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>
23 #define RAID_1_10_NAME "raid10"
26 #include "md-bitmap.h"
29 * RAID10 provides a combination of RAID0 and RAID1 functionality.
30 * The layout of data is defined by
33 * near_copies (stored in low byte of layout)
34 * far_copies (stored in second byte of layout)
35 * far_offset (stored in bit 16 of layout )
36 * use_far_sets (stored in bit 17 of layout )
37 * use_far_sets_bugfixed (stored in bit 18 of layout )
39 * The data to be stored is divided into chunks using chunksize. Each device
40 * is divided into far_copies sections. In each section, chunks are laid out
41 * in a style similar to raid0, but near_copies copies of each chunk is stored
42 * (each on a different drive). The starting device for each section is offset
43 * near_copies from the starting device of the previous section. Thus there
44 * are (near_copies * far_copies) of each chunk, and each is on a different
45 * drive. near_copies and far_copies must be at least one, and their product
46 * is at most raid_disks.
48 * If far_offset is true, then the far_copies are handled a bit differently.
49 * The copies are still in different stripes, but instead of being very far
50 * apart on disk, there are adjacent stripes.
52 * The far and offset algorithms are handled slightly differently if
53 * 'use_far_sets' is true. In this case, the array's devices are grouped into
54 * sets that are (near_copies * far_copies) in size. The far copied stripes
55 * are still shifted by 'near_copies' devices, but this shifting stays confined
56 * to the set rather than the entire array. This is done to improve the number
57 * of device combinations that can fail without causing the array to fail.
58 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
64 * [A B] [C D] [A B] [C D E]
65 * |...| |...| |...| | ... |
66 * [B A] [D C] [B A] [E C D]
69 static void allow_barrier(struct r10conf *conf);
70 static void lower_barrier(struct r10conf *conf);
71 static int _enough(struct r10conf *conf, int previous, int ignore);
72 static int enough(struct r10conf *conf, int ignore);
73 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
75 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
76 static void end_reshape_write(struct bio *bio);
77 static void end_reshape(struct r10conf *conf);
82 #define cmd_before(conf, cmd) \
84 write_sequnlock_irq(&(conf)->resync_lock); \
87 #define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
89 #define wait_event_barrier_cmd(conf, cond, cmd) \
90 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
93 #define wait_event_barrier(conf, cond) \
94 wait_event_barrier_cmd(conf, cond, NULL_CMD)
97 * for resync bio, r10bio pointer can be retrieved from the per-bio
98 * 'struct resync_pages'.
100 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
102 return get_resync_pages(bio)->raid_bio;
105 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
107 struct r10conf *conf = data;
108 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
110 /* allocate a r10bio with room for raid_disks entries in the
112 return kzalloc(size, gfp_flags);
115 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
116 /* amount of memory to reserve for resync requests */
117 #define RESYNC_WINDOW (1024*1024)
118 /* maximum number of concurrent requests, memory permitting */
119 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
120 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
121 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
124 * When performing a resync, we need to read and compare, so
125 * we need as many pages are there are copies.
126 * When performing a recovery, we need 2 bios, one for read,
127 * one for write (we recover only one drive per r10buf)
130 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
132 struct r10conf *conf = data;
133 struct r10bio *r10_bio;
136 int nalloc, nalloc_rp;
137 struct resync_pages *rps;
139 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
143 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
144 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
145 nalloc = conf->copies; /* resync */
147 nalloc = 2; /* recovery */
149 /* allocate once for all bios */
150 if (!conf->have_replacement)
153 nalloc_rp = nalloc * 2;
154 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
156 goto out_free_r10bio;
161 for (j = nalloc ; j-- ; ) {
162 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
165 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
166 r10_bio->devs[j].bio = bio;
167 if (!conf->have_replacement)
169 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
172 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
173 r10_bio->devs[j].repl_bio = bio;
176 * Allocate RESYNC_PAGES data pages and attach them
179 for (j = 0; j < nalloc; j++) {
180 struct bio *rbio = r10_bio->devs[j].repl_bio;
181 struct resync_pages *rp, *rp_repl;
185 rp_repl = &rps[nalloc + j];
187 bio = r10_bio->devs[j].bio;
189 if (!j || test_bit(MD_RECOVERY_SYNC,
190 &conf->mddev->recovery)) {
191 if (resync_alloc_pages(rp, gfp_flags))
194 memcpy(rp, &rps[0], sizeof(*rp));
195 resync_get_all_pages(rp);
198 rp->raid_bio = r10_bio;
199 bio->bi_private = rp;
201 memcpy(rp_repl, rp, sizeof(*rp));
202 rbio->bi_private = rp_repl;
210 resync_free_pages(&rps[j]);
214 for ( ; j < nalloc; j++) {
215 if (r10_bio->devs[j].bio)
216 bio_uninit(r10_bio->devs[j].bio);
217 kfree(r10_bio->devs[j].bio);
218 if (r10_bio->devs[j].repl_bio)
219 bio_uninit(r10_bio->devs[j].repl_bio);
220 kfree(r10_bio->devs[j].repl_bio);
224 rbio_pool_free(r10_bio, conf);
228 static void r10buf_pool_free(void *__r10_bio, void *data)
230 struct r10conf *conf = data;
231 struct r10bio *r10bio = __r10_bio;
233 struct resync_pages *rp = NULL;
235 for (j = conf->copies; j--; ) {
236 struct bio *bio = r10bio->devs[j].bio;
239 rp = get_resync_pages(bio);
240 resync_free_pages(rp);
245 bio = r10bio->devs[j].repl_bio;
252 /* resync pages array stored in the 1st bio's .bi_private */
255 rbio_pool_free(r10bio, conf);
258 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
262 for (i = 0; i < conf->geo.raid_disks; i++) {
263 struct bio **bio = & r10_bio->devs[i].bio;
264 if (!BIO_SPECIAL(*bio))
267 bio = &r10_bio->devs[i].repl_bio;
268 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
274 static void free_r10bio(struct r10bio *r10_bio)
276 struct r10conf *conf = r10_bio->mddev->private;
278 put_all_bios(conf, r10_bio);
279 mempool_free(r10_bio, &conf->r10bio_pool);
282 static void put_buf(struct r10bio *r10_bio)
284 struct r10conf *conf = r10_bio->mddev->private;
286 mempool_free(r10_bio, &conf->r10buf_pool);
291 static void wake_up_barrier(struct r10conf *conf)
293 if (wq_has_sleeper(&conf->wait_barrier))
294 wake_up(&conf->wait_barrier);
297 static void reschedule_retry(struct r10bio *r10_bio)
300 struct mddev *mddev = r10_bio->mddev;
301 struct r10conf *conf = mddev->private;
303 spin_lock_irqsave(&conf->device_lock, flags);
304 list_add(&r10_bio->retry_list, &conf->retry_list);
306 spin_unlock_irqrestore(&conf->device_lock, flags);
308 /* wake up frozen array... */
309 wake_up(&conf->wait_barrier);
311 md_wakeup_thread(mddev->thread);
315 * raid_end_bio_io() is called when we have finished servicing a mirrored
316 * operation and are ready to return a success/failure code to the buffer
319 static void raid_end_bio_io(struct r10bio *r10_bio)
321 struct bio *bio = r10_bio->master_bio;
322 struct r10conf *conf = r10_bio->mddev->private;
324 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
325 bio->bi_status = BLK_STS_IOERR;
329 * Wake up any possible resync thread that waits for the device
334 free_r10bio(r10_bio);
338 * Update disk head position estimator based on IRQ completion info.
340 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
342 struct r10conf *conf = r10_bio->mddev->private;
344 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
345 r10_bio->devs[slot].addr + (r10_bio->sectors);
349 * Find the disk number which triggered given bio
351 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
352 struct bio *bio, int *slotp, int *replp)
357 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
358 if (r10_bio->devs[slot].bio == bio)
360 if (r10_bio->devs[slot].repl_bio == bio) {
366 update_head_pos(slot, r10_bio);
372 return r10_bio->devs[slot].devnum;
375 static void raid10_end_read_request(struct bio *bio)
377 int uptodate = !bio->bi_status;
378 struct r10bio *r10_bio = bio->bi_private;
380 struct md_rdev *rdev;
381 struct r10conf *conf = r10_bio->mddev->private;
383 slot = r10_bio->read_slot;
384 rdev = r10_bio->devs[slot].rdev;
386 * this branch is our 'one mirror IO has finished' event handler:
388 update_head_pos(slot, r10_bio);
392 * Set R10BIO_Uptodate in our master bio, so that
393 * we will return a good error code to the higher
394 * levels even if IO on some other mirrored buffer fails.
396 * The 'master' represents the composite IO operation to
397 * user-side. So if something waits for IO, then it will
398 * wait for the 'master' bio.
400 set_bit(R10BIO_Uptodate, &r10_bio->state);
402 /* If all other devices that store this block have
403 * failed, we want to return the error upwards rather
404 * than fail the last device. Here we redefine
405 * "uptodate" to mean "Don't want to retry"
407 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
412 raid_end_bio_io(r10_bio);
413 rdev_dec_pending(rdev, conf->mddev);
416 * oops, read error - keep the refcount on the rdev
418 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
421 (unsigned long long)r10_bio->sector);
422 set_bit(R10BIO_ReadError, &r10_bio->state);
423 reschedule_retry(r10_bio);
427 static void close_write(struct r10bio *r10_bio)
429 /* clear the bitmap if all writes complete successfully */
430 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
432 !test_bit(R10BIO_Degraded, &r10_bio->state),
434 md_write_end(r10_bio->mddev);
437 static void one_write_done(struct r10bio *r10_bio)
439 if (atomic_dec_and_test(&r10_bio->remaining)) {
440 if (test_bit(R10BIO_WriteError, &r10_bio->state))
441 reschedule_retry(r10_bio);
443 close_write(r10_bio);
444 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
445 reschedule_retry(r10_bio);
447 raid_end_bio_io(r10_bio);
452 static void raid10_end_write_request(struct bio *bio)
454 struct r10bio *r10_bio = bio->bi_private;
457 struct r10conf *conf = r10_bio->mddev->private;
459 struct md_rdev *rdev = NULL;
460 struct bio *to_put = NULL;
463 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
465 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
468 rdev = conf->mirrors[dev].replacement;
472 rdev = conf->mirrors[dev].rdev;
475 * this branch is our 'one mirror IO has finished' event handler:
477 if (bio->bi_status && !discard_error) {
479 /* Never record new bad blocks to replacement,
482 md_error(rdev->mddev, rdev);
484 set_bit(WriteErrorSeen, &rdev->flags);
485 if (!test_and_set_bit(WantReplacement, &rdev->flags))
486 set_bit(MD_RECOVERY_NEEDED,
487 &rdev->mddev->recovery);
490 if (test_bit(FailFast, &rdev->flags) &&
491 (bio->bi_opf & MD_FAILFAST)) {
492 md_error(rdev->mddev, rdev);
496 * When the device is faulty, it is not necessary to
497 * handle write error.
499 if (!test_bit(Faulty, &rdev->flags))
500 set_bit(R10BIO_WriteError, &r10_bio->state);
502 /* Fail the request */
503 set_bit(R10BIO_Degraded, &r10_bio->state);
504 r10_bio->devs[slot].bio = NULL;
511 * Set R10BIO_Uptodate in our master bio, so that
512 * we will return a good error code for to the higher
513 * levels even if IO on some other mirrored buffer fails.
515 * The 'master' represents the composite IO operation to
516 * user-side. So if something waits for IO, then it will
517 * wait for the 'master' bio.
519 * Do not set R10BIO_Uptodate if the current device is
520 * rebuilding or Faulty. This is because we cannot use
521 * such device for properly reading the data back (we could
522 * potentially use it, if the current write would have felt
523 * before rdev->recovery_offset, but for simplicity we don't
526 if (test_bit(In_sync, &rdev->flags) &&
527 !test_bit(Faulty, &rdev->flags))
528 set_bit(R10BIO_Uptodate, &r10_bio->state);
530 /* Maybe we can clear some bad blocks. */
531 if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
536 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
538 r10_bio->devs[slot].bio = IO_MADE_GOOD;
540 set_bit(R10BIO_MadeGood, &r10_bio->state);
546 * Let's see if all mirrored write operations have finished
549 one_write_done(r10_bio);
551 rdev_dec_pending(rdev, conf->mddev);
557 * RAID10 layout manager
558 * As well as the chunksize and raid_disks count, there are two
559 * parameters: near_copies and far_copies.
560 * near_copies * far_copies must be <= raid_disks.
561 * Normally one of these will be 1.
562 * If both are 1, we get raid0.
563 * If near_copies == raid_disks, we get raid1.
565 * Chunks are laid out in raid0 style with near_copies copies of the
566 * first chunk, followed by near_copies copies of the next chunk and
568 * If far_copies > 1, then after 1/far_copies of the array has been assigned
569 * as described above, we start again with a device offset of near_copies.
570 * So we effectively have another copy of the whole array further down all
571 * the drives, but with blocks on different drives.
572 * With this layout, and block is never stored twice on the one device.
574 * raid10_find_phys finds the sector offset of a given virtual sector
575 * on each device that it is on.
577 * raid10_find_virt does the reverse mapping, from a device and a
578 * sector offset to a virtual address
581 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
589 int last_far_set_start, last_far_set_size;
591 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
592 last_far_set_start *= geo->far_set_size;
594 last_far_set_size = geo->far_set_size;
595 last_far_set_size += (geo->raid_disks % geo->far_set_size);
597 /* now calculate first sector/dev */
598 chunk = r10bio->sector >> geo->chunk_shift;
599 sector = r10bio->sector & geo->chunk_mask;
601 chunk *= geo->near_copies;
603 dev = sector_div(stripe, geo->raid_disks);
605 stripe *= geo->far_copies;
607 sector += stripe << geo->chunk_shift;
609 /* and calculate all the others */
610 for (n = 0; n < geo->near_copies; n++) {
614 r10bio->devs[slot].devnum = d;
615 r10bio->devs[slot].addr = s;
618 for (f = 1; f < geo->far_copies; f++) {
619 set = d / geo->far_set_size;
620 d += geo->near_copies;
622 if ((geo->raid_disks % geo->far_set_size) &&
623 (d > last_far_set_start)) {
624 d -= last_far_set_start;
625 d %= last_far_set_size;
626 d += last_far_set_start;
628 d %= geo->far_set_size;
629 d += geo->far_set_size * set;
632 r10bio->devs[slot].devnum = d;
633 r10bio->devs[slot].addr = s;
637 if (dev >= geo->raid_disks) {
639 sector += (geo->chunk_mask + 1);
644 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
646 struct geom *geo = &conf->geo;
648 if (conf->reshape_progress != MaxSector &&
649 ((r10bio->sector >= conf->reshape_progress) !=
650 conf->mddev->reshape_backwards)) {
651 set_bit(R10BIO_Previous, &r10bio->state);
654 clear_bit(R10BIO_Previous, &r10bio->state);
656 __raid10_find_phys(geo, r10bio);
659 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
661 sector_t offset, chunk, vchunk;
662 /* Never use conf->prev as this is only called during resync
663 * or recovery, so reshape isn't happening
665 struct geom *geo = &conf->geo;
666 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
667 int far_set_size = geo->far_set_size;
668 int last_far_set_start;
670 if (geo->raid_disks % geo->far_set_size) {
671 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
672 last_far_set_start *= geo->far_set_size;
674 if (dev >= last_far_set_start) {
675 far_set_size = geo->far_set_size;
676 far_set_size += (geo->raid_disks % geo->far_set_size);
677 far_set_start = last_far_set_start;
681 offset = sector & geo->chunk_mask;
682 if (geo->far_offset) {
684 chunk = sector >> geo->chunk_shift;
685 fc = sector_div(chunk, geo->far_copies);
686 dev -= fc * geo->near_copies;
687 if (dev < far_set_start)
690 while (sector >= geo->stride) {
691 sector -= geo->stride;
692 if (dev < (geo->near_copies + far_set_start))
693 dev += far_set_size - geo->near_copies;
695 dev -= geo->near_copies;
697 chunk = sector >> geo->chunk_shift;
699 vchunk = chunk * geo->raid_disks + dev;
700 sector_div(vchunk, geo->near_copies);
701 return (vchunk << geo->chunk_shift) + offset;
705 * This routine returns the disk from which the requested read should
706 * be done. There is a per-array 'next expected sequential IO' sector
707 * number - if this matches on the next IO then we use the last disk.
708 * There is also a per-disk 'last know head position' sector that is
709 * maintained from IRQ contexts, both the normal and the resync IO
710 * completion handlers update this position correctly. If there is no
711 * perfect sequential match then we pick the disk whose head is closest.
713 * If there are 2 mirrors in the same 2 devices, performance degrades
714 * because position is mirror, not device based.
