r10_bio->devs[slot].rdev;
1160 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1161 rdev = read_balance(conf, r10_bio, &max_sectors);
1164 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1166 (unsigned long long)r10_bio->sector);
1168 raid_end_bio_io(r10_bio);
1172 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1174 bdevname(rdev->bdev, b),
1175 (unsigned long long)r10_bio->sector);
1176 if (max_sectors < bio_sectors(bio)) {
1177 struct bio *split = bio_split(bio, max_sectors,
1178 gfp, &conf->bio_split);
1179 bio_chain(split, bio);
1180 allow_barrier(conf);
1181 submit_bio_noacct(bio);
1184 r10_bio->master_bio = bio;
1185 r10_bio->sectors = max_sectors;
1187 slot = r10_bio->read_slot;
1189 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1190 r10_bio->start_time = bio_start_io_acct(bio);
1191 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1193 r10_bio->devs[slot].bio = read_bio;
1194 r10_bio->devs[slot].rdev = rdev;
1196 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1197 choose_data_offset(r10_bio, rdev);
1198 bio_set_dev(read_bio, rdev->bdev);
1199 read_bio->bi_end_io = raid10_end_read_request;
1200 bio_set_op_attrs(read_bio, op, do_sync);
1201 if (test_bit(FailFast, &rdev->flags) &&
1202 test_bit(R10BIO_FailFast, &r10_bio->state))
1203 read_bio->bi_opf |= MD_FAILFAST;
1204 read_bio->bi_private = r10_bio;
1207 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1209 submit_bio_noacct(read_bio);
1213 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1214 struct bio *bio, bool replacement,
1217 const int op = bio_op(bio);
1218 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1219 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1220 unsigned long flags;
1221 struct blk_plug_cb *cb;
1222 struct raid10_plug_cb *plug = NULL;
1223 struct r10conf *conf = mddev->private;
1224 struct md_rdev *rdev;
1225 int devnum = r10_bio->devs[n_copy].devnum;
1229 rdev = conf->mirrors[devnum].replacement;
1231 /* Replacement just got moved to main 'rdev' */
1233 rdev = conf->mirrors[devnum].rdev;
1236 rdev = conf->mirrors[devnum].rdev;
1238 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1240 r10_bio->devs[n_copy].repl_bio = mbio;
1242 r10_bio->devs[n_copy].bio = mbio;
1244 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1245 choose_data_offset(r10_bio, rdev));
1246 bio_set_dev(mbio, rdev->bdev);
1247 mbio->bi_end_io = raid10_end_write_request;
1248 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1249 if (!replacement && test_bit(FailFast,
1250 &conf->mirrors[devnum].rdev->flags)
1251 && enough(conf, devnum))
1252 mbio->bi_opf |= MD_FAILFAST;
1253 mbio->bi_private = r10_bio;
1255 if (conf->mddev->gendisk)
1256 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1258 /* flush_pending_writes() needs access to the rdev so...*/
1259 mbio->bi_bdev = (void *)rdev;
1261 atomic_inc(&r10_bio->remaining);
1263 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1265 plug = container_of(cb, struct raid10_plug_cb, cb);
1269 bio_list_add(&plug->pending, mbio);
1270 plug->pending_cnt++;
1272 spin_lock_irqsave(&conf->device_lock, flags);
1273 bio_list_add(&conf->pending_bio_list, mbio);
1274 conf->pending_count++;
1275 spin_unlock_irqrestore(&conf->device_lock, flags);
1276 md_wakeup_thread(mddev->thread);
1280 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1283 struct r10conf *conf = mddev->private;
1284 struct md_rdev *blocked_rdev;
1287 blocked_rdev = NULL;
1289 for (i = 0; i < conf->copies; i++) {
1290 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1291 struct md_rdev *rrdev = rcu_dereference(
1292 conf->mirrors[i].replacement);
1295 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1296 atomic_inc(&rdev->nr_pending);
1297 blocked_rdev = rdev;
1300 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1301 atomic_inc(&rrdev->nr_pending);
1302 blocked_rdev = rrdev;
1306 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1308 sector_t dev_sector = r10_bio->devs[i].addr;
1313 * Discard request doesn't care the write result
1314 * so it doesn't need to wait blocked disk here.
1316 if (!r10_bio->sectors)
1319 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1320 &first_bad, &bad_sectors);
1323 * Mustn't write here until the bad block
1326 atomic_inc(&rdev->nr_pending);
1327 set_bit(BlockedBadBlocks, &rdev->flags);
1328 blocked_rdev = rdev;
1335 if (unlikely(blocked_rdev)) {
1336 /* Have to wait for this device to get unblocked, then retry */
1337 allow_barrier(conf);
1338 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1339 __func__, blocked_rdev->raid_disk);
1340 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1346 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1347 struct r10bio *r10_bio)
1349 struct r10conf *conf = mddev->private;
1354 if ((mddev_is_clustered(mddev) &&
1355 md_cluster_ops->area_resyncing(mddev, WRITE,
1356 bio->bi_iter.bi_sector,
1357 bio_end_sector(bio)))) {
1360 prepare_to_wait(&conf->wait_barrier,
1362 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1363 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1367 finish_wait(&conf->wait_barrier, &w);
1370 sectors = r10_bio->sectors;
1371 regular_request_wait(mddev, conf, bio, sectors);
1372 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1373 (mddev->reshape_backwards
1374 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1375 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1376 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1377 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1378 /* Need to update reshape_position in metadata */
1379 mddev->reshape_position = conf->reshape_progress;
1380 set_mask_bits(&mddev->sb_flags, 0,
1381 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1382 md_wakeup_thread(mddev->thread);
1383 raid10_log(conf->mddev, "wait reshape metadata");
1384 wait_event(mddev->sb_wait,
1385 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1387 conf->reshape_safe = mddev->reshape_position;
1390 if (conf->pending_count >= max_queued_requests) {
1391 md_wakeup_thread(mddev->thread);
1392 raid10_log(mddev, "wait queued");
1393 wait_event(conf->wait_barrier,
1394 conf->pending_count < max_queued_requests);
1396 /* first select target devices under rcu_lock and
1397 * inc refcount on their rdev. Record them by setting
1399 * If there are known/acknowledged bad blocks on any device
1400 * on which we have seen a write error, we want to avoid
1401 * writing to those blocks. This potentially requires several
1402 * writes to write around the bad blocks. Each set of writes
1403 * gets its own r10_bio with a set of bios attached.
1406 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1407 raid10_find_phys(conf, r10_bio);
1409 wait_blocked_dev(mddev, r10_bio);
1412 max_sectors = r10_bio->sectors;
1414 for (i = 0; i < conf->copies; i++) {
1415 int d = r10_bio->devs[i].devnum;
1416 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1417 struct md_rdev *rrdev = rcu_dereference(
1418 conf->mirrors[d].replacement);
1421 if (rdev && (test_bit(Faulty, &rdev->flags)))
1423 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1426 r10_bio->devs[i].bio = NULL;
1427 r10_bio->devs[i].repl_bio = NULL;
1429 if (!rdev && !rrdev) {
1430 set_bit(R10BIO_Degraded, &r10_bio->state);
1433 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1435 sector_t dev_sector = r10_bio->devs[i].addr;
1439 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1440 &first_bad, &bad_sectors);
1441 if (is_bad && first_bad <= dev_sector) {
1442 /* Cannot write here at all */
1443 bad_sectors -= (dev_sector - first_bad);
1444 if (bad_sectors < max_sectors)
1445 /* Mustn't write more than bad_sectors
1446 * to other devices yet
1448 max_sectors = bad_sectors;
1449 /* We don't set R10BIO_Degraded as that
1450 * only applies if the disk is missing,
1451 * so it might be re-added, and we want to
1452 * know to recover this chunk.
1453 * In this case the device is here, and the
1454 * fact that this chunk is not in-sync is
1455 * recorded in the bad block log.
1460 int good_sectors = first_bad - dev_sector;
1461 if (good_sectors < max_sectors)
1462 max_sectors = good_sectors;
1466 r10_bio->devs[i].bio = bio;
1467 atomic_inc(&rdev->nr_pending);
1470 r10_bio->devs[i].repl_bio = bio;
1471 atomic_inc(&rrdev->nr_pending);
1476 if (max_sectors < r10_bio->sectors)
1477 r10_bio->sectors = max_sectors;
1479 if (r10_bio->sectors < bio_sectors(bio)) {
1480 struct bio *split = bio_split(bio, r10_bio->sectors,
1481 GFP_NOIO, &conf->bio_split);
1482 bio_chain(split, bio);
1483 allow_barrier(conf);
1484 submit_bio_noacct(bio);
1487 r10_bio->master_bio = bio;
1490 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1491 r10_bio->start_time = bio_start_io_acct(bio);
1492 atomic_set(&r10_bio->remaining, 1);
1493 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1495 for (i = 0; i < conf->copies; i++) {
1496 if (r10_bio->devs[i].bio)
1497 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1498 if (r10_bio->devs[i].repl_bio)
1499 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1501 one_write_done(r10_bio);
1504 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1506 struct r10conf *conf = mddev->private;
1507 struct r10bio *r10_bio;
1509 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1511 r10_bio->master_bio = bio;
1512 r10_bio->sectors = sectors;
1514 r10_bio->mddev = mddev;
1515 r10_bio->sector = bio->bi_iter.bi_sector;
1517 r10_bio->read_slot = -1;
1518 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1519 conf->geo.raid_disks);
1521 if (bio_data_dir(bio) == READ)
1522 raid10_read_request(mddev, bio, r10_bio);
1524 raid10_write_request(mddev, bio, r10_bio);
1527 static void raid_end_discard_bio(struct r10bio *r10bio)
1529 struct r10conf *conf = r10bio->mddev->private;
1530 struct r10bio *first_r10bio;
1532 while (atomic_dec_and_test(&r10bio->remaining)) {
1534 allow_barrier(conf);
1536 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1537 first_r10bio = (struct r10bio *)r10bio->master_bio;
1538 free_r10bio(r10bio);
1539 r10bio = first_r10bio;
1541 md_write_end(r10bio->mddev);
1542 bio_endio(r10bio->master_bio);
1543 free_r10bio(r10bio);
1549 static void raid10_end_discard_request(struct bio *bio)
1551 struct r10bio *r10_bio = bio->bi_private;
1552 struct r10conf *conf = r10_bio->mddev->private;
1553 struct md_rdev *rdev = NULL;
1558 * We don't care the return value of discard bio
1560 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1561 set_bit(R10BIO_Uptodate, &r10_bio->state);
1563 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1565 rdev = conf->mirrors[dev].replacement;
1568 * raid10_remove_disk uses smp_mb to make sure rdev is set to
1569 * replacement before setting replacement to NULL. It can read
1570 * rdev first without barrier protect even replacment is NULL
1573 rdev = conf->mirrors[dev].rdev;
1576 raid_end_discard_bio(r10_bio);
1577 rdev_dec_pending(rdev, conf->mddev);
1581 * There are some limitations to handle discard bio
1582 * 1st, the discard size is bigger than stripe_size*2.
1583 * 2st, if the discard bio spans reshape progress, we use the old way to
1584 * handle discard bio
1586 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1588 struct r10conf *conf = mddev->private;
1589 struct geom *geo = &conf->geo;
1590 int far_copies = geo->far_copies;
1591 bool first_copy = true;
1592 struct r10bio *r10_bio, *first_r10bio;
1596 unsigned int stripe_size;
1597 unsigned int stripe_data_disks;
1598 sector_t split_size;
1599 sector_t bio_start, bio_end;
1600 sector_t first_stripe_index, last_stripe_index;
1601 sector_t start_disk_offset;
1602 unsigned int start_disk_index;
1603 sector_t end_disk_offset;
1604 unsigned int end_disk_index;
1605 unsigned int remainder;
1607 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1613 * Check reshape again to avoid reshape happens after checking
1614 * MD_RECOVERY_RESHAPE and before wait_barrier
1616 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1619 if (geo->near_copies)
1620 stripe_data_disks = geo->raid_disks / geo->near_copies +
1621 geo->raid_disks % geo->near_copies;
1623 stripe_data_disks = geo->raid_disks;
1625 stripe_size = stripe_data_disks << geo->chunk_shift;
1627 bio_start = bio->bi_iter.bi_sector;
1628 bio_end = bio_end_sector(bio);
1631 * Maybe one discard bio is smaller than strip size or across one
1632 * stripe and discard region is larger than one stripe size. For far
1633 * offset layout, if the discard region is not aligned with stripe
1634 * size, there is hole when we submit discard bio to member disk.
1635 * For simplicity, we only handle discard bio which discard region
1636 * is bigger than stripe_size * 2
1638 if (bio_sectors(bio) < stripe_size*2)
1642 * Keep bio aligned with strip size.
