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 >
785 rdev->recovery_offset) {
787 * Read replacement first to prevent reading both rdev
788 * and replacement as NULL during replacement replace
792 rdev = rcu_dereference(conf->mirrors[disk].rdev);
795 test_bit(Faulty, &rdev->flags))
797 if (!test_bit(In_sync, &rdev->flags) &&
798 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
801 dev_sector = r10_bio->devs[slot].addr;
802 if (is_badblock(rdev, dev_sector, sectors,
803 &first_bad, &bad_sectors)) {
804 if (best_dist < MaxSector)
805 /* Already have a better slot */
807 if (first_bad <= dev_sector) {
808 /* Cannot read here. If this is the
809 * 'primary' device, then we must not read
810 * beyond 'bad_sectors' from another device.
812 bad_sectors -= (dev_sector - first_bad);
813 if (!do_balance && sectors > bad_sectors)
814 sectors = bad_sectors;
815 if (best_good_sectors > sectors)
816 best_good_sectors = sectors;
818 sector_t good_sectors =
819 first_bad - dev_sector;
820 if (good_sectors > best_good_sectors) {
821 best_good_sectors = good_sectors;
826 /* Must read from here */
831 best_good_sectors = sectors;
837 /* At least 2 disks to choose from so failfast is OK */
838 set_bit(R10BIO_FailFast, &r10_bio->state);
839 /* This optimisation is debatable, and completely destroys
840 * sequential read speed for 'far copies' arrays. So only
841 * keep it for 'near' arrays, and review those later.
843 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
846 /* for far > 1 always use the lowest address */
847 else if (geo->far_copies > 1)
848 new_distance = r10_bio->devs[slot].addr;
850 new_distance = abs(r10_bio->devs[slot].addr -
851 conf->mirrors[disk].head_position);
852 if (new_distance < best_dist) {
853 best_dist = new_distance;
858 if (slot >= conf->copies) {
864 atomic_inc(&rdev->nr_pending);
865 r10_bio->read_slot = slot;
869 *max_sectors = best_good_sectors;
874 static int raid10_congested(struct mddev *mddev, int bits)
876 struct r10conf *conf = mddev->private;
879 if ((bits & (1 << WB_async_congested)) &&
880 conf->pending_count >= max_queued_requests)
885 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
888 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
889 if (rdev && !test_bit(Faulty, &rdev->flags)) {
890 struct request_queue *q = bdev_get_queue(rdev->bdev);
892 ret |= bdi_congested(q->backing_dev_info, bits);
899 static void flush_pending_writes(struct r10conf *conf)
901 /* Any writes that have been queued but are awaiting
902 * bitmap updates get flushed here.
904 spin_lock_irq(&conf->device_lock);
906 if (conf->pending_bio_list.head) {
907 struct blk_plug plug;
910 bio = bio_list_get(&conf->pending_bio_list);
911 conf->pending_count = 0;
912 spin_unlock_irq(&conf->device_lock);
915 * As this is called in a wait_event() loop (see freeze_array),
916 * current->state might be TASK_UNINTERRUPTIBLE which will
917 * cause a warning when we prepare to wait again. As it is
918 * rare that this path is taken, it is perfectly safe to force
919 * us to go around the wait_event() loop again, so the warning
920 * is a false-positive. Silence the warning by resetting
923 __set_current_state(TASK_RUNNING);
925 blk_start_plug(&plug);
926 /* flush any pending bitmap writes to disk
927 * before proceeding w/ I/O */
928 md_bitmap_unplug(conf->mddev->bitmap);
929 wake_up(&conf->wait_barrier);
931 while (bio) { /* submit pending writes */
932 struct bio *next = bio->bi_next;
933 struct md_rdev *rdev = (void*)bio->bi_disk;
935 bio_set_dev(bio, rdev->bdev);
936 if (test_bit(Faulty, &rdev->flags)) {
938 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
939 !blk_queue_discard(bio->bi_disk->queue)))
943 generic_make_request(bio);
947 blk_finish_plug(&plug);
949 spin_unlock_irq(&conf->device_lock);
953 * Sometimes we need to suspend IO while we do something else,
954 * either some resync/recovery, or reconfigure the array.
955 * To do this we raise a 'barrier'.
956 * The 'barrier' is a counter that can be raised multiple times
957 * to count how many activities are happening which preclude
959 * We can only raise the barrier if there is no pending IO.
960 * i.e. if nr_pending == 0.
961 * We choose only to raise the barrier if no-one is waiting for the
962 * barrier to go down. This means that as soon as an IO request
963 * is ready, no other operations which require a barrier will start
964 * until the IO request has had a chance.
966 * So: regular IO calls 'wait_barrier'. When that returns there
967 * is no backgroup IO happening, It must arrange to call
968 * allow_barrier when it has finished its IO.
969 * backgroup IO calls must call raise_barrier. Once that returns
970 * there is no normal IO happeing. It must arrange to call
971 * lower_barrier when the particular background IO completes.
