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]);
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 md_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 md_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 md_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 (slot >= 0 && 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 allow_barrier(conf);
1212 generic_make_request(bio);
1215 r10_bio->master_bio = bio;
1216 r10_bio->sectors = max_sectors;
1218 slot = r10_bio->read_slot;
1220 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1222 r10_bio->devs[slot].bio = read_bio;
1223 r10_bio->devs[slot].rdev = rdev;
1225 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1226 choose_data_offset(r10_bio, rdev);
1227 bio_set_dev(read_bio, rdev->bdev);
1228 read_bio->bi_end_io = raid10_end_read_request;
1229 bio_set_op_attrs(read_bio, op, do_sync);
1230 if (test_bit(FailFast, &rdev->flags) &&
1231 test_bit(R10BIO_FailFast, &r10_bio->state))
1232 read_bio->bi_opf |= MD_FAILFAST;
1233 read_bio->bi_private = r10_bio;
1236 trace_block_bio_remap(read_bio->bi_disk->queue,
1237 read_bio, disk_devt(mddev->gendisk),
1239 generic_make_request(read_bio);
1243 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1244 struct bio *bio, bool replacement,
1247 const int op = bio_op(bio);
1248 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1249 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1250 unsigned long flags;
1251 struct blk_plug_cb *cb;
1252 struct raid10_plug_cb *plug = NULL;
1253 struct r10conf *conf = mddev->private;
1254 struct md_rdev *rdev;
1255 int devnum = r10_bio->devs[n_copy].devnum;
1259 rdev = conf->mirrors[devnum].replacement;
1261 /* Replacement just got moved to main 'rdev' */
1263 rdev = conf->mirrors[devnum].rdev;
1266 rdev = conf->mirrors[devnum].rdev;
1268 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1270 r10_bio->devs[n_copy].repl_bio = mbio;
1272 r10_bio->devs[n_copy].bio = mbio;
1274 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1275 choose_data_offset(r10_bio, rdev));
1276 bio_set_dev(mbio, rdev->bdev);
1277 mbio->bi_end_io = raid10_end_write_request;
1278 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1279 if (!replacement && test_bit(FailFast,
1280 &conf->mirrors[devnum].rdev->flags)
1281 && enough(conf, devnum))
1282 mbio->bi_opf |= MD_FAILFAST;
1283 mbio->bi_private = r10_bio;
1285 if (conf->mddev->gendisk)
1286 trace_block_bio_remap(mbio->bi_disk->queue,
1287 mbio, disk_devt(conf->mddev->gendisk),
1289 /* flush_pending_writes() needs access to the rdev so...*/
1290 mbio->bi_disk = (void *)rdev;
1292 atomic_inc(&r10_bio->remaining);
1294 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1296 plug = container_of(cb, struct raid10_plug_cb, cb);
1300 bio_list_add(&plug->pending, mbio);
1301 plug->pending_cnt++;
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 bio_list_add(&conf->pending_bio_list, mbio);
1305 conf->pending_count++;
1306 spin_unlock_irqrestore(&conf->device_lock, flags);
1307 md_wakeup_thread(mddev->thread);
1311 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1312 struct r10bio *r10_bio)
1314 struct r10conf *conf = mddev->private;
1316 struct md_rdev *blocked_rdev;
1320 if ((mddev_is_clustered(mddev) &&
1321 md_cluster_ops->area_resyncing(mddev, WRITE,
1322 bio->bi_iter.bi_sector,
1323 bio_end_sector(bio)))) {
1326 prepare_to_wait(&conf->wait_barrier,
1328 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1329 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1333 finish_wait(&conf->wait_barrier, &w);
1337 * Register the new request and wait if the reconstruction
1338 * thread has put up a bar for new requests.
1339 * Continue immediately if no resync is active currently.
1343 sectors = r10_bio->sectors;
1344 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1345 bio->bi_iter.bi_sector < conf->reshape_progress &&
1346 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1348 * IO spans the reshape position. Need to wait for reshape to
1351 raid10_log(conf->mddev, "wait reshape");
1352 allow_barrier(conf);
1353 wait_event(conf->wait_barrier,
1354 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1355 conf->reshape_progress >= bio->bi_iter.bi_sector +
1360 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1361 (mddev->reshape_backwards
1362 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1363 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1364 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1365 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1366 /* Need to update reshape_position in metadata */
1367 mddev->reshape_position = conf->reshape_progress;
1368 set_mask_bits(&mddev->sb_flags, 0,
1369 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1370 md_wakeup_thread(mddev->thread);
1371 raid10_log(conf->mddev, "wait reshape metadata");
1372 wait_event(mddev->sb_wait,
1373 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1375 conf->reshape_safe = mddev->reshape_position;
1378 if (conf->pending_count >= max_queued_requests) {
1379 md_wakeup_thread(mddev->thread);
1380 raid10_log(mddev, "wait queued");
1381 wait_event(conf->wait_barrier,
1382 conf->pending_count < max_queued_requests);
1384 /* first select target devices under rcu_lock and
1385 * inc refcount on their rdev. Record them by setting
1387 * If there are known/acknowledged bad blocks on any device
1388 * on which we have seen a write error, we want to avoid
1389 * writing to those blocks. This potentially requires several
1390 * writes to write around the bad blocks. Each set of writes
1391 * gets its own r10_bio with a set of bios attached.
1394 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1395 raid10_find_phys(conf, r10_bio);
1397 blocked_rdev = NULL;
1399 max_sectors = r10_bio->sectors;
1401 for (i = 0; i < conf->copies; i++) {
1402 int d = r10_bio->devs[i].devnum;
1403 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1404 struct md_rdev *rrdev = rcu_dereference(
1405 conf->mirrors[d].replacement);
1408 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1409 atomic_inc(&rdev->nr_pending);
1410 blocked_rdev = rdev;
1413 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1414 atomic_inc(&rrdev->nr_pending);
1415 blocked_rdev = rrdev;
1418 if (rdev && (test_bit(Faulty, &rdev->flags)))
1420 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1423 r10_bio->devs[i].bio = NULL;
1424 r10_bio->devs[i].repl_bio = NULL;
1426 if (!rdev && !rrdev) {
1427 set_bit(R10BIO_Degraded, &r10_bio->state);
1430 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1432 sector_t dev_sector = r10_bio->devs[i].addr;
1436 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1437 &first_bad, &bad_sectors);
1439 /* Mustn't write here until the bad block
1442 atomic_inc(&rdev->nr_pending);
1443 set_bit(BlockedBadBlocks, &rdev->flags);
1444 blocked_rdev = rdev;
1447 if (is_bad && first_bad <= dev_sector) {
1448 /* Cannot write here at all */
1449 bad_sectors -= (dev_sector - first_bad);
1450 if (bad_sectors < max_sectors)
1451 /* Mustn't write more than bad_sectors
1452 * to other devices yet
1454 max_sectors = bad_sectors;
1455 /* We don't set R10BIO_Degraded as that
1456 * only applies if the disk is missing,
1457 * so it might be re-added, and we want to
1458 * know to recover this chunk.
1459 * In this case the device is here, and the
1460 * fact that this chunk is not in-sync is
1461 * recorded in the bad block log.
1466 int good_sectors = first_bad - dev_sector;
1467 if (good_sectors < max_sectors)
1468 max_sectors = good_sectors;
1472 r10_bio->devs[i].bio = bio;
1473 atomic_inc(&rdev->nr_pending);
1476 r10_bio->devs[i].repl_bio = bio;
1477 atomic_inc(&rrdev->nr_pending);
1482 if (unlikely(blocked_rdev)) {
1483 /* Have to wait for this device to get unblocked, then retry */
1487 for (j = 0; j < i; j++) {
1488 if (r10_bio->devs[j].bio) {
1489 d = r10_bio->devs[j].devnum;
1490 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492 if (r10_bio->devs[j].repl_bio) {
1493 struct md_rdev *rdev;
1494 d = r10_bio->devs[j].devnum;
1495 rdev = conf->mirrors[d].replacement;
1497 /* Race with remove_disk */
1499 rdev = conf->mirrors[d].rdev;
1501 rdev_dec_pending(rdev, mddev);
1504 allow_barrier(conf);
1505 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1506 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1511 if (max_sectors < r10_bio->sectors)
1512 r10_bio->sectors = max_sectors;
1514 if (r10_bio->sectors < bio_sectors(bio)) {
1515 struct bio *split = bio_split(bio, r10_bio->sectors,
1516 GFP_NOIO, &conf->bio_split);
1517 bio_chain(split, bio);
1518 allow_barrier(conf);
1519 generic_make_request(bio);
1522 r10_bio->master_bio = bio;
1525 atomic_set(&r10_bio->remaining, 1);
1526 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1528 for (i = 0; i < conf->copies; i++) {
1529 if (r10_bio->devs[i].bio)
1530 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1531 if (r10_bio->devs[i].repl_bio)
1532 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1534 one_write_done(r10_bio);
1537 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1539 struct r10conf *conf = mddev->private;
1540 struct r10bio *r10_bio;
1542 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1544 r10_bio->master_bio = bio;
1545 r10_bio->sectors = sectors;
1547 r10_bio->mddev = mddev;
1548 r10_bio->sector = bio->bi_iter.bi_sector;
1550 r10_bio->read_slot = -1;
1551 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1553 if (bio_data_dir(bio) == READ)
1554 raid10_read_request(mddev, bio, r10_bio);
1556 raid10_write_request(mddev, bio, r10_bio);
1559 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1561 struct r10conf *conf = mddev->private;
1562 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1563 int chunk_sects = chunk_mask + 1;
1564 int sectors = bio_sectors(bio);
1566 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1567 && md_flush_request(mddev, bio))
1570 if (!md_write_start(mddev, bio))
1574 * If this request crosses a chunk boundary, we need to split
1577 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1578 sectors > chunk_sects
1579 && (conf->geo.near_copies < conf->geo.raid_disks
1580 || conf->prev.near_copies <
1581 conf->prev.raid_disks)))
1582 sectors = chunk_sects -
1583 (bio->bi_iter.bi_sector &
1585 __make_request(mddev, bio, sectors);
1587 /* In case raid10d snuck in to freeze_array */
1588 wake_up(&conf->wait_barrier);
1592 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1594 struct r10conf *conf = mddev->private;
1597 if (conf->geo.near_copies < conf->geo.raid_disks)
1598 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1599 if (conf->geo.near_copies > 1)
1600 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1601 if (conf->geo.far_copies > 1) {
1602 if (conf->geo.far_offset)
1603 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1605 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1606 if (conf->geo.far_set_size != conf->geo.raid_disks)
1607 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1609 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1610 conf->geo.raid_disks - mddev->degraded);
1612 for (i = 0; i < conf->geo.raid_disks; i++) {
1613 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1614 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1617 seq_printf(seq, "]");
1620 /* check if there are enough drives for
1621 * every block to appear on atleast one.
1622 * Don't consider the device numbered 'ignore'
1623 * as we might be about to remove it.
1625 static int _enough(struct r10conf *conf, int previous, int ignore)
1631 disks = conf->prev.raid_disks;
1632 ncopies = conf->prev.near_copies;
1634 disks = conf->geo.raid_disks;
1635 ncopies = conf->geo.near_copies;
1640 int n = conf->copies;
1644 struct md_rdev *rdev;
1645 if (this != ignore &&
1646 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1647 test_bit(In_sync, &rdev->flags))
1649 this = (this+1) % disks;
1653 first = (first + ncopies) % disks;
1654 } while (first != 0);
1661 static int enough(struct r10conf *conf, int ignore)
1663 /* when calling 'enough', both 'prev' and 'geo' must
1665 * This is ensured if ->reconfig_mutex or ->device_lock
1668 return _enough(conf, 0, ignore) &&
1669 _enough(conf, 1, ignore);
1672 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1674 char b[BDEVNAME_SIZE];
1675 struct r10conf *conf = mddev->private;
1676 unsigned long flags;
1679 * If it is not operational, then we have already marked it as dead
1680 * else if it is the last working disks, ignore the error, let the
1681 * next level up know.
