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 = r10_bio->devs[1].bio;
2270 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2272 /* Need to test wbio2->bi_end_io before we call
2273 * generic_make_request as if the former is NULL,
2274 * the latter is free to free wbio2.
2276 if (wbio2 && !wbio2->bi_end_io)
2279 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2280 fix_recovery_read_error(r10_bio);
2281 if (wbio->bi_end_io)
2282 end_sync_request(r10_bio);
2284 end_sync_request(r10_bio);
2289 * share the pages with the first bio
2290 * and submit the write request
2292 d = r10_bio->devs[1].devnum;
2293 if (wbio->bi_end_io) {
2294 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2295 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2296 generic_make_request(wbio);
2299 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2300 md_sync_acct(conf->mirrors[d].replacement->bdev,
2301 bio_sectors(wbio2));
2302 generic_make_request(wbio2);
2307 * Used by fix_read_error() to decay the per rdev read_errors.
2308 * We halve the read error count for every hour that has elapsed
2309 * since the last recorded read error.
2312 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2315 unsigned long hours_since_last;
2316 unsigned int read_errors = atomic_read(&rdev->read_errors);
2318 cur_time_mon = ktime_get_seconds();
2320 if (rdev->last_read_error == 0) {
2321 /* first time we've seen a read error */
2322 rdev->last_read_error = cur_time_mon;
2326 hours_since_last = (long)(cur_time_mon -
2327 rdev->last_read_error) / 3600;
2329 rdev->last_read_error = cur_time_mon;
2332 * if hours_since_last is > the number of bits in read_errors
2333 * just set read errors to 0. We do this to avoid
2334 * overflowing the shift of read_errors by hours_since_last.
2336 if (hours_since_last >= 8 * sizeof(read_errors))
2337 atomic_set(&rdev->read_errors, 0);
2339 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2342 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2343 int sectors, struct page *page, int rw)
2348 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2349 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2351 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2355 set_bit(WriteErrorSeen, &rdev->flags);
2356 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2357 set_bit(MD_RECOVERY_NEEDED,
2358 &rdev->mddev->recovery);
2360 /* need to record an error - either for the block or the device */
2361 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2362 md_error(rdev->mddev, rdev);
2367 * This is a kernel thread which:
2369 * 1. Retries failed read operations on working mirrors.
2370 * 2. Updates the raid superblock when problems encounter.
2371 * 3. Performs writes following reads for array synchronising.
2374 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2376 int sect = 0; /* Offset from r10_bio->sector */
2377 int sectors = r10_bio->sectors;
2378 struct md_rdev *rdev;
2379 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2380 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2382 /* still own a reference to this rdev, so it cannot
2383 * have been cleared recently.
2385 rdev = conf->mirrors[d].rdev;
2387 if (test_bit(Faulty, &rdev->flags))
2388 /* drive has already been failed, just ignore any
2389 more fix_read_error() attempts */
2392 check_decay_read_errors(mddev, rdev);
2393 atomic_inc(&rdev->read_errors);
2394 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2395 char b[BDEVNAME_SIZE];
2396 bdevname(rdev->bdev, b);
2398 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2400 atomic_read(&rdev->read_errors), max_read_errors);
2401 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2403 md_error(mddev, rdev);
2404 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2410 int sl = r10_bio->read_slot;
2414 if (s > (PAGE_SIZE>>9))
2422 d = r10_bio->devs[sl].devnum;
2423 rdev = rcu_dereference(conf->mirrors[d].rdev);
2425 test_bit(In_sync, &rdev->flags) &&
2426 !test_bit(Faulty, &rdev->flags) &&
2427 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2428 &first_bad, &bad_sectors) == 0) {
2429 atomic_inc(&rdev->nr_pending);
2431 success = sync_page_io(rdev,
2432 r10_bio->devs[sl].addr +
2436 REQ_OP_READ, 0, false);
2437 rdev_dec_pending(rdev, mddev);
2443 if (sl == conf->copies)
2445 } while (!success && sl != r10_bio->read_slot);
2449 /* Cannot read from anywhere, just mark the block
2450 * as bad on the first device to discourage future
2453 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2454 rdev = conf->mirrors[dn].rdev;
2456 if (!rdev_set_badblocks(
2458 r10_bio->devs[r10_bio->read_slot].addr
2461 md_error(mddev, rdev);
2462 r10_bio->devs[r10_bio->read_slot].bio
2469 /* write it back and re-read */
2471 while (sl != r10_bio->read_slot) {
2472 char b[BDEVNAME_SIZE];
2477 d = r10_bio->devs[sl].devnum;
2478 rdev = rcu_dereference(conf->mirrors[d].rdev);
2480 test_bit(Faulty, &rdev->flags) ||
2481 !test_bit(In_sync, &rdev->flags))
2484 atomic_inc(&rdev->nr_pending);
2486 if (r10_sync_page_io(rdev,
2487 r10_bio->devs[sl].addr +
2489 s, conf->tmppage, WRITE)
2491 /* Well, this device is dead */
2492 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2494 (unsigned long long)(
2496 choose_data_offset(r10_bio,
2498 bdevname(rdev->bdev, b));
2499 pr_notice("md/raid10:%s: %s: failing drive\n",
2501 bdevname(rdev->bdev, b));
2503 rdev_dec_pending(rdev, mddev);
2507 while (sl != r10_bio->read_slot) {
2508 char b[BDEVNAME_SIZE];
2513 d = r10_bio->devs[sl].devnum;
2514 rdev = rcu_dereference(conf->mirrors[d].rdev);
2516 test_bit(Faulty, &rdev->flags) ||
2517 !test_bit(In_sync, &rdev->flags))
2520 atomic_inc(&rdev->nr_pending);
2522 switch (r10_sync_page_io(rdev,
2523 r10_bio->devs[sl].addr +
2528 /* Well, this device is dead */
2529 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2531 (unsigned long long)(
2533 choose_data_offset(r10_bio, rdev)),
2534 bdevname(rdev->bdev, b));
2535 pr_notice("md/raid10:%s: %s: failing drive\n",
2537 bdevname(rdev->bdev, b));
2540 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2542 (unsigned long long)(
2544 choose_data_offset(r10_bio, rdev)),
2545 bdevname(rdev->bdev, b));
2546 atomic_add(s, &rdev->corrected_errors);
2549 rdev_dec_pending(rdev, mddev);
2559 static int narrow_write_error(struct r10bio *r10_bio, int i)
2561 struct bio *bio = r10_bio->master_bio;
2562 struct mddev *mddev = r10_bio->mddev;
2563 struct r10conf *conf = mddev->private;
2564 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2565 /* bio has the data to be written to slot 'i' where
2566 * we just recently had a write error.
2567 * We repeatedly clone the bio and trim down to one block,
2568 * then try the write. Where the write fails we record
2570 * It is conceivable that the bio doesn't exactly align with
2571 * blocks. We must handle this.
2573 * We currently own a reference to the rdev.
2579 int sect_to_write = r10_bio->sectors;
2582 if (rdev->badblocks.shift < 0)
2585 block_sectors = roundup(1 << rdev->badblocks.shift,
2586 bdev_logical_block_size(rdev->bdev) >> 9);
2587 sector = r10_bio->sector;
2588 sectors = ((r10_bio->sector + block_sectors)
2589 & ~(sector_t)(block_sectors - 1))
2592 while (sect_to_write) {
2595 if (sectors > sect_to_write)
2596 sectors = sect_to_write;
2597 /* Write at 'sector' for 'sectors' */
2598 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2599 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2600 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2601 wbio->bi_iter.bi_sector = wsector +
2602 choose_data_offset(r10_bio, rdev);
2603 bio_set_dev(wbio, rdev->bdev);
2604 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2606 if (submit_bio_wait(wbio) < 0)
2608 ok = rdev_set_badblocks(rdev, wsector,
2613 sect_to_write -= sectors;
2615 sectors = block_sectors;
2620 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2622 int slot = r10_bio->read_slot;
2624 struct r10conf *conf = mddev->private;
2625 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2627 /* we got a read error. Maybe the drive is bad. Maybe just
2628 * the block and we can fix it.
2629 * We freeze all other IO, and try reading the block from
2630 * other devices. When we find one, we re-write
2631 * and check it that fixes the read error.