974 static void raise_barrier(struct r10conf *conf, int force)
976 BUG_ON(force && !conf->barrier);
977 spin_lock_irq(&conf->resync_lock);
979 /* Wait until no block IO is waiting (unless 'force') */
980 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
983 /* block any new IO from starting */
986 /* Now wait for all pending IO to complete */
987 wait_event_lock_irq(conf->wait_barrier,
988 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
991 spin_unlock_irq(&conf->resync_lock);
994 static void lower_barrier(struct r10conf *conf)
997 spin_lock_irqsave(&conf->resync_lock, flags);
999 spin_unlock_irqrestore(&conf->resync_lock, flags);
1000 wake_up(&conf->wait_barrier);
1003 static void wait_barrier(struct r10conf *conf)
1005 spin_lock_irq(&conf->resync_lock);
1006 if (conf->barrier) {
1008 /* Wait for the barrier to drop.
1009 * However if there are already pending
1010 * requests (preventing the barrier from
1011 * rising completely), and the
1012 * pre-process bio queue isn't empty,
1013 * then don't wait, as we need to empty
1014 * that queue to get the nr_pending
1017 raid10_log(conf->mddev, "wait barrier");
1018 wait_event_lock_irq(conf->wait_barrier,
1020 (atomic_read(&conf->nr_pending) &&
1021 current->bio_list &&
1022 (!bio_list_empty(¤t->bio_list[0]) ||
1023 !bio_list_empty(¤t->bio_list[1]))),
1026 if (!conf->nr_waiting)
1027 wake_up(&conf->wait_barrier);
1029 atomic_inc(&conf->nr_pending);
1030 spin_unlock_irq(&conf->resync_lock);
1033 static void allow_barrier(struct r10conf *conf)
1035 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1036 (conf->array_freeze_pending))
1037 wake_up(&conf->wait_barrier);
1040 static void freeze_array(struct r10conf *conf, int extra)
1042 /* stop syncio and normal IO and wait for everything to
1044 * We increment barrier and nr_waiting, and then
1045 * wait until nr_pending match nr_queued+extra
1046 * This is called in the context of one normal IO request
1047 * that has failed. Thus any sync request that might be pending
1048 * will be blocked by nr_pending, and we need to wait for
1049 * pending IO requests to complete or be queued for re-try.
1050 * Thus the number queued (nr_queued) plus this request (extra)
1051 * must match the number of pending IOs (nr_pending) before
1054 spin_lock_irq(&conf->resync_lock);
1055 conf->array_freeze_pending++;
1058 wait_event_lock_irq_cmd(conf->wait_barrier,
1059 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1061 flush_pending_writes(conf));
1063 conf->array_freeze_pending--;
1064 spin_unlock_irq(&conf->resync_lock);
1067 static void unfreeze_array(struct r10conf *conf)
1069 /* reverse the effect of the freeze */
1070 spin_lock_irq(&conf->resync_lock);
1073 wake_up(&conf->wait_barrier);
1074 spin_unlock_irq(&conf->resync_lock);
1077 static sector_t choose_data_offset(struct r10bio *r10_bio,
1078 struct md_rdev *rdev)
1080 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1081 test_bit(R10BIO_Previous, &r10_bio->state))
1082 return rdev->data_offset;
1084 return rdev->new_data_offset;
1087 struct raid10_plug_cb {
1088 struct blk_plug_cb cb;
1089 struct bio_list pending;
1093 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1095 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1097 struct mddev *mddev = plug->cb.data;
1098 struct r10conf *conf = mddev->private;
1101 if (from_schedule || current->bio_list) {
1102 spin_lock_irq(&conf->device_lock);
1103 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1104 conf->pending_count += plug->pending_cnt;
1105 spin_unlock_irq(&conf->device_lock);
1106 wake_up(&conf->wait_barrier);
1107 md_wakeup_thread(mddev->thread);
1112 /* we aren't scheduling, so we can do the write-out directly. */
1113 bio = bio_list_get(&plug->pending);
1114 md_bitmap_unplug(mddev->bitmap);
1115 wake_up(&conf->wait_barrier);
1117 while (bio) { /* submit pending writes */
1118 struct bio *next = bio->bi_next;
1119 struct md_rdev *rdev = (void*)bio->bi_disk;
1120 bio->bi_next = NULL;
1121 bio_set_dev(bio, rdev->bdev);
1122 if (test_bit(Faulty, &rdev->flags)) {
1124 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1125 !blk_queue_discard(bio->bi_disk->queue)))
1126 /* Just ignore it */
1129 generic_make_request(bio);
1136 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1137 struct r10bio *r10_bio)
1139 struct r10conf *conf = mddev->private;
1140 struct bio *read_bio;
1141 const int op = bio_op(bio);
1142 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1145 struct md_rdev *rdev;
1146 char b[BDEVNAME_SIZE];
1147 int slot = r10_bio->read_slot;
1148 struct md_rdev *err_rdev = NULL;
1149 gfp_t gfp = GFP_NOIO;
1151 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1153 * This is an error retry, but we cannot
1154 * safely dereference the rdev in the r10_bio,
1155 * we must use the one in conf.
1156 * If it has already been disconnected (unlikely)
1157 * we lose the device name in error messages.
