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.
474 if (!test_bit(Faulty, &rdev->flags))
475 set_bit(R10BIO_WriteError, &r10_bio->state);
477 /* Fail the request */
478 set_bit(R10BIO_Degraded, &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 void flush_pending_writes(struct r10conf *conf)
853 /* Any writes that have been queued but are awaiting
854 * bitmap updates get flushed here.
856 spin_lock_irq(&conf->device_lock);
858 if (conf->pending_bio_list.head) {
859 struct blk_plug plug;
862 bio = bio_list_get(&conf->pending_bio_list);
863 conf->pending_count = 0;
864 spin_unlock_irq(&conf->device_lock);
867 * As this is called in a wait_event() loop (see freeze_array),
868 * current->state might be TASK_UNINTERRUPTIBLE which will
869 * cause a warning when we prepare to wait again. As it is
870 * rare that this path is taken, it is perfectly safe to force
871 * us to go around the wait_event() loop again, so the warning
872 * is a false-positive. Silence the warning by resetting
875 __set_current_state(TASK_RUNNING);
877 blk_start_plug(&plug);
878 /* flush any pending bitmap writes to disk
879 * before proceeding w/ I/O */
880 md_bitmap_unplug(conf->mddev->bitmap);
881 wake_up(&conf->wait_barrier);
883 while (bio) { /* submit pending writes */
884 struct bio *next = bio->bi_next;
885 struct md_rdev *rdev = (void*)bio->bi_disk;
887 bio_set_dev(bio, rdev->bdev);
888 if (test_bit(Faulty, &rdev->flags)) {
890 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
891 !blk_queue_discard(bio->bi_disk->queue)))
895 submit_bio_noacct(bio);
898 blk_finish_plug(&plug);
900 spin_unlock_irq(&conf->device_lock);
904 * Sometimes we need to suspend IO while we do something else,
905 * either some resync/recovery, or reconfigure the array.
906 * To do this we raise a 'barrier'.
907 * The 'barrier' is a counter that can be raised multiple times
908 * to count how many activities are happening which preclude
910 * We can only raise the barrier if there is no pending IO.
911 * i.e. if nr_pending == 0.
912 * We choose only to raise the barrier if no-one is waiting for the
913 * barrier to go down. This means that as soon as an IO request
914 * is ready, no other operations which require a barrier will start
915 * until the IO request has had a chance.
917 * So: regular IO calls 'wait_barrier'. When that returns there
918 * is no backgroup IO happening, It must arrange to call
919 * allow_barrier when it has finished its IO.
920 * backgroup IO calls must call raise_barrier. Once that returns
921 * there is no normal IO happeing. It must arrange to call
922 * lower_barrier when the particular background IO completes.
925 static void raise_barrier(struct r10conf *conf, int force)
927 BUG_ON(force && !conf->barrier);
928 spin_lock_irq(&conf->resync_lock);
930 /* Wait until no block IO is waiting (unless 'force') */
931 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
934 /* block any new IO from starting */
937 /* Now wait for all pending IO to complete */
938 wait_event_lock_irq(conf->wait_barrier,
939 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
942 spin_unlock_irq(&conf->resync_lock);
945 static void lower_barrier(struct r10conf *conf)
948 spin_lock_irqsave(&conf->resync_lock, flags);
950 spin_unlock_irqrestore(&conf->resync_lock, flags);
951 wake_up(&conf->wait_barrier);
954 static void wait_barrier(struct r10conf *conf)
956 spin_lock_irq(&conf->resync_lock);
958 struct bio_list *bio_list = current->bio_list;
960 /* Wait for the barrier to drop.
961 * However if there are already pending
962 * requests (preventing the barrier from
963 * rising completely), and the
964 * pre-process bio queue isn't empty,
965 * then don't wait, as we need to empty
966 * that queue to get the nr_pending
969 raid10_log(conf->mddev, "wait barrier");
970 wait_event_lock_irq(conf->wait_barrier,
972 (atomic_read(&conf->nr_pending) &&
974 (!bio_list_empty(&bio_list[0]) ||
975 !bio_list_empty(&bio_list[1]))) ||
976 /* move on if recovery thread is
979 (conf->mddev->thread->tsk == current &&
980 test_bit(MD_RECOVERY_RUNNING,
981 &conf->mddev->recovery) &&
982 conf->nr_queued > 0),
985 if (!conf->nr_waiting)
986 wake_up(&conf->wait_barrier);
988 atomic_inc(&conf->nr_pending);
989 spin_unlock_irq(&conf->resync_lock);
992 static void allow_barrier(struct r10conf *conf)
994 if ((atomic_dec_and_test(&conf->nr_pending)) ||
995 (conf->array_freeze_pending))
996 wake_up(&conf->wait_barrier);
999 static void freeze_array(struct r10conf *conf, int extra)
1001 /* stop syncio and normal IO and wait for everything to
1003 * We increment barrier and nr_waiting, and then
1004 * wait until nr_pending match nr_queued+extra
1005 * This is called in the context of one normal IO request
1006 * that has failed. Thus any sync request that might be pending
1007 * will be blocked by nr_pending, and we need to wait for
1008 * pending IO requests to complete or be queued for re-try.
1009 * Thus the number queued (nr_queued) plus this request (extra)
1010 * must match the number of pending IOs (nr_pending) before
1013 spin_lock_irq(&conf->resync_lock);
1014 conf->array_freeze_pending++;
1017 wait_event_lock_irq_cmd(conf->wait_barrier,
1018 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1020 flush_pending_writes(conf));
1022 conf->array_freeze_pending--;
1023 spin_unlock_irq(&conf->resync_lock);
1026 static void unfreeze_array(struct r10conf *conf)
1028 /* reverse the effect of the freeze */
1029 spin_lock_irq(&conf->resync_lock);
1032 wake_up(&conf->wait_barrier);
1033 spin_unlock_irq(&conf->resync_lock);
1036 static sector_t choose_data_offset(struct r10bio *r10_bio,
1037 struct md_rdev *rdev)
1039 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1040 test_bit(R10BIO_Previous, &r10_bio->state))
1041 return rdev->data_offset;
1043 return rdev->new_data_offset;
1046 struct raid10_plug_cb {
1047 struct blk_plug_cb cb;
1048 struct bio_list pending;
1052 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1054 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1056 struct mddev *mddev = plug->cb.data;
1057 struct r10conf *conf = mddev->private;
1060 if (from_schedule || current->bio_list) {
1061 spin_lock_irq(&conf->device_lock);
1062 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1063 conf->pending_count += plug->pending_cnt;
1064 spin_unlock_irq(&conf->device_lock);
1065 wake_up(&conf->wait_barrier);
1066 md_wakeup_thread(mddev->thread);
1071 /* we aren't scheduling, so we can do the write-out directly. */
1072 bio = bio_list_get(&plug->pending);
1073 md_bitmap_unplug(mddev->bitmap);
1074 wake_up(&conf->wait_barrier);
1076 while (bio) { /* submit pending writes */
1077 struct bio *next = bio->bi_next;
1078 struct md_rdev *rdev = (void*)bio->bi_disk;
1079 bio->bi_next = NULL;
1080 bio_set_dev(bio, rdev->bdev);
1081 if (test_bit(Faulty, &rdev->flags)) {
1083 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1084 !blk_queue_discard(bio->bi_disk->queue)))
1085 /* Just ignore it */
1088 submit_bio_noacct(bio);
1095 * 1. Register the new request and wait if the reconstruction thread has put
1096 * up a bar for new requests. Continue immediately if no resync is active
1098 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1100 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1101 struct bio *bio, sector_t sectors)
1104 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1105 bio->bi_iter.bi_sector < conf->reshape_progress &&
1106 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1107 raid10_log(conf->mddev, "wait reshape");
1108 allow_barrier(conf);
1109 wait_event(conf->wait_barrier,
1110 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1111 conf->reshape_progress >= bio->bi_iter.bi_sector +
1117 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1118 struct r10bio *r10_bio)
1120 struct r10conf *conf = mddev->private;
1121 struct bio *read_bio;
1122 const int op = bio_op(bio);
1123 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1125 struct md_rdev *rdev;
1126 char b[BDEVNAME_SIZE];
1127 int slot = r10_bio->read_slot;
1128 struct md_rdev *err_rdev = NULL;
1129 gfp_t gfp = GFP_NOIO;
1131 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1133 * This is an error retry, but we cannot
1134 * safely dereference the rdev in the r10_bio,
1135 * we must use the one in conf.
1136 * If it has already been disconnected (unlikely)
1137 * we lose the device name in error messages.
1141 * As we are blocking raid10, it is a little safer to
1144 gfp = GFP_NOIO | __GFP_HIGH;
1147 disk = r10_bio->devs[slot].devnum;
1148 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1150 bdevname(err_rdev->bdev, b);
1153 /* This never gets dereferenced */
1154 err_rdev = r10_bio->devs[slot].rdev;
1159 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1160 rdev = read_balance(conf, r10_bio, &max_sectors);
1163 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1165 (unsigned long long)r10_bio->sector);
1167 raid_end_bio_io(r10_bio);
1171 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1173 bdevname(rdev->bdev, b),
1174 (unsigned long long)r10_bio->sector);
1175 if (max_sectors < bio_sectors(bio)) {
1176 struct bio *split = bio_split(bio, max_sectors,
1177 gfp, &conf->bio_split);
1178 bio_chain(split, bio);
1179 allow_barrier(conf);
1180 submit_bio_noacct(bio);
1183 r10_bio->master_bio = bio;
1184 r10_bio->sectors = max_sectors;
1186 slot = r10_bio->read_slot;
1188 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1190 r10_bio->devs[slot].bio = read_bio;
1191 r10_bio->devs[slot].rdev = rdev;
1193 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1194 choose_data_offset(r10_bio, rdev);
1195 bio_set_dev(read_bio, rdev->bdev);
1196 read_bio->bi_end_io = raid10_end_read_request;
1197 bio_set_op_attrs(read_bio, op, do_sync);
1198 if (test_bit(FailFast, &rdev->flags) &&
1199 test_bit(R10BIO_FailFast, &r10_bio->state))
1200 read_bio->bi_opf |= MD_FAILFAST;
1201 read_bio->bi_private = r10_bio;
1204 trace_block_bio_remap(read_bio->bi_disk->queue,
1205 read_bio, disk_devt(mddev->gendisk),
1207 submit_bio_noacct(read_bio);
1211 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1212 struct bio *bio, bool replacement,
1215 const int op = bio_op(bio);
1216 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1217 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1218 unsigned long flags;
1219 struct blk_plug_cb *cb;
1220 struct raid10_plug_cb *plug = NULL;
1221 struct r10conf *conf = mddev->private;
1222 struct md_rdev *rdev;
1223 int devnum = r10_bio->devs[n_copy].devnum;
1227 rdev = conf->mirrors[devnum].replacement;
1229 /* Replacement just got moved to main 'rdev' */
1231 rdev = conf->mirrors[devnum].rdev;
1234 rdev = conf->mirrors[devnum].rdev;
1236 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1238 r10_bio->devs[n_copy].repl_bio = mbio;
1240 r10_bio->devs[n_copy].bio = mbio;
1242 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1243 choose_data_offset(r10_bio, rdev));
1244 bio_set_dev(mbio, rdev->bdev);
1245 mbio->bi_end_io = raid10_end_write_request;
1246 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1247 if (!replacement && test_bit(FailFast,
1248 &conf->mirrors[devnum].rdev->flags)
1249 && enough(conf, devnum))
1250 mbio->bi_opf |= MD_FAILFAST;
1251 mbio->bi_private = r10_bio;
1253 if (conf->mddev->gendisk)
1254 trace_block_bio_remap(mbio->bi_disk->queue,
1255 mbio, disk_devt(conf->mddev->gendisk),
1257 /* flush_pending_writes() needs access to the rdev so...*/
1258 mbio->bi_disk = (void *)rdev;
1260 atomic_inc(&r10_bio->remaining);
1262 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1264 plug = container_of(cb, struct raid10_plug_cb, cb);
1268 bio_list_add(&plug->pending, mbio);
1269 plug->pending_cnt++;
1271 spin_lock_irqsave(&conf->device_lock, flags);
1272 bio_list_add(&conf->pending_bio_list, mbio);
1273 conf->pending_count++;
1274 spin_unlock_irqrestore(&conf->device_lock, flags);
1275 md_wakeup_thread(mddev->thread);
1279 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1280 struct r10bio *r10_bio)
1282 struct r10conf *conf = mddev->private;
1284 struct md_rdev *blocked_rdev;
1288 if ((mddev_is_clustered(mddev) &&
1289 md_cluster_ops->area_resyncing(mddev, WRITE,
1290 bio->bi_iter.bi_sector,
1291 bio_end_sector(bio)))) {
1294 prepare_to_wait(&conf->wait_barrier,
1296 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1297 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1301 finish_wait(&conf->wait_barrier, &w);
1304 sectors = r10_bio->sectors;
1305 regular_request_wait(mddev, conf, bio, sectors);
1306 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1307 (mddev->reshape_backwards
1308 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1309 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1310 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1311 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1312 /* Need to update reshape_position in metadata */
1313 mddev->reshape_position = conf->reshape_progress;
1314 set_mask_bits(&mddev->sb_flags, 0,
1315 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1316 md_wakeup_thread(mddev->thread);
1317 raid10_log(conf->mddev, "wait reshape metadata");
1318 wait_event(mddev->sb_wait,
1319 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1321 conf->reshape_safe = mddev->reshape_position;
1324 if (conf->pending_count >= max_queued_requests) {
1325 md_wakeup_thread(mddev->thread);
1326 raid10_log(mddev, "wait queued");
1327 wait_event(conf->wait_barrier,
1328 conf->pending_count < max_queued_requests);
1330 /* first select target devices under rcu_lock and
1331 * inc refcount on their rdev. Record them by setting
1333 * If there are known/acknowledged bad blocks on any device
1334 * on which we have seen a write error, we want to avoid
1335 * writing to those blocks. This potentially requires several
1336 * writes to write around the bad blocks. Each set of writes
1337 * gets its own r10_bio with a set of bios attached.