716 * The rdev for the device selected will have nr_pending incremented.
720 * FIXME: possibly should rethink readbalancing and do it differently
721 * depending on near_copies / far_copies geometry.
723 static struct md_rdev *read_balance(struct r10conf *conf,
724 struct r10bio *r10_bio,
727 const sector_t this_sector = r10_bio->sector;
729 int sectors = r10_bio->sectors;
730 int best_good_sectors;
731 sector_t new_distance, best_dist;
732 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
734 int best_dist_slot, best_pending_slot;
735 bool has_nonrot_disk = false;
736 unsigned int min_pending;
737 struct geom *geo = &conf->geo;
739 raid10_find_phys(conf, r10_bio);
741 min_pending = UINT_MAX;
742 best_dist_rdev = NULL;
743 best_pending_rdev = NULL;
744 best_dist = MaxSector;
745 best_good_sectors = 0;
747 clear_bit(R10BIO_FailFast, &r10_bio->state);
749 if (raid1_should_read_first(conf->mddev, this_sector, sectors))
752 for (slot = 0; slot < conf->copies ; slot++) {
756 unsigned int pending;
759 if (r10_bio->devs[slot].bio == IO_BLOCKED)
761 disk = r10_bio->devs[slot].devnum;
762 rdev = conf->mirrors[disk].replacement;
763 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
764 r10_bio->devs[slot].addr + sectors >
765 rdev->recovery_offset)
766 rdev = conf->mirrors[disk].rdev;
768 test_bit(Faulty, &rdev->flags))
770 if (!test_bit(In_sync, &rdev->flags) &&
771 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
774 dev_sector = r10_bio->devs[slot].addr;
775 if (is_badblock(rdev, dev_sector, sectors,
776 &first_bad, &bad_sectors)) {
777 if (best_dist < MaxSector)
778 /* Already have a better slot */
780 if (first_bad <= dev_sector) {
781 /* Cannot read here. If this is the
782 * 'primary' device, then we must not read
783 * beyond 'bad_sectors' from another device.
785 bad_sectors -= (dev_sector - first_bad);
786 if (!do_balance && sectors > bad_sectors)
787 sectors = bad_sectors;
788 if (best_good_sectors > sectors)
789 best_good_sectors = sectors;
791 sector_t good_sectors =
792 first_bad - dev_sector;
793 if (good_sectors > best_good_sectors) {
794 best_good_sectors = good_sectors;
795 best_dist_slot = slot;
796 best_dist_rdev = rdev;
799 /* Must read from here */
804 best_good_sectors = sectors;
809 nonrot = bdev_nonrot(rdev->bdev);
810 has_nonrot_disk |= nonrot;
811 pending = atomic_read(&rdev->nr_pending);
812 if (min_pending > pending && nonrot) {
813 min_pending = pending;
814 best_pending_slot = slot;
815 best_pending_rdev = rdev;
818 if (best_dist_slot >= 0)
819 /* At least 2 disks to choose from so failfast is OK */
820 set_bit(R10BIO_FailFast, &r10_bio->state);
821 /* This optimisation is debatable, and completely destroys
822 * sequential read speed for 'far copies' arrays. So only
823 * keep it for 'near' arrays, and review those later.
825 if (geo->near_copies > 1 && !pending)
828 /* for far > 1 always use the lowest address */
829 else if (geo->far_copies > 1)
830 new_distance = r10_bio->devs[slot].addr;
832 new_distance = abs(r10_bio->devs[slot].addr -
833 conf->mirrors[disk].head_position);
835 if (new_distance < best_dist) {
836 best_dist = new_distance;
837 best_dist_slot = slot;
838 best_dist_rdev = rdev;
841 if (slot >= conf->copies) {
842 if (has_nonrot_disk) {
843 slot = best_pending_slot;
844 rdev = best_pending_rdev;
846 slot = best_dist_slot;
847 rdev = best_dist_rdev;
852 atomic_inc(&rdev->nr_pending);
853 r10_bio->read_slot = slot;
856 *max_sectors = best_good_sectors;
861 static void flush_pending_writes(struct r10conf *conf)
863 /* Any writes that have been queued but are awaiting
864 * bitmap updates get flushed here.
866 spin_lock_irq(&conf->device_lock);
868 if (conf->pending_bio_list.head) {
869 struct blk_plug plug;
872 bio = bio_list_get(&conf->pending_bio_list);
873 spin_unlock_irq(&conf->device_lock);
876 * As this is called in a wait_event() loop (see freeze_array),
877 * current->state might be TASK_UNINTERRUPTIBLE which will
878 * cause a warning when we prepare to wait again. As it is
879 * rare that this path is taken, it is perfectly safe to force
880 * us to go around the wait_event() loop again, so the warning
881 * is a false-positive. Silence the warning by resetting
884 __set_current_state(TASK_RUNNING);
886 blk_start_plug(&plug);
887 raid1_prepare_flush_writes(conf->mddev->bitmap);
888 wake_up(&conf->wait_barrier);
890 while (bio) { /* submit pending writes */
891 struct bio *next = bio->bi_next;
893 raid1_submit_write(bio);
897 blk_finish_plug(&plug);
899 spin_unlock_irq(&conf->device_lock);
903 * Sometimes we need to suspend IO while we do something else,
904 * either some resync/recovery, or reconfigure the array.
905 * To do this we raise a 'barrier'.
906 * The 'barrier' is a counter that can be raised multiple times
907 * to count how many activities are happening which preclude
909 * We can only raise the barrier if there is no pending IO.
910 * i.e. if nr_pending == 0.
911 * We choose only to raise the barrier if no-one is waiting for the
912 * barrier to go down. This means that as soon as an IO request
913 * is ready, no other operations which require a barrier will start
914 * until the IO request has had a chance.
916 * So: regular IO calls 'wait_barrier'. When that returns there
917 * is no backgroup IO happening, It must arrange to call
918 * allow_barrier when it has finished its IO.
919 * backgroup IO calls must call raise_barrier. Once that returns
920 * there is no normal IO happeing. It must arrange to call
921 * lower_barrier when the particular background IO completes.
924 static void raise_barrier(struct r10conf *conf, int force)
926 write_seqlock_irq(&conf->resync_lock);
928 if (WARN_ON_ONCE(force && !conf->barrier))
931 /* Wait until no block IO is waiting (unless 'force') */
932 wait_event_barrier(conf, force || !conf->nr_waiting);
934 /* block any new IO from starting */
935 WRITE_ONCE(conf->barrier, conf->barrier + 1);
937 /* Now wait for all pending IO to complete */
938 wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
939 conf->barrier < RESYNC_DEPTH);
941 write_sequnlock_irq(&conf->resync_lock);
944 static void lower_barrier(struct r10conf *conf)
948 write_seqlock_irqsave(&conf->resync_lock, flags);
949 WRITE_ONCE(conf->barrier, conf->barrier - 1);
950 write_sequnlock_irqrestore(&conf->resync_lock, flags);
951 wake_up(&conf->wait_barrier);
954 static bool stop_waiting_barrier(struct r10conf *conf)
956 struct bio_list *bio_list = current->bio_list;
957 struct md_thread *thread;
959 /* barrier is dropped */
964 * If there are already pending requests (preventing the barrier from
965 * rising completely), and the pre-process bio queue isn't empty, then
966 * don't wait, as we need to empty that queue to get the nr_pending
969 if (atomic_read(&conf->nr_pending) && bio_list &&
970 (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
973 /* daemon thread must exist while handling io */
974 thread = rcu_dereference_protected(conf->mddev->thread, true);
976 * move on if io is issued from raid10d(), nr_pending is not released
977 * from original io(see handle_read_error()). All raise barrier is
978 * blocked until this io is done.
980 if (thread->tsk == current) {
981 WARN_ON_ONCE(atomic_read(&conf->nr_pending) == 0);
988 static bool wait_barrier_nolock(struct r10conf *conf)
990 unsigned int seq = read_seqbegin(&conf->resync_lock);
992 if (READ_ONCE(conf->barrier))
995 atomic_inc(&conf->nr_pending);
996 if (!read_seqretry(&conf->resync_lock, seq))
999 if (atomic_dec_and_test(&conf->nr_pending))
1000 wake_up_barrier(conf);
1005 static bool wait_barrier(struct r10conf *conf, bool nowait)
1009 if (wait_barrier_nolock(conf))
1012 write_seqlock_irq(&conf->resync_lock);
1013 if (conf->barrier) {
1014 /* Return false when nowait flag is set */
1019 mddev_add_trace_msg(conf->mddev, "raid10 wait barrier");
1020 wait_event_barrier(conf, stop_waiting_barrier(conf));
1023 if (!conf->nr_waiting)
1024 wake_up(&conf->wait_barrier);
1026 /* Only increment nr_pending when we wait */
1028 atomic_inc(&conf->nr_pending);
1029 write_sequnlock_irq(&conf->resync_lock);
1033 static void allow_barrier(struct r10conf *conf)
1035 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1036 (conf->array_freeze_pending))
1037 wake_up_barrier(conf);
1040 static void freeze_array(struct r10conf *conf, int extra)
1042 /* stop syncio and normal IO and wait for everything to
1044 * We increment barrier and nr_waiting, and then
1045 * wait until nr_pending match nr_queued+extra
1046 * This is called in the context of one normal IO request
1047 * that has failed. Thus any sync request that might be pending
1048 * will be blocked by nr_pending, and we need to wait for
1049 * pending IO requests to complete or be queued for re-try.
1050 * Thus the number queued (nr_queued) plus this request (extra)
1051 * must match the number of pending IOs (nr_pending) before
1054 write_seqlock_irq(&conf->resync_lock);
1055 conf->array_freeze_pending++;
1056 WRITE_ONCE(conf->barrier, conf->barrier + 1);
1058 wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1059 conf->nr_queued + extra, flush_pending_writes(conf));
1060 conf->array_freeze_pending--;
1061 write_sequnlock_irq(&conf->resync_lock);
1064 static void unfreeze_array(struct r10conf *conf)
1066 /* reverse the effect of the freeze */
1067 write_seqlock_irq(&conf->resync_lock);
1068 WRITE_ONCE(conf->barrier, conf->barrier - 1);
1070 wake_up(&conf->wait_barrier);
1071 write_sequnlock_irq(&conf->resync_lock);
1074 static sector_t choose_data_offset(struct r10bio *r10_bio,
1075 struct md_rdev *rdev)
1077 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1078 test_bit(R10BIO_Previous, &r10_bio->state))
1079 return rdev->data_offset;
1081 return rdev->new_data_offset;
1084 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1086 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1087 struct mddev *mddev = plug->cb.data;
1088 struct r10conf *conf = mddev->private;
1091 if (from_schedule) {
1092 spin_lock_irq(&conf->device_lock);
1093 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1094 spin_unlock_irq(&conf->device_lock);
1095 wake_up_barrier(conf);
1096 md_wakeup_thread(mddev->thread);
1101 /* we aren't scheduling, so we can do the write-out directly. */
1102 bio = bio_list_get(&plug->pending);
1103 raid1_prepare_flush_writes(mddev->bitmap);
1104 wake_up_barrier(conf);
1106 while (bio) { /* submit pending writes */
1107 struct bio *next = bio->bi_next;
1109 raid1_submit_write(bio);
1117 * 1. Register the new request and wait if the reconstruction thread has put
1118 * up a bar for new requests. Continue immediately if no resync is active
1120 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1122 static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1123 struct bio *bio, sector_t sectors)
1125 /* Bail out if REQ_NOWAIT is set for the bio */
1126 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1127 bio_wouldblock_error(bio);
1130 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1131 bio->bi_iter.bi_sector < conf->reshape_progress &&
1132 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1133 allow_barrier(conf);
1134 if (bio->bi_opf & REQ_NOWAIT) {
1135 bio_wouldblock_error(bio);
1138 mddev_add_trace_msg(conf->mddev, "raid10 wait reshape");
1139 wait_event(conf->wait_barrier,
1140 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1141 conf->reshape_progress >= bio->bi_iter.bi_sector +
1143 wait_barrier(conf, false);
1148 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1149 struct r10bio *r10_bio, bool io_accounting)
1151 struct r10conf *conf = mddev->private;
1152 struct bio *read_bio;
1153 const enum req_op op = bio_op(bio);
1154 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1156 struct md_rdev *rdev;
1157 char b[BDEVNAME_SIZE];
1158 int slot = r10_bio->read_slot;
1159 struct md_rdev *err_rdev = NULL;
1160 gfp_t gfp = GFP_NOIO;
1162 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1164 * This is an error retry, but we cannot
1165 * safely dereference the rdev in the r10_bio,
1166 * we must use the one in conf.
1167 * If it has already been disconnected (unlikely)
1168 * we lose the device name in error messages.
1172 * As we are blocking raid10, it is a little safer to
1175 gfp = GFP_NOIO | __GFP_HIGH;
1177 disk = r10_bio->devs[slot].devnum;
1178 err_rdev = conf->mirrors[disk].rdev;
1180 snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1183 /* This never gets dereferenced */
1184 err_rdev = r10_bio->devs[slot].rdev;
1188 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1190 rdev = read_balance(conf, r10_bio, &max_sectors);
1193 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1195 (unsigned long long)r10_bio->sector);
1197 raid_end_bio_io(r10_bio);
1201 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1204 (unsigned long long)r10_bio->sector);
1205 if (max_sectors < bio_sectors(bio)) {
1206 struct bio *split = bio_split(bio, max_sectors,
1207 gfp, &conf->bio_split);
1208 bio_chain(split, bio);
1209 allow_barrier(conf);
1210 submit_bio_noacct(bio);
1211 wait_barrier(conf, false);
1213 r10_bio->master_bio = bio;
1214 r10_bio->sectors = max_sectors;
1216 slot = r10_bio->read_slot;
1218 if (io_accounting) {
1219 md_account_bio(mddev, &bio);
1220 r10_bio->master_bio = bio;
1222 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1224 r10_bio->devs[slot].bio = read_bio;
1225 r10_bio->devs[slot].rdev = rdev;
1227 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1228 choose_data_offset(r10_bio, rdev);
1229 read_bio->bi_end_io = raid10_end_read_request;
1230 read_bio->bi_opf = op | do_sync;
1231 if (test_bit(FailFast, &rdev->flags) &&
1232 test_bit(R10BIO_FailFast, &r10_bio->state))
1233 read_bio->bi_opf |= MD_FAILFAST;
1234 read_bio->bi_private = r10_bio;
1235 mddev_trace_remap(mddev, read_bio, r10_bio->sector);
1236 submit_bio_noacct(read_bio);
1240 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1241 struct bio *bio, bool replacement,
1244 const enum req_op op = bio_op(bio);
1245 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1246 const blk_opf_t do_fua = bio->bi_opf & REQ_FUA;
1247 unsigned long flags;
1248 struct r10conf *conf = mddev->private;
1249 struct md_rdev *rdev;
1250 int devnum = r10_bio->devs[n_copy].devnum;
1253 rdev = replacement ? conf->mirrors[devnum].replacement :
1254 conf->mirrors[devnum].rdev;
1256 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1258 r10_bio->devs[n_copy].repl_bio = mbio;
1260 r10_bio->devs[n_copy].bio = mbio;
1262 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1263 choose_data_offset(r10_bio, rdev));
1264 mbio->bi_end_io = raid10_end_write_request;
1265 mbio->bi_opf = op | do_sync | do_fua;
1266 if (!replacement && test_bit(FailFast,
1267 &conf->mirrors[devnum].rdev->flags)
1268 && enough(conf, devnum))
1269 mbio->bi_opf |= MD_FAILFAST;
1270 mbio->bi_private = r10_bio;
1271 mddev_trace_remap(mddev, mbio, r10_bio->sector);
1272 /* flush_pending_writes() needs access to the rdev so...*/
1273 mbio->bi_bdev = (void *)rdev;
1275 atomic_inc(&r10_bio->remaining);
1277 if (!raid1_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) {
1278 spin_lock_irqsave(&conf->device_lock, flags);
1279 bio_list_add(&conf->pending_bio_list, mbio);
1280 spin_unlock_irqrestore(&conf->device_lock, flags);
1281 md_wakeup_thread(mddev->thread);
1285 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1288 struct r10conf *conf = mddev->private;
1289 struct md_rdev *blocked_rdev;
1292 blocked_rdev = NULL;
1293 for (i = 0; i < conf->copies; i++) {
1294 struct md_rdev *rdev, *rrdev;
1296 rdev = conf->mirrors[i].rdev;
1297 rrdev = conf->mirrors[i].replacement;
1298 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1299 atomic_inc(&rdev->nr_pending);
1300 blocked_rdev = rdev;
1303 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1304 atomic_inc(&rrdev->nr_pending);
1305 blocked_rdev = rrdev;
1309 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1310 sector_t dev_sector = r10_bio->devs[i].addr;
1313 * Discard request doesn't care the write result
1314 * so it doesn't need to wait blocked disk here.