1161 * As we are blocking raid10, it is a little safer to
1164 gfp = GFP_NOIO | __GFP_HIGH;
1167 disk = r10_bio->devs[slot].devnum;
1168 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1170 bdevname(err_rdev->bdev, b);
1173 /* This never gets dereferenced */
1174 err_rdev = r10_bio->devs[slot].rdev;
1179 * Register the new request and wait if the reconstruction
1180 * thread has put up a bar for new requests.
1181 * Continue immediately if no resync is active currently.
1185 sectors = r10_bio->sectors;
1186 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1187 bio->bi_iter.bi_sector < conf->reshape_progress &&
1188 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1190 * IO spans the reshape position. Need to wait for reshape to
1193 raid10_log(conf->mddev, "wait reshape");
1194 allow_barrier(conf);
1195 wait_event(conf->wait_barrier,
1196 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1197 conf->reshape_progress >= bio->bi_iter.bi_sector +
1202 rdev = read_balance(conf, r10_bio, &max_sectors);
1205 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1207 (unsigned long long)r10_bio->sector);
1209 raid_end_bio_io(r10_bio);
1213 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1215 bdevname(rdev->bdev, b),
1216 (unsigned long long)r10_bio->sector);
1217 if (max_sectors < bio_sectors(bio)) {
1218 struct bio *split = bio_split(bio, max_sectors,
1219 gfp, &conf->bio_split);
1220 bio_chain(split, bio);
1221 allow_barrier(conf);
1222 generic_make_request(bio);
1225 r10_bio->master_bio = bio;
1226 r10_bio->sectors = max_sectors;
1228 slot = r10_bio->read_slot;
1230 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1232 r10_bio->devs[slot].bio = read_bio;
1233 r10_bio->devs[slot].rdev = rdev;
1235 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1236 choose_data_offset(r10_bio, rdev);
1237 bio_set_dev(read_bio, rdev->bdev);
1238 read_bio->bi_end_io = raid10_end_read_request;
1239 bio_set_op_attrs(read_bio, op, do_sync);
1240 if (test_bit(FailFast, &rdev->flags) &&
1241 test_bit(R10BIO_FailFast, &r10_bio->state))
1242 read_bio->bi_opf |= MD_FAILFAST;
1243 read_bio->bi_private = r10_bio;
1246 trace_block_bio_remap(read_bio->bi_disk->queue,
1247 read_bio, disk_devt(mddev->gendisk),
1249 generic_make_request(read_bio);
1253 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1254 struct bio *bio, bool replacement,
1257 const int op = bio_op(bio);
1258 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1259 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1260 unsigned long flags;
1261 struct blk_plug_cb *cb;
1262 struct raid10_plug_cb *plug = NULL;
1263 struct r10conf *conf = mddev->private;
1264 struct md_rdev *rdev;
1265 int devnum = r10_bio->devs[n_copy].devnum;
1269 rdev = conf->mirrors[devnum].replacement;
1271 /* Replacement just got moved to main 'rdev' */
1273 rdev = conf->mirrors[devnum].rdev;
1276 rdev = conf->mirrors[devnum].rdev;
1278 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1280 r10_bio->devs[n_copy].repl_bio = mbio;
1282 r10_bio->devs[n_copy].bio = mbio;
1284 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1285 choose_data_offset(r10_bio, rdev));
1286 bio_set_dev(mbio, rdev->bdev);
1287 mbio->bi_end_io = raid10_end_write_request;
1288 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1289 if (!replacement && test_bit(FailFast,
1290 &conf->mirrors[devnum].rdev->flags)
1291 && enough(conf, devnum))
1292 mbio->bi_opf |= MD_FAILFAST;
1293 mbio->bi_private = r10_bio;
1295 if (conf->mddev->gendisk)
1296 trace_block_bio_remap(mbio->bi_disk->queue,
1297 mbio, disk_devt(conf->mddev->gendisk),
1299 /* flush_pending_writes() needs access to the rdev so...*/
1300 mbio->bi_disk = (void *)rdev;
1302 atomic_inc(&r10_bio->remaining);
1304 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1306 plug = container_of(cb, struct raid10_plug_cb, cb);
1310 bio_list_add(&plug->pending, mbio);
1311 plug->pending_cnt++;
1313 spin_lock_irqsave(&conf->device_lock, flags);
1314 bio_list_add(&conf->pending_bio_list, mbio);
1315 conf->pending_count++;
1316 spin_unlock_irqrestore(&conf->device_lock, flags);
1317 md_wakeup_thread(mddev->thread);
1321 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1322 struct r10bio *r10_bio)
1324 struct r10conf *conf = mddev->private;
1326 struct md_rdev *blocked_rdev;
1330 if ((mddev_is_clustered(mddev) &&
1331 md_cluster_ops->area_resyncing(mddev, WRITE,
1332 bio->bi_iter.bi_sector,
1333 bio_end_sector(bio)))) {
1336 prepare_to_wait(&conf->wait_barrier,
1338 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1339 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1343 finish_wait(&conf->wait_barrier, &w);
1347 * Register the new request and wait if the reconstruction
1348 * thread has put up a bar for new requests.
1349 * Continue immediately if no resync is active currently.