1340 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1341 raid10_find_phys(conf, r10_bio);
1343 blocked_rdev = NULL;
1345 max_sectors = r10_bio->sectors;
1347 for (i = 0; i < conf->copies; i++) {
1348 int d = r10_bio->devs[i].devnum;
1349 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1350 struct md_rdev *rrdev = rcu_dereference(
1351 conf->mirrors[d].replacement);
1354 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1355 atomic_inc(&rdev->nr_pending);
1356 blocked_rdev = rdev;
1359 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1360 atomic_inc(&rrdev->nr_pending);
1361 blocked_rdev = rrdev;
1364 if (rdev && (test_bit(Faulty, &rdev->flags)))
1366 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1369 r10_bio->devs[i].bio = NULL;
1370 r10_bio->devs[i].repl_bio = NULL;
1372 if (!rdev && !rrdev) {
1373 set_bit(R10BIO_Degraded, &r10_bio->state);
1376 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1378 sector_t dev_sector = r10_bio->devs[i].addr;
1382 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1383 &first_bad, &bad_sectors);
1385 /* Mustn't write here until the bad block
1388 atomic_inc(&rdev->nr_pending);
1389 set_bit(BlockedBadBlocks, &rdev->flags);
1390 blocked_rdev = rdev;
1393 if (is_bad && first_bad <= dev_sector) {
1394 /* Cannot write here at all */
1395 bad_sectors -= (dev_sector - first_bad);
1396 if (bad_sectors < max_sectors)
1397 /* Mustn't write more than bad_sectors
1398 * to other devices yet
1400 max_sectors = bad_sectors;
1401 /* We don't set R10BIO_Degraded as that
1402 * only applies if the disk is missing,
1403 * so it might be re-added, and we want to
1404 * know to recover this chunk.
1405 * In this case the device is here, and the
1406 * fact that this chunk is not in-sync is
1407 * recorded in the bad block log.
1412 int good_sectors = first_bad - dev_sector;
1413 if (good_sectors < max_sectors)
1414 max_sectors = good_sectors;
1418 r10_bio->devs[i].bio = bio;
1419 atomic_inc(&rdev->nr_pending);
1422 r10_bio->devs[i].repl_bio = bio;
1423 atomic_inc(&rrdev->nr_pending);
1428 if (unlikely(blocked_rdev)) {
1429 /* Have to wait for this device to get unblocked, then retry */
1433 for (j = 0; j < i; j++) {
1434 if (r10_bio->devs[j].bio) {
1435 d = r10_bio->devs[j].devnum;
1436 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1438 if (r10_bio->devs[j].repl_bio) {
1439 struct md_rdev *rdev;
1440 d = r10_bio->devs[j].devnum;
1441 rdev = conf->mirrors[d].replacement;
1443 /* Race with remove_disk */
1445 rdev = conf->mirrors[d].rdev;
1447 rdev_dec_pending(rdev, mddev);
1450 allow_barrier(conf);
1451 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1452 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1457 if (max_sectors < r10_bio->sectors)
1458 r10_bio->sectors = max_sectors;
1460 if (r10_bio->sectors < bio_sectors(bio)) {
1461 struct bio *split = bio_split(bio, r10_bio->sectors,
1462 GFP_NOIO, &conf->bio_split);
1463 bio_chain(split, bio);
1464 allow_barrier(conf);
1465 submit_bio_noacct(bio);
1468 r10_bio->master_bio = bio;
1471 atomic_set(&r10_bio->remaining, 1);
1472 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1474 for (i = 0; i < conf->copies; i++) {
1475 if (r10_bio->devs[i].bio)
1476 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1477 if (r10_bio->devs[i].repl_bio)
1478 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1480 one_write_done(r10_bio);
1483 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1485 struct r10conf *conf = mddev->private;
1486 struct r10bio *r10_bio;
1488 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1490 r10_bio->master_bio = bio;
1491 r10_bio->sectors = sectors;
1493 r10_bio->mddev = mddev;
1494 r10_bio->sector = bio->bi_iter.bi_sector;
1496 r10_bio->read_slot = -1;
1497 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1499 if (bio_data_dir(bio) == READ)
1500 raid10_read_request(mddev, bio, r10_bio);
1502 raid10_write_request(mddev, bio, r10_bio);
1505 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1507 struct r10conf *conf = mddev->private;
1508 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1509 int chunk_sects = chunk_mask + 1;
1510 int sectors = bio_sectors(bio);
1512 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1513 && md_flush_request(mddev, bio))
1516 if (!md_write_start(mddev, bio))
1520 * If this request crosses a chunk boundary, we need to split
1523 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1524 sectors > chunk_sects
1525 && (conf->geo.near_copies < conf->geo.raid_disks
1526 || conf->prev.near_copies <
1527 conf->prev.raid_disks)))
1528 sectors = chunk_sects -
1529 (bio->bi_iter.bi_sector &
1531 __make_request(mddev, bio, sectors);
1533 /* In case raid10d snuck in to freeze_array */
1534 wake_up(&conf->wait_barrier);
1538 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1540 struct r10conf *conf = mddev->private;
1543 if (conf->geo.near_copies < conf->geo.raid_disks)
1544 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1545 if (conf->geo.near_copies > 1)
1546 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1547 if (conf->geo.far_copies > 1) {
1548 if (conf->geo.far_offset)
1549 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1551 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1552 if (conf->geo.far_set_size != conf->geo.raid_disks)
1553 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1555 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1556 conf->geo.raid_disks - mddev->degraded);
1558 for (i = 0; i < conf->geo.raid_disks; i++) {
1559 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1560 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1563 seq_printf(seq, "]");
1566 /* check if there are enough drives for
1567 * every block to appear on atleast one.
1568 * Don't consider the device numbered 'ignore'
1569 * as we might be about to remove it.
1571 static int _enough(struct r10conf *conf, int previous, int ignore)
1577 disks = conf->prev.raid_disks;
1578 ncopies = conf->prev.near_copies;
1580 disks = conf->geo.raid_disks;
1581 ncopies = conf->geo.near_copies;
1586 int n = conf->copies;
1590 struct md_rdev *rdev;
1591 if (this != ignore &&
1592 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1593 test_bit(In_sync, &rdev->flags))
1595 this = (this+1) % disks;
1599 first = (first + ncopies) % disks;
1600 } while (first != 0);
1607 static int enough(struct r10conf *conf, int ignore)
1609 /* when calling 'enough', both 'prev' and 'geo' must
1611 * This is ensured if ->reconfig_mutex or ->device_lock
1614 return _enough(conf, 0, ignore) &&
1615 _enough(conf, 1, ignore);
1618 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1620 char b[BDEVNAME_SIZE];
1621 struct r10conf *conf = mddev->private;
1622 unsigned long flags;
1625 * If it is not operational, then we have already marked it as dead
1626 * else if it is the last working disks with "fail_last_dev == false",
1627 * ignore the error, let the next level up know.
1628 * else mark the drive as failed
1630 spin_lock_irqsave(&conf->device_lock, flags);
1631 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1632 && !enough(conf, rdev->raid_disk)) {
1634 * Don't fail the drive, just return an IO error.
1636 spin_unlock_irqrestore(&conf->device_lock, flags);
1639 if (test_and_clear_bit(In_sync, &rdev->flags))
1642 * If recovery is running, make sure it aborts.
1644 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1645 set_bit(Blocked, &rdev->flags);
1646 set_bit(Faulty, &rdev->flags);
1647 set_mask_bits(&mddev->sb_flags, 0,
1648 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1649 spin_unlock_irqrestore(&conf->device_lock, flags);
1650 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1651 "md/raid10:%s: Operation continuing on %d devices.\n",
1652 mdname(mddev), bdevname(rdev->bdev, b),
1653 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1656 static void print_conf(struct r10conf *conf)
1659 struct md_rdev *rdev;
1661 pr_debug("RAID10 conf printout:\n");
1663 pr_debug("(!conf)\n");
1666 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1667 conf->geo.raid_disks);
1669 /* This is only called with ->reconfix_mutex held, so
1670 * rcu protection of rdev is not needed */
1671 for (i = 0; i < conf->geo.raid_disks; i++) {
1672 char b[BDEVNAME_SIZE];
1673 rdev = conf->mirrors[i].rdev;
1675 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1676 i, !test_bit(In_sync, &rdev->flags),
1677 !test_bit(Faulty, &rdev->flags),
1678 bdevname(rdev->bdev,b));
1682 static void close_sync(struct r10conf *conf)
1685 allow_barrier(conf);
1687 mempool_exit(&conf->r10buf_pool);
1690 static int raid10_spare_active(struct mddev *mddev)
1693 struct r10conf *conf = mddev->private;
1694 struct raid10_info *tmp;
1696 unsigned long flags;
1699 * Find all non-in_sync disks within the RAID10 configuration
1700 * and mark them in_sync
1702 for (i = 0; i < conf->geo.raid_disks; i++) {
1703 tmp = conf->mirrors + i;
1704 if (tmp->replacement
1705 && tmp->replacement->recovery_offset == MaxSector
1706 && !test_bit(Faulty, &tmp->replacement->flags)
1707 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1708 /* Replacement has just become active */
1710 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1713 /* Replaced device not technically faulty,
1714 * but we need to be sure it gets removed
1715 * and never re-added.
1717 set_bit(Faulty, &tmp->rdev->flags);
1718 sysfs_notify_dirent_safe(
1719 tmp->rdev->sysfs_state);
1721 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1722 } else if (tmp->rdev
1723 && tmp->rdev->recovery_offset == MaxSector
1724 && !test_bit(Faulty, &tmp->rdev->flags)
1725 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1727 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1730 spin_lock_irqsave(&conf->device_lock, flags);
1731 mddev->degraded -= count;
1732 spin_unlock_irqrestore(&conf->device_lock, flags);
1738 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1740 struct r10conf *conf = mddev->private;
1744 int last = conf->geo.raid_disks - 1;
1746 if (mddev->recovery_cp < MaxSector)
1747 /* only hot-add to in-sync arrays, as recovery is
1748 * very different from resync
1751 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1754 if (md_integrity_add_rdev(rdev, mddev))
1757 if (rdev->raid_disk >= 0)
1758 first = last = rdev->raid_disk;
1760 if (rdev->saved_raid_disk >= first &&
1761 rdev->saved_raid_disk < conf->geo.raid_disks &&
1762 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1763 mirror = rdev->saved_raid_disk;
1766 for ( ; mirror <= last ; mirror++) {
1767 struct raid10_info *p = &conf->mirrors[mirror];
1768 if (p->recovery_disabled == mddev->recovery_disabled)
1771 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1772 p->replacement != NULL)
1774 clear_bit(In_sync, &rdev->flags);
1775 set_bit(Replacement, &rdev->flags);
1776 rdev->raid_disk = mirror;
1779 disk_stack_limits(mddev->gendisk, rdev->bdev,
1780 rdev->data_offset << 9);
1782 rcu_assign_pointer(p->replacement, rdev);
1787 disk_stack_limits(mddev->gendisk, rdev->bdev,
1788 rdev->data_offset << 9);
1790 p->head_position = 0;
1791 p->recovery_disabled = mddev->recovery_disabled - 1;
1792 rdev->raid_disk = mirror;
1794 if (rdev->saved_raid_disk != mirror)
1796 rcu_assign_pointer(p->rdev, rdev);
1799 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1800 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1806 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1808 struct r10conf *conf = mddev->private;
1810 int number = rdev->raid_disk;
1811 struct md_rdev **rdevp;
1812 struct raid10_info *p;
1815 if (unlikely(number >= mddev->raid_disks))
1817 p = conf->mirrors + number;
1818 if (rdev == p->rdev)
1820 else if (rdev == p->replacement)
1821 rdevp = &p->replacement;
1825 if (test_bit(In_sync, &rdev->flags) ||
1826 atomic_read(&rdev->nr_pending)) {
1830 /* Only remove non-faulty devices if recovery
1833 if (!test_bit(Faulty, &rdev->flags) &&
1834 mddev->recovery_disabled != p->recovery_disabled &&
1835 (!p->replacement || p->replacement == rdev) &&
1836 number < conf->geo.raid_disks &&
1842 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1844 if (atomic_read(&rdev->nr_pending)) {
1845 /* lost the race, try later */
1851 if (p->replacement) {
1852 /* We must have just cleared 'rdev' */
1853 p->rdev = p->replacement;
1854 clear_bit(Replacement, &p->replacement->flags);
1855 smp_mb(); /* Make sure other CPUs may see both as identical
1856 * but will never see neither -- if they are careful.
