2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include <trace/events/block.h>
32 #include "md-bitmap.h"
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
76 * Number of guaranteed r10bios in case of extreme VM load:
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
98 static int max_queued_requests = 1024;
100 static void allow_barrier(struct r10conf *conf);
101 static void lower_barrier(struct r10conf *conf);
102 static int _enough(struct r10conf *conf, int previous, int ignore);
103 static int enough(struct r10conf *conf, int ignore);
104 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
106 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107 static void end_reshape_write(struct bio *bio);
108 static void end_reshape(struct r10conf *conf);
110 #define raid10_log(md, fmt, args...) \
111 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
113 #include "raid1-10.c"
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
119 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
121 return get_resync_pages(bio)->raid_bio;
124 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
126 struct r10conf *conf = data;
127 int size = offsetof(struct r10bio, devs[conf->copies]);
129 /* allocate a r10bio with room for raid_disks entries in the
131 return kzalloc(size, gfp_flags);
134 static void r10bio_pool_free(void *r10_bio, void *data)
139 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140 /* amount of memory to reserve for resync requests */
141 #define RESYNC_WINDOW (1024*1024)
142 /* maximum number of concurrent requests, memory permitting */
143 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
148 * When performing a resync, we need to read and compare, so
149 * we need as many pages are there are copies.
150 * When performing a recovery, we need 2 bios, one for read,
151 * one for write (we recover only one drive per r10buf)
154 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
156 struct r10conf *conf = data;
157 struct r10bio *r10_bio;
160 int nalloc, nalloc_rp;
161 struct resync_pages *rps;
163 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
167 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
168 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
169 nalloc = conf->copies; /* resync */
171 nalloc = 2; /* recovery */
173 /* allocate once for all bios */
174 if (!conf->have_replacement)
177 nalloc_rp = nalloc * 2;
178 rps = kmalloc(sizeof(struct resync_pages) * nalloc_rp, gfp_flags);
180 goto out_free_r10bio;
185 for (j = nalloc ; j-- ; ) {
186 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
189 r10_bio->devs[j].bio = bio;
190 if (!conf->have_replacement)
192 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
195 r10_bio->devs[j].repl_bio = bio;
198 * Allocate RESYNC_PAGES data pages and attach them
201 for (j = 0; j < nalloc; j++) {
202 struct bio *rbio = r10_bio->devs[j].repl_bio;
203 struct resync_pages *rp, *rp_repl;
207 rp_repl = &rps[nalloc + j];
209 bio = r10_bio->devs[j].bio;
211 if (!j || test_bit(MD_RECOVERY_SYNC,
212 &conf->mddev->recovery)) {
213 if (resync_alloc_pages(rp, gfp_flags))
216 memcpy(rp, &rps[0], sizeof(*rp));
217 resync_get_all_pages(rp);
220 rp->raid_bio = r10_bio;
221 bio->bi_private = rp;
223 memcpy(rp_repl, rp, sizeof(*rp));
224 rbio->bi_private = rp_repl;
232 resync_free_pages(&rps[j * 2]);
236 for ( ; j < nalloc; j++) {
237 if (r10_bio->devs[j].bio)
238 bio_put(r10_bio->devs[j].bio);
239 if (r10_bio->devs[j].repl_bio)
240 bio_put(r10_bio->devs[j].repl_bio);
244 r10bio_pool_free(r10_bio, conf);
248 static void r10buf_pool_free(void *__r10_bio, void *data)
250 struct r10conf *conf = data;
251 struct r10bio *r10bio = __r10_bio;
253 struct resync_pages *rp = NULL;
255 for (j = conf->copies; j--; ) {
256 struct bio *bio = r10bio->devs[j].bio;
258 rp = get_resync_pages(bio);
259 resync_free_pages(rp);
262 bio = r10bio->devs[j].repl_bio;
267 /* resync pages array stored in the 1st bio's .bi_private */
270 r10bio_pool_free(r10bio, conf);
273 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
277 for (i = 0; i < conf->copies; i++) {
278 struct bio **bio = & r10_bio->devs[i].bio;
279 if (!BIO_SPECIAL(*bio))
282 bio = &r10_bio->devs[i].repl_bio;
283 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
289 static void free_r10bio(struct r10bio *r10_bio)
291 struct r10conf *conf = r10_bio->mddev->private;
293 put_all_bios(conf, r10_bio);
294 mempool_free(r10_bio, &conf->r10bio_pool);
297 static void put_buf(struct r10bio *r10_bio)
299 struct r10conf *conf = r10_bio->mddev->private;
301 mempool_free(r10_bio, &conf->r10buf_pool);
306 static void reschedule_retry(struct r10bio *r10_bio)
309 struct mddev *mddev = r10_bio->mddev;
310 struct r10conf *conf = mddev->private;
312 spin_lock_irqsave(&conf->device_lock, flags);
313 list_add(&r10_bio->retry_list, &conf->retry_list);
315 spin_unlock_irqrestore(&conf->device_lock, flags);
317 /* wake up frozen array... */
318 wake_up(&conf->wait_barrier);
320 md_wakeup_thread(mddev->thread);
324 * raid_end_bio_io() is called when we have finished servicing a mirrored
325 * operation and are ready to return a success/failure code to the buffer
328 static void raid_end_bio_io(struct r10bio *r10_bio)
330 struct bio *bio = r10_bio->master_bio;
331 struct r10conf *conf = r10_bio->mddev->private;
333 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
334 bio->bi_status = BLK_STS_IOERR;
338 * Wake up any possible resync thread that waits for the device
343 free_r10bio(r10_bio);
347 * Update disk head position estimator based on IRQ completion info.
349 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
351 struct r10conf *conf = r10_bio->mddev->private;
353 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
354 r10_bio->devs[slot].addr + (r10_bio->sectors);
358 * Find the disk number which triggered given bio
360 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
361 struct bio *bio, int *slotp, int *replp)
366 for (slot = 0; slot < conf->copies; slot++) {
367 if (r10_bio->devs[slot].bio == bio)
369 if (r10_bio->devs[slot].repl_bio == bio) {
375 BUG_ON(slot == conf->copies);
376 update_head_pos(slot, r10_bio);
382 return r10_bio->devs[slot].devnum;
385 static void raid10_end_read_request(struct bio *bio)
387 int uptodate = !bio->bi_status;
388 struct r10bio *r10_bio = bio->bi_private;
390 struct md_rdev *rdev;
391 struct r10conf *conf = r10_bio->mddev->private;
393 slot = r10_bio->read_slot;
394 rdev = r10_bio->devs[slot].rdev;
396 * this branch is our 'one mirror IO has finished' event handler:
398 update_head_pos(slot, r10_bio);
402 * Set R10BIO_Uptodate in our master bio, so that
403 * we will return a good error code to the higher
404 * levels even if IO on some other mirrored buffer fails.
406 * The 'master' represents the composite IO operation to
407 * user-side. So if something waits for IO, then it will
408 * wait for the 'master' bio.
410 set_bit(R10BIO_Uptodate, &r10_bio->state);
412 /* If all other devices that store this block have
413 * failed, we want to return the error upwards rather
414 * than fail the last device. Here we redefine
415 * "uptodate" to mean "Don't want to retry"
417 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
422 raid_end_bio_io(r10_bio);
423 rdev_dec_pending(rdev, conf->mddev);
426 * oops, read error - keep the refcount on the rdev
428 char b[BDEVNAME_SIZE];
429 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
431 bdevname(rdev->bdev, b),
432 (unsigned long long)r10_bio->sector);
433 set_bit(R10BIO_ReadError, &r10_bio->state);
434 reschedule_retry(r10_bio);
438 static void close_write(struct r10bio *r10_bio)
440 /* clear the bitmap if all writes complete successfully */
441 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
443 !test_bit(R10BIO_Degraded, &r10_bio->state),
445 md_write_end(r10_bio->mddev);
448 static void one_write_done(struct r10bio *r10_bio)
450 if (atomic_dec_and_test(&r10_bio->remaining)) {
451 if (test_bit(R10BIO_WriteError, &r10_bio->state))
452 reschedule_retry(r10_bio);
454 close_write(r10_bio);
455 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
456 reschedule_retry(r10_bio);
458 raid_end_bio_io(r10_bio);
463 static void raid10_end_write_request(struct bio *bio)
465 struct r10bio *r10_bio = bio->bi_private;
468 struct r10conf *conf = r10_bio->mddev->private;
470 struct md_rdev *rdev = NULL;
471 struct bio *to_put = NULL;
474 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
476 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
479 rdev = conf->mirrors[dev].replacement;
483 rdev = conf->mirrors[dev].rdev;
486 * this branch is our 'one mirror IO has finished' event handler:
488 if (bio->bi_status && !discard_error) {
490 /* Never record new bad blocks to replacement,
493 md_error(rdev->mddev, rdev);
495 set_bit(WriteErrorSeen, &rdev->flags);
496 if (!test_and_set_bit(WantReplacement, &rdev->flags))
497 set_bit(MD_RECOVERY_NEEDED,
498 &rdev->mddev->recovery);
501 if (test_bit(FailFast, &rdev->flags) &&
502 (bio->bi_opf & MD_FAILFAST)) {
503 md_error(rdev->mddev, rdev);
504 if (!test_bit(Faulty, &rdev->flags))
505 /* This is the only remaining device,
506 * We need to retry the write without
509 set_bit(R10BIO_WriteError, &r10_bio->state);
511 r10_bio->devs[slot].bio = NULL;
516 set_bit(R10BIO_WriteError, &r10_bio->state);
520 * Set R10BIO_Uptodate in our master bio, so that
521 * we will return a good error code for to the higher
522 * levels even if IO on some other mirrored buffer fails.
524 * The 'master' represents the composite IO operation to
525 * user-side. So if something waits for IO, then it will
526 * wait for the 'master' bio.
532 * Do not set R10BIO_Uptodate if the current device is
533 * rebuilding or Faulty. This is because we cannot use
534 * such device for properly reading the data back (we could
535 * potentially use it, if the current write would have felt
536 * before rdev->recovery_offset, but for simplicity we don't
539 if (test_bit(In_sync, &rdev->flags) &&
540 !test_bit(Faulty, &rdev->flags))
541 set_bit(R10BIO_Uptodate, &r10_bio->state);
543 /* Maybe we can clear some bad blocks. */
544 if (is_badblock(rdev,
545 r10_bio->devs[slot].addr,
547 &first_bad, &bad_sectors) && !discard_error) {
550 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
552 r10_bio->devs[slot].bio = IO_MADE_GOOD;
554 set_bit(R10BIO_MadeGood, &r10_bio->state);
560 * Let's see if all mirrored write operations have finished
563 one_write_done(r10_bio);
565 rdev_dec_pending(rdev, conf->mddev);
571 * RAID10 layout manager
572 * As well as the chunksize and raid_disks count, there are two
573 * parameters: near_copies and far_copies.
574 * near_copies * far_copies must be <= raid_disks.
575 * Normally one of these will be 1.
576 * If both are 1, we get raid0.
577 * If near_copies == raid_disks, we get raid1.
579 * Chunks are laid out in raid0 style with near_copies copies of the
580 * first chunk, followed by near_copies copies of the next chunk and
582 * If far_copies > 1, then after 1/far_copies of the array has been assigned
583 * as described above, we start again with a device offset of near_copies.
584 * So we effectively have another copy of the whole array further down all
585 * the drives, but with blocks on different drives.
586 * With this layout, and block is never stored twice on the one device.
588 * raid10_find_phys finds the sector offset of a given virtual sector
589 * on each device that it is on.
591 * raid10_find_virt does the reverse mapping, from a device and a
592 * sector offset to a virtual address
595 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
603 int last_far_set_start, last_far_set_size;
605 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
606 last_far_set_start *= geo->far_set_size;
608 last_far_set_size = geo->far_set_size;
609 last_far_set_size += (geo->raid_disks % geo->far_set_size);
611 /* now calculate first sector/dev */
612 chunk = r10bio->sector >> geo->chunk_shift;
613 sector = r10bio->sector & geo->chunk_mask;
615 chunk *= geo->near_copies;
617 dev = sector_div(stripe, geo->raid_disks);
619 stripe *= geo->far_copies;
621 sector += stripe << geo->chunk_shift;
623 /* and calculate all the others */
624 for (n = 0; n < geo->near_copies; n++) {
628 r10bio->devs[slot].devnum = d;
629 r10bio->devs[slot].addr = s;
632 for (f = 1; f < geo->far_copies; f++) {
633 set = d / geo->far_set_size;
634 d += geo->near_copies;
636 if ((geo->raid_disks % geo->far_set_size) &&
637 (d > last_far_set_start)) {
638 d -= last_far_set_start;
639 d %= last_far_set_size;
640 d += last_far_set_start;
642 d %= geo->far_set_size;
643 d += geo->far_set_size * set;
646 r10bio->devs[slot].devnum = d;
647 r10bio->devs[slot].addr = s;
651 if (dev >= geo->raid_disks) {
653 sector += (geo->chunk_mask + 1);
658 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
660 struct geom *geo = &conf->geo;
662 if (conf->reshape_progress != MaxSector &&
663 ((r10bio->sector >= conf->reshape_progress) !=
664 conf->mddev->reshape_backwards)) {
665 set_bit(R10BIO_Previous, &r10bio->state);
668 clear_bit(R10BIO_Previous, &r10bio->state);
670 __raid10_find_phys(geo, r10bio);
673 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
675 sector_t offset, chunk, vchunk;
676 /* Never use conf->prev as this is only called during resync
677 * or recovery, so reshape isn't happening
679 struct geom *geo = &conf->geo;
680 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
681 int far_set_size = geo->far_set_size;
682 int last_far_set_start;
684 if (geo->raid_disks % geo->far_set_size) {
685 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
686 last_far_set_start *= geo->far_set_size;
688 if (dev >= last_far_set_start) {
689 far_set_size = geo->far_set_size;
690 far_set_size += (geo->raid_disks % geo->far_set_size);
691 far_set_start = last_far_set_start;
695 offset = sector & geo->chunk_mask;
696 if (geo->far_offset) {
698 chunk = sector >> geo->chunk_shift;
699 fc = sector_div(chunk, geo->far_copies);
700 dev -= fc * geo->near_copies;
701 if (dev < far_set_start)
704 while (sector >= geo->stride) {
705 sector -= geo->stride;
706 if (dev < (geo->near_copies + far_set_start))
707 dev += far_set_size - geo->near_copies;
709 dev -= geo->near_copies;
711 chunk = sector >> geo->chunk_shift;
713 vchunk = chunk * geo->raid_disks + dev;
714 sector_div(vchunk, geo->near_copies);
715 return (vchunk << geo->chunk_shift) + offset;
719 * This routine returns the disk from which the requested read should
720 * be done. There is a per-array 'next expected sequential IO' sector
721 * number - if this matches on the next IO then we use the last disk.
722 * There is also a per-disk 'last know head position' sector that is
723 * maintained from IRQ contexts, both the normal and the resync IO
724 * completion handlers update this position correctly. If there is no
725 * perfect sequential match then we pick the disk whose head is closest.