1644 div_u64_rem(bio_start, stripe_size, &remainder);
1646 split_size = stripe_size - remainder;
1647 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1648 bio_chain(split, bio);
1649 allow_barrier(conf);
1650 /* Resend the fist split part */
1651 submit_bio_noacct(split);
1654 div_u64_rem(bio_end, stripe_size, &remainder);
1656 split_size = bio_sectors(bio) - remainder;
1657 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1658 bio_chain(split, bio);
1659 allow_barrier(conf);
1660 /* Resend the second split part */
1661 submit_bio_noacct(bio);
1666 bio_start = bio->bi_iter.bi_sector;
1667 bio_end = bio_end_sector(bio);
1670 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1671 * One stripe contains the chunks from all member disk (one chunk from
1672 * one disk at the same HBA address). For layout detail, see 'man md 4'
1674 chunk = bio_start >> geo->chunk_shift;
1675 chunk *= geo->near_copies;
1676 first_stripe_index = chunk;
1677 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1678 if (geo->far_offset)
1679 first_stripe_index *= geo->far_copies;
1680 start_disk_offset = (bio_start & geo->chunk_mask) +
1681 (first_stripe_index << geo->chunk_shift);
1683 chunk = bio_end >> geo->chunk_shift;
1684 chunk *= geo->near_copies;
1685 last_stripe_index = chunk;
1686 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1687 if (geo->far_offset)
1688 last_stripe_index *= geo->far_copies;
1689 end_disk_offset = (bio_end & geo->chunk_mask) +
1690 (last_stripe_index << geo->chunk_shift);
1693 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1694 r10_bio->mddev = mddev;
1696 r10_bio->sectors = 0;
1697 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1698 wait_blocked_dev(mddev, r10_bio);
1701 * For far layout it needs more than one r10bio to cover all regions.
1702 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1703 * to record the discard bio. Other r10bio->master_bio record the first
1704 * r10bio. The first r10bio only release after all other r10bios finish.
1705 * The discard bio returns only first r10bio finishes
1708 r10_bio->master_bio = bio;
1709 set_bit(R10BIO_Discard, &r10_bio->state);
1711 first_r10bio = r10_bio;
1713 r10_bio->master_bio = (struct bio *)first_r10bio;
1716 for (disk = 0; disk < geo->raid_disks; disk++) {
1717 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1718 struct md_rdev *rrdev = rcu_dereference(
1719 conf->mirrors[disk].replacement);
1721 r10_bio->devs[disk].bio = NULL;
1722 r10_bio->devs[disk].repl_bio = NULL;
1724 if (rdev && (test_bit(Faulty, &rdev->flags)))
1726 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1728 if (!rdev && !rrdev)
1732 r10_bio->devs[disk].bio = bio;
1733 atomic_inc(&rdev->nr_pending);
1736 r10_bio->devs[disk].repl_bio = bio;
1737 atomic_inc(&rrdev->nr_pending);
1742 atomic_set(&r10_bio->remaining, 1);
1743 for (disk = 0; disk < geo->raid_disks; disk++) {
1744 sector_t dev_start, dev_end;
1745 struct bio *mbio, *rbio = NULL;
1746 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1747 struct md_rdev *rrdev = rcu_dereference(
1748 conf->mirrors[disk].replacement);
1751 * Now start to calculate the start and end address for each disk.
1752 * The space between dev_start and dev_end is the discard region.
1754 * For dev_start, it needs to consider three conditions:
1755 * 1st, the disk is before start_disk, you can imagine the disk in
1756 * the next stripe. So the dev_start is the start address of next
1758 * 2st, the disk is after start_disk, it means the disk is at the
1759 * same stripe of first disk
1760 * 3st, the first disk itself, we can use start_disk_offset directly
1762 if (disk < start_disk_index)
1763 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1764 else if (disk > start_disk_index)
1765 dev_start = first_stripe_index * mddev->chunk_sectors;
1767 dev_start = start_disk_offset;
1769 if (disk < end_disk_index)
1770 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1771 else if (disk > end_disk_index)
1772 dev_end = last_stripe_index * mddev->chunk_sectors;
1774 dev_end = end_disk_offset;
1777 * It only handles discard bio which size is >= stripe size, so
1778 * dev_end > dev_start all the time
1780 if (r10_bio->devs[disk].bio) {
1781 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1782 mbio->bi_end_io = raid10_end_discard_request;
1783 mbio->bi_private = r10_bio;
1784 r10_bio->devs[disk].bio = mbio;
1785 r10_bio->devs[disk].devnum = disk;
1786 atomic_inc(&r10_bio->remaining);
1787 md_submit_discard_bio(mddev, rdev, mbio,
1788 dev_start + choose_data_offset(r10_bio, rdev),
1789 dev_end - dev_start);
1792 if (r10_bio->devs[disk].repl_bio) {
1793 rbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1794 rbio->bi_end_io = raid10_end_discard_request;
1795 rbio->bi_private = r10_bio;
1796 r10_bio->devs[disk].repl_bio = rbio;
1797 r10_bio->devs[disk].devnum = disk;
1798 atomic_inc(&r10_bio->remaining);
1799 md_submit_discard_bio(mddev, rrdev, rbio,
1800 dev_start + choose_data_offset(r10_bio, rrdev),
1801 dev_end - dev_start);
1806 if (!geo->far_offset && --far_copies) {
1807 first_stripe_index += geo->stride >> geo->chunk_shift;
1808 start_disk_offset += geo->stride;
1809 last_stripe_index += geo->stride >> geo->chunk_shift;
1810 end_disk_offset += geo->stride;
1811 atomic_inc(&first_r10bio->remaining);
1812 raid_end_discard_bio(r10_bio);
1817 raid_end_discard_bio(r10_bio);
1821 allow_barrier(conf);
1825 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1827 struct r10conf *conf = mddev->private;
1828 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1829 int chunk_sects = chunk_mask + 1;
1830 int sectors = bio_sectors(bio);
1832 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1833 && md_flush_request(mddev, bio))
1836 if (!md_write_start(mddev, bio))
1839 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1840 if (!raid10_handle_discard(mddev, bio))
1844 * If this request crosses a chunk boundary, we need to split
1847 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1848 sectors > chunk_sects
1849 && (conf->geo.near_copies < conf->geo.raid_disks
1850 || conf->prev.near_copies <
1851 conf->prev.raid_disks)))
1852 sectors = chunk_sects -
1853 (bio->bi_iter.bi_sector &
1855 __make_request(mddev, bio, sectors);
1857 /* In case raid10d snuck in to freeze_array */
1858 wake_up(&conf->wait_barrier);
1862 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1864 struct r10conf *conf = mddev->private;
1867 if (conf->geo.near_copies < conf->geo.raid_disks)
1868 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1869 if (conf->geo.near_copies > 1)
1870 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1871 if (conf->geo.far_copies > 1) {
1872 if (conf->geo.far_offset)
1873 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1875 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1876 if (conf->geo.far_set_size != conf->geo.raid_disks)
1877 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1879 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1880 conf->geo.raid_disks - mddev->degraded);
1882 for (i = 0; i < conf->geo.raid_disks; i++) {
1883 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1884 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1887 seq_printf(seq, "]");
1890 /* check if there are enough drives for
1891 * every block to appear on atleast one.
1892 * Don't consider the device numbered 'ignore'
1893 * as we might be about to remove it.
1895 static int _enough(struct r10conf *conf, int previous, int ignore)
1901 disks = conf->prev.raid_disks;
1902 ncopies = conf->prev.near_copies;
1904 disks = conf->geo.raid_disks;
1905 ncopies = conf->geo.near_copies;
1910 int n = conf->copies;
1914 struct md_rdev *rdev;
1915 if (this != ignore &&
1916 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1917 test_bit(In_sync, &rdev->flags))
1919 this = (this+1) % disks;
1923 first = (first + ncopies) % disks;
1924 } while (first != 0);
1931 static int enough(struct r10conf *conf, int ignore)
1933 /* when calling 'enough', both 'prev' and 'geo' must
1935 * This is ensured if ->reconfig_mutex or ->device_lock
1938 return _enough(conf, 0, ignore) &&
1939 _enough(conf, 1, ignore);
1942 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1944 char b[BDEVNAME_SIZE];
1945 struct r10conf *conf = mddev->private;
1946 unsigned long flags;
1949 * If it is not operational, then we have already marked it as dead
1950 * else if it is the last working disks with "fail_last_dev == false",
1951 * ignore the error, let the next level up know.
1952 * else mark the drive as failed
1954 spin_lock_irqsave(&conf->device_lock, flags);
1955 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1956 && !enough(conf, rdev->raid_disk)) {
1958 * Don't fail the drive, just return an IO error.
1960 spin_unlock_irqrestore(&conf->device_lock, flags);
1963 if (test_and_clear_bit(In_sync, &rdev->flags))
1966 * If recovery is running, make sure it aborts.
1968 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1969 set_bit(Blocked, &rdev->flags);
1970 set_bit(Faulty, &rdev->flags);
1971 set_mask_bits(&mddev->sb_flags, 0,
1972 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1973 spin_unlock_irqrestore(&conf->device_lock, flags);
1974 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1975 "md/raid10:%s: Operation continuing on %d devices.\n",
1976 mdname(mddev), bdevname(rdev->bdev, b),
1977 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1980 static void print_conf(struct r10conf *conf)
1983 struct md_rdev *rdev;
1985 pr_debug("RAID10 conf printout:\n");
1987 pr_debug("(!conf)\n");
1990 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1991 conf->geo.raid_disks);
1993 /* This is only called with ->reconfix_mutex held, so
1994 * rcu protection of rdev is not needed */
1995 for (i = 0; i < conf->geo.raid_disks; i++) {
1996 char b[BDEVNAME_SIZE];
1997 rdev = conf->mirrors[i].rdev;
1999 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
2000 i, !test_bit(In_sync, &rdev->flags),
2001 !test_bit(Faulty, &rdev->flags),
2002 bdevname(rdev->bdev,b));
2006 static void close_sync(struct r10conf *conf)
2009 allow_barrier(conf);
2011 mempool_exit(&conf->r10buf_pool);
2014 static int raid10_spare_active(struct mddev *mddev)
2017 struct r10conf *conf = mddev->private;
2018 struct raid10_info *tmp;
2020 unsigned long flags;
2023 * Find all non-in_sync disks within the RAID10 configuration
2024 * and mark them in_sync
2026 for (i = 0; i < conf->geo.raid_disks; i++) {
2027 tmp = conf->mirrors + i;
2028 if (tmp->replacement
2029 && tmp->replacement->recovery_offset == MaxSector
2030 && !test_bit(Faulty, &tmp->replacement->flags)
2031 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2032 /* Replacement has just become active */
2034 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2037 /* Replaced device not technically faulty,
2038 * but we need to be sure it gets removed
2039 * and never re-added.
2041 set_bit(Faulty, &tmp->rdev->flags);
2042 sysfs_notify_dirent_safe(
2043 tmp->rdev->sysfs_state);
2045 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2046 } else if (tmp->rdev
2047 && tmp->rdev->recovery_offset == MaxSector
2048 && !test_bit(Faulty, &tmp->rdev->flags)
2049 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2051 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2054 spin_lock_irqsave(&conf->device_lock, flags);
2055 mddev->degraded -= count;
2056 spin_unlock_irqrestore(&conf->device_lock, flags);
2062 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2064 struct r10conf *conf = mddev->private;
2068 int last = conf->geo.raid_disks - 1;
2070 if (mddev->recovery_cp < MaxSector)
2071 /* only hot-add to in-sync arrays, as recovery is
2072 * very different from resync
2075 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2078 if (md_integrity_add_rdev(rdev, mddev))
2081 if (rdev->raid_disk >= 0)
2082 first = last = rdev->raid_disk;
2084 if (rdev->saved_raid_disk >= first &&
2085 rdev->saved_raid_disk < conf->geo.raid_disks &&
2086 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2087 mirror = rdev->saved_raid_disk;
2090 for ( ; mirror <= last ; mirror++) {
2091 struct raid10_info *p = &conf->mirrors[mirror];
2092 if (p->recovery_disabled == mddev->recovery_disabled)
2095 if (!test_bit(WantReplacement, &p->rdev->flags) ||
2096 p->replacement != NULL)
2098 clear_bit(In_sync, &rdev->flags);
2099 set_bit(Replacement, &rdev->flags);
2100 rdev->raid_disk = mirror;
2103 disk_stack_limits(mddev->gendisk, rdev->bdev,
2104 rdev->data_offset << 9);
2106 rcu_assign_pointer(p->replacement, rdev);
2111 disk_stack_limits(mddev->gendisk, rdev->bdev,
2112 rdev->data_offset << 9);
2114 p->head_position = 0;
2115 p->recovery_disabled = mddev->recovery_disabled - 1;
2116 rdev->raid_disk = mirror;
2118 if (rdev->saved_raid_disk != mirror)
2120 rcu_assign_pointer(p->rdev, rdev);
2123 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
2124 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
2130 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2132 struct r10conf *conf = mddev->private;
2134 int number = rdev->raid_disk;
2135 struct md_rdev **rdevp;
2136 struct raid10_info *p = conf->mirrors + number;
2139 if (rdev == p->rdev)
2141 else if (rdev == p->replacement)
2142 rdevp = &p->replacement;
2146 if (test_bit(In_sync, &rdev->flags) ||
2147 atomic_read(&rdev->nr_pending)) {
2151 /* Only remove non-faulty devices if recovery
2154 if (!test_bit(Faulty, &rdev->flags) &&
2155 mddev->recovery_disabled != p->recovery_disabled &&
2156 (!p->replacement || p->replacement == rdev) &&
2157 number < conf->geo.raid_disks &&
2163 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
2165 if (atomic_read(&rdev->nr_pending)) {
2166 /* lost the race, try later */
2172 if (p->replacement) {
2173 /* We must have just cleared 'rdev' */
2174 p->rdev = p->replacement;
2175 clear_bit(Replacement, &p->replacement->flags);
2176 smp_mb(); /* Make sure other CPUs may see both as identical
2177 * but will never see neither -- if they are careful.
