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_array(nalloc_rp, sizeof(struct resync_pages), 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;
259 rp = get_resync_pages(bio);
260 resync_free_pages(rp);
264 bio = r10bio->devs[j].repl_bio;
269 /* resync pages array stored in the 1st bio's .bi_private */
272 r10bio_pool_free(r10bio, conf);
275 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
279 for (i = 0; i < conf->copies; i++) {
280 struct bio **bio = & r10_bio->devs[i].bio;
281 if (!BIO_SPECIAL(*bio))
284 bio = &r10_bio->devs[i].repl_bio;
285 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
291 static void free_r10bio(struct r10bio *r10_bio)
293 struct r10conf *conf = r10_bio->mddev->private;
295 put_all_bios(conf, r10_bio);
296 mempool_free(r10_bio, &conf->r10bio_pool);
299 static void put_buf(struct r10bio *r10_bio)
301 struct r10conf *conf = r10_bio->mddev->private;
303 mempool_free(r10_bio, &conf->r10buf_pool);
308 static void reschedule_retry(struct r10bio *r10_bio)
311 struct mddev *mddev = r10_bio->mddev;
312 struct r10conf *conf = mddev->private;
314 spin_lock_irqsave(&conf->device_lock, flags);
315 list_add(&r10_bio->retry_list, &conf->retry_list);
317 spin_unlock_irqrestore(&conf->device_lock, flags);
319 /* wake up frozen array... */
320 wake_up(&conf->wait_barrier);
322 md_wakeup_thread(mddev->thread);
326 * raid_end_bio_io() is called when we have finished servicing a mirrored
327 * operation and are ready to return a success/failure code to the buffer
330 static void raid_end_bio_io(struct r10bio *r10_bio)
332 struct bio *bio = r10_bio->master_bio;
333 struct r10conf *conf = r10_bio->mddev->private;
335 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
336 bio->bi_status = BLK_STS_IOERR;
340 * Wake up any possible resync thread that waits for the device
345 free_r10bio(r10_bio);
349 * Update disk head position estimator based on IRQ completion info.
351 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
353 struct r10conf *conf = r10_bio->mddev->private;
355 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
356 r10_bio->devs[slot].addr + (r10_bio->sectors);
360 * Find the disk number which triggered given bio
362 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
363 struct bio *bio, int *slotp, int *replp)
368 for (slot = 0; slot < conf->copies; slot++) {
369 if (r10_bio->devs[slot].bio == bio)
371 if (r10_bio->devs[slot].repl_bio == bio) {
377 BUG_ON(slot == conf->copies);
378 update_head_pos(slot, r10_bio);
384 return r10_bio->devs[slot].devnum;
387 static void raid10_end_read_request(struct bio *bio)
389 int uptodate = !bio->bi_status;
390 struct r10bio *r10_bio = bio->bi_private;
392 struct md_rdev *rdev;
393 struct r10conf *conf = r10_bio->mddev->private;
395 slot = r10_bio->read_slot;
396 rdev = r10_bio->devs[slot].rdev;
398 * this branch is our 'one mirror IO has finished' event handler:
400 update_head_pos(slot, r10_bio);
404 * Set R10BIO_Uptodate in our master bio, so that
405 * we will return a good error code to the higher
406 * levels even if IO on some other mirrored buffer fails.
408 * The 'master' represents the composite IO operation to
409 * user-side. So if something waits for IO, then it will
410 * wait for the 'master' bio.
412 set_bit(R10BIO_Uptodate, &r10_bio->state);
414 /* If all other devices that store this block have
415 * failed, we want to return the error upwards rather
416 * than fail the last device. Here we redefine
417 * "uptodate" to mean "Don't want to retry"
419 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
424 raid_end_bio_io(r10_bio);
425 rdev_dec_pending(rdev, conf->mddev);
428 * oops, read error - keep the refcount on the rdev
430 char b[BDEVNAME_SIZE];
431 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
433 bdevname(rdev->bdev, b),
434 (unsigned long long)r10_bio->sector);
435 set_bit(R10BIO_ReadError, &r10_bio->state);
436 reschedule_retry(r10_bio);
440 static void close_write(struct r10bio *r10_bio)
442 /* clear the bitmap if all writes complete successfully */
443 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
445 !test_bit(R10BIO_Degraded, &r10_bio->state),
447 md_write_end(r10_bio->mddev);
450 static void one_write_done(struct r10bio *r10_bio)
452 if (atomic_dec_and_test(&r10_bio->remaining)) {
453 if (test_bit(R10BIO_WriteError, &r10_bio->state))
454 reschedule_retry(r10_bio);
456 close_write(r10_bio);
457 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
458 reschedule_retry(r10_bio);
460 raid_end_bio_io(r10_bio);
465 static void raid10_end_write_request(struct bio *bio)
467 struct r10bio *r10_bio = bio->bi_private;
470 struct r10conf *conf = r10_bio->mddev->private;
472 struct md_rdev *rdev = NULL;
473 struct bio *to_put = NULL;
476 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
478 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
481 rdev = conf->mirrors[dev].replacement;
485 rdev = conf->mirrors[dev].rdev;
488 * this branch is our 'one mirror IO has finished' event handler:
490 if (bio->bi_status && !discard_error) {
492 /* Never record new bad blocks to replacement,
495 md_error(rdev->mddev, rdev);
497 set_bit(WriteErrorSeen, &rdev->flags);
498 if (!test_and_set_bit(WantReplacement, &rdev->flags))
499 set_bit(MD_RECOVERY_NEEDED,
500 &rdev->mddev->recovery);
503 if (test_bit(FailFast, &rdev->flags) &&
504 (bio->bi_opf & MD_FAILFAST)) {
505 md_error(rdev->mddev, rdev);
506 if (!test_bit(Faulty, &rdev->flags))
507 /* This is the only remaining device,
508 * We need to retry the write without
511 set_bit(R10BIO_WriteError, &r10_bio->state);
513 r10_bio->devs[slot].bio = NULL;
518 set_bit(R10BIO_WriteError, &r10_bio->state);
522 * Set R10BIO_Uptodate in our master bio, so that
523 * we will return a good error code for to the higher
524 * levels even if IO on some other mirrored buffer fails.
526 * The 'master' represents the composite IO operation to
527 * user-side. So if something waits for IO, then it will
528 * wait for the 'master' bio.
534 * Do not set R10BIO_Uptodate if the current device is
535 * rebuilding or Faulty. This is because we cannot use
536 * such device for properly reading the data back (we could
537 * potentially use it, if the current write would have felt
538 * before rdev->recovery_offset, but for simplicity we don't
541 if (test_bit(In_sync, &rdev->flags) &&
542 !test_bit(Faulty, &rdev->flags))
543 set_bit(R10BIO_Uptodate, &r10_bio->state);
545 /* Maybe we can clear some bad blocks. */
546 if (is_badblock(rdev,
547 r10_bio->devs[slot].addr,
549 &first_bad, &bad_sectors) && !discard_error) {
552 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
554 r10_bio->devs[slot].bio = IO_MADE_GOOD;
556 set_bit(R10BIO_MadeGood, &r10_bio->state);
562 * Let's see if all mirrored write operations have finished
565 one_write_done(r10_bio);
567 rdev_dec_pending(rdev, conf->mddev);
573 * RAID10 layout manager
574 * As well as the chunksize and raid_disks count, there are two
575 * parameters: near_copies and far_copies.
576 * near_copies * far_copies must be <= raid_disks.
577 * Normally one of these will be 1.
578 * If both are 1, we get raid0.
579 * If near_copies == raid_disks, we get raid1.
581 * Chunks are laid out in raid0 style with near_copies copies of the
582 * first chunk, followed by near_copies copies of the next chunk and
584 * If far_copies > 1, then after 1/far_copies of the array has been assigned
585 * as described above, we start again with a device offset of near_copies.
586 * So we effectively have another copy of the whole array further down all
587 * the drives, but with blocks on different drives.
588 * With this layout, and block is never stored twice on the one device.
590 * raid10_find_phys finds the sector offset of a given virtual sector
591 * on each device that it is on.
593 * raid10_find_virt does the reverse mapping, from a device and a
594 * sector offset to a virtual address
597 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
605 int last_far_set_start, last_far_set_size;
607 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
608 last_far_set_start *= geo->far_set_size;
610 last_far_set_size = geo->far_set_size;
611 last_far_set_size += (geo->raid_disks % geo->far_set_size);
613 /* now calculate first sector/dev */
614 chunk = r10bio->sector >> geo->chunk_shift;
615 sector = r10bio->sector & geo->chunk_mask;
617 chunk *= geo->near_copies;
619 dev = sector_div(stripe, geo->raid_disks);
621 stripe *= geo->far_copies;
623 sector += stripe << geo->chunk_shift;
625 /* and calculate all the others */
626 for (n = 0; n < geo->near_copies; n++) {
630 r10bio->devs[slot].devnum = d;
631 r10bio->devs[slot].addr = s;
634 for (f = 1; f < geo->far_copies; f++) {
635 set = d / geo->far_set_size;
636 d += geo->near_copies;
638 if ((geo->raid_disks % geo->far_set_size) &&
639 (d > last_far_set_start)) {
640 d -= last_far_set_start;
641 d %= last_far_set_size;
642 d += last_far_set_start;
644 d %= geo->far_set_size;
645 d += geo->far_set_size * set;
648 r10bio->devs[slot].devnum = d;
649 r10bio->devs[slot].addr = s;
653 if (dev >= geo->raid_disks) {
655 sector += (geo->chunk_mask + 1);
660 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
662 struct geom *geo = &conf->geo;
664 if (conf->reshape_progress != MaxSector &&
665 ((r10bio->sector >= conf->reshape_progress) !=
666 conf->mddev->reshape_backwards)) {
667 set_bit(R10BIO_Previous, &r10bio->state);
670 clear_bit(R10BIO_Previous, &r10bio->state);
672 __raid10_find_phys(geo, r10bio);
675 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
677 sector_t offset, chunk, vchunk;
678 /* Never use conf->prev as this is only called during resync
679 * or recovery, so reshape isn't happening
681 struct geom *geo = &conf->geo;
682 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
683 int far_set_size = geo->far_set_size;
684 int last_far_set_start;
686 if (geo->raid_disks % geo->far_set_size) {
687 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
688 last_far_set_start *= geo->far_set_size;
690 if (dev >= last_far_set_start) {
691 far_set_size = geo->far_set_size;
692 far_set_size += (geo->raid_disks % geo->far_set_size);
693 far_set_start = last_far_set_start;
697 offset = sector & geo->chunk_mask;
698 if (geo->far_offset) {
700 chunk = sector >> geo->chunk_shift;
701 fc = sector_div(chunk, geo->far_copies);
702 dev -= fc * geo->near_copies;
703 if (dev < far_set_start)
706 while (sector >= geo->stride) {
707 sector -= geo->stride;
708 if (dev < (geo->near_copies + far_set_start))
709 dev += far_set_size - geo->near_copies;
711 dev -= geo->near_copies;
713 chunk = sector >> geo->chunk_shift;
715 vchunk = chunk * geo->raid_disks + dev;
716 sector_div(vchunk, geo->near_copies);
717 return (vchunk << geo->chunk_shift) + offset;
721 * This routine returns the disk from which the requested read should
722 * be done. There is a per-array 'next expected sequential IO' sector
723 * number - if this matches on the next IO then we use the last disk.
724 * There is also a per-disk 'last know head position' sector that is
725 * maintained from IRQ contexts, both the normal and the resync IO
726 * completion handlers update this position correctly. If there is no
727 * perfect sequential match then we pick the disk whose head is closest.
729 * If there are 2 mirrors in the same 2 devices, performance degrades
730 * because position is mirror, not device based.
732 * The rdev for the device selected will have nr_pending incremented.
736 * FIXME: possibly should rethink readbalancing and do it differently
737 * depending on near_copies / far_copies geometry.
739 static struct md_rdev *read_balance(struct r10conf *conf,
740 struct r10bio *r10_bio,
743 const sector_t this_sector = r10_bio->sector;
745 int sectors = r10_bio->sectors;
746 int best_good_sectors;
747 sector_t new_distance, best_dist;
748 struct md_rdev *best_rdev, *rdev = NULL;
751 struct geom *geo = &conf->geo;
753 raid10_find_phys(conf, r10_bio);
757 best_dist = MaxSector;
758 best_good_sectors = 0;
760 clear_bit(R10BIO_FailFast, &r10_bio->state);
762 * Check if we can balance. We can balance on the whole
763 * device if no resync is going on (recovery is ok), or below
764 * the resync window. We take the first readable disk when
765 * above the resync window.
767 if ((conf->mddev->recovery_cp < MaxSector
768 && (this_sector + sectors >= conf->next_resync)) ||
769 (mddev_is_clustered(conf->mddev) &&
770 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
771 this_sector + sectors)))
774 for (slot = 0; slot < conf->copies ; slot++) {
779 if (r10_bio->devs[slot].bio == IO_BLOCKED)
781 disk = r10_bio->devs[slot].devnum;
782 rdev = rcu_dereference(conf->mirrors[disk].replacement);
783 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
784 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
785 rdev = rcu_dereference(conf->mirrors[disk].rdev);
787 test_bit(Faulty, &rdev->flags))
789 if (!test_bit(In_sync, &rdev->flags) &&
790 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
793 dev_sector = r10_bio->devs[slot].addr;
794 if (is_badblock(rdev, dev_sector, sectors,
795 &first_bad, &bad_sectors)) {
796 if (best_dist < MaxSector)
797 /* Already have a better slot */
799 if (first_bad <= dev_sector) {
800 /* Cannot read here. If this is the
801 * 'primary' device, then we must not read
802 * beyond 'bad_sectors' from another device.
804 bad_sectors -= (dev_sector - first_bad);
805 if (!do_balance && sectors > bad_sectors)
806 sectors = bad_sectors;
807 if (best_good_sectors > sectors)
808 best_good_sectors = sectors;
810 sector_t good_sectors =
811 first_bad - dev_sector;
812 if (good_sectors > best_good_sectors) {
813 best_good_sectors = good_sectors;
818 /* Must read from here */
823 best_good_sectors = sectors;
829 /* At least 2 disks to choose from so failfast is OK */
830 set_bit(R10BIO_FailFast, &r10_bio->state);
831 /* This optimisation is debatable, and completely destroys
832 * sequential read speed for 'far copies' arrays. So only
833 * keep it for 'near' arrays, and review those later.