1353 sectors = r10_bio->sectors;
1354 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1355 bio->bi_iter.bi_sector < conf->reshape_progress &&
1356 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1358 * IO spans the reshape position. Need to wait for reshape to
1361 raid10_log(conf->mddev, "wait reshape");
1362 allow_barrier(conf);
1363 wait_event(conf->wait_barrier,
1364 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1365 conf->reshape_progress >= bio->bi_iter.bi_sector +
1370 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1371 (mddev->reshape_backwards
1372 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1373 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1374 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1375 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1376 /* Need to update reshape_position in metadata */
1377 mddev->reshape_position = conf->reshape_progress;
1378 set_mask_bits(&mddev->sb_flags, 0,
1379 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1380 md_wakeup_thread(mddev->thread);
1381 raid10_log(conf->mddev, "wait reshape metadata");
1382 wait_event(mddev->sb_wait,
1383 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1385 conf->reshape_safe = mddev->reshape_position;
1388 if (conf->pending_count >= max_queued_requests) {
1389 md_wakeup_thread(mddev->thread);
1390 raid10_log(mddev, "wait queued");
1391 wait_event(conf->wait_barrier,
1392 conf->pending_count < max_queued_requests);
1394 /* first select target devices under rcu_lock and
1395 * inc refcount on their rdev. Record them by setting
1397 * If there are known/acknowledged bad blocks on any device
1398 * on which we have seen a write error, we want to avoid
1399 * writing to those blocks. This potentially requires several
1400 * writes to write around the bad blocks. Each set of writes
1401 * gets its own r10_bio with a set of bios attached.
1404 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1405 raid10_find_phys(conf, r10_bio);
1407 blocked_rdev = NULL;
1409 max_sectors = r10_bio->sectors;
1411 for (i = 0; i < conf->copies; i++) {
1412 int d = r10_bio->devs[i].devnum;
1413 struct md_rdev *rdev, *rrdev;
1415 rrdev = rcu_dereference(conf->mirrors[d].replacement);
1417 * Read replacement first to prevent reading both rdev and
1418 * replacement as NULL during replacement replace rdev.
1421 rdev = rcu_dereference(conf->mirrors[d].rdev);
1424 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1425 atomic_inc(&rdev->nr_pending);
1426 blocked_rdev = rdev;
1429 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1430 atomic_inc(&rrdev->nr_pending);
1431 blocked_rdev = rrdev;
1434 if (rdev && (test_bit(Faulty, &rdev->flags)))
1436 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1439 r10_bio->devs[i].bio = NULL;
1440 r10_bio->devs[i].repl_bio = NULL;
1442 if (!rdev && !rrdev) {
1443 set_bit(R10BIO_Degraded, &r10_bio->state);
1446 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1448 sector_t dev_sector = r10_bio->devs[i].addr;
1452 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1453 &first_bad, &bad_sectors);
1455 /* Mustn't write here until the bad block
1458 atomic_inc(&rdev->nr_pending);
1459 set_bit(BlockedBadBlocks, &rdev->flags);
1460 blocked_rdev = rdev;
1463 if (is_bad && first_bad <= dev_sector) {
1464 /* Cannot write here at all */
1465 bad_sectors -= (dev_sector - first_bad);
1466 if (bad_sectors < max_sectors)
1467 /* Mustn't write more than bad_sectors
1468 * to other devices yet
1470 max_sectors = bad_sectors;
1471 /* We don't set R10BIO_Degraded as that
1472 * only applies if the disk is missing,
1473 * so it might be re-added, and we want to
1474 * know to recover this chunk.
1475 * In this case the device is here, and the
1476 * fact that this chunk is not in-sync is
1477 * recorded in the bad block log.
1482 int good_sectors = first_bad - dev_sector;
1483 if (good_sectors < max_sectors)
1484 max_sectors = good_sectors;
1488 r10_bio->devs[i].bio = bio;
1489 atomic_inc(&rdev->nr_pending);
1492 r10_bio->devs[i].repl_bio = bio;
1493 atomic_inc(&rrdev->nr_pending);
1498 if (unlikely(blocked_rdev)) {
1499 /* Have to wait for this device to get unblocked, then retry */
1503 for (j = 0; j < i; j++) {
1504 if (r10_bio->devs[j].bio) {
1505 d = r10_bio->devs[j].devnum;
1506 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1508 if (r10_bio->devs[j].repl_bio) {
1509 struct md_rdev *rdev;
1510 d = r10_bio->devs[j].devnum;
1511 rdev = conf->mirrors[d].replacement;
1513 /* Race with remove_disk */
1515 rdev = conf->mirrors[d].rdev;
1517 rdev_dec_pending(rdev, mddev);
1520 allow_barrier(conf);
1521 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1522 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1527 if (max_sectors < r10_bio->sectors)
1528 r10_bio->sectors = max_sectors;
1530 if (r10_bio->sectors < bio_sectors(bio)) {
1531 struct bio *split = bio_split(bio, r10_bio->sectors,
1532 GFP_NOIO, &conf->bio_split);
1533 bio_chain(split, bio);
1534 allow_barrier(conf);
1535 generic_make_request(bio);
1538 r10_bio->master_bio = bio;
1541 atomic_set(&r10_bio->remaining, 1);
1542 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1544 for (i = 0; i < conf->copies; i++) {
1545 if (r10_bio->devs[i].bio)
1546 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1547 if (r10_bio->devs[i].repl_bio)
1548 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1550 one_write_done(r10_bio);
1553 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1555 struct r10conf *conf = mddev->private;
1556 struct r10bio *r10_bio;
1558 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1560 r10_bio->master_bio = bio;