1682 * else mark the drive as failed
1684 spin_lock_irqsave(&conf->device_lock, flags);
1685 if (test_bit(In_sync, &rdev->flags)
1686 && !enough(conf, rdev->raid_disk)) {
1688 * Don't fail the drive, just return an IO error.
1690 spin_unlock_irqrestore(&conf->device_lock, flags);
1693 if (test_and_clear_bit(In_sync, &rdev->flags))
1696 * If recovery is running, make sure it aborts.
1698 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1699 set_bit(Blocked, &rdev->flags);
1700 set_bit(Faulty, &rdev->flags);
1701 set_mask_bits(&mddev->sb_flags, 0,
1702 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1703 spin_unlock_irqrestore(&conf->device_lock, flags);
1704 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1705 "md/raid10:%s: Operation continuing on %d devices.\n",
1706 mdname(mddev), bdevname(rdev->bdev, b),
1707 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1710 static void print_conf(struct r10conf *conf)
1713 struct md_rdev *rdev;
1715 pr_debug("RAID10 conf printout:\n");
1717 pr_debug("(!conf)\n");
1720 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1721 conf->geo.raid_disks);
1723 /* This is only called with ->reconfix_mutex held, so
1724 * rcu protection of rdev is not needed */
1725 for (i = 0; i < conf->geo.raid_disks; i++) {
1726 char b[BDEVNAME_SIZE];
1727 rdev = conf->mirrors[i].rdev;
1729 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1730 i, !test_bit(In_sync, &rdev->flags),
1731 !test_bit(Faulty, &rdev->flags),
1732 bdevname(rdev->bdev,b));
1736 static void close_sync(struct r10conf *conf)
1739 allow_barrier(conf);
1741 mempool_exit(&conf->r10buf_pool);
1744 static int raid10_spare_active(struct mddev *mddev)
1747 struct r10conf *conf = mddev->private;
1748 struct raid10_info *tmp;
1750 unsigned long flags;
1753 * Find all non-in_sync disks within the RAID10 configuration
1754 * and mark them in_sync
1756 for (i = 0; i < conf->geo.raid_disks; i++) {
1757 tmp = conf->mirrors + i;
1758 if (tmp->replacement
1759 && tmp->replacement->recovery_offset == MaxSector
1760 && !test_bit(Faulty, &tmp->replacement->flags)
1761 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1762 /* Replacement has just become active */
1764 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1767 /* Replaced device not technically faulty,
1768 * but we need to be sure it gets removed
1769 * and never re-added.
1771 set_bit(Faulty, &tmp->rdev->flags);
1772 sysfs_notify_dirent_safe(
1773 tmp->rdev->sysfs_state);
1775 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1776 } else if (tmp->rdev
1777 && tmp->rdev->recovery_offset == MaxSector
1778 && !test_bit(Faulty, &tmp->rdev->flags)
1779 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1781 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1784 spin_lock_irqsave(&conf->device_lock, flags);
1785 mddev->degraded -= count;
1786 spin_unlock_irqrestore(&conf->device_lock, flags);
1792 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1794 struct r10conf *conf = mddev->private;
1798 int last = conf->geo.raid_disks - 1;
1800 if (mddev->recovery_cp < MaxSector)
1801 /* only hot-add to in-sync arrays, as recovery is
1802 * very different from resync
1805 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1808 if (md_integrity_add_rdev(rdev, mddev))
1811 if (rdev->raid_disk >= 0)
1812 first = last = rdev->raid_disk;
1814 if (rdev->saved_raid_disk >= first &&
1815 rdev->saved_raid_disk < conf->geo.raid_disks &&
1816 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1817 mirror = rdev->saved_raid_disk;
1820 for ( ; mirror <= last ; mirror++) {
1821 struct raid10_info *p = &conf->mirrors[mirror];
1822 if (p->recovery_disabled == mddev->recovery_disabled)
1825 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1826 p->replacement != NULL)
1828 clear_bit(In_sync, &rdev->flags);
1829 set_bit(Replacement, &rdev->flags);
1830 rdev->raid_disk = mirror;
1833 disk_stack_limits(mddev->gendisk, rdev->bdev,
1834 rdev->data_offset << 9);
1836 rcu_assign_pointer(p->replacement, rdev);
1841 disk_stack_limits(mddev->gendisk, rdev->bdev,
1842 rdev->data_offset << 9);
1844 p->head_position = 0;
1845 p->recovery_disabled = mddev->recovery_disabled - 1;
1846 rdev->raid_disk = mirror;
1848 if (rdev->saved_raid_disk != mirror)
1850 rcu_assign_pointer(p->rdev, rdev);
1853 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1854 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1860 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1862 struct r10conf *conf = mddev->private;
1864 int number = rdev->raid_disk;
1865 struct md_rdev **rdevp;
1866 struct raid10_info *p;
1869 if (unlikely(number >= mddev->raid_disks))
1871 p = conf->mirrors + number;
1872 if (rdev == p->rdev)
1874 else if (rdev == p->replacement)
1875 rdevp = &p->replacement;
1879 if (test_bit(In_sync, &rdev->flags) ||
1880 atomic_read(&rdev->nr_pending)) {
1884 /* Only remove non-faulty devices if recovery
1887 if (!test_bit(Faulty, &rdev->flags) &&
1888 mddev->recovery_disabled != p->recovery_disabled &&
1889 (!p->replacement || p->replacement == rdev) &&
1890 number < conf->geo.raid_disks &&
1896 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1898 if (atomic_read(&rdev->nr_pending)) {
1899 /* lost the race, try later */
1905 if (p->replacement) {
1906 /* We must have just cleared 'rdev' */
1907 p->rdev = p->replacement;
1908 clear_bit(Replacement, &p->replacement->flags);
1909 smp_mb(); /* Make sure other CPUs may see both as identical
1910 * but will never see neither -- if they are careful.
1912 p->replacement = NULL;
1915 clear_bit(WantReplacement, &rdev->flags);
1916 err = md_integrity_register(mddev);
1924 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1926 struct r10conf *conf = r10_bio->mddev->private;
1928 if (!bio->bi_status)
1929 set_bit(R10BIO_Uptodate, &r10_bio->state);
1931 /* The write handler will notice the lack of
1932 * R10BIO_Uptodate and record any errors etc
1934 atomic_add(r10_bio->sectors,
1935 &conf->mirrors[d].rdev->corrected_errors);
1937 /* for reconstruct, we always reschedule after a read.
1938 * for resync, only after all reads
1940 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1941 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1942 atomic_dec_and_test(&r10_bio->remaining)) {
1943 /* we have read all the blocks,
1944 * do the comparison in process context in raid10d
1946 reschedule_retry(r10_bio);
1950 static void end_sync_read(struct bio *bio)
1952 struct r10bio *r10_bio = get_resync_r10bio(bio);
1953 struct r10conf *conf = r10_bio->mddev->private;
1954 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1956 __end_sync_read(r10_bio, bio, d);
1959 static void end_reshape_read(struct bio *bio)
1961 /* reshape read bio isn't allocated from r10buf_pool */
1962 struct r10bio *r10_bio = bio->bi_private;
1964 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1967 static void end_sync_request(struct r10bio *r10_bio)
1969 struct mddev *mddev = r10_bio->mddev;
1971 while (atomic_dec_and_test(&r10_bio->remaining)) {
1972 if (r10_bio->master_bio == NULL) {
1973 /* the primary of several recovery bios */
1974 sector_t s = r10_bio->sectors;
1975 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1976 test_bit(R10BIO_WriteError, &r10_bio->state))
1977 reschedule_retry(r10_bio);
1980 md_done_sync(mddev, s, 1);
1983 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1984 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1985 test_bit(R10BIO_WriteError, &r10_bio->state))
1986 reschedule_retry(r10_bio);
1994 static void end_sync_write(struct bio *bio)
1996 struct r10bio *r10_bio = get_resync_r10bio(bio);
1997 struct mddev *mddev = r10_bio->mddev;
1998 struct r10conf *conf = mddev->private;
2004 struct md_rdev *rdev = NULL;
2006 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2008 rdev = conf->mirrors[d].replacement;
2010 rdev = conf->mirrors[d].rdev;
2012 if (bio->bi_status) {
2014 md_error(mddev, rdev);
2016 set_bit(WriteErrorSeen, &rdev->flags);
2017 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2018 set_bit(MD_RECOVERY_NEEDED,
2019 &rdev->mddev->recovery);
2020 set_bit(R10BIO_WriteError, &r10_bio->state);
2022 } else if (is_badblock(rdev,
2023 r10_bio->devs[slot].addr,
2025 &first_bad, &bad_sectors))
2026 set_bit(R10BIO_MadeGood, &r10_bio->state);
2028 rdev_dec_pending(rdev, mddev);
2030 end_sync_request(r10_bio);
2034 * Note: sync and recover and handled very differently for raid10
2035 * This code is for resync.
2036 * For resync, we read through virtual addresses and read all blocks.
2037 * If there is any error, we schedule a write. The lowest numbered
2038 * drive is authoritative.
2039 * However requests come for physical address, so we need to map.
2040 * For every physical address there are raid_disks/copies virtual addresses,
2041 * which is always are least one, but is not necessarly an integer.
2042 * This means that a physical address can span multiple chunks, so we may
2043 * have to submit multiple io requests for a single sync request.
2046 * We check if all blocks are in-sync and only write to blocks that
2049 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2051 struct r10conf *conf = mddev->private;
2053 struct bio *tbio, *fbio;
2055 struct page **tpages, **fpages;
2057 atomic_set(&r10_bio->remaining, 1);
2059 /* find the first device with a block */
2060 for (i=0; i<conf->copies; i++)
2061 if (!r10_bio->devs[i].bio->bi_status)
2064 if (i == conf->copies)
2068 fbio = r10_bio->devs[i].bio;
2069 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2070 fbio->bi_iter.bi_idx = 0;
2071 fpages = get_resync_pages(fbio)->pages;
2073 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2074 /* now find blocks with errors */
2075 for (i=0 ; i < conf->copies ; i++) {
2077 struct md_rdev *rdev;
2078 struct resync_pages *rp;
2080 tbio = r10_bio->devs[i].bio;
2082 if (tbio->bi_end_io != end_sync_read)
2087 tpages = get_resync_pages(tbio)->pages;
2088 d = r10_bio->devs[i].devnum;
2089 rdev = conf->mirrors[d].rdev;
2090 if (!r10_bio->devs[i].bio->bi_status) {
2091 /* We know that the bi_io_vec layout is the same for
2092 * both 'first' and 'i', so we just compare them.
2093 * All vec entries are PAGE_SIZE;
2095 int sectors = r10_bio->sectors;
2096 for (j = 0; j < vcnt; j++) {
2097 int len = PAGE_SIZE;
2098 if (sectors < (len / 512))
2099 len = sectors * 512;
2100 if (memcmp(page_address(fpages[j]),
2101 page_address(tpages[j]),
2108 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2109 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2110 /* Don't fix anything. */
2112 } else if (test_bit(FailFast, &rdev->flags)) {
2113 /* Just give up on this device */
2114 md_error(rdev->mddev, rdev);
2117 /* Ok, we need to write this bio, either to correct an
2118 * inconsistency or to correct an unreadable block.
