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->copies]);
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->copies; 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;
302 * Wake up any possible resync thread that waits for the device
307 free_r10bio(r10_bio);
311 * Update disk head position estimator based on IRQ completion info.
313 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
315 struct r10conf *conf = r10_bio->mddev->private;
317 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
318 r10_bio->devs[slot].addr + (r10_bio->sectors);
322 * Find the disk number which triggered given bio
324 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
325 struct bio *bio, int *slotp, int *replp)
330 for (slot = 0; slot < conf->copies; slot++) {
331 if (r10_bio->devs[slot].bio == bio)
333 if (r10_bio->devs[slot].repl_bio == bio) {
339 BUG_ON(slot == conf->copies);
340 update_head_pos(slot, r10_bio);
346 return r10_bio->devs[slot].devnum;
349 static void raid10_end_read_request(struct bio *bio)
351 int uptodate = !bio->bi_status;
352 struct r10bio *r10_bio = bio->bi_private;
354 struct md_rdev *rdev;
355 struct r10conf *conf = r10_bio->mddev->private;
357 slot = r10_bio->read_slot;
358 rdev = r10_bio->devs[slot].rdev;
360 * this branch is our 'one mirror IO has finished' event handler:
362 update_head_pos(slot, r10_bio);
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
374 set_bit(R10BIO_Uptodate, &r10_bio->state);
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
381 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
386 raid_end_bio_io(r10_bio);
387 rdev_dec_pending(rdev, conf->mddev);
390 * oops, read error - keep the refcount on the rdev
392 char b[BDEVNAME_SIZE];
393 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
395 bdevname(rdev->bdev, b),
396 (unsigned long long)r10_bio->sector);
397 set_bit(R10BIO_ReadError, &r10_bio->state);
398 reschedule_retry(r10_bio);
402 static void close_write(struct r10bio *r10_bio)
404 /* clear the bitmap if all writes complete successfully */
405 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
407 !test_bit(R10BIO_Degraded, &r10_bio->state),
409 md_write_end(r10_bio->mddev);
412 static void one_write_done(struct r10bio *r10_bio)
414 if (atomic_dec_and_test(&r10_bio->remaining)) {
415 if (test_bit(R10BIO_WriteError, &r10_bio->state))
416 reschedule_retry(r10_bio);
418 close_write(r10_bio);
419 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
420 reschedule_retry(r10_bio);
422 raid_end_bio_io(r10_bio);
427 static void raid10_end_write_request(struct bio *bio)
429 struct r10bio *r10_bio = bio->bi_private;
432 struct r10conf *conf = r10_bio->mddev->private;
434 struct md_rdev *rdev = NULL;
435 struct bio *to_put = NULL;
438 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
440 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
443 rdev = conf->mirrors[dev].replacement;
447 rdev = conf->mirrors[dev].rdev;
450 * this branch is our 'one mirror IO has finished' event handler:
452 if (bio->bi_status && !discard_error) {
454 /* Never record new bad blocks to replacement,
457 md_error(rdev->mddev, rdev);
459 set_bit(WriteErrorSeen, &rdev->flags);
460 if (!test_and_set_bit(WantReplacement, &rdev->flags))
461 set_bit(MD_RECOVERY_NEEDED,
462 &rdev->mddev->recovery);
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST)) {
467 md_error(rdev->mddev, rdev);
471 * When the device is faulty, it is not necessary to
472 * handle write error.
473 * For failfast, this is the only remaining device,
474 * We need to retry the write without FailFast.
476 if (!test_bit(Faulty, &rdev->flags))
477 set_bit(R10BIO_WriteError, &r10_bio->state);
479 r10_bio->devs[slot].bio = NULL;
486 * Set R10BIO_Uptodate in our master bio, so that
487 * we will return a good error code for to the higher
488 * levels even if IO on some other mirrored buffer fails.
490 * The 'master' represents the composite IO operation to
491 * user-side. So if something waits for IO, then it will
492 * wait for the 'master' bio.
498 * Do not set R10BIO_Uptodate if the current device is
499 * rebuilding or Faulty. This is because we cannot use
500 * such device for properly reading the data back (we could
501 * potentially use it, if the current write would have felt
502 * before rdev->recovery_offset, but for simplicity we don't
505 if (test_bit(In_sync, &rdev->flags) &&
506 !test_bit(Faulty, &rdev->flags))
507 set_bit(R10BIO_Uptodate, &r10_bio->state);
509 /* Maybe we can clear some bad blocks. */
510 if (is_badblock(rdev,
511 r10_bio->devs[slot].addr,
513 &first_bad, &bad_sectors) && !discard_error) {
516 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
518 r10_bio->devs[slot].bio = IO_MADE_GOOD;
520 set_bit(R10BIO_MadeGood, &r10_bio->state);
526 * Let's see if all mirrored write operations have finished
529 one_write_done(r10_bio);
531 rdev_dec_pending(rdev, conf->mddev);
537 * RAID10 layout manager
538 * As well as the chunksize and raid_disks count, there are two
539 * parameters: near_copies and far_copies.
540 * near_copies * far_copies must be <= raid_disks.
541 * Normally one of these will be 1.
542 * If both are 1, we get raid0.
543 * If near_copies == raid_disks, we get raid1.
545 * Chunks are laid out in raid0 style with near_copies copies of the
546 * first chunk, followed by near_copies copies of the next chunk and
548 * If far_copies > 1, then after 1/far_copies of the array has been assigned
549 * as described above, we start again with a device offset of near_copies.
550 * So we effectively have another copy of the whole array further down all
551 * the drives, but with blocks on different drives.
552 * With this layout, and block is never stored twice on the one device.
554 * raid10_find_phys finds the sector offset of a given virtual sector
555 * on each device that it is on.
557 * raid10_find_virt does the reverse mapping, from a device and a
558 * sector offset to a virtual address
561 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
569 int last_far_set_start, last_far_set_size;
571 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
572 last_far_set_start *= geo->far_set_size;
574 last_far_set_size = geo->far_set_size;
575 last_far_set_size += (geo->raid_disks % geo->far_set_size);
577 /* now calculate first sector/dev */
578 chunk = r10bio->sector >> geo->chunk_shift;
579 sector = r10bio->sector & geo->chunk_mask;
581 chunk *= geo->near_copies;
583 dev = sector_div(stripe, geo->raid_disks);
585 stripe *= geo->far_copies;
587 sector += stripe << geo->chunk_shift;
589 /* and calculate all the others */
590 for (n = 0; n < geo->near_copies; n++) {
594 r10bio->devs[slot].devnum = d;
595 r10bio->devs[slot].addr = s;
598 for (f = 1; f < geo->far_copies; f++) {
599 set = d / geo->far_set_size;
600 d += geo->near_copies;
602 if ((geo->raid_disks % geo->far_set_size) &&
603 (d > last_far_set_start)) {
604 d -= last_far_set_start;
605 d %= last_far_set_size;
606 d += last_far_set_start;
608 d %= geo->far_set_size;
609 d += geo->far_set_size * set;
612 r10bio->devs[slot].devnum = d;
613 r10bio->devs[slot].addr = s;
617 if (dev >= geo->raid_disks) {
619 sector += (geo->chunk_mask + 1);
624 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
626 struct geom *geo = &conf->geo;
628 if (conf->reshape_progress != MaxSector &&
629 ((r10bio->sector >= conf->reshape_progress) !=
630 conf->mddev->reshape_backwards)) {
631 set_bit(R10BIO_Previous, &r10bio->state);
634 clear_bit(R10BIO_Previous, &r10bio->state);
636 __raid10_find_phys(geo, r10bio);
639 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
641 sector_t offset, chunk, vchunk;
642 /* Never use conf->prev as this is only called during resync
643 * or recovery, so reshape isn't happening
645 struct geom *geo = &conf->geo;
646 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
647 int far_set_size = geo->far_set_size;
648 int last_far_set_start;
650 if (geo->raid_disks % geo->far_set_size) {
651 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
652 last_far_set_start *= geo->far_set_size;
654 if (dev >= last_far_set_start) {
655 far_set_size = geo->far_set_size;
656 far_set_size += (geo->raid_disks % geo->far_set_size);
657 far_set_start = last_far_set_start;
661 offset = sector & geo->chunk_mask;
662 if (geo->far_offset) {
664 chunk = sector >> geo->chunk_shift;
665 fc = sector_div(chunk, geo->far_copies);
666 dev -= fc * geo->near_copies;
667 if (dev < far_set_start)
670 while (sector >= geo->stride) {
671 sector -= geo->stride;
672 if (dev < (geo->near_copies + far_set_start))
673 dev += far_set_size - geo->near_copies;
675 dev -= geo->near_copies;
677 chunk = sector >> geo->chunk_shift;
679 vchunk = chunk * geo->raid_disks + dev;
680 sector_div(vchunk, geo->near_copies);
681 return (vchunk << geo->chunk_shift) + offset;
685 * This routine returns the disk from which the requested read should
686 * be done. There is a per-array 'next expected sequential IO' sector
687 * number - if this matches on the next IO then we use the last disk.
688 * There is also a per-disk 'last know head position' sector that is
689 * maintained from IRQ contexts, both the normal and the resync IO
690 * completion handlers update this position correctly. If there is no
691 * perfect sequential match then we pick the disk whose head is closest.
693 * If there are 2 mirrors in the same 2 devices, performance degrades
694 * because position is mirror, not device based.
696 * The rdev for the device selected will have nr_pending incremented.
700 * FIXME: possibly should rethink readbalancing and do it differently
701 * depending on near_copies / far_copies geometry.
703 static struct md_rdev *read_balance(struct r10conf *conf,
704 struct r10bio *r10_bio,
707 const sector_t this_sector = r10_bio->sector;
709 int sectors = r10_bio->sectors;
710 int best_good_sectors;
711 sector_t new_distance, best_dist;
712 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
714 int best_dist_slot, best_pending_slot;
715 bool has_nonrot_disk = false;
716 unsigned int min_pending;
717 struct geom *geo = &conf->geo;
719 raid10_find_phys(conf, r10_bio);
722 min_pending = UINT_MAX;
723 best_dist_rdev = NULL;
724 best_pending_rdev = NULL;
725 best_dist = MaxSector;
726 best_good_sectors = 0;
728 clear_bit(R10BIO_FailFast, &r10_bio->state);
730 * Check if we can balance. We can balance on the whole
731 * device if no resync is going on (recovery is ok), or below
732 * the resync window. We take the first readable disk when
733 * above the resync window.
735 if ((conf->mddev->recovery_cp < MaxSector
736 && (this_sector + sectors >= conf->next_resync)) ||
737 (mddev_is_clustered(conf->mddev) &&
738 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
739 this_sector + sectors)))
742 for (slot = 0; slot < conf->copies ; slot++) {
746 unsigned int pending;
749 if (r10_bio->devs[slot].bio == IO_BLOCKED)
751 disk = r10_bio->devs[slot].devnum;
752 rdev = rcu_dereference(conf->mirrors[disk].replacement);
753 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
754 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
755 rdev = rcu_dereference(conf->mirrors[disk].rdev);
757 test_bit(Faulty, &rdev->flags))
759 if (!test_bit(In_sync, &rdev->flags) &&
760 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
763 dev_sector = r10_bio->devs[slot].addr;
764 if (is_badblock(rdev, dev_sector, sectors,
765 &first_bad, &bad_sectors)) {
766 if (best_dist < MaxSector)
767 /* Already have a better slot */
769 if (first_bad <= dev_sector) {
770 /* Cannot read here. If this is the
771 * 'primary' device, then we must not read
772 * beyond 'bad_sectors' from another device.
774 bad_sectors -= (dev_sector - first_bad);
775 if (!do_balance && sectors > bad_sectors)
776 sectors = bad_sectors;
777 if (best_good_sectors > sectors)
778 best_good_sectors = sectors;
780 sector_t good_sectors =
781 first_bad - dev_sector;
782 if (good_sectors > best_good_sectors) {
783 best_good_sectors = good_sectors;
784 best_dist_slot = slot;
785 best_dist_rdev = rdev;
788 /* Must read from here */
793 best_good_sectors = sectors;
798 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
799 has_nonrot_disk |= nonrot;
800 pending = atomic_read(&rdev->nr_pending);
801 if (min_pending > pending && nonrot) {
802 min_pending = pending;
803 best_pending_slot = slot;
804 best_pending_rdev = rdev;
807 if (best_dist_slot >= 0)
808 /* At least 2 disks to choose from so failfast is OK */
809 set_bit(R10BIO_FailFast, &r10_bio->state);
810 /* This optimisation is debatable, and completely destroys
811 * sequential read speed for 'far copies' arrays. So only
812 * keep it for 'near' arrays, and review those later.
814 if (geo->near_copies > 1 && !pending)
817 /* for far > 1 always use the lowest address */
818 else if (geo->far_copies > 1)
819 new_distance = r10_bio->devs[slot].addr;
821 new_distance = abs(r10_bio->devs[slot].addr -
822 conf->mirrors[disk].head_position);
824 if (new_distance < best_dist) {
825 best_dist = new_distance;
826 best_dist_slot = slot;
827 best_dist_rdev = rdev;
830 if (slot >= conf->copies) {
831 if (has_nonrot_disk) {
832 slot = best_pending_slot;
833 rdev = best_pending_rdev;
835 slot = best_dist_slot;
836 rdev = best_dist_rdev;
841 atomic_inc(&rdev->nr_pending);
842 r10_bio->read_slot = slot;
846 *max_sectors = best_good_sectors;
851 static int raid10_congested(struct mddev *mddev, int bits)
853 struct r10conf *conf = mddev->private;
856 if ((bits & (1 << WB_async_congested)) &&
857 conf->pending_count >= max_queued_requests)
862 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
865 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
866 if (rdev && !test_bit(Faulty, &rdev->flags)) {
867 struct request_queue *q = bdev_get_queue(rdev->bdev);
869 ret |= bdi_congested(q->backing_dev_info, bits);
876 static void flush_pending_writes(struct r10conf *conf)
878 /* Any writes that have been queued but are awaiting
879 * bitmap updates get flushed here.