727 * If there are 2 mirrors in the same 2 devices, performance degrades
728 * because position is mirror, not device based.
730 * The rdev for the device selected will have nr_pending incremented.
734 * FIXME: possibly should rethink readbalancing and do it differently
735 * depending on near_copies / far_copies geometry.
737 static struct md_rdev *read_balance(struct r10conf *conf,
738 struct r10bio *r10_bio,
741 const sector_t this_sector = r10_bio->sector;
743 int sectors = r10_bio->sectors;
744 int best_good_sectors;
745 sector_t new_distance, best_dist;
746 struct md_rdev *best_rdev, *rdev = NULL;
749 struct geom *geo = &conf->geo;
751 raid10_find_phys(conf, r10_bio);
755 best_dist = MaxSector;
756 best_good_sectors = 0;
758 clear_bit(R10BIO_FailFast, &r10_bio->state);
760 * Check if we can balance. We can balance on the whole
761 * device if no resync is going on (recovery is ok), or below
762 * the resync window. We take the first readable disk when
763 * above the resync window.
765 if ((conf->mddev->recovery_cp < MaxSector
766 && (this_sector + sectors >= conf->next_resync)) ||
767 (mddev_is_clustered(conf->mddev) &&
768 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
769 this_sector + sectors)))
772 for (slot = 0; slot < conf->copies ; slot++) {
777 if (r10_bio->devs[slot].bio == IO_BLOCKED)
779 disk = r10_bio->devs[slot].devnum;
780 rdev = rcu_dereference(conf->mirrors[disk].replacement);
781 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
782 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
783 rdev = rcu_dereference(conf->mirrors[disk].rdev);
785 test_bit(Faulty, &rdev->flags))
787 if (!test_bit(In_sync, &rdev->flags) &&
788 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
791 dev_sector = r10_bio->devs[slot].addr;
792 if (is_badblock(rdev, dev_sector, sectors,
793 &first_bad, &bad_sectors)) {
794 if (best_dist < MaxSector)
795 /* Already have a better slot */
797 if (first_bad <= dev_sector) {
798 /* Cannot read here. If this is the
799 * 'primary' device, then we must not read
800 * beyond 'bad_sectors' from another device.
802 bad_sectors -= (dev_sector - first_bad);
803 if (!do_balance && sectors > bad_sectors)
804 sectors = bad_sectors;
805 if (best_good_sectors > sectors)
806 best_good_sectors = sectors;
808 sector_t good_sectors =
809 first_bad - dev_sector;
810 if (good_sectors > best_good_sectors) {
811 best_good_sectors = good_sectors;
816 /* Must read from here */
821 best_good_sectors = sectors;
827 /* At least 2 disks to choose from so failfast is OK */
828 set_bit(R10BIO_FailFast, &r10_bio->state);
829 /* This optimisation is debatable, and completely destroys
830 * sequential read speed for 'far copies' arrays. So only
831 * keep it for 'near' arrays, and review those later.
833 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
836 /* for far > 1 always use the lowest address */
837 else if (geo->far_copies > 1)
838 new_distance = r10_bio->devs[slot].addr;
840 new_distance = abs(r10_bio->devs[slot].addr -
841 conf->mirrors[disk].head_position);
842 if (new_distance < best_dist) {
843 best_dist = new_distance;
848 if (slot >= conf->copies) {
854 atomic_inc(&rdev->nr_pending);
855 r10_bio->read_slot = slot;
859 *max_sectors = best_good_sectors;
864 static int raid10_congested(struct mddev *mddev, int bits)
866 struct r10conf *conf = mddev->private;
869 if ((bits & (1 << WB_async_congested)) &&
870 conf->pending_count >= max_queued_requests)
875 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
878 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
879 if (rdev && !test_bit(Faulty, &rdev->flags)) {
880 struct request_queue *q = bdev_get_queue(rdev->bdev);
882 ret |= bdi_congested(q->backing_dev_info, bits);
889 static void flush_pending_writes(struct r10conf *conf)
891 /* Any writes that have been queued but are awaiting
892 * bitmap updates get flushed here.
894 spin_lock_irq(&conf->device_lock);
896 if (conf->pending_bio_list.head) {
897 struct blk_plug plug;
900 bio = bio_list_get(&conf->pending_bio_list);
901 conf->pending_count = 0;
902 spin_unlock_irq(&conf->device_lock);
905 * As this is called in a wait_event() loop (see freeze_array),
906 * current->state might be TASK_UNINTERRUPTIBLE which will
907 * cause a warning when we prepare to wait again. As it is
908 * rare that this path is taken, it is perfectly safe to force
909 * us to go around the wait_event() loop again, so the warning
910 * is a false-positive. Silence the warning by resetting
913 __set_current_state(TASK_RUNNING);
915 blk_start_plug(&plug);
916 /* flush any pending bitmap writes to disk
917 * before proceeding w/ I/O */
918 bitmap_unplug(conf->mddev->bitmap);
919 wake_up(&conf->wait_barrier);
921 while (bio) { /* submit pending writes */
922 struct bio *next = bio->bi_next;
923 struct md_rdev *rdev = (void*)bio->bi_disk;
925 bio_set_dev(bio, rdev->bdev);
926 if (test_bit(Faulty, &rdev->flags)) {
928 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
929 !blk_queue_discard(bio->bi_disk->queue)))
933 generic_make_request(bio);
936 blk_finish_plug(&plug);
938 spin_unlock_irq(&conf->device_lock);
942 * Sometimes we need to suspend IO while we do something else,
943 * either some resync/recovery, or reconfigure the array.
944 * To do this we raise a 'barrier'.
945 * The 'barrier' is a counter that can be raised multiple times
946 * to count how many activities are happening which preclude
948 * We can only raise the barrier if there is no pending IO.
949 * i.e. if nr_pending == 0.
950 * We choose only to raise the barrier if no-one is waiting for the
951 * barrier to go down. This means that as soon as an IO request
952 * is ready, no other operations which require a barrier will start
953 * until the IO request has had a chance.
955 * So: regular IO calls 'wait_barrier'. When that returns there
956 * is no backgroup IO happening, It must arrange to call
957 * allow_barrier when it has finished its IO.
958 * backgroup IO calls must call raise_barrier. Once that returns
959 * there is no normal IO happeing. It must arrange to call
960 * lower_barrier when the particular background IO completes.
963 static void raise_barrier(struct r10conf *conf, int force)
965 BUG_ON(force && !conf->barrier);
966 spin_lock_irq(&conf->resync_lock);
968 /* Wait until no block IO is waiting (unless 'force') */
969 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
972 /* block any new IO from starting */
975 /* Now wait for all pending IO to complete */
976 wait_event_lock_irq(conf->wait_barrier,
977 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
980 spin_unlock_irq(&conf->resync_lock);
983 static void lower_barrier(struct r10conf *conf)
986 spin_lock_irqsave(&conf->resync_lock, flags);
988 spin_unlock_irqrestore(&conf->resync_lock, flags);
989 wake_up(&conf->wait_barrier);
992 static void wait_barrier(struct r10conf *conf)
994 spin_lock_irq(&conf->resync_lock);
997 /* Wait for the barrier to drop.
998 * However if there are already pending
999 * requests (preventing the barrier from
1000 * rising completely), and the
1001 * pre-process bio queue isn't empty,
1002 * then don't wait, as we need to empty
1003 * that queue to get the nr_pending
1006 raid10_log(conf->mddev, "wait barrier");
1007 wait_event_lock_irq(conf->wait_barrier,
1009 (atomic_read(&conf->nr_pending) &&
1010 current->bio_list &&
1011 (!bio_list_empty(¤t->bio_list[0]) ||
1012 !bio_list_empty(¤t->bio_list[1]))),
1015 if (!conf->nr_waiting)
1016 wake_up(&conf->wait_barrier);
1018 atomic_inc(&conf->nr_pending);
1019 spin_unlock_irq(&conf->resync_lock);
1022 static void allow_barrier(struct r10conf *conf)
1024 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1025 (conf->array_freeze_pending))
1026 wake_up(&conf->wait_barrier);
1029 static void freeze_array(struct r10conf *conf, int extra)
1031 /* stop syncio and normal IO and wait for everything to
1033 * We increment barrier and nr_waiting, and then
1034 * wait until nr_pending match nr_queued+extra
1035 * This is called in the context of one normal IO request
1036 * that has failed. Thus any sync request that might be pending
1037 * will be blocked by nr_pending, and we need to wait for
1038 * pending IO requests to complete or be queued for re-try.
1039 * Thus the number queued (nr_queued) plus this request (extra)
1040 * must match the number of pending IOs (nr_pending) before
1043 spin_lock_irq(&conf->resync_lock);
1044 conf->array_freeze_pending++;
1047 wait_event_lock_irq_cmd(conf->wait_barrier,
1048 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1050 flush_pending_writes(conf));
1052 conf->array_freeze_pending--;
1053 spin_unlock_irq(&conf->resync_lock);
1056 static void unfreeze_array(struct r10conf *conf)
1058 /* reverse the effect of the freeze */
1059 spin_lock_irq(&conf->resync_lock);
1062 wake_up(&conf->wait_barrier);
1063 spin_unlock_irq(&conf->resync_lock);
1066 static sector_t choose_data_offset(struct r10bio *r10_bio,
1067 struct md_rdev *rdev)
1069 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1070 test_bit(R10BIO_Previous, &r10_bio->state))
1071 return rdev->data_offset;
1073 return rdev->new_data_offset;
1076 struct raid10_plug_cb {
1077 struct blk_plug_cb cb;
1078 struct bio_list pending;
1082 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1084 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1086 struct mddev *mddev = plug->cb.data;
1087 struct r10conf *conf = mddev->private;
1090 if (from_schedule || current->bio_list) {
1091 spin_lock_irq(&conf->device_lock);
1092 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1093 conf->pending_count += plug->pending_cnt;
1094 spin_unlock_irq(&conf->device_lock);
1095 wake_up(&conf->wait_barrier);
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 bitmap_unplug(mddev->bitmap);
1104 wake_up(&conf->wait_barrier);
1106 while (bio) { /* submit pending writes */
1107 struct bio *next = bio->bi_next;
1108 struct md_rdev *rdev = (void*)bio->bi_disk;
1109 bio->bi_next = NULL;
1110 bio_set_dev(bio, rdev->bdev);
1111 if (test_bit(Faulty, &rdev->flags)) {
1113 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1114 !blk_queue_discard(bio->bi_disk->queue)))
1115 /* Just ignore it */
1118 generic_make_request(bio);
1124 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1125 struct r10bio *r10_bio)
1127 struct r10conf *conf = mddev->private;
1128 struct bio *read_bio;
1129 const int op = bio_op(bio);
1130 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1133 struct md_rdev *rdev;
1134 char b[BDEVNAME_SIZE];
1135 int slot = r10_bio->read_slot;
1136 struct md_rdev *err_rdev = NULL;
1137 gfp_t gfp = GFP_NOIO;
1139 if (r10_bio->devs[slot].rdev) {
1141 * This is an error retry, but we cannot
1142 * safely dereference the rdev in the r10_bio,
1143 * we must use the one in conf.
1144 * If it has already been disconnected (unlikely)
1145 * we lose the device name in error messages.
1149 * As we are blocking raid10, it is a little safer to
1152 gfp = GFP_NOIO | __GFP_HIGH;
1155 disk = r10_bio->devs[slot].devnum;
1156 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1158 bdevname(err_rdev->bdev, b);
1161 /* This never gets dereferenced */
1162 err_rdev = r10_bio->devs[slot].rdev;
1167 * Register the new request and wait if the reconstruction
1168 * thread has put up a bar for new requests.
1169 * Continue immediately if no resync is active currently.
1173 sectors = r10_bio->sectors;
1174 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1175 bio->bi_iter.bi_sector < conf->reshape_progress &&
1176 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1178 * IO spans the reshape position. Need to wait for reshape to
1181 raid10_log(conf->mddev, "wait reshape");
1182 allow_barrier(conf);
1183 wait_event(conf->wait_barrier,
1184 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1185 conf->reshape_progress >= bio->bi_iter.bi_sector +
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: %s: redirecting sector %llu to another mirror\n",
1203 bdevname(rdev->bdev, b),
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 generic_make_request(bio);
1211 r10_bio->master_bio = bio;
1212 r10_bio->sectors = max_sectors;
1214 slot = r10_bio->read_slot;
1216 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1218 r10_bio->devs[slot].bio = read_bio;
1219 r10_bio->devs[slot].rdev = rdev;
1221 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1222 choose_data_offset(r10_bio, rdev);
1223 bio_set_dev(read_bio, rdev->bdev);
1224 read_bio->bi_end_io = raid10_end_read_request;
1225 bio_set_op_attrs(read_bio, op, do_sync);
1226 if (test_bit(FailFast, &rdev->flags) &&
1227 test_bit(R10BIO_FailFast, &r10_bio->state))
1228 read_bio->bi_opf |= MD_FAILFAST;
1229 read_bio->bi_private = r10_bio;
1232 trace_block_bio_remap(read_bio->bi_disk->queue,
1233 read_bio, disk_devt(mddev->gendisk),
1235 generic_make_request(read_bio);
1239 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1240 struct bio *bio, bool replacement,
1243 const int op = bio_op(bio);
1244 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1245 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1246 unsigned long flags;
1247 struct blk_plug_cb *cb;
1248 struct raid10_plug_cb *plug = NULL;
1249 struct r10conf *conf = mddev->private;
1250 struct md_rdev *rdev;
1251 int devnum = r10_bio->devs[n_copy].devnum;
1255 rdev = conf->mirrors[devnum].replacement;
1257 /* Replacement just got moved to main 'rdev' */
1259 rdev = conf->mirrors[devnum].rdev;
1262 rdev = conf->mirrors[devnum].rdev;
1264 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1266 r10_bio->devs[n_copy].repl_bio = mbio;
1268 r10_bio->devs[n_copy].bio = mbio;
1270 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1271 choose_data_offset(r10_bio, rdev));
1272 bio_set_dev(mbio, rdev->bdev);
1273 mbio->bi_end_io = raid10_end_write_request;
1274 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1275 if (!replacement && test_bit(FailFast,
1276 &conf->mirrors[devnum].rdev->flags)
1277 && enough(conf, devnum))
1278 mbio->bi_opf |= MD_FAILFAST;
1279 mbio->bi_private = r10_bio;
1281 if (conf->mddev->gendisk)
1282 trace_block_bio_remap(mbio->bi_disk->queue,
1283 mbio, disk_devt(conf->mddev->gendisk),
1285 /* flush_pending_writes() needs access to the rdev so...*/
1286 mbio->bi_disk = (void *)rdev;
1288 atomic_inc(&r10_bio->remaining);
1290 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1292 plug = container_of(cb, struct raid10_plug_cb, cb);
1296 bio_list_add(&plug->pending, mbio);
1297 plug->pending_cnt++;
1299 spin_lock_irqsave(&conf->device_lock, flags);
1300 bio_list_add(&conf->pending_bio_list, mbio);
1301 conf->pending_count++;
1302 spin_unlock_irqrestore(&conf->device_lock, flags);
1303 md_wakeup_thread(mddev->thread);
1307 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1308 struct r10bio *r10_bio)
1310 struct r10conf *conf = mddev->private;
1312 struct md_rdev *blocked_rdev;
1316 if ((mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector,
1319 bio_end_sector(bio)))) {
1322 prepare_to_wait(&conf->wait_barrier,
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1329 finish_wait(&conf->wait_barrier, &w);
1333 * Register the new request and wait if the reconstruction
1334 * thread has put up a bar for new requests.