2179 p->replacement = NULL;
2182 clear_bit(WantReplacement, &rdev->flags);
2183 err = md_integrity_register(mddev);
2191 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2193 struct r10conf *conf = r10_bio->mddev->private;
2195 if (!bio->bi_status)
2196 set_bit(R10BIO_Uptodate, &r10_bio->state);
2198 /* The write handler will notice the lack of
2199 * R10BIO_Uptodate and record any errors etc
2201 atomic_add(r10_bio->sectors,
2202 &conf->mirrors[d].rdev->corrected_errors);
2204 /* for reconstruct, we always reschedule after a read.
2205 * for resync, only after all reads
2207 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2208 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2209 atomic_dec_and_test(&r10_bio->remaining)) {
2210 /* we have read all the blocks,
2211 * do the comparison in process context in raid10d
2213 reschedule_retry(r10_bio);
2217 static void end_sync_read(struct bio *bio)
2219 struct r10bio *r10_bio = get_resync_r10bio(bio);
2220 struct r10conf *conf = r10_bio->mddev->private;
2221 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2223 __end_sync_read(r10_bio, bio, d);
2226 static void end_reshape_read(struct bio *bio)
2228 /* reshape read bio isn't allocated from r10buf_pool */
2229 struct r10bio *r10_bio = bio->bi_private;
2231 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2234 static void end_sync_request(struct r10bio *r10_bio)
2236 struct mddev *mddev = r10_bio->mddev;
2238 while (atomic_dec_and_test(&r10_bio->remaining)) {
2239 if (r10_bio->master_bio == NULL) {
2240 /* the primary of several recovery bios */
2241 sector_t s = r10_bio->sectors;
2242 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2243 test_bit(R10BIO_WriteError, &r10_bio->state))
2244 reschedule_retry(r10_bio);
2247 md_done_sync(mddev, s, 1);
2250 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2251 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2252 test_bit(R10BIO_WriteError, &r10_bio->state))
2253 reschedule_retry(r10_bio);
2261 static void end_sync_write(struct bio *bio)
2263 struct r10bio *r10_bio = get_resync_r10bio(bio);
2264 struct mddev *mddev = r10_bio->mddev;
2265 struct r10conf *conf = mddev->private;
2271 struct md_rdev *rdev = NULL;
2273 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2275 rdev = conf->mirrors[d].replacement;
2277 rdev = conf->mirrors[d].rdev;
2279 if (bio->bi_status) {
2281 md_error(mddev, rdev);
2283 set_bit(WriteErrorSeen, &rdev->flags);
2284 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2285 set_bit(MD_RECOVERY_NEEDED,
2286 &rdev->mddev->recovery);
2287 set_bit(R10BIO_WriteError, &r10_bio->state);
2289 } else if (is_badblock(rdev,
2290 r10_bio->devs[slot].addr,
2292 &first_bad, &bad_sectors))
2293 set_bit(R10BIO_MadeGood, &r10_bio->state);
2295 rdev_dec_pending(rdev, mddev);
2297 end_sync_request(r10_bio);
2301 * Note: sync and recover and handled very differently for raid10
2302 * This code is for resync.
2303 * For resync, we read through virtual addresses and read all blocks.
2304 * If there is any error, we schedule a write. The lowest numbered
2305 * drive is authoritative.
2306 * However requests come for physical address, so we need to map.
2307 * For every physical address there are raid_disks/copies virtual addresses,
2308 * which is always are least one, but is not necessarly an integer.
2309 * This means that a physical address can span multiple chunks, so we may
2310 * have to submit multiple io requests for a single sync request.
2313 * We check if all blocks are in-sync and only write to blocks that
2316 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2318 struct r10conf *conf = mddev->private;
2320 struct bio *tbio, *fbio;
2322 struct page **tpages, **fpages;
2324 atomic_set(&r10_bio->remaining, 1);
2326 /* find the first device with a block */
2327 for (i=0; i<conf->copies; i++)
2328 if (!r10_bio->devs[i].bio->bi_status)
2331 if (i == conf->copies)
2335 fbio = r10_bio->devs[i].bio;
2336 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2337 fbio->bi_iter.bi_idx = 0;
2338 fpages = get_resync_pages(fbio)->pages;
2340 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2341 /* now find blocks with errors */
2342 for (i=0 ; i < conf->copies ; i++) {
2344 struct md_rdev *rdev;
2345 struct resync_pages *rp;
2347 tbio = r10_bio->devs[i].bio;
2349 if (tbio->bi_end_io != end_sync_read)
2354 tpages = get_resync_pages(tbio)->pages;
2355 d = r10_bio->devs[i].devnum;
2356 rdev = conf->mirrors[d].rdev;
2357 if (!r10_bio->devs[i].bio->bi_status) {
2358 /* We know that the bi_io_vec layout is the same for
2359 * both 'first' and 'i', so we just compare them.
2360 * All vec entries are PAGE_SIZE;
2362 int sectors = r10_bio->sectors;
2363 for (j = 0; j < vcnt; j++) {
2364 int len = PAGE_SIZE;
2365 if (sectors < (len / 512))
2366 len = sectors * 512;
2367 if (memcmp(page_address(fpages[j]),
2368 page_address(tpages[j]),
2375 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2376 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2377 /* Don't fix anything. */
2379 } else if (test_bit(FailFast, &rdev->flags)) {
2380 /* Just give up on this device */
2381 md_error(rdev->mddev, rdev);
2384 /* Ok, we need to write this bio, either to correct an
2385 * inconsistency or to correct an unreadable block.
2386 * First we need to fixup bv_offset, bv_len and
2387 * bi_vecs, as the read request might have corrupted these
2389 rp = get_resync_pages(tbio);
2392 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2394 rp->raid_bio = r10_bio;
2395 tbio->bi_private = rp;
2396 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2397 tbio->bi_end_io = end_sync_write;
2398 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2400 bio_copy_data(tbio, fbio);
2402 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2403 atomic_inc(&r10_bio->remaining);
2404 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2406 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2407 tbio->bi_opf |= MD_FAILFAST;
2408 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2409 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2410 submit_bio_noacct(tbio);
2413 /* Now write out to any replacement devices
2416 for (i = 0; i < conf->copies; i++) {
2419 tbio = r10_bio->devs[i].repl_bio;
2420 if (!tbio || !tbio->bi_end_io)
2422 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2423 && r10_bio->devs[i].bio != fbio)
2424 bio_copy_data(tbio, fbio);
2425 d = r10_bio->devs[i].devnum;
2426 atomic_inc(&r10_bio->remaining);
2427 md_sync_acct(conf->mirrors[d].replacement->bdev,
2429 submit_bio_noacct(tbio);
2433 if (atomic_dec_and_test(&r10_bio->remaining)) {
2434 md_done_sync(mddev, r10_bio->sectors, 1);
2440 * Now for the recovery code.
2441 * Recovery happens across physical sectors.
2442 * We recover all non-is_sync drives by finding the virtual address of
2443 * each, and then choose a working drive that also has that virt address.
2444 * There is a separate r10_bio for each non-in_sync drive.
2445 * Only the first two slots are in use. The first for reading,
2446 * The second for writing.
2449 static void fix_recovery_read_error(struct r10bio *r10_bio)
2451 /* We got a read error during recovery.
2452 * We repeat the read in smaller page-sized sections.
2453 * If a read succeeds, write it to the new device or record
2454 * a bad block if we cannot.
2455 * If a read fails, record a bad block on both old and
2458 struct mddev *mddev = r10_bio->mddev;
2459 struct r10conf *conf = mddev->private;
2460 struct bio *bio = r10_bio->devs[0].bio;
2462 int sectors = r10_bio->sectors;
2464 int dr = r10_bio->devs[0].devnum;
2465 int dw = r10_bio->devs[1].devnum;
2466 struct page **pages = get_resync_pages(bio)->pages;
2470 struct md_rdev *rdev;
2474 if (s > (PAGE_SIZE>>9))
2477 rdev = conf->mirrors[dr].rdev;
2478 addr = r10_bio->devs[0].addr + sect,
2479 ok = sync_page_io(rdev,
2483 REQ_OP_READ, 0, false);
2485 rdev = conf->mirrors[dw].rdev;
2486 addr = r10_bio->devs[1].addr + sect;
2487 ok = sync_page_io(rdev,
2491 REQ_OP_WRITE, 0, false);
2493 set_bit(WriteErrorSeen, &rdev->flags);
2494 if (!test_and_set_bit(WantReplacement,
2496 set_bit(MD_RECOVERY_NEEDED,
2497 &rdev->mddev->recovery);
2501 /* We don't worry if we cannot set a bad block -
2502 * it really is bad so there is no loss in not
2505 rdev_set_badblocks(rdev, addr, s, 0);
2507 if (rdev != conf->mirrors[dw].rdev) {
2508 /* need bad block on destination too */
2509 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2510 addr = r10_bio->devs[1].addr + sect;
2511 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2513 /* just abort the recovery */
2514 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2517 conf->mirrors[dw].recovery_disabled
2518 = mddev->recovery_disabled;
2519 set_bit(MD_RECOVERY_INTR,
2532 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2534 struct r10conf *conf = mddev->private;
2536 struct bio *wbio, *wbio2;
2538 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2539 fix_recovery_read_error(r10_bio);
2540 end_sync_request(r10_bio);
2545 * share the pages with the first bio
2546 * and submit the write request
2548 d = r10_bio->devs[1].devnum;
2549 wbio = r10_bio->devs[1].bio;
2550 wbio2 = r10_bio->devs[1].repl_bio;
2551 /* Need to test wbio2->bi_end_io before we call
2552 * submit_bio_noacct as if the former is NULL,
2553 * the latter is free to free wbio2.
2555 if (wbio2 && !wbio2->bi_end_io)
2557 if (wbio->bi_end_io) {
2558 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2559 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2560 submit_bio_noacct(wbio);
2563 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2564 md_sync_acct(conf->mirrors[d].replacement->bdev,
2565 bio_sectors(wbio2));
2566 submit_bio_noacct(wbio2);
2571 * Used by fix_read_error() to decay the per rdev read_errors.
2572 * We halve the read error count for every hour that has elapsed
2573 * since the last recorded read error.
2576 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2579 unsigned long hours_since_last;
2580 unsigned int read_errors = atomic_read(&rdev->read_errors);
2582 cur_time_mon = ktime_get_seconds();
2584 if (rdev->last_read_error == 0) {
2585 /* first time we've seen a read error */
2586 rdev->last_read_error = cur_time_mon;
2590 hours_since_last = (long)(cur_time_mon -
2591 rdev->last_read_error) / 3600;
2593 rdev->last_read_error = cur_time_mon;
2596 * if hours_since_last is > the number of bits in read_errors
2597 * just set read errors to 0. We do this to avoid
2598 * overflowing the shift of read_errors by hours_since_last.
2600 if (hours_since_last >= 8 * sizeof(read_errors))
2601 atomic_set(&rdev->read_errors, 0);
2603 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2606 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2607 int sectors, struct page *page, int rw)
2612 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2613 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2615 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2619 set_bit(WriteErrorSeen, &rdev->flags);
2620 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2621 set_bit(MD_RECOVERY_NEEDED,
2622 &rdev->mddev->recovery);
2624 /* need to record an error - either for the block or the device */
2625 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2626 md_error(rdev->mddev, rdev);
2631 * This is a kernel thread which:
2633 * 1. Retries failed read operations on working mirrors.
2634 * 2. Updates the raid superblock when problems encounter.
2635 * 3. Performs writes following reads for array synchronising.
2638 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2640 int sect = 0; /* Offset from r10_bio->sector */
2641 int sectors = r10_bio->sectors;
2642 struct md_rdev *rdev;
2643 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2644 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2646 /* still own a reference to this rdev, so it cannot
2647 * have been cleared recently.