2119 * First we need to fixup bv_offset, bv_len and
2120 * bi_vecs, as the read request might have corrupted these
2122 rp = get_resync_pages(tbio);
2125 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2127 rp->raid_bio = r10_bio;
2128 tbio->bi_private = rp;
2129 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2130 tbio->bi_end_io = end_sync_write;
2131 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2133 bio_copy_data(tbio, fbio);
2135 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2136 atomic_inc(&r10_bio->remaining);
2137 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2139 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2140 tbio->bi_opf |= MD_FAILFAST;
2141 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2142 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2143 generic_make_request(tbio);
2146 /* Now write out to any replacement devices
2149 for (i = 0; i < conf->copies; i++) {
2152 tbio = r10_bio->devs[i].repl_bio;
2153 if (!tbio || !tbio->bi_end_io)
2155 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2156 && r10_bio->devs[i].bio != fbio)
2157 bio_copy_data(tbio, fbio);
2158 d = r10_bio->devs[i].devnum;
2159 atomic_inc(&r10_bio->remaining);
2160 md_sync_acct(conf->mirrors[d].replacement->bdev,
2162 generic_make_request(tbio);
2166 if (atomic_dec_and_test(&r10_bio->remaining)) {
2167 md_done_sync(mddev, r10_bio->sectors, 1);
2173 * Now for the recovery code.
2174 * Recovery happens across physical sectors.
2175 * We recover all non-is_sync drives by finding the virtual address of
2176 * each, and then choose a working drive that also has that virt address.
2177 * There is a separate r10_bio for each non-in_sync drive.
2178 * Only the first two slots are in use. The first for reading,
2179 * The second for writing.
2182 static void fix_recovery_read_error(struct r10bio *r10_bio)
2184 /* We got a read error during recovery.
2185 * We repeat the read in smaller page-sized sections.
2186 * If a read succeeds, write it to the new device or record
2187 * a bad block if we cannot.
2188 * If a read fails, record a bad block on both old and
2191 struct mddev *mddev = r10_bio->mddev;
2192 struct r10conf *conf = mddev->private;
2193 struct bio *bio = r10_bio->devs[0].bio;
2195 int sectors = r10_bio->sectors;
2197 int dr = r10_bio->devs[0].devnum;
2198 int dw = r10_bio->devs[1].devnum;
2199 struct page **pages = get_resync_pages(bio)->pages;
2203 struct md_rdev *rdev;
2207 if (s > (PAGE_SIZE>>9))
2210 rdev = conf->mirrors[dr].rdev;
2211 addr = r10_bio->devs[0].addr + sect,
2212 ok = sync_page_io(rdev,
2216 REQ_OP_READ, 0, false);
2218 rdev = conf->mirrors[dw].rdev;
2219 addr = r10_bio->devs[1].addr + sect;
2220 ok = sync_page_io(rdev,
2224 REQ_OP_WRITE, 0, false);
2226 set_bit(WriteErrorSeen, &rdev->flags);
2227 if (!test_and_set_bit(WantReplacement,
2229 set_bit(MD_RECOVERY_NEEDED,
2230 &rdev->mddev->recovery);
2234 /* We don't worry if we cannot set a bad block -
2235 * it really is bad so there is no loss in not
2238 rdev_set_badblocks(rdev, addr, s, 0);
2240 if (rdev != conf->mirrors[dw].rdev) {
2241 /* need bad block on destination too */
2242 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2243 addr = r10_bio->devs[1].addr + sect;
2244 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2246 /* just abort the recovery */
2247 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2250 conf->mirrors[dw].recovery_disabled
2251 = mddev->recovery_disabled;
2252 set_bit(MD_RECOVERY_INTR,
2265 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2267 struct r10conf *conf = mddev->private;
2269 struct bio *wbio, *wbio2;
2271 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2272 fix_recovery_read_error(r10_bio);
2273 end_sync_request(r10_bio);
2278 * share the pages with the first bio
2279 * and submit the write request
2281 d = r10_bio->devs[1].devnum;
2282 wbio = r10_bio->devs[1].bio;
2283 wbio2 = r10_bio->devs[1].repl_bio;
2284 /* Need to test wbio2->bi_end_io before we call
2285 * generic_make_request as if the former is NULL,
2286 * the latter is free to free wbio2.
2288 if (wbio2 && !wbio2->bi_end_io)
2290 if (wbio->bi_end_io) {
2291 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2292 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2293 generic_make_request(wbio);
2296 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2297 md_sync_acct(conf->mirrors[d].replacement->bdev,
2298 bio_sectors(wbio2));
2299 generic_make_request(wbio2);
2304 * Used by fix_read_error() to decay the per rdev read_errors.
2305 * We halve the read error count for every hour that has elapsed
2306 * since the last recorded read error.
2309 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2312 unsigned long hours_since_last;
2313 unsigned int read_errors = atomic_read(&rdev->read_errors);
2315 cur_time_mon = ktime_get_seconds();
2317 if (rdev->last_read_error == 0) {
2318 /* first time we've seen a read error */
2319 rdev->last_read_error = cur_time_mon;
2323 hours_since_last = (long)(cur_time_mon -
2324 rdev->last_read_error) / 3600;
2326 rdev->last_read_error = cur_time_mon;
2329 * if hours_since_last is > the number of bits in read_errors
2330 * just set read errors to 0. We do this to avoid
2331 * overflowing the shift of read_errors by hours_since_last.
2333 if (hours_since_last >= 8 * sizeof(read_errors))
2334 atomic_set(&rdev->read_errors, 0);
2336 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2339 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2340 int sectors, struct page *page, int rw)
2345 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2346 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2348 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2352 set_bit(WriteErrorSeen, &rdev->flags);
2353 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2354 set_bit(MD_RECOVERY_NEEDED,
2355 &rdev->mddev->recovery);
2357 /* need to record an error - either for the block or the device */
2358 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2359 md_error(rdev->mddev, rdev);
2364 * This is a kernel thread which:
2366 * 1. Retries failed read operations on working mirrors.
2367 * 2. Updates the raid superblock when problems encounter.
2368 * 3. Performs writes following reads for array synchronising.
2371 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2373 int sect = 0; /* Offset from r10_bio->sector */
2374 int sectors = r10_bio->sectors;
2375 struct md_rdev *rdev;
2376 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2377 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2379 /* still own a reference to this rdev, so it cannot
2380 * have been cleared recently.
2382 rdev = conf->mirrors[d].rdev;
2384 if (test_bit(Faulty, &rdev->flags))
2385 /* drive has already been failed, just ignore any
2386 more fix_read_error() attempts */
2389 check_decay_read_errors(mddev, rdev);
2390 atomic_inc(&rdev->read_errors);
2391 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2392 char b[BDEVNAME_SIZE];
2393 bdevname(rdev->bdev, b);
2395 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2397 atomic_read(&rdev->read_errors), max_read_errors);
2398 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2400 md_error(mddev, rdev);
2401 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2407 int sl = r10_bio->read_slot;
2411 if (s > (PAGE_SIZE>>9))
2419 d = r10_bio->devs[sl].devnum;
2420 rdev = rcu_dereference(conf->mirrors[d].rdev);
2422 test_bit(In_sync, &rdev->flags) &&
2423 !test_bit(Faulty, &rdev->flags) &&
2424 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2425 &first_bad, &bad_sectors) == 0) {
2426 atomic_inc(&rdev->nr_pending);
2428 success = sync_page_io(rdev,
2429 r10_bio->devs[sl].addr +
2433 REQ_OP_READ, 0, false);
2434 rdev_dec_pending(rdev, mddev);
2440 if (sl == conf->copies)
2442 } while (!success && sl != r10_bio->read_slot);
2446 /* Cannot read from anywhere, just mark the block
2447 * as bad on the first device to discourage future
2450 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2451 rdev = conf->mirrors[dn].rdev;
2453 if (!rdev_set_badblocks(
2455 r10_bio->devs[r10_bio->read_slot].addr
2458 md_error(mddev, rdev);
2459 r10_bio->devs[r10_bio->read_slot].bio
2466 /* write it back and re-read */
2468 while (sl != r10_bio->read_slot) {
2469 char b[BDEVNAME_SIZE];
2474 d = r10_bio->devs[sl].devnum;
2475 rdev = rcu_dereference(conf->mirrors[d].rdev);
2477 test_bit(Faulty, &rdev->flags) ||
2478 !test_bit(In_sync, &rdev->flags))
2481 atomic_inc(&rdev->nr_pending);
2483 if (r10_sync_page_io(rdev,
2484 r10_bio->devs[sl].addr +
2486 s, conf->tmppage, WRITE)
2488 /* Well, this device is dead */
2489 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2491 (unsigned long long)(
2493 choose_data_offset(r10_bio,
2495 bdevname(rdev->bdev, b));
2496 pr_notice("md/raid10:%s: %s: failing drive\n",
2498 bdevname(rdev->bdev, b));
2500 rdev_dec_pending(rdev, mddev);
2504 while (sl != r10_bio->read_slot) {
2505 char b[BDEVNAME_SIZE];
2510 d = r10_bio->devs[sl].devnum;
2511 rdev = rcu_dereference(conf->mirrors[d].rdev);
2513 test_bit(Faulty, &rdev->flags) ||
2514 !test_bit(In_sync, &rdev->flags))
2517 atomic_inc(&rdev->nr_pending);
2519 switch (r10_sync_page_io(rdev,
2520 r10_bio->devs[sl].addr +
2525 /* Well, this device is dead */
2526 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2528 (unsigned long long)(
2530 choose_data_offset(r10_bio, rdev)),
2531 bdevname(rdev->bdev, b));
2532 pr_notice("md/raid10:%s: %s: failing drive\n",
2534 bdevname(rdev->bdev, b));
2537 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2539 (unsigned long long)(
2541 choose_data_offset(r10_bio, rdev)),
2542 bdevname(rdev->bdev, b));
2543 atomic_add(s, &rdev->corrected_errors);
2546 rdev_dec_pending(rdev, mddev);
2556 static int narrow_write_error(struct r10bio *r10_bio, int i)
2558 struct bio *bio = r10_bio->master_bio;
2559 struct mddev *mddev = r10_bio->mddev;
2560 struct r10conf *conf = mddev->private;
2561 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2562 /* bio has the data to be written to slot 'i' where
2563 * we just recently had a write error.
2564 * We repeatedly clone the bio and trim down to one block,
2565 * then try the write. Where the write fails we record
2567 * It is conceivable that the bio doesn't exactly align with
2568 * blocks. We must handle this.
2570 * We currently own a reference to the rdev.
2576 int sect_to_write = r10_bio->sectors;
2579 if (rdev->badblocks.shift < 0)
2582 block_sectors = roundup(1 << rdev->badblocks.shift,
2583 bdev_logical_block_size(rdev->bdev) >> 9);
2584 sector = r10_bio->sector;
2585 sectors = ((r10_bio->sector + block_sectors)
2586 & ~(sector_t)(block_sectors - 1))
2589 while (sect_to_write) {
2592 if (sectors > sect_to_write)
2593 sectors = sect_to_write;
2594 /* Write at 'sector' for 'sectors' */
2595 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2596 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2597 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2598 wbio->bi_iter.bi_sector = wsector +
2599 choose_data_offset(r10_bio, rdev);
2600 bio_set_dev(wbio, rdev->bdev);
2601 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2603 if (submit_bio_wait(wbio) < 0)
2605 ok = rdev_set_badblocks(rdev, wsector,
2610 sect_to_write -= sectors;
2612 sectors = block_sectors;
2617 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2619 int slot = r10_bio->read_slot;
2621 struct r10conf *conf = mddev->private;
2622 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2624 /* we got a read error. Maybe the drive is bad. Maybe just
2625 * the block and we can fix it.
2626 * We freeze all other IO, and try reading the block from
2627 * other devices. When we find one, we re-write
2628 * and check it that fixes the read error.
2629 * This is all done synchronously while the array is
2632 bio = r10_bio->devs[slot].bio;
2634 r10_bio->devs[slot].bio = NULL;
2637 r10_bio->devs[slot].bio = IO_BLOCKED;
2638 else if (!test_bit(FailFast, &rdev->flags)) {
2639 freeze_array(conf, 1);
2640 fix_read_error(conf, mddev, r10_bio);
2641 unfreeze_array(conf);
2643 md_error(mddev, rdev);
2645 rdev_dec_pending(rdev, mddev);
2646 allow_barrier(conf);
2648 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2651 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2653 /* Some sort of write request has finished and it
2654 * succeeded in writing where we thought there was a
2655 * bad block. So forget the bad block.