2632 * This is all done synchronously while the array is
2635 bio = r10_bio->devs[slot].bio;
2637 r10_bio->devs[slot].bio = NULL;
2640 r10_bio->devs[slot].bio = IO_BLOCKED;
2641 else if (!test_bit(FailFast, &rdev->flags)) {
2642 freeze_array(conf, 1);
2643 fix_read_error(conf, mddev, r10_bio);
2644 unfreeze_array(conf);
2646 md_error(mddev, rdev);
2648 rdev_dec_pending(rdev, mddev);
2649 allow_barrier(conf);
2651 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2654 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2656 /* Some sort of write request has finished and it
2657 * succeeded in writing where we thought there was a
2658 * bad block. So forget the bad block.
2659 * Or possibly if failed and we need to record
2663 struct md_rdev *rdev;
2665 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2666 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2667 for (m = 0; m < conf->copies; m++) {
2668 int dev = r10_bio->devs[m].devnum;
2669 rdev = conf->mirrors[dev].rdev;
2670 if (r10_bio->devs[m].bio == NULL ||
2671 r10_bio->devs[m].bio->bi_end_io == NULL)
2673 if (!r10_bio->devs[m].bio->bi_status) {
2674 rdev_clear_badblocks(
2676 r10_bio->devs[m].addr,
2677 r10_bio->sectors, 0);
2679 if (!rdev_set_badblocks(
2681 r10_bio->devs[m].addr,
2682 r10_bio->sectors, 0))
2683 md_error(conf->mddev, rdev);
2685 rdev = conf->mirrors[dev].replacement;
2686 if (r10_bio->devs[m].repl_bio == NULL ||
2687 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2690 if (!r10_bio->devs[m].repl_bio->bi_status) {
2691 rdev_clear_badblocks(
2693 r10_bio->devs[m].addr,
2694 r10_bio->sectors, 0);
2696 if (!rdev_set_badblocks(
2698 r10_bio->devs[m].addr,
2699 r10_bio->sectors, 0))
2700 md_error(conf->mddev, rdev);
2706 for (m = 0; m < conf->copies; m++) {
2707 int dev = r10_bio->devs[m].devnum;
2708 struct bio *bio = r10_bio->devs[m].bio;
2709 rdev = conf->mirrors[dev].rdev;
2710 if (bio == IO_MADE_GOOD) {
2711 rdev_clear_badblocks(
2713 r10_bio->devs[m].addr,
2714 r10_bio->sectors, 0);
2715 rdev_dec_pending(rdev, conf->mddev);
2716 } else if (bio != NULL && bio->bi_status) {
2718 if (!narrow_write_error(r10_bio, m)) {
2719 md_error(conf->mddev, rdev);
2720 set_bit(R10BIO_Degraded,
2723 rdev_dec_pending(rdev, conf->mddev);
2725 bio = r10_bio->devs[m].repl_bio;
2726 rdev = conf->mirrors[dev].replacement;
2727 if (rdev && bio == IO_MADE_GOOD) {
2728 rdev_clear_badblocks(
2730 r10_bio->devs[m].addr,
2731 r10_bio->sectors, 0);
2732 rdev_dec_pending(rdev, conf->mddev);
2736 spin_lock_irq(&conf->device_lock);
2737 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2739 spin_unlock_irq(&conf->device_lock);
2741 * In case freeze_array() is waiting for condition
2742 * nr_pending == nr_queued + extra to be true.
2744 wake_up(&conf->wait_barrier);
2745 md_wakeup_thread(conf->mddev->thread);
2747 if (test_bit(R10BIO_WriteError,
2749 close_write(r10_bio);
2750 raid_end_bio_io(r10_bio);
2755 static void raid10d(struct md_thread *thread)
2757 struct mddev *mddev = thread->mddev;
2758 struct r10bio *r10_bio;
2759 unsigned long flags;
2760 struct r10conf *conf = mddev->private;
2761 struct list_head *head = &conf->retry_list;
2762 struct blk_plug plug;
2764 md_check_recovery(mddev);
2766 if (!list_empty_careful(&conf->bio_end_io_list) &&
2767 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2769 spin_lock_irqsave(&conf->device_lock, flags);
2770 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2771 while (!list_empty(&conf->bio_end_io_list)) {
2772 list_move(conf->bio_end_io_list.prev, &tmp);
2776 spin_unlock_irqrestore(&conf->device_lock, flags);
2777 while (!list_empty(&tmp)) {
2778 r10_bio = list_first_entry(&tmp, struct r10bio,
2780 list_del(&r10_bio->retry_list);
2781 if (mddev->degraded)
2782 set_bit(R10BIO_Degraded, &r10_bio->state);
2784 if (test_bit(R10BIO_WriteError,
2786 close_write(r10_bio);
2787 raid_end_bio_io(r10_bio);
2791 blk_start_plug(&plug);
2794 flush_pending_writes(conf);
2796 spin_lock_irqsave(&conf->device_lock, flags);
2797 if (list_empty(head)) {
2798 spin_unlock_irqrestore(&conf->device_lock, flags);
2801 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2802 list_del(head->prev);
2804 spin_unlock_irqrestore(&conf->device_lock, flags);
2806 mddev = r10_bio->mddev;
2807 conf = mddev->private;
2808 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2809 test_bit(R10BIO_WriteError, &r10_bio->state))
2810 handle_write_completed(conf, r10_bio);
2811 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2812 reshape_request_write(mddev, r10_bio);
2813 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2814 sync_request_write(mddev, r10_bio);
2815 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2816 recovery_request_write(mddev, r10_bio);
2817 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2818 handle_read_error(mddev, r10_bio);
2823 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2824 md_check_recovery(mddev);
2826 blk_finish_plug(&plug);
2829 static int init_resync(struct r10conf *conf)
2833 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2834 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2835 conf->have_replacement = 0;
2836 for (i = 0; i < conf->geo.raid_disks; i++)
2837 if (conf->mirrors[i].replacement)
2838 conf->have_replacement = 1;
2839 ret = mempool_init(&conf->r10buf_pool, buffs,
2840 r10buf_pool_alloc, r10buf_pool_free, conf);
2843 conf->next_resync = 0;
2847 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2849 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2850 struct rsync_pages *rp;
2855 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2856 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2857 nalloc = conf->copies; /* resync */
2859 nalloc = 2; /* recovery */
2861 for (i = 0; i < nalloc; i++) {
2862 bio = r10bio->devs[i].bio;
2863 rp = bio->bi_private;
2865 bio->bi_private = rp;
2866 bio = r10bio->devs[i].repl_bio;
2868 rp = bio->bi_private;
2870 bio->bi_private = rp;
2877 * Set cluster_sync_high since we need other nodes to add the
2878 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2880 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2882 sector_t window_size;
2883 int extra_chunk, chunks;
2886 * First, here we define "stripe" as a unit which across
2887 * all member devices one time, so we get chunks by use
2888 * raid_disks / near_copies. Otherwise, if near_copies is
2889 * close to raid_disks, then resync window could increases
2890 * linearly with the increase of raid_disks, which means
2891 * we will suspend a really large IO window while it is not
2892 * necessary. If raid_disks is not divisible by near_copies,
2893 * an extra chunk is needed to ensure the whole "stripe" is
2897 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2898 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2902 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2905 * At least use a 32M window to align with raid1's resync window
2907 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2908 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2910 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2914 * perform a "sync" on one "block"
2916 * We need to make sure that no normal I/O request - particularly write
2917 * requests - conflict with active sync requests.
2919 * This is achieved by tracking pending requests and a 'barrier' concept
2920 * that can be installed to exclude normal IO requests.
2922 * Resync and recovery are handled very differently.
2923 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2925 * For resync, we iterate over virtual addresses, read all copies,
2926 * and update if there are differences. If only one copy is live,
2928 * For recovery, we iterate over physical addresses, read a good
2929 * value for each non-in_sync drive, and over-write.
2931 * So, for recovery we may have several outstanding complex requests for a
2932 * given address, one for each out-of-sync device. We model this by allocating
2933 * a number of r10_bio structures, one for each out-of-sync device.
2934 * As we setup these structures, we collect all bio's together into a list
2935 * which we then process collectively to add pages, and then process again
2936 * to pass to generic_make_request.
2938 * The r10_bio structures are linked using a borrowed master_bio pointer.
2939 * This link is counted in ->remaining. When the r10_bio that points to NULL
2940 * has its remaining count decremented to 0, the whole complex operation
2945 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2948 struct r10conf *conf = mddev->private;
2949 struct r10bio *r10_bio;
2950 struct bio *biolist = NULL, *bio;
2951 sector_t max_sector, nr_sectors;
2954 sector_t sync_blocks;
2955 sector_t sectors_skipped = 0;
2956 int chunks_skipped = 0;
2957 sector_t chunk_mask = conf->geo.chunk_mask;
2961 * Allow skipping a full rebuild for incremental assembly
2962 * of a clean array, like RAID1 does.