2649 rdev = conf->mirrors[d].rdev;
2651 if (test_bit(Faulty, &rdev->flags))
2652 /* drive has already been failed, just ignore any
2653 more fix_read_error() attempts */
2656 check_decay_read_errors(mddev, rdev);
2657 atomic_inc(&rdev->read_errors);
2658 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2659 char b[BDEVNAME_SIZE];
2660 bdevname(rdev->bdev, b);
2662 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2664 atomic_read(&rdev->read_errors), max_read_errors);
2665 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2667 md_error(mddev, rdev);
2668 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2674 int sl = r10_bio->read_slot;
2678 if (s > (PAGE_SIZE>>9))
2686 d = r10_bio->devs[sl].devnum;
2687 rdev = rcu_dereference(conf->mirrors[d].rdev);
2689 test_bit(In_sync, &rdev->flags) &&
2690 !test_bit(Faulty, &rdev->flags) &&
2691 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2692 &first_bad, &bad_sectors) == 0) {
2693 atomic_inc(&rdev->nr_pending);
2695 success = sync_page_io(rdev,
2696 r10_bio->devs[sl].addr +
2700 REQ_OP_READ, 0, false);
2701 rdev_dec_pending(rdev, mddev);
2707 if (sl == conf->copies)
2709 } while (!success && sl != r10_bio->read_slot);
2713 /* Cannot read from anywhere, just mark the block
2714 * as bad on the first device to discourage future
2717 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2718 rdev = conf->mirrors[dn].rdev;
2720 if (!rdev_set_badblocks(
2722 r10_bio->devs[r10_bio->read_slot].addr
2725 md_error(mddev, rdev);
2726 r10_bio->devs[r10_bio->read_slot].bio
2733 /* write it back and re-read */
2735 while (sl != r10_bio->read_slot) {
2736 char b[BDEVNAME_SIZE];
2741 d = r10_bio->devs[sl].devnum;
2742 rdev = rcu_dereference(conf->mirrors[d].rdev);
2744 test_bit(Faulty, &rdev->flags) ||
2745 !test_bit(In_sync, &rdev->flags))
2748 atomic_inc(&rdev->nr_pending);
2750 if (r10_sync_page_io(rdev,
2751 r10_bio->devs[sl].addr +
2753 s, conf->tmppage, WRITE)
2755 /* Well, this device is dead */
2756 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2758 (unsigned long long)(
2760 choose_data_offset(r10_bio,
2762 bdevname(rdev->bdev, b));
2763 pr_notice("md/raid10:%s: %s: failing drive\n",
2765 bdevname(rdev->bdev, b));
2767 rdev_dec_pending(rdev, mddev);
2771 while (sl != r10_bio->read_slot) {
2772 char b[BDEVNAME_SIZE];
2777 d = r10_bio->devs[sl].devnum;
2778 rdev = rcu_dereference(conf->mirrors[d].rdev);
2780 test_bit(Faulty, &rdev->flags) ||
2781 !test_bit(In_sync, &rdev->flags))
2784 atomic_inc(&rdev->nr_pending);
2786 switch (r10_sync_page_io(rdev,
2787 r10_bio->devs[sl].addr +
2792 /* Well, this device is dead */
2793 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2795 (unsigned long long)(
2797 choose_data_offset(r10_bio, rdev)),
2798 bdevname(rdev->bdev, b));
2799 pr_notice("md/raid10:%s: %s: failing drive\n",
2801 bdevname(rdev->bdev, b));
2804 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2806 (unsigned long long)(
2808 choose_data_offset(r10_bio, rdev)),
2809 bdevname(rdev->bdev, b));
2810 atomic_add(s, &rdev->corrected_errors);
2813 rdev_dec_pending(rdev, mddev);
2823 static int narrow_write_error(struct r10bio *r10_bio, int i)
2825 struct bio *bio = r10_bio->master_bio;
2826 struct mddev *mddev = r10_bio->mddev;
2827 struct r10conf *conf = mddev->private;
2828 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2829 /* bio has the data to be written to slot 'i' where
2830 * we just recently had a write error.
2831 * We repeatedly clone the bio and trim down to one block,
2832 * then try the write. Where the write fails we record
2834 * It is conceivable that the bio doesn't exactly align with
2835 * blocks. We must handle this.
2837 * We currently own a reference to the rdev.
2843 int sect_to_write = r10_bio->sectors;
2846 if (rdev->badblocks.shift < 0)
2849 block_sectors = roundup(1 << rdev->badblocks.shift,
2850 bdev_logical_block_size(rdev->bdev) >> 9);
2851 sector = r10_bio->sector;
2852 sectors = ((r10_bio->sector + block_sectors)
2853 & ~(sector_t)(block_sectors - 1))
2856 while (sect_to_write) {
2859 if (sectors > sect_to_write)
2860 sectors = sect_to_write;
2861 /* Write at 'sector' for 'sectors' */
2862 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2863 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2864 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2865 wbio->bi_iter.bi_sector = wsector +
2866 choose_data_offset(r10_bio, rdev);
2867 bio_set_dev(wbio, rdev->bdev);
2868 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2870 if (submit_bio_wait(wbio) < 0)
2872 ok = rdev_set_badblocks(rdev, wsector,
2877 sect_to_write -= sectors;
2879 sectors = block_sectors;
2884 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2886 int slot = r10_bio->read_slot;
2888 struct r10conf *conf = mddev->private;
2889 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2891 /* we got a read error. Maybe the drive is bad. Maybe just
2892 * the block and we can fix it.
2893 * We freeze all other IO, and try reading the block from
2894 * other devices. When we find one, we re-write
2895 * and check it that fixes the read error.
2896 * This is all done synchronously while the array is
2899 bio = r10_bio->devs[slot].bio;
2901 r10_bio->devs[slot].bio = NULL;
2904 r10_bio->devs[slot].bio = IO_BLOCKED;
2905 else if (!test_bit(FailFast, &rdev->flags)) {
2906 freeze_array(conf, 1);
2907 fix_read_error(conf, mddev, r10_bio);
2908 unfreeze_array(conf);
2910 md_error(mddev, rdev);
2912 rdev_dec_pending(rdev, mddev);
2913 allow_barrier(conf);
2915 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2918 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2920 /* Some sort of write request has finished and it
2921 * succeeded in writing where we thought there was a
2922 * bad block. So forget the bad block.
2923 * Or possibly if failed and we need to record
2927 struct md_rdev *rdev;
2929 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2930 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2931 for (m = 0; m < conf->copies; m++) {
2932 int dev = r10_bio->devs[m].devnum;
2933 rdev = conf->mirrors[dev].rdev;
2934 if (r10_bio->devs[m].bio == NULL ||
2935 r10_bio->devs[m].bio->bi_end_io == NULL)
2937 if (!r10_bio->devs[m].bio->bi_status) {
2938 rdev_clear_badblocks(
2940 r10_bio->devs[m].addr,
2941 r10_bio->sectors, 0);
2943 if (!rdev_set_badblocks(
2945 r10_bio->devs[m].addr,
2946 r10_bio->sectors, 0))
2947 md_error(conf->mddev, rdev);
2949 rdev = conf->mirrors[dev].replacement;
2950 if (r10_bio->devs[m].repl_bio == NULL ||
2951 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2954 if (!r10_bio->devs[m].repl_bio->bi_status) {
2955 rdev_clear_badblocks(
2957 r10_bio->devs[m].addr,
2958 r10_bio->sectors, 0);
2960 if (!rdev_set_badblocks(
2962 r10_bio->devs[m].addr,
2963 r10_bio->sectors, 0))
2964 md_error(conf->mddev, rdev);
2970 for (m = 0; m < conf->copies; m++) {
2971 int dev = r10_bio->devs[m].devnum;
2972 struct bio *bio = r10_bio->devs[m].bio;
2973 rdev = conf->mirrors[dev].rdev;
2974 if (bio == IO_MADE_GOOD) {
2975 rdev_clear_badblocks(
2977 r10_bio->devs[m].addr,
2978 r10_bio->sectors, 0);
2979 rdev_dec_pending(rdev, conf->mddev);
2980 } else if (bio != NULL && bio->bi_status) {
2982 if (!narrow_write_error(r10_bio, m)) {
2983 md_error(conf->mddev, rdev);
2984 set_bit(R10BIO_Degraded,
2987 rdev_dec_pending(rdev, conf->mddev);
2989 bio = r10_bio->devs[m].repl_bio;
2990 rdev = conf->mirrors[dev].replacement;
2991 if (rdev && bio == IO_MADE_GOOD) {
2992 rdev_clear_badblocks(
2994 r10_bio->devs[m].addr,
2995 r10_bio->sectors, 0);
2996 rdev_dec_pending(rdev, conf->mddev);
3000 spin_lock_irq(&conf->device_lock);
3001 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
3003 spin_unlock_irq(&conf->device_lock);
3005 * In case freeze_array() is waiting for condition
3006 * nr_pending == nr_queued + extra to be true.
3008 wake_up(&conf->wait_barrier);
3009 md_wakeup_thread(conf->mddev->thread);
3011 if (test_bit(R10BIO_WriteError,
3013 close_write(r10_bio);
3014 raid_end_bio_io(r10_bio);
3019 static void raid10d(struct md_thread *thread)
3021 struct mddev *mddev = thread->mddev;
3022 struct r10bio *r10_bio;
3023 unsigned long flags;
3024 struct r10conf *conf = mddev->private;
3025 struct list_head *head = &conf->retry_list;
3026 struct blk_plug plug;
3028 md_check_recovery(mddev);
3030 if (!list_empty_careful(&conf->bio_end_io_list) &&
3031 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3033 spin_lock_irqsave(&conf->device_lock, flags);
3034 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3035 while (!list_empty(&conf->bio_end_io_list)) {
3036 list_move(conf->bio_end_io_list.prev, &tmp);
3040 spin_unlock_irqrestore(&conf->device_lock, flags);
3041 while (!list_empty(&tmp)) {
3042 r10_bio = list_first_entry(&tmp, struct r10bio,
3044 list_del(&r10_bio->retry_list);
3045 if (mddev->degraded)
3046 set_bit(R10BIO_Degraded, &r10_bio->state);
3048 if (test_bit(R10BIO_WriteError,
3050 close_write(r10_bio);
3051 raid_end_bio_io(r10_bio);
3055 blk_start_plug(&plug);
3058 flush_pending_writes(conf);
3060 spin_lock_irqsave(&conf->device_lock, flags);
3061 if (list_empty(head)) {
3062 spin_unlock_irqrestore(&conf->device_lock, flags);
3065 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3066 list_del(head->prev);
3068 spin_unlock_irqrestore(&conf->device_lock, flags);
3070 mddev = r10_bio->mddev;
3071 conf = mddev->private;
3072 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3073 test_bit(R10BIO_WriteError, &r10_bio->state))
3074 handle_write_completed(conf, r10_bio);
3075 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3076 reshape_request_write(mddev, r10_bio);
3077 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3078 sync_request_write(mddev, r10_bio);
3079 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3080 recovery_request_write(mddev, r10_bio);
3081 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3082 handle_read_error(mddev, r10_bio);
3087 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3088 md_check_recovery(mddev);
3090 blk_finish_plug(&plug);
3093 static int init_resync(struct r10conf *conf)
3097 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3098 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3099 conf->have_replacement = 0;
3100 for (i = 0; i < conf->geo.raid_disks; i++)
3101 if (conf->mirrors[i].replacement)
3102 conf->have_replacement = 1;
3103 ret = mempool_init(&conf->r10buf_pool, buffs,
3104 r10buf_pool_alloc, r10buf_pool_free, conf);
3107 conf->next_resync = 0;
3111 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3113 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3114 struct rsync_pages *rp;
3119 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3120 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3121 nalloc = conf->copies; /* resync */
3123 nalloc = 2; /* recovery */
3125 for (i = 0; i < nalloc; i++) {
3126 bio = r10bio->devs[i].bio;
3127 rp = bio->bi_private;
3129 bio->bi_private = rp;
3130 bio = r10bio->devs[i].repl_bio;
3132 rp = bio->bi_private;
3134 bio->bi_private = rp;
3141 * Set cluster_sync_high since we need other nodes to add the
3142 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3144 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3146 sector_t window_size;
3147 int extra_chunk, chunks;
3150 * First, here we define "stripe" as a unit which across
3151 * all member devices one time, so we get chunks by use
3152 * raid_disks / near_copies. Otherwise, if near_copies is
3153 * close to raid_disks, then resync window could increases
3154 * linearly with the increase of raid_disks, which means
3155 * we will suspend a really large IO window while it is not
3156 * necessary. If raid_disks is not divisible by near_copies,
3157 * an extra chunk is needed to ensure the whole "stripe" is
3161 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3162 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3166 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3169 * At least use a 32M window to align with raid1's resync window
3171 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3172 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3174 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3178 * perform a "sync" on one "block"
3180 * We need to make sure that no normal I/O request - particularly write
3181 * requests - conflict with active sync requests.
3183 * This is achieved by tracking pending requests and a 'barrier' concept
3184 * that can be installed to exclude normal IO requests.
3186 * Resync and recovery are handled very differently.
3187 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3189 * For resync, we iterate over virtual addresses, read all copies,
3190 * and update if there are differences. If only one copy is live,
3192 * For recovery, we iterate over physical addresses, read a good
3193 * value for each non-in_sync drive, and over-write.
3195 * So, for recovery we may have several outstanding complex requests for a
3196 * given address, one for each out-of-sync device. We model this by allocating
3197 * a number of r10_bio structures, one for each out-of-sync device.
3198 * As we setup these structures, we collect all bio's together into a list
3199 * which we then process collectively to add pages, and then process again
3200 * to pass to submit_bio_noacct.
3202 * The r10_bio structures are linked using a borrowed master_bio pointer.