835 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
838 /* for far > 1 always use the lowest address */
839 else if (geo->far_copies > 1)
840 new_distance = r10_bio->devs[slot].addr;
842 new_distance = abs(r10_bio->devs[slot].addr -
843 conf->mirrors[disk].head_position);
844 if (new_distance < best_dist) {
845 best_dist = new_distance;
850 if (slot >= conf->copies) {
856 atomic_inc(&rdev->nr_pending);
857 r10_bio->read_slot = slot;
861 *max_sectors = best_good_sectors;
866 static int raid10_congested(struct mddev *mddev, int bits)
868 struct r10conf *conf = mddev->private;
871 if ((bits & (1 << WB_async_congested)) &&
872 conf->pending_count >= max_queued_requests)
877 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
880 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
881 if (rdev && !test_bit(Faulty, &rdev->flags)) {
882 struct request_queue *q = bdev_get_queue(rdev->bdev);
884 ret |= bdi_congested(q->backing_dev_info, bits);
891 static void flush_pending_writes(struct r10conf *conf)
893 /* Any writes that have been queued but are awaiting
894 * bitmap updates get flushed here.
896 spin_lock_irq(&conf->device_lock);
898 if (conf->pending_bio_list.head) {
899 struct blk_plug plug;
902 bio = bio_list_get(&conf->pending_bio_list);
903 conf->pending_count = 0;
904 spin_unlock_irq(&conf->device_lock);
907 * As this is called in a wait_event() loop (see freeze_array),
908 * current->state might be TASK_UNINTERRUPTIBLE which will
909 * cause a warning when we prepare to wait again. As it is
910 * rare that this path is taken, it is perfectly safe to force
911 * us to go around the wait_event() loop again, so the warning
912 * is a false-positive. Silence the warning by resetting
915 __set_current_state(TASK_RUNNING);
917 blk_start_plug(&plug);
918 /* flush any pending bitmap writes to disk
919 * before proceeding w/ I/O */
920 bitmap_unplug(conf->mddev->bitmap);
921 wake_up(&conf->wait_barrier);
923 while (bio) { /* submit pending writes */
924 struct bio *next = bio->bi_next;
925 struct md_rdev *rdev = (void*)bio->bi_disk;
927 bio_set_dev(bio, rdev->bdev);
928 if (test_bit(Faulty, &rdev->flags)) {
930 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
931 !blk_queue_discard(bio->bi_disk->queue)))
935 generic_make_request(bio);
938 blk_finish_plug(&plug);
940 spin_unlock_irq(&conf->device_lock);
944 * Sometimes we need to suspend IO while we do something else,
945 * either some resync/recovery, or reconfigure the array.
946 * To do this we raise a 'barrier'.
947 * The 'barrier' is a counter that can be raised multiple times
948 * to count how many activities are happening which preclude
950 * We can only raise the barrier if there is no pending IO.
951 * i.e. if nr_pending == 0.
952 * We choose only to raise the barrier if no-one is waiting for the
953 * barrier to go down. This means that as soon as an IO request
954 * is ready, no other operations which require a barrier will start
955 * until the IO request has had a chance.
957 * So: regular IO calls 'wait_barrier'. When that returns there
958 * is no backgroup IO happening, It must arrange to call
959 * allow_barrier when it has finished its IO.
960 * backgroup IO calls must call raise_barrier. Once that returns
961 * there is no normal IO happeing. It must arrange to call
962 * lower_barrier when the particular background IO completes.
965 static void raise_barrier(struct r10conf *conf, int force)
967 BUG_ON(force && !conf->barrier);
968 spin_lock_irq(&conf->resync_lock);
970 /* Wait until no block IO is waiting (unless 'force') */
971 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
974 /* block any new IO from starting */
977 /* Now wait for all pending IO to complete */
978 wait_event_lock_irq(conf->wait_barrier,
979 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
982 spin_unlock_irq(&conf->resync_lock);
985 static void lower_barrier(struct r10conf *conf)
988 spin_lock_irqsave(&conf->resync_lock, flags);
990 spin_unlock_irqrestore(&conf->resync_lock, flags);
991 wake_up(&conf->wait_barrier);
994 static void wait_barrier(struct r10conf *conf)
996 spin_lock_irq(&conf->resync_lock);
999 /* Wait for the barrier to drop.
1000 * However if there are already pending
1001 * requests (preventing the barrier from
1002 * rising completely), and the
1003 * pre-process bio queue isn't empty,
1004 * then don't wait, as we need to empty
1005 * that queue to get the nr_pending
1008 raid10_log(conf->mddev, "wait barrier");
1009 wait_event_lock_irq(conf->wait_barrier,
1011 (atomic_read(&conf->nr_pending) &&
1012 current->bio_list &&
1013 (!bio_list_empty(¤t->bio_list[0]) ||
1014 !bio_list_empty(¤t->bio_list[1]))),
1017 if (!conf->nr_waiting)
1018 wake_up(&conf->wait_barrier);
1020 atomic_inc(&conf->nr_pending);
1021 spin_unlock_irq(&conf->resync_lock);
1024 static void allow_barrier(struct r10conf *conf)
1026 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1027 (conf->array_freeze_pending))
1028 wake_up(&conf->wait_barrier);
1031 static void freeze_array(struct r10conf *conf, int extra)
1033 /* stop syncio and normal IO and wait for everything to
1035 * We increment barrier and nr_waiting, and then
1036 * wait until nr_pending match nr_queued+extra
1037 * This is called in the context of one normal IO request
1038 * that has failed. Thus any sync request that might be pending
1039 * will be blocked by nr_pending, and we need to wait for
1040 * pending IO requests to complete or be queued for re-try.
1041 * Thus the number queued (nr_queued) plus this request (extra)
1042 * must match the number of pending IOs (nr_pending) before
1045 spin_lock_irq(&conf->resync_lock);
1046 conf->array_freeze_pending++;
1049 wait_event_lock_irq_cmd(conf->wait_barrier,
1050 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1052 flush_pending_writes(conf));
1054 conf->array_freeze_pending--;
1055 spin_unlock_irq(&conf->resync_lock);
1058 static void unfreeze_array(struct r10conf *conf)
1060 /* reverse the effect of the freeze */
1061 spin_lock_irq(&conf->resync_lock);
1064 wake_up(&conf->wait_barrier);
1065 spin_unlock_irq(&conf->resync_lock);
1068 static sector_t choose_data_offset(struct r10bio *r10_bio,
1069 struct md_rdev *rdev)
1071 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1072 test_bit(R10BIO_Previous, &r10_bio->state))
1073 return rdev->data_offset;
1075 return rdev->new_data_offset;
1078 struct raid10_plug_cb {
1079 struct blk_plug_cb cb;
1080 struct bio_list pending;
1084 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1086 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1088 struct mddev *mddev = plug->cb.data;
1089 struct r10conf *conf = mddev->private;
1092 if (from_schedule || current->bio_list) {
1093 spin_lock_irq(&conf->device_lock);
1094 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1095 conf->pending_count += plug->pending_cnt;
1096 spin_unlock_irq(&conf->device_lock);
1097 wake_up(&conf->wait_barrier);
1098 md_wakeup_thread(mddev->thread);
1103 /* we aren't scheduling, so we can do the write-out directly. */
1104 bio = bio_list_get(&plug->pending);
1105 bitmap_unplug(mddev->bitmap);
1106 wake_up(&conf->wait_barrier);
1108 while (bio) { /* submit pending writes */
1109 struct bio *next = bio->bi_next;
1110 struct md_rdev *rdev = (void*)bio->bi_disk;
1111 bio->bi_next = NULL;
1112 bio_set_dev(bio, rdev->bdev);
1113 if (test_bit(Faulty, &rdev->flags)) {
1115 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1116 !blk_queue_discard(bio->bi_disk->queue)))
1117 /* Just ignore it */
1120 generic_make_request(bio);
1126 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1127 struct r10bio *r10_bio)
1129 struct r10conf *conf = mddev->private;
1130 struct bio *read_bio;
1131 const int op = bio_op(bio);
1132 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1135 struct md_rdev *rdev;
1136 char b[BDEVNAME_SIZE];
1137 int slot = r10_bio->read_slot;
1138 struct md_rdev *err_rdev = NULL;
1139 gfp_t gfp = GFP_NOIO;
1141 if (r10_bio->devs[slot].rdev) {
1143 * This is an error retry, but we cannot
1144 * safely dereference the rdev in the r10_bio,
1145 * we must use the one in conf.
1146 * If it has already been disconnected (unlikely)
1147 * we lose the device name in error messages.
1151 * As we are blocking raid10, it is a little safer to
1154 gfp = GFP_NOIO | __GFP_HIGH;
1157 disk = r10_bio->devs[slot].devnum;
1158 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1160 bdevname(err_rdev->bdev, b);
1163 /* This never gets dereferenced */
1164 err_rdev = r10_bio->devs[slot].rdev;
1169 * Register the new request and wait if the reconstruction
1170 * thread has put up a bar for new requests.
1171 * Continue immediately if no resync is active currently.
1175 sectors = r10_bio->sectors;
1176 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1177 bio->bi_iter.bi_sector < conf->reshape_progress &&
1178 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1180 * IO spans the reshape position. Need to wait for reshape to
1183 raid10_log(conf->mddev, "wait reshape");
1184 allow_barrier(conf);
1185 wait_event(conf->wait_barrier,
1186 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1187 conf->reshape_progress >= bio->bi_iter.bi_sector +
1192 rdev = read_balance(conf, r10_bio, &max_sectors);
1195 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1197 (unsigned long long)r10_bio->sector);
1199 raid_end_bio_io(r10_bio);
1203 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1205 bdevname(rdev->bdev, b),
1206 (unsigned long long)r10_bio->sector);
1207 if (max_sectors < bio_sectors(bio)) {
1208 struct bio *split = bio_split(bio, max_sectors,
1209 gfp, &conf->bio_split);
1210 bio_chain(split, bio);
1211 generic_make_request(bio);
1213 r10_bio->master_bio = bio;
1214 r10_bio->sectors = max_sectors;
1216 slot = r10_bio->read_slot;
1218 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1220 r10_bio->devs[slot].bio = read_bio;
1221 r10_bio->devs[slot].rdev = rdev;
1223 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1224 choose_data_offset(r10_bio, rdev);
1225 bio_set_dev(read_bio, rdev->bdev);
1226 read_bio->bi_end_io = raid10_end_read_request;
1227 bio_set_op_attrs(read_bio, op, do_sync);
1228 if (test_bit(FailFast, &rdev->flags) &&
1229 test_bit(R10BIO_FailFast, &r10_bio->state))
1230 read_bio->bi_opf |= MD_FAILFAST;
1231 read_bio->bi_private = r10_bio;
1234 trace_block_bio_remap(read_bio->bi_disk->queue,
1235 read_bio, disk_devt(mddev->gendisk),
1237 generic_make_request(read_bio);
1241 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1242 struct bio *bio, bool replacement,
1245 const int op = bio_op(bio);
1246 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1247 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1248 unsigned long flags;
1249 struct blk_plug_cb *cb;
1250 struct raid10_plug_cb *plug = NULL;
1251 struct r10conf *conf = mddev->private;
1252 struct md_rdev *rdev;
1253 int devnum = r10_bio->devs[n_copy].devnum;
1257 rdev = conf->mirrors[devnum].replacement;
1259 /* Replacement just got moved to main 'rdev' */
1261 rdev = conf->mirrors[devnum].rdev;
1264 rdev = conf->mirrors[devnum].rdev;
1266 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1268 r10_bio->devs[n_copy].repl_bio = mbio;
1270 r10_bio->devs[n_copy].bio = mbio;
1272 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1273 choose_data_offset(r10_bio, rdev));
1274 bio_set_dev(mbio, rdev->bdev);
1275 mbio->bi_end_io = raid10_end_write_request;
1276 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1277 if (!replacement && test_bit(FailFast,
1278 &conf->mirrors[devnum].rdev->flags)
1279 && enough(conf, devnum))
1280 mbio->bi_opf |= MD_FAILFAST;
1281 mbio->bi_private = r10_bio;
1283 if (conf->mddev->gendisk)
1284 trace_block_bio_remap(mbio->bi_disk->queue,
1285 mbio, disk_devt(conf->mddev->gendisk),
1287 /* flush_pending_writes() needs access to the rdev so...*/
1288 mbio->bi_disk = (void *)rdev;
1290 atomic_inc(&r10_bio->remaining);
1292 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1294 plug = container_of(cb, struct raid10_plug_cb, cb);
1298 bio_list_add(&plug->pending, mbio);
1299 plug->pending_cnt++;
1301 spin_lock_irqsave(&conf->device_lock, flags);
1302 bio_list_add(&conf->pending_bio_list, mbio);
1303 conf->pending_count++;
1304 spin_unlock_irqrestore(&conf->device_lock, flags);
1305 md_wakeup_thread(mddev->thread);
1309 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1310 struct r10bio *r10_bio)
1312 struct r10conf *conf = mddev->private;
1314 struct md_rdev *blocked_rdev;
1318 if ((mddev_is_clustered(mddev) &&
1319 md_cluster_ops->area_resyncing(mddev, WRITE,
1320 bio->bi_iter.bi_sector,
1321 bio_end_sector(bio)))) {
1324 prepare_to_wait(&conf->wait_barrier,
1326 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1327 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1331 finish_wait(&conf->wait_barrier, &w);
1335 * Register the new request and wait if the reconstruction
1336 * thread has put up a bar for new requests.
1337 * Continue immediately if no resync is active currently.
1341 sectors = r10_bio->sectors;
1342 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1343 bio->bi_iter.bi_sector < conf->reshape_progress &&
1344 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1346 * IO spans the reshape position. Need to wait for reshape to
1349 raid10_log(conf->mddev, "wait reshape");
1350 allow_barrier(conf);
1351 wait_event(conf->wait_barrier,
1352 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1353 conf->reshape_progress >= bio->bi_iter.bi_sector +
1358 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1359 (mddev->reshape_backwards
1360 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1361 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1362 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1363 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1364 /* Need to update reshape_position in metadata */
1365 mddev->reshape_position = conf->reshape_progress;
1366 set_mask_bits(&mddev->sb_flags, 0,
1367 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1368 md_wakeup_thread(mddev->thread);
1369 raid10_log(conf->mddev, "wait reshape metadata");
1370 wait_event(mddev->sb_wait,
1371 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1373 conf->reshape_safe = mddev->reshape_position;
1376 if (conf->pending_count >= max_queued_requests) {
1377 md_wakeup_thread(mddev->thread);
1378 raid10_log(mddev, "wait queued");
1379 wait_event(conf->wait_barrier,
1380 conf->pending_count < max_queued_requests);
1382 /* first select target devices under rcu_lock and
1383 * inc refcount on their rdev. Record them by setting
1385 * If there are known/acknowledged bad blocks on any device
1386 * on which we have seen a write error, we want to avoid
1387 * writing to those blocks. This potentially requires several
1388 * writes to write around the bad blocks. Each set of writes
1389 * gets its own r10_bio with a set of bios attached.