1316 if (!r10_bio->sectors)
1319 if (rdev_has_badblock(rdev, dev_sector,
1320 r10_bio->sectors) < 0) {
1322 * Mustn't write here until the bad block
1325 atomic_inc(&rdev->nr_pending);
1326 set_bit(BlockedBadBlocks, &rdev->flags);
1327 blocked_rdev = rdev;
1333 if (unlikely(blocked_rdev)) {
1334 /* Have to wait for this device to get unblocked, then retry */
1335 allow_barrier(conf);
1336 mddev_add_trace_msg(conf->mddev,
1337 "raid10 %s wait rdev %d blocked",
1338 __func__, blocked_rdev->raid_disk);
1339 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1340 wait_barrier(conf, false);
1345 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1346 struct r10bio *r10_bio)
1348 struct r10conf *conf = mddev->private;
1353 if ((mddev_is_clustered(mddev) &&
1354 md_cluster_ops->area_resyncing(mddev, WRITE,
1355 bio->bi_iter.bi_sector,
1356 bio_end_sector(bio)))) {
1358 /* Bail out if REQ_NOWAIT is set for the bio */
1359 if (bio->bi_opf & REQ_NOWAIT) {
1360 bio_wouldblock_error(bio);
1364 prepare_to_wait(&conf->wait_barrier,
1366 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1367 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1371 finish_wait(&conf->wait_barrier, &w);
1374 sectors = r10_bio->sectors;
1375 if (!regular_request_wait(mddev, conf, bio, sectors))
1377 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1378 (mddev->reshape_backwards
1379 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1380 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1381 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1382 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1383 /* Need to update reshape_position in metadata */
1384 mddev->reshape_position = conf->reshape_progress;
1385 set_mask_bits(&mddev->sb_flags, 0,
1386 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1387 md_wakeup_thread(mddev->thread);
1388 if (bio->bi_opf & REQ_NOWAIT) {
1389 allow_barrier(conf);
1390 bio_wouldblock_error(bio);
1393 mddev_add_trace_msg(conf->mddev,
1394 "raid10 wait reshape metadata");
1395 wait_event(mddev->sb_wait,
1396 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1398 conf->reshape_safe = mddev->reshape_position;
1401 /* first select target devices under rcu_lock and
1402 * inc refcount on their rdev. Record them by setting
1404 * If there are known/acknowledged bad blocks on any device
1405 * on which we have seen a write error, we want to avoid
1406 * writing to those blocks. This potentially requires several
1407 * writes to write around the bad blocks. Each set of writes
1408 * gets its own r10_bio with a set of bios attached.
1411 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1412 raid10_find_phys(conf, r10_bio);
1414 wait_blocked_dev(mddev, r10_bio);
1416 max_sectors = r10_bio->sectors;
1418 for (i = 0; i < conf->copies; i++) {
1419 int d = r10_bio->devs[i].devnum;
1420 struct md_rdev *rdev, *rrdev;
1422 rdev = conf->mirrors[d].rdev;
1423 rrdev = conf->mirrors[d].replacement;
1424 if (rdev && (test_bit(Faulty, &rdev->flags)))
1426 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1429 r10_bio->devs[i].bio = NULL;
1430 r10_bio->devs[i].repl_bio = NULL;
1432 if (!rdev && !rrdev) {
1433 set_bit(R10BIO_Degraded, &r10_bio->state);
1436 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1438 sector_t dev_sector = r10_bio->devs[i].addr;
1442 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1443 &first_bad, &bad_sectors);
1444 if (is_bad && first_bad <= dev_sector) {
1445 /* Cannot write here at all */
1446 bad_sectors -= (dev_sector - first_bad);
1447 if (bad_sectors < max_sectors)
1448 /* Mustn't write more than bad_sectors
1449 * to other devices yet
1451 max_sectors = bad_sectors;
1452 /* We don't set R10BIO_Degraded as that
1453 * only applies if the disk is missing,
1454 * so it might be re-added, and we want to
1455 * know to recover this chunk.
1456 * In this case the device is here, and the
1457 * fact that this chunk is not in-sync is
1458 * recorded in the bad block log.
1463 int good_sectors = first_bad - dev_sector;
1464 if (good_sectors < max_sectors)
1465 max_sectors = good_sectors;
1469 r10_bio->devs[i].bio = bio;
1470 atomic_inc(&rdev->nr_pending);
1473 r10_bio->devs[i].repl_bio = bio;
1474 atomic_inc(&rrdev->nr_pending);
1478 if (max_sectors < r10_bio->sectors)
1479 r10_bio->sectors = max_sectors;
1481 if (r10_bio->sectors < bio_sectors(bio)) {
1482 struct bio *split = bio_split(bio, r10_bio->sectors,
1483 GFP_NOIO, &conf->bio_split);
1484 bio_chain(split, bio);
1485 allow_barrier(conf);
1486 submit_bio_noacct(bio);
1487 wait_barrier(conf, false);
1489 r10_bio->master_bio = bio;
1492 md_account_bio(mddev, &bio);
1493 r10_bio->master_bio = bio;
1494 atomic_set(&r10_bio->remaining, 1);
1495 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1497 for (i = 0; i < conf->copies; i++) {
1498 if (r10_bio->devs[i].bio)
1499 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1500 if (r10_bio->devs[i].repl_bio)
1501 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1503 one_write_done(r10_bio);
1506 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1508 struct r10conf *conf = mddev->private;
1509 struct r10bio *r10_bio;
1511 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1513 r10_bio->master_bio = bio;
1514 r10_bio->sectors = sectors;
1516 r10_bio->mddev = mddev;
1517 r10_bio->sector = bio->bi_iter.bi_sector;
1519 r10_bio->read_slot = -1;
1520 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1521 conf->geo.raid_disks);
1523 if (bio_data_dir(bio) == READ)
1524 raid10_read_request(mddev, bio, r10_bio, true);
1526 raid10_write_request(mddev, bio, r10_bio);
1529 static void raid_end_discard_bio(struct r10bio *r10bio)
1531 struct r10conf *conf = r10bio->mddev->private;
1532 struct r10bio *first_r10bio;
1534 while (atomic_dec_and_test(&r10bio->remaining)) {
1536 allow_barrier(conf);
1538 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1539 first_r10bio = (struct r10bio *)r10bio->master_bio;
1540 free_r10bio(r10bio);
1541 r10bio = first_r10bio;
1543 md_write_end(r10bio->mddev);
1544 bio_endio(r10bio->master_bio);
1545 free_r10bio(r10bio);
1551 static void raid10_end_discard_request(struct bio *bio)
1553 struct r10bio *r10_bio = bio->bi_private;
1554 struct r10conf *conf = r10_bio->mddev->private;
1555 struct md_rdev *rdev = NULL;
1560 * We don't care the return value of discard bio
1562 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1563 set_bit(R10BIO_Uptodate, &r10_bio->state);
1565 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1566 rdev = repl ? conf->mirrors[dev].replacement :
1567 conf->mirrors[dev].rdev;
1569 raid_end_discard_bio(r10_bio);
1570 rdev_dec_pending(rdev, conf->mddev);
1574 * There are some limitations to handle discard bio
1575 * 1st, the discard size is bigger than stripe_size*2.
1576 * 2st, if the discard bio spans reshape progress, we use the old way to
1577 * handle discard bio
1579 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1581 struct r10conf *conf = mddev->private;
1582 struct geom *geo = &conf->geo;
1583 int far_copies = geo->far_copies;
1584 bool first_copy = true;
1585 struct r10bio *r10_bio, *first_r10bio;
1589 unsigned int stripe_size;
1590 unsigned int stripe_data_disks;
1591 sector_t split_size;
1592 sector_t bio_start, bio_end;
1593 sector_t first_stripe_index, last_stripe_index;
1594 sector_t start_disk_offset;
1595 unsigned int start_disk_index;
1596 sector_t end_disk_offset;
1597 unsigned int end_disk_index;
1598 unsigned int remainder;
1600 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1603 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1604 bio_wouldblock_error(bio);
1607 wait_barrier(conf, false);
1610 * Check reshape again to avoid reshape happens after checking
1611 * MD_RECOVERY_RESHAPE and before wait_barrier
1613 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1616 if (geo->near_copies)
1617 stripe_data_disks = geo->raid_disks / geo->near_copies +
1618 geo->raid_disks % geo->near_copies;
1620 stripe_data_disks = geo->raid_disks;
1622 stripe_size = stripe_data_disks << geo->chunk_shift;
1624 bio_start = bio->bi_iter.bi_sector;
1625 bio_end = bio_end_sector(bio);
1628 * Maybe one discard bio is smaller than strip size or across one
1629 * stripe and discard region is larger than one stripe size. For far
1630 * offset layout, if the discard region is not aligned with stripe
1631 * size, there is hole when we submit discard bio to member disk.
1632 * For simplicity, we only handle discard bio which discard region
1633 * is bigger than stripe_size * 2
1635 if (bio_sectors(bio) < stripe_size*2)
1639 * Keep bio aligned with strip size.
1641 div_u64_rem(bio_start, stripe_size, &remainder);
1643 split_size = stripe_size - remainder;
1644 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1645 bio_chain(split, bio);
1646 allow_barrier(conf);
1647 /* Resend the fist split part */
1648 submit_bio_noacct(split);
1649 wait_barrier(conf, false);
1651 div_u64_rem(bio_end, stripe_size, &remainder);
1653 split_size = bio_sectors(bio) - remainder;
1654 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1655 bio_chain(split, bio);
1656 allow_barrier(conf);
1657 /* Resend the second split part */
1658 submit_bio_noacct(bio);
1660 wait_barrier(conf, false);
1663 bio_start = bio->bi_iter.bi_sector;
1664 bio_end = bio_end_sector(bio);
1667 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1668 * One stripe contains the chunks from all member disk (one chunk from
1669 * one disk at the same HBA address). For layout detail, see 'man md 4'
1671 chunk = bio_start >> geo->chunk_shift;
1672 chunk *= geo->near_copies;
1673 first_stripe_index = chunk;
1674 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1675 if (geo->far_offset)
1676 first_stripe_index *= geo->far_copies;
1677 start_disk_offset = (bio_start & geo->chunk_mask) +
1678 (first_stripe_index << geo->chunk_shift);
1680 chunk = bio_end >> geo->chunk_shift;
1681 chunk *= geo->near_copies;
1682 last_stripe_index = chunk;
1683 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1684 if (geo->far_offset)
1685 last_stripe_index *= geo->far_copies;
1686 end_disk_offset = (bio_end & geo->chunk_mask) +
1687 (last_stripe_index << geo->chunk_shift);
1690 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1691 r10_bio->mddev = mddev;
1693 r10_bio->sectors = 0;
1694 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1695 wait_blocked_dev(mddev, r10_bio);
1698 * For far layout it needs more than one r10bio to cover all regions.
1699 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1700 * to record the discard bio. Other r10bio->master_bio record the first
1701 * r10bio. The first r10bio only release after all other r10bios finish.
1702 * The discard bio returns only first r10bio finishes
1705 r10_bio->master_bio = bio;
1706 set_bit(R10BIO_Discard, &r10_bio->state);
1708 first_r10bio = r10_bio;
1710 r10_bio->master_bio = (struct bio *)first_r10bio;
1713 * first select target devices under rcu_lock and
1714 * inc refcount on their rdev. Record them by setting
1717 for (disk = 0; disk < geo->raid_disks; disk++) {
1718 struct md_rdev *rdev, *rrdev;
1720 rdev = conf->mirrors[disk].rdev;
1721 rrdev = conf->mirrors[disk].replacement;
1722 r10_bio->devs[disk].bio = NULL;
1723 r10_bio->devs[disk].repl_bio = NULL;
1725 if (rdev && (test_bit(Faulty, &rdev->flags)))
1727 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1729 if (!rdev && !rrdev)
1733 r10_bio->devs[disk].bio = bio;
1734 atomic_inc(&rdev->nr_pending);
1737 r10_bio->devs[disk].repl_bio = bio;
1738 atomic_inc(&rrdev->nr_pending);
1742 atomic_set(&r10_bio->remaining, 1);
1743 for (disk = 0; disk < geo->raid_disks; disk++) {
1744 sector_t dev_start, dev_end;
1745 struct bio *mbio, *rbio = NULL;
1748 * Now start to calculate the start and end address for each disk.
1749 * The space between dev_start and dev_end is the discard region.
1751 * For dev_start, it needs to consider three conditions:
1752 * 1st, the disk is before start_disk, you can imagine the disk in
1753 * the next stripe. So the dev_start is the start address of next
1755 * 2st, the disk is after start_disk, it means the disk is at the
1756 * same stripe of first disk
1757 * 3st, the first disk itself, we can use start_disk_offset directly
1759 if (disk < start_disk_index)
1760 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1761 else if (disk > start_disk_index)
1762 dev_start = first_stripe_index * mddev->chunk_sectors;
1764 dev_start = start_disk_offset;
1766 if (disk < end_disk_index)
1767 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1768 else if (disk > end_disk_index)
1769 dev_end = last_stripe_index * mddev->chunk_sectors;
1771 dev_end = end_disk_offset;
1774 * It only handles discard bio which size is >= stripe size, so
1775 * dev_end > dev_start all the time.
1776 * It doesn't need to use rcu lock to get rdev here. We already
1777 * add rdev->nr_pending in the first loop.
1779 if (r10_bio->devs[disk].bio) {
1780 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1781 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1783 mbio->bi_end_io = raid10_end_discard_request;
1784 mbio->bi_private = r10_bio;
1785 r10_bio->devs[disk].bio = mbio;
1786 r10_bio->devs[disk].devnum = disk;
1787 atomic_inc(&r10_bio->remaining);
1788 md_submit_discard_bio(mddev, rdev, mbio,
1789 dev_start + choose_data_offset(r10_bio, rdev),
1790 dev_end - dev_start);
1793 if (r10_bio->devs[disk].repl_bio) {
1794 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1795 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1797 rbio->bi_end_io = raid10_end_discard_request;
1798 rbio->bi_private = r10_bio;
1799 r10_bio->devs[disk].repl_bio = rbio;
1800 r10_bio->devs[disk].devnum = disk;
1801 atomic_inc(&r10_bio->remaining);
1802 md_submit_discard_bio(mddev, rrdev, rbio,
1803 dev_start + choose_data_offset(r10_bio, rrdev),
1804 dev_end - dev_start);
1809 if (!geo->far_offset && --far_copies) {
1810 first_stripe_index += geo->stride >> geo->chunk_shift;
1811 start_disk_offset += geo->stride;
1812 last_stripe_index += geo->stride >> geo->chunk_shift;
1813 end_disk_offset += geo->stride;
1814 atomic_inc(&first_r10bio->remaining);
1815 raid_end_discard_bio(r10_bio);
1816 wait_barrier(conf, false);
1820 raid_end_discard_bio(r10_bio);
1824 allow_barrier(conf);
1828 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1830 struct r10conf *conf = mddev->private;
1831 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1832 int chunk_sects = chunk_mask + 1;
1833 int sectors = bio_sectors(bio);
1835 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1836 && md_flush_request(mddev, bio))
1839 if (!md_write_start(mddev, bio))
1842 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1843 if (!raid10_handle_discard(mddev, bio))
1847 * If this request crosses a chunk boundary, we need to split
1850 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1851 sectors > chunk_sects
1852 && (conf->geo.near_copies < conf->geo.raid_disks
1853 || conf->prev.near_copies <
1854 conf->prev.raid_disks)))
1855 sectors = chunk_sects -
1856 (bio->bi_iter.bi_sector &
1858 __make_request(mddev, bio, sectors);
1860 /* In case raid10d snuck in to freeze_array */
1861 wake_up_barrier(conf);
1865 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1867 struct r10conf *conf = mddev->private;
1870 lockdep_assert_held(&mddev->lock);
1872 if (conf->geo.near_copies < conf->geo.raid_disks)
1873 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1874 if (conf->geo.near_copies > 1)
1875 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1876 if (conf->geo.far_copies > 1) {
1877 if (conf->geo.far_offset)
1878 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1880 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1881 if (conf->geo.far_set_size != conf->geo.raid_disks)
1882 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1884 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1885 conf->geo.raid_disks - mddev->degraded);
1886 for (i = 0; i < conf->geo.raid_disks; i++) {
1887 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1889 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1891 seq_printf(seq, "]");
1894 /* check if there are enough drives for
1895 * every block to appear on atleast one.
1896 * Don't consider the device numbered 'ignore'
1897 * as we might be about to remove it.