2964 if (mddev->bitmap == NULL &&
2965 mddev->recovery_cp == MaxSector &&
2966 mddev->reshape_position == MaxSector &&
2967 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2968 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2969 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2970 conf->fullsync == 0) {
2972 return mddev->dev_sectors - sector_nr;
2975 if (!mempool_initialized(&conf->r10buf_pool))
2976 if (init_resync(conf))
2980 max_sector = mddev->dev_sectors;
2981 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2982 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2983 max_sector = mddev->resync_max_sectors;
2984 if (sector_nr >= max_sector) {
2985 conf->cluster_sync_low = 0;
2986 conf->cluster_sync_high = 0;
2988 /* If we aborted, we need to abort the
2989 * sync on the 'current' bitmap chucks (there can
2990 * be several when recovering multiple devices).
2991 * as we may have started syncing it but not finished.
2992 * We can find the current address in
2993 * mddev->curr_resync, but for recovery,
2994 * we need to convert that to several
2995 * virtual addresses.
2997 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3003 if (mddev->curr_resync < max_sector) { /* aborted */
3004 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3005 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3007 else for (i = 0; i < conf->geo.raid_disks; i++) {
3009 raid10_find_virt(conf, mddev->curr_resync, i);
3010 md_bitmap_end_sync(mddev->bitmap, sect,
3014 /* completed sync */
3015 if ((!mddev->bitmap || conf->fullsync)
3016 && conf->have_replacement
3017 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3018 /* Completed a full sync so the replacements
3019 * are now fully recovered.
3022 for (i = 0; i < conf->geo.raid_disks; i++) {
3023 struct md_rdev *rdev =
3024 rcu_dereference(conf->mirrors[i].replacement);
3026 rdev->recovery_offset = MaxSector;
3032 md_bitmap_close_sync(mddev->bitmap);
3035 return sectors_skipped;
3038 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3039 return reshape_request(mddev, sector_nr, skipped);
3041 if (chunks_skipped >= conf->geo.raid_disks) {
3042 /* if there has been nothing to do on any drive,
3043 * then there is nothing to do at all..
3046 return (max_sector - sector_nr) + sectors_skipped;
3049 if (max_sector > mddev->resync_max)
3050 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3052 /* make sure whole request will fit in a chunk - if chunks
3055 if (conf->geo.near_copies < conf->geo.raid_disks &&
3056 max_sector > (sector_nr | chunk_mask))
3057 max_sector = (sector_nr | chunk_mask) + 1;
3060 * If there is non-resync activity waiting for a turn, then let it
3061 * though before starting on this new sync request.
3063 if (conf->nr_waiting)
3064 schedule_timeout_uninterruptible(1);
3066 /* Again, very different code for resync and recovery.
3067 * Both must result in an r10bio with a list of bios that
3068 * have bi_end_io, bi_sector, bi_disk set,
3069 * and bi_private set to the r10bio.
3070 * For recovery, we may actually create several r10bios
3071 * with 2 bios in each, that correspond to the bios in the main one.
3072 * In this case, the subordinate r10bios link back through a
3073 * borrowed master_bio pointer, and the counter in the master
3074 * includes a ref from each subordinate.
3076 /* First, we decide what to do and set ->bi_end_io
3077 * To end_sync_read if we want to read, and
3078 * end_sync_write if we will want to write.
3081 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3082 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3083 /* recovery... the complicated one */
3087 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3093 struct raid10_info *mirror = &conf->mirrors[i];
3094 struct md_rdev *mrdev, *mreplace;
3097 mrdev = rcu_dereference(mirror->rdev);
3098 mreplace = rcu_dereference(mirror->replacement);
3100 if ((mrdev == NULL ||
3101 test_bit(Faulty, &mrdev->flags) ||
3102 test_bit(In_sync, &mrdev->flags)) &&
3103 (mreplace == NULL ||
3104 test_bit(Faulty, &mreplace->flags))) {
3110 /* want to reconstruct this device */
3112 sect = raid10_find_virt(conf, sector_nr, i);
3113 if (sect >= mddev->resync_max_sectors) {
3114 /* last stripe is not complete - don't
3115 * try to recover this sector.
3120 if (mreplace && test_bit(Faulty, &mreplace->flags))
3122 /* Unless we are doing a full sync, or a replacement
3123 * we only need to recover the block if it is set in
3126 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3128 if (sync_blocks < max_sync)
3129 max_sync = sync_blocks;
3133 /* yep, skip the sync_blocks here, but don't assume
3134 * that there will never be anything to do here
3136 chunks_skipped = -1;
3140 atomic_inc(&mrdev->nr_pending);
3142 atomic_inc(&mreplace->nr_pending);
3145 r10_bio = raid10_alloc_init_r10buf(conf);
3147 raise_barrier(conf, rb2 != NULL);
3148 atomic_set(&r10_bio->remaining, 0);
3150 r10_bio->master_bio = (struct bio*)rb2;
3152 atomic_inc(&rb2->remaining);
3153 r10_bio->mddev = mddev;
3154 set_bit(R10BIO_IsRecover, &r10_bio->state);
3155 r10_bio->sector = sect;
3157 raid10_find_phys(conf, r10_bio);
3159 /* Need to check if the array will still be
3163 for (j = 0; j < conf->geo.raid_disks; j++) {
3164 struct md_rdev *rdev = rcu_dereference(
3165 conf->mirrors[j].rdev);
3166 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3172 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3173 &sync_blocks, still_degraded);
3176 for (j=0; j<conf->copies;j++) {
3178 int d = r10_bio->devs[j].devnum;
3179 sector_t from_addr, to_addr;
3180 struct md_rdev *rdev =
3181 rcu_dereference(conf->mirrors[d].rdev);
3182 sector_t sector, first_bad;
3185 !test_bit(In_sync, &rdev->flags))
3187 /* This is where we read from */
3189 sector = r10_bio->devs[j].addr;
3191 if (is_badblock(rdev, sector, max_sync,
3192 &first_bad, &bad_sectors)) {
3193 if (first_bad > sector)
3194 max_sync = first_bad - sector;
3196 bad_sectors -= (sector
3198 if (max_sync > bad_sectors)
3199 max_sync = bad_sectors;
3203 bio = r10_bio->devs[0].bio;
3204 bio->bi_next = biolist;
3206 bio->bi_end_io = end_sync_read;
3207 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3208 if (test_bit(FailFast, &rdev->flags))
3209 bio->bi_opf |= MD_FAILFAST;
3210 from_addr = r10_bio->devs[j].addr;
3211 bio->bi_iter.bi_sector = from_addr +
3213 bio_set_dev(bio, rdev->bdev);
3214 atomic_inc(&rdev->nr_pending);
3215 /* and we write to 'i' (if not in_sync) */
3217 for (k=0; k<conf->copies; k++)
3218 if (r10_bio->devs[k].devnum == i)
3220 BUG_ON(k == conf->copies);
3221 to_addr = r10_bio->devs[k].addr;
3222 r10_bio->devs[0].devnum = d;
3223 r10_bio->devs[0].addr = from_addr;
3224 r10_bio->devs[1].devnum = i;
3225 r10_bio->devs[1].addr = to_addr;
3227 if (!test_bit(In_sync, &mrdev->flags)) {
3228 bio = r10_bio->devs[1].bio;
3229 bio->bi_next = biolist;
3231 bio->bi_end_io = end_sync_write;
3232 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3233 bio->bi_iter.bi_sector = to_addr
3234 + mrdev->data_offset;
3235 bio_set_dev(bio, mrdev->bdev);
3236 atomic_inc(&r10_bio->remaining);
3238 r10_bio->devs[1].bio->bi_end_io = NULL;
3240 /* and maybe write to replacement */
3241 bio = r10_bio->devs[1].repl_bio;
3243 bio->bi_end_io = NULL;
3244 /* Note: if mreplace != NULL, then bio
3245 * cannot be NULL as r10buf_pool_alloc will
3246 * have allocated it.
3247 * So the second test here is pointless.