1392 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1393 raid10_find_phys(conf, r10_bio);
1395 blocked_rdev = NULL;
1397 max_sectors = r10_bio->sectors;
1399 for (i = 0; i < conf->copies; i++) {
1400 int d = r10_bio->devs[i].devnum;
1401 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1402 struct md_rdev *rrdev = rcu_dereference(
1403 conf->mirrors[d].replacement);
1406 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407 atomic_inc(&rdev->nr_pending);
1408 blocked_rdev = rdev;
1411 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1412 atomic_inc(&rrdev->nr_pending);
1413 blocked_rdev = rrdev;
1416 if (rdev && (test_bit(Faulty, &rdev->flags)))
1418 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1421 r10_bio->devs[i].bio = NULL;
1422 r10_bio->devs[i].repl_bio = NULL;
1424 if (!rdev && !rrdev) {
1425 set_bit(R10BIO_Degraded, &r10_bio->state);
1428 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1430 sector_t dev_sector = r10_bio->devs[i].addr;
1434 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1435 &first_bad, &bad_sectors);
1437 /* Mustn't write here until the bad block
1440 atomic_inc(&rdev->nr_pending);
1441 set_bit(BlockedBadBlocks, &rdev->flags);
1442 blocked_rdev = rdev;
1445 if (is_bad && first_bad <= dev_sector) {
1446 /* Cannot write here at all */
1447 bad_sectors -= (dev_sector - first_bad);
1448 if (bad_sectors < max_sectors)
1449 /* Mustn't write more than bad_sectors
1450 * to other devices yet
1452 max_sectors = bad_sectors;
1453 /* We don't set R10BIO_Degraded as that
1454 * only applies if the disk is missing,
1455 * so it might be re-added, and we want to
1456 * know to recover this chunk.
1457 * In this case the device is here, and the
1458 * fact that this chunk is not in-sync is
1459 * recorded in the bad block log.
1464 int good_sectors = first_bad - dev_sector;
1465 if (good_sectors < max_sectors)
1466 max_sectors = good_sectors;
1470 r10_bio->devs[i].bio = bio;
1471 atomic_inc(&rdev->nr_pending);
1474 r10_bio->devs[i].repl_bio = bio;
1475 atomic_inc(&rrdev->nr_pending);
1480 if (unlikely(blocked_rdev)) {
1481 /* Have to wait for this device to get unblocked, then retry */
1485 for (j = 0; j < i; j++) {
1486 if (r10_bio->devs[j].bio) {
1487 d = r10_bio->devs[j].devnum;
1488 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1490 if (r10_bio->devs[j].repl_bio) {
1491 struct md_rdev *rdev;
1492 d = r10_bio->devs[j].devnum;
1493 rdev = conf->mirrors[d].replacement;
1495 /* Race with remove_disk */
1497 rdev = conf->mirrors[d].rdev;
1499 rdev_dec_pending(rdev, mddev);
1502 allow_barrier(conf);
1503 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1504 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1509 if (max_sectors < r10_bio->sectors)
1510 r10_bio->sectors = max_sectors;
1512 if (r10_bio->sectors < bio_sectors(bio)) {
1513 struct bio *split = bio_split(bio, r10_bio->sectors,
1514 GFP_NOIO, &conf->bio_split);
1515 bio_chain(split, bio);
1516 generic_make_request(bio);
1518 r10_bio->master_bio = bio;
1521 atomic_set(&r10_bio->remaining, 1);
1522 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1524 for (i = 0; i < conf->copies; i++) {
1525 if (r10_bio->devs[i].bio)
1526 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1527 if (r10_bio->devs[i].repl_bio)
1528 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1530 one_write_done(r10_bio);
1533 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1535 struct r10conf *conf = mddev->private;
1536 struct r10bio *r10_bio;
1538 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1540 r10_bio->master_bio = bio;
1541 r10_bio->sectors = sectors;
1543 r10_bio->mddev = mddev;
1544 r10_bio->sector = bio->bi_iter.bi_sector;
1546 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1548 if (bio_data_dir(bio) == READ)
1549 raid10_read_request(mddev, bio, r10_bio);
1551 raid10_write_request(mddev, bio, r10_bio);
1554 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1556 struct r10conf *conf = mddev->private;
1557 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1558 int chunk_sects = chunk_mask + 1;
1559 int sectors = bio_sectors(bio);
1561 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1562 md_flush_request(mddev, bio);
1566 if (!md_write_start(mddev, bio))
1570 * If this request crosses a chunk boundary, we need to split
1573 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1574 sectors > chunk_sects
1575 && (conf->geo.near_copies < conf->geo.raid_disks
1576 || conf->prev.near_copies <
1577 conf->prev.raid_disks)))
1578 sectors = chunk_sects -
1579 (bio->bi_iter.bi_sector &
1581 __make_request(mddev, bio, sectors);
1583 /* In case raid10d snuck in to freeze_array */
1584 wake_up(&conf->wait_barrier);
1588 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1590 struct r10conf *conf = mddev->private;
1593 if (conf->geo.near_copies < conf->geo.raid_disks)
1594 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1595 if (conf->geo.near_copies > 1)
1596 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1597 if (conf->geo.far_copies > 1) {
1598 if (conf->geo.far_offset)
1599 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1601 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1602 if (conf->geo.far_set_size != conf->geo.raid_disks)
1603 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1605 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1606 conf->geo.raid_disks - mddev->degraded);
1608 for (i = 0; i < conf->geo.raid_disks; i++) {
1609 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1610 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1613 seq_printf(seq, "]");
1616 /* check if there are enough drives for
1617 * every block to appear on atleast one.
1618 * Don't consider the device numbered 'ignore'
1619 * as we might be about to remove it.
1621 static int _enough(struct r10conf *conf, int previous, int ignore)
1627 disks = conf->prev.raid_disks;
1628 ncopies = conf->prev.near_copies;
1630 disks = conf->geo.raid_disks;
1631 ncopies = conf->geo.near_copies;
1636 int n = conf->copies;
1640 struct md_rdev *rdev;
1641 if (this != ignore &&
1642 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1643 test_bit(In_sync, &rdev->flags))
1645 this = (this+1) % disks;
1649 first = (first + ncopies) % disks;
1650 } while (first != 0);
1657 static int enough(struct r10conf *conf, int ignore)
1659 /* when calling 'enough', both 'prev' and 'geo' must
1661 * This is ensured if ->reconfig_mutex or ->device_lock
1664 return _enough(conf, 0, ignore) &&
1665 _enough(conf, 1, ignore);
1668 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1670 char b[BDEVNAME_SIZE];
1671 struct r10conf *conf = mddev->private;
1672 unsigned long flags;
1675 * If it is not operational, then we have already marked it as dead
1676 * else if it is the last working disks, ignore the error, let the
1677 * next level up know.
1678 * else mark the drive as failed
1680 spin_lock_irqsave(&conf->device_lock, flags);
1681 if (test_bit(In_sync, &rdev->flags)
1682 && !enough(conf, rdev->raid_disk)) {
1684 * Don't fail the drive, just return an IO error.
1686 spin_unlock_irqrestore(&conf->device_lock, flags);
1689 if (test_and_clear_bit(In_sync, &rdev->flags))
1692 * If recovery is running, make sure it aborts.
1694 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1695 set_bit(Blocked, &rdev->flags);
1696 set_bit(Faulty, &rdev->flags);
1697 set_mask_bits(&mddev->sb_flags, 0,
1698 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1699 spin_unlock_irqrestore(&conf->device_lock, flags);
1700 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1701 "md/raid10:%s: Operation continuing on %d devices.\n",
1702 mdname(mddev), bdevname(rdev->bdev, b),
1703 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1706 static void print_conf(struct r10conf *conf)
1709 struct md_rdev *rdev;
1711 pr_debug("RAID10 conf printout:\n");
1713 pr_debug("(!conf)\n");
1716 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1717 conf->geo.raid_disks);
1719 /* This is only called with ->reconfix_mutex held, so
1720 * rcu protection of rdev is not needed */
1721 for (i = 0; i < conf->geo.raid_disks; i++) {
1722 char b[BDEVNAME_SIZE];
1723 rdev = conf->mirrors[i].rdev;
1725 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1726 i, !test_bit(In_sync, &rdev->flags),
1727 !test_bit(Faulty, &rdev->flags),
1728 bdevname(rdev->bdev,b));
1732 static void close_sync(struct r10conf *conf)
1735 allow_barrier(conf);
1737 mempool_exit(&conf->r10buf_pool);
1740 static int raid10_spare_active(struct mddev *mddev)
1743 struct r10conf *conf = mddev->private;
1744 struct raid10_info *tmp;
1746 unsigned long flags;
1749 * Find all non-in_sync disks within the RAID10 configuration
1750 * and mark them in_sync
1752 for (i = 0; i < conf->geo.raid_disks; i++) {
1753 tmp = conf->mirrors + i;
1754 if (tmp->replacement
1755 && tmp->replacement->recovery_offset == MaxSector
1756 && !test_bit(Faulty, &tmp->replacement->flags)
1757 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1758 /* Replacement has just become active */
1760 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1763 /* Replaced device not technically faulty,
1764 * but we need to be sure it gets removed
1765 * and never re-added.
1767 set_bit(Faulty, &tmp->rdev->flags);
1768 sysfs_notify_dirent_safe(
1769 tmp->rdev->sysfs_state);
1771 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1772 } else if (tmp->rdev
1773 && tmp->rdev->recovery_offset == MaxSector
1774 && !test_bit(Faulty, &tmp->rdev->flags)
1775 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1777 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1780 spin_lock_irqsave(&conf->device_lock, flags);
1781 mddev->degraded -= count;
1782 spin_unlock_irqrestore(&conf->device_lock, flags);
1788 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1790 struct r10conf *conf = mddev->private;
1794 int last = conf->geo.raid_disks - 1;
1796 if (mddev->recovery_cp < MaxSector)
1797 /* only hot-add to in-sync arrays, as recovery is
1798 * very different from resync
1801 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1804 if (md_integrity_add_rdev(rdev, mddev))
1807 if (rdev->raid_disk >= 0)
1808 first = last = rdev->raid_disk;
1810 if (rdev->saved_raid_disk >= first &&
1811 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1812 mirror = rdev->saved_raid_disk;
1815 for ( ; mirror <= last ; mirror++) {
1816 struct raid10_info *p = &conf->mirrors[mirror];
1817 if (p->recovery_disabled == mddev->recovery_disabled)
1820 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1821 p->replacement != NULL)
1823 clear_bit(In_sync, &rdev->flags);
1824 set_bit(Replacement, &rdev->flags);
1825 rdev->raid_disk = mirror;
1828 disk_stack_limits(mddev->gendisk, rdev->bdev,
1829 rdev->data_offset << 9);
1831 rcu_assign_pointer(p->replacement, rdev);
1836 disk_stack_limits(mddev->gendisk, rdev->bdev,
1837 rdev->data_offset << 9);
1839 p->head_position = 0;
1840 p->recovery_disabled = mddev->recovery_disabled - 1;
1841 rdev->raid_disk = mirror;
1843 if (rdev->saved_raid_disk != mirror)
1845 rcu_assign_pointer(p->rdev, rdev);
1848 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1849 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1855 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1857 struct r10conf *conf = mddev->private;
1859 int number = rdev->raid_disk;
1860 struct md_rdev **rdevp;
1861 struct raid10_info *p = conf->mirrors + number;
1864 if (rdev == p->rdev)
1866 else if (rdev == p->replacement)
1867 rdevp = &p->replacement;
1871 if (test_bit(In_sync, &rdev->flags) ||
1872 atomic_read(&rdev->nr_pending)) {
1876 /* Only remove non-faulty devices if recovery
1879 if (!test_bit(Faulty, &rdev->flags) &&
1880 mddev->recovery_disabled != p->recovery_disabled &&
1881 (!p->replacement || p->replacement == rdev) &&
1882 number < conf->geo.raid_disks &&
1888 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1890 if (atomic_read(&rdev->nr_pending)) {
1891 /* lost the race, try later */
1897 if (p->replacement) {
1898 /* We must have just cleared 'rdev' */
1899 p->rdev = p->replacement;
1900 clear_bit(Replacement, &p->replacement->flags);
1901 smp_mb(); /* Make sure other CPUs may see both as identical
1902 * but will never see neither -- if they are careful.
1904 p->replacement = NULL;
1907 clear_bit(WantReplacement, &rdev->flags);
1908 err = md_integrity_register(mddev);
1916 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1918 struct r10conf *conf = r10_bio->mddev->private;
1920 if (!bio->bi_status)
1921 set_bit(R10BIO_Uptodate, &r10_bio->state);
1923 /* The write handler will notice the lack of
1924 * R10BIO_Uptodate and record any errors etc
1926 atomic_add(r10_bio->sectors,
1927 &conf->mirrors[d].rdev->corrected_errors);
1929 /* for reconstruct, we always reschedule after a read.
1930 * for resync, only after all reads
1932 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1933 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1934 atomic_dec_and_test(&r10_bio->remaining)) {
1935 /* we have read all the blocks,
1936 * do the comparison in process context in raid10d
1938 reschedule_retry(r10_bio);
1942 static void end_sync_read(struct bio *bio)
1944 struct r10bio *r10_bio = get_resync_r10bio(bio);
1945 struct r10conf *conf = r10_bio->mddev->private;
1946 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1948 __end_sync_read(r10_bio, bio, d);
1951 static void end_reshape_read(struct bio *bio)
1953 /* reshape read bio isn't allocated from r10buf_pool */
1954 struct r10bio *r10_bio = bio->bi_private;
1956 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1959 static void end_sync_request(struct r10bio *r10_bio)
1961 struct mddev *mddev = r10_bio->mddev;
1963 while (atomic_dec_and_test(&r10_bio->remaining)) {
1964 if (r10_bio->master_bio == NULL) {
1965 /* the primary of several recovery bios */
1966 sector_t s = r10_bio->sectors;
1967 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1968 test_bit(R10BIO_WriteError, &r10_bio->state))
1969 reschedule_retry(r10_bio);
1972 md_done_sync(mddev, s, 1);
1975 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1976 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1977 test_bit(R10BIO_WriteError, &r10_bio->state))
1978 reschedule_retry(r10_bio);
1986 static void end_sync_write(struct bio *bio)
1988 struct r10bio *r10_bio = get_resync_r10bio(bio);
1989 struct mddev *mddev = r10_bio->mddev;
1990 struct r10conf *conf = mddev->private;
1996 struct md_rdev *rdev = NULL;
1998 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2000 rdev = conf->mirrors[d].replacement;
2002 rdev = conf->mirrors[d].rdev;
2004 if (bio->bi_status) {
2006 md_error(mddev, rdev);
2008 set_bit(WriteErrorSeen, &rdev->flags);
2009 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2010 set_bit(MD_RECOVERY_NEEDED,
2011 &rdev->mddev->recovery);
2012 set_bit(R10BIO_WriteError, &r10_bio->state);
2014 } else if (is_badblock(rdev,
2015 r10_bio->devs[slot].addr,
2017 &first_bad, &bad_sectors))
2018 set_bit(R10BIO_MadeGood, &r10_bio->state);
2020 rdev_dec_pending(rdev, mddev);
2022 end_sync_request(r10_bio);
2026 * Note: sync and recover and handled very differently for raid10
2027 * This code is for resync.