1899 static int _enough(struct r10conf *conf, int previous, int ignore)
1905 disks = conf->prev.raid_disks;
1906 ncopies = conf->prev.near_copies;
1908 disks = conf->geo.raid_disks;
1909 ncopies = conf->geo.near_copies;
1913 int n = conf->copies;
1917 struct md_rdev *rdev;
1918 if (this != ignore &&
1919 (rdev = conf->mirrors[this].rdev) &&
1920 test_bit(In_sync, &rdev->flags))
1922 this = (this+1) % disks;
1926 first = (first + ncopies) % disks;
1927 } while (first != 0);
1933 static int enough(struct r10conf *conf, int ignore)
1935 /* when calling 'enough', both 'prev' and 'geo' must
1937 * This is ensured if ->reconfig_mutex or ->device_lock
1940 return _enough(conf, 0, ignore) &&
1941 _enough(conf, 1, ignore);
1945 * raid10_error() - RAID10 error handler.
1946 * @mddev: affected md device.
1947 * @rdev: member device to fail.
1949 * The routine acknowledges &rdev failure and determines new @mddev state.
1950 * If it failed, then:
1951 * - &MD_BROKEN flag is set in &mddev->flags.
1952 * Otherwise, it must be degraded:
1953 * - recovery is interrupted.
1954 * - &mddev->degraded is bumped.
1956 * @rdev is marked as &Faulty excluding case when array is failed and
1957 * &mddev->fail_last_dev is off.
1959 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1961 struct r10conf *conf = mddev->private;
1962 unsigned long flags;
1964 spin_lock_irqsave(&conf->device_lock, flags);
1966 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
1967 set_bit(MD_BROKEN, &mddev->flags);
1969 if (!mddev->fail_last_dev) {
1970 spin_unlock_irqrestore(&conf->device_lock, flags);
1974 if (test_and_clear_bit(In_sync, &rdev->flags))
1977 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1978 set_bit(Blocked, &rdev->flags);
1979 set_bit(Faulty, &rdev->flags);
1980 set_mask_bits(&mddev->sb_flags, 0,
1981 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1982 spin_unlock_irqrestore(&conf->device_lock, flags);
1983 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
1984 "md/raid10:%s: Operation continuing on %d devices.\n",
1985 mdname(mddev), rdev->bdev,
1986 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1989 static void print_conf(struct r10conf *conf)
1992 struct md_rdev *rdev;
1994 pr_debug("RAID10 conf printout:\n");
1996 pr_debug("(!conf)\n");
1999 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2000 conf->geo.raid_disks);
2002 lockdep_assert_held(&conf->mddev->reconfig_mutex);
2003 for (i = 0; i < conf->geo.raid_disks; i++) {
2004 rdev = conf->mirrors[i].rdev;
2006 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2007 i, !test_bit(In_sync, &rdev->flags),
2008 !test_bit(Faulty, &rdev->flags),
2013 static void close_sync(struct r10conf *conf)
2015 wait_barrier(conf, false);
2016 allow_barrier(conf);
2018 mempool_exit(&conf->r10buf_pool);
2021 static int raid10_spare_active(struct mddev *mddev)
2024 struct r10conf *conf = mddev->private;
2025 struct raid10_info *tmp;
2027 unsigned long flags;
2030 * Find all non-in_sync disks within the RAID10 configuration
2031 * and mark them in_sync
2033 for (i = 0; i < conf->geo.raid_disks; i++) {
2034 tmp = conf->mirrors + i;
2035 if (tmp->replacement
2036 && tmp->replacement->recovery_offset == MaxSector
2037 && !test_bit(Faulty, &tmp->replacement->flags)
2038 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2039 /* Replacement has just become active */
2041 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2044 /* Replaced device not technically faulty,
2045 * but we need to be sure it gets removed
2046 * and never re-added.
2048 set_bit(Faulty, &tmp->rdev->flags);
2049 sysfs_notify_dirent_safe(
2050 tmp->rdev->sysfs_state);
2052 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2053 } else if (tmp->rdev
2054 && tmp->rdev->recovery_offset == MaxSector
2055 && !test_bit(Faulty, &tmp->rdev->flags)
2056 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2058 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2061 spin_lock_irqsave(&conf->device_lock, flags);
2062 mddev->degraded -= count;
2063 spin_unlock_irqrestore(&conf->device_lock, flags);
2069 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2071 struct r10conf *conf = mddev->private;
2073 int mirror, repl_slot = -1;
2075 int last = conf->geo.raid_disks - 1;
2076 struct raid10_info *p;
2078 if (mddev->recovery_cp < MaxSector)
2079 /* only hot-add to in-sync arrays, as recovery is
2080 * very different from resync
2083 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2086 if (md_integrity_add_rdev(rdev, mddev))
2089 if (rdev->raid_disk >= 0)
2090 first = last = rdev->raid_disk;
2092 if (rdev->saved_raid_disk >= first &&
2093 rdev->saved_raid_disk < conf->geo.raid_disks &&
2094 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2095 mirror = rdev->saved_raid_disk;
2098 for ( ; mirror <= last ; mirror++) {
2099 p = &conf->mirrors[mirror];
2100 if (p->recovery_disabled == mddev->recovery_disabled)
2103 if (test_bit(WantReplacement, &p->rdev->flags) &&
2104 p->replacement == NULL && repl_slot < 0)
2109 err = mddev_stack_new_rdev(mddev, rdev);
2112 p->head_position = 0;
2113 p->recovery_disabled = mddev->recovery_disabled - 1;
2114 rdev->raid_disk = mirror;
2116 if (rdev->saved_raid_disk != mirror)
2118 WRITE_ONCE(p->rdev, rdev);
2122 if (err && repl_slot >= 0) {
2123 p = &conf->mirrors[repl_slot];
2124 clear_bit(In_sync, &rdev->flags);
2125 set_bit(Replacement, &rdev->flags);
2126 rdev->raid_disk = repl_slot;
2127 err = mddev_stack_new_rdev(mddev, rdev);
2131 WRITE_ONCE(p->replacement, rdev);
2138 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2140 struct r10conf *conf = mddev->private;
2142 int number = rdev->raid_disk;
2143 struct md_rdev **rdevp;
2144 struct raid10_info *p;
2147 if (unlikely(number >= mddev->raid_disks))
2149 p = conf->mirrors + number;
2150 if (rdev == p->rdev)
2152 else if (rdev == p->replacement)
2153 rdevp = &p->replacement;
2157 if (test_bit(In_sync, &rdev->flags) ||
2158 atomic_read(&rdev->nr_pending)) {
2162 /* Only remove non-faulty devices if recovery
2165 if (!test_bit(Faulty, &rdev->flags) &&
2166 mddev->recovery_disabled != p->recovery_disabled &&
2167 (!p->replacement || p->replacement == rdev) &&
2168 number < conf->geo.raid_disks &&
2173 WRITE_ONCE(*rdevp, NULL);
2174 if (p->replacement) {
2175 /* We must have just cleared 'rdev' */
2176 WRITE_ONCE(p->rdev, p->replacement);
2177 clear_bit(Replacement, &p->replacement->flags);
2178 WRITE_ONCE(p->replacement, NULL);
2181 clear_bit(WantReplacement, &rdev->flags);
2182 err = md_integrity_register(mddev);
2190 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2192 struct r10conf *conf = r10_bio->mddev->private;
2194 if (!bio->bi_status)
2195 set_bit(R10BIO_Uptodate, &r10_bio->state);
2197 /* The write handler will notice the lack of
2198 * R10BIO_Uptodate and record any errors etc
2200 atomic_add(r10_bio->sectors,
2201 &conf->mirrors[d].rdev->corrected_errors);
2203 /* for reconstruct, we always reschedule after a read.
2204 * for resync, only after all reads
2206 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2207 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2208 atomic_dec_and_test(&r10_bio->remaining)) {
2209 /* we have read all the blocks,
2210 * do the comparison in process context in raid10d
2212 reschedule_retry(r10_bio);
2216 static void end_sync_read(struct bio *bio)
2218 struct r10bio *r10_bio = get_resync_r10bio(bio);
2219 struct r10conf *conf = r10_bio->mddev->private;
2220 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2222 __end_sync_read(r10_bio, bio, d);
2225 static void end_reshape_read(struct bio *bio)
2227 /* reshape read bio isn't allocated from r10buf_pool */
2228 struct r10bio *r10_bio = bio->bi_private;
2230 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2233 static void end_sync_request(struct r10bio *r10_bio)
2235 struct mddev *mddev = r10_bio->mddev;
2237 while (atomic_dec_and_test(&r10_bio->remaining)) {
2238 if (r10_bio->master_bio == NULL) {
2239 /* the primary of several recovery bios */
2240 sector_t s = r10_bio->sectors;
2241 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2242 test_bit(R10BIO_WriteError, &r10_bio->state))
2243 reschedule_retry(r10_bio);
2246 md_done_sync(mddev, s, 1);
2249 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2250 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2251 test_bit(R10BIO_WriteError, &r10_bio->state))
2252 reschedule_retry(r10_bio);
2260 static void end_sync_write(struct bio *bio)
2262 struct r10bio *r10_bio = get_resync_r10bio(bio);
2263 struct mddev *mddev = r10_bio->mddev;
2264 struct r10conf *conf = mddev->private;
2268 struct md_rdev *rdev = NULL;
2270 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2272 rdev = conf->mirrors[d].replacement;
2274 rdev = conf->mirrors[d].rdev;
2276 if (bio->bi_status) {
2278 md_error(mddev, rdev);
2280 set_bit(WriteErrorSeen, &rdev->flags);
2281 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2282 set_bit(MD_RECOVERY_NEEDED,
2283 &rdev->mddev->recovery);
2284 set_bit(R10BIO_WriteError, &r10_bio->state);
2286 } else if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
2287 r10_bio->sectors)) {
2288 set_bit(R10BIO_MadeGood, &r10_bio->state);
2291 rdev_dec_pending(rdev, mddev);
2293 end_sync_request(r10_bio);
2297 * Note: sync and recover and handled very differently for raid10
2298 * This code is for resync.
2299 * For resync, we read through virtual addresses and read all blocks.
2300 * If there is any error, we schedule a write. The lowest numbered
2301 * drive is authoritative.
2302 * However requests come for physical address, so we need to map.
2303 * For every physical address there are raid_disks/copies virtual addresses,
2304 * which is always are least one, but is not necessarly an integer.
2305 * This means that a physical address can span multiple chunks, so we may
2306 * have to submit multiple io requests for a single sync request.
2309 * We check if all blocks are in-sync and only write to blocks that
2312 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2314 struct r10conf *conf = mddev->private;
2316 struct bio *tbio, *fbio;
2318 struct page **tpages, **fpages;
2320 atomic_set(&r10_bio->remaining, 1);
2322 /* find the first device with a block */
2323 for (i=0; i<conf->copies; i++)
2324 if (!r10_bio->devs[i].bio->bi_status)
2327 if (i == conf->copies)
2331 fbio = r10_bio->devs[i].bio;
2332 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2333 fbio->bi_iter.bi_idx = 0;
2334 fpages = get_resync_pages(fbio)->pages;
2336 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2337 /* now find blocks with errors */
2338 for (i=0 ; i < conf->copies ; i++) {
2340 struct md_rdev *rdev;
2341 struct resync_pages *rp;
2343 tbio = r10_bio->devs[i].bio;
2345 if (tbio->bi_end_io != end_sync_read)
2350 tpages = get_resync_pages(tbio)->pages;
2351 d = r10_bio->devs[i].devnum;
2352 rdev = conf->mirrors[d].rdev;
2353 if (!r10_bio->devs[i].bio->bi_status) {
2354 /* We know that the bi_io_vec layout is the same for
2355 * both 'first' and 'i', so we just compare them.
2356 * All vec entries are PAGE_SIZE;
2358 int sectors = r10_bio->sectors;
2359 for (j = 0; j < vcnt; j++) {
2360 int len = PAGE_SIZE;
2361 if (sectors < (len / 512))
2362 len = sectors * 512;
2363 if (memcmp(page_address(fpages[j]),
2364 page_address(tpages[j]),
2371 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2372 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2373 /* Don't fix anything. */
2375 } else if (test_bit(FailFast, &rdev->flags)) {
2376 /* Just give up on this device */
2377 md_error(rdev->mddev, rdev);
2380 /* Ok, we need to write this bio, either to correct an
2381 * inconsistency or to correct an unreadable block.
2382 * First we need to fixup bv_offset, bv_len and
2383 * bi_vecs, as the read request might have corrupted these
2385 rp = get_resync_pages(tbio);
2386 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2388 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2390 rp->raid_bio = r10_bio;
2391 tbio->bi_private = rp;
2392 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2393 tbio->bi_end_io = end_sync_write;
2395 bio_copy_data(tbio, fbio);
2397 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2398 atomic_inc(&r10_bio->remaining);
2399 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2401 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2402 tbio->bi_opf |= MD_FAILFAST;
2403 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2404 submit_bio_noacct(tbio);
2407 /* Now write out to any replacement devices
2410 for (i = 0; i < conf->copies; i++) {
2413 tbio = r10_bio->devs[i].repl_bio;
2414 if (!tbio || !tbio->bi_end_io)
2416 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2417 && r10_bio->devs[i].bio != fbio)
2418 bio_copy_data(tbio, fbio);
2419 d = r10_bio->devs[i].devnum;
2420 atomic_inc(&r10_bio->remaining);
2421 md_sync_acct(conf->mirrors[d].replacement->bdev,
2423 submit_bio_noacct(tbio);
2427 if (atomic_dec_and_test(&r10_bio->remaining)) {
2428 md_done_sync(mddev, r10_bio->sectors, 1);
2434 * Now for the recovery code.
2435 * Recovery happens across physical sectors.
2436 * We recover all non-is_sync drives by finding the virtual address of
2437 * each, and then choose a working drive that also has that virt address.
2438 * There is a separate r10_bio for each non-in_sync drive.
2439 * Only the first two slots are in use. The first for reading,
2440 * The second for writing.
2443 static void fix_recovery_read_error(struct r10bio *r10_bio)
2445 /* We got a read error during recovery.
2446 * We repeat the read in smaller page-sized sections.
2447 * If a read succeeds, write it to the new device or record
2448 * a bad block if we cannot.
2449 * If a read fails, record a bad block on both old and
2452 struct mddev *mddev = r10_bio->mddev;
2453 struct r10conf *conf = mddev->private;
2454 struct bio *bio = r10_bio->devs[0].bio;
2456 int sectors = r10_bio->sectors;
2458 int dr = r10_bio->devs[0].devnum;
2459 int dw = r10_bio->devs[1].devnum;
2460 struct page **pages = get_resync_pages(bio)->pages;
2464 struct md_rdev *rdev;
2468 if (s > (PAGE_SIZE>>9))
2471 rdev = conf->mirrors[dr].rdev;
2472 addr = r10_bio->devs[0].addr + sect,
2473 ok = sync_page_io(rdev,
2477 REQ_OP_READ, false);
2479 rdev = conf->mirrors[dw].rdev;
2480 addr = r10_bio->devs[1].addr + sect;
2481 ok = sync_page_io(rdev,
2485 REQ_OP_WRITE, false);
2487 set_bit(WriteErrorSeen, &rdev->flags);
2488 if (!test_and_set_bit(WantReplacement,
2490 set_bit(MD_RECOVERY_NEEDED,
2491 &rdev->mddev->recovery);
2495 /* We don't worry if we cannot set a bad block -
2496 * it really is bad so there is no loss in not
2499 rdev_set_badblocks(rdev, addr, s, 0);
2501 if (rdev != conf->mirrors[dw].rdev) {
2502 /* need bad block on destination too */
2503 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2504 addr = r10_bio->devs[1].addr + sect;
2505 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2507 /* just abort the recovery */
2508 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2511 conf->mirrors[dw].recovery_disabled
2512 = mddev->recovery_disabled;
2513 set_bit(MD_RECOVERY_INTR,
2526 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2528 struct r10conf *conf = mddev->private;
2530 struct bio *wbio = r10_bio->devs[1].bio;
2531 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2533 /* Need to test wbio2->bi_end_io before we call
2534 * submit_bio_noacct as if the former is NULL,
2535 * the latter is free to free wbio2.
2537 if (wbio2 && !wbio2->bi_end_io)
2540 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2541 fix_recovery_read_error(r10_bio);
2542 if (wbio->bi_end_io)
2543 end_sync_request(r10_bio);
2545 end_sync_request(r10_bio);
2550 * share the pages with the first bio
2551 * and submit the write request
2553 d = r10_bio->devs[1].devnum;
2554 if (wbio->bi_end_io) {
2555 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2556 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2557 submit_bio_noacct(wbio);
2560 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2561 md_sync_acct(conf->mirrors[d].replacement->bdev,
2562 bio_sectors(wbio2));
2563 submit_bio_noacct(wbio2);
2567 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2568 int sectors, struct page *page, enum req_op op)
2570 if (rdev_has_badblock(rdev, sector, sectors) &&
2571 (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2573 if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2576 if (op == REQ_OP_WRITE) {
2577 set_bit(WriteErrorSeen, &rdev->flags);
2578 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2579 set_bit(MD_RECOVERY_NEEDED,
2580 &rdev->mddev->recovery);
2582 /* need to record an error - either for the block or the device */
2583 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2584 md_error(rdev->mddev, rdev);
2589 * This is a kernel thread which:
2591 * 1. Retries failed read operations on working mirrors.