2656 * Or possibly if failed and we need to record
2660 struct md_rdev *rdev;
2662 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2663 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2664 for (m = 0; m < conf->copies; m++) {
2665 int dev = r10_bio->devs[m].devnum;
2666 rdev = conf->mirrors[dev].rdev;
2667 if (r10_bio->devs[m].bio == NULL ||
2668 r10_bio->devs[m].bio->bi_end_io == NULL)
2670 if (!r10_bio->devs[m].bio->bi_status) {
2671 rdev_clear_badblocks(
2673 r10_bio->devs[m].addr,
2674 r10_bio->sectors, 0);
2676 if (!rdev_set_badblocks(
2678 r10_bio->devs[m].addr,
2679 r10_bio->sectors, 0))
2680 md_error(conf->mddev, rdev);
2682 rdev = conf->mirrors[dev].replacement;
2683 if (r10_bio->devs[m].repl_bio == NULL ||
2684 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2687 if (!r10_bio->devs[m].repl_bio->bi_status) {
2688 rdev_clear_badblocks(
2690 r10_bio->devs[m].addr,
2691 r10_bio->sectors, 0);
2693 if (!rdev_set_badblocks(
2695 r10_bio->devs[m].addr,
2696 r10_bio->sectors, 0))
2697 md_error(conf->mddev, rdev);
2703 for (m = 0; m < conf->copies; m++) {
2704 int dev = r10_bio->devs[m].devnum;
2705 struct bio *bio = r10_bio->devs[m].bio;
2706 rdev = conf->mirrors[dev].rdev;
2707 if (bio == IO_MADE_GOOD) {
2708 rdev_clear_badblocks(
2710 r10_bio->devs[m].addr,
2711 r10_bio->sectors, 0);
2712 rdev_dec_pending(rdev, conf->mddev);
2713 } else if (bio != NULL && bio->bi_status) {
2715 if (!narrow_write_error(r10_bio, m)) {
2716 md_error(conf->mddev, rdev);
2717 set_bit(R10BIO_Degraded,
2720 rdev_dec_pending(rdev, conf->mddev);
2722 bio = r10_bio->devs[m].repl_bio;
2723 rdev = conf->mirrors[dev].replacement;
2724 if (rdev && bio == IO_MADE_GOOD) {
2725 rdev_clear_badblocks(
2727 r10_bio->devs[m].addr,
2728 r10_bio->sectors, 0);
2729 rdev_dec_pending(rdev, conf->mddev);
2733 spin_lock_irq(&conf->device_lock);
2734 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2736 spin_unlock_irq(&conf->device_lock);
2738 * In case freeze_array() is waiting for condition
2739 * nr_pending == nr_queued + extra to be true.
2741 wake_up(&conf->wait_barrier);
2742 md_wakeup_thread(conf->mddev->thread);
2744 if (test_bit(R10BIO_WriteError,
2746 close_write(r10_bio);
2747 raid_end_bio_io(r10_bio);
2752 static void raid10d(struct md_thread *thread)
2754 struct mddev *mddev = thread->mddev;
2755 struct r10bio *r10_bio;
2756 unsigned long flags;
2757 struct r10conf *conf = mddev->private;
2758 struct list_head *head = &conf->retry_list;
2759 struct blk_plug plug;
2761 md_check_recovery(mddev);
2763 if (!list_empty_careful(&conf->bio_end_io_list) &&
2764 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2766 spin_lock_irqsave(&conf->device_lock, flags);
2767 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2768 while (!list_empty(&conf->bio_end_io_list)) {
2769 list_move(conf->bio_end_io_list.prev, &tmp);
2773 spin_unlock_irqrestore(&conf->device_lock, flags);
2774 while (!list_empty(&tmp)) {
2775 r10_bio = list_first_entry(&tmp, struct r10bio,
2777 list_del(&r10_bio->retry_list);
2778 if (mddev->degraded)
2779 set_bit(R10BIO_Degraded, &r10_bio->state);
2781 if (test_bit(R10BIO_WriteError,
2783 close_write(r10_bio);
2784 raid_end_bio_io(r10_bio);
2788 blk_start_plug(&plug);
2791 flush_pending_writes(conf);
2793 spin_lock_irqsave(&conf->device_lock, flags);
2794 if (list_empty(head)) {
2795 spin_unlock_irqrestore(&conf->device_lock, flags);
2798 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2799 list_del(head->prev);
2801 spin_unlock_irqrestore(&conf->device_lock, flags);
2803 mddev = r10_bio->mddev;
2804 conf = mddev->private;
2805 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2806 test_bit(R10BIO_WriteError, &r10_bio->state))
2807 handle_write_completed(conf, r10_bio);
2808 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2809 reshape_request_write(mddev, r10_bio);
2810 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2811 sync_request_write(mddev, r10_bio);
2812 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2813 recovery_request_write(mddev, r10_bio);
2814 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2815 handle_read_error(mddev, r10_bio);
2820 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2821 md_check_recovery(mddev);
2823 blk_finish_plug(&plug);
2826 static int init_resync(struct r10conf *conf)
2830 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2831 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2832 conf->have_replacement = 0;
2833 for (i = 0; i < conf->geo.raid_disks; i++)
2834 if (conf->mirrors[i].replacement)
2835 conf->have_replacement = 1;
2836 ret = mempool_init(&conf->r10buf_pool, buffs,
2837 r10buf_pool_alloc, r10buf_pool_free, conf);
2840 conf->next_resync = 0;
2844 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2846 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2847 struct rsync_pages *rp;
2852 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2853 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2854 nalloc = conf->copies; /* resync */
2856 nalloc = 2; /* recovery */
2858 for (i = 0; i < nalloc; i++) {
2859 bio = r10bio->devs[i].bio;
2860 rp = bio->bi_private;
2862 bio->bi_private = rp;
2863 bio = r10bio->devs[i].repl_bio;
2865 rp = bio->bi_private;
2867 bio->bi_private = rp;
2874 * Set cluster_sync_high since we need other nodes to add the
2875 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2877 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2879 sector_t window_size;
2880 int extra_chunk, chunks;
2883 * First, here we define "stripe" as a unit which across
2884 * all member devices one time, so we get chunks by use
2885 * raid_disks / near_copies. Otherwise, if near_copies is
2886 * close to raid_disks, then resync window could increases
2887 * linearly with the increase of raid_disks, which means
2888 * we will suspend a really large IO window while it is not
2889 * necessary. If raid_disks is not divisible by near_copies,
2890 * an extra chunk is needed to ensure the whole "stripe" is
2894 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2895 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2899 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2902 * At least use a 32M window to align with raid1's resync window
2904 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2905 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2907 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2911 * perform a "sync" on one "block"
2913 * We need to make sure that no normal I/O request - particularly write
2914 * requests - conflict with active sync requests.
2916 * This is achieved by tracking pending requests and a 'barrier' concept
2917 * that can be installed to exclude normal IO requests.
2919 * Resync and recovery are handled very differently.
2920 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2922 * For resync, we iterate over virtual addresses, read all copies,
2923 * and update if there are differences. If only one copy is live,
2925 * For recovery, we iterate over physical addresses, read a good
2926 * value for each non-in_sync drive, and over-write.
2928 * So, for recovery we may have several outstanding complex requests for a
2929 * given address, one for each out-of-sync device. We model this by allocating
2930 * a number of r10_bio structures, one for each out-of-sync device.
2931 * As we setup these structures, we collect all bio's together into a list
2932 * which we then process collectively to add pages, and then process again
2933 * to pass to generic_make_request.
2935 * The r10_bio structures are linked using a borrowed master_bio pointer.
2936 * This link is counted in ->remaining. When the r10_bio that points to NULL
2937 * has its remaining count decremented to 0, the whole complex operation
2942 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2945 struct r10conf *conf = mddev->private;
2946 struct r10bio *r10_bio;
2947 struct bio *biolist = NULL, *bio;
2948 sector_t max_sector, nr_sectors;
2951 sector_t sync_blocks;
2952 sector_t sectors_skipped = 0;
2953 int chunks_skipped = 0;
2954 sector_t chunk_mask = conf->geo.chunk_mask;
2957 if (!mempool_initialized(&conf->r10buf_pool))
2958 if (init_resync(conf))
2962 * Allow skipping a full rebuild for incremental assembly
2963 * of a clean array, like RAID1 does.
2965 if (mddev->bitmap == NULL &&
2966 mddev->recovery_cp == MaxSector &&
2967 mddev->reshape_position == MaxSector &&
2968 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2969 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2970 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2971 conf->fullsync == 0) {
2973 return mddev->dev_sectors - sector_nr;
2977 max_sector = mddev->dev_sectors;
2978 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2979 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2980 max_sector = mddev->resync_max_sectors;
2981 if (sector_nr >= max_sector) {
2982 conf->cluster_sync_low = 0;
2983 conf->cluster_sync_high = 0;
2985 /* If we aborted, we need to abort the
2986 * sync on the 'current' bitmap chucks (there can
2987 * be several when recovering multiple devices).
2988 * as we may have started syncing it but not finished.
2989 * We can find the current address in
2990 * mddev->curr_resync, but for recovery,
2991 * we need to convert that to several
2992 * virtual addresses.
2994 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3000 if (mddev->curr_resync < max_sector) { /* aborted */
3001 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3002 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3004 else for (i = 0; i < conf->geo.raid_disks; i++) {
3006 raid10_find_virt(conf, mddev->curr_resync, i);
3007 md_bitmap_end_sync(mddev->bitmap, sect,
3011 /* completed sync */
3012 if ((!mddev->bitmap || conf->fullsync)
3013 && conf->have_replacement
3014 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3015 /* Completed a full sync so the replacements
3016 * are now fully recovered.
3019 for (i = 0; i < conf->geo.raid_disks; i++) {
3020 struct md_rdev *rdev =
3021 rcu_dereference(conf->mirrors[i].replacement);
3023 rdev->recovery_offset = MaxSector;
3029 md_bitmap_close_sync(mddev->bitmap);
3032 return sectors_skipped;
3035 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3036 return reshape_request(mddev, sector_nr, skipped);
3038 if (chunks_skipped >= conf->geo.raid_disks) {
3039 /* if there has been nothing to do on any drive,
3040 * then there is nothing to do at all..
3043 return (max_sector - sector_nr) + sectors_skipped;
3046 if (max_sector > mddev->resync_max)
3047 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3049 /* make sure whole request will fit in a chunk - if chunks
3052 if (conf->geo.near_copies < conf->geo.raid_disks &&
3053 max_sector > (sector_nr | chunk_mask))
3054 max_sector = (sector_nr | chunk_mask) + 1;
3057 * If there is non-resync activity waiting for a turn, then let it
3058 * though before starting on this new sync request.
3060 if (conf->nr_waiting)
3061 schedule_timeout_uninterruptible(1);
3063 /* Again, very different code for resync and recovery.
3064 * Both must result in an r10bio with a list of bios that
3065 * have bi_end_io, bi_sector, bi_disk set,
3066 * and bi_private set to the r10bio.
3067 * For recovery, we may actually create several r10bios
3068 * with 2 bios in each, that correspond to the bios in the main one.
3069 * In this case, the subordinate r10bios link back through a
3070 * borrowed master_bio pointer, and the counter in the master
3071 * includes a ref from each subordinate.
3073 /* First, we decide what to do and set ->bi_end_io
3074 * To end_sync_read if we want to read, and
3075 * end_sync_write if we will want to write.