2028 * For resync, we read through virtual addresses and read all blocks.
2029 * If there is any error, we schedule a write. The lowest numbered
2030 * drive is authoritative.
2031 * However requests come for physical address, so we need to map.
2032 * For every physical address there are raid_disks/copies virtual addresses,
2033 * which is always are least one, but is not necessarly an integer.
2034 * This means that a physical address can span multiple chunks, so we may
2035 * have to submit multiple io requests for a single sync request.
2038 * We check if all blocks are in-sync and only write to blocks that
2041 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2043 struct r10conf *conf = mddev->private;
2045 struct bio *tbio, *fbio;
2047 struct page **tpages, **fpages;
2049 atomic_set(&r10_bio->remaining, 1);
2051 /* find the first device with a block */
2052 for (i=0; i<conf->copies; i++)
2053 if (!r10_bio->devs[i].bio->bi_status)
2056 if (i == conf->copies)
2060 fbio = r10_bio->devs[i].bio;
2061 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2062 fbio->bi_iter.bi_idx = 0;
2063 fpages = get_resync_pages(fbio)->pages;
2065 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2066 /* now find blocks with errors */
2067 for (i=0 ; i < conf->copies ; i++) {
2069 struct md_rdev *rdev;
2070 struct resync_pages *rp;
2072 tbio = r10_bio->devs[i].bio;
2074 if (tbio->bi_end_io != end_sync_read)
2079 tpages = get_resync_pages(tbio)->pages;
2080 d = r10_bio->devs[i].devnum;
2081 rdev = conf->mirrors[d].rdev;
2082 if (!r10_bio->devs[i].bio->bi_status) {
2083 /* We know that the bi_io_vec layout is the same for
2084 * both 'first' and 'i', so we just compare them.
2085 * All vec entries are PAGE_SIZE;
2087 int sectors = r10_bio->sectors;
2088 for (j = 0; j < vcnt; j++) {
2089 int len = PAGE_SIZE;
2090 if (sectors < (len / 512))
2091 len = sectors * 512;
2092 if (memcmp(page_address(fpages[j]),
2093 page_address(tpages[j]),
2100 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2101 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2102 /* Don't fix anything. */
2104 } else if (test_bit(FailFast, &rdev->flags)) {
2105 /* Just give up on this device */
2106 md_error(rdev->mddev, rdev);
2109 /* Ok, we need to write this bio, either to correct an
2110 * inconsistency or to correct an unreadable block.
2111 * First we need to fixup bv_offset, bv_len and
2112 * bi_vecs, as the read request might have corrupted these
2114 rp = get_resync_pages(tbio);
2117 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2119 rp->raid_bio = r10_bio;
2120 tbio->bi_private = rp;
2121 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2122 tbio->bi_end_io = end_sync_write;
2123 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2125 bio_copy_data(tbio, fbio);
2127 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2128 atomic_inc(&r10_bio->remaining);
2129 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2131 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2132 tbio->bi_opf |= MD_FAILFAST;
2133 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2134 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2135 generic_make_request(tbio);
2138 /* Now write out to any replacement devices
2141 for (i = 0; i < conf->copies; i++) {
2144 tbio = r10_bio->devs[i].repl_bio;
2145 if (!tbio || !tbio->bi_end_io)
2147 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2148 && r10_bio->devs[i].bio != fbio)
2149 bio_copy_data(tbio, fbio);
2150 d = r10_bio->devs[i].devnum;
2151 atomic_inc(&r10_bio->remaining);
2152 md_sync_acct(conf->mirrors[d].replacement->bdev,
2154 generic_make_request(tbio);
2158 if (atomic_dec_and_test(&r10_bio->remaining)) {
2159 md_done_sync(mddev, r10_bio->sectors, 1);
2165 * Now for the recovery code.
2166 * Recovery happens across physical sectors.
2167 * We recover all non-is_sync drives by finding the virtual address of
2168 * each, and then choose a working drive that also has that virt address.
2169 * There is a separate r10_bio for each non-in_sync drive.
2170 * Only the first two slots are in use. The first for reading,
2171 * The second for writing.
2174 static void fix_recovery_read_error(struct r10bio *r10_bio)
2176 /* We got a read error during recovery.
2177 * We repeat the read in smaller page-sized sections.
2178 * If a read succeeds, write it to the new device or record
2179 * a bad block if we cannot.
2180 * If a read fails, record a bad block on both old and
2183 struct mddev *mddev = r10_bio->mddev;
2184 struct r10conf *conf = mddev->private;
2185 struct bio *bio = r10_bio->devs[0].bio;
2187 int sectors = r10_bio->sectors;
2189 int dr = r10_bio->devs[0].devnum;
2190 int dw = r10_bio->devs[1].devnum;
2191 struct page **pages = get_resync_pages(bio)->pages;
2195 struct md_rdev *rdev;
2199 if (s > (PAGE_SIZE>>9))
2202 rdev = conf->mirrors[dr].rdev;
2203 addr = r10_bio->devs[0].addr + sect,
2204 ok = sync_page_io(rdev,
2208 REQ_OP_READ, 0, false);
2210 rdev = conf->mirrors[dw].rdev;
2211 addr = r10_bio->devs[1].addr + sect;
2212 ok = sync_page_io(rdev,
2216 REQ_OP_WRITE, 0, false);
2218 set_bit(WriteErrorSeen, &rdev->flags);
2219 if (!test_and_set_bit(WantReplacement,
2221 set_bit(MD_RECOVERY_NEEDED,
2222 &rdev->mddev->recovery);
2226 /* We don't worry if we cannot set a bad block -
2227 * it really is bad so there is no loss in not
2230 rdev_set_badblocks(rdev, addr, s, 0);
2232 if (rdev != conf->mirrors[dw].rdev) {
2233 /* need bad block on destination too */
2234 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2235 addr = r10_bio->devs[1].addr + sect;
2236 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2238 /* just abort the recovery */
2239 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2242 conf->mirrors[dw].recovery_disabled
2243 = mddev->recovery_disabled;
2244 set_bit(MD_RECOVERY_INTR,
2257 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2259 struct r10conf *conf = mddev->private;
2261 struct bio *wbio, *wbio2;
2263 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2264 fix_recovery_read_error(r10_bio);
2265 end_sync_request(r10_bio);
2270 * share the pages with the first bio
2271 * and submit the write request
2273 d = r10_bio->devs[1].devnum;
2274 wbio = r10_bio->devs[1].bio;
2275 wbio2 = r10_bio->devs[1].repl_bio;
2276 /* Need to test wbio2->bi_end_io before we call
2277 * generic_make_request as if the former is NULL,
2278 * the latter is free to free wbio2.
2280 if (wbio2 && !wbio2->bi_end_io)
2282 if (wbio->bi_end_io) {
2283 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2284 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2285 generic_make_request(wbio);
2288 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2289 md_sync_acct(conf->mirrors[d].replacement->bdev,
2290 bio_sectors(wbio2));
2291 generic_make_request(wbio2);
2296 * Used by fix_read_error() to decay the per rdev read_errors.
2297 * We halve the read error count for every hour that has elapsed
2298 * since the last recorded read error.
2301 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2304 unsigned long hours_since_last;
2305 unsigned int read_errors = atomic_read(&rdev->read_errors);
2307 cur_time_mon = ktime_get_seconds();
2309 if (rdev->last_read_error == 0) {
2310 /* first time we've seen a read error */
2311 rdev->last_read_error = cur_time_mon;
2315 hours_since_last = (long)(cur_time_mon -
2316 rdev->last_read_error) / 3600;
2318 rdev->last_read_error = cur_time_mon;
2321 * if hours_since_last is > the number of bits in read_errors
2322 * just set read errors to 0. We do this to avoid
2323 * overflowing the shift of read_errors by hours_since_last.
2325 if (hours_since_last >= 8 * sizeof(read_errors))
2326 atomic_set(&rdev->read_errors, 0);
2328 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2331 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2332 int sectors, struct page *page, int rw)
2337 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2338 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2340 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2344 set_bit(WriteErrorSeen, &rdev->flags);
2345 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2346 set_bit(MD_RECOVERY_NEEDED,
2347 &rdev->mddev->recovery);
2349 /* need to record an error - either for the block or the device */
2350 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2351 md_error(rdev->mddev, rdev);
2356 * This is a kernel thread which:
2358 * 1. Retries failed read operations on working mirrors.
2359 * 2. Updates the raid superblock when problems encounter.
2360 * 3. Performs writes following reads for array synchronising.
2363 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2365 int sect = 0; /* Offset from r10_bio->sector */
2366 int sectors = r10_bio->sectors;
2367 struct md_rdev *rdev;
2368 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2369 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2371 /* still own a reference to this rdev, so it cannot
2372 * have been cleared recently.
2374 rdev = conf->mirrors[d].rdev;
2376 if (test_bit(Faulty, &rdev->flags))
2377 /* drive has already been failed, just ignore any
2378 more fix_read_error() attempts */
2381 check_decay_read_errors(mddev, rdev);
2382 atomic_inc(&rdev->read_errors);
2383 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2384 char b[BDEVNAME_SIZE];
2385 bdevname(rdev->bdev, b);
2387 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2389 atomic_read(&rdev->read_errors), max_read_errors);
2390 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2392 md_error(mddev, rdev);
2393 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2399 int sl = r10_bio->read_slot;
2403 if (s > (PAGE_SIZE>>9))
2411 d = r10_bio->devs[sl].devnum;
2412 rdev = rcu_dereference(conf->mirrors[d].rdev);
2414 test_bit(In_sync, &rdev->flags) &&
2415 !test_bit(Faulty, &rdev->flags) &&
2416 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2417 &first_bad, &bad_sectors) == 0) {
2418 atomic_inc(&rdev->nr_pending);
2420 success = sync_page_io(rdev,
2421 r10_bio->devs[sl].addr +
2425 REQ_OP_READ, 0, false);
2426 rdev_dec_pending(rdev, mddev);
2432 if (sl == conf->copies)
2434 } while (!success && sl != r10_bio->read_slot);
2438 /* Cannot read from anywhere, just mark the block
2439 * as bad on the first device to discourage future
2442 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2443 rdev = conf->mirrors[dn].rdev;
2445 if (!rdev_set_badblocks(
2447 r10_bio->devs[r10_bio->read_slot].addr
2450 md_error(mddev, rdev);
2451 r10_bio->devs[r10_bio->read_slot].bio
2458 /* write it back and re-read */
2460 while (sl != r10_bio->read_slot) {
2461 char b[BDEVNAME_SIZE];
2466 d = r10_bio->devs[sl].devnum;
2467 rdev = rcu_dereference(conf->mirrors[d].rdev);
2469 test_bit(Faulty, &rdev->flags) ||
2470 !test_bit(In_sync, &rdev->flags))
2473 atomic_inc(&rdev->nr_pending);
2475 if (r10_sync_page_io(rdev,
2476 r10_bio->devs[sl].addr +
2478 s, conf->tmppage, WRITE)
2480 /* Well, this device is dead */
2481 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2483 (unsigned long long)(
2485 choose_data_offset(r10_bio,
2487 bdevname(rdev->bdev, b));
2488 pr_notice("md/raid10:%s: %s: failing drive\n",
2490 bdevname(rdev->bdev, b));
2492 rdev_dec_pending(rdev, mddev);
2496 while (sl != r10_bio->read_slot) {
2497 char b[BDEVNAME_SIZE];
2502 d = r10_bio->devs[sl].devnum;
2503 rdev = rcu_dereference(conf->mirrors[d].rdev);
2505 test_bit(Faulty, &rdev->flags) ||
2506 !test_bit(In_sync, &rdev->flags))
2509 atomic_inc(&rdev->nr_pending);
2511 switch (r10_sync_page_io(rdev,
2512 r10_bio->devs[sl].addr +
2517 /* Well, this device is dead */
2518 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2520 (unsigned long long)(
2522 choose_data_offset(r10_bio, rdev)),
2523 bdevname(rdev->bdev, b));
2524 pr_notice("md/raid10:%s: %s: failing drive\n",
2526 bdevname(rdev->bdev, b));
2529 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2531 (unsigned long long)(
2533 choose_data_offset(r10_bio, rdev)),
2534 bdevname(rdev->bdev, b));
2535 atomic_add(s, &rdev->corrected_errors);
2538 rdev_dec_pending(rdev, mddev);
2548 static int narrow_write_error(struct r10bio *r10_bio, int i)
2550 struct bio *bio = r10_bio->master_bio;
2551 struct mddev *mddev = r10_bio->mddev;
2552 struct r10conf *conf = mddev->private;
2553 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2554 /* bio has the data to be written to slot 'i' where
2555 * we just recently had a write error.
2556 * We repeatedly clone the bio and trim down to one block,
2557 * then try the write. Where the write fails we record
2559 * It is conceivable that the bio doesn't exactly align with
2560 * blocks. We must handle this.
2562 * We currently own a reference to the rdev.
2568 int sect_to_write = r10_bio->sectors;
2571 if (rdev->badblocks.shift < 0)
2574 block_sectors = roundup(1 << rdev->badblocks.shift,
2575 bdev_logical_block_size(rdev->bdev) >> 9);
2576 sector = r10_bio->sector;
2577 sectors = ((r10_bio->sector + block_sectors)
2578 & ~(sector_t)(block_sectors - 1))
2581 while (sect_to_write) {
2584 if (sectors > sect_to_write)
2585 sectors = sect_to_write;
2586 /* Write at 'sector' for 'sectors' */
2587 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2588 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2589 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2590 wbio->bi_iter.bi_sector = wsector +
2591 choose_data_offset(r10_bio, rdev);
2592 bio_set_dev(wbio, rdev->bdev);
2593 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2595 if (submit_bio_wait(wbio) < 0)
2597 ok = rdev_set_badblocks(rdev, wsector,
2602 sect_to_write -= sectors;
2604 sectors = block_sectors;
2609 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2611 int slot = r10_bio->read_slot;
2613 struct r10conf *conf = mddev->private;
2614 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2616 /* we got a read error. Maybe the drive is bad. Maybe just
2617 * the block and we can fix it.
2618 * We freeze all other IO, and try reading the block from
2619 * other devices. When we find one, we re-write
2620 * and check it that fixes the read error.