1858 p->replacement = NULL;
1861 clear_bit(WantReplacement, &rdev->flags);
1862 err = md_integrity_register(mddev);
1870 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1872 struct r10conf *conf = r10_bio->mddev->private;
1874 if (!bio->bi_status)
1875 set_bit(R10BIO_Uptodate, &r10_bio->state);
1877 /* The write handler will notice the lack of
1878 * R10BIO_Uptodate and record any errors etc
1880 atomic_add(r10_bio->sectors,
1881 &conf->mirrors[d].rdev->corrected_errors);
1883 /* for reconstruct, we always reschedule after a read.
1884 * for resync, only after all reads
1886 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1887 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1888 atomic_dec_and_test(&r10_bio->remaining)) {
1889 /* we have read all the blocks,
1890 * do the comparison in process context in raid10d
1892 reschedule_retry(r10_bio);
1896 static void end_sync_read(struct bio *bio)
1898 struct r10bio *r10_bio = get_resync_r10bio(bio);
1899 struct r10conf *conf = r10_bio->mddev->private;
1900 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1902 __end_sync_read(r10_bio, bio, d);
1905 static void end_reshape_read(struct bio *bio)
1907 /* reshape read bio isn't allocated from r10buf_pool */
1908 struct r10bio *r10_bio = bio->bi_private;
1910 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1913 static void end_sync_request(struct r10bio *r10_bio)
1915 struct mddev *mddev = r10_bio->mddev;
1917 while (atomic_dec_and_test(&r10_bio->remaining)) {
1918 if (r10_bio->master_bio == NULL) {
1919 /* the primary of several recovery bios */
1920 sector_t s = r10_bio->sectors;
1921 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1922 test_bit(R10BIO_WriteError, &r10_bio->state))
1923 reschedule_retry(r10_bio);
1926 md_done_sync(mddev, s, 1);
1929 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1930 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1931 test_bit(R10BIO_WriteError, &r10_bio->state))
1932 reschedule_retry(r10_bio);
1940 static void end_sync_write(struct bio *bio)
1942 struct r10bio *r10_bio = get_resync_r10bio(bio);
1943 struct mddev *mddev = r10_bio->mddev;
1944 struct r10conf *conf = mddev->private;
1950 struct md_rdev *rdev = NULL;
1952 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1954 rdev = conf->mirrors[d].replacement;
1956 rdev = conf->mirrors[d].rdev;
1958 if (bio->bi_status) {
1960 md_error(mddev, rdev);
1962 set_bit(WriteErrorSeen, &rdev->flags);
1963 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1964 set_bit(MD_RECOVERY_NEEDED,
1965 &rdev->mddev->recovery);
1966 set_bit(R10BIO_WriteError, &r10_bio->state);
1968 } else if (is_badblock(rdev,
1969 r10_bio->devs[slot].addr,
1971 &first_bad, &bad_sectors))
1972 set_bit(R10BIO_MadeGood, &r10_bio->state);
1974 rdev_dec_pending(rdev, mddev);
1976 end_sync_request(r10_bio);
1980 * Note: sync and recover and handled very differently for raid10
1981 * This code is for resync.
1982 * For resync, we read through virtual addresses and read all blocks.
1983 * If there is any error, we schedule a write. The lowest numbered
1984 * drive is authoritative.
1985 * However requests come for physical address, so we need to map.
1986 * For every physical address there are raid_disks/copies virtual addresses,
1987 * which is always are least one, but is not necessarly an integer.
1988 * This means that a physical address can span multiple chunks, so we may
1989 * have to submit multiple io requests for a single sync request.
1992 * We check if all blocks are in-sync and only write to blocks that
1995 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1997 struct r10conf *conf = mddev->private;
1999 struct bio *tbio, *fbio;
2001 struct page **tpages, **fpages;
2003 atomic_set(&r10_bio->remaining, 1);
2005 /* find the first device with a block */
2006 for (i=0; i<conf->copies; i++)
2007 if (!r10_bio->devs[i].bio->bi_status)
2010 if (i == conf->copies)
2014 fbio = r10_bio->devs[i].bio;
2015 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2016 fbio->bi_iter.bi_idx = 0;
2017 fpages = get_resync_pages(fbio)->pages;
2019 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2020 /* now find blocks with errors */
2021 for (i=0 ; i < conf->copies ; i++) {
2023 struct md_rdev *rdev;
2024 struct resync_pages *rp;
2026 tbio = r10_bio->devs[i].bio;
2028 if (tbio->bi_end_io != end_sync_read)
2033 tpages = get_resync_pages(tbio)->pages;
2034 d = r10_bio->devs[i].devnum;
2035 rdev = conf->mirrors[d].rdev;
2036 if (!r10_bio->devs[i].bio->bi_status) {
2037 /* We know that the bi_io_vec layout is the same for
2038 * both 'first' and 'i', so we just compare them.
2039 * All vec entries are PAGE_SIZE;
2041 int sectors = r10_bio->sectors;
2042 for (j = 0; j < vcnt; j++) {
2043 int len = PAGE_SIZE;
2044 if (sectors < (len / 512))
2045 len = sectors * 512;
2046 if (memcmp(page_address(fpages[j]),
2047 page_address(tpages[j]),
2054 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2055 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2056 /* Don't fix anything. */
2058 } else if (test_bit(FailFast, &rdev->flags)) {
2059 /* Just give up on this device */
2060 md_error(rdev->mddev, rdev);
2063 /* Ok, we need to write this bio, either to correct an
2064 * inconsistency or to correct an unreadable block.
2065 * First we need to fixup bv_offset, bv_len and
2066 * bi_vecs, as the read request might have corrupted these
2068 rp = get_resync_pages(tbio);
2071 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2073 rp->raid_bio = r10_bio;
2074 tbio->bi_private = rp;
2075 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2076 tbio->bi_end_io = end_sync_write;
2077 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2079 bio_copy_data(tbio, fbio);
2081 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2082 atomic_inc(&r10_bio->remaining);
2083 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2085 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2086 tbio->bi_opf |= MD_FAILFAST;
2087 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2088 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2089 submit_bio_noacct(tbio);
2092 /* Now write out to any replacement devices
2095 for (i = 0; i < conf->copies; i++) {
2098 tbio = r10_bio->devs[i].repl_bio;
2099 if (!tbio || !tbio->bi_end_io)
2101 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2102 && r10_bio->devs[i].bio != fbio)
2103 bio_copy_data(tbio, fbio);
2104 d = r10_bio->devs[i].devnum;
2105 atomic_inc(&r10_bio->remaining);
2106 md_sync_acct(conf->mirrors[d].replacement->bdev,
2108 submit_bio_noacct(tbio);
2112 if (atomic_dec_and_test(&r10_bio->remaining)) {
2113 md_done_sync(mddev, r10_bio->sectors, 1);
2119 * Now for the recovery code.
2120 * Recovery happens across physical sectors.
2121 * We recover all non-is_sync drives by finding the virtual address of
2122 * each, and then choose a working drive that also has that virt address.
2123 * There is a separate r10_bio for each non-in_sync drive.
2124 * Only the first two slots are in use. The first for reading,
2125 * The second for writing.
2128 static void fix_recovery_read_error(struct r10bio *r10_bio)
2130 /* We got a read error during recovery.
2131 * We repeat the read in smaller page-sized sections.
2132 * If a read succeeds, write it to the new device or record
2133 * a bad block if we cannot.
2134 * If a read fails, record a bad block on both old and
2137 struct mddev *mddev = r10_bio->mddev;
2138 struct r10conf *conf = mddev->private;
2139 struct bio *bio = r10_bio->devs[0].bio;
2141 int sectors = r10_bio->sectors;
2143 int dr = r10_bio->devs[0].devnum;
2144 int dw = r10_bio->devs[1].devnum;
2145 struct page **pages = get_resync_pages(bio)->pages;
2149 struct md_rdev *rdev;
2153 if (s > (PAGE_SIZE>>9))
2156 rdev = conf->mirrors[dr].rdev;
2157 addr = r10_bio->devs[0].addr + sect,
2158 ok = sync_page_io(rdev,
2162 REQ_OP_READ, 0, false);
2164 rdev = conf->mirrors[dw].rdev;
2165 addr = r10_bio->devs[1].addr + sect;
2166 ok = sync_page_io(rdev,
2170 REQ_OP_WRITE, 0, false);
2172 set_bit(WriteErrorSeen, &rdev->flags);
2173 if (!test_and_set_bit(WantReplacement,
2175 set_bit(MD_RECOVERY_NEEDED,
2176 &rdev->mddev->recovery);
2180 /* We don't worry if we cannot set a bad block -
2181 * it really is bad so there is no loss in not
2184 rdev_set_badblocks(rdev, addr, s, 0);
2186 if (rdev != conf->mirrors[dw].rdev) {
2187 /* need bad block on destination too */
2188 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2189 addr = r10_bio->devs[1].addr + sect;
2190 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2192 /* just abort the recovery */
2193 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2196 conf->mirrors[dw].recovery_disabled
2197 = mddev->recovery_disabled;
2198 set_bit(MD_RECOVERY_INTR,
2211 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2213 struct r10conf *conf = mddev->private;
2215 struct bio *wbio, *wbio2;
2217 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2218 fix_recovery_read_error(r10_bio);
2219 end_sync_request(r10_bio);
2224 * share the pages with the first bio
2225 * and submit the write request
2227 d = r10_bio->devs[1].devnum;
2228 wbio = r10_bio->devs[1].bio;
2229 wbio2 = r10_bio->devs[1].repl_bio;
2230 /* Need to test wbio2->bi_end_io before we call
2231 * submit_bio_noacct as if the former is NULL,
2232 * the latter is free to free wbio2.
2234 if (wbio2 && !wbio2->bi_end_io)
2236 if (wbio->bi_end_io) {
2237 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2238 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2239 submit_bio_noacct(wbio);
2242 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2243 md_sync_acct(conf->mirrors[d].replacement->bdev,
2244 bio_sectors(wbio2));
2245 submit_bio_noacct(wbio2);
2250 * Used by fix_read_error() to decay the per rdev read_errors.
2251 * We halve the read error count for every hour that has elapsed
2252 * since the last recorded read error.
2255 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2258 unsigned long hours_since_last;
2259 unsigned int read_errors = atomic_read(&rdev->read_errors);
2261 cur_time_mon = ktime_get_seconds();
2263 if (rdev->last_read_error == 0) {
2264 /* first time we've seen a read error */
2265 rdev->last_read_error = cur_time_mon;
2269 hours_since_last = (long)(cur_time_mon -
2270 rdev->last_read_error) / 3600;
2272 rdev->last_read_error = cur_time_mon;
2275 * if hours_since_last is > the number of bits in read_errors
2276 * just set read errors to 0. We do this to avoid
2277 * overflowing the shift of read_errors by hours_since_last.
2279 if (hours_since_last >= 8 * sizeof(read_errors))
2280 atomic_set(&rdev->read_errors, 0);
2282 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2285 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2286 int sectors, struct page *page, int rw)
2291 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2292 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2294 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2298 set_bit(WriteErrorSeen, &rdev->flags);
2299 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2300 set_bit(MD_RECOVERY_NEEDED,
2301 &rdev->mddev->recovery);
2303 /* need to record an error - either for the block or the device */
2304 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2305 md_error(rdev->mddev, rdev);
2310 * This is a kernel thread which:
2312 * 1. Retries failed read operations on working mirrors.
2313 * 2. Updates the raid superblock when problems encounter.
2314 * 3. Performs writes following reads for array synchronising.
2317 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2319 int sect = 0; /* Offset from r10_bio->sector */
2320 int sectors = r10_bio->sectors;
2321 struct md_rdev *rdev;
2322 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2323 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2325 /* still own a reference to this rdev, so it cannot
2326 * have been cleared recently.