3203 * This link is counted in ->remaining. When the r10_bio that points to NULL
3204 * has its remaining count decremented to 0, the whole complex operation
3209 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3212 struct r10conf *conf = mddev->private;
3213 struct r10bio *r10_bio;
3214 struct bio *biolist = NULL, *bio;
3215 sector_t max_sector, nr_sectors;
3218 sector_t sync_blocks;
3219 sector_t sectors_skipped = 0;
3220 int chunks_skipped = 0;
3221 sector_t chunk_mask = conf->geo.chunk_mask;
3224 if (!mempool_initialized(&conf->r10buf_pool))
3225 if (init_resync(conf))
3229 * Allow skipping a full rebuild for incremental assembly
3230 * of a clean array, like RAID1 does.
3232 if (mddev->bitmap == NULL &&
3233 mddev->recovery_cp == MaxSector &&
3234 mddev->reshape_position == MaxSector &&
3235 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3236 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3237 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3238 conf->fullsync == 0) {
3240 return mddev->dev_sectors - sector_nr;
3244 max_sector = mddev->dev_sectors;
3245 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3246 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3247 max_sector = mddev->resync_max_sectors;
3248 if (sector_nr >= max_sector) {
3249 conf->cluster_sync_low = 0;
3250 conf->cluster_sync_high = 0;
3252 /* If we aborted, we need to abort the
3253 * sync on the 'current' bitmap chucks (there can
3254 * be several when recovering multiple devices).
3255 * as we may have started syncing it but not finished.
3256 * We can find the current address in
3257 * mddev->curr_resync, but for recovery,
3258 * we need to convert that to several
3259 * virtual addresses.
3261 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3267 if (mddev->curr_resync < max_sector) { /* aborted */
3268 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3269 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3271 else for (i = 0; i < conf->geo.raid_disks; i++) {
3273 raid10_find_virt(conf, mddev->curr_resync, i);
3274 md_bitmap_end_sync(mddev->bitmap, sect,
3278 /* completed sync */
3279 if ((!mddev->bitmap || conf->fullsync)
3280 && conf->have_replacement
3281 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3282 /* Completed a full sync so the replacements
3283 * are now fully recovered.
3286 for (i = 0; i < conf->geo.raid_disks; i++) {
3287 struct md_rdev *rdev =
3288 rcu_dereference(conf->mirrors[i].replacement);
3290 rdev->recovery_offset = MaxSector;
3296 md_bitmap_close_sync(mddev->bitmap);
3299 return sectors_skipped;
3302 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3303 return reshape_request(mddev, sector_nr, skipped);
3305 if (chunks_skipped >= conf->geo.raid_disks) {
3306 /* if there has been nothing to do on any drive,
3307 * then there is nothing to do at all..
3310 return (max_sector - sector_nr) + sectors_skipped;
3313 if (max_sector > mddev->resync_max)
3314 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3316 /* make sure whole request will fit in a chunk - if chunks
3319 if (conf->geo.near_copies < conf->geo.raid_disks &&
3320 max_sector > (sector_nr | chunk_mask))
3321 max_sector = (sector_nr | chunk_mask) + 1;
3324 * If there is non-resync activity waiting for a turn, then let it
3325 * though before starting on this new sync request.
3327 if (conf->nr_waiting)
3328 schedule_timeout_uninterruptible(1);
3330 /* Again, very different code for resync and recovery.
3331 * Both must result in an r10bio with a list of bios that
3332 * have bi_end_io, bi_sector, bi_bdev set,
3333 * and bi_private set to the r10bio.
3334 * For recovery, we may actually create several r10bios
3335 * with 2 bios in each, that correspond to the bios in the main one.
3336 * In this case, the subordinate r10bios link back through a
3337 * borrowed master_bio pointer, and the counter in the master
3338 * includes a ref from each subordinate.
3340 /* First, we decide what to do and set ->bi_end_io
3341 * To end_sync_read if we want to read, and
3342 * end_sync_write if we will want to write.
3345 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3346 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3347 /* recovery... the complicated one */
3351 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3357 int need_recover = 0;
3358 int need_replace = 0;
3359 struct raid10_info *mirror = &conf->mirrors[i];
3360 struct md_rdev *mrdev, *mreplace;
3363 mrdev = rcu_dereference(mirror->rdev);
3364 mreplace = rcu_dereference(mirror->replacement);
3366 if (mrdev != NULL &&
3367 !test_bit(Faulty, &mrdev->flags) &&
3368 !test_bit(In_sync, &mrdev->flags))
3370 if (mreplace != NULL &&
3371 !test_bit(Faulty, &mreplace->flags))
3374 if (!need_recover && !need_replace) {
3380 /* want to reconstruct this device */
3382 sect = raid10_find_virt(conf, sector_nr, i);
3383 if (sect >= mddev->resync_max_sectors) {
3384 /* last stripe is not complete - don't
3385 * try to recover this sector.
3390 if (mreplace && test_bit(Faulty, &mreplace->flags))
3392 /* Unless we are doing a full sync, or a replacement
3393 * we only need to recover the block if it is set in
3396 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3398 if (sync_blocks < max_sync)
3399 max_sync = sync_blocks;
3403 /* yep, skip the sync_blocks here, but don't assume
3404 * that there will never be anything to do here
3406 chunks_skipped = -1;
3410 atomic_inc(&mrdev->nr_pending);
3412 atomic_inc(&mreplace->nr_pending);
3415 r10_bio = raid10_alloc_init_r10buf(conf);
3417 raise_barrier(conf, rb2 != NULL);
3418 atomic_set(&r10_bio->remaining, 0);
3420 r10_bio->master_bio = (struct bio*)rb2;
3422 atomic_inc(&rb2->remaining);
3423 r10_bio->mddev = mddev;
3424 set_bit(R10BIO_IsRecover, &r10_bio->state);
3425 r10_bio->sector = sect;
3427 raid10_find_phys(conf, r10_bio);
3429 /* Need to check if the array will still be
3433 for (j = 0; j < conf->geo.raid_disks; j++) {
3434 struct md_rdev *rdev = rcu_dereference(
3435 conf->mirrors[j].rdev);
3436 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3442 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3443 &sync_blocks, still_degraded);
3446 for (j=0; j<conf->copies;j++) {
3448 int d = r10_bio->devs[j].devnum;
3449 sector_t from_addr, to_addr;
3450 struct md_rdev *rdev =
3451 rcu_dereference(conf->mirrors[d].rdev);
3452 sector_t sector, first_bad;
3455 !test_bit(In_sync, &rdev->flags))
3457 /* This is where we read from */
3459 sector = r10_bio->devs[j].addr;
3461 if (is_badblock(rdev, sector, max_sync,
3462 &first_bad, &bad_sectors)) {
3463 if (first_bad > sector)
3464 max_sync = first_bad - sector;
3466 bad_sectors -= (sector
3468 if (max_sync > bad_sectors)
3469 max_sync = bad_sectors;
3473 bio = r10_bio->devs[0].bio;
3474 bio->bi_next = biolist;
3476 bio->bi_end_io = end_sync_read;
3477 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3478 if (test_bit(FailFast, &rdev->flags))
3479 bio->bi_opf |= MD_FAILFAST;
3480 from_addr = r10_bio->devs[j].addr;
3481 bio->bi_iter.bi_sector = from_addr +
3483 bio_set_dev(bio, rdev->bdev);
3484 atomic_inc(&rdev->nr_pending);
3485 /* and we write to 'i' (if not in_sync) */
3487 for (k=0; k<conf->copies; k++)
3488 if (r10_bio->devs[k].devnum == i)
3490 BUG_ON(k == conf->copies);
3491 to_addr = r10_bio->devs[k].addr;
3492 r10_bio->devs[0].devnum = d;
3493 r10_bio->devs[0].addr = from_addr;
3494 r10_bio->devs[1].devnum = i;
3495 r10_bio->devs[1].addr = to_addr;
3498 bio = r10_bio->devs[1].bio;
3499 bio->bi_next = biolist;
3501 bio->bi_end_io = end_sync_write;
3502 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3503 bio->bi_iter.bi_sector = to_addr
3504 + mrdev->data_offset;
3505 bio_set_dev(bio, mrdev->bdev);
3506 atomic_inc(&r10_bio->remaining);
3508 r10_bio->devs[1].bio->bi_end_io = NULL;
3510 /* and maybe write to replacement */
3511 bio = r10_bio->devs[1].repl_bio;
3513 bio->bi_end_io = NULL;
3514 /* Note: if need_replace, then bio
3515 * cannot be NULL as r10buf_pool_alloc will
3516 * have allocated it.
3520 bio->bi_next = biolist;
3522 bio->bi_end_io = end_sync_write;
3523 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3524 bio->bi_iter.bi_sector = to_addr +
3525 mreplace->data_offset;
3526 bio_set_dev(bio, mreplace->bdev);
3527 atomic_inc(&r10_bio->remaining);
3531 if (j == conf->copies) {
3532 /* Cannot recover, so abort the recovery or
3533 * record a bad block */
3535 /* problem is that there are bad blocks
3536 * on other device(s)
3539 for (k = 0; k < conf->copies; k++)
3540 if (r10_bio->devs[k].devnum == i)
3542 if (!test_bit(In_sync,
3544 && !rdev_set_badblocks(
3546 r10_bio->devs[k].addr,
3550 !rdev_set_badblocks(
3552 r10_bio->devs[k].addr,
3557 if (!test_and_set_bit(MD_RECOVERY_INTR,
3559 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3561 mirror->recovery_disabled
3562 = mddev->recovery_disabled;
3566 atomic_dec(&rb2->remaining);
3568 rdev_dec_pending(mrdev, mddev);
3570 rdev_dec_pending(mreplace, mddev);
3573 rdev_dec_pending(mrdev, mddev);
3575 rdev_dec_pending(mreplace, mddev);
3576 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3577 /* Only want this if there is elsewhere to
3578 * read from. 'j' is currently the first
3582 for (; j < conf->copies; j++) {
3583 int d = r10_bio->devs[j].devnum;
3584 if (conf->mirrors[d].rdev &&
3586 &conf->mirrors[d].rdev->flags))
3590 r10_bio->devs[0].bio->bi_opf
3594 if (biolist == NULL) {
3596 struct r10bio *rb2 = r10_bio;
3597 r10_bio = (struct r10bio*) rb2->master_bio;
3598 rb2->master_bio = NULL;
3604 /* resync. Schedule a read for every block at this virt offset */
3608 * Since curr_resync_completed could probably not update in
3609 * time, and we will set cluster_sync_low based on it.
3610 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3611 * safety reason, which ensures curr_resync_completed is
3612 * updated in bitmap_cond_end_sync.
3614 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3615 mddev_is_clustered(mddev) &&
3616 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3618 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3619 &sync_blocks, mddev->degraded) &&
3620 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3621 &mddev->recovery)) {
3622 /* We can skip this block */
3624 return sync_blocks + sectors_skipped;
3626 if (sync_blocks < max_sync)
3627 max_sync = sync_blocks;
3628 r10_bio = raid10_alloc_init_r10buf(conf);
3631 r10_bio->mddev = mddev;
3632 atomic_set(&r10_bio->remaining, 0);
3633 raise_barrier(conf, 0);
3634 conf->next_resync = sector_nr;
3636 r10_bio->master_bio = NULL;
3637 r10_bio->sector = sector_nr;
3638 set_bit(R10BIO_IsSync, &r10_bio->state);
3639 raid10_find_phys(conf, r10_bio);
3640 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3642 for (i = 0; i < conf->copies; i++) {
3643 int d = r10_bio->devs[i].devnum;
3644 sector_t first_bad, sector;
3646 struct md_rdev *rdev;
3648 if (r10_bio->devs[i].repl_bio)
3649 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3651 bio = r10_bio->devs[i].bio;
3652 bio->bi_status = BLK_STS_IOERR;
3654 rdev = rcu_dereference(conf->mirrors[d].rdev);
3655 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3659 sector = r10_bio->devs[i].addr;
3660 if (is_badblock(rdev, sector, max_sync,
3661 &first_bad, &bad_sectors)) {
3662 if (first_bad > sector)
3663 max_sync = first_bad - sector;
3665 bad_sectors -= (sector - first_bad);
3666 if (max_sync > bad_sectors)
3667 max_sync = bad_sectors;
3672 atomic_inc(&rdev->nr_pending);
3673 atomic_inc(&r10_bio->remaining);
3674 bio->bi_next = biolist;
3676 bio->bi_end_io = end_sync_read;
3677 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3678 if (test_bit(FailFast, &rdev->flags))
3679 bio->bi_opf |= MD_FAILFAST;
3680 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3681 bio_set_dev(bio, rdev->bdev);
3684 rdev = rcu_dereference(conf->mirrors[d].replacement);
3685 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3689 atomic_inc(&rdev->nr_pending);
3691 /* Need to set up for writing to the replacement */
3692 bio = r10_bio->devs[i].repl_bio;
3693 bio->bi_status = BLK_STS_IOERR;
3695 sector = r10_bio->devs[i].addr;
3696 bio->bi_next = biolist;
3698 bio->bi_end_io = end_sync_write;
3699 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3700 if (test_bit(FailFast, &rdev->flags))
3701 bio->bi_opf |= MD_FAILFAST;
3702 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3703 bio_set_dev(bio, rdev->bdev);
3709 for (i=0; i<conf->copies; i++) {
3710 int d = r10_bio->devs[i].devnum;
3711 if (r10_bio->devs[i].bio->bi_end_io)
3712 rdev_dec_pending(conf->mirrors[d].rdev,
3714 if (r10_bio->devs[i].repl_bio &&
3715 r10_bio->devs[i].repl_bio->bi_end_io)
3717 conf->mirrors[d].replacement,
3727 if (sector_nr + max_sync < max_sector)
3728 max_sector = sector_nr + max_sync;
3731 int len = PAGE_SIZE;
3732 if (sector_nr + (len>>9) > max_sector)
3733 len = (max_sector - sector_nr) << 9;
3736 for (bio= biolist ; bio ; bio=bio->bi_next) {
3737 struct resync_pages *rp = get_resync_pages(bio);
3738 page = resync_fetch_page(rp, page_idx);
3740 * won't fail because the vec table is big enough
3741 * to hold all these pages
3743 bio_add_page(bio, page, len, 0);
3745 nr_sectors += len>>9;
3746 sector_nr += len>>9;
3747 } while (++page_idx < RESYNC_PAGES);
3748 r10_bio->sectors = nr_sectors;
3750 if (mddev_is_clustered(mddev) &&
3751 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3752 /* It is resync not recovery */
3753 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3754 conf->cluster_sync_low = mddev->curr_resync_completed;
3755 raid10_set_cluster_sync_high(conf);
3756 /* Send resync message */
3757 md_cluster_ops->resync_info_update(mddev,
3758 conf->cluster_sync_low,
3759 conf->cluster_sync_high);
3761 } else if (mddev_is_clustered(mddev)) {
3762 /* This is recovery not resync */
3763 sector_t sect_va1, sect_va2;
3764 bool broadcast_msg = false;
3766 for (i = 0; i < conf->geo.raid_disks; i++) {
3768 * sector_nr is a device address for recovery, so we
3769 * need translate it to array address before compare
3770 * with cluster_sync_high.