881 spin_lock_irq(&conf->device_lock);
883 if (conf->pending_bio_list.head) {
884 struct blk_plug plug;
887 bio = bio_list_get(&conf->pending_bio_list);
888 conf->pending_count = 0;
889 spin_unlock_irq(&conf->device_lock);
892 * As this is called in a wait_event() loop (see freeze_array),
893 * current->state might be TASK_UNINTERRUPTIBLE which will
894 * cause a warning when we prepare to wait again. As it is
895 * rare that this path is taken, it is perfectly safe to force
896 * us to go around the wait_event() loop again, so the warning
897 * is a false-positive. Silence the warning by resetting
900 __set_current_state(TASK_RUNNING);
902 blk_start_plug(&plug);
903 /* flush any pending bitmap writes to disk
904 * before proceeding w/ I/O */
905 md_bitmap_unplug(conf->mddev->bitmap);
906 wake_up(&conf->wait_barrier);
908 while (bio) { /* submit pending writes */
909 struct bio *next = bio->bi_next;
910 struct md_rdev *rdev = (void*)bio->bi_disk;
912 bio_set_dev(bio, rdev->bdev);
913 if (test_bit(Faulty, &rdev->flags)) {
915 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
916 !blk_queue_discard(bio->bi_disk->queue)))
920 generic_make_request(bio);
923 blk_finish_plug(&plug);
925 spin_unlock_irq(&conf->device_lock);
929 * Sometimes we need to suspend IO while we do something else,
930 * either some resync/recovery, or reconfigure the array.
931 * To do this we raise a 'barrier'.
932 * The 'barrier' is a counter that can be raised multiple times
933 * to count how many activities are happening which preclude
935 * We can only raise the barrier if there is no pending IO.
936 * i.e. if nr_pending == 0.
937 * We choose only to raise the barrier if no-one is waiting for the
938 * barrier to go down. This means that as soon as an IO request
939 * is ready, no other operations which require a barrier will start
940 * until the IO request has had a chance.
942 * So: regular IO calls 'wait_barrier'. When that returns there
943 * is no backgroup IO happening, It must arrange to call
944 * allow_barrier when it has finished its IO.
945 * backgroup IO calls must call raise_barrier. Once that returns
946 * there is no normal IO happeing. It must arrange to call
947 * lower_barrier when the particular background IO completes.
950 static void raise_barrier(struct r10conf *conf, int force)
952 BUG_ON(force && !conf->barrier);
953 spin_lock_irq(&conf->resync_lock);
955 /* Wait until no block IO is waiting (unless 'force') */
956 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
959 /* block any new IO from starting */
962 /* Now wait for all pending IO to complete */
963 wait_event_lock_irq(conf->wait_barrier,
964 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
967 spin_unlock_irq(&conf->resync_lock);
970 static void lower_barrier(struct r10conf *conf)
973 spin_lock_irqsave(&conf->resync_lock, flags);
975 spin_unlock_irqrestore(&conf->resync_lock, flags);
976 wake_up(&conf->wait_barrier);
979 static void wait_barrier(struct r10conf *conf)
981 spin_lock_irq(&conf->resync_lock);
984 /* Wait for the barrier to drop.
985 * However if there are already pending
986 * requests (preventing the barrier from
987 * rising completely), and the
988 * pre-process bio queue isn't empty,
989 * then don't wait, as we need to empty
990 * that queue to get the nr_pending
993 raid10_log(conf->mddev, "wait barrier");
994 wait_event_lock_irq(conf->wait_barrier,
996 (atomic_read(&conf->nr_pending) &&
998 (!bio_list_empty(¤t->bio_list[0]) ||
999 !bio_list_empty(¤t->bio_list[1]))),
1002 if (!conf->nr_waiting)
1003 wake_up(&conf->wait_barrier);
1005 atomic_inc(&conf->nr_pending);
1006 spin_unlock_irq(&conf->resync_lock);
1009 static void allow_barrier(struct r10conf *conf)
1011 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1012 (conf->array_freeze_pending))
1013 wake_up(&conf->wait_barrier);
1016 static void freeze_array(struct r10conf *conf, int extra)
1018 /* stop syncio and normal IO and wait for everything to
1020 * We increment barrier and nr_waiting, and then
1021 * wait until nr_pending match nr_queued+extra
1022 * This is called in the context of one normal IO request
1023 * that has failed. Thus any sync request that might be pending
1024 * will be blocked by nr_pending, and we need to wait for
1025 * pending IO requests to complete or be queued for re-try.
1026 * Thus the number queued (nr_queued) plus this request (extra)
1027 * must match the number of pending IOs (nr_pending) before
1030 spin_lock_irq(&conf->resync_lock);
1031 conf->array_freeze_pending++;
1034 wait_event_lock_irq_cmd(conf->wait_barrier,
1035 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1037 flush_pending_writes(conf));
1039 conf->array_freeze_pending--;
1040 spin_unlock_irq(&conf->resync_lock);
1043 static void unfreeze_array(struct r10conf *conf)
1045 /* reverse the effect of the freeze */
1046 spin_lock_irq(&conf->resync_lock);
1049 wake_up(&conf->wait_barrier);
1050 spin_unlock_irq(&conf->resync_lock);
1053 static sector_t choose_data_offset(struct r10bio *r10_bio,
1054 struct md_rdev *rdev)
1056 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1057 test_bit(R10BIO_Previous, &r10_bio->state))
1058 return rdev->data_offset;
1060 return rdev->new_data_offset;
1063 struct raid10_plug_cb {
1064 struct blk_plug_cb cb;
1065 struct bio_list pending;
1069 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1071 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1073 struct mddev *mddev = plug->cb.data;
1074 struct r10conf *conf = mddev->private;
1077 if (from_schedule || current->bio_list) {
1078 spin_lock_irq(&conf->device_lock);
1079 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1080 conf->pending_count += plug->pending_cnt;
1081 spin_unlock_irq(&conf->device_lock);
1082 wake_up(&conf->wait_barrier);
1083 md_wakeup_thread(mddev->thread);
1088 /* we aren't scheduling, so we can do the write-out directly. */
1089 bio = bio_list_get(&plug->pending);
1090 md_bitmap_unplug(mddev->bitmap);
1091 wake_up(&conf->wait_barrier);
1093 while (bio) { /* submit pending writes */
1094 struct bio *next = bio->bi_next;
1095 struct md_rdev *rdev = (void*)bio->bi_disk;
1096 bio->bi_next = NULL;
1097 bio_set_dev(bio, rdev->bdev);
1098 if (test_bit(Faulty, &rdev->flags)) {
1100 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1101 !blk_queue_discard(bio->bi_disk->queue)))
1102 /* Just ignore it */
1105 generic_make_request(bio);
1112 * 1. Register the new request and wait if the reconstruction thread has put
1113 * up a bar for new requests. Continue immediately if no resync is active
1115 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1117 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1118 struct bio *bio, sector_t sectors)
1121 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1122 bio->bi_iter.bi_sector < conf->reshape_progress &&
1123 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1124 raid10_log(conf->mddev, "wait reshape");
1125 allow_barrier(conf);
1126 wait_event(conf->wait_barrier,
1127 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1128 conf->reshape_progress >= bio->bi_iter.bi_sector +
1134 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1135 struct r10bio *r10_bio)
1137 struct r10conf *conf = mddev->private;
1138 struct bio *read_bio;
1139 const int op = bio_op(bio);
1140 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1142 struct md_rdev *rdev;
1143 char b[BDEVNAME_SIZE];
1144 int slot = r10_bio->read_slot;
1145 struct md_rdev *err_rdev = NULL;
1146 gfp_t gfp = GFP_NOIO;
1148 if (r10_bio->devs[slot].rdev) {
1150 * This is an error retry, but we cannot
1151 * safely dereference the rdev in the r10_bio,
1152 * we must use the one in conf.
1153 * If it has already been disconnected (unlikely)
1154 * we lose the device name in error messages.
1158 * As we are blocking raid10, it is a little safer to
1161 gfp = GFP_NOIO | __GFP_HIGH;
1164 disk = r10_bio->devs[slot].devnum;
1165 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1167 bdevname(err_rdev->bdev, b);
1170 /* This never gets dereferenced */
1171 err_rdev = r10_bio->devs[slot].rdev;
1176 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1177 rdev = read_balance(conf, r10_bio, &max_sectors);
1180 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1182 (unsigned long long)r10_bio->sector);
1184 raid_end_bio_io(r10_bio);
1188 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1190 bdevname(rdev->bdev, b),
1191 (unsigned long long)r10_bio->sector);
1192 if (max_sectors < bio_sectors(bio)) {
1193 struct bio *split = bio_split(bio, max_sectors,
1194 gfp, &conf->bio_split);
1195 bio_chain(split, bio);
1196 allow_barrier(conf);
1197 generic_make_request(bio);
1200 r10_bio->master_bio = bio;
1201 r10_bio->sectors = max_sectors;
1203 slot = r10_bio->read_slot;
1205 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1207 r10_bio->devs[slot].bio = read_bio;
1208 r10_bio->devs[slot].rdev = rdev;
1210 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1211 choose_data_offset(r10_bio, rdev);
1212 bio_set_dev(read_bio, rdev->bdev);
1213 read_bio->bi_end_io = raid10_end_read_request;
1214 bio_set_op_attrs(read_bio, op, do_sync);
1215 if (test_bit(FailFast, &rdev->flags) &&
1216 test_bit(R10BIO_FailFast, &r10_bio->state))
1217 read_bio->bi_opf |= MD_FAILFAST;
1218 read_bio->bi_private = r10_bio;
1221 trace_block_bio_remap(read_bio->bi_disk->queue,
1222 read_bio, disk_devt(mddev->gendisk),
1224 generic_make_request(read_bio);
1228 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1229 struct bio *bio, bool replacement,
1232 const int op = bio_op(bio);
1233 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1234 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1235 unsigned long flags;
1236 struct blk_plug_cb *cb;
1237 struct raid10_plug_cb *plug = NULL;
1238 struct r10conf *conf = mddev->private;
1239 struct md_rdev *rdev;
1240 int devnum = r10_bio->devs[n_copy].devnum;
1244 rdev = conf->mirrors[devnum].replacement;
1246 /* Replacement just got moved to main 'rdev' */
1248 rdev = conf->mirrors[devnum].rdev;
1251 rdev = conf->mirrors[devnum].rdev;
1253 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1255 r10_bio->devs[n_copy].repl_bio = mbio;
1257 r10_bio->devs[n_copy].bio = mbio;
1259 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1260 choose_data_offset(r10_bio, rdev));
1261 bio_set_dev(mbio, rdev->bdev);
1262 mbio->bi_end_io = raid10_end_write_request;
1263 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1264 if (!replacement && test_bit(FailFast,
1265 &conf->mirrors[devnum].rdev->flags)
1266 && enough(conf, devnum))
1267 mbio->bi_opf |= MD_FAILFAST;
1268 mbio->bi_private = r10_bio;
1270 if (conf->mddev->gendisk)
1271 trace_block_bio_remap(mbio->bi_disk->queue,
1272 mbio, disk_devt(conf->mddev->gendisk),
1274 /* flush_pending_writes() needs access to the rdev so...*/
1275 mbio->bi_disk = (void *)rdev;
1277 atomic_inc(&r10_bio->remaining);
1279 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1281 plug = container_of(cb, struct raid10_plug_cb, cb);
1285 bio_list_add(&plug->pending, mbio);
1286 plug->pending_cnt++;
1288 spin_lock_irqsave(&conf->device_lock, flags);
1289 bio_list_add(&conf->pending_bio_list, mbio);
1290 conf->pending_count++;
1291 spin_unlock_irqrestore(&conf->device_lock, flags);
1292 md_wakeup_thread(mddev->thread);
1296 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1297 struct r10bio *r10_bio)
1299 struct r10conf *conf = mddev->private;
1301 struct md_rdev *blocked_rdev;
1305 if ((mddev_is_clustered(mddev) &&
1306 md_cluster_ops->area_resyncing(mddev, WRITE,
1307 bio->bi_iter.bi_sector,
1308 bio_end_sector(bio)))) {
1311 prepare_to_wait(&conf->wait_barrier,
1313 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1314 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1318 finish_wait(&conf->wait_barrier, &w);
1321 sectors = r10_bio->sectors;
1322 regular_request_wait(mddev, conf, bio, sectors);
1323 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1324 (mddev->reshape_backwards
1325 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1326 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1327 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1328 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1329 /* Need to update reshape_position in metadata */
1330 mddev->reshape_position = conf->reshape_progress;
1331 set_mask_bits(&mddev->sb_flags, 0,
1332 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1333 md_wakeup_thread(mddev->thread);
1334 raid10_log(conf->mddev, "wait reshape metadata");
1335 wait_event(mddev->sb_wait,
1336 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1338 conf->reshape_safe = mddev->reshape_position;
1341 if (conf->pending_count >= max_queued_requests) {
1342 md_wakeup_thread(mddev->thread);
1343 raid10_log(mddev, "wait queued");
1344 wait_event(conf->wait_barrier,
1345 conf->pending_count < max_queued_requests);
1347 /* first select target devices under rcu_lock and
1348 * inc refcount on their rdev. Record them by setting
1350 * If there are known/acknowledged bad blocks on any device
1351 * on which we have seen a write error, we want to avoid
1352 * writing to those blocks. This potentially requires several
1353 * writes to write around the bad blocks. Each set of writes
1354 * gets its own r10_bio with a set of bios attached.
1357 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1358 raid10_find_phys(conf, r10_bio);
1360 blocked_rdev = NULL;
1362 max_sectors = r10_bio->sectors;
1364 for (i = 0; i < conf->copies; i++) {
1365 int d = r10_bio->devs[i].devnum;
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1367 struct md_rdev *rrdev = rcu_dereference(
1368 conf->mirrors[d].replacement);
1371 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1372 atomic_inc(&rdev->nr_pending);
1373 blocked_rdev = rdev;
1376 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1377 atomic_inc(&rrdev->nr_pending);
1378 blocked_rdev = rrdev;
1381 if (rdev && (test_bit(Faulty, &rdev->flags)))
1383 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1386 r10_bio->devs[i].bio = NULL;
1387 r10_bio->devs[i].repl_bio = NULL;
1389 if (!rdev && !rrdev) {
1390 set_bit(R10BIO_Degraded, &r10_bio->state);
1393 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1395 sector_t dev_sector = r10_bio->devs[i].addr;
1399 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1400 &first_bad, &bad_sectors);
1402 /* Mustn't write here until the bad block
1405 atomic_inc(&rdev->nr_pending);
1406 set_bit(BlockedBadBlocks, &rdev->flags);
1407 blocked_rdev = rdev;
1410 if (is_bad && first_bad <= dev_sector) {
1411 /* Cannot write here at all */
1412 bad_sectors -= (dev_sector - first_bad);
1413 if (bad_sectors < max_sectors)
1414 /* Mustn't write more than bad_sectors
1415 * to other devices yet
1417 max_sectors = bad_sectors;
1418 /* We don't set R10BIO_Degraded as that
1419 * only applies if the disk is missing,
1420 * so it might be re-added, and we want to
1421 * know to recover this chunk.
1422 * In this case the device is here, and the
1423 * fact that this chunk is not in-sync is
1424 * recorded in the bad block log.