2328 rdev = conf->mirrors[d].rdev;
2330 if (test_bit(Faulty, &rdev->flags))
2331 /* drive has already been failed, just ignore any
2332 more fix_read_error() attempts */
2335 check_decay_read_errors(mddev, rdev);
2336 atomic_inc(&rdev->read_errors);
2337 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2338 char b[BDEVNAME_SIZE];
2339 bdevname(rdev->bdev, b);
2341 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2343 atomic_read(&rdev->read_errors), max_read_errors);
2344 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2346 md_error(mddev, rdev);
2347 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2353 int sl = r10_bio->read_slot;
2357 if (s > (PAGE_SIZE>>9))
2365 d = r10_bio->devs[sl].devnum;
2366 rdev = rcu_dereference(conf->mirrors[d].rdev);
2368 test_bit(In_sync, &rdev->flags) &&
2369 !test_bit(Faulty, &rdev->flags) &&
2370 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2371 &first_bad, &bad_sectors) == 0) {
2372 atomic_inc(&rdev->nr_pending);
2374 success = sync_page_io(rdev,
2375 r10_bio->devs[sl].addr +
2379 REQ_OP_READ, 0, false);
2380 rdev_dec_pending(rdev, mddev);
2386 if (sl == conf->copies)
2388 } while (!success && sl != r10_bio->read_slot);
2392 /* Cannot read from anywhere, just mark the block
2393 * as bad on the first device to discourage future
2396 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2397 rdev = conf->mirrors[dn].rdev;
2399 if (!rdev_set_badblocks(
2401 r10_bio->devs[r10_bio->read_slot].addr
2404 md_error(mddev, rdev);
2405 r10_bio->devs[r10_bio->read_slot].bio
2412 /* write it back and re-read */
2414 while (sl != r10_bio->read_slot) {
2415 char b[BDEVNAME_SIZE];
2420 d = r10_bio->devs[sl].devnum;
2421 rdev = rcu_dereference(conf->mirrors[d].rdev);
2423 test_bit(Faulty, &rdev->flags) ||
2424 !test_bit(In_sync, &rdev->flags))
2427 atomic_inc(&rdev->nr_pending);
2429 if (r10_sync_page_io(rdev,
2430 r10_bio->devs[sl].addr +
2432 s, conf->tmppage, WRITE)
2434 /* Well, this device is dead */
2435 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2437 (unsigned long long)(
2439 choose_data_offset(r10_bio,
2441 bdevname(rdev->bdev, b));
2442 pr_notice("md/raid10:%s: %s: failing drive\n",
2444 bdevname(rdev->bdev, b));
2446 rdev_dec_pending(rdev, mddev);
2450 while (sl != r10_bio->read_slot) {
2451 char b[BDEVNAME_SIZE];
2456 d = r10_bio->devs[sl].devnum;
2457 rdev = rcu_dereference(conf->mirrors[d].rdev);
2459 test_bit(Faulty, &rdev->flags) ||
2460 !test_bit(In_sync, &rdev->flags))
2463 atomic_inc(&rdev->nr_pending);
2465 switch (r10_sync_page_io(rdev,
2466 r10_bio->devs[sl].addr +
2471 /* Well, this device is dead */
2472 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2474 (unsigned long long)(
2476 choose_data_offset(r10_bio, rdev)),
2477 bdevname(rdev->bdev, b));
2478 pr_notice("md/raid10:%s: %s: failing drive\n",
2480 bdevname(rdev->bdev, b));
2483 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2485 (unsigned long long)(
2487 choose_data_offset(r10_bio, rdev)),
2488 bdevname(rdev->bdev, b));
2489 atomic_add(s, &rdev->corrected_errors);
2492 rdev_dec_pending(rdev, mddev);
2502 static int narrow_write_error(struct r10bio *r10_bio, int i)
2504 struct bio *bio = r10_bio->master_bio;
2505 struct mddev *mddev = r10_bio->mddev;
2506 struct r10conf *conf = mddev->private;
2507 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2508 /* bio has the data to be written to slot 'i' where
2509 * we just recently had a write error.
2510 * We repeatedly clone the bio and trim down to one block,
2511 * then try the write. Where the write fails we record
2513 * It is conceivable that the bio doesn't exactly align with
2514 * blocks. We must handle this.
2516 * We currently own a reference to the rdev.
2522 int sect_to_write = r10_bio->sectors;
2525 if (rdev->badblocks.shift < 0)
2528 block_sectors = roundup(1 << rdev->badblocks.shift,
2529 bdev_logical_block_size(rdev->bdev) >> 9);
2530 sector = r10_bio->sector;
2531 sectors = ((r10_bio->sector + block_sectors)
2532 & ~(sector_t)(block_sectors - 1))
2535 while (sect_to_write) {
2538 if (sectors > sect_to_write)
2539 sectors = sect_to_write;
2540 /* Write at 'sector' for 'sectors' */
2541 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2542 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2543 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2544 wbio->bi_iter.bi_sector = wsector +
2545 choose_data_offset(r10_bio, rdev);
2546 bio_set_dev(wbio, rdev->bdev);
2547 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2549 if (submit_bio_wait(wbio) < 0)
2551 ok = rdev_set_badblocks(rdev, wsector,
2556 sect_to_write -= sectors;
2558 sectors = block_sectors;
2563 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2565 int slot = r10_bio->read_slot;
2567 struct r10conf *conf = mddev->private;
2568 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2570 /* we got a read error. Maybe the drive is bad. Maybe just
2571 * the block and we can fix it.
2572 * We freeze all other IO, and try reading the block from
2573 * other devices. When we find one, we re-write
2574 * and check it that fixes the read error.
2575 * This is all done synchronously while the array is
2578 bio = r10_bio->devs[slot].bio;
2580 r10_bio->devs[slot].bio = NULL;
2583 r10_bio->devs[slot].bio = IO_BLOCKED;
2584 else if (!test_bit(FailFast, &rdev->flags)) {
2585 freeze_array(conf, 1);
2586 fix_read_error(conf, mddev, r10_bio);
2587 unfreeze_array(conf);
2589 md_error(mddev, rdev);
2591 rdev_dec_pending(rdev, mddev);
2592 allow_barrier(conf);
2594 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2597 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2599 /* Some sort of write request has finished and it
2600 * succeeded in writing where we thought there was a
2601 * bad block. So forget the bad block.
2602 * Or possibly if failed and we need to record
2606 struct md_rdev *rdev;
2608 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2609 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2610 for (m = 0; m < conf->copies; m++) {
2611 int dev = r10_bio->devs[m].devnum;
2612 rdev = conf->mirrors[dev].rdev;
2613 if (r10_bio->devs[m].bio == NULL ||
2614 r10_bio->devs[m].bio->bi_end_io == NULL)
2616 if (!r10_bio->devs[m].bio->bi_status) {
2617 rdev_clear_badblocks(
2619 r10_bio->devs[m].addr,
2620 r10_bio->sectors, 0);
2622 if (!rdev_set_badblocks(
2624 r10_bio->devs[m].addr,
2625 r10_bio->sectors, 0))
2626 md_error(conf->mddev, rdev);
2628 rdev = conf->mirrors[dev].replacement;
2629 if (r10_bio->devs[m].repl_bio == NULL ||
2630 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2633 if (!r10_bio->devs[m].repl_bio->bi_status) {
2634 rdev_clear_badblocks(
2636 r10_bio->devs[m].addr,
2637 r10_bio->sectors, 0);
2639 if (!rdev_set_badblocks(
2641 r10_bio->devs[m].addr,
2642 r10_bio->sectors, 0))
2643 md_error(conf->mddev, rdev);
2649 for (m = 0; m < conf->copies; m++) {
2650 int dev = r10_bio->devs[m].devnum;
2651 struct bio *bio = r10_bio->devs[m].bio;
2652 rdev = conf->mirrors[dev].rdev;
2653 if (bio == IO_MADE_GOOD) {
2654 rdev_clear_badblocks(
2656 r10_bio->devs[m].addr,
2657 r10_bio->sectors, 0);
2658 rdev_dec_pending(rdev, conf->mddev);
2659 } else if (bio != NULL && bio->bi_status) {
2661 if (!narrow_write_error(r10_bio, m)) {
2662 md_error(conf->mddev, rdev);
2663 set_bit(R10BIO_Degraded,
2666 rdev_dec_pending(rdev, conf->mddev);
2668 bio = r10_bio->devs[m].repl_bio;
2669 rdev = conf->mirrors[dev].replacement;
2670 if (rdev && bio == IO_MADE_GOOD) {
2671 rdev_clear_badblocks(
2673 r10_bio->devs[m].addr,
2674 r10_bio->sectors, 0);
2675 rdev_dec_pending(rdev, conf->mddev);
2679 spin_lock_irq(&conf->device_lock);
2680 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2682 spin_unlock_irq(&conf->device_lock);
2684 * In case freeze_array() is waiting for condition
2685 * nr_pending == nr_queued + extra to be true.
2687 wake_up(&conf->wait_barrier);
2688 md_wakeup_thread(conf->mddev->thread);
2690 if (test_bit(R10BIO_WriteError,
2692 close_write(r10_bio);
2693 raid_end_bio_io(r10_bio);
2698 static void raid10d(struct md_thread *thread)
2700 struct mddev *mddev = thread->mddev;
2701 struct r10bio *r10_bio;
2702 unsigned long flags;
2703 struct r10conf *conf = mddev->private;
2704 struct list_head *head = &conf->retry_list;
2705 struct blk_plug plug;
2707 md_check_recovery(mddev);
2709 if (!list_empty_careful(&conf->bio_end_io_list) &&
2710 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2712 spin_lock_irqsave(&conf->device_lock, flags);
2713 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2714 while (!list_empty(&conf->bio_end_io_list)) {
2715 list_move(conf->bio_end_io_list.prev, &tmp);
2719 spin_unlock_irqrestore(&conf->device_lock, flags);
2720 while (!list_empty(&tmp)) {
2721 r10_bio = list_first_entry(&tmp, struct r10bio,
2723 list_del(&r10_bio->retry_list);
2724 if (mddev->degraded)
2725 set_bit(R10BIO_Degraded, &r10_bio->state);
2727 if (test_bit(R10BIO_WriteError,
2729 close_write(r10_bio);
2730 raid_end_bio_io(r10_bio);
2734 blk_start_plug(&plug);
2737 flush_pending_writes(conf);
2739 spin_lock_irqsave(&conf->device_lock, flags);
2740 if (list_empty(head)) {
2741 spin_unlock_irqrestore(&conf->device_lock, flags);
2744 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2745 list_del(head->prev);
2747 spin_unlock_irqrestore(&conf->device_lock, flags);
2749 mddev = r10_bio->mddev;
2750 conf = mddev->private;
2751 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2752 test_bit(R10BIO_WriteError, &r10_bio->state))
2753 handle_write_completed(conf, r10_bio);
2754 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2755 reshape_request_write(mddev, r10_bio);
2756 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2757 sync_request_write(mddev, r10_bio);
2758 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2759 recovery_request_write(mddev, r10_bio);
2760 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2761 handle_read_error(mddev, r10_bio);
2766 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2767 md_check_recovery(mddev);
2769 blk_finish_plug(&plug);
2772 static int init_resync(struct r10conf *conf)
2776 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2777 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2778 conf->have_replacement = 0;
2779 for (i = 0; i < conf->geo.raid_disks; i++)
2780 if (conf->mirrors[i].replacement)
2781 conf->have_replacement = 1;
2782 ret = mempool_init(&conf->r10buf_pool, buffs,
2783 r10buf_pool_alloc, r10buf_pool_free, conf);
2786 conf->next_resync = 0;
2790 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2792 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2793 struct rsync_pages *rp;
2798 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2799 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2800 nalloc = conf->copies; /* resync */
2802 nalloc = 2; /* recovery */
2804 for (i = 0; i < nalloc; i++) {
2805 bio = r10bio->devs[i].bio;
2806 rp = bio->bi_private;
2808 bio->bi_private = rp;
2809 bio = r10bio->devs[i].repl_bio;
2811 rp = bio->bi_private;
2813 bio->bi_private = rp;
2820 * Set cluster_sync_high since we need other nodes to add the
2821 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2823 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2825 sector_t window_size;
2826 int extra_chunk, chunks;
2829 * First, here we define "stripe" as a unit which across
2830 * all member devices one time, so we get chunks by use
2831 * raid_disks / near_copies. Otherwise, if near_copies is
2832 * close to raid_disks, then resync window could increases
2833 * linearly with the increase of raid_disks, which means
2834 * we will suspend a really large IO window while it is not
2835 * necessary. If raid_disks is not divisible by near_copies,
2836 * an extra chunk is needed to ensure the whole "stripe" is
2840 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2841 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2845 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2848 * At least use a 32M window to align with raid1's resync window
2850 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2851 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2853 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2857 * perform a "sync" on one "block"
2859 * We need to make sure that no normal I/O request - particularly write
2860 * requests - conflict with active sync requests.
2862 * This is achieved by tracking pending requests and a 'barrier' concept
2863 * that can be installed to exclude normal IO requests.
2865 * Resync and recovery are handled very differently.
2866 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2868 * For resync, we iterate over virtual addresses, read all copies,
2869 * and update if there are differences. If only one copy is live,
2871 * For recovery, we iterate over physical addresses, read a good
2872 * value for each non-in_sync drive, and over-write.
2874 * So, for recovery we may have several outstanding complex requests for a
2875 * given address, one for each out-of-sync device. We model this by allocating
2876 * a number of r10_bio structures, one for each out-of-sync device.
2877 * As we setup these structures, we collect all bio's together into a list
2878 * which we then process collectively to add pages, and then process again
2879 * to pass to submit_bio_noacct.
2881 * The r10_bio structures are linked using a borrowed master_bio pointer.