3772 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3774 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3775 broadcast_msg = true;
3777 * curr_resync_completed is similar as
3778 * sector_nr, so make the translation too.
3780 sect_va2 = raid10_find_virt(conf,
3781 mddev->curr_resync_completed, i);
3783 if (conf->cluster_sync_low == 0 ||
3784 conf->cluster_sync_low > sect_va2)
3785 conf->cluster_sync_low = sect_va2;
3788 if (broadcast_msg) {
3789 raid10_set_cluster_sync_high(conf);
3790 md_cluster_ops->resync_info_update(mddev,
3791 conf->cluster_sync_low,
3792 conf->cluster_sync_high);
3798 biolist = biolist->bi_next;
3800 bio->bi_next = NULL;
3801 r10_bio = get_resync_r10bio(bio);
3802 r10_bio->sectors = nr_sectors;
3804 if (bio->bi_end_io == end_sync_read) {
3805 md_sync_acct_bio(bio, nr_sectors);
3807 submit_bio_noacct(bio);
3811 if (sectors_skipped)
3812 /* pretend they weren't skipped, it makes
3813 * no important difference in this case
3815 md_done_sync(mddev, sectors_skipped, 1);
3817 return sectors_skipped + nr_sectors;
3819 /* There is nowhere to write, so all non-sync
3820 * drives must be failed or in resync, all drives
3821 * have a bad block, so try the next chunk...
3823 if (sector_nr + max_sync < max_sector)
3824 max_sector = sector_nr + max_sync;
3826 sectors_skipped += (max_sector - sector_nr);
3828 sector_nr = max_sector;
3833 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3836 struct r10conf *conf = mddev->private;
3839 raid_disks = min(conf->geo.raid_disks,
3840 conf->prev.raid_disks);
3842 sectors = conf->dev_sectors;
3844 size = sectors >> conf->geo.chunk_shift;
3845 sector_div(size, conf->geo.far_copies);
3846 size = size * raid_disks;
3847 sector_div(size, conf->geo.near_copies);
3849 return size << conf->geo.chunk_shift;
3852 static void calc_sectors(struct r10conf *conf, sector_t size)
3854 /* Calculate the number of sectors-per-device that will
3855 * actually be used, and set conf->dev_sectors and
3859 size = size >> conf->geo.chunk_shift;
3860 sector_div(size, conf->geo.far_copies);
3861 size = size * conf->geo.raid_disks;
3862 sector_div(size, conf->geo.near_copies);
3863 /* 'size' is now the number of chunks in the array */
3864 /* calculate "used chunks per device" */
3865 size = size * conf->copies;
3867 /* We need to round up when dividing by raid_disks to
3868 * get the stride size.
3870 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3872 conf->dev_sectors = size << conf->geo.chunk_shift;
3874 if (conf->geo.far_offset)
3875 conf->geo.stride = 1 << conf->geo.chunk_shift;
3877 sector_div(size, conf->geo.far_copies);
3878 conf->geo.stride = size << conf->geo.chunk_shift;
3882 enum geo_type {geo_new, geo_old, geo_start};
3883 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3886 int layout, chunk, disks;
3889 layout = mddev->layout;
3890 chunk = mddev->chunk_sectors;
3891 disks = mddev->raid_disks - mddev->delta_disks;
3894 layout = mddev->new_layout;
3895 chunk = mddev->new_chunk_sectors;
3896 disks = mddev->raid_disks;
3898 default: /* avoid 'may be unused' warnings */
3899 case geo_start: /* new when starting reshape - raid_disks not
3901 layout = mddev->new_layout;
3902 chunk = mddev->new_chunk_sectors;
3903 disks = mddev->raid_disks + mddev->delta_disks;
3908 if (chunk < (PAGE_SIZE >> 9) ||
3909 !is_power_of_2(chunk))
3912 fc = (layout >> 8) & 255;
3913 fo = layout & (1<<16);
3914 geo->raid_disks = disks;
3915 geo->near_copies = nc;
3916 geo->far_copies = fc;
3917 geo->far_offset = fo;
3918 switch (layout >> 17) {
3919 case 0: /* original layout. simple but not always optimal */
3920 geo->far_set_size = disks;
3922 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3923 * actually using this, but leave code here just in case.*/
3924 geo->far_set_size = disks/fc;
3925 WARN(geo->far_set_size < fc,
3926 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3928 case 2: /* "improved" layout fixed to match documentation */
3929 geo->far_set_size = fc * nc;
3931 default: /* Not a valid layout */
3934 geo->chunk_mask = chunk - 1;
3935 geo->chunk_shift = ffz(~chunk);
3939 static struct r10conf *setup_conf(struct mddev *mddev)
3941 struct r10conf *conf = NULL;
3946 copies = setup_geo(&geo, mddev, geo_new);
3949 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3950 mdname(mddev), PAGE_SIZE);
3954 if (copies < 2 || copies > mddev->raid_disks) {
3955 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3956 mdname(mddev), mddev->new_layout);
3961 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3965 /* FIXME calc properly */
3966 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3967 sizeof(struct raid10_info),
3972 conf->tmppage = alloc_page(GFP_KERNEL);
3977 conf->copies = copies;
3978 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3979 rbio_pool_free, conf);
3983 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3987 calc_sectors(conf, mddev->dev_sectors);
3988 if (mddev->reshape_position == MaxSector) {
3989 conf->prev = conf->geo;
3990 conf->reshape_progress = MaxSector;
3992 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3996 conf->reshape_progress = mddev->reshape_position;
3997 if (conf->prev.far_offset)
3998 conf->prev.stride = 1 << conf->prev.chunk_shift;
4000 /* far_copies must be 1 */
4001 conf->prev.stride = conf->dev_sectors;
4003 conf->reshape_safe = conf->reshape_progress;
4004 spin_lock_init(&conf->device_lock);
4005 INIT_LIST_HEAD(&conf->retry_list);
4006 INIT_LIST_HEAD(&conf->bio_end_io_list);
4008 spin_lock_init(&conf->resync_lock);
4009 init_waitqueue_head(&conf->wait_barrier);
4010 atomic_set(&conf->nr_pending, 0);
4013 conf->thread = md_register_thread(raid10d, mddev, "raid10");
4017 conf->mddev = mddev;
4022 mempool_exit(&conf->r10bio_pool);
4023 kfree(conf->mirrors);
4024 safe_put_page(conf->tmppage);
4025 bioset_exit(&conf->bio_split);
4028 return ERR_PTR(err);
4031 static void raid10_set_io_opt(struct r10conf *conf)
4033 int raid_disks = conf->geo.raid_disks;
4035 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4036 raid_disks /= conf->geo.near_copies;
4037 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4041 static int raid10_run(struct mddev *mddev)
4043 struct r10conf *conf;
4045 struct raid10_info *disk;
4046 struct md_rdev *rdev;
4048 sector_t min_offset_diff = 0;
4050 bool discard_supported = false;
4052 if (mddev_init_writes_pending(mddev) < 0)
4055 if (mddev->private == NULL) {
4056 conf = setup_conf(mddev);
4058 return PTR_ERR(conf);
4059 mddev->private = conf;
4061 conf = mddev->private;
4065 if (mddev_is_clustered(conf->mddev)) {
4068 fc = (mddev->layout >> 8) & 255;
4069 fo = mddev->layout & (1<<16);
4070 if (fc > 1 || fo > 0) {
4071 pr_err("only near layout is supported by clustered"
4077 mddev->thread = conf->thread;
4078 conf->thread = NULL;
4081 blk_queue_max_discard_sectors(mddev->queue,
4083 blk_queue_max_write_same_sectors(mddev->queue, 0);
4084 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4085 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4086 raid10_set_io_opt(conf);
4089 rdev_for_each(rdev, mddev) {
4092 disk_idx = rdev->raid_disk;
4095 if (disk_idx >= conf->geo.raid_disks &&
4096 disk_idx >= conf->prev.raid_disks)
4098 disk = conf->mirrors + disk_idx;
4100 if (test_bit(Replacement, &rdev->flags)) {
4101 if (disk->replacement)
4103 disk->replacement = rdev;
4109 diff = (rdev->new_data_offset - rdev->data_offset);
4110 if (!mddev->reshape_backwards)
4114 if (first || diff < min_offset_diff)
4115 min_offset_diff = diff;
4118 disk_stack_limits(mddev->gendisk, rdev->bdev,
4119 rdev->data_offset << 9);
4121 disk->head_position = 0;
4123 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
4124 discard_supported = true;
4129 if (discard_supported)
4130 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
4133 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
4136 /* need to check that every block has at least one working mirror */
4137 if (!enough(conf, -1)) {
4138 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4143 if (conf->reshape_progress != MaxSector) {
4144 /* must ensure that shape change is supported */
4145 if (conf->geo.far_copies != 1 &&
4146 conf->geo.far_offset == 0)
4148 if (conf->prev.far_copies != 1 &&
4149 conf->prev.far_offset == 0)
4153 mddev->degraded = 0;
4155 i < conf->geo.raid_disks
4156 || i < conf->prev.raid_disks;
4159 disk = conf->mirrors + i;
4161 if (!disk->rdev && disk->replacement) {
4162 /* The replacement is all we have - use it */
4163 disk->rdev = disk->replacement;
4164 disk->replacement = NULL;
4165 clear_bit(Replacement, &disk->rdev->flags);
4169 !test_bit(In_sync, &disk->rdev->flags)) {
4170 disk->head_position = 0;
4173 disk->rdev->saved_raid_disk < 0)
4177 if (disk->replacement &&
4178 !test_bit(In_sync, &disk->replacement->flags) &&
4179 disk->replacement->saved_raid_disk < 0) {
4183 disk->recovery_disabled = mddev->recovery_disabled - 1;
4186 if (mddev->recovery_cp != MaxSector)
4187 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4189 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4190 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4191 conf->geo.raid_disks);
4193 * Ok, everything is just fine now
4195 mddev->dev_sectors = conf->dev_sectors;
4196 size = raid10_size(mddev, 0, 0);
4197 md_set_array_sectors(mddev, size);
4198 mddev->resync_max_sectors = size;
4199 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4201 if (md_integrity_register(mddev))
4204 if (conf->reshape_progress != MaxSector) {
4205 unsigned long before_length, after_length;
4207 before_length = ((1 << conf->prev.chunk_shift) *
4208 conf->prev.far_copies);
4209 after_length = ((1 << conf->geo.chunk_shift) *
4210 conf->geo.far_copies);
4212 if (max(before_length, after_length) > min_offset_diff) {
4213 /* This cannot work */
4214 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4217 conf->offset_diff = min_offset_diff;
4219 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4220 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4221 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4222 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4223 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4225 if (!mddev->sync_thread)
4232 md_unregister_thread(&mddev->thread);
4233 mempool_exit(&conf->r10bio_pool);
4234 safe_put_page(conf->tmppage);
4235 kfree(conf->mirrors);
4237 mddev->private = NULL;
4242 static void raid10_free(struct mddev *mddev, void *priv)
4244 struct r10conf *conf = priv;
4246 mempool_exit(&conf->r10bio_pool);
4247 safe_put_page(conf->tmppage);
4248 kfree(conf->mirrors);
4249 kfree(conf->mirrors_old);
4250 kfree(conf->mirrors_new);
4251 bioset_exit(&conf->bio_split);
4255 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4257 struct r10conf *conf = mddev->private;
4260 raise_barrier(conf, 0);
4262 lower_barrier(conf);
4265 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4267 /* Resize of 'far' arrays is not supported.