3078 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3079 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3080 /* recovery... the complicated one */
3084 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3090 struct raid10_info *mirror = &conf->mirrors[i];
3091 struct md_rdev *mrdev, *mreplace;
3094 mrdev = rcu_dereference(mirror->rdev);
3095 mreplace = rcu_dereference(mirror->replacement);
3097 if ((mrdev == NULL ||
3098 test_bit(Faulty, &mrdev->flags) ||
3099 test_bit(In_sync, &mrdev->flags)) &&
3100 (mreplace == NULL ||
3101 test_bit(Faulty, &mreplace->flags))) {
3107 /* want to reconstruct this device */
3109 sect = raid10_find_virt(conf, sector_nr, i);
3110 if (sect >= mddev->resync_max_sectors) {
3111 /* last stripe is not complete - don't
3112 * try to recover this sector.
3117 if (mreplace && test_bit(Faulty, &mreplace->flags))
3119 /* Unless we are doing a full sync, or a replacement
3120 * we only need to recover the block if it is set in
3123 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3125 if (sync_blocks < max_sync)
3126 max_sync = sync_blocks;
3130 /* yep, skip the sync_blocks here, but don't assume
3131 * that there will never be anything to do here
3133 chunks_skipped = -1;
3137 atomic_inc(&mrdev->nr_pending);
3139 atomic_inc(&mreplace->nr_pending);
3142 r10_bio = raid10_alloc_init_r10buf(conf);
3144 raise_barrier(conf, rb2 != NULL);
3145 atomic_set(&r10_bio->remaining, 0);
3147 r10_bio->master_bio = (struct bio*)rb2;
3149 atomic_inc(&rb2->remaining);
3150 r10_bio->mddev = mddev;
3151 set_bit(R10BIO_IsRecover, &r10_bio->state);
3152 r10_bio->sector = sect;
3154 raid10_find_phys(conf, r10_bio);
3156 /* Need to check if the array will still be
3160 for (j = 0; j < conf->geo.raid_disks; j++) {
3161 struct md_rdev *rdev = rcu_dereference(
3162 conf->mirrors[j].rdev);
3163 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3169 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3170 &sync_blocks, still_degraded);
3173 for (j=0; j<conf->copies;j++) {
3175 int d = r10_bio->devs[j].devnum;
3176 sector_t from_addr, to_addr;
3177 struct md_rdev *rdev =
3178 rcu_dereference(conf->mirrors[d].rdev);
3179 sector_t sector, first_bad;
3182 !test_bit(In_sync, &rdev->flags))
3184 /* This is where we read from */
3186 sector = r10_bio->devs[j].addr;
3188 if (is_badblock(rdev, sector, max_sync,
3189 &first_bad, &bad_sectors)) {
3190 if (first_bad > sector)
3191 max_sync = first_bad - sector;
3193 bad_sectors -= (sector
3195 if (max_sync > bad_sectors)
3196 max_sync = bad_sectors;
3200 bio = r10_bio->devs[0].bio;
3201 bio->bi_next = biolist;
3203 bio->bi_end_io = end_sync_read;
3204 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3205 if (test_bit(FailFast, &rdev->flags))
3206 bio->bi_opf |= MD_FAILFAST;
3207 from_addr = r10_bio->devs[j].addr;
3208 bio->bi_iter.bi_sector = from_addr +
3210 bio_set_dev(bio, rdev->bdev);
3211 atomic_inc(&rdev->nr_pending);
3212 /* and we write to 'i' (if not in_sync) */
3214 for (k=0; k<conf->copies; k++)
3215 if (r10_bio->devs[k].devnum == i)
3217 BUG_ON(k == conf->copies);
3218 to_addr = r10_bio->devs[k].addr;
3219 r10_bio->devs[0].devnum = d;
3220 r10_bio->devs[0].addr = from_addr;
3221 r10_bio->devs[1].devnum = i;
3222 r10_bio->devs[1].addr = to_addr;
3224 if (!test_bit(In_sync, &mrdev->flags)) {
3225 bio = r10_bio->devs[1].bio;
3226 bio->bi_next = biolist;
3228 bio->bi_end_io = end_sync_write;
3229 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3230 bio->bi_iter.bi_sector = to_addr
3231 + mrdev->data_offset;
3232 bio_set_dev(bio, mrdev->bdev);
3233 atomic_inc(&r10_bio->remaining);
3235 r10_bio->devs[1].bio->bi_end_io = NULL;
3237 /* and maybe write to replacement */
3238 bio = r10_bio->devs[1].repl_bio;
3240 bio->bi_end_io = NULL;
3241 /* Note: if mreplace != NULL, then bio
3242 * cannot be NULL as r10buf_pool_alloc will
3243 * have allocated it.
3244 * So the second test here is pointless.
3245 * But it keeps semantic-checkers happy, and
3246 * this comment keeps human reviewers
3249 if (mreplace == NULL || bio == NULL ||
3250 test_bit(Faulty, &mreplace->flags))
3252 bio->bi_next = biolist;
3254 bio->bi_end_io = end_sync_write;
3255 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3256 bio->bi_iter.bi_sector = to_addr +
3257 mreplace->data_offset;
3258 bio_set_dev(bio, mreplace->bdev);
3259 atomic_inc(&r10_bio->remaining);
3263 if (j == conf->copies) {
3264 /* Cannot recover, so abort the recovery or
3265 * record a bad block */
3267 /* problem is that there are bad blocks
3268 * on other device(s)
3271 for (k = 0; k < conf->copies; k++)
3272 if (r10_bio->devs[k].devnum == i)
3274 if (!test_bit(In_sync,
3276 && !rdev_set_badblocks(
3278 r10_bio->devs[k].addr,
3282 !rdev_set_badblocks(
3284 r10_bio->devs[k].addr,
3289 if (!test_and_set_bit(MD_RECOVERY_INTR,
3291 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3293 mirror->recovery_disabled
3294 = mddev->recovery_disabled;
3298 atomic_dec(&rb2->remaining);
3300 rdev_dec_pending(mrdev, mddev);
3302 rdev_dec_pending(mreplace, mddev);
3305 rdev_dec_pending(mrdev, mddev);
3307 rdev_dec_pending(mreplace, mddev);
3308 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3309 /* Only want this if there is elsewhere to
3310 * read from. 'j' is currently the first
3314 for (; j < conf->copies; j++) {
3315 int d = r10_bio->devs[j].devnum;
3316 if (conf->mirrors[d].rdev &&
3318 &conf->mirrors[d].rdev->flags))
3322 r10_bio->devs[0].bio->bi_opf
3326 if (biolist == NULL) {
3328 struct r10bio *rb2 = r10_bio;
3329 r10_bio = (struct r10bio*) rb2->master_bio;
3330 rb2->master_bio = NULL;
3336 /* resync. Schedule a read for every block at this virt offset */
3340 * Since curr_resync_completed could probably not update in
3341 * time, and we will set cluster_sync_low based on it.
3342 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3343 * safety reason, which ensures curr_resync_completed is
3344 * updated in bitmap_cond_end_sync.
3346 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3347 mddev_is_clustered(mddev) &&
3348 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3350 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3351 &sync_blocks, mddev->degraded) &&
3352 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3353 &mddev->recovery)) {
3354 /* We can skip this block */
3356 return sync_blocks + sectors_skipped;
3358 if (sync_blocks < max_sync)
3359 max_sync = sync_blocks;
3360 r10_bio = raid10_alloc_init_r10buf(conf);
3363 r10_bio->mddev = mddev;
3364 atomic_set(&r10_bio->remaining, 0);
3365 raise_barrier(conf, 0);
3366 conf->next_resync = sector_nr;
3368 r10_bio->master_bio = NULL;
3369 r10_bio->sector = sector_nr;
3370 set_bit(R10BIO_IsSync, &r10_bio->state);
3371 raid10_find_phys(conf, r10_bio);
3372 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3374 for (i = 0; i < conf->copies; i++) {
3375 int d = r10_bio->devs[i].devnum;
3376 sector_t first_bad, sector;
3378 struct md_rdev *rdev;
3380 if (r10_bio->devs[i].repl_bio)
3381 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3383 bio = r10_bio->devs[i].bio;
3384 bio->bi_status = BLK_STS_IOERR;
3386 rdev = rcu_dereference(conf->mirrors[d].rdev);
3387 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3391 sector = r10_bio->devs[i].addr;
3392 if (is_badblock(rdev, sector, max_sync,
3393 &first_bad, &bad_sectors)) {
3394 if (first_bad > sector)
3395 max_sync = first_bad - sector;
3397 bad_sectors -= (sector - first_bad);
3398 if (max_sync > bad_sectors)
3399 max_sync = bad_sectors;
3404 atomic_inc(&rdev->nr_pending);
3405 atomic_inc(&r10_bio->remaining);
3406 bio->bi_next = biolist;
3408 bio->bi_end_io = end_sync_read;
3409 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3410 if (test_bit(FailFast, &rdev->flags))
3411 bio->bi_opf |= MD_FAILFAST;
3412 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3413 bio_set_dev(bio, rdev->bdev);
3416 rdev = rcu_dereference(conf->mirrors[d].replacement);
3417 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3421 atomic_inc(&rdev->nr_pending);
3423 /* Need to set up for writing to the replacement */
3424 bio = r10_bio->devs[i].repl_bio;
3425 bio->bi_status = BLK_STS_IOERR;
3427 sector = r10_bio->devs[i].addr;
3428 bio->bi_next = biolist;
3430 bio->bi_end_io = end_sync_write;
3431 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3432 if (test_bit(FailFast, &rdev->flags))
3433 bio->bi_opf |= MD_FAILFAST;
3434 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3435 bio_set_dev(bio, rdev->bdev);
3441 for (i=0; i<conf->copies; i++) {
3442 int d = r10_bio->devs[i].devnum;
3443 if (r10_bio->devs[i].bio->bi_end_io)
3444 rdev_dec_pending(conf->mirrors[d].rdev,
3446 if (r10_bio->devs[i].repl_bio &&
3447 r10_bio->devs[i].repl_bio->bi_end_io)
3449 conf->mirrors[d].replacement,
3459 if (sector_nr + max_sync < max_sector)
3460 max_sector = sector_nr + max_sync;
3463 int len = PAGE_SIZE;
3464 if (sector_nr + (len>>9) > max_sector)
3465 len = (max_sector - sector_nr) << 9;
3468 for (bio= biolist ; bio ; bio=bio->bi_next) {
3469 struct resync_pages *rp = get_resync_pages(bio);
3470 page = resync_fetch_page(rp, page_idx);
3472 * won't fail because the vec table is big enough
3473 * to hold all these pages
3475 bio_add_page(bio, page, len, 0);
3477 nr_sectors += len>>9;
3478 sector_nr += len>>9;
3479 } while (++page_idx < RESYNC_PAGES);
3480 r10_bio->sectors = nr_sectors;
3482 if (mddev_is_clustered(mddev) &&
3483 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3484 /* It is resync not recovery */
3485 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3486 conf->cluster_sync_low = mddev->curr_resync_completed;
3487 raid10_set_cluster_sync_high(conf);
3488 /* Send resync message */
3489 md_cluster_ops->resync_info_update(mddev,
3490 conf->cluster_sync_low,
3491 conf->cluster_sync_high);
3493 } else if (mddev_is_clustered(mddev)) {
3494 /* This is recovery not resync */
3495 sector_t sect_va1, sect_va2;
3496 bool broadcast_msg = false;
3498 for (i = 0; i < conf->geo.raid_disks; i++) {
3500 * sector_nr is a device address for recovery, so we
3501 * need translate it to array address before compare
3502 * with cluster_sync_high.
3504 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3506 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3507 broadcast_msg = true;
3509 * curr_resync_completed is similar as
3510 * sector_nr, so make the translation too.