2882 * This link is counted in ->remaining. When the r10_bio that points to NULL
2883 * has its remaining count decremented to 0, the whole complex operation
2888 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2891 struct r10conf *conf = mddev->private;
2892 struct r10bio *r10_bio;
2893 struct bio *biolist = NULL, *bio;
2894 sector_t max_sector, nr_sectors;
2897 sector_t sync_blocks;
2898 sector_t sectors_skipped = 0;
2899 int chunks_skipped = 0;
2900 sector_t chunk_mask = conf->geo.chunk_mask;
2903 if (!mempool_initialized(&conf->r10buf_pool))
2904 if (init_resync(conf))
2908 * Allow skipping a full rebuild for incremental assembly
2909 * of a clean array, like RAID1 does.
2911 if (mddev->bitmap == NULL &&
2912 mddev->recovery_cp == MaxSector &&
2913 mddev->reshape_position == MaxSector &&
2914 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2915 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2916 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2917 conf->fullsync == 0) {
2919 return mddev->dev_sectors - sector_nr;
2923 max_sector = mddev->dev_sectors;
2924 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2925 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2926 max_sector = mddev->resync_max_sectors;
2927 if (sector_nr >= max_sector) {
2928 conf->cluster_sync_low = 0;
2929 conf->cluster_sync_high = 0;
2931 /* If we aborted, we need to abort the
2932 * sync on the 'current' bitmap chucks (there can
2933 * be several when recovering multiple devices).
2934 * as we may have started syncing it but not finished.
2935 * We can find the current address in
2936 * mddev->curr_resync, but for recovery,
2937 * we need to convert that to several
2938 * virtual addresses.
2940 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2946 if (mddev->curr_resync < max_sector) { /* aborted */
2947 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2948 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2950 else for (i = 0; i < conf->geo.raid_disks; i++) {
2952 raid10_find_virt(conf, mddev->curr_resync, i);
2953 md_bitmap_end_sync(mddev->bitmap, sect,
2957 /* completed sync */
2958 if ((!mddev->bitmap || conf->fullsync)
2959 && conf->have_replacement
2960 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2961 /* Completed a full sync so the replacements
2962 * are now fully recovered.
2965 for (i = 0; i < conf->geo.raid_disks; i++) {
2966 struct md_rdev *rdev =
2967 rcu_dereference(conf->mirrors[i].replacement);
2969 rdev->recovery_offset = MaxSector;
2975 md_bitmap_close_sync(mddev->bitmap);
2978 return sectors_skipped;
2981 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2982 return reshape_request(mddev, sector_nr, skipped);
2984 if (chunks_skipped >= conf->geo.raid_disks) {
2985 /* if there has been nothing to do on any drive,
2986 * then there is nothing to do at all..
2989 return (max_sector - sector_nr) + sectors_skipped;
2992 if (max_sector > mddev->resync_max)
2993 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2995 /* make sure whole request will fit in a chunk - if chunks
2998 if (conf->geo.near_copies < conf->geo.raid_disks &&
2999 max_sector > (sector_nr | chunk_mask))
3000 max_sector = (sector_nr | chunk_mask) + 1;
3003 * If there is non-resync activity waiting for a turn, then let it
3004 * though before starting on this new sync request.
3006 if (conf->nr_waiting)
3007 schedule_timeout_uninterruptible(1);
3009 /* Again, very different code for resync and recovery.
3010 * Both must result in an r10bio with a list of bios that
3011 * have bi_end_io, bi_sector, bi_disk set,
3012 * and bi_private set to the r10bio.
3013 * For recovery, we may actually create several r10bios
3014 * with 2 bios in each, that correspond to the bios in the main one.
3015 * In this case, the subordinate r10bios link back through a
3016 * borrowed master_bio pointer, and the counter in the master
3017 * includes a ref from each subordinate.
3019 /* First, we decide what to do and set ->bi_end_io
3020 * To end_sync_read if we want to read, and
3021 * end_sync_write if we will want to write.
3024 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3025 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3026 /* recovery... the complicated one */
3030 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3036 int need_recover = 0;
3037 int need_replace = 0;
3038 struct raid10_info *mirror = &conf->mirrors[i];
3039 struct md_rdev *mrdev, *mreplace;
3042 mrdev = rcu_dereference(mirror->rdev);
3043 mreplace = rcu_dereference(mirror->replacement);
3045 if (mrdev != NULL &&
3046 !test_bit(Faulty, &mrdev->flags) &&
3047 !test_bit(In_sync, &mrdev->flags))
3049 if (mreplace != NULL &&
3050 !test_bit(Faulty, &mreplace->flags))
3053 if (!need_recover && !need_replace) {
3059 /* want to reconstruct this device */
3061 sect = raid10_find_virt(conf, sector_nr, i);
3062 if (sect >= mddev->resync_max_sectors) {
3063 /* last stripe is not complete - don't
3064 * try to recover this sector.
3069 if (mreplace && test_bit(Faulty, &mreplace->flags))
3071 /* Unless we are doing a full sync, or a replacement
3072 * we only need to recover the block if it is set in
3075 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3077 if (sync_blocks < max_sync)
3078 max_sync = sync_blocks;
3082 /* yep, skip the sync_blocks here, but don't assume
3083 * that there will never be anything to do here
3085 chunks_skipped = -1;
3089 atomic_inc(&mrdev->nr_pending);
3091 atomic_inc(&mreplace->nr_pending);
3094 r10_bio = raid10_alloc_init_r10buf(conf);
3096 raise_barrier(conf, rb2 != NULL);
3097 atomic_set(&r10_bio->remaining, 0);
3099 r10_bio->master_bio = (struct bio*)rb2;
3101 atomic_inc(&rb2->remaining);
3102 r10_bio->mddev = mddev;
3103 set_bit(R10BIO_IsRecover, &r10_bio->state);
3104 r10_bio->sector = sect;
3106 raid10_find_phys(conf, r10_bio);
3108 /* Need to check if the array will still be
3112 for (j = 0; j < conf->geo.raid_disks; j++) {
3113 struct md_rdev *rdev = rcu_dereference(
3114 conf->mirrors[j].rdev);
3115 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3121 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3122 &sync_blocks, still_degraded);
3125 for (j=0; j<conf->copies;j++) {
3127 int d = r10_bio->devs[j].devnum;
3128 sector_t from_addr, to_addr;
3129 struct md_rdev *rdev =
3130 rcu_dereference(conf->mirrors[d].rdev);
3131 sector_t sector, first_bad;
3134 !test_bit(In_sync, &rdev->flags))
3136 /* This is where we read from */
3138 sector = r10_bio->devs[j].addr;
3140 if (is_badblock(rdev, sector, max_sync,
3141 &first_bad, &bad_sectors)) {
3142 if (first_bad > sector)
3143 max_sync = first_bad - sector;
3145 bad_sectors -= (sector
3147 if (max_sync > bad_sectors)
3148 max_sync = bad_sectors;
3152 bio = r10_bio->devs[0].bio;
3153 bio->bi_next = biolist;
3155 bio->bi_end_io = end_sync_read;
3156 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3157 if (test_bit(FailFast, &rdev->flags))
3158 bio->bi_opf |= MD_FAILFAST;
3159 from_addr = r10_bio->devs[j].addr;
3160 bio->bi_iter.bi_sector = from_addr +
3162 bio_set_dev(bio, rdev->bdev);
3163 atomic_inc(&rdev->nr_pending);
3164 /* and we write to 'i' (if not in_sync) */
3166 for (k=0; k<conf->copies; k++)
3167 if (r10_bio->devs[k].devnum == i)
3169 BUG_ON(k == conf->copies);
3170 to_addr = r10_bio->devs[k].addr;
3171 r10_bio->devs[0].devnum = d;
3172 r10_bio->devs[0].addr = from_addr;
3173 r10_bio->devs[1].devnum = i;
3174 r10_bio->devs[1].addr = to_addr;
3177 bio = r10_bio->devs[1].bio;
3178 bio->bi_next = biolist;
3180 bio->bi_end_io = end_sync_write;
3181 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3182 bio->bi_iter.bi_sector = to_addr
3183 + mrdev->data_offset;
3184 bio_set_dev(bio, mrdev->bdev);
3185 atomic_inc(&r10_bio->remaining);
3187 r10_bio->devs[1].bio->bi_end_io = NULL;
3189 /* and maybe write to replacement */
3190 bio = r10_bio->devs[1].repl_bio;
3192 bio->bi_end_io = NULL;
3193 /* Note: if need_replace, then bio
3194 * cannot be NULL as r10buf_pool_alloc will
3195 * have allocated it.
3199 bio->bi_next = biolist;
3201 bio->bi_end_io = end_sync_write;
3202 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3203 bio->bi_iter.bi_sector = to_addr +
3204 mreplace->data_offset;
3205 bio_set_dev(bio, mreplace->bdev);
3206 atomic_inc(&r10_bio->remaining);
3210 if (j == conf->copies) {
3211 /* Cannot recover, so abort the recovery or
3212 * record a bad block */
3214 /* problem is that there are bad blocks
3215 * on other device(s)
3218 for (k = 0; k < conf->copies; k++)
3219 if (r10_bio->devs[k].devnum == i)
3221 if (!test_bit(In_sync,
3223 && !rdev_set_badblocks(
3225 r10_bio->devs[k].addr,
3229 !rdev_set_badblocks(
3231 r10_bio->devs[k].addr,
3236 if (!test_and_set_bit(MD_RECOVERY_INTR,
3238 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3240 mirror->recovery_disabled
3241 = mddev->recovery_disabled;
3245 atomic_dec(&rb2->remaining);
3247 rdev_dec_pending(mrdev, mddev);
3249 rdev_dec_pending(mreplace, mddev);
3252 rdev_dec_pending(mrdev, mddev);
3254 rdev_dec_pending(mreplace, mddev);
3255 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3256 /* Only want this if there is elsewhere to
3257 * read from. 'j' is currently the first
3261 for (; j < conf->copies; j++) {
3262 int d = r10_bio->devs[j].devnum;
3263 if (conf->mirrors[d].rdev &&
3265 &conf->mirrors[d].rdev->flags))
3269 r10_bio->devs[0].bio->bi_opf
3273 if (biolist == NULL) {
3275 struct r10bio *rb2 = r10_bio;
3276 r10_bio = (struct r10bio*) rb2->master_bio;
3277 rb2->master_bio = NULL;
3283 /* resync. Schedule a read for every block at this virt offset */
3287 * Since curr_resync_completed could probably not update in
3288 * time, and we will set cluster_sync_low based on it.
3289 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3290 * safety reason, which ensures curr_resync_completed is
3291 * updated in bitmap_cond_end_sync.
3293 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3294 mddev_is_clustered(mddev) &&
3295 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3297 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3298 &sync_blocks, mddev->degraded) &&
3299 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3300 &mddev->recovery)) {
3301 /* We can skip this block */
3303 return sync_blocks + sectors_skipped;
3305 if (sync_blocks < max_sync)
3306 max_sync = sync_blocks;
3307 r10_bio = raid10_alloc_init_r10buf(conf);
3310 r10_bio->mddev = mddev;
3311 atomic_set(&r10_bio->remaining, 0);
3312 raise_barrier(conf, 0);
3313 conf->next_resync = sector_nr;
3315 r10_bio->master_bio = NULL;
3316 r10_bio->sector = sector_nr;
3317 set_bit(R10BIO_IsSync, &r10_bio->state);
3318 raid10_find_phys(conf, r10_bio);
3319 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3321 for (i = 0; i < conf->copies; i++) {
3322 int d = r10_bio->devs[i].devnum;
3323 sector_t first_bad, sector;
3325 struct md_rdev *rdev;
3327 if (r10_bio->devs[i].repl_bio)
3328 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3330 bio = r10_bio->devs[i].bio;
3331 bio->bi_status = BLK_STS_IOERR;
3333 rdev = rcu_dereference(conf->mirrors[d].rdev);
3334 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3338 sector = r10_bio->devs[i].addr;
3339 if (is_badblock(rdev, sector, max_sync,
3340 &first_bad, &bad_sectors)) {
3341 if (first_bad > sector)
3342 max_sync = first_bad - sector;
3344 bad_sectors -= (sector - first_bad);
3345 if (max_sync > bad_sectors)
3346 max_sync = bad_sectors;
3351 atomic_inc(&rdev->nr_pending);
3352 atomic_inc(&r10_bio->remaining);
3353 bio->bi_next = biolist;
3355 bio->bi_end_io = end_sync_read;
3356 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3357 if (test_bit(FailFast, &rdev->flags))
3358 bio->bi_opf |= MD_FAILFAST;
3359 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3360 bio_set_dev(bio, rdev->bdev);
3363 rdev = rcu_dereference(conf->mirrors[d].replacement);
3364 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3368 atomic_inc(&rdev->nr_pending);
3370 /* Need to set up for writing to the replacement */
3371 bio = r10_bio->devs[i].repl_bio;
3372 bio->bi_status = BLK_STS_IOERR;
3374 sector = r10_bio->devs[i].addr;
3375 bio->bi_next = biolist;
3377 bio->bi_end_io = end_sync_write;
3378 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3379 if (test_bit(FailFast, &rdev->flags))
3380 bio->bi_opf |= MD_FAILFAST;
3381 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3382 bio_set_dev(bio, rdev->bdev);
3388 for (i=0; i<conf->copies; i++) {
3389 int d = r10_bio->devs[i].devnum;
3390 if (r10_bio->devs[i].bio->bi_end_io)
3391 rdev_dec_pending(conf->mirrors[d].rdev,
3393 if (r10_bio->devs[i].repl_bio &&
3394 r10_bio->devs[i].repl_bio->bi_end_io)
3396 conf->mirrors[d].replacement,
3406 if (sector_nr + max_sync < max_sector)
3407 max_sector = sector_nr + max_sync;
3410 int len = PAGE_SIZE;
3411 if (sector_nr + (len>>9) > max_sector)
3412 len = (max_sector - sector_nr) << 9;
3415 for (bio= biolist ; bio ; bio=bio->bi_next) {
3416 struct resync_pages *rp = get_resync_pages(bio);
3417 page = resync_fetch_page(rp, page_idx);
3419 * won't fail because the vec table is big enough
3420 * to hold all these pages
3422 bio_add_page(bio, page, len, 0);
3424 nr_sectors += len>>9;
3425 sector_nr += len>>9;
3426 } while (++page_idx < RESYNC_PAGES);
3427 r10_bio->sectors = nr_sectors;
3429 if (mddev_is_clustered(mddev) &&
3430 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3431 /* It is resync not recovery */
3432 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3433 conf->cluster_sync_low = mddev->curr_resync_completed;
3434 raid10_set_cluster_sync_high(conf);
3435 /* Send resync message */
3436 md_cluster_ops->resync_info_update(mddev,
3437 conf->cluster_sync_low,
3438 conf->cluster_sync_high);
3440 } else if (mddev_is_clustered(mddev)) {
3441 /* This is recovery not resync */
3442 sector_t sect_va1, sect_va2;
3443 bool broadcast_msg = false;
3445 for (i = 0; i < conf->geo.raid_disks; i++) {
3447 * sector_nr is a device address for recovery, so we
3448 * need translate it to array address before compare
3449 * with cluster_sync_high.