4268 * For 'near' and 'offset' arrays we can set the
4269 * number of sectors used to be an appropriate multiple
4270 * of the chunk size.
4271 * For 'offset', this is far_copies*chunksize.
4272 * For 'near' the multiplier is the LCM of
4273 * near_copies and raid_disks.
4274 * So if far_copies > 1 && !far_offset, fail.
4275 * Else find LCM(raid_disks, near_copy)*far_copies and
4276 * multiply by chunk_size. Then round to this number.
4277 * This is mostly done by raid10_size()
4279 struct r10conf *conf = mddev->private;
4280 sector_t oldsize, size;
4282 if (mddev->reshape_position != MaxSector)
4285 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4288 oldsize = raid10_size(mddev, 0, 0);
4289 size = raid10_size(mddev, sectors, 0);
4290 if (mddev->external_size &&
4291 mddev->array_sectors > size)
4293 if (mddev->bitmap) {
4294 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4298 md_set_array_sectors(mddev, size);
4299 if (sectors > mddev->dev_sectors &&
4300 mddev->recovery_cp > oldsize) {
4301 mddev->recovery_cp = oldsize;
4302 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4304 calc_sectors(conf, sectors);
4305 mddev->dev_sectors = conf->dev_sectors;
4306 mddev->resync_max_sectors = size;
4310 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4312 struct md_rdev *rdev;
4313 struct r10conf *conf;
4315 if (mddev->degraded > 0) {
4316 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4318 return ERR_PTR(-EINVAL);
4320 sector_div(size, devs);
4322 /* Set new parameters */
4323 mddev->new_level = 10;
4324 /* new layout: far_copies = 1, near_copies = 2 */
4325 mddev->new_layout = (1<<8) + 2;
4326 mddev->new_chunk_sectors = mddev->chunk_sectors;
4327 mddev->delta_disks = mddev->raid_disks;
4328 mddev->raid_disks *= 2;
4329 /* make sure it will be not marked as dirty */
4330 mddev->recovery_cp = MaxSector;
4331 mddev->dev_sectors = size;
4333 conf = setup_conf(mddev);
4334 if (!IS_ERR(conf)) {
4335 rdev_for_each(rdev, mddev)
4336 if (rdev->raid_disk >= 0) {
4337 rdev->new_raid_disk = rdev->raid_disk * 2;
4338 rdev->sectors = size;
4346 static void *raid10_takeover(struct mddev *mddev)
4348 struct r0conf *raid0_conf;
4350 /* raid10 can take over:
4351 * raid0 - providing it has only two drives
4353 if (mddev->level == 0) {
4354 /* for raid0 takeover only one zone is supported */
4355 raid0_conf = mddev->private;
4356 if (raid0_conf->nr_strip_zones > 1) {
4357 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4359 return ERR_PTR(-EINVAL);
4361 return raid10_takeover_raid0(mddev,
4362 raid0_conf->strip_zone->zone_end,
4363 raid0_conf->strip_zone->nb_dev);
4365 return ERR_PTR(-EINVAL);
4368 static int raid10_check_reshape(struct mddev *mddev)
4370 /* Called when there is a request to change
4371 * - layout (to ->new_layout)
4372 * - chunk size (to ->new_chunk_sectors)
4373 * - raid_disks (by delta_disks)
4374 * or when trying to restart a reshape that was ongoing.
4376 * We need to validate the request and possibly allocate
4377 * space if that might be an issue later.
4379 * Currently we reject any reshape of a 'far' mode array,
4380 * allow chunk size to change if new is generally acceptable,
4381 * allow raid_disks to increase, and allow
4382 * a switch between 'near' mode and 'offset' mode.
4384 struct r10conf *conf = mddev->private;
4387 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4390 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4391 /* mustn't change number of copies */
4393 if (geo.far_copies > 1 && !geo.far_offset)
4394 /* Cannot switch to 'far' mode */
4397 if (mddev->array_sectors & geo.chunk_mask)
4398 /* not factor of array size */
4401 if (!enough(conf, -1))
4404 kfree(conf->mirrors_new);
4405 conf->mirrors_new = NULL;
4406 if (mddev->delta_disks > 0) {
4407 /* allocate new 'mirrors' list */
4409 kcalloc(mddev->raid_disks + mddev->delta_disks,
4410 sizeof(struct raid10_info),
4412 if (!conf->mirrors_new)
4419 * Need to check if array has failed when deciding whether to:
4421 * - remove non-faulty devices
4424 * This determination is simple when no reshape is happening.
4425 * However if there is a reshape, we need to carefully check
4426 * both the before and after sections.
4427 * This is because some failed devices may only affect one
4428 * of the two sections, and some non-in_sync devices may
4429 * be insync in the section most affected by failed devices.
4431 static int calc_degraded(struct r10conf *conf)
4433 int degraded, degraded2;
4438 /* 'prev' section first */
4439 for (i = 0; i < conf->prev.raid_disks; i++) {
4440 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4441 if (!rdev || test_bit(Faulty, &rdev->flags))
4443 else if (!test_bit(In_sync, &rdev->flags))
4444 /* When we can reduce the number of devices in
4445 * an array, this might not contribute to
4446 * 'degraded'. It does now.
4451 if (conf->geo.raid_disks == conf->prev.raid_disks)
4455 for (i = 0; i < conf->geo.raid_disks; i++) {
4456 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4457 if (!rdev || test_bit(Faulty, &rdev->flags))
4459 else if (!test_bit(In_sync, &rdev->flags)) {
4460 /* If reshape is increasing the number of devices,
4461 * this section has already been recovered, so
4462 * it doesn't contribute to degraded.
4465 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4470 if (degraded2 > degraded)
4475 static int raid10_start_reshape(struct mddev *mddev)
4477 /* A 'reshape' has been requested. This commits
4478 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4479 * This also checks if there are enough spares and adds them
4481 * We currently require enough spares to make the final
4482 * array non-degraded. We also require that the difference
4483 * between old and new data_offset - on each device - is
4484 * enough that we never risk over-writing.
4487 unsigned long before_length, after_length;
4488 sector_t min_offset_diff = 0;
4491 struct r10conf *conf = mddev->private;
4492 struct md_rdev *rdev;
4496 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4499 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4502 before_length = ((1 << conf->prev.chunk_shift) *
4503 conf->prev.far_copies);
4504 after_length = ((1 << conf->geo.chunk_shift) *
4505 conf->geo.far_copies);
4507 rdev_for_each(rdev, mddev) {
4508 if (!test_bit(In_sync, &rdev->flags)
4509 && !test_bit(Faulty, &rdev->flags))
4511 if (rdev->raid_disk >= 0) {
4512 long long diff = (rdev->new_data_offset
4513 - rdev->data_offset);
4514 if (!mddev->reshape_backwards)
4518 if (first || diff < min_offset_diff)
4519 min_offset_diff = diff;
4524 if (max(before_length, after_length) > min_offset_diff)
4527 if (spares < mddev->delta_disks)
4530 conf->offset_diff = min_offset_diff;
4531 spin_lock_irq(&conf->device_lock);
4532 if (conf->mirrors_new) {
4533 memcpy(conf->mirrors_new, conf->mirrors,
4534 sizeof(struct raid10_info)*conf->prev.raid_disks);
4536 kfree(conf->mirrors_old);
4537 conf->mirrors_old = conf->mirrors;
4538 conf->mirrors = conf->mirrors_new;
4539 conf->mirrors_new = NULL;
4541 setup_geo(&conf->geo, mddev, geo_start);
4543 if (mddev->reshape_backwards) {
4544 sector_t size = raid10_size(mddev, 0, 0);
4545 if (size < mddev->array_sectors) {
4546 spin_unlock_irq(&conf->device_lock);
4547 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4551 mddev->resync_max_sectors = size;
4552 conf->reshape_progress = size;
4554 conf->reshape_progress = 0;
4555 conf->reshape_safe = conf->reshape_progress;
4556 spin_unlock_irq(&conf->device_lock);
4558 if (mddev->delta_disks && mddev->bitmap) {
4559 struct mdp_superblock_1 *sb = NULL;
4560 sector_t oldsize, newsize;
4562 oldsize = raid10_size(mddev, 0, 0);
4563 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4565 if (!mddev_is_clustered(mddev)) {
4566 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4573 rdev_for_each(rdev, mddev) {
4574 if (rdev->raid_disk > -1 &&
4575 !test_bit(Faulty, &rdev->flags))
4576 sb = page_address(rdev->sb_page);
4580 * some node is already performing reshape, and no need to
4581 * call md_bitmap_resize again since it should be called when
4582 * receiving BITMAP_RESIZE msg
4584 if ((sb && (le32_to_cpu(sb->feature_map) &
4585 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4588 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4592 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4594 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4599 if (mddev->delta_disks > 0) {
4600 rdev_for_each(rdev, mddev)
4601 if (rdev->raid_disk < 0 &&
4602 !test_bit(Faulty, &rdev->flags)) {
4603 if (raid10_add_disk(mddev, rdev) == 0) {
4604 if (rdev->raid_disk >=
4605 conf->prev.raid_disks)
4606 set_bit(In_sync, &rdev->flags);
4608 rdev->recovery_offset = 0;
4610 /* Failure here is OK */
4611 sysfs_link_rdev(mddev, rdev);
4613 } else if (rdev->raid_disk >= conf->prev.raid_disks
4614 && !test_bit(Faulty, &rdev->flags)) {
4615 /* This is a spare that was manually added */
4616 set_bit(In_sync, &rdev->flags);
4619 /* When a reshape changes the number of devices,
4620 * ->degraded is measured against the larger of the
4621 * pre and post numbers.
4623 spin_lock_irq(&conf->device_lock);
4624 mddev->degraded = calc_degraded(conf);
4625 spin_unlock_irq(&conf->device_lock);
4626 mddev->raid_disks = conf->geo.raid_disks;
4627 mddev->reshape_position = conf->reshape_progress;
4628 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4630 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4631 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4632 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4633 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4634 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4636 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4638 if (!mddev->sync_thread) {
4642 conf->reshape_checkpoint = jiffies;
4643 md_wakeup_thread(mddev->sync_thread);
4644 md_new_event(mddev);
4648 mddev->recovery = 0;
4649 spin_lock_irq(&conf->device_lock);
4650 conf->geo = conf->prev;
4651 mddev->raid_disks = conf->geo.raid_disks;
4652 rdev_for_each(rdev, mddev)
4653 rdev->new_data_offset = rdev->data_offset;
4655 conf->reshape_progress = MaxSector;
4656 conf->reshape_safe = MaxSector;
4657 mddev->reshape_position = MaxSector;
4658 spin_unlock_irq(&conf->device_lock);
4662 /* Calculate the last device-address that could contain
4663 * any block from the chunk that includes the array-address 's'
4664 * and report the next address.
4665 * i.e. the address returned will be chunk-aligned and after
4666 * any data that is in the chunk containing 's'.
4668 static sector_t last_dev_address(sector_t s, struct geom *geo)
4670 s = (s | geo->chunk_mask) + 1;
4671 s >>= geo->chunk_shift;
4672 s *= geo->near_copies;
4673 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4674 s *= geo->far_copies;
4675 s <<= geo->chunk_shift;
4679 /* Calculate the first device-address that could contain
4680 * any block from the chunk that includes the array-address 's'.
4681 * This too will be the start of a chunk
4683 static sector_t first_dev_address(sector_t s, struct geom *geo)
4685 s >>= geo->chunk_shift;
4686 s *= geo->near_copies;
4687 sector_div(s, geo->raid_disks);
4688 s *= geo->far_copies;
4689 s <<= geo->chunk_shift;
4693 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4696 /* We simply copy at most one chunk (smallest of old and new)
4697 * at a time, possibly less if that exceeds RESYNC_PAGES,
4698 * or we hit a bad block or something.
4699 * This might mean we pause for normal IO in the middle of
4700 * a chunk, but that is not a problem as mddev->reshape_position
4701 * can record any location.
4703 * If we will want to write to a location that isn't
4704 * yet recorded as 'safe' (i.e. in metadata on disk) then
4705 * we need to flush all reshape requests and update the metadata.
4707 * When reshaping forwards (e.g. to more devices), we interpret
4708 * 'safe' as the earliest block which might not have been copied
4709 * down yet. We divide this by previous stripe size and multiply
4710 * by previous stripe length to get lowest device offset that we
4711 * cannot write to yet.
4712 * We interpret 'sector_nr' as an address that we want to write to.
4713 * From this we use last_device_address() to find where we might
4714 * write to, and first_device_address on the 'safe' position.
4715 * If this 'next' write position is after the 'safe' position,
4716 * we must update the metadata to increase the 'safe' position.
4718 * When reshaping backwards, we round in the opposite direction
4719 * and perform the reverse test: next write position must not be
4720 * less than current safe position.