3512 sect_va2 = raid10_find_virt(conf,
3513 mddev->curr_resync_completed, i);
3515 if (conf->cluster_sync_low == 0 ||
3516 conf->cluster_sync_low > sect_va2)
3517 conf->cluster_sync_low = sect_va2;
3520 if (broadcast_msg) {
3521 raid10_set_cluster_sync_high(conf);
3522 md_cluster_ops->resync_info_update(mddev,
3523 conf->cluster_sync_low,
3524 conf->cluster_sync_high);
3530 biolist = biolist->bi_next;
3532 bio->bi_next = NULL;
3533 r10_bio = get_resync_r10bio(bio);
3534 r10_bio->sectors = nr_sectors;
3536 if (bio->bi_end_io == end_sync_read) {
3537 md_sync_acct_bio(bio, nr_sectors);
3539 generic_make_request(bio);
3543 if (sectors_skipped)
3544 /* pretend they weren't skipped, it makes
3545 * no important difference in this case
3547 md_done_sync(mddev, sectors_skipped, 1);
3549 return sectors_skipped + nr_sectors;
3551 /* There is nowhere to write, so all non-sync
3552 * drives must be failed or in resync, all drives
3553 * have a bad block, so try the next chunk...
3555 if (sector_nr + max_sync < max_sector)
3556 max_sector = sector_nr + max_sync;
3558 sectors_skipped += (max_sector - sector_nr);
3560 sector_nr = max_sector;
3565 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3568 struct r10conf *conf = mddev->private;
3571 raid_disks = min(conf->geo.raid_disks,
3572 conf->prev.raid_disks);
3574 sectors = conf->dev_sectors;
3576 size = sectors >> conf->geo.chunk_shift;
3577 sector_div(size, conf->geo.far_copies);
3578 size = size * raid_disks;
3579 sector_div(size, conf->geo.near_copies);
3581 return size << conf->geo.chunk_shift;
3584 static void calc_sectors(struct r10conf *conf, sector_t size)
3586 /* Calculate the number of sectors-per-device that will
3587 * actually be used, and set conf->dev_sectors and
3591 size = size >> conf->geo.chunk_shift;
3592 sector_div(size, conf->geo.far_copies);
3593 size = size * conf->geo.raid_disks;
3594 sector_div(size, conf->geo.near_copies);
3595 /* 'size' is now the number of chunks in the array */
3596 /* calculate "used chunks per device" */
3597 size = size * conf->copies;
3599 /* We need to round up when dividing by raid_disks to
3600 * get the stride size.
3602 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3604 conf->dev_sectors = size << conf->geo.chunk_shift;
3606 if (conf->geo.far_offset)
3607 conf->geo.stride = 1 << conf->geo.chunk_shift;
3609 sector_div(size, conf->geo.far_copies);
3610 conf->geo.stride = size << conf->geo.chunk_shift;
3614 enum geo_type {geo_new, geo_old, geo_start};
3615 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3618 int layout, chunk, disks;
3621 layout = mddev->layout;
3622 chunk = mddev->chunk_sectors;
3623 disks = mddev->raid_disks - mddev->delta_disks;
3626 layout = mddev->new_layout;
3627 chunk = mddev->new_chunk_sectors;
3628 disks = mddev->raid_disks;
3630 default: /* avoid 'may be unused' warnings */
3631 case geo_start: /* new when starting reshape - raid_disks not
3633 layout = mddev->new_layout;
3634 chunk = mddev->new_chunk_sectors;
3635 disks = mddev->raid_disks + mddev->delta_disks;
3640 if (chunk < (PAGE_SIZE >> 9) ||
3641 !is_power_of_2(chunk))
3644 fc = (layout >> 8) & 255;
3645 fo = layout & (1<<16);
3646 geo->raid_disks = disks;
3647 geo->near_copies = nc;
3648 geo->far_copies = fc;
3649 geo->far_offset = fo;
3650 switch (layout >> 17) {
3651 case 0: /* original layout. simple but not always optimal */
3652 geo->far_set_size = disks;
3654 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3655 * actually using this, but leave code here just in case.*/
3656 geo->far_set_size = disks/fc;
3657 WARN(geo->far_set_size < fc,
3658 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3660 case 2: /* "improved" layout fixed to match documentation */
3661 geo->far_set_size = fc * nc;
3663 default: /* Not a valid layout */
3666 geo->chunk_mask = chunk - 1;
3667 geo->chunk_shift = ffz(~chunk);
3671 static struct r10conf *setup_conf(struct mddev *mddev)
3673 struct r10conf *conf = NULL;
3678 copies = setup_geo(&geo, mddev, geo_new);
3681 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3682 mdname(mddev), PAGE_SIZE);
3686 if (copies < 2 || copies > mddev->raid_disks) {
3687 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3688 mdname(mddev), mddev->new_layout);
3693 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3697 /* FIXME calc properly */
3698 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3699 sizeof(struct raid10_info),
3704 conf->tmppage = alloc_page(GFP_KERNEL);
3709 conf->copies = copies;
3710 err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3711 r10bio_pool_free, conf);
3715 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3719 calc_sectors(conf, mddev->dev_sectors);
3720 if (mddev->reshape_position == MaxSector) {
3721 conf->prev = conf->geo;
3722 conf->reshape_progress = MaxSector;
3724 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3728 conf->reshape_progress = mddev->reshape_position;
3729 if (conf->prev.far_offset)
3730 conf->prev.stride = 1 << conf->prev.chunk_shift;
3732 /* far_copies must be 1 */
3733 conf->prev.stride = conf->dev_sectors;
3735 conf->reshape_safe = conf->reshape_progress;
3736 spin_lock_init(&conf->device_lock);
3737 INIT_LIST_HEAD(&conf->retry_list);
3738 INIT_LIST_HEAD(&conf->bio_end_io_list);
3740 spin_lock_init(&conf->resync_lock);
3741 init_waitqueue_head(&conf->wait_barrier);
3742 atomic_set(&conf->nr_pending, 0);
3745 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3749 conf->mddev = mddev;
3754 mempool_exit(&conf->r10bio_pool);
3755 kfree(conf->mirrors);
3756 safe_put_page(conf->tmppage);
3757 bioset_exit(&conf->bio_split);
3760 return ERR_PTR(err);
3763 static int raid10_run(struct mddev *mddev)
3765 struct r10conf *conf;
3766 int i, disk_idx, chunk_size;
3767 struct raid10_info *disk;
3768 struct md_rdev *rdev;
3770 sector_t min_offset_diff = 0;
3772 bool discard_supported = false;
3774 if (mddev_init_writes_pending(mddev) < 0)
3777 if (mddev->private == NULL) {
3778 conf = setup_conf(mddev);
3780 return PTR_ERR(conf);
3781 mddev->private = conf;
3783 conf = mddev->private;
3787 if (mddev_is_clustered(conf->mddev)) {
3790 fc = (mddev->layout >> 8) & 255;
3791 fo = mddev->layout & (1<<16);
3792 if (fc > 1 || fo > 0) {
3793 pr_err("only near layout is supported by clustered"
3799 mddev->thread = conf->thread;
3800 conf->thread = NULL;
3802 chunk_size = mddev->chunk_sectors << 9;
3804 blk_queue_max_discard_sectors(mddev->queue,
3805 mddev->chunk_sectors);
3806 blk_queue_max_write_same_sectors(mddev->queue, 0);
3807 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3808 blk_queue_io_min(mddev->queue, chunk_size);
3809 if (conf->geo.raid_disks % conf->geo.near_copies)
3810 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3812 blk_queue_io_opt(mddev->queue, chunk_size *
3813 (conf->geo.raid_disks / conf->geo.near_copies));
3816 rdev_for_each(rdev, mddev) {
3819 disk_idx = rdev->raid_disk;
3822 if (disk_idx >= conf->geo.raid_disks &&
3823 disk_idx >= conf->prev.raid_disks)
3825 disk = conf->mirrors + disk_idx;
3827 if (test_bit(Replacement, &rdev->flags)) {
3828 if (disk->replacement)
3830 disk->replacement = rdev;
3836 diff = (rdev->new_data_offset - rdev->data_offset);
3837 if (!mddev->reshape_backwards)
3841 if (first || diff < min_offset_diff)
3842 min_offset_diff = diff;
3845 disk_stack_limits(mddev->gendisk, rdev->bdev,
3846 rdev->data_offset << 9);
3848 disk->head_position = 0;
3850 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3851 discard_supported = true;
3856 if (discard_supported)
3857 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3860 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3863 /* need to check that every block has at least one working mirror */
3864 if (!enough(conf, -1)) {
3865 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3870 if (conf->reshape_progress != MaxSector) {
3871 /* must ensure that shape change is supported */
3872 if (conf->geo.far_copies != 1 &&
3873 conf->geo.far_offset == 0)
3875 if (conf->prev.far_copies != 1 &&
3876 conf->prev.far_offset == 0)
3880 mddev->degraded = 0;
3882 i < conf->geo.raid_disks
3883 || i < conf->prev.raid_disks;
3886 disk = conf->mirrors + i;
3888 if (!disk->rdev && disk->replacement) {
3889 /* The replacement is all we have - use it */
3890 disk->rdev = disk->replacement;
3891 disk->replacement = NULL;
3892 clear_bit(Replacement, &disk->rdev->flags);
3896 !test_bit(In_sync, &disk->rdev->flags)) {
3897 disk->head_position = 0;
3900 disk->rdev->saved_raid_disk < 0)
3904 if (disk->replacement &&
3905 !test_bit(In_sync, &disk->replacement->flags) &&
3906 disk->replacement->saved_raid_disk < 0) {
3910 disk->recovery_disabled = mddev->recovery_disabled - 1;
3913 if (mddev->recovery_cp != MaxSector)
3914 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3916 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3917 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3918 conf->geo.raid_disks);
3920 * Ok, everything is just fine now
3922 mddev->dev_sectors = conf->dev_sectors;
3923 size = raid10_size(mddev, 0, 0);
3924 md_set_array_sectors(mddev, size);
3925 mddev->resync_max_sectors = size;
3926 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3929 int stripe = conf->geo.raid_disks *
3930 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3932 /* Calculate max read-ahead size.
3933 * We need to readahead at least twice a whole stripe....
3936 stripe /= conf->geo.near_copies;
3937 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3938 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3941 if (md_integrity_register(mddev))
3944 if (conf->reshape_progress != MaxSector) {
3945 unsigned long before_length, after_length;
3947 before_length = ((1 << conf->prev.chunk_shift) *
3948 conf->prev.far_copies);
3949 after_length = ((1 << conf->geo.chunk_shift) *
3950 conf->geo.far_copies);
3952 if (max(before_length, after_length) > min_offset_diff) {
3953 /* This cannot work */
3954 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3957 conf->offset_diff = min_offset_diff;
3959 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3960 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3961 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3962 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3963 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3965 if (!mddev->sync_thread)
3972 md_unregister_thread(&mddev->thread);
3973 mempool_exit(&conf->r10bio_pool);
3974 safe_put_page(conf->tmppage);
3975 kfree(conf->mirrors);
3977 mddev->private = NULL;
3982 static void raid10_free(struct mddev *mddev, void *priv)
3984 struct r10conf *conf = priv;
3986 mempool_exit(&conf->r10bio_pool);
3987 safe_put_page(conf->tmppage);
3988 kfree(conf->mirrors);
3989 kfree(conf->mirrors_old);
3990 kfree(conf->mirrors_new);
3991 bioset_exit(&conf->bio_split);
3995 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3997 struct r10conf *conf = mddev->private;
4000 raise_barrier(conf, 0);
4002 lower_barrier(conf);
4005 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4007 /* Resize of 'far' arrays is not supported.
4008 * For 'near' and 'offset' arrays we can set the
4009 * number of sectors used to be an appropriate multiple
4010 * of the chunk size.
4011 * For 'offset', this is far_copies*chunksize.
4012 * For 'near' the multiplier is the LCM of
4013 * near_copies and raid_disks.
4014 * So if far_copies > 1 && !far_offset, fail.
4015 * Else find LCM(raid_disks, near_copy)*far_copies and
4016 * multiply by chunk_size. Then round to this number.