3451 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3453 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3454 broadcast_msg = true;
3456 * curr_resync_completed is similar as
3457 * sector_nr, so make the translation too.
3459 sect_va2 = raid10_find_virt(conf,
3460 mddev->curr_resync_completed, i);
3462 if (conf->cluster_sync_low == 0 ||
3463 conf->cluster_sync_low > sect_va2)
3464 conf->cluster_sync_low = sect_va2;
3467 if (broadcast_msg) {
3468 raid10_set_cluster_sync_high(conf);
3469 md_cluster_ops->resync_info_update(mddev,
3470 conf->cluster_sync_low,
3471 conf->cluster_sync_high);
3477 biolist = biolist->bi_next;
3479 bio->bi_next = NULL;
3480 r10_bio = get_resync_r10bio(bio);
3481 r10_bio->sectors = nr_sectors;
3483 if (bio->bi_end_io == end_sync_read) {
3484 md_sync_acct_bio(bio, nr_sectors);
3486 submit_bio_noacct(bio);
3490 if (sectors_skipped)
3491 /* pretend they weren't skipped, it makes
3492 * no important difference in this case
3494 md_done_sync(mddev, sectors_skipped, 1);
3496 return sectors_skipped + nr_sectors;
3498 /* There is nowhere to write, so all non-sync
3499 * drives must be failed or in resync, all drives
3500 * have a bad block, so try the next chunk...
3502 if (sector_nr + max_sync < max_sector)
3503 max_sector = sector_nr + max_sync;
3505 sectors_skipped += (max_sector - sector_nr);
3507 sector_nr = max_sector;
3512 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3515 struct r10conf *conf = mddev->private;
3518 raid_disks = min(conf->geo.raid_disks,
3519 conf->prev.raid_disks);
3521 sectors = conf->dev_sectors;
3523 size = sectors >> conf->geo.chunk_shift;
3524 sector_div(size, conf->geo.far_copies);
3525 size = size * raid_disks;
3526 sector_div(size, conf->geo.near_copies);
3528 return size << conf->geo.chunk_shift;
3531 static void calc_sectors(struct r10conf *conf, sector_t size)
3533 /* Calculate the number of sectors-per-device that will
3534 * actually be used, and set conf->dev_sectors and
3538 size = size >> conf->geo.chunk_shift;
3539 sector_div(size, conf->geo.far_copies);
3540 size = size * conf->geo.raid_disks;
3541 sector_div(size, conf->geo.near_copies);
3542 /* 'size' is now the number of chunks in the array */
3543 /* calculate "used chunks per device" */
3544 size = size * conf->copies;
3546 /* We need to round up when dividing by raid_disks to
3547 * get the stride size.
3549 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3551 conf->dev_sectors = size << conf->geo.chunk_shift;
3553 if (conf->geo.far_offset)
3554 conf->geo.stride = 1 << conf->geo.chunk_shift;
3556 sector_div(size, conf->geo.far_copies);
3557 conf->geo.stride = size << conf->geo.chunk_shift;
3561 enum geo_type {geo_new, geo_old, geo_start};
3562 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3565 int layout, chunk, disks;
3568 layout = mddev->layout;
3569 chunk = mddev->chunk_sectors;
3570 disks = mddev->raid_disks - mddev->delta_disks;
3573 layout = mddev->new_layout;
3574 chunk = mddev->new_chunk_sectors;
3575 disks = mddev->raid_disks;
3577 default: /* avoid 'may be unused' warnings */
3578 case geo_start: /* new when starting reshape - raid_disks not
3580 layout = mddev->new_layout;
3581 chunk = mddev->new_chunk_sectors;
3582 disks = mddev->raid_disks + mddev->delta_disks;
3587 if (chunk < (PAGE_SIZE >> 9) ||
3588 !is_power_of_2(chunk))
3591 fc = (layout >> 8) & 255;
3592 fo = layout & (1<<16);
3593 geo->raid_disks = disks;
3594 geo->near_copies = nc;
3595 geo->far_copies = fc;
3596 geo->far_offset = fo;
3597 switch (layout >> 17) {
3598 case 0: /* original layout. simple but not always optimal */
3599 geo->far_set_size = disks;
3601 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3602 * actually using this, but leave code here just in case.*/
3603 geo->far_set_size = disks/fc;
3604 WARN(geo->far_set_size < fc,
3605 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3607 case 2: /* "improved" layout fixed to match documentation */
3608 geo->far_set_size = fc * nc;
3610 default: /* Not a valid layout */
3613 geo->chunk_mask = chunk - 1;
3614 geo->chunk_shift = ffz(~chunk);
3618 static struct r10conf *setup_conf(struct mddev *mddev)
3620 struct r10conf *conf = NULL;
3625 copies = setup_geo(&geo, mddev, geo_new);
3628 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3629 mdname(mddev), PAGE_SIZE);
3633 if (copies < 2 || copies > mddev->raid_disks) {
3634 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3635 mdname(mddev), mddev->new_layout);
3640 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3644 /* FIXME calc properly */
3645 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3646 sizeof(struct raid10_info),
3651 conf->tmppage = alloc_page(GFP_KERNEL);
3656 conf->copies = copies;
3657 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3658 rbio_pool_free, conf);
3662 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3666 calc_sectors(conf, mddev->dev_sectors);
3667 if (mddev->reshape_position == MaxSector) {
3668 conf->prev = conf->geo;
3669 conf->reshape_progress = MaxSector;
3671 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3675 conf->reshape_progress = mddev->reshape_position;
3676 if (conf->prev.far_offset)
3677 conf->prev.stride = 1 << conf->prev.chunk_shift;
3679 /* far_copies must be 1 */
3680 conf->prev.stride = conf->dev_sectors;
3682 conf->reshape_safe = conf->reshape_progress;
3683 spin_lock_init(&conf->device_lock);
3684 INIT_LIST_HEAD(&conf->retry_list);
3685 INIT_LIST_HEAD(&conf->bio_end_io_list);
3687 spin_lock_init(&conf->resync_lock);
3688 init_waitqueue_head(&conf->wait_barrier);
3689 atomic_set(&conf->nr_pending, 0);
3692 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3696 conf->mddev = mddev;
3701 mempool_exit(&conf->r10bio_pool);
3702 kfree(conf->mirrors);
3703 safe_put_page(conf->tmppage);
3704 bioset_exit(&conf->bio_split);
3707 return ERR_PTR(err);
3710 static void raid10_set_io_opt(struct r10conf *conf)
3712 int raid_disks = conf->geo.raid_disks;
3714 if (!(conf->geo.raid_disks % conf->geo.near_copies))
3715 raid_disks /= conf->geo.near_copies;
3716 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
3720 static int raid10_run(struct mddev *mddev)
3722 struct r10conf *conf;
3724 struct raid10_info *disk;
3725 struct md_rdev *rdev;
3727 sector_t min_offset_diff = 0;
3729 bool discard_supported = false;
3731 if (mddev_init_writes_pending(mddev) < 0)
3734 if (mddev->private == NULL) {
3735 conf = setup_conf(mddev);
3737 return PTR_ERR(conf);
3738 mddev->private = conf;
3740 conf = mddev->private;
3744 if (mddev_is_clustered(conf->mddev)) {
3747 fc = (mddev->layout >> 8) & 255;
3748 fo = mddev->layout & (1<<16);
3749 if (fc > 1 || fo > 0) {
3750 pr_err("only near layout is supported by clustered"
3756 mddev->thread = conf->thread;
3757 conf->thread = NULL;
3760 blk_queue_max_discard_sectors(mddev->queue,
3761 mddev->chunk_sectors);
3762 blk_queue_max_write_same_sectors(mddev->queue, 0);
3763 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3764 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
3765 raid10_set_io_opt(conf);
3768 rdev_for_each(rdev, mddev) {
3771 disk_idx = rdev->raid_disk;
3774 if (disk_idx >= conf->geo.raid_disks &&
3775 disk_idx >= conf->prev.raid_disks)
3777 disk = conf->mirrors + disk_idx;
3779 if (test_bit(Replacement, &rdev->flags)) {
3780 if (disk->replacement)
3782 disk->replacement = rdev;
3788 diff = (rdev->new_data_offset - rdev->data_offset);
3789 if (!mddev->reshape_backwards)
3793 if (first || diff < min_offset_diff)
3794 min_offset_diff = diff;
3797 disk_stack_limits(mddev->gendisk, rdev->bdev,
3798 rdev->data_offset << 9);
3800 disk->head_position = 0;
3802 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3803 discard_supported = true;
3808 if (discard_supported)
3809 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3812 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3815 /* need to check that every block has at least one working mirror */
3816 if (!enough(conf, -1)) {
3817 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3822 if (conf->reshape_progress != MaxSector) {
3823 /* must ensure that shape change is supported */
3824 if (conf->geo.far_copies != 1 &&
3825 conf->geo.far_offset == 0)
3827 if (conf->prev.far_copies != 1 &&
3828 conf->prev.far_offset == 0)
3832 mddev->degraded = 0;
3834 i < conf->geo.raid_disks
3835 || i < conf->prev.raid_disks;
3838 disk = conf->mirrors + i;
3840 if (!disk->rdev && disk->replacement) {
3841 /* The replacement is all we have - use it */
3842 disk->rdev = disk->replacement;
3843 disk->replacement = NULL;
3844 clear_bit(Replacement, &disk->rdev->flags);
3848 !test_bit(In_sync, &disk->rdev->flags)) {
3849 disk->head_position = 0;
3852 disk->rdev->saved_raid_disk < 0)
3856 if (disk->replacement &&
3857 !test_bit(In_sync, &disk->replacement->flags) &&
3858 disk->replacement->saved_raid_disk < 0) {
3862 disk->recovery_disabled = mddev->recovery_disabled - 1;
3865 if (mddev->recovery_cp != MaxSector)
3866 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3868 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3869 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3870 conf->geo.raid_disks);
3872 * Ok, everything is just fine now
3874 mddev->dev_sectors = conf->dev_sectors;
3875 size = raid10_size(mddev, 0, 0);
3876 md_set_array_sectors(mddev, size);
3877 mddev->resync_max_sectors = size;
3878 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3880 if (md_integrity_register(mddev))
3883 if (conf->reshape_progress != MaxSector) {
3884 unsigned long before_length, after_length;
3886 before_length = ((1 << conf->prev.chunk_shift) *
3887 conf->prev.far_copies);
3888 after_length = ((1 << conf->geo.chunk_shift) *
3889 conf->geo.far_copies);
3891 if (max(before_length, after_length) > min_offset_diff) {
3892 /* This cannot work */
3893 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3896 conf->offset_diff = min_offset_diff;
3898 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3899 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3900 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3901 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3902 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3904 if (!mddev->sync_thread)
3911 md_unregister_thread(&mddev->thread);
3912 mempool_exit(&conf->r10bio_pool);
3913 safe_put_page(conf->tmppage);
3914 kfree(conf->mirrors);
3916 mddev->private = NULL;
3921 static void raid10_free(struct mddev *mddev, void *priv)
3923 struct r10conf *conf = priv;
3925 mempool_exit(&conf->r10bio_pool);
3926 safe_put_page(conf->tmppage);
3927 kfree(conf->mirrors);
3928 kfree(conf->mirrors_old);
3929 kfree(conf->mirrors_new);
3930 bioset_exit(&conf->bio_split);
3934 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3936 struct r10conf *conf = mddev->private;
3939 raise_barrier(conf, 0);
3941 lower_barrier(conf);
3944 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3946 /* Resize of 'far' arrays is not supported.