4722 * In all this the minimum difference in data offsets
4723 * (conf->offset_diff - always positive) allows a bit of slack,
4724 * so next can be after 'safe', but not by more than offset_diff
4726 * We need to prepare all the bios here before we start any IO
4727 * to ensure the size we choose is acceptable to all devices.
4728 * The means one for each copy for write-out and an extra one for
4730 * We store the read-in bio in ->master_bio and the others in
4731 * ->devs[x].bio and ->devs[x].repl_bio.
4733 struct r10conf *conf = mddev->private;
4734 struct r10bio *r10_bio;
4735 sector_t next, safe, last;
4739 struct md_rdev *rdev;
4742 struct bio *bio, *read_bio;
4743 int sectors_done = 0;
4744 struct page **pages;
4746 if (sector_nr == 0) {
4747 /* If restarting in the middle, skip the initial sectors */
4748 if (mddev->reshape_backwards &&
4749 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4750 sector_nr = (raid10_size(mddev, 0, 0)
4751 - conf->reshape_progress);
4752 } else if (!mddev->reshape_backwards &&
4753 conf->reshape_progress > 0)
4754 sector_nr = conf->reshape_progress;
4756 mddev->curr_resync_completed = sector_nr;
4757 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4763 /* We don't use sector_nr to track where we are up to
4764 * as that doesn't work well for ->reshape_backwards.
4765 * So just use ->reshape_progress.
4767 if (mddev->reshape_backwards) {
4768 /* 'next' is the earliest device address that we might
4769 * write to for this chunk in the new layout
4771 next = first_dev_address(conf->reshape_progress - 1,
4774 /* 'safe' is the last device address that we might read from
4775 * in the old layout after a restart
4777 safe = last_dev_address(conf->reshape_safe - 1,
4780 if (next + conf->offset_diff < safe)
4783 last = conf->reshape_progress - 1;
4784 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4785 & conf->prev.chunk_mask);
4786 if (sector_nr + RESYNC_SECTORS < last)
4787 sector_nr = last + 1 - RESYNC_SECTORS;
4789 /* 'next' is after the last device address that we
4790 * might write to for this chunk in the new layout
4792 next = last_dev_address(conf->reshape_progress, &conf->geo);
4794 /* 'safe' is the earliest device address that we might
4795 * read from in the old layout after a restart
4797 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4799 /* Need to update metadata if 'next' might be beyond 'safe'
4800 * as that would possibly corrupt data
4802 if (next > safe + conf->offset_diff)
4805 sector_nr = conf->reshape_progress;
4806 last = sector_nr | (conf->geo.chunk_mask
4807 & conf->prev.chunk_mask);
4809 if (sector_nr + RESYNC_SECTORS <= last)
4810 last = sector_nr + RESYNC_SECTORS - 1;
4814 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4815 /* Need to update reshape_position in metadata */
4817 mddev->reshape_position = conf->reshape_progress;
4818 if (mddev->reshape_backwards)
4819 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4820 - conf->reshape_progress;
4822 mddev->curr_resync_completed = conf->reshape_progress;
4823 conf->reshape_checkpoint = jiffies;
4824 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4825 md_wakeup_thread(mddev->thread);
4826 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4827 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4828 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4829 allow_barrier(conf);
4830 return sectors_done;
4832 conf->reshape_safe = mddev->reshape_position;
4833 allow_barrier(conf);
4836 raise_barrier(conf, 0);
4838 /* Now schedule reads for blocks from sector_nr to last */
4839 r10_bio = raid10_alloc_init_r10buf(conf);
4841 raise_barrier(conf, 1);
4842 atomic_set(&r10_bio->remaining, 0);
4843 r10_bio->mddev = mddev;
4844 r10_bio->sector = sector_nr;
4845 set_bit(R10BIO_IsReshape, &r10_bio->state);
4846 r10_bio->sectors = last - sector_nr + 1;
4847 rdev = read_balance(conf, r10_bio, &max_sectors);
4848 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4851 /* Cannot read from here, so need to record bad blocks
4852 * on all the target devices.
4855 mempool_free(r10_bio, &conf->r10buf_pool);
4856 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4857 return sectors_done;
4860 read_bio = bio_alloc_bioset(GFP_KERNEL, RESYNC_PAGES, &mddev->bio_set);
4862 bio_set_dev(read_bio, rdev->bdev);
4863 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4864 + rdev->data_offset);
4865 read_bio->bi_private = r10_bio;
4866 read_bio->bi_end_io = end_reshape_read;
4867 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4868 r10_bio->master_bio = read_bio;
4869 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4872 * Broadcast RESYNC message to other nodes, so all nodes would not
4873 * write to the region to avoid conflict.
4875 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4876 struct mdp_superblock_1 *sb = NULL;
4877 int sb_reshape_pos = 0;
4879 conf->cluster_sync_low = sector_nr;
4880 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4881 sb = page_address(rdev->sb_page);
4883 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4885 * Set cluster_sync_low again if next address for array
4886 * reshape is less than cluster_sync_low. Since we can't
4887 * update cluster_sync_low until it has finished reshape.
4889 if (sb_reshape_pos < conf->cluster_sync_low)
4890 conf->cluster_sync_low = sb_reshape_pos;
4893 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4894 conf->cluster_sync_high);
4897 /* Now find the locations in the new layout */
4898 __raid10_find_phys(&conf->geo, r10_bio);
4901 read_bio->bi_next = NULL;
4904 for (s = 0; s < conf->copies*2; s++) {
4906 int d = r10_bio->devs[s/2].devnum;
4907 struct md_rdev *rdev2;
4909 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4910 b = r10_bio->devs[s/2].repl_bio;
4912 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4913 b = r10_bio->devs[s/2].bio;
4915 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4918 bio_set_dev(b, rdev2->bdev);
4919 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4920 rdev2->new_data_offset;
4921 b->bi_end_io = end_reshape_write;
4922 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4927 /* Now add as many pages as possible to all of these bios. */
4930 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4931 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4932 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4933 int len = (max_sectors - s) << 9;
4934 if (len > PAGE_SIZE)
4936 for (bio = blist; bio ; bio = bio->bi_next) {
4938 * won't fail because the vec table is big enough
4939 * to hold all these pages
4941 bio_add_page(bio, page, len, 0);
4943 sector_nr += len >> 9;
4944 nr_sectors += len >> 9;
4947 r10_bio->sectors = nr_sectors;
4949 /* Now submit the read */
4950 md_sync_acct_bio(read_bio, r10_bio->sectors);
4951 atomic_inc(&r10_bio->remaining);
4952 read_bio->bi_next = NULL;
4953 submit_bio_noacct(read_bio);
4954 sectors_done += nr_sectors;
4955 if (sector_nr <= last)
4958 lower_barrier(conf);
4960 /* Now that we have done the whole section we can
4961 * update reshape_progress
4963 if (mddev->reshape_backwards)
4964 conf->reshape_progress -= sectors_done;
4966 conf->reshape_progress += sectors_done;
4968 return sectors_done;
4971 static void end_reshape_request(struct r10bio *r10_bio);
4972 static int handle_reshape_read_error(struct mddev *mddev,
4973 struct r10bio *r10_bio);
4974 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4976 /* Reshape read completed. Hopefully we have a block
4978 * If we got a read error then we do sync 1-page reads from
4979 * elsewhere until we find the data - or give up.
4981 struct r10conf *conf = mddev->private;
4984 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4985 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4986 /* Reshape has been aborted */
4987 md_done_sync(mddev, r10_bio->sectors, 0);
4991 /* We definitely have the data in the pages, schedule the
4994 atomic_set(&r10_bio->remaining, 1);
4995 for (s = 0; s < conf->copies*2; s++) {
4997 int d = r10_bio->devs[s/2].devnum;
4998 struct md_rdev *rdev;
5001 rdev = rcu_dereference(conf->mirrors[d].replacement);
5002 b = r10_bio->devs[s/2].repl_bio;
5004 rdev = rcu_dereference(conf->mirrors[d].rdev);
5005 b = r10_bio->devs[s/2].bio;
5007 if (!rdev || test_bit(Faulty, &rdev->flags)) {
5011 atomic_inc(&rdev->nr_pending);
5013 md_sync_acct_bio(b, r10_bio->sectors);
5014 atomic_inc(&r10_bio->remaining);
5016 submit_bio_noacct(b);
5018 end_reshape_request(r10_bio);
5021 static void end_reshape(struct r10conf *conf)
5023 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
5026 spin_lock_irq(&conf->device_lock);
5027 conf->prev = conf->geo;
5028 md_finish_reshape(conf->mddev);
5030 conf->reshape_progress = MaxSector;
5031 conf->reshape_safe = MaxSector;
5032 spin_unlock_irq(&conf->device_lock);
5034 if (conf->mddev->queue)
5035 raid10_set_io_opt(conf);
5039 static void raid10_update_reshape_pos(struct mddev *mddev)
5041 struct r10conf *conf = mddev->private;
5044 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
5045 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
5046 || mddev->reshape_position == MaxSector)
5047 conf->reshape_progress = mddev->reshape_position;
5052 static int handle_reshape_read_error(struct mddev *mddev,
5053 struct r10bio *r10_bio)
5055 /* Use sync reads to get the blocks from somewhere else */
5056 int sectors = r10_bio->sectors;
5057 struct r10conf *conf = mddev->private;
5058 struct r10bio *r10b;
5061 struct page **pages;
5063 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5065 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5069 /* reshape IOs share pages from .devs[0].bio */
5070 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5072 r10b->sector = r10_bio->sector;
5073 __raid10_find_phys(&conf->prev, r10b);
5078 int first_slot = slot;
5080 if (s > (PAGE_SIZE >> 9))
5085 int d = r10b->devs[slot].devnum;
5086 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5089 test_bit(Faulty, &rdev->flags) ||
5090 !test_bit(In_sync, &rdev->flags))
5093 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5094 atomic_inc(&rdev->nr_pending);
5096 success = sync_page_io(rdev,
5100 REQ_OP_READ, 0, false);
5101 rdev_dec_pending(rdev, mddev);
5107 if (slot >= conf->copies)
5109 if (slot == first_slot)
5114 /* couldn't read this block, must give up */
5115 set_bit(MD_RECOVERY_INTR,
5127 static void end_reshape_write(struct bio *bio)
5129 struct r10bio *r10_bio = get_resync_r10bio(bio);
5130 struct mddev *mddev = r10_bio->mddev;
5131 struct r10conf *conf = mddev->private;
5135 struct md_rdev *rdev = NULL;
5137 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5139 rdev = conf->mirrors[d].replacement;
5142 rdev = conf->mirrors[d].rdev;
5145 if (bio->bi_status) {
5146 /* FIXME should record badblock */
5147 md_error(mddev, rdev);
5150 rdev_dec_pending(rdev, mddev);
5151 end_reshape_request(r10_bio);
5154 static void end_reshape_request(struct r10bio *r10_bio)
5156 if (!atomic_dec_and_test(&r10_bio->remaining))
5158 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5159 bio_put(r10_bio->master_bio);
5163 static void raid10_finish_reshape(struct mddev *mddev)
5165 struct r10conf *conf = mddev->private;
5167 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5170 if (mddev->delta_disks > 0) {
5171 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5172 mddev->recovery_cp = mddev->resync_max_sectors;
5173 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5175 mddev->resync_max_sectors = mddev->array_sectors;
5179 for (d = conf->geo.raid_disks ;
5180 d < conf->geo.raid_disks - mddev->delta_disks;
5182 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5184 clear_bit(In_sync, &rdev->flags);
5185 rdev = rcu_dereference(conf->mirrors[d].replacement);
5187 clear_bit(In_sync, &rdev->flags);
5191 mddev->layout = mddev->new_layout;
5192 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5193 mddev->reshape_position = MaxSector;
5194 mddev->delta_disks = 0;
5195 mddev->reshape_backwards = 0;
5198 static struct md_personality raid10_personality =
5202 .owner = THIS_MODULE,
5203 .make_request = raid10_make_request,
5205 .free = raid10_free,
5206 .status = raid10_status,
5207 .error_handler = raid10_error,
5208 .hot_add_disk = raid10_add_disk,
5209 .hot_remove_disk= raid10_remove_disk,
5210 .spare_active = raid10_spare_active,
5211 .sync_request = raid10_sync_request,
5212 .quiesce = raid10_quiesce,
5213 .size = raid10_size,
5214 .resize = raid10_resize,
5215 .takeover = raid10_takeover,
5216 .check_reshape = raid10_check_reshape,
5217 .start_reshape = raid10_start_reshape,
5218 .finish_reshape = raid10_finish_reshape,
5219 .update_reshape_pos = raid10_update_reshape_pos,
5222 static int __init raid_init(void)
5224 return register_md_personality(&raid10_personality);
5227 static void raid_exit(void)
5229 unregister_md_personality(&raid10_personality);
5232 module_init(raid_init);
5233 module_exit(raid_exit);
5234 MODULE_LICENSE("GPL");
5235 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5236 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5237 MODULE_ALIAS("md-raid10");
5238 MODULE_ALIAS("md-level-10");
5240 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);