2592 * 2. Updates the raid superblock when problems encounter.
2593 * 3. Performs writes following reads for array synchronising.
2596 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2598 int sect = 0; /* Offset from r10_bio->sector */
2599 int sectors = r10_bio->sectors, slot = r10_bio->read_slot;
2600 struct md_rdev *rdev;
2601 int d = r10_bio->devs[slot].devnum;
2603 /* still own a reference to this rdev, so it cannot
2604 * have been cleared recently.
2606 rdev = conf->mirrors[d].rdev;
2608 if (test_bit(Faulty, &rdev->flags))
2609 /* drive has already been failed, just ignore any
2610 more fix_read_error() attempts */
2613 if (exceed_read_errors(mddev, rdev)) {
2614 r10_bio->devs[slot].bio = IO_BLOCKED;
2624 if (s > (PAGE_SIZE>>9))
2628 d = r10_bio->devs[sl].devnum;
2629 rdev = conf->mirrors[d].rdev;
2631 test_bit(In_sync, &rdev->flags) &&
2632 !test_bit(Faulty, &rdev->flags) &&
2633 rdev_has_badblock(rdev,
2634 r10_bio->devs[sl].addr + sect,
2636 atomic_inc(&rdev->nr_pending);
2637 success = sync_page_io(rdev,
2638 r10_bio->devs[sl].addr +
2642 REQ_OP_READ, false);
2643 rdev_dec_pending(rdev, mddev);
2648 if (sl == conf->copies)
2650 } while (sl != slot);
2653 /* Cannot read from anywhere, just mark the block
2654 * as bad on the first device to discourage future
2657 int dn = r10_bio->devs[slot].devnum;
2658 rdev = conf->mirrors[dn].rdev;
2660 if (!rdev_set_badblocks(
2662 r10_bio->devs[slot].addr
2665 md_error(mddev, rdev);
2666 r10_bio->devs[slot].bio
2673 /* write it back and re-read */
2674 while (sl != slot) {
2678 d = r10_bio->devs[sl].devnum;
2679 rdev = conf->mirrors[d].rdev;
2681 test_bit(Faulty, &rdev->flags) ||
2682 !test_bit(In_sync, &rdev->flags))
2685 atomic_inc(&rdev->nr_pending);
2686 if (r10_sync_page_io(rdev,
2687 r10_bio->devs[sl].addr +
2689 s, conf->tmppage, REQ_OP_WRITE)
2691 /* Well, this device is dead */
2692 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2694 (unsigned long long)(
2696 choose_data_offset(r10_bio,
2699 pr_notice("md/raid10:%s: %pg: failing drive\n",
2703 rdev_dec_pending(rdev, mddev);
2706 while (sl != slot) {
2710 d = r10_bio->devs[sl].devnum;
2711 rdev = conf->mirrors[d].rdev;
2713 test_bit(Faulty, &rdev->flags) ||
2714 !test_bit(In_sync, &rdev->flags))
2717 atomic_inc(&rdev->nr_pending);
2718 switch (r10_sync_page_io(rdev,
2719 r10_bio->devs[sl].addr +
2721 s, conf->tmppage, REQ_OP_READ)) {
2723 /* Well, this device is dead */
2724 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2726 (unsigned long long)(
2728 choose_data_offset(r10_bio, rdev)),
2730 pr_notice("md/raid10:%s: %pg: failing drive\n",
2735 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2737 (unsigned long long)(
2739 choose_data_offset(r10_bio, rdev)),
2741 atomic_add(s, &rdev->corrected_errors);
2744 rdev_dec_pending(rdev, mddev);
2752 static int narrow_write_error(struct r10bio *r10_bio, int i)
2754 struct bio *bio = r10_bio->master_bio;
2755 struct mddev *mddev = r10_bio->mddev;
2756 struct r10conf *conf = mddev->private;
2757 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2758 /* bio has the data to be written to slot 'i' where
2759 * we just recently had a write error.
2760 * We repeatedly clone the bio and trim down to one block,
2761 * then try the write. Where the write fails we record
2763 * It is conceivable that the bio doesn't exactly align with
2764 * blocks. We must handle this.
2766 * We currently own a reference to the rdev.
2772 int sect_to_write = r10_bio->sectors;
2775 if (rdev->badblocks.shift < 0)
2778 block_sectors = roundup(1 << rdev->badblocks.shift,
2779 bdev_logical_block_size(rdev->bdev) >> 9);
2780 sector = r10_bio->sector;
2781 sectors = ((r10_bio->sector + block_sectors)
2782 & ~(sector_t)(block_sectors - 1))
2785 while (sect_to_write) {
2788 if (sectors > sect_to_write)
2789 sectors = sect_to_write;
2790 /* Write at 'sector' for 'sectors' */
2791 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2793 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2794 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2795 wbio->bi_iter.bi_sector = wsector +
2796 choose_data_offset(r10_bio, rdev);
2797 wbio->bi_opf = REQ_OP_WRITE;
2799 if (submit_bio_wait(wbio) < 0)
2801 ok = rdev_set_badblocks(rdev, wsector,
2806 sect_to_write -= sectors;
2808 sectors = block_sectors;
2813 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2815 int slot = r10_bio->read_slot;
2817 struct r10conf *conf = mddev->private;
2818 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2820 /* we got a read error. Maybe the drive is bad. Maybe just
2821 * the block and we can fix it.
2822 * We freeze all other IO, and try reading the block from
2823 * other devices. When we find one, we re-write
2824 * and check it that fixes the read error.
2825 * This is all done synchronously while the array is
2828 bio = r10_bio->devs[slot].bio;
2830 r10_bio->devs[slot].bio = NULL;
2833 r10_bio->devs[slot].bio = IO_BLOCKED;
2834 else if (!test_bit(FailFast, &rdev->flags)) {
2835 freeze_array(conf, 1);
2836 fix_read_error(conf, mddev, r10_bio);
2837 unfreeze_array(conf);
2839 md_error(mddev, rdev);
2841 rdev_dec_pending(rdev, mddev);
2843 raid10_read_request(mddev, r10_bio->master_bio, r10_bio, false);
2845 * allow_barrier after re-submit to ensure no sync io
2846 * can be issued while regular io pending.
2848 allow_barrier(conf);
2851 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2853 /* Some sort of write request has finished and it
2854 * succeeded in writing where we thought there was a
2855 * bad block. So forget the bad block.
2856 * Or possibly if failed and we need to record
2860 struct md_rdev *rdev;
2862 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2863 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2864 for (m = 0; m < conf->copies; m++) {
2865 int dev = r10_bio->devs[m].devnum;
2866 rdev = conf->mirrors[dev].rdev;
2867 if (r10_bio->devs[m].bio == NULL ||
2868 r10_bio->devs[m].bio->bi_end_io == NULL)
2870 if (!r10_bio->devs[m].bio->bi_status) {
2871 rdev_clear_badblocks(
2873 r10_bio->devs[m].addr,
2874 r10_bio->sectors, 0);
2876 if (!rdev_set_badblocks(
2878 r10_bio->devs[m].addr,
2879 r10_bio->sectors, 0))
2880 md_error(conf->mddev, rdev);
2882 rdev = conf->mirrors[dev].replacement;
2883 if (r10_bio->devs[m].repl_bio == NULL ||
2884 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2887 if (!r10_bio->devs[m].repl_bio->bi_status) {
2888 rdev_clear_badblocks(
2890 r10_bio->devs[m].addr,
2891 r10_bio->sectors, 0);
2893 if (!rdev_set_badblocks(
2895 r10_bio->devs[m].addr,
2896 r10_bio->sectors, 0))
2897 md_error(conf->mddev, rdev);
2903 for (m = 0; m < conf->copies; m++) {
2904 int dev = r10_bio->devs[m].devnum;
2905 struct bio *bio = r10_bio->devs[m].bio;
2906 rdev = conf->mirrors[dev].rdev;
2907 if (bio == IO_MADE_GOOD) {
2908 rdev_clear_badblocks(
2910 r10_bio->devs[m].addr,
2911 r10_bio->sectors, 0);
2912 rdev_dec_pending(rdev, conf->mddev);
2913 } else if (bio != NULL && bio->bi_status) {
2915 if (!narrow_write_error(r10_bio, m)) {
2916 md_error(conf->mddev, rdev);
2917 set_bit(R10BIO_Degraded,
2920 rdev_dec_pending(rdev, conf->mddev);
2922 bio = r10_bio->devs[m].repl_bio;
2923 rdev = conf->mirrors[dev].replacement;
2924 if (rdev && bio == IO_MADE_GOOD) {
2925 rdev_clear_badblocks(
2927 r10_bio->devs[m].addr,
2928 r10_bio->sectors, 0);
2929 rdev_dec_pending(rdev, conf->mddev);
2933 spin_lock_irq(&conf->device_lock);
2934 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2936 spin_unlock_irq(&conf->device_lock);
2938 * In case freeze_array() is waiting for condition
2939 * nr_pending == nr_queued + extra to be true.
2941 wake_up(&conf->wait_barrier);
2942 md_wakeup_thread(conf->mddev->thread);
2944 if (test_bit(R10BIO_WriteError,
2946 close_write(r10_bio);
2947 raid_end_bio_io(r10_bio);
2952 static void raid10d(struct md_thread *thread)
2954 struct mddev *mddev = thread->mddev;
2955 struct r10bio *r10_bio;
2956 unsigned long flags;
2957 struct r10conf *conf = mddev->private;
2958 struct list_head *head = &conf->retry_list;
2959 struct blk_plug plug;
2961 md_check_recovery(mddev);
2963 if (!list_empty_careful(&conf->bio_end_io_list) &&
2964 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2966 spin_lock_irqsave(&conf->device_lock, flags);
2967 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2968 while (!list_empty(&conf->bio_end_io_list)) {
2969 list_move(conf->bio_end_io_list.prev, &tmp);
2973 spin_unlock_irqrestore(&conf->device_lock, flags);
2974 while (!list_empty(&tmp)) {
2975 r10_bio = list_first_entry(&tmp, struct r10bio,
2977 list_del(&r10_bio->retry_list);
2978 if (mddev->degraded)
2979 set_bit(R10BIO_Degraded, &r10_bio->state);
2981 if (test_bit(R10BIO_WriteError,
2983 close_write(r10_bio);
2984 raid_end_bio_io(r10_bio);
2988 blk_start_plug(&plug);
2991 flush_pending_writes(conf);
2993 spin_lock_irqsave(&conf->device_lock, flags);
2994 if (list_empty(head)) {
2995 spin_unlock_irqrestore(&conf->device_lock, flags);
2998 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2999 list_del(head->prev);
3001 spin_unlock_irqrestore(&conf->device_lock, flags);
3003 mddev = r10_bio->mddev;
3004 conf = mddev->private;
3005 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3006 test_bit(R10BIO_WriteError, &r10_bio->state))
3007 handle_write_completed(conf, r10_bio);
3008 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3009 reshape_request_write(mddev, r10_bio);
3010 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3011 sync_request_write(mddev, r10_bio);
3012 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3013 recovery_request_write(mddev, r10_bio);
3014 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3015 handle_read_error(mddev, r10_bio);
3020 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3021 md_check_recovery(mddev);
3023 blk_finish_plug(&plug);
3026 static int init_resync(struct r10conf *conf)
3030 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3031 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3032 conf->have_replacement = 0;
3033 for (i = 0; i < conf->geo.raid_disks; i++)
3034 if (conf->mirrors[i].replacement)
3035 conf->have_replacement = 1;
3036 ret = mempool_init(&conf->r10buf_pool, buffs,
3037 r10buf_pool_alloc, r10buf_pool_free, conf);
3040 conf->next_resync = 0;
3044 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3046 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3047 struct rsync_pages *rp;
3052 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3053 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3054 nalloc = conf->copies; /* resync */
3056 nalloc = 2; /* recovery */
3058 for (i = 0; i < nalloc; i++) {
3059 bio = r10bio->devs[i].bio;
3060 rp = bio->bi_private;
3061 bio_reset(bio, NULL, 0);
3062 bio->bi_private = rp;
3063 bio = r10bio->devs[i].repl_bio;
3065 rp = bio->bi_private;
3066 bio_reset(bio, NULL, 0);
3067 bio->bi_private = rp;
3074 * Set cluster_sync_high since we need other nodes to add the
3075 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3077 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3079 sector_t window_size;
3080 int extra_chunk, chunks;
3083 * First, here we define "stripe" as a unit which across
3084 * all member devices one time, so we get chunks by use
3085 * raid_disks / near_copies. Otherwise, if near_copies is
3086 * close to raid_disks, then resync window could increases
3087 * linearly with the increase of raid_disks, which means
3088 * we will suspend a really large IO window while it is not
3089 * necessary. If raid_disks is not divisible by near_copies,
3090 * an extra chunk is needed to ensure the whole "stripe" is
3094 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3095 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3099 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3102 * At least use a 32M window to align with raid1's resync window
3104 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3105 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3107 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3111 * perform a "sync" on one "block"
3113 * We need to make sure that no normal I/O request - particularly write
3114 * requests - conflict with active sync requests.
3116 * This is achieved by tracking pending requests and a 'barrier' concept
3117 * that can be installed to exclude normal IO requests.
3119 * Resync and recovery are handled very differently.
3120 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3122 * For resync, we iterate over virtual addresses, read all copies,
3123 * and update if there are differences. If only one copy is live,
3125 * For recovery, we iterate over physical addresses, read a good
3126 * value for each non-in_sync drive, and over-write.
3128 * So, for recovery we may have several outstanding complex requests for a
3129 * given address, one for each out-of-sync device. We model this by allocating
3130 * a number of r10_bio structures, one for each out-of-sync device.
3131 * As we setup these structures, we collect all bio's together into a list
3132 * which we then process collectively to add pages, and then process again
3133 * to pass to submit_bio_noacct.
3135 * The r10_bio structures are linked using a borrowed master_bio pointer.
3136 * This link is counted in ->remaining. When the r10_bio that points to NULL
3137 * has its remaining count decremented to 0, the whole complex operation
3142 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3145 struct r10conf *conf = mddev->private;
3146 struct r10bio *r10_bio;
3147 struct bio *biolist = NULL, *bio;
3148 sector_t max_sector, nr_sectors;
3151 sector_t sync_blocks;
3152 sector_t sectors_skipped = 0;
3153 int chunks_skipped = 0;
3154 sector_t chunk_mask = conf->geo.chunk_mask;
3156 int error_disk = -1;
3159 * Allow skipping a full rebuild for incremental assembly
3160 * of a clean array, like RAID1 does.
3162 if (mddev->bitmap == NULL &&
3163 mddev->recovery_cp == MaxSector &&
3164 mddev->reshape_position == MaxSector &&
3165 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3166 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3167 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3168 conf->fullsync == 0) {
3170 return mddev->dev_sectors - sector_nr;
3173 if (!mempool_initialized(&conf->r10buf_pool))
3174 if (init_resync(conf))
3178 max_sector = mddev->dev_sectors;
3179 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3180 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3181 max_sector = mddev->resync_max_sectors;
3182 if (sector_nr >= max_sector) {
3183 conf->cluster_sync_low = 0;
3184 conf->cluster_sync_high = 0;
3186 /* If we aborted, we need to abort the
3187 * sync on the 'current' bitmap chucks (there can
3188 * be several when recovering multiple devices).
3189 * as we may have started syncing it but not finished.
3190 * We can find the current address in
3191 * mddev->curr_resync, but for recovery,
3192 * we need to convert that to several
3193 * virtual addresses.
3195 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3201 if (mddev->curr_resync < max_sector) { /* aborted */
3202 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3203 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3205 else for (i = 0; i < conf->geo.raid_disks; i++) {
3207 raid10_find_virt(conf, mddev->curr_resync, i);
3208 md_bitmap_end_sync(mddev->bitmap, sect,
3212 /* completed sync */
3213 if ((!mddev->bitmap || conf->fullsync)
3214 && conf->have_replacement
3215 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3216 /* Completed a full sync so the replacements
3217 * are now fully recovered.