2621 * This is all done synchronously while the array is
2624 bio = r10_bio->devs[slot].bio;
2626 r10_bio->devs[slot].bio = NULL;
2629 r10_bio->devs[slot].bio = IO_BLOCKED;
2630 else if (!test_bit(FailFast, &rdev->flags)) {
2631 freeze_array(conf, 1);
2632 fix_read_error(conf, mddev, r10_bio);
2633 unfreeze_array(conf);
2635 md_error(mddev, rdev);
2637 rdev_dec_pending(rdev, mddev);
2638 allow_barrier(conf);
2640 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2643 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2645 /* Some sort of write request has finished and it
2646 * succeeded in writing where we thought there was a
2647 * bad block. So forget the bad block.
2648 * Or possibly if failed and we need to record
2652 struct md_rdev *rdev;
2654 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2655 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2656 for (m = 0; m < conf->copies; m++) {
2657 int dev = r10_bio->devs[m].devnum;
2658 rdev = conf->mirrors[dev].rdev;
2659 if (r10_bio->devs[m].bio == NULL ||
2660 r10_bio->devs[m].bio->bi_end_io == NULL)
2662 if (!r10_bio->devs[m].bio->bi_status) {
2663 rdev_clear_badblocks(
2665 r10_bio->devs[m].addr,
2666 r10_bio->sectors, 0);
2668 if (!rdev_set_badblocks(
2670 r10_bio->devs[m].addr,
2671 r10_bio->sectors, 0))
2672 md_error(conf->mddev, rdev);
2674 rdev = conf->mirrors[dev].replacement;
2675 if (r10_bio->devs[m].repl_bio == NULL ||
2676 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2679 if (!r10_bio->devs[m].repl_bio->bi_status) {
2680 rdev_clear_badblocks(
2682 r10_bio->devs[m].addr,
2683 r10_bio->sectors, 0);
2685 if (!rdev_set_badblocks(
2687 r10_bio->devs[m].addr,
2688 r10_bio->sectors, 0))
2689 md_error(conf->mddev, rdev);
2695 for (m = 0; m < conf->copies; m++) {
2696 int dev = r10_bio->devs[m].devnum;
2697 struct bio *bio = r10_bio->devs[m].bio;
2698 rdev = conf->mirrors[dev].rdev;
2699 if (bio == IO_MADE_GOOD) {
2700 rdev_clear_badblocks(
2702 r10_bio->devs[m].addr,
2703 r10_bio->sectors, 0);
2704 rdev_dec_pending(rdev, conf->mddev);
2705 } else if (bio != NULL && bio->bi_status) {
2707 if (!narrow_write_error(r10_bio, m)) {
2708 md_error(conf->mddev, rdev);
2709 set_bit(R10BIO_Degraded,
2712 rdev_dec_pending(rdev, conf->mddev);
2714 bio = r10_bio->devs[m].repl_bio;
2715 rdev = conf->mirrors[dev].replacement;
2716 if (rdev && bio == IO_MADE_GOOD) {
2717 rdev_clear_badblocks(
2719 r10_bio->devs[m].addr,
2720 r10_bio->sectors, 0);
2721 rdev_dec_pending(rdev, conf->mddev);
2725 spin_lock_irq(&conf->device_lock);
2726 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2728 spin_unlock_irq(&conf->device_lock);
2730 * In case freeze_array() is waiting for condition
2731 * nr_pending == nr_queued + extra to be true.
2733 wake_up(&conf->wait_barrier);
2734 md_wakeup_thread(conf->mddev->thread);
2736 if (test_bit(R10BIO_WriteError,
2738 close_write(r10_bio);
2739 raid_end_bio_io(r10_bio);
2744 static void raid10d(struct md_thread *thread)
2746 struct mddev *mddev = thread->mddev;
2747 struct r10bio *r10_bio;
2748 unsigned long flags;
2749 struct r10conf *conf = mddev->private;
2750 struct list_head *head = &conf->retry_list;
2751 struct blk_plug plug;
2753 md_check_recovery(mddev);
2755 if (!list_empty_careful(&conf->bio_end_io_list) &&
2756 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2758 spin_lock_irqsave(&conf->device_lock, flags);
2759 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2760 while (!list_empty(&conf->bio_end_io_list)) {
2761 list_move(conf->bio_end_io_list.prev, &tmp);
2765 spin_unlock_irqrestore(&conf->device_lock, flags);
2766 while (!list_empty(&tmp)) {
2767 r10_bio = list_first_entry(&tmp, struct r10bio,
2769 list_del(&r10_bio->retry_list);
2770 if (mddev->degraded)
2771 set_bit(R10BIO_Degraded, &r10_bio->state);
2773 if (test_bit(R10BIO_WriteError,
2775 close_write(r10_bio);
2776 raid_end_bio_io(r10_bio);
2780 blk_start_plug(&plug);
2783 flush_pending_writes(conf);
2785 spin_lock_irqsave(&conf->device_lock, flags);
2786 if (list_empty(head)) {
2787 spin_unlock_irqrestore(&conf->device_lock, flags);
2790 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2791 list_del(head->prev);
2793 spin_unlock_irqrestore(&conf->device_lock, flags);
2795 mddev = r10_bio->mddev;
2796 conf = mddev->private;
2797 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2798 test_bit(R10BIO_WriteError, &r10_bio->state))
2799 handle_write_completed(conf, r10_bio);
2800 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2801 reshape_request_write(mddev, r10_bio);
2802 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2803 sync_request_write(mddev, r10_bio);
2804 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2805 recovery_request_write(mddev, r10_bio);
2806 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2807 handle_read_error(mddev, r10_bio);
2812 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2813 md_check_recovery(mddev);
2815 blk_finish_plug(&plug);
2818 static int init_resync(struct r10conf *conf)
2822 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2823 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2824 conf->have_replacement = 0;
2825 for (i = 0; i < conf->geo.raid_disks; i++)
2826 if (conf->mirrors[i].replacement)
2827 conf->have_replacement = 1;
2828 ret = mempool_init(&conf->r10buf_pool, buffs,
2829 r10buf_pool_alloc, r10buf_pool_free, conf);
2832 conf->next_resync = 0;
2836 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2838 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2839 struct rsync_pages *rp;
2844 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2845 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2846 nalloc = conf->copies; /* resync */
2848 nalloc = 2; /* recovery */
2850 for (i = 0; i < nalloc; i++) {
2851 bio = r10bio->devs[i].bio;
2852 rp = bio->bi_private;
2854 bio->bi_private = rp;
2855 bio = r10bio->devs[i].repl_bio;
2857 rp = bio->bi_private;
2859 bio->bi_private = rp;
2866 * Set cluster_sync_high since we need other nodes to add the
2867 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2869 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2871 sector_t window_size;
2872 int extra_chunk, chunks;
2875 * First, here we define "stripe" as a unit which across
2876 * all member devices one time, so we get chunks by use
2877 * raid_disks / near_copies. Otherwise, if near_copies is
2878 * close to raid_disks, then resync window could increases
2879 * linearly with the increase of raid_disks, which means
2880 * we will suspend a really large IO window while it is not
2881 * necessary. If raid_disks is not divisible by near_copies,
2882 * an extra chunk is needed to ensure the whole "stripe" is
2886 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2887 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2891 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2894 * At least use a 32M window to align with raid1's resync window
2896 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2897 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2899 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2903 * perform a "sync" on one "block"
2905 * We need to make sure that no normal I/O request - particularly write
2906 * requests - conflict with active sync requests.
2908 * This is achieved by tracking pending requests and a 'barrier' concept
2909 * that can be installed to exclude normal IO requests.
2911 * Resync and recovery are handled very differently.
2912 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2914 * For resync, we iterate over virtual addresses, read all copies,
2915 * and update if there are differences. If only one copy is live,
2917 * For recovery, we iterate over physical addresses, read a good
2918 * value for each non-in_sync drive, and over-write.
2920 * So, for recovery we may have several outstanding complex requests for a
2921 * given address, one for each out-of-sync device. We model this by allocating
2922 * a number of r10_bio structures, one for each out-of-sync device.
2923 * As we setup these structures, we collect all bio's together into a list
2924 * which we then process collectively to add pages, and then process again
2925 * to pass to generic_make_request.
2927 * The r10_bio structures are linked using a borrowed master_bio pointer.
2928 * This link is counted in ->remaining. When the r10_bio that points to NULL
2929 * has its remaining count decremented to 0, the whole complex operation
2934 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2937 struct r10conf *conf = mddev->private;
2938 struct r10bio *r10_bio;
2939 struct bio *biolist = NULL, *bio;
2940 sector_t max_sector, nr_sectors;
2943 sector_t sync_blocks;
2944 sector_t sectors_skipped = 0;
2945 int chunks_skipped = 0;
2946 sector_t chunk_mask = conf->geo.chunk_mask;
2949 if (!mempool_initialized(&conf->r10buf_pool))
2950 if (init_resync(conf))
2954 * Allow skipping a full rebuild for incremental assembly
2955 * of a clean array, like RAID1 does.
2957 if (mddev->bitmap == NULL &&
2958 mddev->recovery_cp == MaxSector &&
2959 mddev->reshape_position == MaxSector &&
2960 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2961 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2962 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2963 conf->fullsync == 0) {
2965 return mddev->dev_sectors - sector_nr;
2969 max_sector = mddev->dev_sectors;
2970 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2971 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2972 max_sector = mddev->resync_max_sectors;
2973 if (sector_nr >= max_sector) {
2974 conf->cluster_sync_low = 0;
2975 conf->cluster_sync_high = 0;
2977 /* If we aborted, we need to abort the
2978 * sync on the 'current' bitmap chucks (there can
2979 * be several when recovering multiple devices).
2980 * as we may have started syncing it but not finished.
2981 * We can find the current address in
2982 * mddev->curr_resync, but for recovery,
2983 * we need to convert that to several
2984 * virtual addresses.
2986 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2992 if (mddev->curr_resync < max_sector) { /* aborted */
2993 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2994 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2996 else for (i = 0; i < conf->geo.raid_disks; i++) {
2998 raid10_find_virt(conf, mddev->curr_resync, i);
2999 bitmap_end_sync(mddev->bitmap, sect,
3003 /* completed sync */
3004 if ((!mddev->bitmap || conf->fullsync)
3005 && conf->have_replacement
3006 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3007 /* Completed a full sync so the replacements
3008 * are now fully recovered.
3011 for (i = 0; i < conf->geo.raid_disks; i++) {
3012 struct md_rdev *rdev =
3013 rcu_dereference(conf->mirrors[i].replacement);
3015 rdev->recovery_offset = MaxSector;
3021 bitmap_close_sync(mddev->bitmap);
3024 return sectors_skipped;
3027 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3028 return reshape_request(mddev, sector_nr, skipped);
3030 if (chunks_skipped >= conf->geo.raid_disks) {
3031 /* if there has been nothing to do on any drive,
3032 * then there is nothing to do at all..
3035 return (max_sector - sector_nr) + sectors_skipped;
3038 if (max_sector > mddev->resync_max)
3039 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3041 /* make sure whole request will fit in a chunk - if chunks
3044 if (conf->geo.near_copies < conf->geo.raid_disks &&
3045 max_sector > (sector_nr | chunk_mask))
3046 max_sector = (sector_nr | chunk_mask) + 1;
3049 * If there is non-resync activity waiting for a turn, then let it
3050 * though before starting on this new sync request.
3052 if (conf->nr_waiting)
3053 schedule_timeout_uninterruptible(1);
3055 /* Again, very different code for resync and recovery.
3056 * Both must result in an r10bio with a list of bios that
3057 * have bi_end_io, bi_sector, bi_disk set,
3058 * and bi_private set to the r10bio.
3059 * For recovery, we may actually create several r10bios
3060 * with 2 bios in each, that correspond to the bios in the main one.
3061 * In this case, the subordinate r10bios link back through a
3062 * borrowed master_bio pointer, and the counter in the master
3063 * includes a ref from each subordinate.
3065 /* First, we decide what to do and set ->bi_end_io
3066 * To end_sync_read if we want to read, and
3067 * end_sync_write if we will want to write.
3070 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3071 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3072 /* recovery... the complicated one */
3076 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3082 struct raid10_info *mirror = &conf->mirrors[i];
3083 struct md_rdev *mrdev, *mreplace;
3086 mrdev = rcu_dereference(mirror->rdev);
3087 mreplace = rcu_dereference(mirror->replacement);
3089 if ((mrdev == NULL ||
3090 test_bit(Faulty, &mrdev->flags) ||
3091 test_bit(In_sync, &mrdev->flags)) &&
3092 (mreplace == NULL ||
3093 test_bit(Faulty, &mreplace->flags))) {
3099 /* want to reconstruct this device */
3101 sect = raid10_find_virt(conf, sector_nr, i);
3102 if (sect >= mddev->resync_max_sectors) {
3103 /* last stripe is not complete - don't
3104 * try to recover this sector.
3109 if (mreplace && test_bit(Faulty, &mreplace->flags))
3111 /* Unless we are doing a full sync, or a replacement
3112 * we only need to recover the block if it is set in
3115 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3117 if (sync_blocks < max_sync)
3118 max_sync = sync_blocks;
3122 /* yep, skip the sync_blocks here, but don't assume
3123 * that there will never be anything to do here
3125 chunks_skipped = -1;
3129 atomic_inc(&mrdev->nr_pending);
3131 atomic_inc(&mreplace->nr_pending);
3134 r10_bio = raid10_alloc_init_r10buf(conf);
3136 raise_barrier(conf, rb2 != NULL);
3137 atomic_set(&r10_bio->remaining, 0);
3139 r10_bio->master_bio = (struct bio*)rb2;
3141 atomic_inc(&rb2->remaining);
3142 r10_bio->mddev = mddev;
3143 set_bit(R10BIO_IsRecover, &r10_bio->state);
3144 r10_bio->sector = sect;
3146 raid10_find_phys(conf, r10_bio);
3148 /* Need to check if the array will still be
3152 for (j = 0; j < conf->geo.raid_disks; j++) {
3153 struct md_rdev *rdev = rcu_dereference(
3154 conf->mirrors[j].rdev);
3155 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3161 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3162 &sync_blocks, still_degraded);
3165 for (j=0; j<conf->copies;j++) {
3167 int d = r10_bio->devs[j].devnum;
3168 sector_t from_addr, to_addr;
3169 struct md_rdev *rdev =
3170 rcu_dereference(conf->mirrors[d].rdev);
3171 sector_t sector, first_bad;
3174 !test_bit(In_sync, &rdev->flags))
3176 /* This is where we read from */
3178 sector = r10_bio->devs[j].addr;
3180 if (is_badblock(rdev, sector, max_sync,
3181 &first_bad, &bad_sectors)) {
3182 if (first_bad > sector)
3183 max_sync = first_bad - sector;
3185 bad_sectors -= (sector
3187 if (max_sync > bad_sectors)
3188 max_sync = bad_sectors;
3192 bio = r10_bio->devs[0].bio;
3193 bio->bi_next = biolist;
3195 bio->bi_end_io = end_sync_read;
3196 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3197 if (test_bit(FailFast, &rdev->flags))
3198 bio->bi_opf |= MD_FAILFAST;
3199 from_addr = r10_bio->devs[j].addr;
3200 bio->bi_iter.bi_sector = from_addr +
3202 bio_set_dev(bio, rdev->bdev);
3203 atomic_inc(&rdev->nr_pending);
3204 /* and we write to 'i' (if not in_sync) */
3206 for (k=0; k<conf->copies; k++)
3207 if (r10_bio->devs[k].devnum == i)
3209 BUG_ON(k == conf->copies);
3210 to_addr = r10_bio->devs[k].addr;
3211 r10_bio->devs[0].devnum = d;
3212 r10_bio->devs[0].addr = from_addr;
3213 r10_bio->devs[1].devnum = i;
3214 r10_bio->devs[1].addr = to_addr;
3216 if (!test_bit(In_sync, &mrdev->flags)) {
3217 bio = r10_bio->devs[1].bio;
3218 bio->bi_next = biolist;
3220 bio->bi_end_io = end_sync_write;
3221 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3222 bio->bi_iter.bi_sector = to_addr
3223 + mrdev->data_offset;
3224 bio_set_dev(bio, mrdev->bdev);
3225 atomic_inc(&r10_bio->remaining);
3227 r10_bio->devs[1].bio->bi_end_io = NULL;
3229 /* and maybe write to replacement */
3230 bio = r10_bio->devs[1].repl_bio;
3232 bio->bi_end_io = NULL;
3233 /* Note: if mreplace != NULL, then bio
3234 * cannot be NULL as r10buf_pool_alloc will
3235 * have allocated it.