1429 int good_sectors = first_bad - dev_sector;
1430 if (good_sectors < max_sectors)
1431 max_sectors = good_sectors;
1435 r10_bio->devs[i].bio = bio;
1436 atomic_inc(&rdev->nr_pending);
1439 r10_bio->devs[i].repl_bio = bio;
1440 atomic_inc(&rrdev->nr_pending);
1445 if (unlikely(blocked_rdev)) {
1446 /* Have to wait for this device to get unblocked, then retry */
1450 for (j = 0; j < i; j++) {
1451 if (r10_bio->devs[j].bio) {
1452 d = r10_bio->devs[j].devnum;
1453 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1455 if (r10_bio->devs[j].repl_bio) {
1456 struct md_rdev *rdev;
1457 d = r10_bio->devs[j].devnum;
1458 rdev = conf->mirrors[d].replacement;
1460 /* Race with remove_disk */
1462 rdev = conf->mirrors[d].rdev;
1464 rdev_dec_pending(rdev, mddev);
1467 allow_barrier(conf);
1468 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1469 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1474 if (max_sectors < r10_bio->sectors)
1475 r10_bio->sectors = max_sectors;
1477 if (r10_bio->sectors < bio_sectors(bio)) {
1478 struct bio *split = bio_split(bio, r10_bio->sectors,
1479 GFP_NOIO, &conf->bio_split);
1480 bio_chain(split, bio);
1481 allow_barrier(conf);
1482 generic_make_request(bio);
1485 r10_bio->master_bio = bio;
1488 atomic_set(&r10_bio->remaining, 1);
1489 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1491 for (i = 0; i < conf->copies; i++) {
1492 if (r10_bio->devs[i].bio)
1493 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1494 if (r10_bio->devs[i].repl_bio)
1495 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1497 one_write_done(r10_bio);
1500 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1502 struct r10conf *conf = mddev->private;
1503 struct r10bio *r10_bio;
1505 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1507 r10_bio->master_bio = bio;
1508 r10_bio->sectors = sectors;
1510 r10_bio->mddev = mddev;
1511 r10_bio->sector = bio->bi_iter.bi_sector;
1513 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1515 if (bio_data_dir(bio) == READ)
1516 raid10_read_request(mddev, bio, r10_bio);
1518 raid10_write_request(mddev, bio, r10_bio);
1521 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1523 struct r10conf *conf = mddev->private;
1524 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1525 int chunk_sects = chunk_mask + 1;
1526 int sectors = bio_sectors(bio);
1528 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1529 && md_flush_request(mddev, bio))
1532 if (!md_write_start(mddev, bio))
1536 * If this request crosses a chunk boundary, we need to split
1539 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1540 sectors > chunk_sects
1541 && (conf->geo.near_copies < conf->geo.raid_disks
1542 || conf->prev.near_copies <
1543 conf->prev.raid_disks)))
1544 sectors = chunk_sects -
1545 (bio->bi_iter.bi_sector &
1547 __make_request(mddev, bio, sectors);
1549 /* In case raid10d snuck in to freeze_array */
1550 wake_up(&conf->wait_barrier);
1554 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1556 struct r10conf *conf = mddev->private;
1559 if (conf->geo.near_copies < conf->geo.raid_disks)
1560 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1561 if (conf->geo.near_copies > 1)
1562 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1563 if (conf->geo.far_copies > 1) {
1564 if (conf->geo.far_offset)
1565 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1567 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1568 if (conf->geo.far_set_size != conf->geo.raid_disks)
1569 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1571 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1572 conf->geo.raid_disks - mddev->degraded);
1574 for (i = 0; i < conf->geo.raid_disks; i++) {
1575 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1576 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1579 seq_printf(seq, "]");
1582 /* check if there are enough drives for
1583 * every block to appear on atleast one.
1584 * Don't consider the device numbered 'ignore'
1585 * as we might be about to remove it.
1587 static int _enough(struct r10conf *conf, int previous, int ignore)
1593 disks = conf->prev.raid_disks;
1594 ncopies = conf->prev.near_copies;
1596 disks = conf->geo.raid_disks;
1597 ncopies = conf->geo.near_copies;
1602 int n = conf->copies;
1606 struct md_rdev *rdev;
1607 if (this != ignore &&
1608 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1609 test_bit(In_sync, &rdev->flags))
1611 this = (this+1) % disks;
1615 first = (first + ncopies) % disks;
1616 } while (first != 0);
1623 static int enough(struct r10conf *conf, int ignore)
1625 /* when calling 'enough', both 'prev' and 'geo' must
1627 * This is ensured if ->reconfig_mutex or ->device_lock
1630 return _enough(conf, 0, ignore) &&
1631 _enough(conf, 1, ignore);
1634 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1636 char b[BDEVNAME_SIZE];
1637 struct r10conf *conf = mddev->private;
1638 unsigned long flags;
1641 * If it is not operational, then we have already marked it as dead
1642 * else if it is the last working disks with "fail_last_dev == false",
1643 * ignore the error, let the next level up know.
1644 * else mark the drive as failed
1646 spin_lock_irqsave(&conf->device_lock, flags);
1647 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1648 && !enough(conf, rdev->raid_disk)) {
1650 * Don't fail the drive, just return an IO error.
1652 spin_unlock_irqrestore(&conf->device_lock, flags);
1655 if (test_and_clear_bit(In_sync, &rdev->flags))
1658 * If recovery is running, make sure it aborts.
1660 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1661 set_bit(Blocked, &rdev->flags);
1662 set_bit(Faulty, &rdev->flags);
1663 set_mask_bits(&mddev->sb_flags, 0,
1664 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1665 spin_unlock_irqrestore(&conf->device_lock, flags);
1666 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1667 "md/raid10:%s: Operation continuing on %d devices.\n",
1668 mdname(mddev), bdevname(rdev->bdev, b),
1669 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1672 static void print_conf(struct r10conf *conf)
1675 struct md_rdev *rdev;
1677 pr_debug("RAID10 conf printout:\n");
1679 pr_debug("(!conf)\n");
1682 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1683 conf->geo.raid_disks);
1685 /* This is only called with ->reconfix_mutex held, so
1686 * rcu protection of rdev is not needed */
1687 for (i = 0; i < conf->geo.raid_disks; i++) {
1688 char b[BDEVNAME_SIZE];
1689 rdev = conf->mirrors[i].rdev;
1691 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1692 i, !test_bit(In_sync, &rdev->flags),
1693 !test_bit(Faulty, &rdev->flags),
1694 bdevname(rdev->bdev,b));
1698 static void close_sync(struct r10conf *conf)
1701 allow_barrier(conf);
1703 mempool_exit(&conf->r10buf_pool);
1706 static int raid10_spare_active(struct mddev *mddev)
1709 struct r10conf *conf = mddev->private;
1710 struct raid10_info *tmp;
1712 unsigned long flags;
1715 * Find all non-in_sync disks within the RAID10 configuration
1716 * and mark them in_sync
1718 for (i = 0; i < conf->geo.raid_disks; i++) {
1719 tmp = conf->mirrors + i;
1720 if (tmp->replacement
1721 && tmp->replacement->recovery_offset == MaxSector
1722 && !test_bit(Faulty, &tmp->replacement->flags)
1723 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1724 /* Replacement has just become active */
1726 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1729 /* Replaced device not technically faulty,
1730 * but we need to be sure it gets removed
1731 * and never re-added.
1733 set_bit(Faulty, &tmp->rdev->flags);
1734 sysfs_notify_dirent_safe(
1735 tmp->rdev->sysfs_state);
1737 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1738 } else if (tmp->rdev
1739 && tmp->rdev->recovery_offset == MaxSector
1740 && !test_bit(Faulty, &tmp->rdev->flags)
1741 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1743 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1746 spin_lock_irqsave(&conf->device_lock, flags);
1747 mddev->degraded -= count;
1748 spin_unlock_irqrestore(&conf->device_lock, flags);
1754 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1756 struct r10conf *conf = mddev->private;
1760 int last = conf->geo.raid_disks - 1;
1762 if (mddev->recovery_cp < MaxSector)
1763 /* only hot-add to in-sync arrays, as recovery is
1764 * very different from resync
1767 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1770 if (md_integrity_add_rdev(rdev, mddev))
1773 if (rdev->raid_disk >= 0)
1774 first = last = rdev->raid_disk;
1776 if (rdev->saved_raid_disk >= first &&
1777 rdev->saved_raid_disk < conf->geo.raid_disks &&
1778 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1779 mirror = rdev->saved_raid_disk;
1782 for ( ; mirror <= last ; mirror++) {
1783 struct raid10_info *p = &conf->mirrors[mirror];
1784 if (p->recovery_disabled == mddev->recovery_disabled)
1787 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1788 p->replacement != NULL)
1790 clear_bit(In_sync, &rdev->flags);
1791 set_bit(Replacement, &rdev->flags);
1792 rdev->raid_disk = mirror;
1795 disk_stack_limits(mddev->gendisk, rdev->bdev,
1796 rdev->data_offset << 9);
1798 rcu_assign_pointer(p->replacement, rdev);
1803 disk_stack_limits(mddev->gendisk, rdev->bdev,
1804 rdev->data_offset << 9);
1806 p->head_position = 0;
1807 p->recovery_disabled = mddev->recovery_disabled - 1;
1808 rdev->raid_disk = mirror;
1810 if (rdev->saved_raid_disk != mirror)
1812 rcu_assign_pointer(p->rdev, rdev);
1815 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1816 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1822 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1824 struct r10conf *conf = mddev->private;
1826 int number = rdev->raid_disk;
1827 struct md_rdev **rdevp;
1828 struct raid10_info *p = conf->mirrors + number;
1831 if (rdev == p->rdev)
1833 else if (rdev == p->replacement)
1834 rdevp = &p->replacement;
1838 if (test_bit(In_sync, &rdev->flags) ||
1839 atomic_read(&rdev->nr_pending)) {
1843 /* Only remove non-faulty devices if recovery
1846 if (!test_bit(Faulty, &rdev->flags) &&
1847 mddev->recovery_disabled != p->recovery_disabled &&
1848 (!p->replacement || p->replacement == rdev) &&
1849 number < conf->geo.raid_disks &&
1855 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1857 if (atomic_read(&rdev->nr_pending)) {
1858 /* lost the race, try later */
1864 if (p->replacement) {
1865 /* We must have just cleared 'rdev' */
1866 p->rdev = p->replacement;
1867 clear_bit(Replacement, &p->replacement->flags);
1868 smp_mb(); /* Make sure other CPUs may see both as identical
1869 * but will never see neither -- if they are careful.
1871 p->replacement = NULL;
1874 clear_bit(WantReplacement, &rdev->flags);
1875 err = md_integrity_register(mddev);
1883 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1885 struct r10conf *conf = r10_bio->mddev->private;
1887 if (!bio->bi_status)
1888 set_bit(R10BIO_Uptodate, &r10_bio->state);
1890 /* The write handler will notice the lack of
1891 * R10BIO_Uptodate and record any errors etc
1893 atomic_add(r10_bio->sectors,
1894 &conf->mirrors[d].rdev->corrected_errors);
1896 /* for reconstruct, we always reschedule after a read.
1897 * for resync, only after all reads
1899 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1900 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1901 atomic_dec_and_test(&r10_bio->remaining)) {
1902 /* we have read all the blocks,
1903 * do the comparison in process context in raid10d
1905 reschedule_retry(r10_bio);
1909 static void end_sync_read(struct bio *bio)
1911 struct r10bio *r10_bio = get_resync_r10bio(bio);
1912 struct r10conf *conf = r10_bio->mddev->private;
1913 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1915 __end_sync_read(r10_bio, bio, d);
1918 static void end_reshape_read(struct bio *bio)
1920 /* reshape read bio isn't allocated from r10buf_pool */
1921 struct r10bio *r10_bio = bio->bi_private;
1923 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1926 static void end_sync_request(struct r10bio *r10_bio)
1928 struct mddev *mddev = r10_bio->mddev;
1930 while (atomic_dec_and_test(&r10_bio->remaining)) {
1931 if (r10_bio->master_bio == NULL) {
1932 /* the primary of several recovery bios */
1933 sector_t s = r10_bio->sectors;
1934 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1935 test_bit(R10BIO_WriteError, &r10_bio->state))
1936 reschedule_retry(r10_bio);
1939 md_done_sync(mddev, s, 1);
1942 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1943 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1944 test_bit(R10BIO_WriteError, &r10_bio->state))
1945 reschedule_retry(r10_bio);
1953 static void end_sync_write(struct bio *bio)
1955 struct r10bio *r10_bio = get_resync_r10bio(bio);
1956 struct mddev *mddev = r10_bio->mddev;
1957 struct r10conf *conf = mddev->private;
1963 struct md_rdev *rdev = NULL;
1965 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1967 rdev = conf->mirrors[d].replacement;
1969 rdev = conf->mirrors[d].rdev;
1971 if (bio->bi_status) {
1973 md_error(mddev, rdev);
1975 set_bit(WriteErrorSeen, &rdev->flags);
1976 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1977 set_bit(MD_RECOVERY_NEEDED,
1978 &rdev->mddev->recovery);
1979 set_bit(R10BIO_WriteError, &r10_bio->state);
1981 } else if (is_badblock(rdev,
1982 r10_bio->devs[slot].addr,
1984 &first_bad, &bad_sectors))
1985 set_bit(R10BIO_MadeGood, &r10_bio->state);
1987 rdev_dec_pending(rdev, mddev);
1989 end_sync_request(r10_bio);
1993 * Note: sync and recover and handled very differently for raid10
1994 * This code is for resync.
1995 * For resync, we read through virtual addresses and read all blocks.
1996 * If there is any error, we schedule a write. The lowest numbered
1997 * drive is authoritative.
1998 * However requests come for physical address, so we need to map.
1999 * For every physical address there are raid_disks/copies virtual addresses,
2000 * which is always are least one, but is not necessarly an integer.
2001 * This means that a physical address can span multiple chunks, so we may
2002 * have to submit multiple io requests for a single sync request.
2005 * We check if all blocks are in-sync and only write to blocks that
2008 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2010 struct r10conf *conf = mddev->private;
2012 struct bio *tbio, *fbio;
2014 struct page **tpages, **fpages;
2016 atomic_set(&r10_bio->remaining, 1);
2018 /* find the first device with a block */
2019 for (i=0; i<conf->copies; i++)
2020 if (!r10_bio->devs[i].bio->bi_status)
2023 if (i == conf->copies)
2027 fbio = r10_bio->devs[i].bio;
2028 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2029 fbio->bi_iter.bi_idx = 0;
2030 fpages = get_resync_pages(fbio)->pages;
2032 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2033 /* now find blocks with errors */
2034 for (i=0 ; i < conf->copies ; i++) {
2036 struct md_rdev *rdev;
2037 struct resync_pages *rp;
2039 tbio = r10_bio->devs[i].bio;
2041 if (tbio->bi_end_io != end_sync_read)
2046 tpages = get_resync_pages(tbio)->pages;
2047 d = r10_bio->devs[i].devnum;
2048 rdev = conf->mirrors[d].rdev;
2049 if (!r10_bio->devs[i].bio->bi_status) {
2050 /* We know that the bi_io_vec layout is the same for
2051 * both 'first' and 'i', so we just compare them.