3947 * For 'near' and 'offset' arrays we can set the
3948 * number of sectors used to be an appropriate multiple
3949 * of the chunk size.
3950 * For 'offset', this is far_copies*chunksize.
3951 * For 'near' the multiplier is the LCM of
3952 * near_copies and raid_disks.
3953 * So if far_copies > 1 && !far_offset, fail.
3954 * Else find LCM(raid_disks, near_copy)*far_copies and
3955 * multiply by chunk_size. Then round to this number.
3956 * This is mostly done by raid10_size()
3958 struct r10conf *conf = mddev->private;
3959 sector_t oldsize, size;
3961 if (mddev->reshape_position != MaxSector)
3964 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3967 oldsize = raid10_size(mddev, 0, 0);
3968 size = raid10_size(mddev, sectors, 0);
3969 if (mddev->external_size &&
3970 mddev->array_sectors > size)
3972 if (mddev->bitmap) {
3973 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3977 md_set_array_sectors(mddev, size);
3978 if (sectors > mddev->dev_sectors &&
3979 mddev->recovery_cp > oldsize) {
3980 mddev->recovery_cp = oldsize;
3981 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3983 calc_sectors(conf, sectors);
3984 mddev->dev_sectors = conf->dev_sectors;
3985 mddev->resync_max_sectors = size;
3989 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3991 struct md_rdev *rdev;
3992 struct r10conf *conf;
3994 if (mddev->degraded > 0) {
3995 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3997 return ERR_PTR(-EINVAL);
3999 sector_div(size, devs);
4001 /* Set new parameters */
4002 mddev->new_level = 10;
4003 /* new layout: far_copies = 1, near_copies = 2 */
4004 mddev->new_layout = (1<<8) + 2;
4005 mddev->new_chunk_sectors = mddev->chunk_sectors;
4006 mddev->delta_disks = mddev->raid_disks;
4007 mddev->raid_disks *= 2;
4008 /* make sure it will be not marked as dirty */
4009 mddev->recovery_cp = MaxSector;
4010 mddev->dev_sectors = size;
4012 conf = setup_conf(mddev);
4013 if (!IS_ERR(conf)) {
4014 rdev_for_each(rdev, mddev)
4015 if (rdev->raid_disk >= 0) {
4016 rdev->new_raid_disk = rdev->raid_disk * 2;
4017 rdev->sectors = size;
4025 static void *raid10_takeover(struct mddev *mddev)
4027 struct r0conf *raid0_conf;
4029 /* raid10 can take over:
4030 * raid0 - providing it has only two drives
4032 if (mddev->level == 0) {
4033 /* for raid0 takeover only one zone is supported */
4034 raid0_conf = mddev->private;
4035 if (raid0_conf->nr_strip_zones > 1) {
4036 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4038 return ERR_PTR(-EINVAL);
4040 return raid10_takeover_raid0(mddev,
4041 raid0_conf->strip_zone->zone_end,
4042 raid0_conf->strip_zone->nb_dev);
4044 return ERR_PTR(-EINVAL);
4047 static int raid10_check_reshape(struct mddev *mddev)
4049 /* Called when there is a request to change
4050 * - layout (to ->new_layout)
4051 * - chunk size (to ->new_chunk_sectors)
4052 * - raid_disks (by delta_disks)
4053 * or when trying to restart a reshape that was ongoing.
4055 * We need to validate the request and possibly allocate
4056 * space if that might be an issue later.
4058 * Currently we reject any reshape of a 'far' mode array,
4059 * allow chunk size to change if new is generally acceptable,
4060 * allow raid_disks to increase, and allow
4061 * a switch between 'near' mode and 'offset' mode.
4063 struct r10conf *conf = mddev->private;
4066 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4069 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4070 /* mustn't change number of copies */
4072 if (geo.far_copies > 1 && !geo.far_offset)
4073 /* Cannot switch to 'far' mode */
4076 if (mddev->array_sectors & geo.chunk_mask)
4077 /* not factor of array size */
4080 if (!enough(conf, -1))
4083 kfree(conf->mirrors_new);
4084 conf->mirrors_new = NULL;
4085 if (mddev->delta_disks > 0) {
4086 /* allocate new 'mirrors' list */
4088 kcalloc(mddev->raid_disks + mddev->delta_disks,
4089 sizeof(struct raid10_info),
4091 if (!conf->mirrors_new)
4098 * Need to check if array has failed when deciding whether to:
4100 * - remove non-faulty devices
4103 * This determination is simple when no reshape is happening.
4104 * However if there is a reshape, we need to carefully check
4105 * both the before and after sections.
4106 * This is because some failed devices may only affect one
4107 * of the two sections, and some non-in_sync devices may
4108 * be insync in the section most affected by failed devices.
4110 static int calc_degraded(struct r10conf *conf)
4112 int degraded, degraded2;
4117 /* 'prev' section first */
4118 for (i = 0; i < conf->prev.raid_disks; i++) {
4119 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4120 if (!rdev || test_bit(Faulty, &rdev->flags))
4122 else if (!test_bit(In_sync, &rdev->flags))
4123 /* When we can reduce the number of devices in
4124 * an array, this might not contribute to
4125 * 'degraded'. It does now.
4130 if (conf->geo.raid_disks == conf->prev.raid_disks)
4134 for (i = 0; i < conf->geo.raid_disks; i++) {
4135 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4136 if (!rdev || test_bit(Faulty, &rdev->flags))
4138 else if (!test_bit(In_sync, &rdev->flags)) {
4139 /* If reshape is increasing the number of devices,
4140 * this section has already been recovered, so
4141 * it doesn't contribute to degraded.
4144 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4149 if (degraded2 > degraded)
4154 static int raid10_start_reshape(struct mddev *mddev)
4156 /* A 'reshape' has been requested. This commits
4157 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4158 * This also checks if there are enough spares and adds them
4160 * We currently require enough spares to make the final
4161 * array non-degraded. We also require that the difference
4162 * between old and new data_offset - on each device - is
4163 * enough that we never risk over-writing.
4166 unsigned long before_length, after_length;
4167 sector_t min_offset_diff = 0;
4170 struct r10conf *conf = mddev->private;
4171 struct md_rdev *rdev;
4175 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4178 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4181 before_length = ((1 << conf->prev.chunk_shift) *
4182 conf->prev.far_copies);
4183 after_length = ((1 << conf->geo.chunk_shift) *
4184 conf->geo.far_copies);
4186 rdev_for_each(rdev, mddev) {
4187 if (!test_bit(In_sync, &rdev->flags)
4188 && !test_bit(Faulty, &rdev->flags))
4190 if (rdev->raid_disk >= 0) {
4191 long long diff = (rdev->new_data_offset
4192 - rdev->data_offset);
4193 if (!mddev->reshape_backwards)
4197 if (first || diff < min_offset_diff)
4198 min_offset_diff = diff;
4203 if (max(before_length, after_length) > min_offset_diff)
4206 if (spares < mddev->delta_disks)
4209 conf->offset_diff = min_offset_diff;
4210 spin_lock_irq(&conf->device_lock);
4211 if (conf->mirrors_new) {
4212 memcpy(conf->mirrors_new, conf->mirrors,
4213 sizeof(struct raid10_info)*conf->prev.raid_disks);
4215 kfree(conf->mirrors_old);
4216 conf->mirrors_old = conf->mirrors;
4217 conf->mirrors = conf->mirrors_new;
4218 conf->mirrors_new = NULL;
4220 setup_geo(&conf->geo, mddev, geo_start);
4222 if (mddev->reshape_backwards) {
4223 sector_t size = raid10_size(mddev, 0, 0);
4224 if (size < mddev->array_sectors) {
4225 spin_unlock_irq(&conf->device_lock);
4226 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4230 mddev->resync_max_sectors = size;
4231 conf->reshape_progress = size;
4233 conf->reshape_progress = 0;
4234 conf->reshape_safe = conf->reshape_progress;
4235 spin_unlock_irq(&conf->device_lock);
4237 if (mddev->delta_disks && mddev->bitmap) {
4238 struct mdp_superblock_1 *sb = NULL;
4239 sector_t oldsize, newsize;
4241 oldsize = raid10_size(mddev, 0, 0);
4242 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4244 if (!mddev_is_clustered(mddev)) {
4245 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4252 rdev_for_each(rdev, mddev) {
4253 if (rdev->raid_disk > -1 &&
4254 !test_bit(Faulty, &rdev->flags))
4255 sb = page_address(rdev->sb_page);
4259 * some node is already performing reshape, and no need to
4260 * call md_bitmap_resize again since it should be called when
4261 * receiving BITMAP_RESIZE msg
4263 if ((sb && (le32_to_cpu(sb->feature_map) &
4264 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4267 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4271 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4273 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4278 if (mddev->delta_disks > 0) {
4279 rdev_for_each(rdev, mddev)
4280 if (rdev->raid_disk < 0 &&
4281 !test_bit(Faulty, &rdev->flags)) {
4282 if (raid10_add_disk(mddev, rdev) == 0) {
4283 if (rdev->raid_disk >=
4284 conf->prev.raid_disks)
4285 set_bit(In_sync, &rdev->flags);
4287 rdev->recovery_offset = 0;
4289 /* Failure here is OK */
4290 sysfs_link_rdev(mddev, rdev);
4292 } else if (rdev->raid_disk >= conf->prev.raid_disks
4293 && !test_bit(Faulty, &rdev->flags)) {
4294 /* This is a spare that was manually added */
4295 set_bit(In_sync, &rdev->flags);
4298 /* When a reshape changes the number of devices,
4299 * ->degraded is measured against the larger of the
4300 * pre and post numbers.
4302 spin_lock_irq(&conf->device_lock);
4303 mddev->degraded = calc_degraded(conf);
4304 spin_unlock_irq(&conf->device_lock);
4305 mddev->raid_disks = conf->geo.raid_disks;
4306 mddev->reshape_position = conf->reshape_progress;
4307 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4309 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4310 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4311 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4312 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4313 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4315 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4317 if (!mddev->sync_thread) {
4321 conf->reshape_checkpoint = jiffies;
4322 md_wakeup_thread(mddev->sync_thread);
4323 md_new_event(mddev);
4327 mddev->recovery = 0;
4328 spin_lock_irq(&conf->device_lock);
4329 conf->geo = conf->prev;
4330 mddev->raid_disks = conf->geo.raid_disks;
4331 rdev_for_each(rdev, mddev)
4332 rdev->new_data_offset = rdev->data_offset;
4334 conf->reshape_progress = MaxSector;
4335 conf->reshape_safe = MaxSector;
4336 mddev->reshape_position = MaxSector;
4337 spin_unlock_irq(&conf->device_lock);
4341 /* Calculate the last device-address that could contain
4342 * any block from the chunk that includes the array-address 's'
4343 * and report the next address.
4344 * i.e. the address returned will be chunk-aligned and after
4345 * any data that is in the chunk containing 's'.
4347 static sector_t last_dev_address(sector_t s, struct geom *geo)
4349 s = (s | geo->chunk_mask) + 1;
4350 s >>= geo->chunk_shift;
4351 s *= geo->near_copies;
4352 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4353 s *= geo->far_copies;
4354 s <<= geo->chunk_shift;
4358 /* Calculate the first device-address that could contain
4359 * any block from the chunk that includes the array-address 's'.
4360 * This too will be the start of a chunk
4362 static sector_t first_dev_address(sector_t s, struct geom *geo)
4364 s >>= geo->chunk_shift;
4365 s *= geo->near_copies;
4366 sector_div(s, geo->raid_disks);
4367 s *= geo->far_copies;
4368 s <<= geo->chunk_shift;
4372 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4375 /* We simply copy at most one chunk (smallest of old and new)
4376 * at a time, possibly less if that exceeds RESYNC_PAGES,
4377 * or we hit a bad block or something.
4378 * This might mean we pause for normal IO in the middle of
4379 * a chunk, but that is not a problem as mddev->reshape_position
4380 * can record any location.
4382 * If we will want to write to a location that isn't
4383 * yet recorded as 'safe' (i.e. in metadata on disk) then
4384 * we need to flush all reshape requests and update the metadata.
4386 * When reshaping forwards (e.g. to more devices), we interpret
4387 * 'safe' as the earliest block which might not have been copied
4388 * down yet. We divide this by previous stripe size and multiply
4389 * by previous stripe length to get lowest device offset that we
4390 * cannot write to yet.
4391 * We interpret 'sector_nr' as an address that we want to write to.
4392 * From this we use last_device_address() to find where we might
4393 * write to, and first_device_address on the 'safe' position.
4394 * If this 'next' write position is after the 'safe' position,
4395 * we must update the metadata to increase the 'safe' position.
4397 * When reshaping backwards, we round in the opposite direction
4398 * and perform the reverse test: next write position must not be
4399 * less than current safe position.