3236 * So the second test here is pointless.
3237 * But it keeps semantic-checkers happy, and
3238 * this comment keeps human reviewers
3241 if (mreplace == NULL || bio == NULL ||
3242 test_bit(Faulty, &mreplace->flags))
3244 bio->bi_next = biolist;
3246 bio->bi_end_io = end_sync_write;
3247 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3248 bio->bi_iter.bi_sector = to_addr +
3249 mreplace->data_offset;
3250 bio_set_dev(bio, mreplace->bdev);
3251 atomic_inc(&r10_bio->remaining);
3255 if (j == conf->copies) {
3256 /* Cannot recover, so abort the recovery or
3257 * record a bad block */
3259 /* problem is that there are bad blocks
3260 * on other device(s)
3263 for (k = 0; k < conf->copies; k++)
3264 if (r10_bio->devs[k].devnum == i)
3266 if (!test_bit(In_sync,
3268 && !rdev_set_badblocks(
3270 r10_bio->devs[k].addr,
3274 !rdev_set_badblocks(
3276 r10_bio->devs[k].addr,
3281 if (!test_and_set_bit(MD_RECOVERY_INTR,
3283 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3285 mirror->recovery_disabled
3286 = mddev->recovery_disabled;
3290 atomic_dec(&rb2->remaining);
3292 rdev_dec_pending(mrdev, mddev);
3294 rdev_dec_pending(mreplace, mddev);
3297 rdev_dec_pending(mrdev, mddev);
3299 rdev_dec_pending(mreplace, mddev);
3300 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3301 /* Only want this if there is elsewhere to
3302 * read from. 'j' is currently the first
3306 for (; j < conf->copies; j++) {
3307 int d = r10_bio->devs[j].devnum;
3308 if (conf->mirrors[d].rdev &&
3310 &conf->mirrors[d].rdev->flags))
3314 r10_bio->devs[0].bio->bi_opf
3318 if (biolist == NULL) {
3320 struct r10bio *rb2 = r10_bio;
3321 r10_bio = (struct r10bio*) rb2->master_bio;
3322 rb2->master_bio = NULL;
3328 /* resync. Schedule a read for every block at this virt offset */
3332 * Since curr_resync_completed could probably not update in
3333 * time, and we will set cluster_sync_low based on it.
3334 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3335 * safety reason, which ensures curr_resync_completed is
3336 * updated in bitmap_cond_end_sync.
3338 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3339 mddev_is_clustered(mddev) &&
3340 (sector_nr + 2 * RESYNC_SECTORS >
3341 conf->cluster_sync_high));
3343 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3344 &sync_blocks, mddev->degraded) &&
3345 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3346 &mddev->recovery)) {
3347 /* We can skip this block */
3349 return sync_blocks + sectors_skipped;
3351 if (sync_blocks < max_sync)
3352 max_sync = sync_blocks;
3353 r10_bio = raid10_alloc_init_r10buf(conf);
3356 r10_bio->mddev = mddev;
3357 atomic_set(&r10_bio->remaining, 0);
3358 raise_barrier(conf, 0);
3359 conf->next_resync = sector_nr;
3361 r10_bio->master_bio = NULL;
3362 r10_bio->sector = sector_nr;
3363 set_bit(R10BIO_IsSync, &r10_bio->state);
3364 raid10_find_phys(conf, r10_bio);
3365 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3367 for (i = 0; i < conf->copies; i++) {
3368 int d = r10_bio->devs[i].devnum;
3369 sector_t first_bad, sector;
3371 struct md_rdev *rdev;
3373 if (r10_bio->devs[i].repl_bio)
3374 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3376 bio = r10_bio->devs[i].bio;
3377 bio->bi_status = BLK_STS_IOERR;
3379 rdev = rcu_dereference(conf->mirrors[d].rdev);
3380 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3384 sector = r10_bio->devs[i].addr;
3385 if (is_badblock(rdev, sector, max_sync,
3386 &first_bad, &bad_sectors)) {
3387 if (first_bad > sector)
3388 max_sync = first_bad - sector;
3390 bad_sectors -= (sector - first_bad);
3391 if (max_sync > bad_sectors)
3392 max_sync = bad_sectors;
3397 atomic_inc(&rdev->nr_pending);
3398 atomic_inc(&r10_bio->remaining);
3399 bio->bi_next = biolist;
3401 bio->bi_end_io = end_sync_read;
3402 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3403 if (test_bit(FailFast, &rdev->flags))
3404 bio->bi_opf |= MD_FAILFAST;
3405 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3406 bio_set_dev(bio, rdev->bdev);
3409 rdev = rcu_dereference(conf->mirrors[d].replacement);
3410 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3414 atomic_inc(&rdev->nr_pending);
3416 /* Need to set up for writing to the replacement */
3417 bio = r10_bio->devs[i].repl_bio;
3418 bio->bi_status = BLK_STS_IOERR;
3420 sector = r10_bio->devs[i].addr;
3421 bio->bi_next = biolist;
3423 bio->bi_end_io = end_sync_write;
3424 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3425 if (test_bit(FailFast, &rdev->flags))
3426 bio->bi_opf |= MD_FAILFAST;
3427 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3428 bio_set_dev(bio, rdev->bdev);
3434 for (i=0; i<conf->copies; i++) {
3435 int d = r10_bio->devs[i].devnum;
3436 if (r10_bio->devs[i].bio->bi_end_io)
3437 rdev_dec_pending(conf->mirrors[d].rdev,
3439 if (r10_bio->devs[i].repl_bio &&
3440 r10_bio->devs[i].repl_bio->bi_end_io)
3442 conf->mirrors[d].replacement,
3452 if (sector_nr + max_sync < max_sector)
3453 max_sector = sector_nr + max_sync;
3456 int len = PAGE_SIZE;
3457 if (sector_nr + (len>>9) > max_sector)
3458 len = (max_sector - sector_nr) << 9;
3461 for (bio= biolist ; bio ; bio=bio->bi_next) {
3462 struct resync_pages *rp = get_resync_pages(bio);
3463 page = resync_fetch_page(rp, page_idx);
3465 * won't fail because the vec table is big enough
3466 * to hold all these pages
3468 bio_add_page(bio, page, len, 0);
3470 nr_sectors += len>>9;
3471 sector_nr += len>>9;
3472 } while (++page_idx < RESYNC_PAGES);
3473 r10_bio->sectors = nr_sectors;
3475 if (mddev_is_clustered(mddev) &&
3476 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3477 /* It is resync not recovery */
3478 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3479 conf->cluster_sync_low = mddev->curr_resync_completed;
3480 raid10_set_cluster_sync_high(conf);
3481 /* Send resync message */
3482 md_cluster_ops->resync_info_update(mddev,
3483 conf->cluster_sync_low,
3484 conf->cluster_sync_high);
3486 } else if (mddev_is_clustered(mddev)) {
3487 /* This is recovery not resync */
3488 sector_t sect_va1, sect_va2;
3489 bool broadcast_msg = false;
3491 for (i = 0; i < conf->geo.raid_disks; i++) {
3493 * sector_nr is a device address for recovery, so we
3494 * need translate it to array address before compare
3495 * with cluster_sync_high.
3497 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3499 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3500 broadcast_msg = true;
3502 * curr_resync_completed is similar as
3503 * sector_nr, so make the translation too.
3505 sect_va2 = raid10_find_virt(conf,
3506 mddev->curr_resync_completed, i);
3508 if (conf->cluster_sync_low == 0 ||
3509 conf->cluster_sync_low > sect_va2)
3510 conf->cluster_sync_low = sect_va2;
3513 if (broadcast_msg) {
3514 raid10_set_cluster_sync_high(conf);
3515 md_cluster_ops->resync_info_update(mddev,
3516 conf->cluster_sync_low,
3517 conf->cluster_sync_high);
3523 biolist = biolist->bi_next;
3525 bio->bi_next = NULL;
3526 r10_bio = get_resync_r10bio(bio);
3527 r10_bio->sectors = nr_sectors;
3529 if (bio->bi_end_io == end_sync_read) {
3530 md_sync_acct_bio(bio, nr_sectors);
3532 generic_make_request(bio);
3536 if (sectors_skipped)
3537 /* pretend they weren't skipped, it makes
3538 * no important difference in this case
3540 md_done_sync(mddev, sectors_skipped, 1);
3542 return sectors_skipped + nr_sectors;
3544 /* There is nowhere to write, so all non-sync
3545 * drives must be failed or in resync, all drives
3546 * have a bad block, so try the next chunk...
3548 if (sector_nr + max_sync < max_sector)
3549 max_sector = sector_nr + max_sync;
3551 sectors_skipped += (max_sector - sector_nr);
3553 sector_nr = max_sector;
3558 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3561 struct r10conf *conf = mddev->private;
3564 raid_disks = min(conf->geo.raid_disks,
3565 conf->prev.raid_disks);
3567 sectors = conf->dev_sectors;
3569 size = sectors >> conf->geo.chunk_shift;
3570 sector_div(size, conf->geo.far_copies);
3571 size = size * raid_disks;
3572 sector_div(size, conf->geo.near_copies);
3574 return size << conf->geo.chunk_shift;
3577 static void calc_sectors(struct r10conf *conf, sector_t size)
3579 /* Calculate the number of sectors-per-device that will
3580 * actually be used, and set conf->dev_sectors and
3584 size = size >> conf->geo.chunk_shift;
3585 sector_div(size, conf->geo.far_copies);
3586 size = size * conf->geo.raid_disks;
3587 sector_div(size, conf->geo.near_copies);
3588 /* 'size' is now the number of chunks in the array */
3589 /* calculate "used chunks per device" */
3590 size = size * conf->copies;
3592 /* We need to round up when dividing by raid_disks to
3593 * get the stride size.