1335 * Continue immediately if no resync is active currently.
1339 sectors = r10_bio->sectors;
1340 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1341 bio->bi_iter.bi_sector < conf->reshape_progress &&
1342 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1344 * IO spans the reshape position. Need to wait for reshape to
1347 raid10_log(conf->mddev, "wait reshape");
1348 allow_barrier(conf);
1349 wait_event(conf->wait_barrier,
1350 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1351 conf->reshape_progress >= bio->bi_iter.bi_sector +
1356 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1357 (mddev->reshape_backwards
1358 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1359 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1360 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1361 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1362 /* Need to update reshape_position in metadata */
1363 mddev->reshape_position = conf->reshape_progress;
1364 set_mask_bits(&mddev->sb_flags, 0,
1365 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1366 md_wakeup_thread(mddev->thread);
1367 raid10_log(conf->mddev, "wait reshape metadata");
1368 wait_event(mddev->sb_wait,
1369 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1371 conf->reshape_safe = mddev->reshape_position;
1374 if (conf->pending_count >= max_queued_requests) {
1375 md_wakeup_thread(mddev->thread);
1376 raid10_log(mddev, "wait queued");
1377 wait_event(conf->wait_barrier,
1378 conf->pending_count < max_queued_requests);
1380 /* first select target devices under rcu_lock and
1381 * inc refcount on their rdev. Record them by setting
1383 * If there are known/acknowledged bad blocks on any device
1384 * on which we have seen a write error, we want to avoid
1385 * writing to those blocks. This potentially requires several
1386 * writes to write around the bad blocks. Each set of writes
1387 * gets its own r10_bio with a set of bios attached.
1390 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1391 raid10_find_phys(conf, r10_bio);
1393 blocked_rdev = NULL;
1395 max_sectors = r10_bio->sectors;
1397 for (i = 0; i < conf->copies; i++) {
1398 int d = r10_bio->devs[i].devnum;
1399 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1400 struct md_rdev *rrdev = rcu_dereference(
1401 conf->mirrors[d].replacement);
1404 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1405 atomic_inc(&rdev->nr_pending);
1406 blocked_rdev = rdev;
1409 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1410 atomic_inc(&rrdev->nr_pending);
1411 blocked_rdev = rrdev;
1414 if (rdev && (test_bit(Faulty, &rdev->flags)))
1416 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1419 r10_bio->devs[i].bio = NULL;
1420 r10_bio->devs[i].repl_bio = NULL;
1422 if (!rdev && !rrdev) {
1423 set_bit(R10BIO_Degraded, &r10_bio->state);
1426 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1428 sector_t dev_sector = r10_bio->devs[i].addr;
1432 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1433 &first_bad, &bad_sectors);
1435 /* Mustn't write here until the bad block
1438 atomic_inc(&rdev->nr_pending);
1439 set_bit(BlockedBadBlocks, &rdev->flags);
1440 blocked_rdev = rdev;
1443 if (is_bad && first_bad <= dev_sector) {
1444 /* Cannot write here at all */
1445 bad_sectors -= (dev_sector - first_bad);
1446 if (bad_sectors < max_sectors)
1447 /* Mustn't write more than bad_sectors
1448 * to other devices yet
1450 max_sectors = bad_sectors;
1451 /* We don't set R10BIO_Degraded as that
1452 * only applies if the disk is missing,
1453 * so it might be re-added, and we want to
1454 * know to recover this chunk.
1455 * In this case the device is here, and the
1456 * fact that this chunk is not in-sync is
1457 * recorded in the bad block log.
1462 int good_sectors = first_bad - dev_sector;
1463 if (good_sectors < max_sectors)
1464 max_sectors = good_sectors;
1468 r10_bio->devs[i].bio = bio;
1469 atomic_inc(&rdev->nr_pending);
1472 r10_bio->devs[i].repl_bio = bio;
1473 atomic_inc(&rrdev->nr_pending);
1478 if (unlikely(blocked_rdev)) {
1479 /* Have to wait for this device to get unblocked, then retry */
1483 for (j = 0; j < i; j++) {
1484 if (r10_bio->devs[j].bio) {
1485 d = r10_bio->devs[j].devnum;
1486 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1488 if (r10_bio->devs[j].repl_bio) {
1489 struct md_rdev *rdev;
1490 d = r10_bio->devs[j].devnum;
1491 rdev = conf->mirrors[d].replacement;
1493 /* Race with remove_disk */
1495 rdev = conf->mirrors[d].rdev;
1497 rdev_dec_pending(rdev, mddev);
1500 allow_barrier(conf);
1501 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1502 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1507 if (max_sectors < r10_bio->sectors)
1508 r10_bio->sectors = max_sectors;
1510 if (r10_bio->sectors < bio_sectors(bio)) {
1511 struct bio *split = bio_split(bio, r10_bio->sectors,
1512 GFP_NOIO, &conf->bio_split);
1513 bio_chain(split, bio);
1514 generic_make_request(bio);
1516 r10_bio->master_bio = bio;
1519 atomic_set(&r10_bio->remaining, 1);
1520 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1522 for (i = 0; i < conf->copies; i++) {
1523 if (r10_bio->devs[i].bio)
1524 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1525 if (r10_bio->devs[i].repl_bio)
1526 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1528 one_write_done(r10_bio);
1531 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1533 struct r10conf *conf = mddev->private;
1534 struct r10bio *r10_bio;
1536 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1538 r10_bio->master_bio = bio;
1539 r10_bio->sectors = sectors;
1541 r10_bio->mddev = mddev;
1542 r10_bio->sector = bio->bi_iter.bi_sector;
1544 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1546 if (bio_data_dir(bio) == READ)
1547 raid10_read_request(mddev, bio, r10_bio);
1549 raid10_write_request(mddev, bio, r10_bio);
1552 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1554 struct r10conf *conf = mddev->private;
1555 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1556 int chunk_sects = chunk_mask + 1;
1557 int sectors = bio_sectors(bio);
1559 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1560 md_flush_request(mddev, bio);
1564 if (!md_write_start(mddev, bio))
1568 * If this request crosses a chunk boundary, we need to split
1571 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1572 sectors > chunk_sects
1573 && (conf->geo.near_copies < conf->geo.raid_disks
1574 || conf->prev.near_copies <
1575 conf->prev.raid_disks)))
1576 sectors = chunk_sects -
1577 (bio->bi_iter.bi_sector &
1579 __make_request(mddev, bio, sectors);
1581 /* In case raid10d snuck in to freeze_array */
1582 wake_up(&conf->wait_barrier);
1586 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1588 struct r10conf *conf = mddev->private;
1591 if (conf->geo.near_copies < conf->geo.raid_disks)
1592 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1593 if (conf->geo.near_copies > 1)
1594 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1595 if (conf->geo.far_copies > 1) {
1596 if (conf->geo.far_offset)
1597 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1599 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1600 if (conf->geo.far_set_size != conf->geo.raid_disks)
1601 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1603 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1604 conf->geo.raid_disks - mddev->degraded);
1606 for (i = 0; i < conf->geo.raid_disks; i++) {
1607 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1608 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1611 seq_printf(seq, "]");
1614 /* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1619 static int _enough(struct r10conf *conf, int previous, int ignore)
1625 disks = conf->prev.raid_disks;
1626 ncopies = conf->prev.near_copies;
1628 disks = conf->geo.raid_disks;
1629 ncopies = conf->geo.near_copies;
1634 int n = conf->copies;
1638 struct md_rdev *rdev;
1639 if (this != ignore &&
1640 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1641 test_bit(In_sync, &rdev->flags))
1643 this = (this+1) % disks;
1647 first = (first + ncopies) % disks;
1648 } while (first != 0);
1655 static int enough(struct r10conf *conf, int ignore)
1657 /* when calling 'enough', both 'prev' and 'geo' must
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1662 return _enough(conf, 0, ignore) &&
1663 _enough(conf, 1, ignore);
1666 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1668 char b[BDEVNAME_SIZE];
1669 struct r10conf *conf = mddev->private;
1670 unsigned long flags;
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 if (test_bit(In_sync, &rdev->flags)
1680 && !enough(conf, rdev->raid_disk)) {
1682 * Don't fail the drive, just return an IO error.
1684 spin_unlock_irqrestore(&conf->device_lock, flags);
1687 if (test_and_clear_bit(In_sync, &rdev->flags))
1690 * If recovery is running, make sure it aborts.
1692 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1693 set_bit(Blocked, &rdev->flags);
1694 set_bit(Faulty, &rdev->flags);
1695 set_mask_bits(&mddev->sb_flags, 0,
1696 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1697 spin_unlock_irqrestore(&conf->device_lock, flags);
1698 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1699 "md/raid10:%s: Operation continuing on %d devices.\n",
1700 mdname(mddev), bdevname(rdev->bdev, b),
1701 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1704 static void print_conf(struct r10conf *conf)
1707 struct md_rdev *rdev;
1709 pr_debug("RAID10 conf printout:\n");
1711 pr_debug("(!conf)\n");
1714 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1715 conf->geo.raid_disks);
1717 /* This is only called with ->reconfix_mutex held, so
1718 * rcu protection of rdev is not needed */
1719 for (i = 0; i < conf->geo.raid_disks; i++) {
1720 char b[BDEVNAME_SIZE];
1721 rdev = conf->mirrors[i].rdev;
1723 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1724 i, !test_bit(In_sync, &rdev->flags),
1725 !test_bit(Faulty, &rdev->flags),
1726 bdevname(rdev->bdev,b));
1730 static void close_sync(struct r10conf *conf)
1733 allow_barrier(conf);
1735 mempool_exit(&conf->r10buf_pool);
1738 static int raid10_spare_active(struct mddev *mddev)
1741 struct r10conf *conf = mddev->private;
1742 struct raid10_info *tmp;
1744 unsigned long flags;
1747 * Find all non-in_sync disks within the RAID10 configuration
1748 * and mark them in_sync
1750 for (i = 0; i < conf->geo.raid_disks; i++) {
1751 tmp = conf->mirrors + i;
1752 if (tmp->replacement
1753 && tmp->replacement->recovery_offset == MaxSector
1754 && !test_bit(Faulty, &tmp->replacement->flags)
1755 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1756 /* Replacement has just become active */
1758 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1761 /* Replaced device not technically faulty,
1762 * but we need to be sure it gets removed
1763 * and never re-added.
1765 set_bit(Faulty, &tmp->rdev->flags);
1766 sysfs_notify_dirent_safe(
1767 tmp->rdev->sysfs_state);
1769 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1770 } else if (tmp->rdev
1771 && tmp->rdev->recovery_offset == MaxSector
1772 && !test_bit(Faulty, &tmp->rdev->flags)
1773 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1775 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1778 spin_lock_irqsave(&conf->device_lock, flags);
1779 mddev->degraded -= count;
1780 spin_unlock_irqrestore(&conf->device_lock, flags);
1786 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1788 struct r10conf *conf = mddev->private;
1792 int last = conf->geo.raid_disks - 1;
1794 if (mddev->recovery_cp < MaxSector)
1795 /* only hot-add to in-sync arrays, as recovery is
1796 * very different from resync
1799 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1802 if (md_integrity_add_rdev(rdev, mddev))
1805 if (rdev->raid_disk >= 0)
1806 first = last = rdev->raid_disk;
1808 if (rdev->saved_raid_disk >= first &&
1809 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1810 mirror = rdev->saved_raid_disk;
1813 for ( ; mirror <= last ; mirror++) {
1814 struct raid10_info *p = &conf->mirrors[mirror];
1815 if (p->recovery_disabled == mddev->recovery_disabled)
1818 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1819 p->replacement != NULL)
1821 clear_bit(In_sync, &rdev->flags);
1822 set_bit(Replacement, &rdev->flags);
1823 rdev->raid_disk = mirror;
1826 disk_stack_limits(mddev->gendisk, rdev->bdev,
1827 rdev->data_offset << 9);
1829 rcu_assign_pointer(p->replacement, rdev);
1834 disk_stack_limits(mddev->gendisk, rdev->bdev,
1835 rdev->data_offset << 9);
1837 p->head_position = 0;
1838 p->recovery_disabled = mddev->recovery_disabled - 1;
1839 rdev->raid_disk = mirror;
1841 if (rdev->saved_raid_disk != mirror)
1843 rcu_assign_pointer(p->rdev, rdev);
1846 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1847 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1853 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1855 struct r10conf *conf = mddev->private;
1857 int number = rdev->raid_disk;
1858 struct md_rdev **rdevp;
1859 struct raid10_info *p = conf->mirrors + number;
1862 if (rdev == p->rdev)
1864 else if (rdev == p->replacement)
1865 rdevp = &p->replacement;
1869 if (test_bit(In_sync, &rdev->flags) ||
1870 atomic_read(&rdev->nr_pending)) {
1874 /* Only remove non-faulty devices if recovery
1877 if (!test_bit(Faulty, &rdev->flags) &&
1878 mddev->recovery_disabled != p->recovery_disabled &&
1879 (!p->replacement || p->replacement == rdev) &&
1880 number < conf->geo.raid_disks &&
1886 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1888 if (atomic_read(&rdev->nr_pending)) {
1889 /* lost the race, try later */
1895 if (p->replacement) {
1896 /* We must have just cleared 'rdev' */
1897 p->rdev = p->replacement;
1898 clear_bit(Replacement, &p->replacement->flags);
1899 smp_mb(); /* Make sure other CPUs may see both as identical
1900 * but will never see neither -- if they are careful.
1902 p->replacement = NULL;
1905 clear_bit(WantReplacement, &rdev->flags);
1906 err = md_integrity_register(mddev);
1914 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1916 struct r10conf *conf = r10_bio->mddev->private;
1918 if (!bio->bi_status)
1919 set_bit(R10BIO_Uptodate, &r10_bio->state);
1921 /* The write handler will notice the lack of
1922 * R10BIO_Uptodate and record any errors etc
1924 atomic_add(r10_bio->sectors,
1925 &conf->mirrors[d].rdev->corrected_errors);
1927 /* for reconstruct, we always reschedule after a read.