2052 * All vec entries are PAGE_SIZE;
2054 int sectors = r10_bio->sectors;
2055 for (j = 0; j < vcnt; j++) {
2056 int len = PAGE_SIZE;
2057 if (sectors < (len / 512))
2058 len = sectors * 512;
2059 if (memcmp(page_address(fpages[j]),
2060 page_address(tpages[j]),
2067 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2068 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2069 /* Don't fix anything. */
2071 } else if (test_bit(FailFast, &rdev->flags)) {
2072 /* Just give up on this device */
2073 md_error(rdev->mddev, rdev);
2076 /* Ok, we need to write this bio, either to correct an
2077 * inconsistency or to correct an unreadable block.
2078 * First we need to fixup bv_offset, bv_len and
2079 * bi_vecs, as the read request might have corrupted these
2081 rp = get_resync_pages(tbio);
2084 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2086 rp->raid_bio = r10_bio;
2087 tbio->bi_private = rp;
2088 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2089 tbio->bi_end_io = end_sync_write;
2090 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2092 bio_copy_data(tbio, fbio);
2094 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2095 atomic_inc(&r10_bio->remaining);
2096 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2098 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2099 tbio->bi_opf |= MD_FAILFAST;
2100 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2101 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2102 generic_make_request(tbio);
2105 /* Now write out to any replacement devices
2108 for (i = 0; i < conf->copies; i++) {
2111 tbio = r10_bio->devs[i].repl_bio;
2112 if (!tbio || !tbio->bi_end_io)
2114 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2115 && r10_bio->devs[i].bio != fbio)
2116 bio_copy_data(tbio, fbio);
2117 d = r10_bio->devs[i].devnum;
2118 atomic_inc(&r10_bio->remaining);
2119 md_sync_acct(conf->mirrors[d].replacement->bdev,
2121 generic_make_request(tbio);
2125 if (atomic_dec_and_test(&r10_bio->remaining)) {
2126 md_done_sync(mddev, r10_bio->sectors, 1);
2132 * Now for the recovery code.
2133 * Recovery happens across physical sectors.
2134 * We recover all non-is_sync drives by finding the virtual address of
2135 * each, and then choose a working drive that also has that virt address.
2136 * There is a separate r10_bio for each non-in_sync drive.
2137 * Only the first two slots are in use. The first for reading,
2138 * The second for writing.
2141 static void fix_recovery_read_error(struct r10bio *r10_bio)
2143 /* We got a read error during recovery.
2144 * We repeat the read in smaller page-sized sections.
2145 * If a read succeeds, write it to the new device or record
2146 * a bad block if we cannot.
2147 * If a read fails, record a bad block on both old and
2150 struct mddev *mddev = r10_bio->mddev;
2151 struct r10conf *conf = mddev->private;
2152 struct bio *bio = r10_bio->devs[0].bio;
2154 int sectors = r10_bio->sectors;
2156 int dr = r10_bio->devs[0].devnum;
2157 int dw = r10_bio->devs[1].devnum;
2158 struct page **pages = get_resync_pages(bio)->pages;
2162 struct md_rdev *rdev;
2166 if (s > (PAGE_SIZE>>9))
2169 rdev = conf->mirrors[dr].rdev;
2170 addr = r10_bio->devs[0].addr + sect,
2171 ok = sync_page_io(rdev,
2175 REQ_OP_READ, 0, false);
2177 rdev = conf->mirrors[dw].rdev;
2178 addr = r10_bio->devs[1].addr + sect;
2179 ok = sync_page_io(rdev,
2183 REQ_OP_WRITE, 0, false);
2185 set_bit(WriteErrorSeen, &rdev->flags);
2186 if (!test_and_set_bit(WantReplacement,
2188 set_bit(MD_RECOVERY_NEEDED,
2189 &rdev->mddev->recovery);
2193 /* We don't worry if we cannot set a bad block -
2194 * it really is bad so there is no loss in not
2197 rdev_set_badblocks(rdev, addr, s, 0);
2199 if (rdev != conf->mirrors[dw].rdev) {
2200 /* need bad block on destination too */
2201 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2202 addr = r10_bio->devs[1].addr + sect;
2203 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2205 /* just abort the recovery */
2206 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2209 conf->mirrors[dw].recovery_disabled
2210 = mddev->recovery_disabled;
2211 set_bit(MD_RECOVERY_INTR,
2224 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2226 struct r10conf *conf = mddev->private;
2228 struct bio *wbio, *wbio2;
2230 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2231 fix_recovery_read_error(r10_bio);
2232 end_sync_request(r10_bio);
2237 * share the pages with the first bio
2238 * and submit the write request
2240 d = r10_bio->devs[1].devnum;
2241 wbio = r10_bio->devs[1].bio;
2242 wbio2 = r10_bio->devs[1].repl_bio;
2243 /* Need to test wbio2->bi_end_io before we call
2244 * generic_make_request as if the former is NULL,
2245 * the latter is free to free wbio2.
2247 if (wbio2 && !wbio2->bi_end_io)
2249 if (wbio->bi_end_io) {
2250 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2251 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2252 generic_make_request(wbio);
2255 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2256 md_sync_acct(conf->mirrors[d].replacement->bdev,
2257 bio_sectors(wbio2));
2258 generic_make_request(wbio2);
2263 * Used by fix_read_error() to decay the per rdev read_errors.
2264 * We halve the read error count for every hour that has elapsed
2265 * since the last recorded read error.
2268 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2271 unsigned long hours_since_last;
2272 unsigned int read_errors = atomic_read(&rdev->read_errors);
2274 cur_time_mon = ktime_get_seconds();
2276 if (rdev->last_read_error == 0) {
2277 /* first time we've seen a read error */
2278 rdev->last_read_error = cur_time_mon;
2282 hours_since_last = (long)(cur_time_mon -
2283 rdev->last_read_error) / 3600;
2285 rdev->last_read_error = cur_time_mon;
2288 * if hours_since_last is > the number of bits in read_errors
2289 * just set read errors to 0. We do this to avoid
2290 * overflowing the shift of read_errors by hours_since_last.
2292 if (hours_since_last >= 8 * sizeof(read_errors))
2293 atomic_set(&rdev->read_errors, 0);
2295 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2298 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2299 int sectors, struct page *page, int rw)
2304 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2305 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2307 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2311 set_bit(WriteErrorSeen, &rdev->flags);
2312 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2313 set_bit(MD_RECOVERY_NEEDED,
2314 &rdev->mddev->recovery);
2316 /* need to record an error - either for the block or the device */
2317 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2318 md_error(rdev->mddev, rdev);
2323 * This is a kernel thread which:
2325 * 1. Retries failed read operations on working mirrors.
2326 * 2. Updates the raid superblock when problems encounter.
2327 * 3. Performs writes following reads for array synchronising.
2330 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2332 int sect = 0; /* Offset from r10_bio->sector */
2333 int sectors = r10_bio->sectors;
2334 struct md_rdev *rdev;
2335 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2336 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2338 /* still own a reference to this rdev, so it cannot
2339 * have been cleared recently.
2341 rdev = conf->mirrors[d].rdev;
2343 if (test_bit(Faulty, &rdev->flags))
2344 /* drive has already been failed, just ignore any
2345 more fix_read_error() attempts */
2348 check_decay_read_errors(mddev, rdev);
2349 atomic_inc(&rdev->read_errors);
2350 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2351 char b[BDEVNAME_SIZE];
2352 bdevname(rdev->bdev, b);
2354 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2356 atomic_read(&rdev->read_errors), max_read_errors);
2357 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2359 md_error(mddev, rdev);
2360 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2366 int sl = r10_bio->read_slot;
2370 if (s > (PAGE_SIZE>>9))
2378 d = r10_bio->devs[sl].devnum;
2379 rdev = rcu_dereference(conf->mirrors[d].rdev);
2381 test_bit(In_sync, &rdev->flags) &&
2382 !test_bit(Faulty, &rdev->flags) &&
2383 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2384 &first_bad, &bad_sectors) == 0) {
2385 atomic_inc(&rdev->nr_pending);
2387 success = sync_page_io(rdev,
2388 r10_bio->devs[sl].addr +
2392 REQ_OP_READ, 0, false);
2393 rdev_dec_pending(rdev, mddev);
2399 if (sl == conf->copies)
2401 } while (!success && sl != r10_bio->read_slot);
2405 /* Cannot read from anywhere, just mark the block
2406 * as bad on the first device to discourage future
2409 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2410 rdev = conf->mirrors[dn].rdev;
2412 if (!rdev_set_badblocks(
2414 r10_bio->devs[r10_bio->read_slot].addr
2417 md_error(mddev, rdev);
2418 r10_bio->devs[r10_bio->read_slot].bio
2425 /* write it back and re-read */
2427 while (sl != r10_bio->read_slot) {
2428 char b[BDEVNAME_SIZE];
2433 d = r10_bio->devs[sl].devnum;
2434 rdev = rcu_dereference(conf->mirrors[d].rdev);
2436 test_bit(Faulty, &rdev->flags) ||
2437 !test_bit(In_sync, &rdev->flags))
2440 atomic_inc(&rdev->nr_pending);
2442 if (r10_sync_page_io(rdev,
2443 r10_bio->devs[sl].addr +
2445 s, conf->tmppage, WRITE)
2447 /* Well, this device is dead */
2448 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2450 (unsigned long long)(
2452 choose_data_offset(r10_bio,
2454 bdevname(rdev->bdev, b));
2455 pr_notice("md/raid10:%s: %s: failing drive\n",
2457 bdevname(rdev->bdev, b));
2459 rdev_dec_pending(rdev, mddev);
2463 while (sl != r10_bio->read_slot) {
2464 char b[BDEVNAME_SIZE];
2469 d = r10_bio->devs[sl].devnum;
2470 rdev = rcu_dereference(conf->mirrors[d].rdev);
2472 test_bit(Faulty, &rdev->flags) ||
2473 !test_bit(In_sync, &rdev->flags))
2476 atomic_inc(&rdev->nr_pending);
2478 switch (r10_sync_page_io(rdev,
2479 r10_bio->devs[sl].addr +
2484 /* Well, this device is dead */
2485 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2487 (unsigned long long)(
2489 choose_data_offset(r10_bio, rdev)),
2490 bdevname(rdev->bdev, b));
2491 pr_notice("md/raid10:%s: %s: failing drive\n",
2493 bdevname(rdev->bdev, b));
2496 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2498 (unsigned long long)(
2500 choose_data_offset(r10_bio, rdev)),
2501 bdevname(rdev->bdev, b));
2502 atomic_add(s, &rdev->corrected_errors);
2505 rdev_dec_pending(rdev, mddev);
2515 static int narrow_write_error(struct r10bio *r10_bio, int i)
2517 struct bio *bio = r10_bio->master_bio;
2518 struct mddev *mddev = r10_bio->mddev;
2519 struct r10conf *conf = mddev->private;
2520 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2521 /* bio has the data to be written to slot 'i' where
2522 * we just recently had a write error.
2523 * We repeatedly clone the bio and trim down to one block,
2524 * then try the write. Where the write fails we record
2526 * It is conceivable that the bio doesn't exactly align with
2527 * blocks. We must handle this.
2529 * We currently own a reference to the rdev.
2535 int sect_to_write = r10_bio->sectors;
2538 if (rdev->badblocks.shift < 0)
2541 block_sectors = roundup(1 << rdev->badblocks.shift,
2542 bdev_logical_block_size(rdev->bdev) >> 9);
2543 sector = r10_bio->sector;
2544 sectors = ((r10_bio->sector + block_sectors)
2545 & ~(sector_t)(block_sectors - 1))
2548 while (sect_to_write) {
2551 if (sectors > sect_to_write)
2552 sectors = sect_to_write;
2553 /* Write at 'sector' for 'sectors' */
2554 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2555 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2556 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2557 wbio->bi_iter.bi_sector = wsector +
2558 choose_data_offset(r10_bio, rdev);
2559 bio_set_dev(wbio, rdev->bdev);
2560 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2562 if (submit_bio_wait(wbio) < 0)
2564 ok = rdev_set_badblocks(rdev, wsector,
2569 sect_to_write -= sectors;
2571 sectors = block_sectors;
2576 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2578 int slot = r10_bio->read_slot;
2580 struct r10conf *conf = mddev->private;
2581 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2583 /* we got a read error. Maybe the drive is bad. Maybe just
2584 * the block and we can fix it.
2585 * We freeze all other IO, and try reading the block from
2586 * other devices. When we find one, we re-write
2587 * and check it that fixes the read error.
2588 * This is all done synchronously while the array is
2591 bio = r10_bio->devs[slot].bio;
2593 r10_bio->devs[slot].bio = NULL;
2596 r10_bio->devs[slot].bio = IO_BLOCKED;
2597 else if (!test_bit(FailFast, &rdev->flags)) {
2598 freeze_array(conf, 1);
2599 fix_read_error(conf, mddev, r10_bio);
2600 unfreeze_array(conf);
2602 md_error(mddev, rdev);
2604 rdev_dec_pending(rdev, mddev);
2605 allow_barrier(conf);
2607 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2610 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2612 /* Some sort of write request has finished and it
2613 * succeeded in writing where we thought there was a
2614 * bad block. So forget the bad block.