3248 * But it keeps semantic-checkers happy, and
3249 * this comment keeps human reviewers
3252 if (mreplace == NULL || bio == NULL ||
3253 test_bit(Faulty, &mreplace->flags))
3255 bio->bi_next = biolist;
3257 bio->bi_end_io = end_sync_write;
3258 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3259 bio->bi_iter.bi_sector = to_addr +
3260 mreplace->data_offset;
3261 bio_set_dev(bio, mreplace->bdev);
3262 atomic_inc(&r10_bio->remaining);
3266 if (j == conf->copies) {
3267 /* Cannot recover, so abort the recovery or
3268 * record a bad block */
3270 /* problem is that there are bad blocks
3271 * on other device(s)
3274 for (k = 0; k < conf->copies; k++)
3275 if (r10_bio->devs[k].devnum == i)
3277 if (!test_bit(In_sync,
3279 && !rdev_set_badblocks(
3281 r10_bio->devs[k].addr,
3285 !rdev_set_badblocks(
3287 r10_bio->devs[k].addr,
3292 if (!test_and_set_bit(MD_RECOVERY_INTR,
3294 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3296 mirror->recovery_disabled
3297 = mddev->recovery_disabled;
3301 atomic_dec(&rb2->remaining);
3303 rdev_dec_pending(mrdev, mddev);
3305 rdev_dec_pending(mreplace, mddev);
3308 rdev_dec_pending(mrdev, mddev);
3310 rdev_dec_pending(mreplace, mddev);
3311 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3312 /* Only want this if there is elsewhere to
3313 * read from. 'j' is currently the first
3317 for (; j < conf->copies; j++) {
3318 int d = r10_bio->devs[j].devnum;
3319 if (conf->mirrors[d].rdev &&
3321 &conf->mirrors[d].rdev->flags))
3325 r10_bio->devs[0].bio->bi_opf
3329 if (biolist == NULL) {
3331 struct r10bio *rb2 = r10_bio;
3332 r10_bio = (struct r10bio*) rb2->master_bio;
3333 rb2->master_bio = NULL;
3339 /* resync. Schedule a read for every block at this virt offset */
3343 * Since curr_resync_completed could probably not update in
3344 * time, and we will set cluster_sync_low based on it.
3345 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3346 * safety reason, which ensures curr_resync_completed is
3347 * updated in bitmap_cond_end_sync.
3349 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3350 mddev_is_clustered(mddev) &&
3351 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3353 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3354 &sync_blocks, mddev->degraded) &&
3355 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3356 &mddev->recovery)) {
3357 /* We can skip this block */
3359 return sync_blocks + sectors_skipped;
3361 if (sync_blocks < max_sync)
3362 max_sync = sync_blocks;
3363 r10_bio = raid10_alloc_init_r10buf(conf);
3366 r10_bio->mddev = mddev;
3367 atomic_set(&r10_bio->remaining, 0);
3368 raise_barrier(conf, 0);
3369 conf->next_resync = sector_nr;
3371 r10_bio->master_bio = NULL;
3372 r10_bio->sector = sector_nr;
3373 set_bit(R10BIO_IsSync, &r10_bio->state);
3374 raid10_find_phys(conf, r10_bio);
3375 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3377 for (i = 0; i < conf->copies; i++) {
3378 int d = r10_bio->devs[i].devnum;
3379 sector_t first_bad, sector;
3381 struct md_rdev *rdev;
3383 if (r10_bio->devs[i].repl_bio)
3384 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3386 bio = r10_bio->devs[i].bio;
3387 bio->bi_status = BLK_STS_IOERR;
3389 rdev = rcu_dereference(conf->mirrors[d].rdev);
3390 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3394 sector = r10_bio->devs[i].addr;
3395 if (is_badblock(rdev, sector, max_sync,
3396 &first_bad, &bad_sectors)) {
3397 if (first_bad > sector)
3398 max_sync = first_bad - sector;
3400 bad_sectors -= (sector - first_bad);
3401 if (max_sync > bad_sectors)
3402 max_sync = bad_sectors;
3407 atomic_inc(&rdev->nr_pending);
3408 atomic_inc(&r10_bio->remaining);
3409 bio->bi_next = biolist;
3411 bio->bi_end_io = end_sync_read;
3412 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3413 if (test_bit(FailFast, &rdev->flags))
3414 bio->bi_opf |= MD_FAILFAST;
3415 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3416 bio_set_dev(bio, rdev->bdev);
3419 rdev = rcu_dereference(conf->mirrors[d].replacement);
3420 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3424 atomic_inc(&rdev->nr_pending);
3426 /* Need to set up for writing to the replacement */
3427 bio = r10_bio->devs[i].repl_bio;
3428 bio->bi_status = BLK_STS_IOERR;
3430 sector = r10_bio->devs[i].addr;
3431 bio->bi_next = biolist;
3433 bio->bi_end_io = end_sync_write;
3434 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3435 if (test_bit(FailFast, &rdev->flags))
3436 bio->bi_opf |= MD_FAILFAST;
3437 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3438 bio_set_dev(bio, rdev->bdev);
3444 for (i=0; i<conf->copies; i++) {
3445 int d = r10_bio->devs[i].devnum;
3446 if (r10_bio->devs[i].bio->bi_end_io)
3447 rdev_dec_pending(conf->mirrors[d].rdev,
3449 if (r10_bio->devs[i].repl_bio &&
3450 r10_bio->devs[i].repl_bio->bi_end_io)
3452 conf->mirrors[d].replacement,
3462 if (sector_nr + max_sync < max_sector)
3463 max_sector = sector_nr + max_sync;
3466 int len = PAGE_SIZE;
3467 if (sector_nr + (len>>9) > max_sector)
3468 len = (max_sector - sector_nr) << 9;
3471 for (bio= biolist ; bio ; bio=bio->bi_next) {
3472 struct resync_pages *rp = get_resync_pages(bio);
3473 page = resync_fetch_page(rp, page_idx);
3475 * won't fail because the vec table is big enough
3476 * to hold all these pages
3478 bio_add_page(bio, page, len, 0);
3480 nr_sectors += len>>9;
3481 sector_nr += len>>9;
3482 } while (++page_idx < RESYNC_PAGES);
3483 r10_bio->sectors = nr_sectors;
3485 if (mddev_is_clustered(mddev) &&
3486 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3487 /* It is resync not recovery */
3488 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3489 conf->cluster_sync_low = mddev->curr_resync_completed;
3490 raid10_set_cluster_sync_high(conf);
3491 /* Send resync message */
3492 md_cluster_ops->resync_info_update(mddev,
3493 conf->cluster_sync_low,
3494 conf->cluster_sync_high);
3496 } else if (mddev_is_clustered(mddev)) {
3497 /* This is recovery not resync */
3498 sector_t sect_va1, sect_va2;
3499 bool broadcast_msg = false;
3501 for (i = 0; i < conf->geo.raid_disks; i++) {
3503 * sector_nr is a device address for recovery, so we
3504 * need translate it to array address before compare
3505 * with cluster_sync_high.
3507 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3509 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3510 broadcast_msg = true;
3512 * curr_resync_completed is similar as
3513 * sector_nr, so make the translation too.
3515 sect_va2 = raid10_find_virt(conf,
3516 mddev->curr_resync_completed, i);
3518 if (conf->cluster_sync_low == 0 ||
3519 conf->cluster_sync_low > sect_va2)
3520 conf->cluster_sync_low = sect_va2;
3523 if (broadcast_msg) {
3524 raid10_set_cluster_sync_high(conf);
3525 md_cluster_ops->resync_info_update(mddev,
3526 conf->cluster_sync_low,
3527 conf->cluster_sync_high);
3533 biolist = biolist->bi_next;
3535 bio->bi_next = NULL;
3536 r10_bio = get_resync_r10bio(bio);
3537 r10_bio->sectors = nr_sectors;
3539 if (bio->bi_end_io == end_sync_read) {
3540 md_sync_acct_bio(bio, nr_sectors);
3542 generic_make_request(bio);
3546 if (sectors_skipped)
3547 /* pretend they weren't skipped, it makes
3548 * no important difference in this case
3550 md_done_sync(mddev, sectors_skipped, 1);
3552 return sectors_skipped + nr_sectors;
3554 /* There is nowhere to write, so all non-sync
3555 * drives must be failed or in resync, all drives
3556 * have a bad block, so try the next chunk...
3558 if (sector_nr + max_sync < max_sector)
3559 max_sector = sector_nr + max_sync;
3561 sectors_skipped += (max_sector - sector_nr);
3563 sector_nr = max_sector;
3568 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3571 struct r10conf *conf = mddev->private;
3574 raid_disks = min(conf->geo.raid_disks,
3575 conf->prev.raid_disks);
3577 sectors = conf->dev_sectors;
3579 size = sectors >> conf->geo.chunk_shift;
3580 sector_div(size, conf->geo.far_copies);
3581 size = size * raid_disks;
3582 sector_div(size, conf->geo.near_copies);
3584 return size << conf->geo.chunk_shift;
3587 static void calc_sectors(struct r10conf *conf, sector_t size)
3589 /* Calculate the number of sectors-per-device that will
3590 * actually be used, and set conf->dev_sectors and
3594 size = size >> conf->geo.chunk_shift;
3595 sector_div(size, conf->geo.far_copies);
3596 size = size * conf->geo.raid_disks;
3597 sector_div(size, conf->geo.near_copies);
3598 /* 'size' is now the number of chunks in the array */
3599 /* calculate "used chunks per device" */
3600 size = size * conf->copies;
3602 /* We need to round up when dividing by raid_disks to
3603 * get the stride size.