1928 * for resync, only after all reads
1930 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1931 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1932 atomic_dec_and_test(&r10_bio->remaining)) {
1933 /* we have read all the blocks,
1934 * do the comparison in process context in raid10d
1936 reschedule_retry(r10_bio);
1940 static void end_sync_read(struct bio *bio)
1942 struct r10bio *r10_bio = get_resync_r10bio(bio);
1943 struct r10conf *conf = r10_bio->mddev->private;
1944 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1946 __end_sync_read(r10_bio, bio, d);
1949 static void end_reshape_read(struct bio *bio)
1951 /* reshape read bio isn't allocated from r10buf_pool */
1952 struct r10bio *r10_bio = bio->bi_private;
1954 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1957 static void end_sync_request(struct r10bio *r10_bio)
1959 struct mddev *mddev = r10_bio->mddev;
1961 while (atomic_dec_and_test(&r10_bio->remaining)) {
1962 if (r10_bio->master_bio == NULL) {
1963 /* the primary of several recovery bios */
1964 sector_t s = r10_bio->sectors;
1965 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1966 test_bit(R10BIO_WriteError, &r10_bio->state))
1967 reschedule_retry(r10_bio);
1970 md_done_sync(mddev, s, 1);
1973 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1974 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1975 test_bit(R10BIO_WriteError, &r10_bio->state))
1976 reschedule_retry(r10_bio);
1984 static void end_sync_write(struct bio *bio)
1986 struct r10bio *r10_bio = get_resync_r10bio(bio);
1987 struct mddev *mddev = r10_bio->mddev;
1988 struct r10conf *conf = mddev->private;
1994 struct md_rdev *rdev = NULL;
1996 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1998 rdev = conf->mirrors[d].replacement;
2000 rdev = conf->mirrors[d].rdev;
2002 if (bio->bi_status) {
2004 md_error(mddev, rdev);
2006 set_bit(WriteErrorSeen, &rdev->flags);
2007 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2008 set_bit(MD_RECOVERY_NEEDED,
2009 &rdev->mddev->recovery);
2010 set_bit(R10BIO_WriteError, &r10_bio->state);
2012 } else if (is_badblock(rdev,
2013 r10_bio->devs[slot].addr,
2015 &first_bad, &bad_sectors))
2016 set_bit(R10BIO_MadeGood, &r10_bio->state);
2018 rdev_dec_pending(rdev, mddev);
2020 end_sync_request(r10_bio);
2024 * Note: sync and recover and handled very differently for raid10
2025 * This code is for resync.
2026 * For resync, we read through virtual addresses and read all blocks.
2027 * If there is any error, we schedule a write. The lowest numbered
2028 * drive is authoritative.
2029 * However requests come for physical address, so we need to map.
2030 * For every physical address there are raid_disks/copies virtual addresses,
2031 * which is always are least one, but is not necessarly an integer.
2032 * This means that a physical address can span multiple chunks, so we may
2033 * have to submit multiple io requests for a single sync request.
2036 * We check if all blocks are in-sync and only write to blocks that
2039 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2041 struct r10conf *conf = mddev->private;
2043 struct bio *tbio, *fbio;
2045 struct page **tpages, **fpages;
2047 atomic_set(&r10_bio->remaining, 1);
2049 /* find the first device with a block */
2050 for (i=0; i<conf->copies; i++)
2051 if (!r10_bio->devs[i].bio->bi_status)
2054 if (i == conf->copies)
2058 fbio = r10_bio->devs[i].bio;
2059 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2060 fbio->bi_iter.bi_idx = 0;
2061 fpages = get_resync_pages(fbio)->pages;
2063 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2064 /* now find blocks with errors */
2065 for (i=0 ; i < conf->copies ; i++) {
2067 struct md_rdev *rdev;
2068 struct resync_pages *rp;
2070 tbio = r10_bio->devs[i].bio;
2072 if (tbio->bi_end_io != end_sync_read)
2077 tpages = get_resync_pages(tbio)->pages;
2078 d = r10_bio->devs[i].devnum;
2079 rdev = conf->mirrors[d].rdev;
2080 if (!r10_bio->devs[i].bio->bi_status) {
2081 /* We know that the bi_io_vec layout is the same for
2082 * both 'first' and 'i', so we just compare them.
2083 * All vec entries are PAGE_SIZE;
2085 int sectors = r10_bio->sectors;
2086 for (j = 0; j < vcnt; j++) {
2087 int len = PAGE_SIZE;
2088 if (sectors < (len / 512))
2089 len = sectors * 512;
2090 if (memcmp(page_address(fpages[j]),
2091 page_address(tpages[j]),
2098 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2099 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2100 /* Don't fix anything. */
2102 } else if (test_bit(FailFast, &rdev->flags)) {
2103 /* Just give up on this device */
2104 md_error(rdev->mddev, rdev);
2107 /* Ok, we need to write this bio, either to correct an
2108 * inconsistency or to correct an unreadable block.
2109 * First we need to fixup bv_offset, bv_len and
2110 * bi_vecs, as the read request might have corrupted these
2112 rp = get_resync_pages(tbio);
2115 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2117 rp->raid_bio = r10_bio;
2118 tbio->bi_private = rp;
2119 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2120 tbio->bi_end_io = end_sync_write;
2121 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2123 bio_copy_data(tbio, fbio);
2125 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2126 atomic_inc(&r10_bio->remaining);
2127 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2129 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2130 tbio->bi_opf |= MD_FAILFAST;
2131 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2132 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2133 generic_make_request(tbio);
2136 /* Now write out to any replacement devices
2139 for (i = 0; i < conf->copies; i++) {
2142 tbio = r10_bio->devs[i].repl_bio;
2143 if (!tbio || !tbio->bi_end_io)
2145 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2146 && r10_bio->devs[i].bio != fbio)
2147 bio_copy_data(tbio, fbio);
2148 d = r10_bio->devs[i].devnum;
2149 atomic_inc(&r10_bio->remaining);
2150 md_sync_acct(conf->mirrors[d].replacement->bdev,
2152 generic_make_request(tbio);
2156 if (atomic_dec_and_test(&r10_bio->remaining)) {
2157 md_done_sync(mddev, r10_bio->sectors, 1);
2163 * Now for the recovery code.
2164 * Recovery happens across physical sectors.
2165 * We recover all non-is_sync drives by finding the virtual address of
2166 * each, and then choose a working drive that also has that virt address.
2167 * There is a separate r10_bio for each non-in_sync drive.
2168 * Only the first two slots are in use. The first for reading,
2169 * The second for writing.
2172 static void fix_recovery_read_error(struct r10bio *r10_bio)
2174 /* We got a read error during recovery.
2175 * We repeat the read in smaller page-sized sections.
2176 * If a read succeeds, write it to the new device or record
2177 * a bad block if we cannot.
2178 * If a read fails, record a bad block on both old and
2181 struct mddev *mddev = r10_bio->mddev;
2182 struct r10conf *conf = mddev->private;
2183 struct bio *bio = r10_bio->devs[0].bio;
2185 int sectors = r10_bio->sectors;
2187 int dr = r10_bio->devs[0].devnum;
2188 int dw = r10_bio->devs[1].devnum;
2189 struct page **pages = get_resync_pages(bio)->pages;
2193 struct md_rdev *rdev;
2197 if (s > (PAGE_SIZE>>9))
2200 rdev = conf->mirrors[dr].rdev;
2201 addr = r10_bio->devs[0].addr + sect,
2202 ok = sync_page_io(rdev,
2206 REQ_OP_READ, 0, false);
2208 rdev = conf->mirrors[dw].rdev;
2209 addr = r10_bio->devs[1].addr + sect;
2210 ok = sync_page_io(rdev,
2214 REQ_OP_WRITE, 0, false);
2216 set_bit(WriteErrorSeen, &rdev->flags);
2217 if (!test_and_set_bit(WantReplacement,
2219 set_bit(MD_RECOVERY_NEEDED,
2220 &rdev->mddev->recovery);
2224 /* We don't worry if we cannot set a bad block -
2225 * it really is bad so there is no loss in not
2228 rdev_set_badblocks(rdev, addr, s, 0);
2230 if (rdev != conf->mirrors[dw].rdev) {
2231 /* need bad block on destination too */
2232 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2233 addr = r10_bio->devs[1].addr + sect;
2234 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2236 /* just abort the recovery */
2237 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2240 conf->mirrors[dw].recovery_disabled
2241 = mddev->recovery_disabled;
2242 set_bit(MD_RECOVERY_INTR,
2255 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2257 struct r10conf *conf = mddev->private;
2259 struct bio *wbio, *wbio2;
2261 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2262 fix_recovery_read_error(r10_bio);
2263 end_sync_request(r10_bio);
2268 * share the pages with the first bio
2269 * and submit the write request
2271 d = r10_bio->devs[1].devnum;
2272 wbio = r10_bio->devs[1].bio;
2273 wbio2 = r10_bio->devs[1].repl_bio;
2274 /* Need to test wbio2->bi_end_io before we call
2275 * generic_make_request as if the former is NULL,
2276 * the latter is free to free wbio2.
2278 if (wbio2 && !wbio2->bi_end_io)
2280 if (wbio->bi_end_io) {
2281 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2282 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2283 generic_make_request(wbio);
2286 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2287 md_sync_acct(conf->mirrors[d].replacement->bdev,
2288 bio_sectors(wbio2));
2289 generic_make_request(wbio2);
2294 * Used by fix_read_error() to decay the per rdev read_errors.
2295 * We halve the read error count for every hour that has elapsed
2296 * since the last recorded read error.
2299 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2302 unsigned long hours_since_last;
2303 unsigned int read_errors = atomic_read(&rdev->read_errors);
2305 cur_time_mon = ktime_get_seconds();
2307 if (rdev->last_read_error == 0) {
2308 /* first time we've seen a read error */
2309 rdev->last_read_error = cur_time_mon;
2313 hours_since_last = (long)(cur_time_mon -
2314 rdev->last_read_error) / 3600;
2316 rdev->last_read_error = cur_time_mon;
2319 * if hours_since_last is > the number of bits in read_errors
2320 * just set read errors to 0. We do this to avoid
2321 * overflowing the shift of read_errors by hours_since_last.
2323 if (hours_since_last >= 8 * sizeof(read_errors))
2324 atomic_set(&rdev->read_errors, 0);
2326 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2329 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2330 int sectors, struct page *page, int rw)
2335 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2336 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2338 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2342 set_bit(WriteErrorSeen, &rdev->flags);
2343 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2344 set_bit(MD_RECOVERY_NEEDED,
2345 &rdev->mddev->recovery);
2347 /* need to record an error - either for the block or the device */
2348 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2349 md_error(rdev->mddev, rdev);
2354 * This is a kernel thread which:
2356 * 1. Retries failed read operations on working mirrors.
2357 * 2. Updates the raid superblock when problems encounter.
2358 * 3. Performs writes following reads for array synchronising.
2361 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2363 int sect = 0; /* Offset from r10_bio->sector */
2364 int sectors = r10_bio->sectors;
2365 struct md_rdev*rdev;
2366 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2367 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2369 /* still own a reference to this rdev, so it cannot
2370 * have been cleared recently.
2372 rdev = conf->mirrors[d].rdev;
2374 if (test_bit(Faulty, &rdev->flags))
2375 /* drive has already been failed, just ignore any
2376 more fix_read_error() attempts */
2379 check_decay_read_errors(mddev, rdev);
2380 atomic_inc(&rdev->read_errors);
2381 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2382 char b[BDEVNAME_SIZE];
2383 bdevname(rdev->bdev, b);
2385 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2387 atomic_read(&rdev->read_errors), max_read_errors);
2388 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2390 md_error(mddev, rdev);
2391 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2397 int sl = r10_bio->read_slot;
2401 if (s > (PAGE_SIZE>>9))
2409 d = r10_bio->devs[sl].devnum;
2410 rdev = rcu_dereference(conf->mirrors[d].rdev);
2412 test_bit(In_sync, &rdev->flags) &&
2413 !test_bit(Faulty, &rdev->flags) &&
2414 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2415 &first_bad, &bad_sectors) == 0) {
2416 atomic_inc(&rdev->nr_pending);
2418 success = sync_page_io(rdev,
2419 r10_bio->devs[sl].addr +
2423 REQ_OP_READ, 0, false);
2424 rdev_dec_pending(rdev, mddev);
2430 if (sl == conf->copies)
2432 } while (!success && sl != r10_bio->read_slot);
2436 /* Cannot read from anywhere, just mark the block
2437 * as bad on the first device to discourage future
2440 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2441 rdev = conf->mirrors[dn].rdev;
2443 if (!rdev_set_badblocks(
2445 r10_bio->devs[r10_bio->read_slot].addr
2448 md_error(mddev, rdev);
2449 r10_bio->devs[r10_bio->read_slot].bio
2456 /* write it back and re-read */
2458 while (sl != r10_bio->read_slot) {
2459 char b[BDEVNAME_SIZE];
2464 d = r10_bio->devs[sl].devnum;
2465 rdev = rcu_dereference(conf->mirrors[d].rdev);
2467 test_bit(Faulty, &rdev->flags) ||
2468 !test_bit(In_sync, &rdev->flags))
2471 atomic_inc(&rdev->nr_pending);
2473 if (r10_sync_page_io(rdev,
2474 r10_bio->devs[sl].addr +
2476 s, conf->tmppage, WRITE)
2478 /* Well, this device is dead */
2479 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2481 (unsigned long long)(
2483 choose_data_offset(r10_bio,
2485 bdevname(rdev->bdev, b));
2486 pr_notice("md/raid10:%s: %s: failing drive\n",
2488 bdevname(rdev->bdev, b));
2490 rdev_dec_pending(rdev, mddev);
2494 while (sl != r10_bio->read_slot) {
2495 char b[BDEVNAME_SIZE];
2500 d = r10_bio->devs[sl].devnum;
2501 rdev = rcu_dereference(conf->mirrors[d].rdev);
2503 test_bit(Faulty, &rdev->flags) ||
2504 !test_bit(In_sync, &rdev->flags))
2507 atomic_inc(&rdev->nr_pending);
2509 switch (r10_sync_page_io(rdev,
2510 r10_bio->devs[sl].addr +
2515 /* Well, this device is dead */
2516 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2518 (unsigned long long)(
2520 choose_data_offset(r10_bio, rdev)),
2521 bdevname(rdev->bdev, b));
2522 pr_notice("md/raid10:%s: %s: failing drive\n",
2524 bdevname(rdev->bdev, b));
2527 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2529 (unsigned long long)(
2531 choose_data_offset(r10_bio, rdev)),
2532 bdevname(rdev->bdev, b));
2533 atomic_add(s, &rdev->corrected_errors);
2536 rdev_dec_pending(rdev, mddev);
2546 static int narrow_write_error(struct r10bio *r10_bio, int i)
2548 struct bio *bio = r10_bio->master_bio;
2549 struct mddev *mddev = r10_bio->mddev;
2550 struct r10conf *conf = mddev->private;
2551 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2552 /* bio has the data to be written to slot 'i' where
2553 * we just recently had a write error.
2554 * We repeatedly clone the bio and trim down to one block,
2555 * then try the write. Where the write fails we record
2557 * It is conceivable that the bio doesn't exactly align with
2558 * blocks. We must handle this.
2560 * We currently own a reference to the rdev.