3219 for (i = 0; i < conf->geo.raid_disks; i++) {
3220 struct md_rdev *rdev =
3221 conf->mirrors[i].replacement;
3224 rdev->recovery_offset = MaxSector;
3229 md_bitmap_close_sync(mddev->bitmap);
3232 return sectors_skipped;
3235 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3236 return reshape_request(mddev, sector_nr, skipped);
3238 if (chunks_skipped >= conf->geo.raid_disks) {
3239 pr_err("md/raid10:%s: %s fails\n", mdname(mddev),
3240 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
3241 if (error_disk >= 0 &&
3242 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3244 * recovery fails, set mirrors.recovery_disabled,
3245 * device shouldn't be added to there.
3247 conf->mirrors[error_disk].recovery_disabled =
3248 mddev->recovery_disabled;
3252 * if there has been nothing to do on any drive,
3253 * then there is nothing to do at all.
3256 return (max_sector - sector_nr) + sectors_skipped;
3259 if (max_sector > mddev->resync_max)
3260 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3262 /* make sure whole request will fit in a chunk - if chunks
3265 if (conf->geo.near_copies < conf->geo.raid_disks &&
3266 max_sector > (sector_nr | chunk_mask))
3267 max_sector = (sector_nr | chunk_mask) + 1;
3270 * If there is non-resync activity waiting for a turn, then let it
3271 * though before starting on this new sync request.
3273 if (conf->nr_waiting)
3274 schedule_timeout_uninterruptible(1);
3276 /* Again, very different code for resync and recovery.
3277 * Both must result in an r10bio with a list of bios that
3278 * have bi_end_io, bi_sector, bi_bdev set,
3279 * and bi_private set to the r10bio.
3280 * For recovery, we may actually create several r10bios
3281 * with 2 bios in each, that correspond to the bios in the main one.
3282 * In this case, the subordinate r10bios link back through a
3283 * borrowed master_bio pointer, and the counter in the master
3284 * includes a ref from each subordinate.
3286 /* First, we decide what to do and set ->bi_end_io
3287 * To end_sync_read if we want to read, and
3288 * end_sync_write if we will want to write.
3291 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3292 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3293 /* recovery... the complicated one */
3297 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3303 struct raid10_info *mirror = &conf->mirrors[i];
3304 struct md_rdev *mrdev, *mreplace;
3306 mrdev = mirror->rdev;
3307 mreplace = mirror->replacement;
3309 if (mrdev && (test_bit(Faulty, &mrdev->flags) ||
3310 test_bit(In_sync, &mrdev->flags)))
3312 if (mreplace && test_bit(Faulty, &mreplace->flags))
3315 if (!mrdev && !mreplace)
3319 /* want to reconstruct this device */
3321 sect = raid10_find_virt(conf, sector_nr, i);
3322 if (sect >= mddev->resync_max_sectors)
3323 /* last stripe is not complete - don't
3324 * try to recover this sector.
3327 /* Unless we are doing a full sync, or a replacement
3328 * we only need to recover the block if it is set in
3331 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3333 if (sync_blocks < max_sync)
3334 max_sync = sync_blocks;
3338 /* yep, skip the sync_blocks here, but don't assume
3339 * that there will never be anything to do here
3341 chunks_skipped = -1;
3345 atomic_inc(&mrdev->nr_pending);
3347 atomic_inc(&mreplace->nr_pending);
3349 r10_bio = raid10_alloc_init_r10buf(conf);
3351 raise_barrier(conf, rb2 != NULL);
3352 atomic_set(&r10_bio->remaining, 0);
3354 r10_bio->master_bio = (struct bio*)rb2;
3356 atomic_inc(&rb2->remaining);
3357 r10_bio->mddev = mddev;
3358 set_bit(R10BIO_IsRecover, &r10_bio->state);
3359 r10_bio->sector = sect;
3361 raid10_find_phys(conf, r10_bio);
3363 /* Need to check if the array will still be
3366 for (j = 0; j < conf->geo.raid_disks; j++) {
3367 struct md_rdev *rdev = conf->mirrors[j].rdev;
3369 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3375 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3376 &sync_blocks, still_degraded);
3379 for (j=0; j<conf->copies;j++) {
3381 int d = r10_bio->devs[j].devnum;
3382 sector_t from_addr, to_addr;
3383 struct md_rdev *rdev = conf->mirrors[d].rdev;
3384 sector_t sector, first_bad;
3387 !test_bit(In_sync, &rdev->flags))
3389 /* This is where we read from */
3391 sector = r10_bio->devs[j].addr;
3393 if (is_badblock(rdev, sector, max_sync,
3394 &first_bad, &bad_sectors)) {
3395 if (first_bad > sector)
3396 max_sync = first_bad - sector;
3398 bad_sectors -= (sector
3400 if (max_sync > bad_sectors)
3401 max_sync = bad_sectors;
3405 bio = r10_bio->devs[0].bio;
3406 bio->bi_next = biolist;
3408 bio->bi_end_io = end_sync_read;
3409 bio->bi_opf = REQ_OP_READ;
3410 if (test_bit(FailFast, &rdev->flags))
3411 bio->bi_opf |= MD_FAILFAST;
3412 from_addr = r10_bio->devs[j].addr;
3413 bio->bi_iter.bi_sector = from_addr +
3415 bio_set_dev(bio, rdev->bdev);
3416 atomic_inc(&rdev->nr_pending);
3417 /* and we write to 'i' (if not in_sync) */
3419 for (k=0; k<conf->copies; k++)
3420 if (r10_bio->devs[k].devnum == i)
3422 BUG_ON(k == conf->copies);
3423 to_addr = r10_bio->devs[k].addr;
3424 r10_bio->devs[0].devnum = d;
3425 r10_bio->devs[0].addr = from_addr;
3426 r10_bio->devs[1].devnum = i;
3427 r10_bio->devs[1].addr = to_addr;
3430 bio = r10_bio->devs[1].bio;
3431 bio->bi_next = biolist;
3433 bio->bi_end_io = end_sync_write;
3434 bio->bi_opf = REQ_OP_WRITE;
3435 bio->bi_iter.bi_sector = to_addr
3436 + mrdev->data_offset;
3437 bio_set_dev(bio, mrdev->bdev);
3438 atomic_inc(&r10_bio->remaining);
3440 r10_bio->devs[1].bio->bi_end_io = NULL;
3442 /* and maybe write to replacement */
3443 bio = r10_bio->devs[1].repl_bio;
3445 bio->bi_end_io = NULL;
3446 /* Note: if replace is not NULL, then bio
3447 * cannot be NULL as r10buf_pool_alloc will
3448 * have allocated it.
3452 bio->bi_next = biolist;
3454 bio->bi_end_io = end_sync_write;
3455 bio->bi_opf = REQ_OP_WRITE;
3456 bio->bi_iter.bi_sector = to_addr +
3457 mreplace->data_offset;
3458 bio_set_dev(bio, mreplace->bdev);
3459 atomic_inc(&r10_bio->remaining);
3462 if (j == conf->copies) {
3463 /* Cannot recover, so abort the recovery or
3464 * record a bad block */
3466 /* problem is that there are bad blocks
3467 * on other device(s)
3470 for (k = 0; k < conf->copies; k++)
3471 if (r10_bio->devs[k].devnum == i)
3473 if (mrdev && !test_bit(In_sync,
3475 && !rdev_set_badblocks(
3477 r10_bio->devs[k].addr,
3481 !rdev_set_badblocks(
3483 r10_bio->devs[k].addr,
3488 if (!test_and_set_bit(MD_RECOVERY_INTR,
3490 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3492 mirror->recovery_disabled
3493 = mddev->recovery_disabled;
3499 atomic_dec(&rb2->remaining);
3502 rdev_dec_pending(mrdev, mddev);
3504 rdev_dec_pending(mreplace, mddev);
3508 rdev_dec_pending(mrdev, mddev);
3510 rdev_dec_pending(mreplace, mddev);
3511 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3512 /* Only want this if there is elsewhere to
3513 * read from. 'j' is currently the first
3517 for (; j < conf->copies; j++) {
3518 int d = r10_bio->devs[j].devnum;
3519 if (conf->mirrors[d].rdev &&
3521 &conf->mirrors[d].rdev->flags))
3525 r10_bio->devs[0].bio->bi_opf
3529 if (biolist == NULL) {
3531 struct r10bio *rb2 = r10_bio;
3532 r10_bio = (struct r10bio*) rb2->master_bio;
3533 rb2->master_bio = NULL;
3539 /* resync. Schedule a read for every block at this virt offset */
3543 * Since curr_resync_completed could probably not update in
3544 * time, and we will set cluster_sync_low based on it.
3545 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3546 * safety reason, which ensures curr_resync_completed is
3547 * updated in bitmap_cond_end_sync.
3549 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3550 mddev_is_clustered(mddev) &&
3551 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3553 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3554 &sync_blocks, mddev->degraded) &&
3555 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3556 &mddev->recovery)) {
3557 /* We can skip this block */
3559 return sync_blocks + sectors_skipped;
3561 if (sync_blocks < max_sync)
3562 max_sync = sync_blocks;
3563 r10_bio = raid10_alloc_init_r10buf(conf);
3566 r10_bio->mddev = mddev;
3567 atomic_set(&r10_bio->remaining, 0);
3568 raise_barrier(conf, 0);
3569 conf->next_resync = sector_nr;
3571 r10_bio->master_bio = NULL;
3572 r10_bio->sector = sector_nr;
3573 set_bit(R10BIO_IsSync, &r10_bio->state);
3574 raid10_find_phys(conf, r10_bio);
3575 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3577 for (i = 0; i < conf->copies; i++) {
3578 int d = r10_bio->devs[i].devnum;
3579 sector_t first_bad, sector;
3581 struct md_rdev *rdev;
3583 if (r10_bio->devs[i].repl_bio)
3584 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3586 bio = r10_bio->devs[i].bio;
3587 bio->bi_status = BLK_STS_IOERR;
3588 rdev = conf->mirrors[d].rdev;
3589 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3592 sector = r10_bio->devs[i].addr;
3593 if (is_badblock(rdev, sector, max_sync,
3594 &first_bad, &bad_sectors)) {
3595 if (first_bad > sector)
3596 max_sync = first_bad - sector;
3598 bad_sectors -= (sector - first_bad);
3599 if (max_sync > bad_sectors)
3600 max_sync = bad_sectors;
3604 atomic_inc(&rdev->nr_pending);
3605 atomic_inc(&r10_bio->remaining);
3606 bio->bi_next = biolist;
3608 bio->bi_end_io = end_sync_read;
3609 bio->bi_opf = REQ_OP_READ;
3610 if (test_bit(FailFast, &rdev->flags))
3611 bio->bi_opf |= MD_FAILFAST;
3612 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3613 bio_set_dev(bio, rdev->bdev);
3616 rdev = conf->mirrors[d].replacement;
3617 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3620 atomic_inc(&rdev->nr_pending);
3622 /* Need to set up for writing to the replacement */
3623 bio = r10_bio->devs[i].repl_bio;
3624 bio->bi_status = BLK_STS_IOERR;
3626 sector = r10_bio->devs[i].addr;
3627 bio->bi_next = biolist;
3629 bio->bi_end_io = end_sync_write;
3630 bio->bi_opf = REQ_OP_WRITE;
3631 if (test_bit(FailFast, &rdev->flags))
3632 bio->bi_opf |= MD_FAILFAST;
3633 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3634 bio_set_dev(bio, rdev->bdev);
3639 for (i=0; i<conf->copies; i++) {
3640 int d = r10_bio->devs[i].devnum;
3641 if (r10_bio->devs[i].bio->bi_end_io)
3642 rdev_dec_pending(conf->mirrors[d].rdev,
3644 if (r10_bio->devs[i].repl_bio &&
3645 r10_bio->devs[i].repl_bio->bi_end_io)
3647 conf->mirrors[d].replacement,
3657 if (sector_nr + max_sync < max_sector)
3658 max_sector = sector_nr + max_sync;
3661 int len = PAGE_SIZE;
3662 if (sector_nr + (len>>9) > max_sector)
3663 len = (max_sector - sector_nr) << 9;
3666 for (bio= biolist ; bio ; bio=bio->bi_next) {
3667 struct resync_pages *rp = get_resync_pages(bio);
3668 page = resync_fetch_page(rp, page_idx);
3669 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
3670 bio->bi_status = BLK_STS_RESOURCE;
3675 nr_sectors += len>>9;
3676 sector_nr += len>>9;
3677 } while (++page_idx < RESYNC_PAGES);
3678 r10_bio->sectors = nr_sectors;
3680 if (mddev_is_clustered(mddev) &&
3681 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3682 /* It is resync not recovery */
3683 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3684 conf->cluster_sync_low = mddev->curr_resync_completed;
3685 raid10_set_cluster_sync_high(conf);
3686 /* Send resync message */
3687 md_cluster_ops->resync_info_update(mddev,
3688 conf->cluster_sync_low,
3689 conf->cluster_sync_high);
3691 } else if (mddev_is_clustered(mddev)) {
3692 /* This is recovery not resync */
3693 sector_t sect_va1, sect_va2;
3694 bool broadcast_msg = false;
3696 for (i = 0; i < conf->geo.raid_disks; i++) {
3698 * sector_nr is a device address for recovery, so we
3699 * need translate it to array address before compare
3700 * with cluster_sync_high.
3702 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3704 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3705 broadcast_msg = true;
3707 * curr_resync_completed is similar as
3708 * sector_nr, so make the translation too.
3710 sect_va2 = raid10_find_virt(conf,
3711 mddev->curr_resync_completed, i);
3713 if (conf->cluster_sync_low == 0 ||
3714 conf->cluster_sync_low > sect_va2)
3715 conf->cluster_sync_low = sect_va2;
3718 if (broadcast_msg) {
3719 raid10_set_cluster_sync_high(conf);
3720 md_cluster_ops->resync_info_update(mddev,
3721 conf->cluster_sync_low,
3722 conf->cluster_sync_high);
3728 biolist = biolist->bi_next;
3730 bio->bi_next = NULL;
3731 r10_bio = get_resync_r10bio(bio);
3732 r10_bio->sectors = nr_sectors;
3734 if (bio->bi_end_io == end_sync_read) {
3735 md_sync_acct_bio(bio, nr_sectors);
3737 submit_bio_noacct(bio);
3741 if (sectors_skipped)
3742 /* pretend they weren't skipped, it makes
3743 * no important difference in this case
3745 md_done_sync(mddev, sectors_skipped, 1);
3747 return sectors_skipped + nr_sectors;
3749 /* There is nowhere to write, so all non-sync
3750 * drives must be failed or in resync, all drives
3751 * have a bad block, so try the next chunk...
3753 if (sector_nr + max_sync < max_sector)
3754 max_sector = sector_nr + max_sync;
3756 sectors_skipped += (max_sector - sector_nr);
3758 sector_nr = max_sector;
3763 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3766 struct r10conf *conf = mddev->private;
3769 raid_disks = min(conf->geo.raid_disks,
3770 conf->prev.raid_disks);
3772 sectors = conf->dev_sectors;
3774 size = sectors >> conf->geo.chunk_shift;
3775 sector_div(size, conf->geo.far_copies);
3776 size = size * raid_disks;
3777 sector_div(size, conf->geo.near_copies);
3779 return size << conf->geo.chunk_shift;
3782 static void calc_sectors(struct r10conf *conf, sector_t size)
3784 /* Calculate the number of sectors-per-device that will
3785 * actually be used, and set conf->dev_sectors and
3789 size = size >> conf->geo.chunk_shift;
3790 sector_div(size, conf->geo.far_copies);
3791 size = size * conf->geo.raid_disks;
3792 sector_div(size, conf->geo.near_copies);
3793 /* 'size' is now the number of chunks in the array */
3794 /* calculate "used chunks per device" */
3795 size = size * conf->copies;
3797 /* We need to round up when dividing by raid_disks to
3798 * get the stride size.