3595 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3597 conf->dev_sectors = size << conf->geo.chunk_shift;
3599 if (conf->geo.far_offset)
3600 conf->geo.stride = 1 << conf->geo.chunk_shift;
3602 sector_div(size, conf->geo.far_copies);
3603 conf->geo.stride = size << conf->geo.chunk_shift;
3607 enum geo_type {geo_new, geo_old, geo_start};
3608 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3611 int layout, chunk, disks;
3614 layout = mddev->layout;
3615 chunk = mddev->chunk_sectors;
3616 disks = mddev->raid_disks - mddev->delta_disks;
3619 layout = mddev->new_layout;
3620 chunk = mddev->new_chunk_sectors;
3621 disks = mddev->raid_disks;
3623 default: /* avoid 'may be unused' warnings */
3624 case geo_start: /* new when starting reshape - raid_disks not
3626 layout = mddev->new_layout;
3627 chunk = mddev->new_chunk_sectors;
3628 disks = mddev->raid_disks + mddev->delta_disks;
3633 if (chunk < (PAGE_SIZE >> 9) ||
3634 !is_power_of_2(chunk))
3637 fc = (layout >> 8) & 255;
3638 fo = layout & (1<<16);
3639 geo->raid_disks = disks;
3640 geo->near_copies = nc;
3641 geo->far_copies = fc;
3642 geo->far_offset = fo;
3643 switch (layout >> 17) {
3644 case 0: /* original layout. simple but not always optimal */
3645 geo->far_set_size = disks;
3647 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3648 * actually using this, but leave code here just in case.*/
3649 geo->far_set_size = disks/fc;
3650 WARN(geo->far_set_size < fc,
3651 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3653 case 2: /* "improved" layout fixed to match documentation */
3654 geo->far_set_size = fc * nc;
3656 default: /* Not a valid layout */
3659 geo->chunk_mask = chunk - 1;
3660 geo->chunk_shift = ffz(~chunk);
3664 static struct r10conf *setup_conf(struct mddev *mddev)
3666 struct r10conf *conf = NULL;
3671 copies = setup_geo(&geo, mddev, geo_new);
3674 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3675 mdname(mddev), PAGE_SIZE);
3679 if (copies < 2 || copies > mddev->raid_disks) {
3680 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3681 mdname(mddev), mddev->new_layout);
3686 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3690 /* FIXME calc properly */
3691 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3692 sizeof(struct raid10_info),
3697 conf->tmppage = alloc_page(GFP_KERNEL);
3702 conf->copies = copies;
3703 err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3704 r10bio_pool_free, conf);
3708 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3712 calc_sectors(conf, mddev->dev_sectors);
3713 if (mddev->reshape_position == MaxSector) {
3714 conf->prev = conf->geo;
3715 conf->reshape_progress = MaxSector;
3717 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3721 conf->reshape_progress = mddev->reshape_position;
3722 if (conf->prev.far_offset)
3723 conf->prev.stride = 1 << conf->prev.chunk_shift;
3725 /* far_copies must be 1 */
3726 conf->prev.stride = conf->dev_sectors;
3728 conf->reshape_safe = conf->reshape_progress;
3729 spin_lock_init(&conf->device_lock);
3730 INIT_LIST_HEAD(&conf->retry_list);
3731 INIT_LIST_HEAD(&conf->bio_end_io_list);
3733 spin_lock_init(&conf->resync_lock);
3734 init_waitqueue_head(&conf->wait_barrier);
3735 atomic_set(&conf->nr_pending, 0);
3738 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3742 conf->mddev = mddev;
3747 mempool_exit(&conf->r10bio_pool);
3748 kfree(conf->mirrors);
3749 safe_put_page(conf->tmppage);
3750 bioset_exit(&conf->bio_split);
3753 return ERR_PTR(err);
3756 static int raid10_run(struct mddev *mddev)
3758 struct r10conf *conf;
3759 int i, disk_idx, chunk_size;
3760 struct raid10_info *disk;
3761 struct md_rdev *rdev;
3763 sector_t min_offset_diff = 0;
3765 bool discard_supported = false;
3767 if (mddev_init_writes_pending(mddev) < 0)
3770 if (mddev->private == NULL) {
3771 conf = setup_conf(mddev);
3773 return PTR_ERR(conf);
3774 mddev->private = conf;
3776 conf = mddev->private;
3780 if (mddev_is_clustered(conf->mddev)) {
3783 fc = (mddev->layout >> 8) & 255;
3784 fo = mddev->layout & (1<<16);
3785 if (fc > 1 || fo > 0) {
3786 pr_err("only near layout is supported by clustered"
3792 mddev->thread = conf->thread;
3793 conf->thread = NULL;
3795 chunk_size = mddev->chunk_sectors << 9;
3797 blk_queue_max_discard_sectors(mddev->queue,
3798 mddev->chunk_sectors);
3799 blk_queue_max_write_same_sectors(mddev->queue, 0);
3800 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3801 blk_queue_io_min(mddev->queue, chunk_size);
3802 if (conf->geo.raid_disks % conf->geo.near_copies)
3803 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3805 blk_queue_io_opt(mddev->queue, chunk_size *
3806 (conf->geo.raid_disks / conf->geo.near_copies));
3809 rdev_for_each(rdev, mddev) {
3812 disk_idx = rdev->raid_disk;
3815 if (disk_idx >= conf->geo.raid_disks &&
3816 disk_idx >= conf->prev.raid_disks)
3818 disk = conf->mirrors + disk_idx;
3820 if (test_bit(Replacement, &rdev->flags)) {
3821 if (disk->replacement)
3823 disk->replacement = rdev;
3829 diff = (rdev->new_data_offset - rdev->data_offset);
3830 if (!mddev->reshape_backwards)
3834 if (first || diff < min_offset_diff)
3835 min_offset_diff = diff;
3838 disk_stack_limits(mddev->gendisk, rdev->bdev,
3839 rdev->data_offset << 9);
3841 disk->head_position = 0;
3843 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3844 discard_supported = true;
3849 if (discard_supported)
3850 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3853 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3856 /* need to check that every block has at least one working mirror */
3857 if (!enough(conf, -1)) {
3858 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3863 if (conf->reshape_progress != MaxSector) {
3864 /* must ensure that shape change is supported */
3865 if (conf->geo.far_copies != 1 &&
3866 conf->geo.far_offset == 0)
3868 if (conf->prev.far_copies != 1 &&
3869 conf->prev.far_offset == 0)
3873 mddev->degraded = 0;
3875 i < conf->geo.raid_disks
3876 || i < conf->prev.raid_disks;
3879 disk = conf->mirrors + i;
3881 if (!disk->rdev && disk->replacement) {
3882 /* The replacement is all we have - use it */
3883 disk->rdev = disk->replacement;
3884 disk->replacement = NULL;
3885 clear_bit(Replacement, &disk->rdev->flags);
3889 !test_bit(In_sync, &disk->rdev->flags)) {
3890 disk->head_position = 0;
3893 disk->rdev->saved_raid_disk < 0)
3896 disk->recovery_disabled = mddev->recovery_disabled - 1;
3899 if (mddev->recovery_cp != MaxSector)
3900 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3902 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3903 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3904 conf->geo.raid_disks);
3906 * Ok, everything is just fine now
3908 mddev->dev_sectors = conf->dev_sectors;
3909 size = raid10_size(mddev, 0, 0);
3910 md_set_array_sectors(mddev, size);
3911 mddev->resync_max_sectors = size;
3912 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3915 int stripe = conf->geo.raid_disks *
3916 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3918 /* Calculate max read-ahead size.
3919 * We need to readahead at least twice a whole stripe....
3922 stripe /= conf->geo.near_copies;
3923 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3924 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3927 if (md_integrity_register(mddev))
3930 if (conf->reshape_progress != MaxSector) {
3931 unsigned long before_length, after_length;
3933 before_length = ((1 << conf->prev.chunk_shift) *
3934 conf->prev.far_copies);
3935 after_length = ((1 << conf->geo.chunk_shift) *
3936 conf->geo.far_copies);
3938 if (max(before_length, after_length) > min_offset_diff) {
3939 /* This cannot work */
3940 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3943 conf->offset_diff = min_offset_diff;
3945 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3946 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3947 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3948 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3949 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3956 md_unregister_thread(&mddev->thread);
3957 mempool_exit(&conf->r10bio_pool);
3958 safe_put_page(conf->tmppage);
3959 kfree(conf->mirrors);
3961 mddev->private = NULL;
3966 static void raid10_free(struct mddev *mddev, void *priv)
3968 struct r10conf *conf = priv;
3970 mempool_exit(&conf->r10bio_pool);
3971 safe_put_page(conf->tmppage);
3972 kfree(conf->mirrors);
3973 kfree(conf->mirrors_old);
3974 kfree(conf->mirrors_new);
3975 bioset_exit(&conf->bio_split);
3979 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3981 struct r10conf *conf = mddev->private;
3984 raise_barrier(conf, 0);
3986 lower_barrier(conf);
3989 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3991 /* Resize of 'far' arrays is not supported.
3992 * For 'near' and 'offset' arrays we can set the
3993 * number of sectors used to be an appropriate multiple
3994 * of the chunk size.
3995 * For 'offset', this is far_copies*chunksize.
3996 * For 'near' the multiplier is the LCM of
3997 * near_copies and raid_disks.
3998 * So if far_copies > 1 && !far_offset, fail.
3999 * Else find LCM(raid_disks, near_copy)*far_copies and
4000 * multiply by chunk_size. Then round to this number.
4001 * This is mostly done by raid10_size()
4003 struct r10conf *conf = mddev->private;
4004 sector_t oldsize, size;
4006 if (mddev->reshape_position != MaxSector)
4009 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4012 oldsize = raid10_size(mddev, 0, 0);
4013 size = raid10_size(mddev, sectors, 0);
4014 if (mddev->external_size &&
4015 mddev->array_sectors > size)
4017 if (mddev->bitmap) {
4018 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
4022 md_set_array_sectors(mddev, size);
4023 if (sectors > mddev->dev_sectors &&
4024 mddev->recovery_cp > oldsize) {
4025 mddev->recovery_cp = oldsize;
4026 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4028 calc_sectors(conf, sectors);
4029 mddev->dev_sectors = conf->dev_sectors;
4030 mddev->resync_max_sectors = size;
4034 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4036 struct md_rdev *rdev;
4037 struct r10conf *conf;
4039 if (mddev->degraded > 0) {
4040 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4042 return ERR_PTR(-EINVAL);
4044 sector_div(size, devs);
4046 /* Set new parameters */
4047 mddev->new_level = 10;
4048 /* new layout: far_copies = 1, near_copies = 2 */
4049 mddev->new_layout = (1<<8) + 2;
4050 mddev->new_chunk_sectors = mddev->chunk_sectors;
4051 mddev->delta_disks = mddev->raid_disks;
4052 mddev->raid_disks *= 2;
4053 /* make sure it will be not marked as dirty */
4054 mddev->recovery_cp = MaxSector;
4055 mddev->dev_sectors = size;
4057 conf = setup_conf(mddev);
4058 if (!IS_ERR(conf)) {
4059 rdev_for_each(rdev, mddev)
4060 if (rdev->raid_disk >= 0) {
4061 rdev->new_raid_disk = rdev->raid_disk * 2;
4062 rdev->sectors = size;
4070 static void *raid10_takeover(struct mddev *mddev)
4072 struct r0conf *raid0_conf;
4074 /* raid10 can take over:
4075 * raid0 - providing it has only two drives
4077 if (mddev->level == 0) {
4078 /* for raid0 takeover only one zone is supported */
4079 raid0_conf = mddev->private;
4080 if (raid0_conf->nr_strip_zones > 1) {
4081 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4083 return ERR_PTR(-EINVAL);
4085 return raid10_takeover_raid0(mddev,
4086 raid0_conf->strip_zone->zone_end,
4087 raid0_conf->strip_zone->nb_dev);
4089 return ERR_PTR(-EINVAL);
4092 static int raid10_check_reshape(struct mddev *mddev)
4094 /* Called when there is a request to change
4095 * - layout (to ->new_layout)
4096 * - chunk size (to ->new_chunk_sectors)
4097 * - raid_disks (by delta_disks)
4098 * or when trying to restart a reshape that was ongoing.
4100 * We need to validate the request and possibly allocate
4101 * space if that might be an issue later.
4103 * Currently we reject any reshape of a 'far' mode array,
4104 * allow chunk size to change if new is generally acceptable,
4105 * allow raid_disks to increase, and allow
4106 * a switch between 'near' mode and 'offset' mode.
4108 struct r10conf *conf = mddev->private;
4111 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4114 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4115 /* mustn't change number of copies */
4117 if (geo.far_copies > 1 && !geo.far_offset)
4118 /* Cannot switch to 'far' mode */
4121 if (mddev->array_sectors & geo.chunk_mask)
4122 /* not factor of array size */
4125 if (!enough(conf, -1))
4128 kfree(conf->mirrors_new);
4129 conf->mirrors_new = NULL;
4130 if (mddev->delta_disks > 0) {
4131 /* allocate new 'mirrors' list */
4133 kcalloc(mddev->raid_disks + mddev->delta_disks,
4134 sizeof(struct raid10_info),
4136 if (!conf->mirrors_new)
4143 * Need to check if array has failed when deciding whether to:
4145 * - remove non-faulty devices
4148 * This determination is simple when no reshape is happening.
4149 * However if there is a reshape, we need to carefully check
4150 * both the before and after sections.
4151 * This is because some failed devices may only affect one
4152 * of the two sections, and some non-in_sync devices may
4153 * be insync in the section most affected by failed devices.
4155 static int calc_degraded(struct r10conf *conf)
4157 int degraded, degraded2;
4162 /* 'prev' section first */
4163 for (i = 0; i < conf->prev.raid_disks; i++) {
4164 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4165 if (!rdev || test_bit(Faulty, &rdev->flags))
4167 else if (!test_bit(In_sync, &rdev->flags))
4168 /* When we can reduce the number of devices in
4169 * an array, this might not contribute to
4170 * 'degraded'. It does now.
4175 if (conf->geo.raid_disks == conf->prev.raid_disks)
4179 for (i = 0; i < conf->geo.raid_disks; i++) {
4180 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4181 if (!rdev || test_bit(Faulty, &rdev->flags))
4183 else if (!test_bit(In_sync, &rdev->flags)) {
4184 /* If reshape is increasing the number of devices,
4185 * this section has already been recovered, so
4186 * it doesn't contribute to degraded.
4189 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4194 if (degraded2 > degraded)
4199 static int raid10_start_reshape(struct mddev *mddev)
4201 /* A 'reshape' has been requested. This commits
4202 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4203 * This also checks if there are enough spares and adds them
4205 * We currently require enough spares to make the final
4206 * array non-degraded. We also require that the difference
4207 * between old and new data_offset - on each device - is
4208 * enough that we never risk over-writing.
4211 unsigned long before_length, after_length;
4212 sector_t min_offset_diff = 0;
4215 struct r10conf *conf = mddev->private;
4216 struct md_rdev *rdev;
4220 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4223 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4226 before_length = ((1 << conf->prev.chunk_shift) *
4227 conf->prev.far_copies);
4228 after_length = ((1 << conf->geo.chunk_shift) *
4229 conf->geo.far_copies);
4231 rdev_for_each(rdev, mddev) {
4232 if (!test_bit(In_sync, &rdev->flags)
4233 && !test_bit(Faulty, &rdev->flags))
4235 if (rdev->raid_disk >= 0) {
4236 long long diff = (rdev->new_data_offset
4237 - rdev->data_offset);
4238 if (!mddev->reshape_backwards)
4242 if (first || diff < min_offset_diff)
4243 min_offset_diff = diff;
4248 if (max(before_length, after_length) > min_offset_diff)
4251 if (spares < mddev->delta_disks)
4254 conf->offset_diff = min_offset_diff;
4255 spin_lock_irq(&conf->device_lock);
4256 if (conf->mirrors_new) {
4257 memcpy(conf->mirrors_new, conf->mirrors,
4258 sizeof(struct raid10_info)*conf->prev.raid_disks);
4260 kfree(conf->mirrors_old);
4261 conf->mirrors_old = conf->mirrors;
4262 conf->mirrors = conf->mirrors_new;
4263 conf->mirrors_new = NULL;
4265 setup_geo(&conf->geo, mddev, geo_start);
4267 if (mddev->reshape_backwards) {
4268 sector_t size = raid10_size(mddev, 0, 0);
4269 if (size < mddev->array_sectors) {
4270 spin_unlock_irq(&conf->device_lock);
4271 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4275 mddev->resync_max_sectors = size;
4276 conf->reshape_progress = size;
4278 conf->reshape_progress = 0;
4279 conf->reshape_safe = conf->reshape_progress;
4280 spin_unlock_irq(&conf->device_lock);
4282 if (mddev->delta_disks && mddev->bitmap) {
4283 ret = bitmap_resize(mddev->bitmap,
4284 raid10_size(mddev, 0,
4285 conf->geo.raid_disks),
4290 if (mddev->delta_disks > 0) {
4291 rdev_for_each(rdev, mddev)
4292 if (rdev->raid_disk < 0 &&
4293 !test_bit(Faulty, &rdev->flags)) {
4294 if (raid10_add_disk(mddev, rdev) == 0) {
4295 if (rdev->raid_disk >=
4296 conf->prev.raid_disks)
4297 set_bit(In_sync, &rdev->flags);
4299 rdev->recovery_offset = 0;
4301 if (sysfs_link_rdev(mddev, rdev))
4302 /* Failure here is OK */;
4304 } else if (rdev->raid_disk >= conf->prev.raid_disks
4305 && !test_bit(Faulty, &rdev->flags)) {
4306 /* This is a spare that was manually added */
4307 set_bit(In_sync, &rdev->flags);
4310 /* When a reshape changes the number of devices,
4311 * ->degraded is measured against the larger of the
4312 * pre and post numbers.