2566 int sect_to_write = r10_bio->sectors;
2569 if (rdev->badblocks.shift < 0)
2572 block_sectors = roundup(1 << rdev->badblocks.shift,
2573 bdev_logical_block_size(rdev->bdev) >> 9);
2574 sector = r10_bio->sector;
2575 sectors = ((r10_bio->sector + block_sectors)
2576 & ~(sector_t)(block_sectors - 1))
2579 while (sect_to_write) {
2582 if (sectors > sect_to_write)
2583 sectors = sect_to_write;
2584 /* Write at 'sector' for 'sectors' */
2585 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2586 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2587 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2588 wbio->bi_iter.bi_sector = wsector +
2589 choose_data_offset(r10_bio, rdev);
2590 bio_set_dev(wbio, rdev->bdev);
2591 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2593 if (submit_bio_wait(wbio) < 0)
2595 ok = rdev_set_badblocks(rdev, wsector,
2600 sect_to_write -= sectors;
2602 sectors = block_sectors;
2607 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2609 int slot = r10_bio->read_slot;
2611 struct r10conf *conf = mddev->private;
2612 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2614 /* we got a read error. Maybe the drive is bad. Maybe just
2615 * the block and we can fix it.
2616 * We freeze all other IO, and try reading the block from
2617 * other devices. When we find one, we re-write
2618 * and check it that fixes the read error.
2619 * This is all done synchronously while the array is
2622 bio = r10_bio->devs[slot].bio;
2624 r10_bio->devs[slot].bio = NULL;
2627 r10_bio->devs[slot].bio = IO_BLOCKED;
2628 else if (!test_bit(FailFast, &rdev->flags)) {
2629 freeze_array(conf, 1);
2630 fix_read_error(conf, mddev, r10_bio);
2631 unfreeze_array(conf);
2633 md_error(mddev, rdev);
2635 rdev_dec_pending(rdev, mddev);
2636 allow_barrier(conf);
2638 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2641 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2643 /* Some sort of write request has finished and it
2644 * succeeded in writing where we thought there was a
2645 * bad block. So forget the bad block.
2646 * Or possibly if failed and we need to record
2650 struct md_rdev *rdev;
2652 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2653 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2654 for (m = 0; m < conf->copies; m++) {
2655 int dev = r10_bio->devs[m].devnum;
2656 rdev = conf->mirrors[dev].rdev;
2657 if (r10_bio->devs[m].bio == NULL ||
2658 r10_bio->devs[m].bio->bi_end_io == NULL)
2660 if (!r10_bio->devs[m].bio->bi_status) {
2661 rdev_clear_badblocks(
2663 r10_bio->devs[m].addr,
2664 r10_bio->sectors, 0);
2666 if (!rdev_set_badblocks(
2668 r10_bio->devs[m].addr,
2669 r10_bio->sectors, 0))
2670 md_error(conf->mddev, rdev);
2672 rdev = conf->mirrors[dev].replacement;
2673 if (r10_bio->devs[m].repl_bio == NULL ||
2674 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2677 if (!r10_bio->devs[m].repl_bio->bi_status) {
2678 rdev_clear_badblocks(
2680 r10_bio->devs[m].addr,
2681 r10_bio->sectors, 0);
2683 if (!rdev_set_badblocks(
2685 r10_bio->devs[m].addr,
2686 r10_bio->sectors, 0))
2687 md_error(conf->mddev, rdev);
2693 for (m = 0; m < conf->copies; m++) {
2694 int dev = r10_bio->devs[m].devnum;
2695 struct bio *bio = r10_bio->devs[m].bio;
2696 rdev = conf->mirrors[dev].rdev;
2697 if (bio == IO_MADE_GOOD) {
2698 rdev_clear_badblocks(
2700 r10_bio->devs[m].addr,
2701 r10_bio->sectors, 0);
2702 rdev_dec_pending(rdev, conf->mddev);
2703 } else if (bio != NULL && bio->bi_status) {
2705 if (!narrow_write_error(r10_bio, m)) {
2706 md_error(conf->mddev, rdev);
2707 set_bit(R10BIO_Degraded,
2710 rdev_dec_pending(rdev, conf->mddev);
2712 bio = r10_bio->devs[m].repl_bio;
2713 rdev = conf->mirrors[dev].replacement;
2714 if (rdev && bio == IO_MADE_GOOD) {
2715 rdev_clear_badblocks(
2717 r10_bio->devs[m].addr,
2718 r10_bio->sectors, 0);
2719 rdev_dec_pending(rdev, conf->mddev);
2723 spin_lock_irq(&conf->device_lock);
2724 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2726 spin_unlock_irq(&conf->device_lock);
2728 * In case freeze_array() is waiting for condition
2729 * nr_pending == nr_queued + extra to be true.
2731 wake_up(&conf->wait_barrier);
2732 md_wakeup_thread(conf->mddev->thread);
2734 if (test_bit(R10BIO_WriteError,
2736 close_write(r10_bio);
2737 raid_end_bio_io(r10_bio);
2742 static void raid10d(struct md_thread *thread)
2744 struct mddev *mddev = thread->mddev;
2745 struct r10bio *r10_bio;
2746 unsigned long flags;
2747 struct r10conf *conf = mddev->private;
2748 struct list_head *head = &conf->retry_list;
2749 struct blk_plug plug;
2751 md_check_recovery(mddev);
2753 if (!list_empty_careful(&conf->bio_end_io_list) &&
2754 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2756 spin_lock_irqsave(&conf->device_lock, flags);
2757 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2758 while (!list_empty(&conf->bio_end_io_list)) {
2759 list_move(conf->bio_end_io_list.prev, &tmp);
2763 spin_unlock_irqrestore(&conf->device_lock, flags);
2764 while (!list_empty(&tmp)) {
2765 r10_bio = list_first_entry(&tmp, struct r10bio,
2767 list_del(&r10_bio->retry_list);
2768 if (mddev->degraded)
2769 set_bit(R10BIO_Degraded, &r10_bio->state);
2771 if (test_bit(R10BIO_WriteError,
2773 close_write(r10_bio);
2774 raid_end_bio_io(r10_bio);
2778 blk_start_plug(&plug);
2781 flush_pending_writes(conf);
2783 spin_lock_irqsave(&conf->device_lock, flags);
2784 if (list_empty(head)) {
2785 spin_unlock_irqrestore(&conf->device_lock, flags);
2788 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2789 list_del(head->prev);
2791 spin_unlock_irqrestore(&conf->device_lock, flags);
2793 mddev = r10_bio->mddev;
2794 conf = mddev->private;
2795 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2796 test_bit(R10BIO_WriteError, &r10_bio->state))
2797 handle_write_completed(conf, r10_bio);
2798 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2799 reshape_request_write(mddev, r10_bio);
2800 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2801 sync_request_write(mddev, r10_bio);
2802 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2803 recovery_request_write(mddev, r10_bio);
2804 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2805 handle_read_error(mddev, r10_bio);
2810 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2811 md_check_recovery(mddev);
2813 blk_finish_plug(&plug);
2816 static int init_resync(struct r10conf *conf)
2820 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2821 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2822 conf->have_replacement = 0;
2823 for (i = 0; i < conf->geo.raid_disks; i++)
2824 if (conf->mirrors[i].replacement)
2825 conf->have_replacement = 1;
2826 ret = mempool_init(&conf->r10buf_pool, buffs,
2827 r10buf_pool_alloc, r10buf_pool_free, conf);
2830 conf->next_resync = 0;
2834 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2836 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2837 struct rsync_pages *rp;
2842 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2843 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2844 nalloc = conf->copies; /* resync */
2846 nalloc = 2; /* recovery */
2848 for (i = 0; i < nalloc; i++) {
2849 bio = r10bio->devs[i].bio;
2850 rp = bio->bi_private;
2852 bio->bi_private = rp;
2853 bio = r10bio->devs[i].repl_bio;
2855 rp = bio->bi_private;
2857 bio->bi_private = rp;
2864 * Set cluster_sync_high since we need other nodes to add the
2865 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2867 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2869 sector_t window_size;
2870 int extra_chunk, chunks;
2873 * First, here we define "stripe" as a unit which across
2874 * all member devices one time, so we get chunks by use
2875 * raid_disks / near_copies. Otherwise, if near_copies is
2876 * close to raid_disks, then resync window could increases
2877 * linearly with the increase of raid_disks, which means
2878 * we will suspend a really large IO window while it is not
2879 * necessary. If raid_disks is not divisible by near_copies,
2880 * an extra chunk is needed to ensure the whole "stripe" is
2884 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2885 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2889 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2892 * At least use a 32M window to align with raid1's resync window
2894 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2895 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2897 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2901 * perform a "sync" on one "block"
2903 * We need to make sure that no normal I/O request - particularly write
2904 * requests - conflict with active sync requests.
2906 * This is achieved by tracking pending requests and a 'barrier' concept
2907 * that can be installed to exclude normal IO requests.
2909 * Resync and recovery are handled very differently.
2910 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2912 * For resync, we iterate over virtual addresses, read all copies,
2913 * and update if there are differences. If only one copy is live,
2915 * For recovery, we iterate over physical addresses, read a good
2916 * value for each non-in_sync drive, and over-write.
2918 * So, for recovery we may have several outstanding complex requests for a
2919 * given address, one for each out-of-sync device. We model this by allocating
2920 * a number of r10_bio structures, one for each out-of-sync device.
2921 * As we setup these structures, we collect all bio's together into a list
2922 * which we then process collectively to add pages, and then process again
2923 * to pass to generic_make_request.
2925 * The r10_bio structures are linked using a borrowed master_bio pointer.
2926 * This link is counted in ->remaining. When the r10_bio that points to NULL
2927 * has its remaining count decremented to 0, the whole complex operation
2932 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2935 struct r10conf *conf = mddev->private;
2936 struct r10bio *r10_bio;
2937 struct bio *biolist = NULL, *bio;
2938 sector_t max_sector, nr_sectors;
2941 sector_t sync_blocks;
2942 sector_t sectors_skipped = 0;
2943 int chunks_skipped = 0;
2944 sector_t chunk_mask = conf->geo.chunk_mask;
2947 if (!mempool_initialized(&conf->r10buf_pool))
2948 if (init_resync(conf))
2952 * Allow skipping a full rebuild for incremental assembly
2953 * of a clean array, like RAID1 does.
2955 if (mddev->bitmap == NULL &&
2956 mddev->recovery_cp == MaxSector &&
2957 mddev->reshape_position == MaxSector &&
2958 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2959 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2960 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2961 conf->fullsync == 0) {
2963 return mddev->dev_sectors - sector_nr;
2967 max_sector = mddev->dev_sectors;
2968 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2969 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2970 max_sector = mddev->resync_max_sectors;
2971 if (sector_nr >= max_sector) {
2972 conf->cluster_sync_low = 0;
2973 conf->cluster_sync_high = 0;
2975 /* If we aborted, we need to abort the
2976 * sync on the 'current' bitmap chucks (there can
2977 * be several when recovering multiple devices).
2978 * as we may have started syncing it but not finished.
2979 * We can find the current address in
2980 * mddev->curr_resync, but for recovery,
2981 * we need to convert that to several
2982 * virtual addresses.
2984 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2990 if (mddev->curr_resync < max_sector) { /* aborted */
2991 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2992 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2994 else for (i = 0; i < conf->geo.raid_disks; i++) {
2996 raid10_find_virt(conf, mddev->curr_resync, i);
2997 bitmap_end_sync(mddev->bitmap, sect,
3001 /* completed sync */
3002 if ((!mddev->bitmap || conf->fullsync)
3003 && conf->have_replacement
3004 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3005 /* Completed a full sync so the replacements
3006 * are now fully recovered.
3009 for (i = 0; i < conf->geo.raid_disks; i++) {
3010 struct md_rdev *rdev =
3011 rcu_dereference(conf->mirrors[i].replacement);
3013 rdev->recovery_offset = MaxSector;
3019 bitmap_close_sync(mddev->bitmap);
3022 return sectors_skipped;
3025 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3026 return reshape_request(mddev, sector_nr, skipped);
3028 if (chunks_skipped >= conf->geo.raid_disks) {
3029 /* if there has been nothing to do on any drive,
3030 * then there is nothing to do at all..
3033 return (max_sector - sector_nr) + sectors_skipped;
3036 if (max_sector > mddev->resync_max)
3037 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3039 /* make sure whole request will fit in a chunk - if chunks
3042 if (conf->geo.near_copies < conf->geo.raid_disks &&
3043 max_sector > (sector_nr | chunk_mask))
3044 max_sector = (sector_nr | chunk_mask) + 1;
3047 * If there is non-resync activity waiting for a turn, then let it
3048 * though before starting on this new sync request.
3050 if (conf->nr_waiting)
3051 schedule_timeout_uninterruptible(1);
3053 /* Again, very different code for resync and recovery.
3054 * Both must result in an r10bio with a list of bios that
3055 * have bi_end_io, bi_sector, bi_disk set,
3056 * and bi_private set to the r10bio.
3057 * For recovery, we may actually create several r10bios
3058 * with 2 bios in each, that correspond to the bios in the main one.
3059 * In this case, the subordinate r10bios link back through a
3060 * borrowed master_bio pointer, and the counter in the master
3061 * includes a ref from each subordinate.
3063 /* First, we decide what to do and set ->bi_end_io
3064 * To end_sync_read if we want to read, and
3065 * end_sync_write if we will want to write.
3068 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3069 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3070 /* recovery... the complicated one */
3074 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3080 struct raid10_info *mirror = &conf->mirrors[i];
3081 struct md_rdev *mrdev, *mreplace;
3084 mrdev = rcu_dereference(mirror->rdev);
3085 mreplace = rcu_dereference(mirror->replacement);
3087 if ((mrdev == NULL ||
3088 test_bit(Faulty, &mrdev->flags) ||
3089 test_bit(In_sync, &mrdev->flags)) &&
3090 (mreplace == NULL ||
3091 test_bit(Faulty, &mreplace->flags))) {
3097 /* want to reconstruct this device */
3099 sect = raid10_find_virt(conf, sector_nr, i);
3100 if (sect >= mddev->resync_max_sectors) {
3101 /* last stripe is not complete - don't
3102 * try to recover this sector.