4401 * In all this the minimum difference in data offsets
4402 * (conf->offset_diff - always positive) allows a bit of slack,
4403 * so next can be after 'safe', but not by more than offset_diff
4405 * We need to prepare all the bios here before we start any IO
4406 * to ensure the size we choose is acceptable to all devices.
4407 * The means one for each copy for write-out and an extra one for
4409 * We store the read-in bio in ->master_bio and the others in
4410 * ->devs[x].bio and ->devs[x].repl_bio.
4412 struct r10conf *conf = mddev->private;
4413 struct r10bio *r10_bio;
4414 sector_t next, safe, last;
4418 struct md_rdev *rdev;
4421 struct bio *bio, *read_bio;
4422 int sectors_done = 0;
4423 struct page **pages;
4425 if (sector_nr == 0) {
4426 /* If restarting in the middle, skip the initial sectors */
4427 if (mddev->reshape_backwards &&
4428 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4429 sector_nr = (raid10_size(mddev, 0, 0)
4430 - conf->reshape_progress);
4431 } else if (!mddev->reshape_backwards &&
4432 conf->reshape_progress > 0)
4433 sector_nr = conf->reshape_progress;
4435 mddev->curr_resync_completed = sector_nr;
4436 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4442 /* We don't use sector_nr to track where we are up to
4443 * as that doesn't work well for ->reshape_backwards.
4444 * So just use ->reshape_progress.
4446 if (mddev->reshape_backwards) {
4447 /* 'next' is the earliest device address that we might
4448 * write to for this chunk in the new layout
4450 next = first_dev_address(conf->reshape_progress - 1,
4453 /* 'safe' is the last device address that we might read from
4454 * in the old layout after a restart
4456 safe = last_dev_address(conf->reshape_safe - 1,
4459 if (next + conf->offset_diff < safe)
4462 last = conf->reshape_progress - 1;
4463 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4464 & conf->prev.chunk_mask);
4465 if (sector_nr + RESYNC_SECTORS < last)
4466 sector_nr = last + 1 - RESYNC_SECTORS;
4468 /* 'next' is after the last device address that we
4469 * might write to for this chunk in the new layout
4471 next = last_dev_address(conf->reshape_progress, &conf->geo);
4473 /* 'safe' is the earliest device address that we might
4474 * read from in the old layout after a restart
4476 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4478 /* Need to update metadata if 'next' might be beyond 'safe'
4479 * as that would possibly corrupt data
4481 if (next > safe + conf->offset_diff)
4484 sector_nr = conf->reshape_progress;
4485 last = sector_nr | (conf->geo.chunk_mask
4486 & conf->prev.chunk_mask);
4488 if (sector_nr + RESYNC_SECTORS <= last)
4489 last = sector_nr + RESYNC_SECTORS - 1;
4493 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4494 /* Need to update reshape_position in metadata */
4496 mddev->reshape_position = conf->reshape_progress;
4497 if (mddev->reshape_backwards)
4498 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4499 - conf->reshape_progress;
4501 mddev->curr_resync_completed = conf->reshape_progress;
4502 conf->reshape_checkpoint = jiffies;
4503 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4504 md_wakeup_thread(mddev->thread);
4505 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4506 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4507 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4508 allow_barrier(conf);
4509 return sectors_done;
4511 conf->reshape_safe = mddev->reshape_position;
4512 allow_barrier(conf);
4515 raise_barrier(conf, 0);
4517 /* Now schedule reads for blocks from sector_nr to last */
4518 r10_bio = raid10_alloc_init_r10buf(conf);
4520 raise_barrier(conf, 1);
4521 atomic_set(&r10_bio->remaining, 0);
4522 r10_bio->mddev = mddev;
4523 r10_bio->sector = sector_nr;
4524 set_bit(R10BIO_IsReshape, &r10_bio->state);
4525 r10_bio->sectors = last - sector_nr + 1;
4526 rdev = read_balance(conf, r10_bio, &max_sectors);
4527 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4530 /* Cannot read from here, so need to record bad blocks
4531 * on all the target devices.
4534 mempool_free(r10_bio, &conf->r10buf_pool);
4535 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4536 return sectors_done;
4539 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4541 bio_set_dev(read_bio, rdev->bdev);
4542 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4543 + rdev->data_offset);
4544 read_bio->bi_private = r10_bio;
4545 read_bio->bi_end_io = end_reshape_read;
4546 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4547 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4548 read_bio->bi_status = 0;
4549 read_bio->bi_vcnt = 0;
4550 read_bio->bi_iter.bi_size = 0;
4551 r10_bio->master_bio = read_bio;
4552 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4555 * Broadcast RESYNC message to other nodes, so all nodes would not
4556 * write to the region to avoid conflict.
4558 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4559 struct mdp_superblock_1 *sb = NULL;
4560 int sb_reshape_pos = 0;
4562 conf->cluster_sync_low = sector_nr;
4563 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4564 sb = page_address(rdev->sb_page);
4566 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4568 * Set cluster_sync_low again if next address for array
4569 * reshape is less than cluster_sync_low. Since we can't
4570 * update cluster_sync_low until it has finished reshape.
4572 if (sb_reshape_pos < conf->cluster_sync_low)
4573 conf->cluster_sync_low = sb_reshape_pos;
4576 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4577 conf->cluster_sync_high);
4580 /* Now find the locations in the new layout */
4581 __raid10_find_phys(&conf->geo, r10_bio);
4584 read_bio->bi_next = NULL;
4587 for (s = 0; s < conf->copies*2; s++) {
4589 int d = r10_bio->devs[s/2].devnum;
4590 struct md_rdev *rdev2;
4592 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4593 b = r10_bio->devs[s/2].repl_bio;
4595 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4596 b = r10_bio->devs[s/2].bio;
4598 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4601 bio_set_dev(b, rdev2->bdev);
4602 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4603 rdev2->new_data_offset;
4604 b->bi_end_io = end_reshape_write;
4605 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4610 /* Now add as many pages as possible to all of these bios. */
4613 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4614 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4615 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4616 int len = (max_sectors - s) << 9;
4617 if (len > PAGE_SIZE)
4619 for (bio = blist; bio ; bio = bio->bi_next) {
4621 * won't fail because the vec table is big enough
4622 * to hold all these pages
4624 bio_add_page(bio, page, len, 0);
4626 sector_nr += len >> 9;
4627 nr_sectors += len >> 9;
4630 r10_bio->sectors = nr_sectors;
4632 /* Now submit the read */
4633 md_sync_acct_bio(read_bio, r10_bio->sectors);
4634 atomic_inc(&r10_bio->remaining);
4635 read_bio->bi_next = NULL;
4636 submit_bio_noacct(read_bio);
4637 sectors_done += nr_sectors;
4638 if (sector_nr <= last)
4641 lower_barrier(conf);
4643 /* Now that we have done the whole section we can
4644 * update reshape_progress
4646 if (mddev->reshape_backwards)
4647 conf->reshape_progress -= sectors_done;
4649 conf->reshape_progress += sectors_done;
4651 return sectors_done;
4654 static void end_reshape_request(struct r10bio *r10_bio);
4655 static int handle_reshape_read_error(struct mddev *mddev,
4656 struct r10bio *r10_bio);
4657 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4659 /* Reshape read completed. Hopefully we have a block
4661 * If we got a read error then we do sync 1-page reads from
4662 * elsewhere until we find the data - or give up.
4664 struct r10conf *conf = mddev->private;
4667 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4668 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4669 /* Reshape has been aborted */
4670 md_done_sync(mddev, r10_bio->sectors, 0);
4674 /* We definitely have the data in the pages, schedule the
4677 atomic_set(&r10_bio->remaining, 1);
4678 for (s = 0; s < conf->copies*2; s++) {
4680 int d = r10_bio->devs[s/2].devnum;
4681 struct md_rdev *rdev;
4684 rdev = rcu_dereference(conf->mirrors[d].replacement);
4685 b = r10_bio->devs[s/2].repl_bio;
4687 rdev = rcu_dereference(conf->mirrors[d].rdev);
4688 b = r10_bio->devs[s/2].bio;
4690 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4694 atomic_inc(&rdev->nr_pending);
4696 md_sync_acct_bio(b, r10_bio->sectors);
4697 atomic_inc(&r10_bio->remaining);
4699 submit_bio_noacct(b);
4701 end_reshape_request(r10_bio);
4704 static void end_reshape(struct r10conf *conf)
4706 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4709 spin_lock_irq(&conf->device_lock);
4710 conf->prev = conf->geo;
4711 md_finish_reshape(conf->mddev);
4713 conf->reshape_progress = MaxSector;
4714 conf->reshape_safe = MaxSector;
4715 spin_unlock_irq(&conf->device_lock);
4717 if (conf->mddev->queue)
4718 raid10_set_io_opt(conf);
4722 static void raid10_update_reshape_pos(struct mddev *mddev)
4724 struct r10conf *conf = mddev->private;
4727 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4728 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4729 || mddev->reshape_position == MaxSector)
4730 conf->reshape_progress = mddev->reshape_position;
4735 static int handle_reshape_read_error(struct mddev *mddev,
4736 struct r10bio *r10_bio)
4738 /* Use sync reads to get the blocks from somewhere else */
4739 int sectors = r10_bio->sectors;
4740 struct r10conf *conf = mddev->private;
4741 struct r10bio *r10b;
4744 struct page **pages;
4746 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4748 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4752 /* reshape IOs share pages from .devs[0].bio */
4753 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4755 r10b->sector = r10_bio->sector;
4756 __raid10_find_phys(&conf->prev, r10b);
4761 int first_slot = slot;
4763 if (s > (PAGE_SIZE >> 9))
4768 int d = r10b->devs[slot].devnum;
4769 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4772 test_bit(Faulty, &rdev->flags) ||
4773 !test_bit(In_sync, &rdev->flags))
4776 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4777 atomic_inc(&rdev->nr_pending);
4779 success = sync_page_io(rdev,
4783 REQ_OP_READ, 0, false);
4784 rdev_dec_pending(rdev, mddev);
4790 if (slot >= conf->copies)
4792 if (slot == first_slot)
4797 /* couldn't read this block, must give up */
4798 set_bit(MD_RECOVERY_INTR,
4810 static void end_reshape_write(struct bio *bio)
4812 struct r10bio *r10_bio = get_resync_r10bio(bio);
4813 struct mddev *mddev = r10_bio->mddev;
4814 struct r10conf *conf = mddev->private;
4818 struct md_rdev *rdev = NULL;
4820 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4822 rdev = conf->mirrors[d].replacement;
4825 rdev = conf->mirrors[d].rdev;
4828 if (bio->bi_status) {
4829 /* FIXME should record badblock */
4830 md_error(mddev, rdev);
4833 rdev_dec_pending(rdev, mddev);
4834 end_reshape_request(r10_bio);
4837 static void end_reshape_request(struct r10bio *r10_bio)
4839 if (!atomic_dec_and_test(&r10_bio->remaining))
4841 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4842 bio_put(r10_bio->master_bio);
4846 static void raid10_finish_reshape(struct mddev *mddev)
4848 struct r10conf *conf = mddev->private;
4850 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4853 if (mddev->delta_disks > 0) {
4854 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4855 mddev->recovery_cp = mddev->resync_max_sectors;
4856 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4858 mddev->resync_max_sectors = mddev->array_sectors;
4862 for (d = conf->geo.raid_disks ;
4863 d < conf->geo.raid_disks - mddev->delta_disks;
4865 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4867 clear_bit(In_sync, &rdev->flags);
4868 rdev = rcu_dereference(conf->mirrors[d].replacement);
4870 clear_bit(In_sync, &rdev->flags);
4874 mddev->layout = mddev->new_layout;
4875 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4876 mddev->reshape_position = MaxSector;
4877 mddev->delta_disks = 0;
4878 mddev->reshape_backwards = 0;
4881 static struct md_personality raid10_personality =
4885 .owner = THIS_MODULE,
4886 .make_request = raid10_make_request,
4888 .free = raid10_free,
4889 .status = raid10_status,
4890 .error_handler = raid10_error,
4891 .hot_add_disk = raid10_add_disk,
4892 .hot_remove_disk= raid10_remove_disk,
4893 .spare_active = raid10_spare_active,
4894 .sync_request = raid10_sync_request,
4895 .quiesce = raid10_quiesce,
4896 .size = raid10_size,
4897 .resize = raid10_resize,
4898 .takeover = raid10_takeover,
4899 .check_reshape = raid10_check_reshape,
4900 .start_reshape = raid10_start_reshape,
4901 .finish_reshape = raid10_finish_reshape,
4902 .update_reshape_pos = raid10_update_reshape_pos,
4905 static int __init raid_init(void)
4907 return register_md_personality(&raid10_personality);
4910 static void raid_exit(void)
4912 unregister_md_personality(&raid10_personality);
4915 module_init(raid_init);
4916 module_exit(raid_exit);
4917 MODULE_LICENSE("GPL");
4918 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4919 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4920 MODULE_ALIAS("md-raid10");
4921 MODULE_ALIAS("md-level-10");
4923 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);