3800 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3802 conf->dev_sectors = size << conf->geo.chunk_shift;
3804 if (conf->geo.far_offset)
3805 conf->geo.stride = 1 << conf->geo.chunk_shift;
3807 sector_div(size, conf->geo.far_copies);
3808 conf->geo.stride = size << conf->geo.chunk_shift;
3812 enum geo_type {geo_new, geo_old, geo_start};
3813 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3816 int layout, chunk, disks;
3819 layout = mddev->layout;
3820 chunk = mddev->chunk_sectors;
3821 disks = mddev->raid_disks - mddev->delta_disks;
3824 layout = mddev->new_layout;
3825 chunk = mddev->new_chunk_sectors;
3826 disks = mddev->raid_disks;
3828 default: /* avoid 'may be unused' warnings */
3829 case geo_start: /* new when starting reshape - raid_disks not
3831 layout = mddev->new_layout;
3832 chunk = mddev->new_chunk_sectors;
3833 disks = mddev->raid_disks + mddev->delta_disks;
3838 if (chunk < (PAGE_SIZE >> 9) ||
3839 !is_power_of_2(chunk))
3842 fc = (layout >> 8) & 255;
3843 fo = layout & (1<<16);
3844 geo->raid_disks = disks;
3845 geo->near_copies = nc;
3846 geo->far_copies = fc;
3847 geo->far_offset = fo;
3848 switch (layout >> 17) {
3849 case 0: /* original layout. simple but not always optimal */
3850 geo->far_set_size = disks;
3852 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3853 * actually using this, but leave code here just in case.*/
3854 geo->far_set_size = disks/fc;
3855 WARN(geo->far_set_size < fc,
3856 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3858 case 2: /* "improved" layout fixed to match documentation */
3859 geo->far_set_size = fc * nc;
3861 default: /* Not a valid layout */
3864 geo->chunk_mask = chunk - 1;
3865 geo->chunk_shift = ffz(~chunk);
3869 static void raid10_free_conf(struct r10conf *conf)
3874 mempool_exit(&conf->r10bio_pool);
3875 kfree(conf->mirrors);
3876 kfree(conf->mirrors_old);
3877 kfree(conf->mirrors_new);
3878 safe_put_page(conf->tmppage);
3879 bioset_exit(&conf->bio_split);
3883 static struct r10conf *setup_conf(struct mddev *mddev)
3885 struct r10conf *conf = NULL;
3890 copies = setup_geo(&geo, mddev, geo_new);
3893 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3894 mdname(mddev), PAGE_SIZE);
3898 if (copies < 2 || copies > mddev->raid_disks) {
3899 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3900 mdname(mddev), mddev->new_layout);
3905 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3909 /* FIXME calc properly */
3910 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3911 sizeof(struct raid10_info),
3916 conf->tmppage = alloc_page(GFP_KERNEL);
3921 conf->copies = copies;
3922 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3923 rbio_pool_free, conf);
3927 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3931 calc_sectors(conf, mddev->dev_sectors);
3932 if (mddev->reshape_position == MaxSector) {
3933 conf->prev = conf->geo;
3934 conf->reshape_progress = MaxSector;
3936 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3940 conf->reshape_progress = mddev->reshape_position;
3941 if (conf->prev.far_offset)
3942 conf->prev.stride = 1 << conf->prev.chunk_shift;
3944 /* far_copies must be 1 */
3945 conf->prev.stride = conf->dev_sectors;
3947 conf->reshape_safe = conf->reshape_progress;
3948 spin_lock_init(&conf->device_lock);
3949 INIT_LIST_HEAD(&conf->retry_list);
3950 INIT_LIST_HEAD(&conf->bio_end_io_list);
3952 seqlock_init(&conf->resync_lock);
3953 init_waitqueue_head(&conf->wait_barrier);
3954 atomic_set(&conf->nr_pending, 0);
3957 rcu_assign_pointer(conf->thread,
3958 md_register_thread(raid10d, mddev, "raid10"));
3962 conf->mddev = mddev;
3966 raid10_free_conf(conf);
3967 return ERR_PTR(err);
3970 static unsigned int raid10_nr_stripes(struct r10conf *conf)
3972 unsigned int raid_disks = conf->geo.raid_disks;
3974 if (conf->geo.raid_disks % conf->geo.near_copies)
3976 return raid_disks / conf->geo.near_copies;
3979 static int raid10_set_queue_limits(struct mddev *mddev)
3981 struct r10conf *conf = mddev->private;
3982 struct queue_limits lim;
3984 blk_set_stacking_limits(&lim);
3985 lim.max_write_zeroes_sectors = 0;
3986 lim.io_min = mddev->chunk_sectors << 9;
3987 lim.io_opt = lim.io_min * raid10_nr_stripes(conf);
3988 mddev_stack_rdev_limits(mddev, &lim);
3989 return queue_limits_set(mddev->gendisk->queue, &lim);
3992 static int raid10_run(struct mddev *mddev)
3994 struct r10conf *conf;
3996 struct raid10_info *disk;
3997 struct md_rdev *rdev;
3999 sector_t min_offset_diff = 0;
4003 if (mddev->private == NULL) {
4004 conf = setup_conf(mddev);
4006 return PTR_ERR(conf);
4007 mddev->private = conf;
4009 conf = mddev->private;
4013 rcu_assign_pointer(mddev->thread, conf->thread);
4014 rcu_assign_pointer(conf->thread, NULL);
4016 if (mddev_is_clustered(conf->mddev)) {
4019 fc = (mddev->layout >> 8) & 255;
4020 fo = mddev->layout & (1<<16);
4021 if (fc > 1 || fo > 0) {
4022 pr_err("only near layout is supported by clustered"
4028 rdev_for_each(rdev, mddev) {
4031 disk_idx = rdev->raid_disk;
4034 if (disk_idx >= conf->geo.raid_disks &&
4035 disk_idx >= conf->prev.raid_disks)
4037 disk = conf->mirrors + disk_idx;
4039 if (test_bit(Replacement, &rdev->flags)) {
4040 if (disk->replacement)
4042 disk->replacement = rdev;
4048 diff = (rdev->new_data_offset - rdev->data_offset);
4049 if (!mddev->reshape_backwards)
4053 if (first || diff < min_offset_diff)
4054 min_offset_diff = diff;
4056 disk->head_position = 0;
4060 if (!mddev_is_dm(conf->mddev)) {
4061 ret = raid10_set_queue_limits(mddev);
4066 /* need to check that every block has at least one working mirror */
4067 if (!enough(conf, -1)) {
4068 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4073 if (conf->reshape_progress != MaxSector) {
4074 /* must ensure that shape change is supported */
4075 if (conf->geo.far_copies != 1 &&
4076 conf->geo.far_offset == 0)
4078 if (conf->prev.far_copies != 1 &&
4079 conf->prev.far_offset == 0)
4083 mddev->degraded = 0;
4085 i < conf->geo.raid_disks
4086 || i < conf->prev.raid_disks;
4089 disk = conf->mirrors + i;
4091 if (!disk->rdev && disk->replacement) {
4092 /* The replacement is all we have - use it */
4093 disk->rdev = disk->replacement;
4094 disk->replacement = NULL;
4095 clear_bit(Replacement, &disk->rdev->flags);
4099 !test_bit(In_sync, &disk->rdev->flags)) {
4100 disk->head_position = 0;
4103 disk->rdev->saved_raid_disk < 0)
4107 if (disk->replacement &&
4108 !test_bit(In_sync, &disk->replacement->flags) &&
4109 disk->replacement->saved_raid_disk < 0) {
4113 disk->recovery_disabled = mddev->recovery_disabled - 1;
4116 if (mddev->recovery_cp != MaxSector)
4117 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4119 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4120 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4121 conf->geo.raid_disks);
4123 * Ok, everything is just fine now
4125 mddev->dev_sectors = conf->dev_sectors;
4126 size = raid10_size(mddev, 0, 0);
4127 md_set_array_sectors(mddev, size);
4128 mddev->resync_max_sectors = size;
4129 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4131 if (md_integrity_register(mddev))
4134 if (conf->reshape_progress != MaxSector) {
4135 unsigned long before_length, after_length;
4137 before_length = ((1 << conf->prev.chunk_shift) *
4138 conf->prev.far_copies);
4139 after_length = ((1 << conf->geo.chunk_shift) *
4140 conf->geo.far_copies);
4142 if (max(before_length, after_length) > min_offset_diff) {
4143 /* This cannot work */
4144 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4147 conf->offset_diff = min_offset_diff;
4149 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4150 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4151 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4152 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4158 md_unregister_thread(mddev, &mddev->thread);
4159 raid10_free_conf(conf);
4160 mddev->private = NULL;
4165 static void raid10_free(struct mddev *mddev, void *priv)
4167 raid10_free_conf(priv);
4170 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4172 struct r10conf *conf = mddev->private;
4175 raise_barrier(conf, 0);
4177 lower_barrier(conf);
4180 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4182 /* Resize of 'far' arrays is not supported.
4183 * For 'near' and 'offset' arrays we can set the
4184 * number of sectors used to be an appropriate multiple
4185 * of the chunk size.
4186 * For 'offset', this is far_copies*chunksize.
4187 * For 'near' the multiplier is the LCM of
4188 * near_copies and raid_disks.
4189 * So if far_copies > 1 && !far_offset, fail.
4190 * Else find LCM(raid_disks, near_copy)*far_copies and
4191 * multiply by chunk_size. Then round to this number.
4192 * This is mostly done by raid10_size()
4194 struct r10conf *conf = mddev->private;
4195 sector_t oldsize, size;
4197 if (mddev->reshape_position != MaxSector)
4200 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4203 oldsize = raid10_size(mddev, 0, 0);
4204 size = raid10_size(mddev, sectors, 0);
4205 if (mddev->external_size &&
4206 mddev->array_sectors > size)
4208 if (mddev->bitmap) {
4209 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4213 md_set_array_sectors(mddev, size);
4214 if (sectors > mddev->dev_sectors &&
4215 mddev->recovery_cp > oldsize) {
4216 mddev->recovery_cp = oldsize;
4217 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4219 calc_sectors(conf, sectors);
4220 mddev->dev_sectors = conf->dev_sectors;
4221 mddev->resync_max_sectors = size;
4225 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4227 struct md_rdev *rdev;
4228 struct r10conf *conf;
4230 if (mddev->degraded > 0) {
4231 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4233 return ERR_PTR(-EINVAL);
4235 sector_div(size, devs);
4237 /* Set new parameters */
4238 mddev->new_level = 10;
4239 /* new layout: far_copies = 1, near_copies = 2 */
4240 mddev->new_layout = (1<<8) + 2;
4241 mddev->new_chunk_sectors = mddev->chunk_sectors;
4242 mddev->delta_disks = mddev->raid_disks;
4243 mddev->raid_disks *= 2;
4244 /* make sure it will be not marked as dirty */
4245 mddev->recovery_cp = MaxSector;
4246 mddev->dev_sectors = size;
4248 conf = setup_conf(mddev);
4249 if (!IS_ERR(conf)) {
4250 rdev_for_each(rdev, mddev)
4251 if (rdev->raid_disk >= 0) {
4252 rdev->new_raid_disk = rdev->raid_disk * 2;
4253 rdev->sectors = size;
4260 static void *raid10_takeover(struct mddev *mddev)
4262 struct r0conf *raid0_conf;
4264 /* raid10 can take over:
4265 * raid0 - providing it has only two drives
4267 if (mddev->level == 0) {
4268 /* for raid0 takeover only one zone is supported */
4269 raid0_conf = mddev->private;
4270 if (raid0_conf->nr_strip_zones > 1) {
4271 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4273 return ERR_PTR(-EINVAL);
4275 return raid10_takeover_raid0(mddev,
4276 raid0_conf->strip_zone->zone_end,
4277 raid0_conf->strip_zone->nb_dev);
4279 return ERR_PTR(-EINVAL);
4282 static int raid10_check_reshape(struct mddev *mddev)
4284 /* Called when there is a request to change
4285 * - layout (to ->new_layout)
4286 * - chunk size (to ->new_chunk_sectors)
4287 * - raid_disks (by delta_disks)
4288 * or when trying to restart a reshape that was ongoing.
4290 * We need to validate the request and possibly allocate
4291 * space if that might be an issue later.
4293 * Currently we reject any reshape of a 'far' mode array,
4294 * allow chunk size to change if new is generally acceptable,
4295 * allow raid_disks to increase, and allow
4296 * a switch between 'near' mode and 'offset' mode.
4298 struct r10conf *conf = mddev->private;
4301 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4304 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4305 /* mustn't change number of copies */
4307 if (geo.far_copies > 1 && !geo.far_offset)
4308 /* Cannot switch to 'far' mode */
4311 if (mddev->array_sectors & geo.chunk_mask)
4312 /* not factor of array size */
4315 if (!enough(conf, -1))
4318 kfree(conf->mirrors_new);
4319 conf->mirrors_new = NULL;
4320 if (mddev->delta_disks > 0) {
4321 /* allocate new 'mirrors' list */
4323 kcalloc(mddev->raid_disks + mddev->delta_disks,
4324 sizeof(struct raid10_info),
4326 if (!conf->mirrors_new)
4333 * Need to check if array has failed when deciding whether to:
4335 * - remove non-faulty devices
4338 * This determination is simple when no reshape is happening.
4339 * However if there is a reshape, we need to carefully check
4340 * both the before and after sections.
4341 * This is because some failed devices may only affect one
4342 * of the two sections, and some non-in_sync devices may
4343 * be insync in the section most affected by failed devices.
4345 static int calc_degraded(struct r10conf *conf)
4347 int degraded, degraded2;
4351 /* 'prev' section first */
4352 for (i = 0; i < conf->prev.raid_disks; i++) {
4353 struct md_rdev *rdev = conf->mirrors[i].rdev;
4355 if (!rdev || test_bit(Faulty, &rdev->flags))
4357 else if (!test_bit(In_sync, &rdev->flags))
4358 /* When we can reduce the number of devices in
4359 * an array, this might not contribute to
4360 * 'degraded'. It does now.
4364 if (conf->geo.raid_disks == conf->prev.raid_disks)
4367 for (i = 0; i < conf->geo.raid_disks; i++) {
4368 struct md_rdev *rdev = conf->mirrors[i].rdev;
4370 if (!rdev || test_bit(Faulty, &rdev->flags))
4372 else if (!test_bit(In_sync, &rdev->flags)) {
4373 /* If reshape is increasing the number of devices,
4374 * this section has already been recovered, so
4375 * it doesn't contribute to degraded.
4378 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4382 if (degraded2 > degraded)
4387 static int raid10_start_reshape(struct mddev *mddev)
4389 /* A 'reshape' has been requested. This commits
4390 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4391 * This also checks if there are enough spares and adds them
4393 * We currently require enough spares to make the final
4394 * array non-degraded. We also require that the difference
4395 * between old and new data_offset - on each device - is
4396 * enough that we never risk over-writing.
4399 unsigned long before_length, after_length;
4400 sector_t min_offset_diff = 0;
4403 struct r10conf *conf = mddev->private;
4404 struct md_rdev *rdev;
4408 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4411 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4414 before_length = ((1 << conf->prev.chunk_shift) *
4415 conf->prev.far_copies);
4416 after_length = ((1 << conf->geo.chunk_shift) *
4417 conf->geo.far_copies);
4419 rdev_for_each(rdev, mddev) {
4420 if (!test_bit(In_sync, &rdev->flags)
4421 && !test_bit(Faulty, &rdev->flags))
4423 if (rdev->raid_disk >= 0) {
4424 long long diff = (rdev->new_data_offset
4425 - rdev->data_offset);
4426 if (!mddev->reshape_backwards)
4430 if (first || diff < min_offset_diff)
4431 min_offset_diff = diff;
4436 if (max(before_length, after_length) > min_offset_diff)
4439 if (spares < mddev->delta_disks)
4442 conf->offset_diff = min_offset_diff;
4443 spin_lock_irq(&conf->device_lock);
4444 if (conf->mirrors_new) {
4445 memcpy(conf->mirrors_new, conf->mirrors,
4446 sizeof(struct raid10_info)*conf->prev.raid_disks);
4448 kfree(conf->mirrors_old);
4449 conf->mirrors_old = conf->mirrors;
4450 conf->mirrors = conf->mirrors_new;
4451 conf->mirrors_new = NULL;
4453 setup_geo(&conf->geo, mddev, geo_start);
4455 if (mddev->reshape_backwards) {
4456 sector_t size = raid10_size(mddev, 0, 0);
4457 if (size < mddev->array_sectors) {
4458 spin_unlock_irq(&conf->device_lock);
4459 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4463 mddev->resync_max_sectors = size;
4464 conf->reshape_progress = size;
4466 conf->reshape_progress = 0;
4467 conf->reshape_safe = conf->reshape_progress;
4468 spin_unlock_irq(&conf->device_lock);
4470 if (mddev->delta_disks && mddev->bitmap) {
4471 struct mdp_superblock_1 *sb = NULL;
4472 sector_t oldsize, newsize;
4474 oldsize = raid10_size(mddev, 0, 0);
4475 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4477 if (!mddev_is_clustered(mddev)) {
4478 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4485 rdev_for_each(rdev, mddev) {
4486 if (rdev->raid_disk > -1 &&
4487 !test_bit(Faulty, &rdev->flags))
4488 sb = page_address(rdev->sb_page);
4492 * some node is already performing reshape, and no need to
4493 * call md_bitmap_resize again since it should be called when
4494 * receiving BITMAP_RESIZE msg
4496 if ((sb && (le32_to_cpu(sb->feature_map) &
4497 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4500 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4504 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4506 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4511 if (mddev->delta_disks > 0) {
4512 rdev_for_each(rdev, mddev)
4513 if (rdev->raid_disk < 0 &&
4514 !test_bit(Faulty, &rdev->flags)) {
4515 if (raid10_add_disk(mddev, rdev) == 0) {
4516 if (rdev->raid_disk >=
4517 conf->prev.raid_disks)
4518 set_bit(In_sync, &rdev->flags);
4520 rdev->recovery_offset = 0;
4522 /* Failure here is OK */
4523 sysfs_link_rdev(mddev, rdev);
4525 } else if (rdev->raid_disk >= conf->prev.raid_disks
4526 && !test_bit(Faulty, &rdev->flags)) {
4527 /* This is a spare that was manually added */
4528 set_bit(In_sync, &rdev->flags);
4531 /* When a reshape changes the number of devices,
4532 * ->degraded is measured against the larger of the
4533 * pre and post numbers.