1561 r10_bio->sectors = sectors;
1563 r10_bio->mddev = mddev;
1564 r10_bio->sector = bio->bi_iter.bi_sector;
1566 r10_bio->read_slot = -1;
1567 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1569 if (bio_data_dir(bio) == READ)
1570 raid10_read_request(mddev, bio, r10_bio);
1572 raid10_write_request(mddev, bio, r10_bio);
1575 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1577 struct r10conf *conf = mddev->private;
1578 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1579 int chunk_sects = chunk_mask + 1;
1580 int sectors = bio_sectors(bio);
1582 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1583 && md_flush_request(mddev, bio))
1586 if (!md_write_start(mddev, bio))
1590 * If this request crosses a chunk boundary, we need to split
1593 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1594 sectors > chunk_sects
1595 && (conf->geo.near_copies < conf->geo.raid_disks
1596 || conf->prev.near_copies <
1597 conf->prev.raid_disks)))
1598 sectors = chunk_sects -
1599 (bio->bi_iter.bi_sector &
1601 __make_request(mddev, bio, sectors);
1603 /* In case raid10d snuck in to freeze_array */
1604 wake_up(&conf->wait_barrier);
1608 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1610 struct r10conf *conf = mddev->private;
1613 if (conf->geo.near_copies < conf->geo.raid_disks)
1614 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1615 if (conf->geo.near_copies > 1)
1616 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1617 if (conf->geo.far_copies > 1) {
1618 if (conf->geo.far_offset)
1619 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1621 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1622 if (conf->geo.far_set_size != conf->geo.raid_disks)
1623 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1625 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1626 conf->geo.raid_disks - mddev->degraded);
1628 for (i = 0; i < conf->geo.raid_disks; i++) {
1629 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1630 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1633 seq_printf(seq, "]");
1636 /* check if there are enough drives for
1637 * every block to appear on atleast one.
1638 * Don't consider the device numbered 'ignore'
1639 * as we might be about to remove it.
1641 static int _enough(struct r10conf *conf, int previous, int ignore)
1647 disks = conf->prev.raid_disks;
1648 ncopies = conf->prev.near_copies;
1650 disks = conf->geo.raid_disks;
1651 ncopies = conf->geo.near_copies;
1656 int n = conf->copies;
1660 struct md_rdev *rdev;
1661 if (this != ignore &&
1662 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1663 test_bit(In_sync, &rdev->flags))
1665 this = (this+1) % disks;
1669 first = (first + ncopies) % disks;
1670 } while (first != 0);
1677 static int enough(struct r10conf *conf, int ignore)
1679 /* when calling 'enough', both 'prev' and 'geo' must
1681 * This is ensured if ->reconfig_mutex or ->device_lock
1684 return _enough(conf, 0, ignore) &&
1685 _enough(conf, 1, ignore);
1688 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1690 char b[BDEVNAME_SIZE];
1691 struct r10conf *conf = mddev->private;
1692 unsigned long flags;
1695 * If it is not operational, then we have already marked it as dead
1696 * else if it is the last working disks, ignore the error, let the
1697 * next level up know.
1698 * else mark the drive as failed
1700 spin_lock_irqsave(&conf->device_lock, flags);
1701 if (test_bit(In_sync, &rdev->flags)
1702 && !enough(conf, rdev->raid_disk)) {
1704 * Don't fail the drive, just return an IO error.
1706 spin_unlock_irqrestore(&conf->device_lock, flags);
1709 if (test_and_clear_bit(In_sync, &rdev->flags))
1712 * If recovery is running, make sure it aborts.
1714 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1715 set_bit(Blocked, &rdev->flags);
1716 set_bit(Faulty, &rdev->flags);
1717 set_mask_bits(&mddev->sb_flags, 0,
1718 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1719 spin_unlock_irqrestore(&conf->device_lock, flags);
1720 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1721 "md/raid10:%s: Operation continuing on %d devices.\n",
1722 mdname(mddev), bdevname(rdev->bdev, b),
1723 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1726 static void print_conf(struct r10conf *conf)
1729 struct md_rdev *rdev;
1731 pr_debug("RAID10 conf printout:\n");
1733 pr_debug("(!conf)\n");
1736 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1737 conf->geo.raid_disks);
1739 /* This is only called with ->reconfix_mutex held, so
1740 * rcu protection of rdev is not needed */
1741 for (i = 0; i < conf->geo.raid_disks; i++) {
1742 char b[BDEVNAME_SIZE];
1743 rdev = conf->mirrors[i].rdev;
1745 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1746 i, !test_bit(In_sync, &rdev->flags),
1747 !test_bit(Faulty, &rdev->flags),
1748 bdevname(rdev->bdev,b));
1752 static void close_sync(struct r10conf *conf)
1755 allow_barrier(conf);
1757 mempool_exit(&conf->r10buf_pool);
1760 static int raid10_spare_active(struct mddev *mddev)
1763 struct r10conf *conf = mddev->private;
1764 struct raid10_info *tmp;
1766 unsigned long flags;
1769 * Find all non-in_sync disks within the RAID10 configuration
1770 * and mark them in_sync
1772 for (i = 0; i < conf->geo.raid_disks; i++) {
1773 tmp = conf->mirrors + i;
1774 if (tmp->replacement
1775 && tmp->replacement->recovery_offset == MaxSector
1776 && !test_bit(Faulty, &tmp->replacement->flags)
1777 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1778 /* Replacement has just become active */
1780 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1783 /* Replaced device not technically faulty,
1784 * but we need to be sure it gets removed
1785 * and never re-added.