4017 * This is mostly done by raid10_size()
4019 struct r10conf *conf = mddev->private;
4020 sector_t oldsize, size;
4022 if (mddev->reshape_position != MaxSector)
4025 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4028 oldsize = raid10_size(mddev, 0, 0);
4029 size = raid10_size(mddev, sectors, 0);
4030 if (mddev->external_size &&
4031 mddev->array_sectors > size)
4033 if (mddev->bitmap) {
4034 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4038 md_set_array_sectors(mddev, size);
4039 if (sectors > mddev->dev_sectors &&
4040 mddev->recovery_cp > oldsize) {
4041 mddev->recovery_cp = oldsize;
4042 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4044 calc_sectors(conf, sectors);
4045 mddev->dev_sectors = conf->dev_sectors;
4046 mddev->resync_max_sectors = size;
4050 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4052 struct md_rdev *rdev;
4053 struct r10conf *conf;
4055 if (mddev->degraded > 0) {
4056 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4058 return ERR_PTR(-EINVAL);
4060 sector_div(size, devs);
4062 /* Set new parameters */
4063 mddev->new_level = 10;
4064 /* new layout: far_copies = 1, near_copies = 2 */
4065 mddev->new_layout = (1<<8) + 2;
4066 mddev->new_chunk_sectors = mddev->chunk_sectors;
4067 mddev->delta_disks = mddev->raid_disks;
4068 mddev->raid_disks *= 2;
4069 /* make sure it will be not marked as dirty */
4070 mddev->recovery_cp = MaxSector;
4071 mddev->dev_sectors = size;
4073 conf = setup_conf(mddev);
4074 if (!IS_ERR(conf)) {
4075 rdev_for_each(rdev, mddev)
4076 if (rdev->raid_disk >= 0) {
4077 rdev->new_raid_disk = rdev->raid_disk * 2;
4078 rdev->sectors = size;
4086 static void *raid10_takeover(struct mddev *mddev)
4088 struct r0conf *raid0_conf;
4090 /* raid10 can take over:
4091 * raid0 - providing it has only two drives
4093 if (mddev->level == 0) {
4094 /* for raid0 takeover only one zone is supported */
4095 raid0_conf = mddev->private;
4096 if (raid0_conf->nr_strip_zones > 1) {
4097 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4099 return ERR_PTR(-EINVAL);
4101 return raid10_takeover_raid0(mddev,
4102 raid0_conf->strip_zone->zone_end,
4103 raid0_conf->strip_zone->nb_dev);
4105 return ERR_PTR(-EINVAL);
4108 static int raid10_check_reshape(struct mddev *mddev)
4110 /* Called when there is a request to change
4111 * - layout (to ->new_layout)
4112 * - chunk size (to ->new_chunk_sectors)
4113 * - raid_disks (by delta_disks)
4114 * or when trying to restart a reshape that was ongoing.
4116 * We need to validate the request and possibly allocate
4117 * space if that might be an issue later.
4119 * Currently we reject any reshape of a 'far' mode array,
4120 * allow chunk size to change if new is generally acceptable,
4121 * allow raid_disks to increase, and allow
4122 * a switch between 'near' mode and 'offset' mode.
4124 struct r10conf *conf = mddev->private;
4127 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4130 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4131 /* mustn't change number of copies */
4133 if (geo.far_copies > 1 && !geo.far_offset)
4134 /* Cannot switch to 'far' mode */
4137 if (mddev->array_sectors & geo.chunk_mask)
4138 /* not factor of array size */
4141 if (!enough(conf, -1))
4144 kfree(conf->mirrors_new);
4145 conf->mirrors_new = NULL;
4146 if (mddev->delta_disks > 0) {
4147 /* allocate new 'mirrors' list */
4149 kcalloc(mddev->raid_disks + mddev->delta_disks,
4150 sizeof(struct raid10_info),
4152 if (!conf->mirrors_new)
4159 * Need to check if array has failed when deciding whether to:
4161 * - remove non-faulty devices
4164 * This determination is simple when no reshape is happening.
4165 * However if there is a reshape, we need to carefully check
4166 * both the before and after sections.
4167 * This is because some failed devices may only affect one
4168 * of the two sections, and some non-in_sync devices may
4169 * be insync in the section most affected by failed devices.
4171 static int calc_degraded(struct r10conf *conf)
4173 int degraded, degraded2;
4178 /* 'prev' section first */
4179 for (i = 0; i < conf->prev.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 /* When we can reduce the number of devices in
4185 * an array, this might not contribute to
4186 * 'degraded'. It does now.
4191 if (conf->geo.raid_disks == conf->prev.raid_disks)
4195 for (i = 0; i < conf->geo.raid_disks; i++) {
4196 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4197 if (!rdev || test_bit(Faulty, &rdev->flags))
4199 else if (!test_bit(In_sync, &rdev->flags)) {
4200 /* If reshape is increasing the number of devices,
4201 * this section has already been recovered, so
4202 * it doesn't contribute to degraded.
4205 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4210 if (degraded2 > degraded)
4215 static int raid10_start_reshape(struct mddev *mddev)
4217 /* A 'reshape' has been requested. This commits
4218 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4219 * This also checks if there are enough spares and adds them
4221 * We currently require enough spares to make the final
4222 * array non-degraded. We also require that the difference
4223 * between old and new data_offset - on each device - is
4224 * enough that we never risk over-writing.
4227 unsigned long before_length, after_length;
4228 sector_t min_offset_diff = 0;
4231 struct r10conf *conf = mddev->private;
4232 struct md_rdev *rdev;
4236 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4239 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4242 before_length = ((1 << conf->prev.chunk_shift) *
4243 conf->prev.far_copies);
4244 after_length = ((1 << conf->geo.chunk_shift) *
4245 conf->geo.far_copies);
4247 rdev_for_each(rdev, mddev) {
4248 if (!test_bit(In_sync, &rdev->flags)
4249 && !test_bit(Faulty, &rdev->flags))
4251 if (rdev->raid_disk >= 0) {
4252 long long diff = (rdev->new_data_offset
4253 - rdev->data_offset);
4254 if (!mddev->reshape_backwards)
4258 if (first || diff < min_offset_diff)
4259 min_offset_diff = diff;
4264 if (max(before_length, after_length) > min_offset_diff)
4267 if (spares < mddev->delta_disks)
4270 conf->offset_diff = min_offset_diff;
4271 spin_lock_irq(&conf->device_lock);
4272 if (conf->mirrors_new) {
4273 memcpy(conf->mirrors_new, conf->mirrors,
4274 sizeof(struct raid10_info)*conf->prev.raid_disks);
4276 kfree(conf->mirrors_old);
4277 conf->mirrors_old = conf->mirrors;
4278 conf->mirrors = conf->mirrors_new;
4279 conf->mirrors_new = NULL;
4281 setup_geo(&conf->geo, mddev, geo_start);
4283 if (mddev->reshape_backwards) {
4284 sector_t size = raid10_size(mddev, 0, 0);
4285 if (size < mddev->array_sectors) {
4286 spin_unlock_irq(&conf->device_lock);
4287 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4291 mddev->resync_max_sectors = size;
4292 conf->reshape_progress = size;
4294 conf->reshape_progress = 0;
4295 conf->reshape_safe = conf->reshape_progress;
4296 spin_unlock_irq(&conf->device_lock);
4298 if (mddev->delta_disks && mddev->bitmap) {
4299 ret = md_bitmap_resize(mddev->bitmap,
4300 raid10_size(mddev, 0, conf->geo.raid_disks),
4305 if (mddev->delta_disks > 0) {
4306 rdev_for_each(rdev, mddev)
4307 if (rdev->raid_disk < 0 &&
4308 !test_bit(Faulty, &rdev->flags)) {
4309 if (raid10_add_disk(mddev, rdev) == 0) {
4310 if (rdev->raid_disk >=
4311 conf->prev.raid_disks)
4312 set_bit(In_sync, &rdev->flags);
4314 rdev->recovery_offset = 0;
4316 if (sysfs_link_rdev(mddev, rdev))
4317 /* Failure here is OK */;
4319 } else if (rdev->raid_disk >= conf->prev.raid_disks
4320 && !test_bit(Faulty, &rdev->flags)) {
4321 /* This is a spare that was manually added */
4322 set_bit(In_sync, &rdev->flags);
4325 /* When a reshape changes the number of devices,
4326 * ->degraded is measured against the larger of the
4327 * pre and post numbers.
4329 spin_lock_irq(&conf->device_lock);
4330 mddev->degraded = calc_degraded(conf);
4331 spin_unlock_irq(&conf->device_lock);
4332 mddev->raid_disks = conf->geo.raid_disks;
4333 mddev->reshape_position = conf->reshape_progress;
4334 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4336 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4337 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4338 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4339 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4340 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4342 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4344 if (!mddev->sync_thread) {
4348 conf->reshape_checkpoint = jiffies;
4349 md_wakeup_thread(mddev->sync_thread);
4350 md_new_event(mddev);
4354 mddev->recovery = 0;
4355 spin_lock_irq(&conf->device_lock);
4356 conf->geo = conf->prev;
4357 mddev->raid_disks = conf->geo.raid_disks;
4358 rdev_for_each(rdev, mddev)
4359 rdev->new_data_offset = rdev->data_offset;
4361 conf->reshape_progress = MaxSector;
4362 conf->reshape_safe = MaxSector;
4363 mddev->reshape_position = MaxSector;
4364 spin_unlock_irq(&conf->device_lock);
4368 /* Calculate the last device-address that could contain
4369 * any block from the chunk that includes the array-address 's'
4370 * and report the next address.
4371 * i.e. the address returned will be chunk-aligned and after
4372 * any data that is in the chunk containing 's'.
4374 static sector_t last_dev_address(sector_t s, struct geom *geo)
4376 s = (s | geo->chunk_mask) + 1;
4377 s >>= geo->chunk_shift;
4378 s *= geo->near_copies;
4379 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4380 s *= geo->far_copies;
4381 s <<= geo->chunk_shift;
4385 /* Calculate the first device-address that could contain
4386 * any block from the chunk that includes the array-address 's'.
4387 * This too will be the start of a chunk
4389 static sector_t first_dev_address(sector_t s, struct geom *geo)
4391 s >>= geo->chunk_shift;
4392 s *= geo->near_copies;
4393 sector_div(s, geo->raid_disks);
4394 s *= geo->far_copies;
4395 s <<= geo->chunk_shift;
4399 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4402 /* We simply copy at most one chunk (smallest of old and new)
4403 * at a time, possibly less if that exceeds RESYNC_PAGES,
4404 * or we hit a bad block or something.
4405 * This might mean we pause for normal IO in the middle of
4406 * a chunk, but that is not a problem as mddev->reshape_position
4407 * can record any location.
4409 * If we will want to write to a location that isn't
4410 * yet recorded as 'safe' (i.e. in metadata on disk) then
4411 * we need to flush all reshape requests and update the metadata.
4413 * When reshaping forwards (e.g. to more devices), we interpret
4414 * 'safe' as the earliest block which might not have been copied
4415 * down yet. We divide this by previous stripe size and multiply
4416 * by previous stripe length to get lowest device offset that we
4417 * cannot write to yet.
4418 * We interpret 'sector_nr' as an address that we want to write to.
4419 * From this we use last_device_address() to find where we might
4420 * write to, and first_device_address on the 'safe' position.
4421 * If this 'next' write position is after the 'safe' position,
4422 * we must update the metadata to increase the 'safe' position.
4424 * When reshaping backwards, we round in the opposite direction
4425 * and perform the reverse test: next write position must not be
4426 * less than current safe position.
4428 * In all this the minimum difference in data offsets
4429 * (conf->offset_diff - always positive) allows a bit of slack,
4430 * so next can be after 'safe', but not by more than offset_diff
4432 * We need to prepare all the bios here before we start any IO
4433 * to ensure the size we choose is acceptable to all devices.