3605 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3607 conf->dev_sectors = size << conf->geo.chunk_shift;
3609 if (conf->geo.far_offset)
3610 conf->geo.stride = 1 << conf->geo.chunk_shift;
3612 sector_div(size, conf->geo.far_copies);
3613 conf->geo.stride = size << conf->geo.chunk_shift;
3617 enum geo_type {geo_new, geo_old, geo_start};
3618 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3621 int layout, chunk, disks;
3624 layout = mddev->layout;
3625 chunk = mddev->chunk_sectors;
3626 disks = mddev->raid_disks - mddev->delta_disks;
3629 layout = mddev->new_layout;
3630 chunk = mddev->new_chunk_sectors;
3631 disks = mddev->raid_disks;
3633 default: /* avoid 'may be unused' warnings */
3634 case geo_start: /* new when starting reshape - raid_disks not
3636 layout = mddev->new_layout;
3637 chunk = mddev->new_chunk_sectors;
3638 disks = mddev->raid_disks + mddev->delta_disks;
3643 if (chunk < (PAGE_SIZE >> 9) ||
3644 !is_power_of_2(chunk))
3647 fc = (layout >> 8) & 255;
3648 fo = layout & (1<<16);
3649 geo->raid_disks = disks;
3650 geo->near_copies = nc;
3651 geo->far_copies = fc;
3652 geo->far_offset = fo;
3653 switch (layout >> 17) {
3654 case 0: /* original layout. simple but not always optimal */
3655 geo->far_set_size = disks;
3657 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3658 * actually using this, but leave code here just in case.*/
3659 geo->far_set_size = disks/fc;
3660 WARN(geo->far_set_size < fc,
3661 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3663 case 2: /* "improved" layout fixed to match documentation */
3664 geo->far_set_size = fc * nc;
3666 default: /* Not a valid layout */
3669 geo->chunk_mask = chunk - 1;
3670 geo->chunk_shift = ffz(~chunk);
3674 static void raid10_free_conf(struct r10conf *conf)
3679 mempool_exit(&conf->r10bio_pool);
3680 kfree(conf->mirrors);
3681 kfree(conf->mirrors_old);
3682 kfree(conf->mirrors_new);
3683 safe_put_page(conf->tmppage);
3684 bioset_exit(&conf->bio_split);
3688 static struct r10conf *setup_conf(struct mddev *mddev)
3690 struct r10conf *conf = NULL;
3695 copies = setup_geo(&geo, mddev, geo_new);
3698 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3699 mdname(mddev), PAGE_SIZE);
3703 if (copies < 2 || copies > mddev->raid_disks) {
3704 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3705 mdname(mddev), mddev->new_layout);
3710 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3714 /* FIXME calc properly */
3715 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3716 sizeof(struct raid10_info),
3721 conf->tmppage = alloc_page(GFP_KERNEL);
3726 conf->copies = copies;
3727 err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3728 r10bio_pool_free, conf);
3732 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3736 calc_sectors(conf, mddev->dev_sectors);
3737 if (mddev->reshape_position == MaxSector) {
3738 conf->prev = conf->geo;
3739 conf->reshape_progress = MaxSector;
3741 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3745 conf->reshape_progress = mddev->reshape_position;
3746 if (conf->prev.far_offset)
3747 conf->prev.stride = 1 << conf->prev.chunk_shift;
3749 /* far_copies must be 1 */
3750 conf->prev.stride = conf->dev_sectors;
3752 conf->reshape_safe = conf->reshape_progress;
3753 spin_lock_init(&conf->device_lock);
3754 INIT_LIST_HEAD(&conf->retry_list);
3755 INIT_LIST_HEAD(&conf->bio_end_io_list);
3757 spin_lock_init(&conf->resync_lock);
3758 init_waitqueue_head(&conf->wait_barrier);
3759 atomic_set(&conf->nr_pending, 0);
3762 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3766 conf->mddev = mddev;
3770 raid10_free_conf(conf);
3771 return ERR_PTR(err);
3774 static void raid10_set_io_opt(struct r10conf *conf)
3776 int raid_disks = conf->geo.raid_disks;
3778 if (!(conf->geo.raid_disks % conf->geo.near_copies))
3779 raid_disks /= conf->geo.near_copies;
3780 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
3784 static int raid10_run(struct mddev *mddev)
3786 struct r10conf *conf;
3788 struct raid10_info *disk;
3789 struct md_rdev *rdev;
3791 sector_t min_offset_diff = 0;
3793 bool discard_supported = false;
3795 if (mddev_init_writes_pending(mddev) < 0)
3798 if (mddev->private == NULL) {
3799 conf = setup_conf(mddev);
3801 return PTR_ERR(conf);
3802 mddev->private = conf;
3804 conf = mddev->private;
3808 mddev->thread = conf->thread;
3809 conf->thread = NULL;
3811 if (mddev_is_clustered(conf->mddev)) {
3814 fc = (mddev->layout >> 8) & 255;
3815 fo = mddev->layout & (1<<16);
3816 if (fc > 1 || fo > 0) {
3817 pr_err("only near layout is supported by clustered"
3824 blk_queue_max_discard_sectors(mddev->queue,
3825 mddev->chunk_sectors);
3826 blk_queue_max_write_same_sectors(mddev->queue, 0);
3827 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3828 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
3829 raid10_set_io_opt(conf);
3832 rdev_for_each(rdev, mddev) {
3835 disk_idx = rdev->raid_disk;
3838 if (disk_idx >= conf->geo.raid_disks &&
3839 disk_idx >= conf->prev.raid_disks)
3841 disk = conf->mirrors + disk_idx;
3843 if (test_bit(Replacement, &rdev->flags)) {
3844 if (disk->replacement)
3846 disk->replacement = rdev;
3852 diff = (rdev->new_data_offset - rdev->data_offset);
3853 if (!mddev->reshape_backwards)
3857 if (first || diff < min_offset_diff)
3858 min_offset_diff = diff;
3861 disk_stack_limits(mddev->gendisk, rdev->bdev,
3862 rdev->data_offset << 9);
3864 disk->head_position = 0;
3866 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3867 discard_supported = true;
3872 if (discard_supported)
3873 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3876 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3879 /* need to check that every block has at least one working mirror */
3880 if (!enough(conf, -1)) {
3881 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3886 if (conf->reshape_progress != MaxSector) {
3887 /* must ensure that shape change is supported */
3888 if (conf->geo.far_copies != 1 &&
3889 conf->geo.far_offset == 0)
3891 if (conf->prev.far_copies != 1 &&
3892 conf->prev.far_offset == 0)
3896 mddev->degraded = 0;
3898 i < conf->geo.raid_disks
3899 || i < conf->prev.raid_disks;
3902 disk = conf->mirrors + i;
3904 if (!disk->rdev && disk->replacement) {
3905 /* The replacement is all we have - use it */
3906 disk->rdev = disk->replacement;
3907 disk->replacement = NULL;
3908 clear_bit(Replacement, &disk->rdev->flags);
3912 !test_bit(In_sync, &disk->rdev->flags)) {
3913 disk->head_position = 0;
3916 disk->rdev->saved_raid_disk < 0)
3920 if (disk->replacement &&
3921 !test_bit(In_sync, &disk->replacement->flags) &&
3922 disk->replacement->saved_raid_disk < 0) {
3926 disk->recovery_disabled = mddev->recovery_disabled - 1;
3929 if (mddev->recovery_cp != MaxSector)
3930 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3932 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3933 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3934 conf->geo.raid_disks);
3936 * Ok, everything is just fine now
3938 mddev->dev_sectors = conf->dev_sectors;
3939 size = raid10_size(mddev, 0, 0);
3940 md_set_array_sectors(mddev, size);
3941 mddev->resync_max_sectors = size;
3942 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3945 int stripe = conf->geo.raid_disks *
3946 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3948 /* Calculate max read-ahead size.
3949 * We need to readahead at least twice a whole stripe....