1787 set_bit(Faulty, &tmp->rdev->flags);
1788 sysfs_notify_dirent_safe(
1789 tmp->rdev->sysfs_state);
1791 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1792 } else if (tmp->rdev
1793 && tmp->rdev->recovery_offset == MaxSector
1794 && !test_bit(Faulty, &tmp->rdev->flags)
1795 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1797 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1800 spin_lock_irqsave(&conf->device_lock, flags);
1801 mddev->degraded -= count;
1802 spin_unlock_irqrestore(&conf->device_lock, flags);
1808 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1810 struct r10conf *conf = mddev->private;
1814 int last = conf->geo.raid_disks - 1;
1816 if (mddev->recovery_cp < MaxSector)
1817 /* only hot-add to in-sync arrays, as recovery is
1818 * very different from resync
1821 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1824 if (md_integrity_add_rdev(rdev, mddev))
1827 if (rdev->raid_disk >= 0)
1828 first = last = rdev->raid_disk;
1830 if (rdev->saved_raid_disk >= first &&
1831 rdev->saved_raid_disk < conf->geo.raid_disks &&
1832 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1833 mirror = rdev->saved_raid_disk;
1836 for ( ; mirror <= last ; mirror++) {
1837 struct raid10_info *p = &conf->mirrors[mirror];
1838 if (p->recovery_disabled == mddev->recovery_disabled)
1841 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1842 p->replacement != NULL)
1844 clear_bit(In_sync, &rdev->flags);
1845 set_bit(Replacement, &rdev->flags);
1846 rdev->raid_disk = mirror;
1849 disk_stack_limits(mddev->gendisk, rdev->bdev,
1850 rdev->data_offset << 9);
1852 rcu_assign_pointer(p->replacement, rdev);
1857 disk_stack_limits(mddev->gendisk, rdev->bdev,
1858 rdev->data_offset << 9);
1860 p->head_position = 0;
1861 p->recovery_disabled = mddev->recovery_disabled - 1;
1862 rdev->raid_disk = mirror;
1864 if (rdev->saved_raid_disk != mirror)
1866 rcu_assign_pointer(p->rdev, rdev);
1869 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1870 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1876 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1878 struct r10conf *conf = mddev->private;
1880 int number = rdev->raid_disk;
1881 struct md_rdev **rdevp;
1882 struct raid10_info *p;
1885 if (unlikely(number >= mddev->raid_disks))
1887 p = conf->mirrors + number;
1888 if (rdev == p->rdev)
1890 else if (rdev == p->replacement)
1891 rdevp = &p->replacement;
1895 if (test_bit(In_sync, &rdev->flags) ||
1896 atomic_read(&rdev->nr_pending)) {
1900 /* Only remove non-faulty devices if recovery
1903 if (!test_bit(Faulty, &rdev->flags) &&
1904 mddev->recovery_disabled != p->recovery_disabled &&
1905 (!p->replacement || p->replacement == rdev) &&
1906 number < conf->geo.raid_disks &&
1912 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1914 if (atomic_read(&rdev->nr_pending)) {
1915 /* lost the race, try later */
1921 if (p->replacement) {
1922 /* We must have just cleared 'rdev' */
1923 p->rdev = p->replacement;
1924 clear_bit(Replacement, &p->replacement->flags);
1925 smp_mb(); /* Make sure other CPUs may see both as identical
1926 * but will never see neither -- if they are careful.
1928 p->replacement = NULL;
1931 clear_bit(WantReplacement, &rdev->flags);
1932 err = md_integrity_register(mddev);
1940 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1942 struct r10conf *conf = r10_bio->mddev->private;
1944 if (!bio->bi_status)
1945 set_bit(R10BIO_Uptodate, &r10_bio->state);
1947 /* The write handler will notice the lack of
1948 * R10BIO_Uptodate and record any errors etc
1950 atomic_add(r10_bio->sectors,
1951 &conf->mirrors[d].rdev->corrected_errors);
1953 /* for reconstruct, we always reschedule after a read.