4434 * The means one for each copy for write-out and an extra one for
4436 * We store the read-in bio in ->master_bio and the others in
4437 * ->devs[x].bio and ->devs[x].repl_bio.
4439 struct r10conf *conf = mddev->private;
4440 struct r10bio *r10_bio;
4441 sector_t next, safe, last;
4445 struct md_rdev *rdev;
4448 struct bio *bio, *read_bio;
4449 int sectors_done = 0;
4450 struct page **pages;
4452 if (sector_nr == 0) {
4453 /* If restarting in the middle, skip the initial sectors */
4454 if (mddev->reshape_backwards &&
4455 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4456 sector_nr = (raid10_size(mddev, 0, 0)
4457 - conf->reshape_progress);
4458 } else if (!mddev->reshape_backwards &&
4459 conf->reshape_progress > 0)
4460 sector_nr = conf->reshape_progress;
4462 mddev->curr_resync_completed = sector_nr;
4463 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4469 /* We don't use sector_nr to track where we are up to
4470 * as that doesn't work well for ->reshape_backwards.
4471 * So just use ->reshape_progress.
4473 if (mddev->reshape_backwards) {
4474 /* 'next' is the earliest device address that we might
4475 * write to for this chunk in the new layout
4477 next = first_dev_address(conf->reshape_progress - 1,
4480 /* 'safe' is the last device address that we might read from
4481 * in the old layout after a restart
4483 safe = last_dev_address(conf->reshape_safe - 1,
4486 if (next + conf->offset_diff < safe)
4489 last = conf->reshape_progress - 1;
4490 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4491 & conf->prev.chunk_mask);
4492 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4493 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4495 /* 'next' is after the last device address that we
4496 * might write to for this chunk in the new layout
4498 next = last_dev_address(conf->reshape_progress, &conf->geo);
4500 /* 'safe' is the earliest device address that we might
4501 * read from in the old layout after a restart
4503 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4505 /* Need to update metadata if 'next' might be beyond 'safe'
4506 * as that would possibly corrupt data
4508 if (next > safe + conf->offset_diff)
4511 sector_nr = conf->reshape_progress;
4512 last = sector_nr | (conf->geo.chunk_mask
4513 & conf->prev.chunk_mask);
4515 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4516 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4520 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4521 /* Need to update reshape_position in metadata */
4523 mddev->reshape_position = conf->reshape_progress;
4524 if (mddev->reshape_backwards)
4525 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4526 - conf->reshape_progress;
4528 mddev->curr_resync_completed = conf->reshape_progress;
4529 conf->reshape_checkpoint = jiffies;
4530 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4531 md_wakeup_thread(mddev->thread);
4532 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4533 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4534 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4535 allow_barrier(conf);
4536 return sectors_done;
4538 conf->reshape_safe = mddev->reshape_position;
4539 allow_barrier(conf);
4542 raise_barrier(conf, 0);
4544 /* Now schedule reads for blocks from sector_nr to last */
4545 r10_bio = raid10_alloc_init_r10buf(conf);
4547 raise_barrier(conf, 1);
4548 atomic_set(&r10_bio->remaining, 0);
4549 r10_bio->mddev = mddev;
4550 r10_bio->sector = sector_nr;
4551 set_bit(R10BIO_IsReshape, &r10_bio->state);
4552 r10_bio->sectors = last - sector_nr + 1;
4553 rdev = read_balance(conf, r10_bio, &max_sectors);
4554 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4557 /* Cannot read from here, so need to record bad blocks
4558 * on all the target devices.
4561 mempool_free(r10_bio, &conf->r10buf_pool);
4562 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4563 return sectors_done;
4566 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4568 bio_set_dev(read_bio, rdev->bdev);
4569 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4570 + rdev->data_offset);
4571 read_bio->bi_private = r10_bio;
4572 read_bio->bi_end_io = end_reshape_read;
4573 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4574 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4575 read_bio->bi_status = 0;
4576 read_bio->bi_vcnt = 0;
4577 read_bio->bi_iter.bi_size = 0;
4578 r10_bio->master_bio = read_bio;
4579 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4581 /* Now find the locations in the new layout */
4582 __raid10_find_phys(&conf->geo, r10_bio);
4585 read_bio->bi_next = NULL;
4588 for (s = 0; s < conf->copies*2; s++) {
4590 int d = r10_bio->devs[s/2].devnum;
4591 struct md_rdev *rdev2;
4593 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4594 b = r10_bio->devs[s/2].repl_bio;
4596 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4597 b = r10_bio->devs[s/2].bio;
4599 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4602 bio_set_dev(b, rdev2->bdev);
4603 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4604 rdev2->new_data_offset;
4605 b->bi_end_io = end_reshape_write;
4606 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4611 /* Now add as many pages as possible to all of these bios. */
4614 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4615 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4616 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4617 int len = (max_sectors - s) << 9;
4618 if (len > PAGE_SIZE)
4620 for (bio = blist; bio ; bio = bio->bi_next) {
4622 * won't fail because the vec table is big enough
4623 * to hold all these pages
4625 bio_add_page(bio, page, len, 0);
4627 sector_nr += len >> 9;
4628 nr_sectors += len >> 9;
4631 r10_bio->sectors = nr_sectors;
4633 /* Now submit the read */
4634 md_sync_acct_bio(read_bio, r10_bio->sectors);
4635 atomic_inc(&r10_bio->remaining);
4636 read_bio->bi_next = NULL;
4637 generic_make_request(read_bio);
4638 sectors_done += nr_sectors;
4639 if (sector_nr <= last)
4642 lower_barrier(conf);
4644 /* Now that we have done the whole section we can
4645 * update reshape_progress
4647 if (mddev->reshape_backwards)
4648 conf->reshape_progress -= sectors_done;
4650 conf->reshape_progress += sectors_done;
4652 return sectors_done;
4655 static void end_reshape_request(struct r10bio *r10_bio);
4656 static int handle_reshape_read_error(struct mddev *mddev,
4657 struct r10bio *r10_bio);
4658 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4660 /* Reshape read completed. Hopefully we have a block
4662 * If we got a read error then we do sync 1-page reads from
4663 * elsewhere until we find the data - or give up.
4665 struct r10conf *conf = mddev->private;
4668 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4669 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4670 /* Reshape has been aborted */
4671 md_done_sync(mddev, r10_bio->sectors, 0);
4675 /* We definitely have the data in the pages, schedule the
4678 atomic_set(&r10_bio->remaining, 1);
4679 for (s = 0; s < conf->copies*2; s++) {
4681 int d = r10_bio->devs[s/2].devnum;
4682 struct md_rdev *rdev;
4685 rdev = rcu_dereference(conf->mirrors[d].replacement);
4686 b = r10_bio->devs[s/2].repl_bio;
4688 rdev = rcu_dereference(conf->mirrors[d].rdev);
4689 b = r10_bio->devs[s/2].bio;
4691 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4695 atomic_inc(&rdev->nr_pending);
4697 md_sync_acct_bio(b, r10_bio->sectors);
4698 atomic_inc(&r10_bio->remaining);
4700 generic_make_request(b);
4702 end_reshape_request(r10_bio);
4705 static void end_reshape(struct r10conf *conf)
4707 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4710 spin_lock_irq(&conf->device_lock);
4711 conf->prev = conf->geo;
4712 md_finish_reshape(conf->mddev);
4714 conf->reshape_progress = MaxSector;
4715 conf->reshape_safe = MaxSector;
4716 spin_unlock_irq(&conf->device_lock);
4718 /* read-ahead size must cover two whole stripes, which is
4719 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4721 if (conf->mddev->queue) {
4722 int stripe = conf->geo.raid_disks *
4723 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4724 stripe /= conf->geo.near_copies;
4725 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4726 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4731 static int handle_reshape_read_error(struct mddev *mddev,
4732 struct r10bio *r10_bio)
4734 /* Use sync reads to get the blocks from somewhere else */
4735 int sectors = r10_bio->sectors;
4736 struct r10conf *conf = mddev->private;
4737 struct r10bio *r10b;
4740 struct page **pages;
4742 r10b = kmalloc(sizeof(*r10b) +
4743 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4745 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4749 /* reshape IOs share pages from .devs[0].bio */
4750 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4752 r10b->sector = r10_bio->sector;
4753 __raid10_find_phys(&conf->prev, r10b);
4758 int first_slot = slot;
4760 if (s > (PAGE_SIZE >> 9))
4765 int d = r10b->devs[slot].devnum;
4766 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4769 test_bit(Faulty, &rdev->flags) ||
4770 !test_bit(In_sync, &rdev->flags))
4773 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4774 atomic_inc(&rdev->nr_pending);
4776 success = sync_page_io(rdev,
4780 REQ_OP_READ, 0, false);
4781 rdev_dec_pending(rdev, mddev);
4787 if (slot >= conf->copies)
4789 if (slot == first_slot)
4794 /* couldn't read this block, must give up */
4795 set_bit(MD_RECOVERY_INTR,
4807 static void end_reshape_write(struct bio *bio)
4809 struct r10bio *r10_bio = get_resync_r10bio(bio);
4810 struct mddev *mddev = r10_bio->mddev;
4811 struct r10conf *conf = mddev->private;
4815 struct md_rdev *rdev = NULL;
4817 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4819 rdev = conf->mirrors[d].replacement;
4822 rdev = conf->mirrors[d].rdev;
4825 if (bio->bi_status) {
4826 /* FIXME should record badblock */
4827 md_error(mddev, rdev);
4830 rdev_dec_pending(rdev, mddev);
4831 end_reshape_request(r10_bio);
4834 static void end_reshape_request(struct r10bio *r10_bio)
4836 if (!atomic_dec_and_test(&r10_bio->remaining))
4838 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4839 bio_put(r10_bio->master_bio);
4843 static void raid10_finish_reshape(struct mddev *mddev)
4845 struct r10conf *conf = mddev->private;
4847 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4850 if (mddev->delta_disks > 0) {
4851 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4852 mddev->recovery_cp = mddev->resync_max_sectors;
4853 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4855 mddev->resync_max_sectors = mddev->array_sectors;
4859 for (d = conf->geo.raid_disks ;
4860 d < conf->geo.raid_disks - mddev->delta_disks;
4862 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4864 clear_bit(In_sync, &rdev->flags);
4865 rdev = rcu_dereference(conf->mirrors[d].replacement);
4867 clear_bit(In_sync, &rdev->flags);
4871 mddev->layout = mddev->new_layout;
4872 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4873 mddev->reshape_position = MaxSector;
4874 mddev->delta_disks = 0;
4875 mddev->reshape_backwards = 0;
4878 static struct md_personality raid10_personality =
4882 .owner = THIS_MODULE,
4883 .make_request = raid10_make_request,
4885 .free = raid10_free,
4886 .status = raid10_status,
4887 .error_handler = raid10_error,
4888 .hot_add_disk = raid10_add_disk,
4889 .hot_remove_disk= raid10_remove_disk,
4890 .spare_active = raid10_spare_active,
4891 .sync_request = raid10_sync_request,
4892 .quiesce = raid10_quiesce,
4893 .size = raid10_size,
4894 .resize = raid10_resize,
4895 .takeover = raid10_takeover,
4896 .check_reshape = raid10_check_reshape,
4897 .start_reshape = raid10_start_reshape,
4898 .finish_reshape = raid10_finish_reshape,
4899 .congested = raid10_congested,
4902 static int __init raid_init(void)
4904 return register_md_personality(&raid10_personality);
4907 static void raid_exit(void)
4909 unregister_md_personality(&raid10_personality);
4912 module_init(raid_init);
4913 module_exit(raid_exit);
4914 MODULE_LICENSE("GPL");
4915 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4916 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4917 MODULE_ALIAS("md-raid10");
4918 MODULE_ALIAS("md-level-10");
4920 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);