3952 stripe /= conf->geo.near_copies;
3953 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3954 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3957 if (md_integrity_register(mddev))
3960 if (conf->reshape_progress != MaxSector) {
3961 unsigned long before_length, after_length;
3963 before_length = ((1 << conf->prev.chunk_shift) *
3964 conf->prev.far_copies);
3965 after_length = ((1 << conf->geo.chunk_shift) *
3966 conf->geo.far_copies);
3968 if (max(before_length, after_length) > min_offset_diff) {
3969 /* This cannot work */
3970 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3973 conf->offset_diff = min_offset_diff;
3975 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3976 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3977 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3978 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3979 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3981 if (!mddev->sync_thread)
3988 md_unregister_thread(&mddev->thread);
3989 raid10_free_conf(conf);
3990 mddev->private = NULL;
3995 static void raid10_free(struct mddev *mddev, void *priv)
3997 raid10_free_conf(priv);
4000 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4002 struct r10conf *conf = mddev->private;
4005 raise_barrier(conf, 0);
4007 lower_barrier(conf);
4010 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4012 /* Resize of 'far' arrays is not supported.
4013 * For 'near' and 'offset' arrays we can set the
4014 * number of sectors used to be an appropriate multiple
4015 * of the chunk size.
4016 * For 'offset', this is far_copies*chunksize.
4017 * For 'near' the multiplier is the LCM of
4018 * near_copies and raid_disks.
4019 * So if far_copies > 1 && !far_offset, fail.
4020 * Else find LCM(raid_disks, near_copy)*far_copies and
4021 * multiply by chunk_size. Then round to this number.
4022 * This is mostly done by raid10_size()
4024 struct r10conf *conf = mddev->private;
4025 sector_t oldsize, size;
4027 if (mddev->reshape_position != MaxSector)
4030 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4033 oldsize = raid10_size(mddev, 0, 0);
4034 size = raid10_size(mddev, sectors, 0);
4035 if (mddev->external_size &&
4036 mddev->array_sectors > size)
4038 if (mddev->bitmap) {
4039 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4043 md_set_array_sectors(mddev, size);
4044 if (sectors > mddev->dev_sectors &&
4045 mddev->recovery_cp > oldsize) {
4046 mddev->recovery_cp = oldsize;
4047 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4049 calc_sectors(conf, sectors);
4050 mddev->dev_sectors = conf->dev_sectors;
4051 mddev->resync_max_sectors = size;
4055 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4057 struct md_rdev *rdev;
4058 struct r10conf *conf;
4060 if (mddev->degraded > 0) {
4061 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4063 return ERR_PTR(-EINVAL);
4065 sector_div(size, devs);
4067 /* Set new parameters */
4068 mddev->new_level = 10;
4069 /* new layout: far_copies = 1, near_copies = 2 */
4070 mddev->new_layout = (1<<8) + 2;
4071 mddev->new_chunk_sectors = mddev->chunk_sectors;
4072 mddev->delta_disks = mddev->raid_disks;
4073 mddev->raid_disks *= 2;
4074 /* make sure it will be not marked as dirty */
4075 mddev->recovery_cp = MaxSector;
4076 mddev->dev_sectors = size;
4078 conf = setup_conf(mddev);
4079 if (!IS_ERR(conf)) {
4080 rdev_for_each(rdev, mddev)
4081 if (rdev->raid_disk >= 0) {
4082 rdev->new_raid_disk = rdev->raid_disk * 2;
4083 rdev->sectors = size;
4091 static void *raid10_takeover(struct mddev *mddev)
4093 struct r0conf *raid0_conf;
4095 /* raid10 can take over:
4096 * raid0 - providing it has only two drives
4098 if (mddev->level == 0) {
4099 /* for raid0 takeover only one zone is supported */
4100 raid0_conf = mddev->private;
4101 if (raid0_conf->nr_strip_zones > 1) {
4102 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4104 return ERR_PTR(-EINVAL);
4106 return raid10_takeover_raid0(mddev,
4107 raid0_conf->strip_zone->zone_end,
4108 raid0_conf->strip_zone->nb_dev);
4110 return ERR_PTR(-EINVAL);
4113 static int raid10_check_reshape(struct mddev *mddev)
4115 /* Called when there is a request to change
4116 * - layout (to ->new_layout)
4117 * - chunk size (to ->new_chunk_sectors)
4118 * - raid_disks (by delta_disks)
4119 * or when trying to restart a reshape that was ongoing.
4121 * We need to validate the request and possibly allocate
4122 * space if that might be an issue later.
4124 * Currently we reject any reshape of a 'far' mode array,
4125 * allow chunk size to change if new is generally acceptable,
4126 * allow raid_disks to increase, and allow
4127 * a switch between 'near' mode and 'offset' mode.
4129 struct r10conf *conf = mddev->private;
4132 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4135 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4136 /* mustn't change number of copies */
4138 if (geo.far_copies > 1 && !geo.far_offset)
4139 /* Cannot switch to 'far' mode */
4142 if (mddev->array_sectors & geo.chunk_mask)
4143 /* not factor of array size */
4146 if (!enough(conf, -1))
4149 kfree(conf->mirrors_new);
4150 conf->mirrors_new = NULL;
4151 if (mddev->delta_disks > 0) {
4152 /* allocate new 'mirrors' list */
4154 kcalloc(mddev->raid_disks + mddev->delta_disks,
4155 sizeof(struct raid10_info),
4157 if (!conf->mirrors_new)
4164 * Need to check if array has failed when deciding whether to:
4166 * - remove non-faulty devices
4169 * This determination is simple when no reshape is happening.
4170 * However if there is a reshape, we need to carefully check
4171 * both the before and after sections.
4172 * This is because some failed devices may only affect one
4173 * of the two sections, and some non-in_sync devices may
4174 * be insync in the section most affected by failed devices.
4176 static int calc_degraded(struct r10conf *conf)
4178 int degraded, degraded2;
4183 /* 'prev' section first */
4184 for (i = 0; i < conf->prev.raid_disks; i++) {
4185 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4186 if (!rdev || test_bit(Faulty, &rdev->flags))
4188 else if (!test_bit(In_sync, &rdev->flags))
4189 /* When we can reduce the number of devices in
4190 * an array, this might not contribute to
4191 * 'degraded'. It does now.
4196 if (conf->geo.raid_disks == conf->prev.raid_disks)
4200 for (i = 0; i < conf->geo.raid_disks; i++) {
4201 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4202 if (!rdev || test_bit(Faulty, &rdev->flags))
4204 else if (!test_bit(In_sync, &rdev->flags)) {
4205 /* If reshape is increasing the number of devices,
4206 * this section has already been recovered, so
4207 * it doesn't contribute to degraded.
4210 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4215 if (degraded2 > degraded)
4220 static int raid10_start_reshape(struct mddev *mddev)
4222 /* A 'reshape' has been requested. This commits
4223 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4224 * This also checks if there are enough spares and adds them
4226 * We currently require enough spares to make the final
4227 * array non-degraded. We also require that the difference
4228 * between old and new data_offset - on each device - is
4229 * enough that we never risk over-writing.