1954 * for resync, only after all reads
1956 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1957 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1958 atomic_dec_and_test(&r10_bio->remaining)) {
1959 /* we have read all the blocks,
1960 * do the comparison in process context in raid10d
1962 reschedule_retry(r10_bio);
1966 static void end_sync_read(struct bio *bio)
1968 struct r10bio *r10_bio = get_resync_r10bio(bio);
1969 struct r10conf *conf = r10_bio->mddev->private;
1970 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1972 __end_sync_read(r10_bio, bio, d);
1975 static void end_reshape_read(struct bio *bio)
1977 /* reshape read bio isn't allocated from r10buf_pool */
1978 struct r10bio *r10_bio = bio->bi_private;
1980 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1983 static void end_sync_request(struct r10bio *r10_bio)
1985 struct mddev *mddev = r10_bio->mddev;
1987 while (atomic_dec_and_test(&r10_bio->remaining)) {
1988 if (r10_bio->master_bio == NULL) {
1989 /* the primary of several recovery bios */
1990 sector_t s = r10_bio->sectors;
1991 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1992 test_bit(R10BIO_WriteError, &r10_bio->state))
1993 reschedule_retry(r10_bio);
1996 md_done_sync(mddev, s, 1);
1999 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2000 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2001 test_bit(R10BIO_WriteError, &r10_bio->state))
2002 reschedule_retry(r10_bio);
2010 static void end_sync_write(struct bio *bio)
2012 struct r10bio *r10_bio = get_resync_r10bio(bio);
2013 struct mddev *mddev = r10_bio->mddev;
2014 struct r10conf *conf = mddev->private;
2020 struct md_rdev *rdev = NULL;
2022 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2024 rdev = conf->mirrors[d].replacement;
2026 rdev = conf->mirrors[d].rdev;
2028 if (bio->bi_status) {
2030 md_error(mddev, rdev);
2032 set_bit(WriteErrorSeen, &rdev->flags);
2033 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2034 set_bit(MD_RECOVERY_NEEDED,
2035 &rdev->mddev->recovery);
2036 set_bit(R10BIO_WriteError, &r10_bio->state);
2038 } else if (is_badblock(rdev,
2039 r10_bio->devs[slot].addr,
2041 &first_bad, &bad_sectors))
2042 set_bit(R10BIO_MadeGood, &r10_bio->state);
2044 rdev_dec_pending(rdev, mddev);
2046 end_sync_request(r10_bio);
2050 * Note: sync and recover and handled very differently for raid10
2051 * This code is for resync.
2052 * For resync, we read through virtual addresses and read all blocks.
2053 * If there is any error, we schedule a write. The lowest numbered
2054 * drive is authoritative.
2055 * However requests come for physical address, so we need to map.
2056 * For every physical address there are raid_disks/copies virtual addresses,
2057 * which is always are least one, but is not necessarly an integer.
2058 * This means that a physical address can span multiple chunks, so we may
2059 * have to submit multiple io requests for a single sync request.
2062 * We check if all blocks are in-sync and only write to blocks that
2065 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2067 struct r10conf *conf = mddev->private;
2069 struct bio *tbio, *fbio;
2071 struct page **tpages, **fpages;
2073 atomic_set(&r10_bio->remaining, 1);
2075 /* find the first device with a block */
2076 for (i=0; i<conf->copies; i++)
2077 if (!r10_bio->devs[i].bio->bi_status)
2080 if (i == conf->copies)
2084 fbio = r10_bio->devs[i].bio;
2085 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2086 fbio->bi_iter.bi_idx = 0;
2087 fpages = get_resync_pages(fbio)->pages;
2089 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2090 /* now find blocks with errors */
2091 for (i=0 ; i < conf->copies ; i++) {
2093 struct md_rdev *rdev;
2094 struct resync_pages *rp;
2096 tbio = r10_bio->devs[i].bio;
2098 if (tbio->bi_end_io != end_sync_read)
2103 tpages = get_resync_pages(tbio)->pages;
2104 d = r10_bio->devs[i].devnum;
2105 rdev = conf->mirrors[d].rdev;
2106 if (!r10_bio->devs[i].bio->bi_status) {
2107 /* We know that the bi_io_vec layout is the same for
2108 * both 'first' and 'i', so we just compare them.
2109 * All vec entries are PAGE_SIZE;
2111 int sectors = r10_bio->sectors;
2112 for (j = 0; j < vcnt; j++) {
2113 int len = PAGE_SIZE;
2114 if (sectors < (len / 512))
2115 len = sectors * 512;
2116 if (memcmp(page_address(fpages[j]),
2117 page_address(tpages[j]),
2124 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2125 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2126 /* Don't fix anything. */
2128 } else if (test_bit(FailFast, &rdev->flags)) {
2129 /* Just give up on this device */
2130 md_error(rdev->mddev, rdev);
2133 /* Ok, we need to write this bio, either to correct an
2134 * inconsistency or to correct an unreadable block.