3107 if (mreplace && test_bit(Faulty, &mreplace->flags))
3109 /* Unless we are doing a full sync, or a replacement
3110 * we only need to recover the block if it is set in
3113 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3115 if (sync_blocks < max_sync)
3116 max_sync = sync_blocks;
3120 /* yep, skip the sync_blocks here, but don't assume
3121 * that there will never be anything to do here
3123 chunks_skipped = -1;
3127 atomic_inc(&mrdev->nr_pending);
3129 atomic_inc(&mreplace->nr_pending);
3132 r10_bio = raid10_alloc_init_r10buf(conf);
3134 raise_barrier(conf, rb2 != NULL);
3135 atomic_set(&r10_bio->remaining, 0);
3137 r10_bio->master_bio = (struct bio*)rb2;
3139 atomic_inc(&rb2->remaining);
3140 r10_bio->mddev = mddev;
3141 set_bit(R10BIO_IsRecover, &r10_bio->state);
3142 r10_bio->sector = sect;
3144 raid10_find_phys(conf, r10_bio);
3146 /* Need to check if the array will still be
3150 for (j = 0; j < conf->geo.raid_disks; j++) {
3151 struct md_rdev *rdev = rcu_dereference(
3152 conf->mirrors[j].rdev);
3153 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3159 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3160 &sync_blocks, still_degraded);
3163 for (j=0; j<conf->copies;j++) {
3165 int d = r10_bio->devs[j].devnum;
3166 sector_t from_addr, to_addr;
3167 struct md_rdev *rdev =
3168 rcu_dereference(conf->mirrors[d].rdev);
3169 sector_t sector, first_bad;
3172 !test_bit(In_sync, &rdev->flags))
3174 /* This is where we read from */
3176 sector = r10_bio->devs[j].addr;
3178 if (is_badblock(rdev, sector, max_sync,
3179 &first_bad, &bad_sectors)) {
3180 if (first_bad > sector)
3181 max_sync = first_bad - sector;
3183 bad_sectors -= (sector
3185 if (max_sync > bad_sectors)
3186 max_sync = bad_sectors;
3190 bio = r10_bio->devs[0].bio;
3191 bio->bi_next = biolist;
3193 bio->bi_end_io = end_sync_read;
3194 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3195 if (test_bit(FailFast, &rdev->flags))
3196 bio->bi_opf |= MD_FAILFAST;
3197 from_addr = r10_bio->devs[j].addr;
3198 bio->bi_iter.bi_sector = from_addr +
3200 bio_set_dev(bio, rdev->bdev);
3201 atomic_inc(&rdev->nr_pending);
3202 /* and we write to 'i' (if not in_sync) */
3204 for (k=0; k<conf->copies; k++)
3205 if (r10_bio->devs[k].devnum == i)
3207 BUG_ON(k == conf->copies);
3208 to_addr = r10_bio->devs[k].addr;
3209 r10_bio->devs[0].devnum = d;
3210 r10_bio->devs[0].addr = from_addr;
3211 r10_bio->devs[1].devnum = i;
3212 r10_bio->devs[1].addr = to_addr;
3214 if (!test_bit(In_sync, &mrdev->flags)) {
3215 bio = r10_bio->devs[1].bio;
3216 bio->bi_next = biolist;
3218 bio->bi_end_io = end_sync_write;
3219 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3220 bio->bi_iter.bi_sector = to_addr
3221 + mrdev->data_offset;
3222 bio_set_dev(bio, mrdev->bdev);
3223 atomic_inc(&r10_bio->remaining);
3225 r10_bio->devs[1].bio->bi_end_io = NULL;
3227 /* and maybe write to replacement */
3228 bio = r10_bio->devs[1].repl_bio;
3230 bio->bi_end_io = NULL;
3231 /* Note: if mreplace != NULL, then bio
3232 * cannot be NULL as r10buf_pool_alloc will
3233 * have allocated it.
3234 * So the second test here is pointless.
3235 * But it keeps semantic-checkers happy, and
3236 * this comment keeps human reviewers
3239 if (mreplace == NULL || bio == NULL ||
3240 test_bit(Faulty, &mreplace->flags))
3242 bio->bi_next = biolist;
3244 bio->bi_end_io = end_sync_write;
3245 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3246 bio->bi_iter.bi_sector = to_addr +
3247 mreplace->data_offset;
3248 bio_set_dev(bio, mreplace->bdev);
3249 atomic_inc(&r10_bio->remaining);
3253 if (j == conf->copies) {
3254 /* Cannot recover, so abort the recovery or
3255 * record a bad block */
3257 /* problem is that there are bad blocks
3258 * on other device(s)
3261 for (k = 0; k < conf->copies; k++)
3262 if (r10_bio->devs[k].devnum == i)
3264 if (!test_bit(In_sync,
3266 && !rdev_set_badblocks(
3268 r10_bio->devs[k].addr,
3272 !rdev_set_badblocks(
3274 r10_bio->devs[k].addr,
3279 if (!test_and_set_bit(MD_RECOVERY_INTR,
3281 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3283 mirror->recovery_disabled
3284 = mddev->recovery_disabled;
3288 atomic_dec(&rb2->remaining);
3290 rdev_dec_pending(mrdev, mddev);
3292 rdev_dec_pending(mreplace, mddev);
3295 rdev_dec_pending(mrdev, mddev);
3297 rdev_dec_pending(mreplace, mddev);
3298 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3299 /* Only want this if there is elsewhere to
3300 * read from. 'j' is currently the first
3304 for (; j < conf->copies; j++) {
3305 int d = r10_bio->devs[j].devnum;
3306 if (conf->mirrors[d].rdev &&
3308 &conf->mirrors[d].rdev->flags))
3312 r10_bio->devs[0].bio->bi_opf
3316 if (biolist == NULL) {
3318 struct r10bio *rb2 = r10_bio;
3319 r10_bio = (struct r10bio*) rb2->master_bio;
3320 rb2->master_bio = NULL;
3326 /* resync. Schedule a read for every block at this virt offset */
3330 * Since curr_resync_completed could probably not update in
3331 * time, and we will set cluster_sync_low based on it.
3332 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3333 * safety reason, which ensures curr_resync_completed is
3334 * updated in bitmap_cond_end_sync.
3336 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3337 mddev_is_clustered(mddev) &&
3338 (sector_nr + 2 * RESYNC_SECTORS >
3339 conf->cluster_sync_high));
3341 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3342 &sync_blocks, mddev->degraded) &&
3343 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3344 &mddev->recovery)) {
3345 /* We can skip this block */
3347 return sync_blocks + sectors_skipped;
3349 if (sync_blocks < max_sync)
3350 max_sync = sync_blocks;
3351 r10_bio = raid10_alloc_init_r10buf(conf);
3354 r10_bio->mddev = mddev;
3355 atomic_set(&r10_bio->remaining, 0);
3356 raise_barrier(conf, 0);
3357 conf->next_resync = sector_nr;
3359 r10_bio->master_bio = NULL;
3360 r10_bio->sector = sector_nr;
3361 set_bit(R10BIO_IsSync, &r10_bio->state);
3362 raid10_find_phys(conf, r10_bio);
3363 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3365 for (i = 0; i < conf->copies; i++) {
3366 int d = r10_bio->devs[i].devnum;
3367 sector_t first_bad, sector;
3369 struct md_rdev *rdev;
3371 if (r10_bio->devs[i].repl_bio)
3372 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3374 bio = r10_bio->devs[i].bio;
3375 bio->bi_status = BLK_STS_IOERR;
3377 rdev = rcu_dereference(conf->mirrors[d].rdev);
3378 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3382 sector = r10_bio->devs[i].addr;
3383 if (is_badblock(rdev, sector, max_sync,
3384 &first_bad, &bad_sectors)) {
3385 if (first_bad > sector)
3386 max_sync = first_bad - sector;
3388 bad_sectors -= (sector - first_bad);
3389 if (max_sync > bad_sectors)
3390 max_sync = bad_sectors;
3395 atomic_inc(&rdev->nr_pending);
3396 atomic_inc(&r10_bio->remaining);
3397 bio->bi_next = biolist;
3399 bio->bi_end_io = end_sync_read;
3400 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3401 if (test_bit(FailFast, &rdev->flags))
3402 bio->bi_opf |= MD_FAILFAST;
3403 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3404 bio_set_dev(bio, rdev->bdev);
3407 rdev = rcu_dereference(conf->mirrors[d].replacement);
3408 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3412 atomic_inc(&rdev->nr_pending);
3414 /* Need to set up for writing to the replacement */
3415 bio = r10_bio->devs[i].repl_bio;
3416 bio->bi_status = BLK_STS_IOERR;
3418 sector = r10_bio->devs[i].addr;
3419 bio->bi_next = biolist;
3421 bio->bi_end_io = end_sync_write;
3422 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3423 if (test_bit(FailFast, &rdev->flags))
3424 bio->bi_opf |= MD_FAILFAST;
3425 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3426 bio_set_dev(bio, rdev->bdev);
3432 for (i=0; i<conf->copies; i++) {
3433 int d = r10_bio->devs[i].devnum;
3434 if (r10_bio->devs[i].bio->bi_end_io)
3435 rdev_dec_pending(conf->mirrors[d].rdev,
3437 if (r10_bio->devs[i].repl_bio &&
3438 r10_bio->devs[i].repl_bio->bi_end_io)
3440 conf->mirrors[d].replacement,
3450 if (sector_nr + max_sync < max_sector)
3451 max_sector = sector_nr + max_sync;
3454 int len = PAGE_SIZE;
3455 if (sector_nr + (len>>9) > max_sector)
3456 len = (max_sector - sector_nr) << 9;
3459 for (bio= biolist ; bio ; bio=bio->bi_next) {
3460 struct resync_pages *rp = get_resync_pages(bio);
3461 page = resync_fetch_page(rp, page_idx);
3463 * won't fail because the vec table is big enough
3464 * to hold all these pages
3466 bio_add_page(bio, page, len, 0);
3468 nr_sectors += len>>9;
3469 sector_nr += len>>9;
3470 } while (++page_idx < RESYNC_PAGES);
3471 r10_bio->sectors = nr_sectors;
3473 if (mddev_is_clustered(mddev) &&
3474 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3475 /* It is resync not recovery */
3476 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3477 conf->cluster_sync_low = mddev->curr_resync_completed;
3478 raid10_set_cluster_sync_high(conf);
3479 /* Send resync message */
3480 md_cluster_ops->resync_info_update(mddev,
3481 conf->cluster_sync_low,
3482 conf->cluster_sync_high);
3484 } else if (mddev_is_clustered(mddev)) {
3485 /* This is recovery not resync */
3486 sector_t sect_va1, sect_va2;
3487 bool broadcast_msg = false;
3489 for (i = 0; i < conf->geo.raid_disks; i++) {
3491 * sector_nr is a device address for recovery, so we
3492 * need translate it to array address before compare
3493 * with cluster_sync_high.
3495 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3497 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3498 broadcast_msg = true;
3500 * curr_resync_completed is similar as
3501 * sector_nr, so make the translation too.
3503 sect_va2 = raid10_find_virt(conf,
3504 mddev->curr_resync_completed, i);
3506 if (conf->cluster_sync_low == 0 ||
3507 conf->cluster_sync_low > sect_va2)
3508 conf->cluster_sync_low = sect_va2;
3511 if (broadcast_msg) {
3512 raid10_set_cluster_sync_high(conf);
3513 md_cluster_ops->resync_info_update(mddev,
3514 conf->cluster_sync_low,
3515 conf->cluster_sync_high);
3521 biolist = biolist->bi_next;
3523 bio->bi_next = NULL;
3524 r10_bio = get_resync_r10bio(bio);
3525 r10_bio->sectors = nr_sectors;
3527 if (bio->bi_end_io == end_sync_read) {
3528 md_sync_acct_bio(bio, nr_sectors);
3530 generic_make_request(bio);
3534 if (sectors_skipped)
3535 /* pretend they weren't skipped, it makes
3536 * no important difference in this case
3538 md_done_sync(mddev, sectors_skipped, 1);
3540 return sectors_skipped + nr_sectors;
3542 /* There is nowhere to write, so all non-sync
3543 * drives must be failed or in resync, all drives
3544 * have a bad block, so try the next chunk...
3546 if (sector_nr + max_sync < max_sector)
3547 max_sector = sector_nr + max_sync;
3549 sectors_skipped += (max_sector - sector_nr);
3551 sector_nr = max_sector;
3556 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3559 struct r10conf *conf = mddev->private;
3562 raid_disks = min(conf->geo.raid_disks,
3563 conf->prev.raid_disks);
3565 sectors = conf->dev_sectors;
3567 size = sectors >> conf->geo.chunk_shift;
3568 sector_div(size, conf->geo.far_copies);
3569 size = size * raid_disks;
3570 sector_div(size, conf->geo.near_copies);
3572 return size << conf->geo.chunk_shift;
3575 static void calc_sectors(struct r10conf *conf, sector_t size)
3577 /* Calculate the number of sectors-per-device that will
3578 * actually be used, and set conf->dev_sectors and
3582 size = size >> conf->geo.chunk_shift;
3583 sector_div(size, conf->geo.far_copies);
3584 size = size * conf->geo.raid_disks;
3585 sector_div(size, conf->geo.near_copies);
3586 /* 'size' is now the number of chunks in the array */
3587 /* calculate "used chunks per device" */
3588 size = size * conf->copies;
3590 /* We need to round up when dividing by raid_disks to
3591 * get the stride size.