2615 * Or possibly if failed and we need to record
2619 struct md_rdev *rdev;
2621 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2622 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2623 for (m = 0; m < conf->copies; m++) {
2624 int dev = r10_bio->devs[m].devnum;
2625 rdev = conf->mirrors[dev].rdev;
2626 if (r10_bio->devs[m].bio == NULL ||
2627 r10_bio->devs[m].bio->bi_end_io == NULL)
2629 if (!r10_bio->devs[m].bio->bi_status) {
2630 rdev_clear_badblocks(
2632 r10_bio->devs[m].addr,
2633 r10_bio->sectors, 0);
2635 if (!rdev_set_badblocks(
2637 r10_bio->devs[m].addr,
2638 r10_bio->sectors, 0))
2639 md_error(conf->mddev, rdev);
2641 rdev = conf->mirrors[dev].replacement;
2642 if (r10_bio->devs[m].repl_bio == NULL ||
2643 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2646 if (!r10_bio->devs[m].repl_bio->bi_status) {
2647 rdev_clear_badblocks(
2649 r10_bio->devs[m].addr,
2650 r10_bio->sectors, 0);
2652 if (!rdev_set_badblocks(
2654 r10_bio->devs[m].addr,
2655 r10_bio->sectors, 0))
2656 md_error(conf->mddev, rdev);
2662 for (m = 0; m < conf->copies; m++) {
2663 int dev = r10_bio->devs[m].devnum;
2664 struct bio *bio = r10_bio->devs[m].bio;
2665 rdev = conf->mirrors[dev].rdev;
2666 if (bio == IO_MADE_GOOD) {
2667 rdev_clear_badblocks(
2669 r10_bio->devs[m].addr,
2670 r10_bio->sectors, 0);
2671 rdev_dec_pending(rdev, conf->mddev);
2672 } else if (bio != NULL && bio->bi_status) {
2674 if (!narrow_write_error(r10_bio, m)) {
2675 md_error(conf->mddev, rdev);
2676 set_bit(R10BIO_Degraded,
2679 rdev_dec_pending(rdev, conf->mddev);
2681 bio = r10_bio->devs[m].repl_bio;
2682 rdev = conf->mirrors[dev].replacement;
2683 if (rdev && bio == IO_MADE_GOOD) {
2684 rdev_clear_badblocks(
2686 r10_bio->devs[m].addr,
2687 r10_bio->sectors, 0);
2688 rdev_dec_pending(rdev, conf->mddev);
2692 spin_lock_irq(&conf->device_lock);
2693 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2695 spin_unlock_irq(&conf->device_lock);
2697 * In case freeze_array() is waiting for condition
2698 * nr_pending == nr_queued + extra to be true.
2700 wake_up(&conf->wait_barrier);
2701 md_wakeup_thread(conf->mddev->thread);
2703 if (test_bit(R10BIO_WriteError,
2705 close_write(r10_bio);
2706 raid_end_bio_io(r10_bio);
2711 static void raid10d(struct md_thread *thread)
2713 struct mddev *mddev = thread->mddev;
2714 struct r10bio *r10_bio;
2715 unsigned long flags;
2716 struct r10conf *conf = mddev->private;
2717 struct list_head *head = &conf->retry_list;
2718 struct blk_plug plug;
2720 md_check_recovery(mddev);
2722 if (!list_empty_careful(&conf->bio_end_io_list) &&
2723 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2725 spin_lock_irqsave(&conf->device_lock, flags);
2726 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2727 while (!list_empty(&conf->bio_end_io_list)) {
2728 list_move(conf->bio_end_io_list.prev, &tmp);
2732 spin_unlock_irqrestore(&conf->device_lock, flags);
2733 while (!list_empty(&tmp)) {
2734 r10_bio = list_first_entry(&tmp, struct r10bio,
2736 list_del(&r10_bio->retry_list);
2737 if (mddev->degraded)
2738 set_bit(R10BIO_Degraded, &r10_bio->state);
2740 if (test_bit(R10BIO_WriteError,
2742 close_write(r10_bio);
2743 raid_end_bio_io(r10_bio);
2747 blk_start_plug(&plug);
2750 flush_pending_writes(conf);
2752 spin_lock_irqsave(&conf->device_lock, flags);
2753 if (list_empty(head)) {
2754 spin_unlock_irqrestore(&conf->device_lock, flags);
2757 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2758 list_del(head->prev);
2760 spin_unlock_irqrestore(&conf->device_lock, flags);
2762 mddev = r10_bio->mddev;
2763 conf = mddev->private;
2764 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2765 test_bit(R10BIO_WriteError, &r10_bio->state))
2766 handle_write_completed(conf, r10_bio);
2767 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2768 reshape_request_write(mddev, r10_bio);
2769 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2770 sync_request_write(mddev, r10_bio);
2771 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2772 recovery_request_write(mddev, r10_bio);
2773 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2774 handle_read_error(mddev, r10_bio);
2779 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2780 md_check_recovery(mddev);
2782 blk_finish_plug(&plug);
2785 static int init_resync(struct r10conf *conf)
2789 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2790 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2791 conf->have_replacement = 0;
2792 for (i = 0; i < conf->geo.raid_disks; i++)
2793 if (conf->mirrors[i].replacement)
2794 conf->have_replacement = 1;
2795 ret = mempool_init(&conf->r10buf_pool, buffs,
2796 r10buf_pool_alloc, r10buf_pool_free, conf);
2799 conf->next_resync = 0;
2803 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2805 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2806 struct rsync_pages *rp;
2811 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2812 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2813 nalloc = conf->copies; /* resync */
2815 nalloc = 2; /* recovery */
2817 for (i = 0; i < nalloc; i++) {
2818 bio = r10bio->devs[i].bio;
2819 rp = bio->bi_private;
2821 bio->bi_private = rp;
2822 bio = r10bio->devs[i].repl_bio;
2824 rp = bio->bi_private;
2826 bio->bi_private = rp;
2833 * Set cluster_sync_high since we need other nodes to add the
2834 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2836 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2838 sector_t window_size;
2839 int extra_chunk, chunks;
2842 * First, here we define "stripe" as a unit which across
2843 * all member devices one time, so we get chunks by use
2844 * raid_disks / near_copies. Otherwise, if near_copies is
2845 * close to raid_disks, then resync window could increases
2846 * linearly with the increase of raid_disks, which means
2847 * we will suspend a really large IO window while it is not
2848 * necessary. If raid_disks is not divisible by near_copies,
2849 * an extra chunk is needed to ensure the whole "stripe" is
2853 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2854 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2858 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2861 * At least use a 32M window to align with raid1's resync window
2863 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2864 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2866 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2870 * perform a "sync" on one "block"
2872 * We need to make sure that no normal I/O request - particularly write
2873 * requests - conflict with active sync requests.
2875 * This is achieved by tracking pending requests and a 'barrier' concept
2876 * that can be installed to exclude normal IO requests.
2878 * Resync and recovery are handled very differently.
2879 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2881 * For resync, we iterate over virtual addresses, read all copies,
2882 * and update if there are differences. If only one copy is live,
2884 * For recovery, we iterate over physical addresses, read a good
2885 * value for each non-in_sync drive, and over-write.
2887 * So, for recovery we may have several outstanding complex requests for a
2888 * given address, one for each out-of-sync device. We model this by allocating
2889 * a number of r10_bio structures, one for each out-of-sync device.
2890 * As we setup these structures, we collect all bio's together into a list
2891 * which we then process collectively to add pages, and then process again
2892 * to pass to generic_make_request.
2894 * The r10_bio structures are linked using a borrowed master_bio pointer.
2895 * This link is counted in ->remaining. When the r10_bio that points to NULL
2896 * has its remaining count decremented to 0, the whole complex operation
2901 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2904 struct r10conf *conf = mddev->private;
2905 struct r10bio *r10_bio;
2906 struct bio *biolist = NULL, *bio;
2907 sector_t max_sector, nr_sectors;
2910 sector_t sync_blocks;
2911 sector_t sectors_skipped = 0;
2912 int chunks_skipped = 0;
2913 sector_t chunk_mask = conf->geo.chunk_mask;
2916 if (!mempool_initialized(&conf->r10buf_pool))
2917 if (init_resync(conf))
2921 * Allow skipping a full rebuild for incremental assembly
2922 * of a clean array, like RAID1 does.
2924 if (mddev->bitmap == NULL &&
2925 mddev->recovery_cp == MaxSector &&
2926 mddev->reshape_position == MaxSector &&
2927 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2928 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2929 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2930 conf->fullsync == 0) {
2932 return mddev->dev_sectors - sector_nr;
2936 max_sector = mddev->dev_sectors;
2937 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2938 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2939 max_sector = mddev->resync_max_sectors;
2940 if (sector_nr >= max_sector) {
2941 conf->cluster_sync_low = 0;
2942 conf->cluster_sync_high = 0;
2944 /* If we aborted, we need to abort the
2945 * sync on the 'current' bitmap chucks (there can
2946 * be several when recovering multiple devices).
2947 * as we may have started syncing it but not finished.
2948 * We can find the current address in
2949 * mddev->curr_resync, but for recovery,
2950 * we need to convert that to several
2951 * virtual addresses.
2953 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2959 if (mddev->curr_resync < max_sector) { /* aborted */
2960 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2961 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2963 else for (i = 0; i < conf->geo.raid_disks; i++) {
2965 raid10_find_virt(conf, mddev->curr_resync, i);
2966 md_bitmap_end_sync(mddev->bitmap, sect,
2970 /* completed sync */
2971 if ((!mddev->bitmap || conf->fullsync)
2972 && conf->have_replacement
2973 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2974 /* Completed a full sync so the replacements
2975 * are now fully recovered.
2978 for (i = 0; i < conf->geo.raid_disks; i++) {
2979 struct md_rdev *rdev =
2980 rcu_dereference(conf->mirrors[i].replacement);
2982 rdev->recovery_offset = MaxSector;
2988 md_bitmap_close_sync(mddev->bitmap);
2991 return sectors_skipped;
2994 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2995 return reshape_request(mddev, sector_nr, skipped);
2997 if (chunks_skipped >= conf->geo.raid_disks) {
2998 /* if there has been nothing to do on any drive,
2999 * then there is nothing to do at all..
3002 return (max_sector - sector_nr) + sectors_skipped;
3005 if (max_sector > mddev->resync_max)
3006 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3008 /* make sure whole request will fit in a chunk - if chunks
3011 if (conf->geo.near_copies < conf->geo.raid_disks &&
3012 max_sector > (sector_nr | chunk_mask))
3013 max_sector = (sector_nr | chunk_mask) + 1;
3016 * If there is non-resync activity waiting for a turn, then let it
3017 * though before starting on this new sync request.
3019 if (conf->nr_waiting)
3020 schedule_timeout_uninterruptible(1);
3022 /* Again, very different code for resync and recovery.
3023 * Both must result in an r10bio with a list of bios that
3024 * have bi_end_io, bi_sector, bi_disk set,
3025 * and bi_private set to the r10bio.
3026 * For recovery, we may actually create several r10bios
3027 * with 2 bios in each, that correspond to the bios in the main one.
3028 * In this case, the subordinate r10bios link back through a
3029 * borrowed master_bio pointer, and the counter in the master
3030 * includes a ref from each subordinate.
3032 /* First, we decide what to do and set ->bi_end_io
3033 * To end_sync_read if we want to read, and
3034 * end_sync_write if we will want to write.
3037 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3038 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3039 /* recovery... the complicated one */
3043 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3049 int need_recover = 0;
3050 int need_replace = 0;
3051 struct raid10_info *mirror = &conf->mirrors[i];
3052 struct md_rdev *mrdev, *mreplace;
3055 mrdev = rcu_dereference(mirror->rdev);
3056 mreplace = rcu_dereference(mirror->replacement);
3058 if (mrdev != NULL &&
3059 !test_bit(Faulty, &mrdev->flags) &&
3060 !test_bit(In_sync, &mrdev->flags))
3062 if (mreplace != NULL &&
3063 !test_bit(Faulty, &mreplace->flags))
3066 if (!need_recover && !need_replace) {
3072 /* want to reconstruct this device */
3074 sect = raid10_find_virt(conf, sector_nr, i);
3075 if (sect >= mddev->resync_max_sectors) {
3076 /* last stripe is not complete - don't
3077 * try to recover this sector.
3082 if (mreplace && test_bit(Faulty, &mreplace->flags))
3084 /* Unless we are doing a full sync, or a replacement
3085 * we only need to recover the block if it is set in
3088 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3090 if (sync_blocks < max_sync)
3091 max_sync = sync_blocks;
3095 /* yep, skip the sync_blocks here, but don't assume
3096 * that there will never be anything to do here
3098 chunks_skipped = -1;
3102 atomic_inc(&mrdev->nr_pending);
3104 atomic_inc(&mreplace->nr_pending);
3107 r10_bio = raid10_alloc_init_r10buf(conf);
3109 raise_barrier(conf, rb2 != NULL);
3110 atomic_set(&r10_bio->remaining, 0);
3112 r10_bio->master_bio = (struct bio*)rb2;
3114 atomic_inc(&rb2->remaining);
3115 r10_bio->mddev = mddev;
3116 set_bit(R10BIO_IsRecover, &r10_bio->state);
3117 r10_bio->sector = sect;
3119 raid10_find_phys(conf, r10_bio);
3121 /* Need to check if the array will still be
3125 for (j = 0; j < conf->geo.raid_disks; j++) {
3126 struct md_rdev *rdev = rcu_dereference(
3127 conf->mirrors[j].rdev);
3128 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3134 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3135 &sync_blocks, still_degraded);
3138 for (j=0; j<conf->copies;j++) {
3140 int d = r10_bio->devs[j].devnum;
3141 sector_t from_addr, to_addr;
3142 struct md_rdev *rdev =
3143 rcu_dereference(conf->mirrors[d].rdev);
3144 sector_t sector, first_bad;
3147 !test_bit(In_sync, &rdev->flags))
3149 /* This is where we read from */
3151 sector = r10_bio->devs[j].addr;
3153 if (is_badblock(rdev, sector, max_sync,
3154 &first_bad, &bad_sectors)) {
3155 if (first_bad > sector)
3156 max_sync = first_bad - sector;
3158 bad_sectors -= (sector
3160 if (max_sync > bad_sectors)
3161 max_sync = bad_sectors;
3165 bio = r10_bio->devs[0].bio;
3166 bio->bi_next = biolist;
3168 bio->bi_end_io = end_sync_read;
3169 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3170 if (test_bit(FailFast, &rdev->flags))
3171 bio->bi_opf |= MD_FAILFAST;
3172 from_addr = r10_bio->devs[j].addr;
3173 bio->bi_iter.bi_sector = from_addr +
3175 bio_set_dev(bio, rdev->bdev);
3176 atomic_inc(&rdev->nr_pending);
3177 /* and we write to 'i' (if not in_sync) */
3179 for (k=0; k<conf->copies; k++)
3180 if (r10_bio->devs[k].devnum == i)
3182 BUG_ON(k == conf->copies);
3183 to_addr = r10_bio->devs[k].addr;
3184 r10_bio->devs[0].devnum = d;
3185 r10_bio->devs[0].addr = from_addr;
3186 r10_bio->devs[1].devnum = i;
3187 r10_bio->devs[1].addr = to_addr;
3190 bio = r10_bio->devs[1].bio;
3191 bio->bi_next = biolist;
3193 bio->bi_end_io = end_sync_write;
3194 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3195 bio->bi_iter.bi_sector = to_addr
3196 + mrdev->data_offset;
3197 bio_set_dev(bio, mrdev->bdev);
3198 atomic_inc(&r10_bio->remaining);
3200 r10_bio->devs[1].bio->bi_end_io = NULL;
3202 /* and maybe write to replacement */
3203 bio = r10_bio->devs[1].repl_bio;
3205 bio->bi_end_io = NULL;
3206 /* Note: if need_replace, then bio
3207 * cannot be NULL as r10buf_pool_alloc will
3208 * have allocated it.