4314 spin_lock_irq(&conf->device_lock);
4315 mddev->degraded = calc_degraded(conf);
4316 spin_unlock_irq(&conf->device_lock);
4317 mddev->raid_disks = conf->geo.raid_disks;
4318 mddev->reshape_position = conf->reshape_progress;
4319 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4321 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4322 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4323 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4324 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4325 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4327 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4329 if (!mddev->sync_thread) {
4333 conf->reshape_checkpoint = jiffies;
4334 md_wakeup_thread(mddev->sync_thread);
4335 md_new_event(mddev);
4339 mddev->recovery = 0;
4340 spin_lock_irq(&conf->device_lock);
4341 conf->geo = conf->prev;
4342 mddev->raid_disks = conf->geo.raid_disks;
4343 rdev_for_each(rdev, mddev)
4344 rdev->new_data_offset = rdev->data_offset;
4346 conf->reshape_progress = MaxSector;
4347 conf->reshape_safe = MaxSector;
4348 mddev->reshape_position = MaxSector;
4349 spin_unlock_irq(&conf->device_lock);
4353 /* Calculate the last device-address that could contain
4354 * any block from the chunk that includes the array-address 's'
4355 * and report the next address.
4356 * i.e. the address returned will be chunk-aligned and after
4357 * any data that is in the chunk containing 's'.
4359 static sector_t last_dev_address(sector_t s, struct geom *geo)
4361 s = (s | geo->chunk_mask) + 1;
4362 s >>= geo->chunk_shift;
4363 s *= geo->near_copies;
4364 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4365 s *= geo->far_copies;
4366 s <<= geo->chunk_shift;
4370 /* Calculate the first device-address that could contain
4371 * any block from the chunk that includes the array-address 's'.
4372 * This too will be the start of a chunk
4374 static sector_t first_dev_address(sector_t s, struct geom *geo)
4376 s >>= geo->chunk_shift;
4377 s *= geo->near_copies;
4378 sector_div(s, geo->raid_disks);
4379 s *= geo->far_copies;
4380 s <<= geo->chunk_shift;
4384 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4387 /* We simply copy at most one chunk (smallest of old and new)
4388 * at a time, possibly less if that exceeds RESYNC_PAGES,
4389 * or we hit a bad block or something.
4390 * This might mean we pause for normal IO in the middle of
4391 * a chunk, but that is not a problem as mddev->reshape_position
4392 * can record any location.
4394 * If we will want to write to a location that isn't
4395 * yet recorded as 'safe' (i.e. in metadata on disk) then
4396 * we need to flush all reshape requests and update the metadata.
4398 * When reshaping forwards (e.g. to more devices), we interpret
4399 * 'safe' as the earliest block which might not have been copied
4400 * down yet. We divide this by previous stripe size and multiply
4401 * by previous stripe length to get lowest device offset that we
4402 * cannot write to yet.
4403 * We interpret 'sector_nr' as an address that we want to write to.
4404 * From this we use last_device_address() to find where we might
4405 * write to, and first_device_address on the 'safe' position.
4406 * If this 'next' write position is after the 'safe' position,
4407 * we must update the metadata to increase the 'safe' position.
4409 * When reshaping backwards, we round in the opposite direction
4410 * and perform the reverse test: next write position must not be
4411 * less than current safe position.
4413 * In all this the minimum difference in data offsets
4414 * (conf->offset_diff - always positive) allows a bit of slack,
4415 * so next can be after 'safe', but not by more than offset_diff
4417 * We need to prepare all the bios here before we start any IO
4418 * to ensure the size we choose is acceptable to all devices.
4419 * The means one for each copy for write-out and an extra one for
4421 * We store the read-in bio in ->master_bio and the others in
4422 * ->devs[x].bio and ->devs[x].repl_bio.
4424 struct r10conf *conf = mddev->private;
4425 struct r10bio *r10_bio;
4426 sector_t next, safe, last;
4430 struct md_rdev *rdev;
4433 struct bio *bio, *read_bio;
4434 int sectors_done = 0;
4435 struct page **pages;
4437 if (sector_nr == 0) {
4438 /* If restarting in the middle, skip the initial sectors */
4439 if (mddev->reshape_backwards &&
4440 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4441 sector_nr = (raid10_size(mddev, 0, 0)
4442 - conf->reshape_progress);
4443 } else if (!mddev->reshape_backwards &&
4444 conf->reshape_progress > 0)
4445 sector_nr = conf->reshape_progress;
4447 mddev->curr_resync_completed = sector_nr;
4448 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4454 /* We don't use sector_nr to track where we are up to
4455 * as that doesn't work well for ->reshape_backwards.
4456 * So just use ->reshape_progress.
4458 if (mddev->reshape_backwards) {
4459 /* 'next' is the earliest device address that we might
4460 * write to for this chunk in the new layout
4462 next = first_dev_address(conf->reshape_progress - 1,
4465 /* 'safe' is the last device address that we might read from
4466 * in the old layout after a restart
4468 safe = last_dev_address(conf->reshape_safe - 1,
4471 if (next + conf->offset_diff < safe)
4474 last = conf->reshape_progress - 1;
4475 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4476 & conf->prev.chunk_mask);
4477 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4478 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4480 /* 'next' is after the last device address that we
4481 * might write to for this chunk in the new layout
4483 next = last_dev_address(conf->reshape_progress, &conf->geo);
4485 /* 'safe' is the earliest device address that we might
4486 * read from in the old layout after a restart
4488 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4490 /* Need to update metadata if 'next' might be beyond 'safe'
4491 * as that would possibly corrupt data
4493 if (next > safe + conf->offset_diff)
4496 sector_nr = conf->reshape_progress;
4497 last = sector_nr | (conf->geo.chunk_mask
4498 & conf->prev.chunk_mask);
4500 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4501 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4505 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4506 /* Need to update reshape_position in metadata */
4508 mddev->reshape_position = conf->reshape_progress;
4509 if (mddev->reshape_backwards)
4510 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4511 - conf->reshape_progress;
4513 mddev->curr_resync_completed = conf->reshape_progress;
4514 conf->reshape_checkpoint = jiffies;
4515 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4516 md_wakeup_thread(mddev->thread);
4517 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4518 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4519 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4520 allow_barrier(conf);
4521 return sectors_done;
4523 conf->reshape_safe = mddev->reshape_position;
4524 allow_barrier(conf);
4528 /* Now schedule reads for blocks from sector_nr to last */
4529 r10_bio = raid10_alloc_init_r10buf(conf);
4531 raise_barrier(conf, sectors_done != 0);
4532 atomic_set(&r10_bio->remaining, 0);
4533 r10_bio->mddev = mddev;
4534 r10_bio->sector = sector_nr;
4535 set_bit(R10BIO_IsReshape, &r10_bio->state);
4536 r10_bio->sectors = last - sector_nr + 1;
4537 rdev = read_balance(conf, r10_bio, &max_sectors);
4538 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4541 /* Cannot read from here, so need to record bad blocks
4542 * on all the target devices.
4545 mempool_free(r10_bio, &conf->r10buf_pool);
4546 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4547 return sectors_done;
4550 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4552 bio_set_dev(read_bio, rdev->bdev);
4553 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4554 + rdev->data_offset);
4555 read_bio->bi_private = r10_bio;
4556 read_bio->bi_end_io = end_reshape_read;
4557 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4558 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4559 read_bio->bi_status = 0;
4560 read_bio->bi_vcnt = 0;
4561 read_bio->bi_iter.bi_size = 0;
4562 r10_bio->master_bio = read_bio;
4563 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4565 /* Now find the locations in the new layout */
4566 __raid10_find_phys(&conf->geo, r10_bio);
4569 read_bio->bi_next = NULL;
4572 for (s = 0; s < conf->copies*2; s++) {
4574 int d = r10_bio->devs[s/2].devnum;
4575 struct md_rdev *rdev2;
4577 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4578 b = r10_bio->devs[s/2].repl_bio;
4580 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4581 b = r10_bio->devs[s/2].bio;
4583 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4586 bio_set_dev(b, rdev2->bdev);
4587 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4588 rdev2->new_data_offset;
4589 b->bi_end_io = end_reshape_write;
4590 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4595 /* Now add as many pages as possible to all of these bios. */
4598 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4599 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4600 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4601 int len = (max_sectors - s) << 9;
4602 if (len > PAGE_SIZE)
4604 for (bio = blist; bio ; bio = bio->bi_next) {
4606 * won't fail because the vec table is big enough
4607 * to hold all these pages
4609 bio_add_page(bio, page, len, 0);
4611 sector_nr += len >> 9;
4612 nr_sectors += len >> 9;
4615 r10_bio->sectors = nr_sectors;
4617 /* Now submit the read */
4618 md_sync_acct_bio(read_bio, r10_bio->sectors);
4619 atomic_inc(&r10_bio->remaining);
4620 read_bio->bi_next = NULL;
4621 generic_make_request(read_bio);
4622 sector_nr += nr_sectors;
4623 sectors_done += nr_sectors;
4624 if (sector_nr <= last)
4627 /* Now that we have done the whole section we can
4628 * update reshape_progress
4630 if (mddev->reshape_backwards)
4631 conf->reshape_progress -= sectors_done;
4633 conf->reshape_progress += sectors_done;
4635 return sectors_done;
4638 static void end_reshape_request(struct r10bio *r10_bio);
4639 static int handle_reshape_read_error(struct mddev *mddev,
4640 struct r10bio *r10_bio);
4641 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4643 /* Reshape read completed. Hopefully we have a block
4645 * If we got a read error then we do sync 1-page reads from
4646 * elsewhere until we find the data - or give up.
4648 struct r10conf *conf = mddev->private;
4651 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4652 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4653 /* Reshape has been aborted */
4654 md_done_sync(mddev, r10_bio->sectors, 0);
4658 /* We definitely have the data in the pages, schedule the
4661 atomic_set(&r10_bio->remaining, 1);
4662 for (s = 0; s < conf->copies*2; s++) {
4664 int d = r10_bio->devs[s/2].devnum;
4665 struct md_rdev *rdev;
4668 rdev = rcu_dereference(conf->mirrors[d].replacement);
4669 b = r10_bio->devs[s/2].repl_bio;
4671 rdev = rcu_dereference(conf->mirrors[d].rdev);
4672 b = r10_bio->devs[s/2].bio;
4674 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4678 atomic_inc(&rdev->nr_pending);
4680 md_sync_acct_bio(b, r10_bio->sectors);
4681 atomic_inc(&r10_bio->remaining);
4683 generic_make_request(b);
4685 end_reshape_request(r10_bio);
4688 static void end_reshape(struct r10conf *conf)
4690 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4693 spin_lock_irq(&conf->device_lock);
4694 conf->prev = conf->geo;
4695 md_finish_reshape(conf->mddev);
4697 conf->reshape_progress = MaxSector;
4698 conf->reshape_safe = MaxSector;
4699 spin_unlock_irq(&conf->device_lock);
4701 /* read-ahead size must cover two whole stripes, which is
4702 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4704 if (conf->mddev->queue) {
4705 int stripe = conf->geo.raid_disks *
4706 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4707 stripe /= conf->geo.near_copies;
4708 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4709 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4714 static int handle_reshape_read_error(struct mddev *mddev,
4715 struct r10bio *r10_bio)
4717 /* Use sync reads to get the blocks from somewhere else */
4718 int sectors = r10_bio->sectors;
4719 struct r10conf *conf = mddev->private;
4720 struct r10bio *r10b;
4723 struct page **pages;
4725 r10b = kmalloc(sizeof(*r10b) +
4726 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4728 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4732 /* reshape IOs share pages from .devs[0].bio */
4733 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4735 r10b->sector = r10_bio->sector;
4736 __raid10_find_phys(&conf->prev, r10b);
4741 int first_slot = slot;
4743 if (s > (PAGE_SIZE >> 9))
4748 int d = r10b->devs[slot].devnum;
4749 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4752 test_bit(Faulty, &rdev->flags) ||
4753 !test_bit(In_sync, &rdev->flags))
4756 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4757 atomic_inc(&rdev->nr_pending);
4759 success = sync_page_io(rdev,
4763 REQ_OP_READ, 0, false);
4764 rdev_dec_pending(rdev, mddev);
4770 if (slot >= conf->copies)
4772 if (slot == first_slot)
4777 /* couldn't read this block, must give up */
4778 set_bit(MD_RECOVERY_INTR,
4790 static void end_reshape_write(struct bio *bio)
4792 struct r10bio *r10_bio = get_resync_r10bio(bio);
4793 struct mddev *mddev = r10_bio->mddev;
4794 struct r10conf *conf = mddev->private;
4798 struct md_rdev *rdev = NULL;
4800 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4802 rdev = conf->mirrors[d].replacement;
4805 rdev = conf->mirrors[d].rdev;
4808 if (bio->bi_status) {
4809 /* FIXME should record badblock */
4810 md_error(mddev, rdev);
4813 rdev_dec_pending(rdev, mddev);
4814 end_reshape_request(r10_bio);
4817 static void end_reshape_request(struct r10bio *r10_bio)
4819 if (!atomic_dec_and_test(&r10_bio->remaining))
4821 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4822 bio_put(r10_bio->master_bio);
4826 static void raid10_finish_reshape(struct mddev *mddev)
4828 struct r10conf *conf = mddev->private;
4830 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4833 if (mddev->delta_disks > 0) {
4834 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4835 mddev->recovery_cp = mddev->resync_max_sectors;
4836 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4838 mddev->resync_max_sectors = mddev->array_sectors;
4842 for (d = conf->geo.raid_disks ;
4843 d < conf->geo.raid_disks - mddev->delta_disks;
4845 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4847 clear_bit(In_sync, &rdev->flags);
4848 rdev = rcu_dereference(conf->mirrors[d].replacement);
4850 clear_bit(In_sync, &rdev->flags);
4854 mddev->layout = mddev->new_layout;
4855 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4856 mddev->reshape_position = MaxSector;
4857 mddev->delta_disks = 0;
4858 mddev->reshape_backwards = 0;
4861 static struct md_personality raid10_personality =
4865 .owner = THIS_MODULE,
4866 .make_request = raid10_make_request,
4868 .free = raid10_free,
4869 .status = raid10_status,
4870 .error_handler = raid10_error,
4871 .hot_add_disk = raid10_add_disk,
4872 .hot_remove_disk= raid10_remove_disk,
4873 .spare_active = raid10_spare_active,
4874 .sync_request = raid10_sync_request,
4875 .quiesce = raid10_quiesce,
4876 .size = raid10_size,
4877 .resize = raid10_resize,
4878 .takeover = raid10_takeover,
4879 .check_reshape = raid10_check_reshape,
4880 .start_reshape = raid10_start_reshape,
4881 .finish_reshape = raid10_finish_reshape,
4882 .congested = raid10_congested,
4885 static int __init raid_init(void)
4887 return register_md_personality(&raid10_personality);
4890 static void raid_exit(void)
4892 unregister_md_personality(&raid10_personality);
4895 module_init(raid_init);
4896 module_exit(raid_exit);
4897 MODULE_LICENSE("GPL");
4898 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4899 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4900 MODULE_ALIAS("md-raid10");
4901 MODULE_ALIAS("md-level-10");
4903 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);