3593 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3595 conf->dev_sectors = size << conf->geo.chunk_shift;
3597 if (conf->geo.far_offset)
3598 conf->geo.stride = 1 << conf->geo.chunk_shift;
3600 sector_div(size, conf->geo.far_copies);
3601 conf->geo.stride = size << conf->geo.chunk_shift;
3605 enum geo_type {geo_new, geo_old, geo_start};
3606 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3609 int layout, chunk, disks;
3612 layout = mddev->layout;
3613 chunk = mddev->chunk_sectors;
3614 disks = mddev->raid_disks - mddev->delta_disks;
3617 layout = mddev->new_layout;
3618 chunk = mddev->new_chunk_sectors;
3619 disks = mddev->raid_disks;
3621 default: /* avoid 'may be unused' warnings */
3622 case geo_start: /* new when starting reshape - raid_disks not
3624 layout = mddev->new_layout;
3625 chunk = mddev->new_chunk_sectors;
3626 disks = mddev->raid_disks + mddev->delta_disks;
3631 if (chunk < (PAGE_SIZE >> 9) ||
3632 !is_power_of_2(chunk))
3635 fc = (layout >> 8) & 255;
3636 fo = layout & (1<<16);
3637 geo->raid_disks = disks;
3638 geo->near_copies = nc;
3639 geo->far_copies = fc;
3640 geo->far_offset = fo;
3641 switch (layout >> 17) {
3642 case 0: /* original layout. simple but not always optimal */
3643 geo->far_set_size = disks;
3645 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3646 * actually using this, but leave code here just in case.*/
3647 geo->far_set_size = disks/fc;
3648 WARN(geo->far_set_size < fc,
3649 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3651 case 2: /* "improved" layout fixed to match documentation */
3652 geo->far_set_size = fc * nc;
3654 default: /* Not a valid layout */
3657 geo->chunk_mask = chunk - 1;
3658 geo->chunk_shift = ffz(~chunk);
3662 static struct r10conf *setup_conf(struct mddev *mddev)
3664 struct r10conf *conf = NULL;
3669 copies = setup_geo(&geo, mddev, geo_new);
3672 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3673 mdname(mddev), PAGE_SIZE);
3677 if (copies < 2 || copies > mddev->raid_disks) {
3678 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3679 mdname(mddev), mddev->new_layout);
3684 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3688 /* FIXME calc properly */
3689 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3690 max(0,-mddev->delta_disks)),
3695 conf->tmppage = alloc_page(GFP_KERNEL);
3700 conf->copies = copies;
3701 err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3702 r10bio_pool_free, conf);
3706 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3710 calc_sectors(conf, mddev->dev_sectors);
3711 if (mddev->reshape_position == MaxSector) {
3712 conf->prev = conf->geo;
3713 conf->reshape_progress = MaxSector;
3715 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3719 conf->reshape_progress = mddev->reshape_position;
3720 if (conf->prev.far_offset)
3721 conf->prev.stride = 1 << conf->prev.chunk_shift;
3723 /* far_copies must be 1 */
3724 conf->prev.stride = conf->dev_sectors;
3726 conf->reshape_safe = conf->reshape_progress;
3727 spin_lock_init(&conf->device_lock);
3728 INIT_LIST_HEAD(&conf->retry_list);
3729 INIT_LIST_HEAD(&conf->bio_end_io_list);
3731 spin_lock_init(&conf->resync_lock);
3732 init_waitqueue_head(&conf->wait_barrier);
3733 atomic_set(&conf->nr_pending, 0);
3736 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3740 conf->mddev = mddev;
3745 mempool_exit(&conf->r10bio_pool);
3746 kfree(conf->mirrors);
3747 safe_put_page(conf->tmppage);
3748 bioset_exit(&conf->bio_split);
3751 return ERR_PTR(err);
3754 static int raid10_run(struct mddev *mddev)
3756 struct r10conf *conf;
3757 int i, disk_idx, chunk_size;
3758 struct raid10_info *disk;
3759 struct md_rdev *rdev;
3761 sector_t min_offset_diff = 0;
3763 bool discard_supported = false;
3765 if (mddev_init_writes_pending(mddev) < 0)
3768 if (mddev->private == NULL) {
3769 conf = setup_conf(mddev);
3771 return PTR_ERR(conf);
3772 mddev->private = conf;
3774 conf = mddev->private;
3778 if (mddev_is_clustered(conf->mddev)) {
3781 fc = (mddev->layout >> 8) & 255;
3782 fo = mddev->layout & (1<<16);
3783 if (fc > 1 || fo > 0) {
3784 pr_err("only near layout is supported by clustered"
3790 mddev->thread = conf->thread;
3791 conf->thread = NULL;
3793 chunk_size = mddev->chunk_sectors << 9;
3795 blk_queue_max_discard_sectors(mddev->queue,
3796 mddev->chunk_sectors);
3797 blk_queue_max_write_same_sectors(mddev->queue, 0);
3798 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3799 blk_queue_io_min(mddev->queue, chunk_size);
3800 if (conf->geo.raid_disks % conf->geo.near_copies)
3801 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3803 blk_queue_io_opt(mddev->queue, chunk_size *
3804 (conf->geo.raid_disks / conf->geo.near_copies));
3807 rdev_for_each(rdev, mddev) {
3810 disk_idx = rdev->raid_disk;
3813 if (disk_idx >= conf->geo.raid_disks &&
3814 disk_idx >= conf->prev.raid_disks)
3816 disk = conf->mirrors + disk_idx;
3818 if (test_bit(Replacement, &rdev->flags)) {
3819 if (disk->replacement)
3821 disk->replacement = rdev;
3827 diff = (rdev->new_data_offset - rdev->data_offset);
3828 if (!mddev->reshape_backwards)
3832 if (first || diff < min_offset_diff)
3833 min_offset_diff = diff;
3836 disk_stack_limits(mddev->gendisk, rdev->bdev,
3837 rdev->data_offset << 9);
3839 disk->head_position = 0;
3841 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3842 discard_supported = true;
3847 if (discard_supported)
3848 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3851 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3854 /* need to check that every block has at least one working mirror */
3855 if (!enough(conf, -1)) {
3856 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3861 if (conf->reshape_progress != MaxSector) {
3862 /* must ensure that shape change is supported */
3863 if (conf->geo.far_copies != 1 &&
3864 conf->geo.far_offset == 0)
3866 if (conf->prev.far_copies != 1 &&
3867 conf->prev.far_offset == 0)
3871 mddev->degraded = 0;
3873 i < conf->geo.raid_disks
3874 || i < conf->prev.raid_disks;
3877 disk = conf->mirrors + i;
3879 if (!disk->rdev && disk->replacement) {
3880 /* The replacement is all we have - use it */
3881 disk->rdev = disk->replacement;
3882 disk->replacement = NULL;
3883 clear_bit(Replacement, &disk->rdev->flags);
3887 !test_bit(In_sync, &disk->rdev->flags)) {
3888 disk->head_position = 0;
3891 disk->rdev->saved_raid_disk < 0)
3894 disk->recovery_disabled = mddev->recovery_disabled - 1;
3897 if (mddev->recovery_cp != MaxSector)
3898 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3900 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3901 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3902 conf->geo.raid_disks);
3904 * Ok, everything is just fine now
3906 mddev->dev_sectors = conf->dev_sectors;
3907 size = raid10_size(mddev, 0, 0);
3908 md_set_array_sectors(mddev, size);
3909 mddev->resync_max_sectors = size;
3910 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3913 int stripe = conf->geo.raid_disks *
3914 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3916 /* Calculate max read-ahead size.
3917 * We need to readahead at least twice a whole stripe....
3920 stripe /= conf->geo.near_copies;
3921 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3922 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3925 if (md_integrity_register(mddev))
3928 if (conf->reshape_progress != MaxSector) {
3929 unsigned long before_length, after_length;
3931 before_length = ((1 << conf->prev.chunk_shift) *
3932 conf->prev.far_copies);
3933 after_length = ((1 << conf->geo.chunk_shift) *
3934 conf->geo.far_copies);
3936 if (max(before_length, after_length) > min_offset_diff) {
3937 /* This cannot work */
3938 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3941 conf->offset_diff = min_offset_diff;
3943 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3944 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3945 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3946 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3947 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3954 md_unregister_thread(&mddev->thread);
3955 mempool_exit(&conf->r10bio_pool);
3956 safe_put_page(conf->tmppage);
3957 kfree(conf->mirrors);
3959 mddev->private = NULL;
3964 static void raid10_free(struct mddev *mddev, void *priv)
3966 struct r10conf *conf = priv;
3968 mempool_exit(&conf->r10bio_pool);
3969 safe_put_page(conf->tmppage);
3970 kfree(conf->mirrors);
3971 kfree(conf->mirrors_old);
3972 kfree(conf->mirrors_new);
3973 bioset_exit(&conf->bio_split);
3977 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3979 struct r10conf *conf = mddev->private;
3982 raise_barrier(conf, 0);
3984 lower_barrier(conf);
3987 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3989 /* Resize of 'far' arrays is not supported.
3990 * For 'near' and 'offset' arrays we can set the
3991 * number of sectors used to be an appropriate multiple
3992 * of the chunk size.
3993 * For 'offset', this is far_copies*chunksize.
3994 * For 'near' the multiplier is the LCM of
3995 * near_copies and raid_disks.
3996 * So if far_copies > 1 && !far_offset, fail.
3997 * Else find LCM(raid_disks, near_copy)*far_copies and
3998 * multiply by chunk_size. Then round to this number.
3999 * This is mostly done by raid10_size()
4001 struct r10conf *conf = mddev->private;
4002 sector_t oldsize, size;
4004 if (mddev->reshape_position != MaxSector)
4007 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4010 oldsize = raid10_size(mddev, 0, 0);
4011 size = raid10_size(mddev, sectors, 0);
4012 if (mddev->external_size &&
4013 mddev->array_sectors > size)
4015 if (mddev->bitmap) {
4016 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
4020 md_set_array_sectors(mddev, size);
4021 if (sectors > mddev->dev_sectors &&
4022 mddev->recovery_cp > oldsize) {
4023 mddev->recovery_cp = oldsize;
4024 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4026 calc_sectors(conf, sectors);
4027 mddev->dev_sectors = conf->dev_sectors;
4028 mddev->resync_max_sectors = size;
4032 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4034 struct md_rdev *rdev;
4035 struct r10conf *conf;
4037 if (mddev->degraded > 0) {
4038 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4040 return ERR_PTR(-EINVAL);
4042 sector_div(size, devs);
4044 /* Set new parameters */
4045 mddev->new_level = 10;
4046 /* new layout: far_copies = 1, near_copies = 2 */
4047 mddev->new_layout = (1<<8) + 2;
4048 mddev->new_chunk_sectors = mddev->chunk_sectors;
4049 mddev->delta_disks = mddev->raid_disks;
4050 mddev->raid_disks *= 2;
4051 /* make sure it will be not marked as dirty */
4052 mddev->recovery_cp = MaxSector;
4053 mddev->dev_sectors = size;
4055 conf = setup_conf(mddev);
4056 if (!IS_ERR(conf)) {
4057 rdev_for_each(rdev, mddev)
4058 if (rdev->raid_disk >= 0) {
4059 rdev->new_raid_disk = rdev->raid_disk * 2;
4060 rdev->sectors = size;
4068 static void *raid10_takeover(struct mddev *mddev)
4070 struct r0conf *raid0_conf;
4072 /* raid10 can take over:
4073 * raid0 - providing it has only two drives
4075 if (mddev->level == 0) {
4076 /* for raid0 takeover only one zone is supported */
4077 raid0_conf = mddev->private;
4078 if (raid0_conf->nr_strip_zones > 1) {
4079 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4081 return ERR_PTR(-EINVAL);
4083 return raid10_takeover_raid0(mddev,
4084 raid0_conf->strip_zone->zone_end,
4085 raid0_conf->strip_zone->nb_dev);
4087 return ERR_PTR(-EINVAL);
4090 static int raid10_check_reshape(struct mddev *mddev)
4092 /* Called when there is a request to change
4093 * - layout (to ->new_layout)
4094 * - chunk size (to ->new_chunk_sectors)
4095 * - raid_disks (by delta_disks)
4096 * or when trying to restart a reshape that was ongoing.
4098 * We need to validate the request and possibly allocate
4099 * space if that might be an issue later.
4101 * Currently we reject any reshape of a 'far' mode array,
4102 * allow chunk size to change if new is generally acceptable,
4103 * allow raid_disks to increase, and allow
4104 * a switch between 'near' mode and 'offset' mode.
4106 struct r10conf *conf = mddev->private;
4109 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4112 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4113 /* mustn't change number of copies */
4115 if (geo.far_copies > 1 && !geo.far_offset)
4116 /* Cannot switch to 'far' mode */
4119 if (mddev->array_sectors & geo.chunk_mask)
4120 /* not factor of array size */
4123 if (!enough(conf, -1))
4126 kfree(conf->mirrors_new);
4127 conf->mirrors_new = NULL;
4128 if (mddev->delta_disks > 0) {
4129 /* allocate new 'mirrors' list */
4130 conf->mirrors_new = kzalloc(
4131 sizeof(struct raid10_info)
4132 *(mddev->raid_disks +
4133 mddev->delta_disks),
4135 if (!conf->mirrors_new)
4142 * Need to check if array has failed when deciding whether to:
4144 * - remove non-faulty devices
4147 * This determination is simple when no reshape is happening.
4148 * However if there is a reshape, we need to carefully check
4149 * both the before and after sections.
4150 * This is because some failed devices may only affect one
4151 * of the two sections, and some non-in_sync devices may
4152 * be insync in the section most affected by failed devices.
4154 static int calc_degraded(struct r10conf *conf)
4156 int degraded, degraded2;
4161 /* 'prev' section first */
4162 for (i = 0; i < conf->prev.raid_disks; i++) {
4163 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4164 if (!rdev || test_bit(Faulty, &rdev->flags))
4166 else if (!test_bit(In_sync, &rdev->flags))
4167 /* When we can reduce the number of devices in
4168 * an array, this might not contribute to
4169 * 'degraded'. It does now.
4174 if (conf->geo.raid_disks == conf->prev.raid_disks)
4178 for (i = 0; i < conf->geo.raid_disks; i++) {
4179 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4180 if (!rdev || test_bit(Faulty, &rdev->flags))
4182 else if (!test_bit(In_sync, &rdev->flags)) {
4183 /* If reshape is increasing the number of devices,
4184 * this section has already been recovered, so
4185 * it doesn't contribute to degraded.
4188 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4193 if (degraded2 > degraded)
4198 static int raid10_start_reshape(struct mddev *mddev)
4200 /* A 'reshape' has been requested. This commits
4201 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4202 * This also checks if there are enough spares and adds them
4204 * We currently require enough spares to make the final
4205 * array non-degraded. We also require that the difference
4206 * between old and new data_offset - on each device - is
4207 * enough that we never risk over-writing.
4210 unsigned long before_length, after_length;
4211 sector_t min_offset_diff = 0;
4214 struct r10conf *conf = mddev->private;
4215 struct md_rdev *rdev;
4219 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4222 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4225 before_length = ((1 << conf->prev.chunk_shift) *
4226 conf->prev.far_copies);
4227 after_length = ((1 << conf->geo.chunk_shift) *
4228 conf->geo.far_copies);
4230 rdev_for_each(rdev, mddev) {
4231 if (!test_bit(In_sync, &rdev->flags)
4232 && !test_bit(Faulty, &rdev->flags))
4234 if (rdev->raid_disk >= 0) {
4235 long long diff = (rdev->new_data_offset
4236 - rdev->data_offset);
4237 if (!mddev->reshape_backwards)
4241 if (first || diff < min_offset_diff)
4242 min_offset_diff = diff;
4247 if (max(before_length, after_length) > min_offset_diff)
4250 if (spares < mddev->delta_disks)
4253 conf->offset_diff = min_offset_diff;
4254 spin_lock_irq(&conf->device_lock);
4255 if (conf->mirrors_new) {
4256 memcpy(conf->mirrors_new, conf->mirrors,
4257 sizeof(struct raid10_info)*conf->prev.raid_disks);
4259 kfree(conf->mirrors_old);
4260 conf->mirrors_old = conf->mirrors;
4261 conf->mirrors = conf->mirrors_new;
4262 conf->mirrors_new = NULL;
4264 setup_geo(&conf->geo, mddev, geo_start);
4266 if (mddev->reshape_backwards) {
4267 sector_t size = raid10_size(mddev, 0, 0);
4268 if (size < mddev->array_sectors) {
4269 spin_unlock_irq(&conf->device_lock);
4270 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4274 mddev->resync_max_sectors = size;
4275 conf->reshape_progress = size;
4277 conf->reshape_progress = 0;
4278 conf->reshape_safe = conf->reshape_progress;
4279 spin_unlock_irq(&conf->device_lock);
4281 if (mddev->delta_disks && mddev->bitmap) {
4282 ret = bitmap_resize(mddev->bitmap,
4283 raid10_size(mddev, 0,
4284 conf->geo.raid_disks),
4289 if (mddev->delta_disks > 0) {
4290 rdev_for_each(rdev, mddev)
4291 if (rdev->raid_disk < 0 &&
4292 !test_bit(Faulty, &rdev->flags)) {
4293 if (raid10_add_disk(mddev, rdev) == 0) {
4294 if (rdev->raid_disk >=
4295 conf->prev.raid_disks)
4296 set_bit(In_sync, &rdev->flags);
4298 rdev->recovery_offset = 0;
4300 if (sysfs_link_rdev(mddev, rdev))
4301 /* Failure here is OK */;
4303 } else if (rdev->raid_disk >= conf->prev.raid_disks
4304 && !test_bit(Faulty, &rdev->flags)) {
4305 /* This is a spare that was manually added */
4306 set_bit(In_sync, &rdev->flags);
4309 /* When a reshape changes the number of devices,
4310 * ->degraded is measured against the larger of the
4311 * pre and post numbers.