3212 bio->bi_next = biolist;
3214 bio->bi_end_io = end_sync_write;
3215 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3216 bio->bi_iter.bi_sector = to_addr +
3217 mreplace->data_offset;
3218 bio_set_dev(bio, mreplace->bdev);
3219 atomic_inc(&r10_bio->remaining);
3223 if (j == conf->copies) {
3224 /* Cannot recover, so abort the recovery or
3225 * record a bad block */
3227 /* problem is that there are bad blocks
3228 * on other device(s)
3231 for (k = 0; k < conf->copies; k++)
3232 if (r10_bio->devs[k].devnum == i)
3234 if (!test_bit(In_sync,
3236 && !rdev_set_badblocks(
3238 r10_bio->devs[k].addr,
3242 !rdev_set_badblocks(
3244 r10_bio->devs[k].addr,
3249 if (!test_and_set_bit(MD_RECOVERY_INTR,
3251 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3253 mirror->recovery_disabled
3254 = mddev->recovery_disabled;
3258 atomic_dec(&rb2->remaining);
3260 rdev_dec_pending(mrdev, mddev);
3262 rdev_dec_pending(mreplace, mddev);
3265 rdev_dec_pending(mrdev, mddev);
3267 rdev_dec_pending(mreplace, mddev);
3268 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3269 /* Only want this if there is elsewhere to
3270 * read from. 'j' is currently the first
3274 for (; j < conf->copies; j++) {
3275 int d = r10_bio->devs[j].devnum;
3276 if (conf->mirrors[d].rdev &&
3278 &conf->mirrors[d].rdev->flags))
3282 r10_bio->devs[0].bio->bi_opf
3286 if (biolist == NULL) {
3288 struct r10bio *rb2 = r10_bio;
3289 r10_bio = (struct r10bio*) rb2->master_bio;
3290 rb2->master_bio = NULL;
3296 /* resync. Schedule a read for every block at this virt offset */
3300 * Since curr_resync_completed could probably not update in
3301 * time, and we will set cluster_sync_low based on it.
3302 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3303 * safety reason, which ensures curr_resync_completed is
3304 * updated in bitmap_cond_end_sync.
3306 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3307 mddev_is_clustered(mddev) &&
3308 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3310 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3311 &sync_blocks, mddev->degraded) &&
3312 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3313 &mddev->recovery)) {
3314 /* We can skip this block */
3316 return sync_blocks + sectors_skipped;
3318 if (sync_blocks < max_sync)
3319 max_sync = sync_blocks;
3320 r10_bio = raid10_alloc_init_r10buf(conf);
3323 r10_bio->mddev = mddev;
3324 atomic_set(&r10_bio->remaining, 0);
3325 raise_barrier(conf, 0);
3326 conf->next_resync = sector_nr;
3328 r10_bio->master_bio = NULL;
3329 r10_bio->sector = sector_nr;
3330 set_bit(R10BIO_IsSync, &r10_bio->state);
3331 raid10_find_phys(conf, r10_bio);
3332 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3334 for (i = 0; i < conf->copies; i++) {
3335 int d = r10_bio->devs[i].devnum;
3336 sector_t first_bad, sector;
3338 struct md_rdev *rdev;
3340 if (r10_bio->devs[i].repl_bio)
3341 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3343 bio = r10_bio->devs[i].bio;
3344 bio->bi_status = BLK_STS_IOERR;
3346 rdev = rcu_dereference(conf->mirrors[d].rdev);
3347 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3351 sector = r10_bio->devs[i].addr;
3352 if (is_badblock(rdev, sector, max_sync,
3353 &first_bad, &bad_sectors)) {
3354 if (first_bad > sector)
3355 max_sync = first_bad - sector;
3357 bad_sectors -= (sector - first_bad);
3358 if (max_sync > bad_sectors)
3359 max_sync = bad_sectors;
3364 atomic_inc(&rdev->nr_pending);
3365 atomic_inc(&r10_bio->remaining);
3366 bio->bi_next = biolist;
3368 bio->bi_end_io = end_sync_read;
3369 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3370 if (test_bit(FailFast, &rdev->flags))
3371 bio->bi_opf |= MD_FAILFAST;
3372 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3373 bio_set_dev(bio, rdev->bdev);
3376 rdev = rcu_dereference(conf->mirrors[d].replacement);
3377 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3381 atomic_inc(&rdev->nr_pending);
3383 /* Need to set up for writing to the replacement */
3384 bio = r10_bio->devs[i].repl_bio;
3385 bio->bi_status = BLK_STS_IOERR;
3387 sector = r10_bio->devs[i].addr;
3388 bio->bi_next = biolist;
3390 bio->bi_end_io = end_sync_write;
3391 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3392 if (test_bit(FailFast, &rdev->flags))
3393 bio->bi_opf |= MD_FAILFAST;
3394 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3395 bio_set_dev(bio, rdev->bdev);
3401 for (i=0; i<conf->copies; i++) {
3402 int d = r10_bio->devs[i].devnum;
3403 if (r10_bio->devs[i].bio->bi_end_io)
3404 rdev_dec_pending(conf->mirrors[d].rdev,
3406 if (r10_bio->devs[i].repl_bio &&
3407 r10_bio->devs[i].repl_bio->bi_end_io)
3409 conf->mirrors[d].replacement,
3419 if (sector_nr + max_sync < max_sector)
3420 max_sector = sector_nr + max_sync;
3423 int len = PAGE_SIZE;
3424 if (sector_nr + (len>>9) > max_sector)
3425 len = (max_sector - sector_nr) << 9;
3428 for (bio= biolist ; bio ; bio=bio->bi_next) {
3429 struct resync_pages *rp = get_resync_pages(bio);
3430 page = resync_fetch_page(rp, page_idx);
3432 * won't fail because the vec table is big enough
3433 * to hold all these pages
3435 bio_add_page(bio, page, len, 0);
3437 nr_sectors += len>>9;
3438 sector_nr += len>>9;
3439 } while (++page_idx < RESYNC_PAGES);
3440 r10_bio->sectors = nr_sectors;
3442 if (mddev_is_clustered(mddev) &&
3443 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3444 /* It is resync not recovery */
3445 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3446 conf->cluster_sync_low = mddev->curr_resync_completed;
3447 raid10_set_cluster_sync_high(conf);
3448 /* Send resync message */
3449 md_cluster_ops->resync_info_update(mddev,
3450 conf->cluster_sync_low,
3451 conf->cluster_sync_high);
3453 } else if (mddev_is_clustered(mddev)) {
3454 /* This is recovery not resync */
3455 sector_t sect_va1, sect_va2;
3456 bool broadcast_msg = false;
3458 for (i = 0; i < conf->geo.raid_disks; i++) {
3460 * sector_nr is a device address for recovery, so we
3461 * need translate it to array address before compare
3462 * with cluster_sync_high.
3464 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3466 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3467 broadcast_msg = true;
3469 * curr_resync_completed is similar as
3470 * sector_nr, so make the translation too.
3472 sect_va2 = raid10_find_virt(conf,
3473 mddev->curr_resync_completed, i);
3475 if (conf->cluster_sync_low == 0 ||
3476 conf->cluster_sync_low > sect_va2)
3477 conf->cluster_sync_low = sect_va2;
3480 if (broadcast_msg) {
3481 raid10_set_cluster_sync_high(conf);
3482 md_cluster_ops->resync_info_update(mddev,
3483 conf->cluster_sync_low,
3484 conf->cluster_sync_high);
3490 biolist = biolist->bi_next;
3492 bio->bi_next = NULL;
3493 r10_bio = get_resync_r10bio(bio);
3494 r10_bio->sectors = nr_sectors;
3496 if (bio->bi_end_io == end_sync_read) {
3497 md_sync_acct_bio(bio, nr_sectors);
3499 generic_make_request(bio);
3503 if (sectors_skipped)
3504 /* pretend they weren't skipped, it makes
3505 * no important difference in this case
3507 md_done_sync(mddev, sectors_skipped, 1);
3509 return sectors_skipped + nr_sectors;
3511 /* There is nowhere to write, so all non-sync
3512 * drives must be failed or in resync, all drives
3513 * have a bad block, so try the next chunk...
3515 if (sector_nr + max_sync < max_sector)
3516 max_sector = sector_nr + max_sync;
3518 sectors_skipped += (max_sector - sector_nr);
3520 sector_nr = max_sector;
3525 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3528 struct r10conf *conf = mddev->private;
3531 raid_disks = min(conf->geo.raid_disks,
3532 conf->prev.raid_disks);
3534 sectors = conf->dev_sectors;
3536 size = sectors >> conf->geo.chunk_shift;
3537 sector_div(size, conf->geo.far_copies);
3538 size = size * raid_disks;
3539 sector_div(size, conf->geo.near_copies);
3541 return size << conf->geo.chunk_shift;
3544 static void calc_sectors(struct r10conf *conf, sector_t size)
3546 /* Calculate the number of sectors-per-device that will
3547 * actually be used, and set conf->dev_sectors and
3551 size = size >> conf->geo.chunk_shift;
3552 sector_div(size, conf->geo.far_copies);
3553 size = size * conf->geo.raid_disks;
3554 sector_div(size, conf->geo.near_copies);
3555 /* 'size' is now the number of chunks in the array */
3556 /* calculate "used chunks per device" */
3557 size = size * conf->copies;
3559 /* We need to round up when dividing by raid_disks to
3560 * get the stride size.
3562 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3564 conf->dev_sectors = size << conf->geo.chunk_shift;
3566 if (conf->geo.far_offset)
3567 conf->geo.stride = 1 << conf->geo.chunk_shift;
3569 sector_div(size, conf->geo.far_copies);
3570 conf->geo.stride = size << conf->geo.chunk_shift;
3574 enum geo_type {geo_new, geo_old, geo_start};
3575 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3578 int layout, chunk, disks;
3581 layout = mddev->layout;
3582 chunk = mddev->chunk_sectors;
3583 disks = mddev->raid_disks - mddev->delta_disks;
3586 layout = mddev->new_layout;
3587 chunk = mddev->new_chunk_sectors;
3588 disks = mddev->raid_disks;
3590 default: /* avoid 'may be unused' warnings */
3591 case geo_start: /* new when starting reshape - raid_disks not
3593 layout = mddev->new_layout;
3594 chunk = mddev->new_chunk_sectors;
3595 disks = mddev->raid_disks + mddev->delta_disks;
3600 if (chunk < (PAGE_SIZE >> 9) ||
3601 !is_power_of_2(chunk))
3604 fc = (layout >> 8) & 255;
3605 fo = layout & (1<<16);
3606 geo->raid_disks = disks;
3607 geo->near_copies = nc;
3608 geo->far_copies = fc;
3609 geo->far_offset = fo;
3610 switch (layout >> 17) {
3611 case 0: /* original layout. simple but not always optimal */
3612 geo->far_set_size = disks;
3614 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3615 * actually using this, but leave code here just in case.*/
3616 geo->far_set_size = disks/fc;
3617 WARN(geo->far_set_size < fc,
3618 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3620 case 2: /* "improved" layout fixed to match documentation */
3621 geo->far_set_size = fc * nc;
3623 default: /* Not a valid layout */
3626 geo->chunk_mask = chunk - 1;
3627 geo->chunk_shift = ffz(~chunk);
3631 static struct r10conf *setup_conf(struct mddev *mddev)
3633 struct r10conf *conf = NULL;
3638 copies = setup_geo(&geo, mddev, geo_new);
3641 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3642 mdname(mddev), PAGE_SIZE);
3646 if (copies < 2 || copies > mddev->raid_disks) {
3647 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3648 mdname(mddev), mddev->new_layout);
3653 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3657 /* FIXME calc properly */
3658 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3659 sizeof(struct raid10_info),
3664 conf->tmppage = alloc_page(GFP_KERNEL);
3669 conf->copies = copies;
3670 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3671 rbio_pool_free, conf);
3675 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3679 calc_sectors(conf, mddev->dev_sectors);
3680 if (mddev->reshape_position == MaxSector) {
3681 conf->prev = conf->geo;
3682 conf->reshape_progress = MaxSector;
3684 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3688 conf->reshape_progress = mddev->reshape_position;
3689 if (conf->prev.far_offset)
3690 conf->prev.stride = 1 << conf->prev.chunk_shift;
3692 /* far_copies must be 1 */
3693 conf->prev.stride = conf->dev_sectors;
3695 conf->reshape_safe = conf->reshape_progress;
3696 spin_lock_init(&conf->device_lock);
3697 INIT_LIST_HEAD(&conf->retry_list);
3698 INIT_LIST_HEAD(&conf->bio_end_io_list);
3700 spin_lock_init(&conf->resync_lock);
3701 init_waitqueue_head(&conf->wait_barrier);
3702 atomic_set(&conf->nr_pending, 0);
3705 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3709 conf->mddev = mddev;
3714 mempool_exit(&conf->r10bio_pool);
3715 kfree(conf->mirrors);
3716 safe_put_page(conf->tmppage);
3717 bioset_exit(&conf->bio_split);
3720 return ERR_PTR(err);
3723 static int raid10_run(struct mddev *mddev)
3725 struct r10conf *conf;
3726 int i, disk_idx, chunk_size;
3727 struct raid10_info *disk;
3728 struct md_rdev *rdev;
3730 sector_t min_offset_diff = 0;
3732 bool discard_supported = false;
3734 if (mddev_init_writes_pending(mddev) < 0)
3737 if (mddev->private == NULL) {
3738 conf = setup_conf(mddev);
3740 return PTR_ERR(conf);
3741 mddev->private = conf;
3743 conf = mddev->private;
3747 if (mddev_is_clustered(conf->mddev)) {
3750 fc = (mddev->layout >> 8) & 255;
3751 fo = mddev->layout & (1<<16);
3752 if (fc > 1 || fo > 0) {
3753 pr_err("only near layout is supported by clustered"
3759 mddev->thread = conf->thread;
3760 conf->thread = NULL;
3762 chunk_size = mddev->chunk_sectors << 9;
3764 blk_queue_max_discard_sectors(mddev->queue,
3765 mddev->chunk_sectors);
3766 blk_queue_max_write_same_sectors(mddev->queue, 0);
3767 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3768 blk_queue_io_min(mddev->queue, chunk_size);
3769 if (conf->geo.raid_disks % conf->geo.near_copies)
3770 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3772 blk_queue_io_opt(mddev->queue, chunk_size *
3773 (conf->geo.raid_disks / conf->geo.near_copies));
3776 rdev_for_each(rdev, mddev) {
3779 disk_idx = rdev->raid_disk;
3782 if (disk_idx >= conf->geo.raid_disks &&
3783 disk_idx >= conf->prev.raid_disks)
3785 disk = conf->mirrors + disk_idx;
3787 if (test_bit(Replacement, &rdev->flags)) {
3788 if (disk->replacement)
3790 disk->replacement = rdev;
3796 diff = (rdev->new_data_offset - rdev->data_offset);
3797 if (!mddev->reshape_backwards)
3801 if (first || diff < min_offset_diff)
3802 min_offset_diff = diff;
3805 disk_stack_limits(mddev->gendisk, rdev->bdev,
3806 rdev->data_offset << 9);
3808 disk->head_position = 0;
3810 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3811 discard_supported = true;
3816 if (discard_supported)
3817 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3820 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3823 /* need to check that every block has at least one working mirror */
3824 if (!enough(conf, -1)) {
3825 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3830 if (conf->reshape_progress != MaxSector) {
3831 /* must ensure that shape change is supported */
3832 if (conf->geo.far_copies != 1 &&
3833 conf->geo.far_offset == 0)
3835 if (conf->prev.far_copies != 1 &&
3836 conf->prev.far_offset == 0)
3840 mddev->degraded = 0;
3842 i < conf->geo.raid_disks
3843 || i < conf->prev.raid_disks;
3846 disk = conf->mirrors + i;
3848 if (!disk->rdev && disk->replacement) {
3849 /* The replacement is all we have - use it */
3850 disk->rdev = disk->replacement;
3851 disk->replacement = NULL;
3852 clear_bit(Replacement, &disk->rdev->flags);
3856 !test_bit(In_sync, &disk->rdev->flags)) {
3857 disk->head_position = 0;
3860 disk->rdev->saved_raid_disk < 0)
3864 if (disk->replacement &&
3865 !test_bit(In_sync, &disk->replacement->flags) &&
3866 disk->replacement->saved_raid_disk < 0) {
3870 disk->recovery_disabled = mddev->recovery_disabled - 1;
3873 if (mddev->recovery_cp != MaxSector)
3874 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3876 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3877 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3878 conf->geo.raid_disks);
3880 * Ok, everything is just fine now
3882 mddev->dev_sectors = conf->dev_sectors;
3883 size = raid10_size(mddev, 0, 0);
3884 md_set_array_sectors(mddev, size);
3885 mddev->resync_max_sectors = size;
3886 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3889 int stripe = conf->geo.raid_disks *
3890 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3892 /* Calculate max read-ahead size.