4232 unsigned long before_length, after_length;
4233 sector_t min_offset_diff = 0;
4236 struct r10conf *conf = mddev->private;
4237 struct md_rdev *rdev;
4241 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4244 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4247 before_length = ((1 << conf->prev.chunk_shift) *
4248 conf->prev.far_copies);
4249 after_length = ((1 << conf->geo.chunk_shift) *
4250 conf->geo.far_copies);
4252 rdev_for_each(rdev, mddev) {
4253 if (!test_bit(In_sync, &rdev->flags)
4254 && !test_bit(Faulty, &rdev->flags))
4256 if (rdev->raid_disk >= 0) {
4257 long long diff = (rdev->new_data_offset
4258 - rdev->data_offset);
4259 if (!mddev->reshape_backwards)
4263 if (first || diff < min_offset_diff)
4264 min_offset_diff = diff;
4269 if (max(before_length, after_length) > min_offset_diff)
4272 if (spares < mddev->delta_disks)
4275 conf->offset_diff = min_offset_diff;
4276 spin_lock_irq(&conf->device_lock);
4277 if (conf->mirrors_new) {
4278 memcpy(conf->mirrors_new, conf->mirrors,
4279 sizeof(struct raid10_info)*conf->prev.raid_disks);
4281 kfree(conf->mirrors_old);
4282 conf->mirrors_old = conf->mirrors;
4283 conf->mirrors = conf->mirrors_new;
4284 conf->mirrors_new = NULL;
4286 setup_geo(&conf->geo, mddev, geo_start);
4288 if (mddev->reshape_backwards) {
4289 sector_t size = raid10_size(mddev, 0, 0);
4290 if (size < mddev->array_sectors) {
4291 spin_unlock_irq(&conf->device_lock);
4292 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4296 mddev->resync_max_sectors = size;
4297 conf->reshape_progress = size;
4299 conf->reshape_progress = 0;
4300 conf->reshape_safe = conf->reshape_progress;
4301 spin_unlock_irq(&conf->device_lock);
4303 if (mddev->delta_disks && mddev->bitmap) {
4304 ret = md_bitmap_resize(mddev->bitmap,
4305 raid10_size(mddev, 0, conf->geo.raid_disks),
4310 if (mddev->delta_disks > 0) {
4311 rdev_for_each(rdev, mddev)
4312 if (rdev->raid_disk < 0 &&
4313 !test_bit(Faulty, &rdev->flags)) {
4314 if (raid10_add_disk(mddev, rdev) == 0) {
4315 if (rdev->raid_disk >=
4316 conf->prev.raid_disks)
4317 set_bit(In_sync, &rdev->flags);
4319 rdev->recovery_offset = 0;
4321 if (sysfs_link_rdev(mddev, rdev))
4322 /* Failure here is OK */;
4324 } else if (rdev->raid_disk >= conf->prev.raid_disks
4325 && !test_bit(Faulty, &rdev->flags)) {
4326 /* This is a spare that was manually added */
4327 set_bit(In_sync, &rdev->flags);
4330 /* When a reshape changes the number of devices,
4331 * ->degraded is measured against the larger of the
4332 * pre and post numbers.
4334 spin_lock_irq(&conf->device_lock);
4335 mddev->degraded = calc_degraded(conf);
4336 spin_unlock_irq(&conf->device_lock);
4337 mddev->raid_disks = conf->geo.raid_disks;
4338 mddev->reshape_position = conf->reshape_progress;
4339 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4341 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4342 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4343 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4344 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4345 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4347 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4349 if (!mddev->sync_thread) {
4353 conf->reshape_checkpoint = jiffies;
4354 md_wakeup_thread(mddev->sync_thread);
4355 md_new_event(mddev);
4359 mddev->recovery = 0;
4360 spin_lock_irq(&conf->device_lock);
4361 conf->geo = conf->prev;
4362 mddev->raid_disks = conf->geo.raid_disks;
4363 rdev_for_each(rdev, mddev)
4364 rdev->new_data_offset = rdev->data_offset;
4366 conf->reshape_progress = MaxSector;
4367 conf->reshape_safe = MaxSector;
4368 mddev->reshape_position = MaxSector;
4369 spin_unlock_irq(&conf->device_lock);
4373 /* Calculate the last device-address that could contain
4374 * any block from the chunk that includes the array-address 's'
4375 * and report the next address.
4376 * i.e. the address returned will be chunk-aligned and after
4377 * any data that is in the chunk containing 's'.
4379 static sector_t last_dev_address(sector_t s, struct geom *geo)
4381 s = (s | geo->chunk_mask) + 1;
4382 s >>= geo->chunk_shift;
4383 s *= geo->near_copies;
4384 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4385 s *= geo->far_copies;
4386 s <<= geo->chunk_shift;
4390 /* Calculate the first device-address that could contain
4391 * any block from the chunk that includes the array-address 's'.
4392 * This too will be the start of a chunk
4394 static sector_t first_dev_address(sector_t s, struct geom *geo)
4396 s >>= geo->chunk_shift;
4397 s *= geo->near_copies;
4398 sector_div(s, geo->raid_disks);
4399 s *= geo->far_copies;
4400 s <<= geo->chunk_shift;
4404 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4407 /* We simply copy at most one chunk (smallest of old and new)
4408 * at a time, possibly less if that exceeds RESYNC_PAGES,
4409 * or we hit a bad block or something.
4410 * This might mean we pause for normal IO in the middle of
4411 * a chunk, but that is not a problem as mddev->reshape_position
4412 * can record any location.
4414 * If we will want to write to a location that isn't
4415 * yet recorded as 'safe' (i.e. in metadata on disk) then
4416 * we need to flush all reshape requests and update the metadata.
4418 * When reshaping forwards (e.g. to more devices), we interpret
4419 * 'safe' as the earliest block which might not have been copied
4420 * down yet. We divide this by previous stripe size and multiply
4421 * by previous stripe length to get lowest device offset that we
4422 * cannot write to yet.
4423 * We interpret 'sector_nr' as an address that we want to write to.
4424 * From this we use last_device_address() to find where we might
4425 * write to, and first_device_address on the 'safe' position.
4426 * If this 'next' write position is after the 'safe' position,
4427 * we must update the metadata to increase the 'safe' position.
4429 * When reshaping backwards, we round in the opposite direction
4430 * and perform the reverse test: next write position must not be
4431 * less than current safe position.
4433 * In all this the minimum difference in data offsets
4434 * (conf->offset_diff - always positive) allows a bit of slack,
4435 * so next can be after 'safe', but not by more than offset_diff
4437 * We need to prepare all the bios here before we start any IO
4438 * to ensure the size we choose is acceptable to all devices.
4439 * The means one for each copy for write-out and an extra one for
4441 * We store the read-in bio in ->master_bio and the others in
4442 * ->devs[x].bio and ->devs[x].repl_bio.
4444 struct r10conf *conf = mddev->private;
4445 struct r10bio *r10_bio;
4446 sector_t next, safe, last;
4450 struct md_rdev *rdev;
4453 struct bio *bio, *read_bio;
4454 int sectors_done = 0;
4455 struct page **pages;
4457 if (sector_nr == 0) {
4458 /* If restarting in the middle, skip the initial sectors */
4459 if (mddev->reshape_backwards &&
4460 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4461 sector_nr = (raid10_size(mddev, 0, 0)
4462 - conf->reshape_progress);
4463 } else if (!mddev->reshape_backwards &&
4464 conf->reshape_progress > 0)
4465 sector_nr = conf->reshape_progress;
4467 mddev->curr_resync_completed = sector_nr;
4468 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4474 /* We don't use sector_nr to track where we are up to
4475 * as that doesn't work well for ->reshape_backwards.
4476 * So just use ->reshape_progress.
4478 if (mddev->reshape_backwards) {
4479 /* 'next' is the earliest device address that we might
4480 * write to for this chunk in the new layout
4482 next = first_dev_address(conf->reshape_progress - 1,
4485 /* 'safe' is the last device address that we might read from
4486 * in the old layout after a restart
4488 safe = last_dev_address(conf->reshape_safe - 1,
4491 if (next + conf->offset_diff < safe)
4494 last = conf->reshape_progress - 1;
4495 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4496 & conf->prev.chunk_mask);
4497 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4498 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4500 /* 'next' is after the last device address that we
4501 * might write to for this chunk in the new layout
4503 next = last_dev_address(conf->reshape_progress, &conf->geo);
4505 /* 'safe' is the earliest device address that we might
4506 * read from in the old layout after a restart
4508 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4510 /* Need to update metadata if 'next' might be beyond 'safe'
4511 * as that would possibly corrupt data
4513 if (next > safe + conf->offset_diff)
4516 sector_nr = conf->reshape_progress;
4517 last = sector_nr | (conf->geo.chunk_mask
4518 & conf->prev.chunk_mask);
4520 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4521 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4525 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4526 /* Need to update reshape_position in metadata */
4528 mddev->reshape_position = conf->reshape_progress;
4529 if (mddev->reshape_backwards)
4530 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4531 - conf->reshape_progress;
4533 mddev->curr_resync_completed = conf->reshape_progress;
4534 conf->reshape_checkpoint = jiffies;
4535 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4536 md_wakeup_thread(mddev->thread);
4537 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4538 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4539 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4540 allow_barrier(conf);
4541 return sectors_done;
4543 conf->reshape_safe = mddev->reshape_position;
4544 allow_barrier(conf);
4547 raise_barrier(conf, 0);
4549 /* Now schedule reads for blocks from sector_nr to last */
4550 r10_bio = raid10_alloc_init_r10buf(conf);
4552 raise_barrier(conf, 1);
4553 atomic_set(&r10_bio->remaining, 0);
4554 r10_bio->mddev = mddev;
4555 r10_bio->sector = sector_nr;
4556 set_bit(R10BIO_IsReshape, &r10_bio->state);
4557 r10_bio->sectors = last - sector_nr + 1;
4558 rdev = read_balance(conf, r10_bio, &max_sectors);
4559 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4562 /* Cannot read from here, so need to record bad blocks
4563 * on all the target devices.