4535 spin_lock_irq(&conf->device_lock);
4536 mddev->degraded = calc_degraded(conf);
4537 spin_unlock_irq(&conf->device_lock);
4538 mddev->raid_disks = conf->geo.raid_disks;
4539 mddev->reshape_position = conf->reshape_progress;
4540 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4542 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4543 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4544 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4545 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4546 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4547 conf->reshape_checkpoint = jiffies;
4552 mddev->recovery = 0;
4553 spin_lock_irq(&conf->device_lock);
4554 conf->geo = conf->prev;
4555 mddev->raid_disks = conf->geo.raid_disks;
4556 rdev_for_each(rdev, mddev)
4557 rdev->new_data_offset = rdev->data_offset;
4559 conf->reshape_progress = MaxSector;
4560 conf->reshape_safe = MaxSector;
4561 mddev->reshape_position = MaxSector;
4562 spin_unlock_irq(&conf->device_lock);
4566 /* Calculate the last device-address that could contain
4567 * any block from the chunk that includes the array-address 's'
4568 * and report the next address.
4569 * i.e. the address returned will be chunk-aligned and after
4570 * any data that is in the chunk containing 's'.
4572 static sector_t last_dev_address(sector_t s, struct geom *geo)
4574 s = (s | geo->chunk_mask) + 1;
4575 s >>= geo->chunk_shift;
4576 s *= geo->near_copies;
4577 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4578 s *= geo->far_copies;
4579 s <<= geo->chunk_shift;
4583 /* Calculate the first device-address that could contain
4584 * any block from the chunk that includes the array-address 's'.
4585 * This too will be the start of a chunk
4587 static sector_t first_dev_address(sector_t s, struct geom *geo)
4589 s >>= geo->chunk_shift;
4590 s *= geo->near_copies;
4591 sector_div(s, geo->raid_disks);
4592 s *= geo->far_copies;
4593 s <<= geo->chunk_shift;
4597 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4600 /* We simply copy at most one chunk (smallest of old and new)
4601 * at a time, possibly less if that exceeds RESYNC_PAGES,
4602 * or we hit a bad block or something.
4603 * This might mean we pause for normal IO in the middle of
4604 * a chunk, but that is not a problem as mddev->reshape_position
4605 * can record any location.
4607 * If we will want to write to a location that isn't
4608 * yet recorded as 'safe' (i.e. in metadata on disk) then
4609 * we need to flush all reshape requests and update the metadata.
4611 * When reshaping forwards (e.g. to more devices), we interpret
4612 * 'safe' as the earliest block which might not have been copied
4613 * down yet. We divide this by previous stripe size and multiply
4614 * by previous stripe length to get lowest device offset that we
4615 * cannot write to yet.
4616 * We interpret 'sector_nr' as an address that we want to write to.
4617 * From this we use last_device_address() to find where we might
4618 * write to, and first_device_address on the 'safe' position.
4619 * If this 'next' write position is after the 'safe' position,
4620 * we must update the metadata to increase the 'safe' position.
4622 * When reshaping backwards, we round in the opposite direction
4623 * and perform the reverse test: next write position must not be
4624 * less than current safe position.
4626 * In all this the minimum difference in data offsets
4627 * (conf->offset_diff - always positive) allows a bit of slack,
4628 * so next can be after 'safe', but not by more than offset_diff
4630 * We need to prepare all the bios here before we start any IO
4631 * to ensure the size we choose is acceptable to all devices.
4632 * The means one for each copy for write-out and an extra one for
4634 * We store the read-in bio in ->master_bio and the others in
4635 * ->devs[x].bio and ->devs[x].repl_bio.
4637 struct r10conf *conf = mddev->private;
4638 struct r10bio *r10_bio;
4639 sector_t next, safe, last;
4643 struct md_rdev *rdev;
4646 struct bio *bio, *read_bio;
4647 int sectors_done = 0;
4648 struct page **pages;
4650 if (sector_nr == 0) {
4651 /* If restarting in the middle, skip the initial sectors */
4652 if (mddev->reshape_backwards &&
4653 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4654 sector_nr = (raid10_size(mddev, 0, 0)
4655 - conf->reshape_progress);
4656 } else if (!mddev->reshape_backwards &&
4657 conf->reshape_progress > 0)
4658 sector_nr = conf->reshape_progress;
4660 mddev->curr_resync_completed = sector_nr;
4661 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4667 /* We don't use sector_nr to track where we are up to
4668 * as that doesn't work well for ->reshape_backwards.
4669 * So just use ->reshape_progress.
4671 if (mddev->reshape_backwards) {
4672 /* 'next' is the earliest device address that we might
4673 * write to for this chunk in the new layout
4675 next = first_dev_address(conf->reshape_progress - 1,
4678 /* 'safe' is the last device address that we might read from
4679 * in the old layout after a restart
4681 safe = last_dev_address(conf->reshape_safe - 1,
4684 if (next + conf->offset_diff < safe)
4687 last = conf->reshape_progress - 1;
4688 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4689 & conf->prev.chunk_mask);
4690 if (sector_nr + RESYNC_SECTORS < last)
4691 sector_nr = last + 1 - RESYNC_SECTORS;
4693 /* 'next' is after the last device address that we
4694 * might write to for this chunk in the new layout
4696 next = last_dev_address(conf->reshape_progress, &conf->geo);
4698 /* 'safe' is the earliest device address that we might
4699 * read from in the old layout after a restart
4701 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4703 /* Need to update metadata if 'next' might be beyond 'safe'
4704 * as that would possibly corrupt data
4706 if (next > safe + conf->offset_diff)
4709 sector_nr = conf->reshape_progress;
4710 last = sector_nr | (conf->geo.chunk_mask
4711 & conf->prev.chunk_mask);
4713 if (sector_nr + RESYNC_SECTORS <= last)
4714 last = sector_nr + RESYNC_SECTORS - 1;
4718 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4719 /* Need to update reshape_position in metadata */
4720 wait_barrier(conf, false);
4721 mddev->reshape_position = conf->reshape_progress;
4722 if (mddev->reshape_backwards)
4723 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4724 - conf->reshape_progress;
4726 mddev->curr_resync_completed = conf->reshape_progress;
4727 conf->reshape_checkpoint = jiffies;
4728 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4729 md_wakeup_thread(mddev->thread);
4730 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4731 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4732 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4733 allow_barrier(conf);
4734 return sectors_done;
4736 conf->reshape_safe = mddev->reshape_position;
4737 allow_barrier(conf);
4740 raise_barrier(conf, 0);
4742 /* Now schedule reads for blocks from sector_nr to last */
4743 r10_bio = raid10_alloc_init_r10buf(conf);
4745 raise_barrier(conf, 1);
4746 atomic_set(&r10_bio->remaining, 0);
4747 r10_bio->mddev = mddev;
4748 r10_bio->sector = sector_nr;
4749 set_bit(R10BIO_IsReshape, &r10_bio->state);
4750 r10_bio->sectors = last - sector_nr + 1;
4751 rdev = read_balance(conf, r10_bio, &max_sectors);
4752 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4755 /* Cannot read from here, so need to record bad blocks
4756 * on all the target devices.
4759 mempool_free(r10_bio, &conf->r10buf_pool);
4760 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4761 return sectors_done;
4764 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4765 GFP_KERNEL, &mddev->bio_set);
4766 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4767 + rdev->data_offset);
4768 read_bio->bi_private = r10_bio;
4769 read_bio->bi_end_io = end_reshape_read;
4770 r10_bio->master_bio = read_bio;
4771 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4774 * Broadcast RESYNC message to other nodes, so all nodes would not
4775 * write to the region to avoid conflict.
4777 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4778 struct mdp_superblock_1 *sb = NULL;
4779 int sb_reshape_pos = 0;
4781 conf->cluster_sync_low = sector_nr;
4782 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4783 sb = page_address(rdev->sb_page);
4785 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4787 * Set cluster_sync_low again if next address for array
4788 * reshape is less than cluster_sync_low. Since we can't
4789 * update cluster_sync_low until it has finished reshape.
4791 if (sb_reshape_pos < conf->cluster_sync_low)
4792 conf->cluster_sync_low = sb_reshape_pos;
4795 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4796 conf->cluster_sync_high);
4799 /* Now find the locations in the new layout */
4800 __raid10_find_phys(&conf->geo, r10_bio);
4803 read_bio->bi_next = NULL;
4805 for (s = 0; s < conf->copies*2; s++) {
4807 int d = r10_bio->devs[s/2].devnum;
4808 struct md_rdev *rdev2;
4810 rdev2 = conf->mirrors[d].replacement;
4811 b = r10_bio->devs[s/2].repl_bio;
4813 rdev2 = conf->mirrors[d].rdev;
4814 b = r10_bio->devs[s/2].bio;
4816 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4819 bio_set_dev(b, rdev2->bdev);
4820 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4821 rdev2->new_data_offset;
4822 b->bi_end_io = end_reshape_write;
4823 b->bi_opf = REQ_OP_WRITE;
4828 /* Now add as many pages as possible to all of these bios. */
4831 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4832 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4833 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4834 int len = (max_sectors - s) << 9;
4835 if (len > PAGE_SIZE)
4837 for (bio = blist; bio ; bio = bio->bi_next) {
4838 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
4839 bio->bi_status = BLK_STS_RESOURCE;
4841 return sectors_done;
4844 sector_nr += len >> 9;
4845 nr_sectors += len >> 9;
4847 r10_bio->sectors = nr_sectors;
4849 /* Now submit the read */
4850 md_sync_acct_bio(read_bio, r10_bio->sectors);
4851 atomic_inc(&r10_bio->remaining);
4852 read_bio->bi_next = NULL;
4853 submit_bio_noacct(read_bio);
4854 sectors_done += nr_sectors;
4855 if (sector_nr <= last)
4858 lower_barrier(conf);
4860 /* Now that we have done the whole section we can
4861 * update reshape_progress
4863 if (mddev->reshape_backwards)
4864 conf->reshape_progress -= sectors_done;
4866 conf->reshape_progress += sectors_done;
4868 return sectors_done;
4871 static void end_reshape_request(struct r10bio *r10_bio);
4872 static int handle_reshape_read_error(struct mddev *mddev,
4873 struct r10bio *r10_bio);
4874 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4876 /* Reshape read completed. Hopefully we have a block
4878 * If we got a read error then we do sync 1-page reads from
4879 * elsewhere until we find the data - or give up.
4881 struct r10conf *conf = mddev->private;
4884 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4885 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4886 /* Reshape has been aborted */
4887 md_done_sync(mddev, r10_bio->sectors, 0);
4891 /* We definitely have the data in the pages, schedule the
4894 atomic_set(&r10_bio->remaining, 1);
4895 for (s = 0; s < conf->copies*2; s++) {
4897 int d = r10_bio->devs[s/2].devnum;
4898 struct md_rdev *rdev;
4900 rdev = conf->mirrors[d].replacement;
4901 b = r10_bio->devs[s/2].repl_bio;
4903 rdev = conf->mirrors[d].rdev;
4904 b = r10_bio->devs[s/2].bio;
4906 if (!rdev || test_bit(Faulty, &rdev->flags))
4909 atomic_inc(&rdev->nr_pending);
4910 md_sync_acct_bio(b, r10_bio->sectors);
4911 atomic_inc(&r10_bio->remaining);
4913 submit_bio_noacct(b);
4915 end_reshape_request(r10_bio);
4918 static void end_reshape(struct r10conf *conf)
4920 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4923 spin_lock_irq(&conf->device_lock);
4924 conf->prev = conf->geo;
4925 md_finish_reshape(conf->mddev);
4927 conf->reshape_progress = MaxSector;
4928 conf->reshape_safe = MaxSector;
4929 spin_unlock_irq(&conf->device_lock);
4931 mddev_update_io_opt(conf->mddev, raid10_nr_stripes(conf));
4935 static void raid10_update_reshape_pos(struct mddev *mddev)
4937 struct r10conf *conf = mddev->private;
4940 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4941 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4942 || mddev->reshape_position == MaxSector)
4943 conf->reshape_progress = mddev->reshape_position;
4948 static int handle_reshape_read_error(struct mddev *mddev,
4949 struct r10bio *r10_bio)
4951 /* Use sync reads to get the blocks from somewhere else */
4952 int sectors = r10_bio->sectors;
4953 struct r10conf *conf = mddev->private;
4954 struct r10bio *r10b;
4957 struct page **pages;
4959 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4961 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4965 /* reshape IOs share pages from .devs[0].bio */
4966 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4968 r10b->sector = r10_bio->sector;
4969 __raid10_find_phys(&conf->prev, r10b);
4974 int first_slot = slot;
4976 if (s > (PAGE_SIZE >> 9))
4980 int d = r10b->devs[slot].devnum;
4981 struct md_rdev *rdev = conf->mirrors[d].rdev;
4984 test_bit(Faulty, &rdev->flags) ||
4985 !test_bit(In_sync, &rdev->flags))
4988 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4989 atomic_inc(&rdev->nr_pending);
4990 success = sync_page_io(rdev,
4994 REQ_OP_READ, false);
4995 rdev_dec_pending(rdev, mddev);
5000 if (slot >= conf->copies)
5002 if (slot == first_slot)
5006 /* couldn't read this block, must give up */
5007 set_bit(MD_RECOVERY_INTR,
5019 static void end_reshape_write(struct bio *bio)
5021 struct r10bio *r10_bio = get_resync_r10bio(bio);
5022 struct mddev *mddev = r10_bio->mddev;
5023 struct r10conf *conf = mddev->private;
5027 struct md_rdev *rdev = NULL;
5029 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5030 rdev = repl ? conf->mirrors[d].replacement :
5031 conf->mirrors[d].rdev;
5033 if (bio->bi_status) {
5034 /* FIXME should record badblock */
5035 md_error(mddev, rdev);
5038 rdev_dec_pending(rdev, mddev);
5039 end_reshape_request(r10_bio);
5042 static void end_reshape_request(struct r10bio *r10_bio)
5044 if (!atomic_dec_and_test(&r10_bio->remaining))
5046 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5047 bio_put(r10_bio->master_bio);
5051 static void raid10_finish_reshape(struct mddev *mddev)
5053 struct r10conf *conf = mddev->private;
5055 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5058 if (mddev->delta_disks > 0) {
5059 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5060 mddev->recovery_cp = mddev->resync_max_sectors;
5061 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5063 mddev->resync_max_sectors = mddev->array_sectors;
5066 for (d = conf->geo.raid_disks ;
5067 d < conf->geo.raid_disks - mddev->delta_disks;
5069 struct md_rdev *rdev = conf->mirrors[d].rdev;
5071 clear_bit(In_sync, &rdev->flags);
5072 rdev = conf->mirrors[d].replacement;
5074 clear_bit(In_sync, &rdev->flags);
5077 mddev->layout = mddev->new_layout;
5078 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5079 mddev->reshape_position = MaxSector;
5080 mddev->delta_disks = 0;
5081 mddev->reshape_backwards = 0;
5084 static struct md_personality raid10_personality =
5088 .owner = THIS_MODULE,
5089 .make_request = raid10_make_request,
5091 .free = raid10_free,
5092 .status = raid10_status,
5093 .error_handler = raid10_error,
5094 .hot_add_disk = raid10_add_disk,
5095 .hot_remove_disk= raid10_remove_disk,
5096 .spare_active = raid10_spare_active,
5097 .sync_request = raid10_sync_request,
5098 .quiesce = raid10_quiesce,
5099 .size = raid10_size,
5100 .resize = raid10_resize,
5101 .takeover = raid10_takeover,
5102 .check_reshape = raid10_check_reshape,
5103 .start_reshape = raid10_start_reshape,
5104 .finish_reshape = raid10_finish_reshape,
5105 .update_reshape_pos = raid10_update_reshape_pos,
5108 static int __init raid_init(void)
5110 return register_md_personality(&raid10_personality);
5113 static void raid_exit(void)
5115 unregister_md_personality(&raid10_personality);
5118 module_init(raid_init);
5119 module_exit(raid_exit);
5120 MODULE_LICENSE("GPL");
5121 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5122 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5123 MODULE_ALIAS("md-raid10");
5124 MODULE_ALIAS("md-level-10");