4313 spin_lock_irq(&conf->device_lock);
4314 mddev->degraded = calc_degraded(conf);
4315 spin_unlock_irq(&conf->device_lock);
4316 mddev->raid_disks = conf->geo.raid_disks;
4317 mddev->reshape_position = conf->reshape_progress;
4318 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4320 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4321 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4322 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4323 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4324 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4326 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4328 if (!mddev->sync_thread) {
4332 conf->reshape_checkpoint = jiffies;
4333 md_wakeup_thread(mddev->sync_thread);
4334 md_new_event(mddev);
4338 mddev->recovery = 0;
4339 spin_lock_irq(&conf->device_lock);
4340 conf->geo = conf->prev;
4341 mddev->raid_disks = conf->geo.raid_disks;
4342 rdev_for_each(rdev, mddev)
4343 rdev->new_data_offset = rdev->data_offset;
4345 conf->reshape_progress = MaxSector;
4346 conf->reshape_safe = MaxSector;
4347 mddev->reshape_position = MaxSector;
4348 spin_unlock_irq(&conf->device_lock);
4352 /* Calculate the last device-address that could contain
4353 * any block from the chunk that includes the array-address 's'
4354 * and report the next address.
4355 * i.e. the address returned will be chunk-aligned and after
4356 * any data that is in the chunk containing 's'.
4358 static sector_t last_dev_address(sector_t s, struct geom *geo)
4360 s = (s | geo->chunk_mask) + 1;
4361 s >>= geo->chunk_shift;
4362 s *= geo->near_copies;
4363 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4364 s *= geo->far_copies;
4365 s <<= geo->chunk_shift;
4369 /* Calculate the first device-address that could contain
4370 * any block from the chunk that includes the array-address 's'.
4371 * This too will be the start of a chunk
4373 static sector_t first_dev_address(sector_t s, struct geom *geo)
4375 s >>= geo->chunk_shift;
4376 s *= geo->near_copies;
4377 sector_div(s, geo->raid_disks);
4378 s *= geo->far_copies;
4379 s <<= geo->chunk_shift;
4383 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4386 /* We simply copy at most one chunk (smallest of old and new)
4387 * at a time, possibly less if that exceeds RESYNC_PAGES,
4388 * or we hit a bad block or something.
4389 * This might mean we pause for normal IO in the middle of
4390 * a chunk, but that is not a problem as mddev->reshape_position
4391 * can record any location.
4393 * If we will want to write to a location that isn't
4394 * yet recorded as 'safe' (i.e. in metadata on disk) then
4395 * we need to flush all reshape requests and update the metadata.
4397 * When reshaping forwards (e.g. to more devices), we interpret
4398 * 'safe' as the earliest block which might not have been copied
4399 * down yet. We divide this by previous stripe size and multiply
4400 * by previous stripe length to get lowest device offset that we
4401 * cannot write to yet.
4402 * We interpret 'sector_nr' as an address that we want to write to.
4403 * From this we use last_device_address() to find where we might
4404 * write to, and first_device_address on the 'safe' position.
4405 * If this 'next' write position is after the 'safe' position,
4406 * we must update the metadata to increase the 'safe' position.
4408 * When reshaping backwards, we round in the opposite direction
4409 * and perform the reverse test: next write position must not be
4410 * less than current safe position.
4412 * In all this the minimum difference in data offsets
4413 * (conf->offset_diff - always positive) allows a bit of slack,
4414 * so next can be after 'safe', but not by more than offset_diff
4416 * We need to prepare all the bios here before we start any IO
4417 * to ensure the size we choose is acceptable to all devices.
4418 * The means one for each copy for write-out and an extra one for
4420 * We store the read-in bio in ->master_bio and the others in
4421 * ->devs[x].bio and ->devs[x].repl_bio.
4423 struct r10conf *conf = mddev->private;
4424 struct r10bio *r10_bio;
4425 sector_t next, safe, last;
4429 struct md_rdev *rdev;
4432 struct bio *bio, *read_bio;
4433 int sectors_done = 0;
4434 struct page **pages;
4436 if (sector_nr == 0) {
4437 /* If restarting in the middle, skip the initial sectors */
4438 if (mddev->reshape_backwards &&
4439 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4440 sector_nr = (raid10_size(mddev, 0, 0)
4441 - conf->reshape_progress);
4442 } else if (!mddev->reshape_backwards &&
4443 conf->reshape_progress > 0)
4444 sector_nr = conf->reshape_progress;
4446 mddev->curr_resync_completed = sector_nr;
4447 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4453 /* We don't use sector_nr to track where we are up to
4454 * as that doesn't work well for ->reshape_backwards.
4455 * So just use ->reshape_progress.
4457 if (mddev->reshape_backwards) {
4458 /* 'next' is the earliest device address that we might
4459 * write to for this chunk in the new layout
4461 next = first_dev_address(conf->reshape_progress - 1,
4464 /* 'safe' is the last device address that we might read from
4465 * in the old layout after a restart
4467 safe = last_dev_address(conf->reshape_safe - 1,
4470 if (next + conf->offset_diff < safe)
4473 last = conf->reshape_progress - 1;
4474 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4475 & conf->prev.chunk_mask);
4476 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4477 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4479 /* 'next' is after the last device address that we
4480 * might write to for this chunk in the new layout
4482 next = last_dev_address(conf->reshape_progress, &conf->geo);
4484 /* 'safe' is the earliest device address that we might
4485 * read from in the old layout after a restart
4487 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4489 /* Need to update metadata if 'next' might be beyond 'safe'
4490 * as that would possibly corrupt data
4492 if (next > safe + conf->offset_diff)
4495 sector_nr = conf->reshape_progress;
4496 last = sector_nr | (conf->geo.chunk_mask
4497 & conf->prev.chunk_mask);
4499 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4500 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4504 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4505 /* Need to update reshape_position in metadata */
4507 mddev->reshape_position = conf->reshape_progress;
4508 if (mddev->reshape_backwards)
4509 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4510 - conf->reshape_progress;
4512 mddev->curr_resync_completed = conf->reshape_progress;
4513 conf->reshape_checkpoint = jiffies;
4514 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4515 md_wakeup_thread(mddev->thread);
4516 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4517 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4518 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4519 allow_barrier(conf);
4520 return sectors_done;
4522 conf->reshape_safe = mddev->reshape_position;
4523 allow_barrier(conf);
4527 /* Now schedule reads for blocks from sector_nr to last */
4528 r10_bio = raid10_alloc_init_r10buf(conf);
4530 raise_barrier(conf, sectors_done != 0);
4531 atomic_set(&r10_bio->remaining, 0);
4532 r10_bio->mddev = mddev;
4533 r10_bio->sector = sector_nr;
4534 set_bit(R10BIO_IsReshape, &r10_bio->state);
4535 r10_bio->sectors = last - sector_nr + 1;
4536 rdev = read_balance(conf, r10_bio, &max_sectors);
4537 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4540 /* Cannot read from here, so need to record bad blocks
4541 * on all the target devices.
4544 mempool_free(r10_bio, &conf->r10buf_pool);
4545 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4546 return sectors_done;
4549 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4551 bio_set_dev(read_bio, rdev->bdev);
4552 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4553 + rdev->data_offset);
4554 read_bio->bi_private = r10_bio;
4555 read_bio->bi_end_io = end_reshape_read;
4556 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4557 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4558 read_bio->bi_status = 0;
4559 read_bio->bi_vcnt = 0;
4560 read_bio->bi_iter.bi_size = 0;
4561 r10_bio->master_bio = read_bio;
4562 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4564 /* Now find the locations in the new layout */
4565 __raid10_find_phys(&conf->geo, r10_bio);
4568 read_bio->bi_next = NULL;
4571 for (s = 0; s < conf->copies*2; s++) {
4573 int d = r10_bio->devs[s/2].devnum;
4574 struct md_rdev *rdev2;
4576 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4577 b = r10_bio->devs[s/2].repl_bio;
4579 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4580 b = r10_bio->devs[s/2].bio;
4582 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4585 bio_set_dev(b, rdev2->bdev);
4586 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4587 rdev2->new_data_offset;
4588 b->bi_end_io = end_reshape_write;
4589 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4594 /* Now add as many pages as possible to all of these bios. */
4597 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4598 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4599 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4600 int len = (max_sectors - s) << 9;
4601 if (len > PAGE_SIZE)
4603 for (bio = blist; bio ; bio = bio->bi_next) {
4605 * won't fail because the vec table is big enough
4606 * to hold all these pages
4608 bio_add_page(bio, page, len, 0);
4610 sector_nr += len >> 9;
4611 nr_sectors += len >> 9;
4614 r10_bio->sectors = nr_sectors;
4616 /* Now submit the read */
4617 md_sync_acct_bio(read_bio, r10_bio->sectors);
4618 atomic_inc(&r10_bio->remaining);
4619 read_bio->bi_next = NULL;
4620 generic_make_request(read_bio);
4621 sector_nr += nr_sectors;
4622 sectors_done += nr_sectors;
4623 if (sector_nr <= last)
4626 /* Now that we have done the whole section we can
4627 * update reshape_progress
4629 if (mddev->reshape_backwards)
4630 conf->reshape_progress -= sectors_done;
4632 conf->reshape_progress += sectors_done;
4634 return sectors_done;
4637 static void end_reshape_request(struct r10bio *r10_bio);
4638 static int handle_reshape_read_error(struct mddev *mddev,
4639 struct r10bio *r10_bio);
4640 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4642 /* Reshape read completed. Hopefully we have a block
4644 * If we got a read error then we do sync 1-page reads from
4645 * elsewhere until we find the data - or give up.
4647 struct r10conf *conf = mddev->private;
4650 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4651 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4652 /* Reshape has been aborted */
4653 md_done_sync(mddev, r10_bio->sectors, 0);
4657 /* We definitely have the data in the pages, schedule the
4660 atomic_set(&r10_bio->remaining, 1);
4661 for (s = 0; s < conf->copies*2; s++) {
4663 int d = r10_bio->devs[s/2].devnum;
4664 struct md_rdev *rdev;
4667 rdev = rcu_dereference(conf->mirrors[d].replacement);
4668 b = r10_bio->devs[s/2].repl_bio;
4670 rdev = rcu_dereference(conf->mirrors[d].rdev);
4671 b = r10_bio->devs[s/2].bio;
4673 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4677 atomic_inc(&rdev->nr_pending);
4679 md_sync_acct_bio(b, r10_bio->sectors);
4680 atomic_inc(&r10_bio->remaining);
4682 generic_make_request(b);
4684 end_reshape_request(r10_bio);
4687 static void end_reshape(struct r10conf *conf)
4689 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4692 spin_lock_irq(&conf->device_lock);
4693 conf->prev = conf->geo;
4694 md_finish_reshape(conf->mddev);
4696 conf->reshape_progress = MaxSector;
4697 conf->reshape_safe = MaxSector;
4698 spin_unlock_irq(&conf->device_lock);
4700 /* read-ahead size must cover two whole stripes, which is
4701 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4703 if (conf->mddev->queue) {
4704 int stripe = conf->geo.raid_disks *
4705 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4706 stripe /= conf->geo.near_copies;
4707 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4708 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4713 static int handle_reshape_read_error(struct mddev *mddev,
4714 struct r10bio *r10_bio)
4716 /* Use sync reads to get the blocks from somewhere else */
4717 int sectors = r10_bio->sectors;
4718 struct r10conf *conf = mddev->private;
4719 struct r10bio *r10b;
4722 struct page **pages;
4724 r10b = kmalloc(sizeof(*r10b) +
4725 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4727 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4731 /* reshape IOs share pages from .devs[0].bio */
4732 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4734 r10b->sector = r10_bio->sector;
4735 __raid10_find_phys(&conf->prev, r10b);
4740 int first_slot = slot;
4742 if (s > (PAGE_SIZE >> 9))
4747 int d = r10b->devs[slot].devnum;
4748 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4751 test_bit(Faulty, &rdev->flags) ||
4752 !test_bit(In_sync, &rdev->flags))
4755 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4756 atomic_inc(&rdev->nr_pending);
4758 success = sync_page_io(rdev,
4762 REQ_OP_READ, 0, false);
4763 rdev_dec_pending(rdev, mddev);
4769 if (slot >= conf->copies)
4771 if (slot == first_slot)
4776 /* couldn't read this block, must give up */
4777 set_bit(MD_RECOVERY_INTR,
4789 static void end_reshape_write(struct bio *bio)
4791 struct r10bio *r10_bio = get_resync_r10bio(bio);
4792 struct mddev *mddev = r10_bio->mddev;
4793 struct r10conf *conf = mddev->private;
4797 struct md_rdev *rdev = NULL;
4799 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4801 rdev = conf->mirrors[d].replacement;
4804 rdev = conf->mirrors[d].rdev;
4807 if (bio->bi_status) {
4808 /* FIXME should record badblock */
4809 md_error(mddev, rdev);
4812 rdev_dec_pending(rdev, mddev);
4813 end_reshape_request(r10_bio);
4816 static void end_reshape_request(struct r10bio *r10_bio)
4818 if (!atomic_dec_and_test(&r10_bio->remaining))
4820 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4821 bio_put(r10_bio->master_bio);
4825 static void raid10_finish_reshape(struct mddev *mddev)
4827 struct r10conf *conf = mddev->private;
4829 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4832 if (mddev->delta_disks > 0) {
4833 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4834 mddev->recovery_cp = mddev->resync_max_sectors;
4835 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4837 mddev->resync_max_sectors = mddev->array_sectors;
4841 for (d = conf->geo.raid_disks ;
4842 d < conf->geo.raid_disks - mddev->delta_disks;
4844 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4846 clear_bit(In_sync, &rdev->flags);
4847 rdev = rcu_dereference(conf->mirrors[d].replacement);
4849 clear_bit(In_sync, &rdev->flags);
4853 mddev->layout = mddev->new_layout;
4854 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4855 mddev->reshape_position = MaxSector;
4856 mddev->delta_disks = 0;
4857 mddev->reshape_backwards = 0;
4860 static struct md_personality raid10_personality =
4864 .owner = THIS_MODULE,
4865 .make_request = raid10_make_request,
4867 .free = raid10_free,
4868 .status = raid10_status,
4869 .error_handler = raid10_error,
4870 .hot_add_disk = raid10_add_disk,
4871 .hot_remove_disk= raid10_remove_disk,
4872 .spare_active = raid10_spare_active,
4873 .sync_request = raid10_sync_request,
4874 .quiesce = raid10_quiesce,
4875 .size = raid10_size,
4876 .resize = raid10_resize,
4877 .takeover = raid10_takeover,
4878 .check_reshape = raid10_check_reshape,
4879 .start_reshape = raid10_start_reshape,
4880 .finish_reshape = raid10_finish_reshape,
4881 .congested = raid10_congested,
4884 static int __init raid_init(void)
4886 return register_md_personality(&raid10_personality);
4889 static void raid_exit(void)
4891 unregister_md_personality(&raid10_personality);
4894 module_init(raid_init);
4895 module_exit(raid_exit);
4896 MODULE_LICENSE("GPL");
4897 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4898 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4899 MODULE_ALIAS("md-raid10");
4900 MODULE_ALIAS("md-level-10");
4902 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);