3893 * We need to readahead at least twice a whole stripe....
3896 stripe /= conf->geo.near_copies;
3897 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3898 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3901 if (md_integrity_register(mddev))
3904 if (conf->reshape_progress != MaxSector) {
3905 unsigned long before_length, after_length;
3907 before_length = ((1 << conf->prev.chunk_shift) *
3908 conf->prev.far_copies);
3909 after_length = ((1 << conf->geo.chunk_shift) *
3910 conf->geo.far_copies);
3912 if (max(before_length, after_length) > min_offset_diff) {
3913 /* This cannot work */
3914 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3917 conf->offset_diff = min_offset_diff;
3919 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3920 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3921 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3922 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3923 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3925 if (!mddev->sync_thread)
3932 md_unregister_thread(&mddev->thread);
3933 mempool_exit(&conf->r10bio_pool);
3934 safe_put_page(conf->tmppage);
3935 kfree(conf->mirrors);
3937 mddev->private = NULL;
3942 static void raid10_free(struct mddev *mddev, void *priv)
3944 struct r10conf *conf = priv;
3946 mempool_exit(&conf->r10bio_pool);
3947 safe_put_page(conf->tmppage);
3948 kfree(conf->mirrors);
3949 kfree(conf->mirrors_old);
3950 kfree(conf->mirrors_new);
3951 bioset_exit(&conf->bio_split);
3955 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3957 struct r10conf *conf = mddev->private;
3960 raise_barrier(conf, 0);
3962 lower_barrier(conf);
3965 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3967 /* Resize of 'far' arrays is not supported.
3968 * For 'near' and 'offset' arrays we can set the
3969 * number of sectors used to be an appropriate multiple
3970 * of the chunk size.
3971 * For 'offset', this is far_copies*chunksize.
3972 * For 'near' the multiplier is the LCM of
3973 * near_copies and raid_disks.
3974 * So if far_copies > 1 && !far_offset, fail.
3975 * Else find LCM(raid_disks, near_copy)*far_copies and
3976 * multiply by chunk_size. Then round to this number.
3977 * This is mostly done by raid10_size()
3979 struct r10conf *conf = mddev->private;
3980 sector_t oldsize, size;
3982 if (mddev->reshape_position != MaxSector)
3985 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3988 oldsize = raid10_size(mddev, 0, 0);
3989 size = raid10_size(mddev, sectors, 0);
3990 if (mddev->external_size &&
3991 mddev->array_sectors > size)
3993 if (mddev->bitmap) {
3994 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3998 md_set_array_sectors(mddev, size);
3999 if (sectors > mddev->dev_sectors &&
4000 mddev->recovery_cp > oldsize) {
4001 mddev->recovery_cp = oldsize;
4002 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4004 calc_sectors(conf, sectors);
4005 mddev->dev_sectors = conf->dev_sectors;
4006 mddev->resync_max_sectors = size;
4010 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4012 struct md_rdev *rdev;
4013 struct r10conf *conf;
4015 if (mddev->degraded > 0) {
4016 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4018 return ERR_PTR(-EINVAL);
4020 sector_div(size, devs);
4022 /* Set new parameters */
4023 mddev->new_level = 10;
4024 /* new layout: far_copies = 1, near_copies = 2 */
4025 mddev->new_layout = (1<<8) + 2;
4026 mddev->new_chunk_sectors = mddev->chunk_sectors;
4027 mddev->delta_disks = mddev->raid_disks;
4028 mddev->raid_disks *= 2;
4029 /* make sure it will be not marked as dirty */
4030 mddev->recovery_cp = MaxSector;
4031 mddev->dev_sectors = size;
4033 conf = setup_conf(mddev);
4034 if (!IS_ERR(conf)) {
4035 rdev_for_each(rdev, mddev)
4036 if (rdev->raid_disk >= 0) {
4037 rdev->new_raid_disk = rdev->raid_disk * 2;
4038 rdev->sectors = size;
4046 static void *raid10_takeover(struct mddev *mddev)
4048 struct r0conf *raid0_conf;
4050 /* raid10 can take over:
4051 * raid0 - providing it has only two drives
4053 if (mddev->level == 0) {
4054 /* for raid0 takeover only one zone is supported */
4055 raid0_conf = mddev->private;
4056 if (raid0_conf->nr_strip_zones > 1) {
4057 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4059 return ERR_PTR(-EINVAL);
4061 return raid10_takeover_raid0(mddev,
4062 raid0_conf->strip_zone->zone_end,
4063 raid0_conf->strip_zone->nb_dev);
4065 return ERR_PTR(-EINVAL);
4068 static int raid10_check_reshape(struct mddev *mddev)
4070 /* Called when there is a request to change
4071 * - layout (to ->new_layout)
4072 * - chunk size (to ->new_chunk_sectors)
4073 * - raid_disks (by delta_disks)
4074 * or when trying to restart a reshape that was ongoing.
4076 * We need to validate the request and possibly allocate
4077 * space if that might be an issue later.
4079 * Currently we reject any reshape of a 'far' mode array,
4080 * allow chunk size to change if new is generally acceptable,
4081 * allow raid_disks to increase, and allow
4082 * a switch between 'near' mode and 'offset' mode.
4084 struct r10conf *conf = mddev->private;
4087 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4090 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4091 /* mustn't change number of copies */
4093 if (geo.far_copies > 1 && !geo.far_offset)
4094 /* Cannot switch to 'far' mode */
4097 if (mddev->array_sectors & geo.chunk_mask)
4098 /* not factor of array size */
4101 if (!enough(conf, -1))
4104 kfree(conf->mirrors_new);
4105 conf->mirrors_new = NULL;
4106 if (mddev->delta_disks > 0) {
4107 /* allocate new 'mirrors' list */
4109 kcalloc(mddev->raid_disks + mddev->delta_disks,
4110 sizeof(struct raid10_info),
4112 if (!conf->mirrors_new)
4119 * Need to check if array has failed when deciding whether to:
4121 * - remove non-faulty devices
4124 * This determination is simple when no reshape is happening.
4125 * However if there is a reshape, we need to carefully check
4126 * both the before and after sections.
4127 * This is because some failed devices may only affect one
4128 * of the two sections, and some non-in_sync devices may
4129 * be insync in the section most affected by failed devices.
4131 static int calc_degraded(struct r10conf *conf)
4133 int degraded, degraded2;
4138 /* 'prev' section first */
4139 for (i = 0; i < conf->prev.raid_disks; i++) {
4140 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4141 if (!rdev || test_bit(Faulty, &rdev->flags))
4143 else if (!test_bit(In_sync, &rdev->flags))
4144 /* When we can reduce the number of devices in
4145 * an array, this might not contribute to
4146 * 'degraded'. It does now.
4151 if (conf->geo.raid_disks == conf->prev.raid_disks)
4155 for (i = 0; i < conf->geo.raid_disks; i++) {
4156 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4157 if (!rdev || test_bit(Faulty, &rdev->flags))
4159 else if (!test_bit(In_sync, &rdev->flags)) {
4160 /* If reshape is increasing the number of devices,
4161 * this section has already been recovered, so
4162 * it doesn't contribute to degraded.
4165 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4170 if (degraded2 > degraded)
4175 static int raid10_start_reshape(struct mddev *mddev)
4177 /* A 'reshape' has been requested. This commits
4178 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4179 * This also checks if there are enough spares and adds them
4181 * We currently require enough spares to make the final
4182 * array non-degraded. We also require that the difference
4183 * between old and new data_offset - on each device - is
4184 * enough that we never risk over-writing.
4187 unsigned long before_length, after_length;
4188 sector_t min_offset_diff = 0;
4191 struct r10conf *conf = mddev->private;
4192 struct md_rdev *rdev;
4196 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4199 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4202 before_length = ((1 << conf->prev.chunk_shift) *
4203 conf->prev.far_copies);
4204 after_length = ((1 << conf->geo.chunk_shift) *
4205 conf->geo.far_copies);
4207 rdev_for_each(rdev, mddev) {
4208 if (!test_bit(In_sync, &rdev->flags)
4209 && !test_bit(Faulty, &rdev->flags))
4211 if (rdev->raid_disk >= 0) {
4212 long long diff = (rdev->new_data_offset
4213 - rdev->data_offset);
4214 if (!mddev->reshape_backwards)
4218 if (first || diff < min_offset_diff)
4219 min_offset_diff = diff;
4224 if (max(before_length, after_length) > min_offset_diff)
4227 if (spares < mddev->delta_disks)
4230 conf->offset_diff = min_offset_diff;
4231 spin_lock_irq(&conf->device_lock);
4232 if (conf->mirrors_new) {
4233 memcpy(conf->mirrors_new, conf->mirrors,
4234 sizeof(struct raid10_info)*conf->prev.raid_disks);
4236 kfree(conf->mirrors_old);
4237 conf->mirrors_old = conf->mirrors;
4238 conf->mirrors = conf->mirrors_new;
4239 conf->mirrors_new = NULL;
4241 setup_geo(&conf->geo, mddev, geo_start);
4243 if (mddev->reshape_backwards) {
4244 sector_t size = raid10_size(mddev, 0, 0);
4245 if (size < mddev->array_sectors) {
4246 spin_unlock_irq(&conf->device_lock);
4247 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4251 mddev->resync_max_sectors = size;
4252 conf->reshape_progress = size;
4254 conf->reshape_progress = 0;
4255 conf->reshape_safe = conf->reshape_progress;
4256 spin_unlock_irq(&conf->device_lock);
4258 if (mddev->delta_disks && mddev->bitmap) {
4259 struct mdp_superblock_1 *sb = NULL;
4260 sector_t oldsize, newsize;
4262 oldsize = raid10_size(mddev, 0, 0);
4263 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4265 if (!mddev_is_clustered(mddev)) {
4266 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4273 rdev_for_each(rdev, mddev) {
4274 if (rdev->raid_disk > -1 &&
4275 !test_bit(Faulty, &rdev->flags))
4276 sb = page_address(rdev->sb_page);
4280 * some node is already performing reshape, and no need to
4281 * call md_bitmap_resize again since it should be called when
4282 * receiving BITMAP_RESIZE msg
4284 if ((sb && (le32_to_cpu(sb->feature_map) &
4285 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4288 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4292 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4294 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4299 if (mddev->delta_disks > 0) {
4300 rdev_for_each(rdev, mddev)
4301 if (rdev->raid_disk < 0 &&
4302 !test_bit(Faulty, &rdev->flags)) {
4303 if (raid10_add_disk(mddev, rdev) == 0) {
4304 if (rdev->raid_disk >=
4305 conf->prev.raid_disks)
4306 set_bit(In_sync, &rdev->flags);
4308 rdev->recovery_offset = 0;
4310 if (sysfs_link_rdev(mddev, rdev))
4311 /* Failure here is OK */;
4313 } else if (rdev->raid_disk >= conf->prev.raid_disks
4314 && !test_bit(Faulty, &rdev->flags)) {
4315 /* This is a spare that was manually added */
4316 set_bit(In_sync, &rdev->flags);
4319 /* When a reshape changes the number of devices,
4320 * ->degraded is measured against the larger of the
4321 * pre and post numbers.
4323 spin_lock_irq(&conf->device_lock);
4324 mddev->degraded = calc_degraded(conf);
4325 spin_unlock_irq(&conf->device_lock);
4326 mddev->raid_disks = conf->geo.raid_disks;
4327 mddev->reshape_position = conf->reshape_progress;
4328 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4330 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4331 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4332 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4333 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4334 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4336 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4338 if (!mddev->sync_thread) {
4342 conf->reshape_checkpoint = jiffies;
4343 md_wakeup_thread(mddev->sync_thread);
4344 md_new_event(mddev);
4348 mddev->recovery = 0;
4349 spin_lock_irq(&conf->device_lock);
4350 conf->geo = conf->prev;
4351 mddev->raid_disks = conf->geo.raid_disks;
4352 rdev_for_each(rdev, mddev)
4353 rdev->new_data_offset = rdev->data_offset;
4355 conf->reshape_progress = MaxSector;
4356 conf->reshape_safe = MaxSector;
4357 mddev->reshape_position = MaxSector;
4358 spin_unlock_irq(&conf->device_lock);
4362 /* Calculate the last device-address that could contain
4363 * any block from the chunk that includes the array-address 's'
4364 * and report the next address.
4365 * i.e. the address returned will be chunk-aligned and after
4366 * any data that is in the chunk containing 's'.