4566 mempool_free(r10_bio, &conf->r10buf_pool);
4567 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4568 return sectors_done;
4571 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4573 bio_set_dev(read_bio, rdev->bdev);
4574 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4575 + rdev->data_offset);
4576 read_bio->bi_private = r10_bio;
4577 read_bio->bi_end_io = end_reshape_read;
4578 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4579 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4580 read_bio->bi_status = 0;
4581 read_bio->bi_vcnt = 0;
4582 read_bio->bi_iter.bi_size = 0;
4583 r10_bio->master_bio = read_bio;
4584 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4586 /* Now find the locations in the new layout */
4587 __raid10_find_phys(&conf->geo, r10_bio);
4590 read_bio->bi_next = NULL;
4593 for (s = 0; s < conf->copies*2; s++) {
4595 int d = r10_bio->devs[s/2].devnum;
4596 struct md_rdev *rdev2;
4598 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4599 b = r10_bio->devs[s/2].repl_bio;
4601 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4602 b = r10_bio->devs[s/2].bio;
4604 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4607 bio_set_dev(b, rdev2->bdev);
4608 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4609 rdev2->new_data_offset;
4610 b->bi_end_io = end_reshape_write;
4611 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4616 /* Now add as many pages as possible to all of these bios. */
4619 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4620 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4621 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4622 int len = (max_sectors - s) << 9;
4623 if (len > PAGE_SIZE)
4625 for (bio = blist; bio ; bio = bio->bi_next) {
4627 * won't fail because the vec table is big enough
4628 * to hold all these pages
4630 bio_add_page(bio, page, len, 0);
4632 sector_nr += len >> 9;
4633 nr_sectors += len >> 9;
4636 r10_bio->sectors = nr_sectors;
4638 /* Now submit the read */
4639 md_sync_acct_bio(read_bio, r10_bio->sectors);
4640 atomic_inc(&r10_bio->remaining);
4641 read_bio->bi_next = NULL;
4642 generic_make_request(read_bio);
4643 sectors_done += nr_sectors;
4644 if (sector_nr <= last)
4647 lower_barrier(conf);
4649 /* Now that we have done the whole section we can
4650 * update reshape_progress
4652 if (mddev->reshape_backwards)
4653 conf->reshape_progress -= sectors_done;
4655 conf->reshape_progress += sectors_done;
4657 return sectors_done;
4660 static void end_reshape_request(struct r10bio *r10_bio);
4661 static int handle_reshape_read_error(struct mddev *mddev,
4662 struct r10bio *r10_bio);
4663 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4665 /* Reshape read completed. Hopefully we have a block
4667 * If we got a read error then we do sync 1-page reads from
4668 * elsewhere until we find the data - or give up.
4670 struct r10conf *conf = mddev->private;
4673 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4674 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4675 /* Reshape has been aborted */
4676 md_done_sync(mddev, r10_bio->sectors, 0);
4680 /* We definitely have the data in the pages, schedule the
4683 atomic_set(&r10_bio->remaining, 1);
4684 for (s = 0; s < conf->copies*2; s++) {
4686 int d = r10_bio->devs[s/2].devnum;
4687 struct md_rdev *rdev;
4690 rdev = rcu_dereference(conf->mirrors[d].replacement);
4691 b = r10_bio->devs[s/2].repl_bio;
4693 rdev = rcu_dereference(conf->mirrors[d].rdev);
4694 b = r10_bio->devs[s/2].bio;
4696 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4700 atomic_inc(&rdev->nr_pending);
4702 md_sync_acct_bio(b, r10_bio->sectors);
4703 atomic_inc(&r10_bio->remaining);
4705 generic_make_request(b);
4707 end_reshape_request(r10_bio);
4710 static void end_reshape(struct r10conf *conf)
4712 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4715 spin_lock_irq(&conf->device_lock);
4716 conf->prev = conf->geo;
4717 md_finish_reshape(conf->mddev);
4719 conf->reshape_progress = MaxSector;
4720 conf->reshape_safe = MaxSector;
4721 spin_unlock_irq(&conf->device_lock);
4723 /* read-ahead size must cover two whole stripes, which is
4724 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4726 if (conf->mddev->queue) {
4727 int stripe = conf->geo.raid_disks *
4728 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4729 stripe /= conf->geo.near_copies;
4730 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4731 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4732 raid10_set_io_opt(conf);
4737 static int handle_reshape_read_error(struct mddev *mddev,
4738 struct r10bio *r10_bio)
4740 /* Use sync reads to get the blocks from somewhere else */
4741 int sectors = r10_bio->sectors;
4742 struct r10conf *conf = mddev->private;
4743 struct r10bio *r10b;
4746 struct page **pages;
4748 r10b = kmalloc(sizeof(*r10b) +
4749 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4751 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4755 /* reshape IOs share pages from .devs[0].bio */
4756 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4758 r10b->sector = r10_bio->sector;
4759 __raid10_find_phys(&conf->prev, r10b);
4764 int first_slot = slot;
4766 if (s > (PAGE_SIZE >> 9))
4771 int d = r10b->devs[slot].devnum;
4772 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4775 test_bit(Faulty, &rdev->flags) ||
4776 !test_bit(In_sync, &rdev->flags))
4779 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4780 atomic_inc(&rdev->nr_pending);
4782 success = sync_page_io(rdev,
4786 REQ_OP_READ, 0, false);
4787 rdev_dec_pending(rdev, mddev);
4793 if (slot >= conf->copies)
4795 if (slot == first_slot)
4800 /* couldn't read this block, must give up */
4801 set_bit(MD_RECOVERY_INTR,
4813 static void end_reshape_write(struct bio *bio)
4815 struct r10bio *r10_bio = get_resync_r10bio(bio);
4816 struct mddev *mddev = r10_bio->mddev;
4817 struct r10conf *conf = mddev->private;
4821 struct md_rdev *rdev = NULL;
4823 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4825 rdev = conf->mirrors[d].replacement;
4828 rdev = conf->mirrors[d].rdev;
4831 if (bio->bi_status) {
4832 /* FIXME should record badblock */
4833 md_error(mddev, rdev);
4836 rdev_dec_pending(rdev, mddev);
4837 end_reshape_request(r10_bio);
4840 static void end_reshape_request(struct r10bio *r10_bio)
4842 if (!atomic_dec_and_test(&r10_bio->remaining))
4844 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4845 bio_put(r10_bio->master_bio);
4849 static void raid10_finish_reshape(struct mddev *mddev)
4851 struct r10conf *conf = mddev->private;
4853 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4856 if (mddev->delta_disks > 0) {
4857 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4858 mddev->recovery_cp = mddev->resync_max_sectors;
4859 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4861 mddev->resync_max_sectors = mddev->array_sectors;
4865 for (d = conf->geo.raid_disks ;
4866 d < conf->geo.raid_disks - mddev->delta_disks;
4868 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4870 clear_bit(In_sync, &rdev->flags);
4871 rdev = rcu_dereference(conf->mirrors[d].replacement);
4873 clear_bit(In_sync, &rdev->flags);
4877 mddev->layout = mddev->new_layout;
4878 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4879 mddev->reshape_position = MaxSector;
4880 mddev->delta_disks = 0;
4881 mddev->reshape_backwards = 0;
4884 static struct md_personality raid10_personality =
4888 .owner = THIS_MODULE,
4889 .make_request = raid10_make_request,
4891 .free = raid10_free,
4892 .status = raid10_status,
4893 .error_handler = raid10_error,
4894 .hot_add_disk = raid10_add_disk,
4895 .hot_remove_disk= raid10_remove_disk,
4896 .spare_active = raid10_spare_active,
4897 .sync_request = raid10_sync_request,
4898 .quiesce = raid10_quiesce,
4899 .size = raid10_size,
4900 .resize = raid10_resize,
4901 .takeover = raid10_takeover,
4902 .check_reshape = raid10_check_reshape,
4903 .start_reshape = raid10_start_reshape,
4904 .finish_reshape = raid10_finish_reshape,
4905 .congested = raid10_congested,
4908 static int __init raid_init(void)
4910 return register_md_personality(&raid10_personality);
4913 static void raid_exit(void)
4915 unregister_md_personality(&raid10_personality);
4918 module_init(raid_init);
4919 module_exit(raid_exit);
4920 MODULE_LICENSE("GPL");
4921 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4922 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4923 MODULE_ALIAS("md-raid10");
4924 MODULE_ALIAS("md-level-10");
4926 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);