4368 static sector_t last_dev_address(sector_t s, struct geom *geo)
4370 s = (s | geo->chunk_mask) + 1;
4371 s >>= geo->chunk_shift;
4372 s *= geo->near_copies;
4373 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4374 s *= geo->far_copies;
4375 s <<= geo->chunk_shift;
4379 /* Calculate the first device-address that could contain
4380 * any block from the chunk that includes the array-address 's'.
4381 * This too will be the start of a chunk
4383 static sector_t first_dev_address(sector_t s, struct geom *geo)
4385 s >>= geo->chunk_shift;
4386 s *= geo->near_copies;
4387 sector_div(s, geo->raid_disks);
4388 s *= geo->far_copies;
4389 s <<= geo->chunk_shift;
4393 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4396 /* We simply copy at most one chunk (smallest of old and new)
4397 * at a time, possibly less if that exceeds RESYNC_PAGES,
4398 * or we hit a bad block or something.
4399 * This might mean we pause for normal IO in the middle of
4400 * a chunk, but that is not a problem as mddev->reshape_position
4401 * can record any location.
4403 * If we will want to write to a location that isn't
4404 * yet recorded as 'safe' (i.e. in metadata on disk) then
4405 * we need to flush all reshape requests and update the metadata.
4407 * When reshaping forwards (e.g. to more devices), we interpret
4408 * 'safe' as the earliest block which might not have been copied
4409 * down yet. We divide this by previous stripe size and multiply
4410 * by previous stripe length to get lowest device offset that we
4411 * cannot write to yet.
4412 * We interpret 'sector_nr' as an address that we want to write to.
4413 * From this we use last_device_address() to find where we might
4414 * write to, and first_device_address on the 'safe' position.
4415 * If this 'next' write position is after the 'safe' position,
4416 * we must update the metadata to increase the 'safe' position.
4418 * When reshaping backwards, we round in the opposite direction
4419 * and perform the reverse test: next write position must not be
4420 * less than current safe position.
4422 * In all this the minimum difference in data offsets
4423 * (conf->offset_diff - always positive) allows a bit of slack,
4424 * so next can be after 'safe', but not by more than offset_diff
4426 * We need to prepare all the bios here before we start any IO
4427 * to ensure the size we choose is acceptable to all devices.
4428 * The means one for each copy for write-out and an extra one for
4430 * We store the read-in bio in ->master_bio and the others in
4431 * ->devs[x].bio and ->devs[x].repl_bio.
4433 struct r10conf *conf = mddev->private;
4434 struct r10bio *r10_bio;
4435 sector_t next, safe, last;
4439 struct md_rdev *rdev;
4442 struct bio *bio, *read_bio;
4443 int sectors_done = 0;
4444 struct page **pages;
4446 if (sector_nr == 0) {
4447 /* If restarting in the middle, skip the initial sectors */
4448 if (mddev->reshape_backwards &&
4449 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4450 sector_nr = (raid10_size(mddev, 0, 0)
4451 - conf->reshape_progress);
4452 } else if (!mddev->reshape_backwards &&
4453 conf->reshape_progress > 0)
4454 sector_nr = conf->reshape_progress;
4456 mddev->curr_resync_completed = sector_nr;
4457 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4463 /* We don't use sector_nr to track where we are up to
4464 * as that doesn't work well for ->reshape_backwards.
4465 * So just use ->reshape_progress.
4467 if (mddev->reshape_backwards) {
4468 /* 'next' is the earliest device address that we might
4469 * write to for this chunk in the new layout
4471 next = first_dev_address(conf->reshape_progress - 1,
4474 /* 'safe' is the last device address that we might read from
4475 * in the old layout after a restart
4477 safe = last_dev_address(conf->reshape_safe - 1,
4480 if (next + conf->offset_diff < safe)
4483 last = conf->reshape_progress - 1;
4484 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4485 & conf->prev.chunk_mask);
4486 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4487 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4489 /* 'next' is after the last device address that we
4490 * might write to for this chunk in the new layout
4492 next = last_dev_address(conf->reshape_progress, &conf->geo);
4494 /* 'safe' is the earliest device address that we might
4495 * read from in the old layout after a restart
4497 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4499 /* Need to update metadata if 'next' might be beyond 'safe'
4500 * as that would possibly corrupt data
4502 if (next > safe + conf->offset_diff)
4505 sector_nr = conf->reshape_progress;
4506 last = sector_nr | (conf->geo.chunk_mask
4507 & conf->prev.chunk_mask);
4509 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4510 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4514 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4515 /* Need to update reshape_position in metadata */
4517 mddev->reshape_position = conf->reshape_progress;
4518 if (mddev->reshape_backwards)
4519 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4520 - conf->reshape_progress;
4522 mddev->curr_resync_completed = conf->reshape_progress;
4523 conf->reshape_checkpoint = jiffies;
4524 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4525 md_wakeup_thread(mddev->thread);
4526 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4527 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4528 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4529 allow_barrier(conf);
4530 return sectors_done;
4532 conf->reshape_safe = mddev->reshape_position;
4533 allow_barrier(conf);
4536 raise_barrier(conf, 0);
4538 /* Now schedule reads for blocks from sector_nr to last */
4539 r10_bio = raid10_alloc_init_r10buf(conf);
4541 raise_barrier(conf, 1);
4542 atomic_set(&r10_bio->remaining, 0);
4543 r10_bio->mddev = mddev;
4544 r10_bio->sector = sector_nr;
4545 set_bit(R10BIO_IsReshape, &r10_bio->state);
4546 r10_bio->sectors = last - sector_nr + 1;
4547 rdev = read_balance(conf, r10_bio, &max_sectors);
4548 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4551 /* Cannot read from here, so need to record bad blocks
4552 * on all the target devices.
4555 mempool_free(r10_bio, &conf->r10buf_pool);
4556 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4557 return sectors_done;
4560 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4562 bio_set_dev(read_bio, rdev->bdev);
4563 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4564 + rdev->data_offset);
4565 read_bio->bi_private = r10_bio;
4566 read_bio->bi_end_io = end_reshape_read;
4567 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4568 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4569 read_bio->bi_status = 0;
4570 read_bio->bi_vcnt = 0;
4571 read_bio->bi_iter.bi_size = 0;
4572 r10_bio->master_bio = read_bio;
4573 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4576 * Broadcast RESYNC message to other nodes, so all nodes would not
4577 * write to the region to avoid conflict.
4579 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4580 struct mdp_superblock_1 *sb = NULL;
4581 int sb_reshape_pos = 0;
4583 conf->cluster_sync_low = sector_nr;
4584 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4585 sb = page_address(rdev->sb_page);
4587 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4589 * Set cluster_sync_low again if next address for array
4590 * reshape is less than cluster_sync_low. Since we can't
4591 * update cluster_sync_low until it has finished reshape.
4593 if (sb_reshape_pos < conf->cluster_sync_low)
4594 conf->cluster_sync_low = sb_reshape_pos;
4597 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4598 conf->cluster_sync_high);
4601 /* Now find the locations in the new layout */
4602 __raid10_find_phys(&conf->geo, r10_bio);
4605 read_bio->bi_next = NULL;
4608 for (s = 0; s < conf->copies*2; s++) {
4610 int d = r10_bio->devs[s/2].devnum;
4611 struct md_rdev *rdev2;
4613 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4614 b = r10_bio->devs[s/2].repl_bio;
4616 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4617 b = r10_bio->devs[s/2].bio;
4619 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4622 bio_set_dev(b, rdev2->bdev);
4623 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4624 rdev2->new_data_offset;
4625 b->bi_end_io = end_reshape_write;
4626 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4631 /* Now add as many pages as possible to all of these bios. */
4634 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4635 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4636 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4637 int len = (max_sectors - s) << 9;
4638 if (len > PAGE_SIZE)
4640 for (bio = blist; bio ; bio = bio->bi_next) {
4642 * won't fail because the vec table is big enough
4643 * to hold all these pages
4645 bio_add_page(bio, page, len, 0);
4647 sector_nr += len >> 9;
4648 nr_sectors += len >> 9;
4651 r10_bio->sectors = nr_sectors;
4653 /* Now submit the read */
4654 md_sync_acct_bio(read_bio, r10_bio->sectors);
4655 atomic_inc(&r10_bio->remaining);
4656 read_bio->bi_next = NULL;
4657 generic_make_request(read_bio);
4658 sectors_done += nr_sectors;
4659 if (sector_nr <= last)
4662 lower_barrier(conf);
4664 /* Now that we have done the whole section we can
4665 * update reshape_progress
4667 if (mddev->reshape_backwards)
4668 conf->reshape_progress -= sectors_done;
4670 conf->reshape_progress += sectors_done;
4672 return sectors_done;
4675 static void end_reshape_request(struct r10bio *r10_bio);
4676 static int handle_reshape_read_error(struct mddev *mddev,
4677 struct r10bio *r10_bio);
4678 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4680 /* Reshape read completed. Hopefully we have a block
4682 * If we got a read error then we do sync 1-page reads from
4683 * elsewhere until we find the data - or give up.
4685 struct r10conf *conf = mddev->private;
4688 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4689 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4690 /* Reshape has been aborted */
4691 md_done_sync(mddev, r10_bio->sectors, 0);
4695 /* We definitely have the data in the pages, schedule the
4698 atomic_set(&r10_bio->remaining, 1);
4699 for (s = 0; s < conf->copies*2; s++) {
4701 int d = r10_bio->devs[s/2].devnum;
4702 struct md_rdev *rdev;
4705 rdev = rcu_dereference(conf->mirrors[d].replacement);
4706 b = r10_bio->devs[s/2].repl_bio;
4708 rdev = rcu_dereference(conf->mirrors[d].rdev);
4709 b = r10_bio->devs[s/2].bio;
4711 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4715 atomic_inc(&rdev->nr_pending);
4717 md_sync_acct_bio(b, r10_bio->sectors);
4718 atomic_inc(&r10_bio->remaining);
4720 generic_make_request(b);
4722 end_reshape_request(r10_bio);
4725 static void end_reshape(struct r10conf *conf)
4727 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4730 spin_lock_irq(&conf->device_lock);
4731 conf->prev = conf->geo;
4732 md_finish_reshape(conf->mddev);
4734 conf->reshape_progress = MaxSector;
4735 conf->reshape_safe = MaxSector;
4736 spin_unlock_irq(&conf->device_lock);
4738 /* read-ahead size must cover two whole stripes, which is
4739 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4741 if (conf->mddev->queue) {
4742 int stripe = conf->geo.raid_disks *
4743 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4744 stripe /= conf->geo.near_copies;
4745 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4746 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4751 static void raid10_update_reshape_pos(struct mddev *mddev)
4753 struct r10conf *conf = mddev->private;
4756 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4757 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4758 || mddev->reshape_position == MaxSector)
4759 conf->reshape_progress = mddev->reshape_position;
4764 static int handle_reshape_read_error(struct mddev *mddev,
4765 struct r10bio *r10_bio)
4767 /* Use sync reads to get the blocks from somewhere else */
4768 int sectors = r10_bio->sectors;
4769 struct r10conf *conf = mddev->private;
4770 struct r10bio *r10b;
4773 struct page **pages;
4775 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4777 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4781 /* reshape IOs share pages from .devs[0].bio */
4782 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4784 r10b->sector = r10_bio->sector;
4785 __raid10_find_phys(&conf->prev, r10b);
4790 int first_slot = slot;
4792 if (s > (PAGE_SIZE >> 9))
4797 int d = r10b->devs[slot].devnum;
4798 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4801 test_bit(Faulty, &rdev->flags) ||
4802 !test_bit(In_sync, &rdev->flags))
4805 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4806 atomic_inc(&rdev->nr_pending);
4808 success = sync_page_io(rdev,
4812 REQ_OP_READ, 0, false);
4813 rdev_dec_pending(rdev, mddev);
4819 if (slot >= conf->copies)
4821 if (slot == first_slot)
4826 /* couldn't read this block, must give up */
4827 set_bit(MD_RECOVERY_INTR,
4839 static void end_reshape_write(struct bio *bio)
4841 struct r10bio *r10_bio = get_resync_r10bio(bio);
4842 struct mddev *mddev = r10_bio->mddev;
4843 struct r10conf *conf = mddev->private;
4847 struct md_rdev *rdev = NULL;
4849 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4851 rdev = conf->mirrors[d].replacement;
4854 rdev = conf->mirrors[d].rdev;
4857 if (bio->bi_status) {
4858 /* FIXME should record badblock */
4859 md_error(mddev, rdev);
4862 rdev_dec_pending(rdev, mddev);
4863 end_reshape_request(r10_bio);
4866 static void end_reshape_request(struct r10bio *r10_bio)
4868 if (!atomic_dec_and_test(&r10_bio->remaining))
4870 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4871 bio_put(r10_bio->master_bio);
4875 static void raid10_finish_reshape(struct mddev *mddev)
4877 struct r10conf *conf = mddev->private;
4879 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4882 if (mddev->delta_disks > 0) {
4883 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4884 mddev->recovery_cp = mddev->resync_max_sectors;
4885 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4887 mddev->resync_max_sectors = mddev->array_sectors;
4891 for (d = conf->geo.raid_disks ;
4892 d < conf->geo.raid_disks - mddev->delta_disks;
4894 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4896 clear_bit(In_sync, &rdev->flags);
4897 rdev = rcu_dereference(conf->mirrors[d].replacement);
4899 clear_bit(In_sync, &rdev->flags);
4903 mddev->layout = mddev->new_layout;
4904 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4905 mddev->reshape_position = MaxSector;
4906 mddev->delta_disks = 0;
4907 mddev->reshape_backwards = 0;
4910 static struct md_personality raid10_personality =
4914 .owner = THIS_MODULE,
4915 .make_request = raid10_make_request,
4917 .free = raid10_free,
4918 .status = raid10_status,
4919 .error_handler = raid10_error,
4920 .hot_add_disk = raid10_add_disk,
4921 .hot_remove_disk= raid10_remove_disk,
4922 .spare_active = raid10_spare_active,
4923 .sync_request = raid10_sync_request,
4924 .quiesce = raid10_quiesce,
4925 .size = raid10_size,
4926 .resize = raid10_resize,
4927 .takeover = raid10_takeover,
4928 .check_reshape = raid10_check_reshape,
4929 .start_reshape = raid10_start_reshape,
4930 .finish_reshape = raid10_finish_reshape,
4931 .update_reshape_pos = raid10_update_reshape_pos,
4932 .congested = raid10_congested,
4935 static int __init raid_init(void)
4937 return register_md_personality(&raid10_personality);
4940 static void raid_exit(void)
4942 unregister_md_personality(&raid10_personality);
4945 module_init(raid_init);
4946 module_exit(raid_exit);
4947 MODULE_LICENSE("GPL");
4948 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4949 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4950 MODULE_ALIAS("md-raid10");
4951 MODULE_ALIAS("md-level-10");
4953 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);