2135 * First we need to fixup bv_offset, bv_len and
2136 * bi_vecs, as the read request might have corrupted these
2138 rp = get_resync_pages(tbio);
2141 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2143 rp->raid_bio = r10_bio;
2144 tbio->bi_private = rp;
2145 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2146 tbio->bi_end_io = end_sync_write;
2147 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2149 bio_copy_data(tbio, fbio);
2151 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2152 atomic_inc(&r10_bio->remaining);
2153 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2155 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2156 tbio->bi_opf |= MD_FAILFAST;
2157 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2158 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2159 generic_make_request(tbio);
2162 /* Now write out to any replacement devices
2165 for (i = 0; i < conf->copies; i++) {
2168 tbio = r10_bio->devs[i].repl_bio;
2169 if (!tbio || !tbio->bi_end_io)
2171 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2172 && r10_bio->devs[i].bio != fbio)
2173 bio_copy_data(tbio, fbio);
2174 d = r10_bio->devs[i].devnum;
2175 atomic_inc(&r10_bio->remaining);
2176 md_sync_acct(conf->mirrors[d].replacement->bdev,
2178 generic_make_request(tbio);
2182 if (atomic_dec_and_test(&r10_bio->remaining)) {
2183 md_done_sync(mddev, r10_bio->sectors, 1);
2189 * Now for the recovery code.
2190 * Recovery happens across physical sectors.
2191 * We recover all non-is_sync drives by finding the virtual address of
2192 * each, and then choose a working drive that also has that virt address.
2193 * There is a separate r10_bio for each non-in_sync drive.
2194 * Only the first two slots are in use. The first for reading,
2195 * The second for writing.
2198 static void fix_recovery_read_error(struct r10bio *r10_bio)
2200 /* We got a read error during recovery.
2201 * We repeat the read in smaller page-sized sections.
2202 * If a read succeeds, write it to the new device or record
2203 * a bad block if we cannot.
2204 * If a read fails, record a bad block on both old and
2207 struct mddev *mddev = r10_bio->mddev;
2208 struct r10conf *conf = mddev->private;
2209 struct bio *bio = r10_bio->devs[0].bio;
2211 int sectors = r10_bio->sectors;
2213 int dr = r10_bio->devs[0].devnum;
2214 int dw = r10_bio->devs[1].devnum;
2215 struct page **pages = get_resync_pages(bio)->pages;
2219 struct md_rdev *rdev;
2223 if (s > (PAGE_SIZE>>9))
2226 rdev = conf->mirrors[dr].rdev;
2227 addr = r10_bio->devs[0].addr + sect,
2228 ok = sync_page_io(rdev,
2232 REQ_OP_READ, 0, false);
2234 rdev = conf->mirrors[dw].rdev;
2235 addr = r10_bio->devs[1].addr + sect;
2236 ok = sync_page_io(rdev,
2240 REQ_OP_WRITE, 0, false);
2242 set_bit(WriteErrorSeen, &rdev->flags);
2243 if (!test_and_set_bit(WantReplacement,
2245 set_bit(MD_RECOVERY_NEEDED,
2246 &rdev->mddev->recovery);
2250 /* We don't worry if we cannot set a bad block -
2251 * it really is bad so there is no loss in not
2254 rdev_set_badblocks(rdev, addr, s, 0);
2256 if (rdev != conf->mirrors[dw].rdev) {
2257 /* need bad block on destination too */
2258 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2259 addr = r10_bio->devs[1].addr + sect;
2260 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2262 /* just abort the recovery */
2263 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2266 conf->mirrors[dw].recovery_disabled
2267 = mddev->recovery_disabled;
2268 set_bit(MD_RECOVERY_INTR,
2281 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2283 struct r10conf *conf = mddev->private;
2285 struct bio *wbio = r10_bio->devs[1].bio;
2286 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2288 /* Need to test wbio2->bi_end_io before we call
2289 * generic_make_request as if the former is NULL,
2290 * the latter is free to free wbio2.
2292 if (wbio2 && !wbio2->bi_end_io)
2295 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2296 fix_recovery_read_error(r10_bio);
2297 if (wbio->bi_end_io)
2298 end_sync_request(r10_bio);
2300 end_sync_request(r10_bio);
2305 * share the pages with the first bio
2306 * and submit the write request
2308 d = r10_bio->devs[1].devnum;
2309 if (wbio->bi_end_io) {
2310 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2311 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2312 generic_make_request(wbio);
2315 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2316 md_sync_acct(conf->mirrors[d].replacement->bdev,
2317 bio_sectors(wbio2));
2318 generic_make_request(wbio2);
2323 * Used by fix_read_error() to decay the per rdev read_errors.
2324 * We halve the read error count for every hour that has elapsed
2325 * since the last recorded read error.
2328 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2331 unsigned long hours_since_last;
2332 unsigned int read_errors = atomic_read(&rdev->read_errors);
2334 cur_time_mon = ktime_get_seconds();
2336 if (rdev->last_read_error == 0) {
2337 /* first time we've seen a read error */
2338 rdev->last_read_error = cur_time_mon;
2342 hours_since_last = (long)(cur_time_mon -
2343 rdev->last_read_error) / 3600;
2345 rdev->last_read_error = cur_time_mon;
2348 * if hours_since_last is > the number of bits in read_errors
2349 * just set read errors to 0. We do this to avoid
2350 * overflowing the shift of read_errors by hours_since_last.
2352 if (hours_since_last >= 8 * sizeof(read_errors))
2353 atomic_set(&rdev->read_errors, 0);
2355 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2358 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2359 int sectors, struct page *page, int rw)
2364 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2365 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2367 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2371 set_bit(WriteErrorSeen, &rdev->flags);
2372 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2373 set_bit(MD_RECOVERY_NEEDED,
2374 &rdev->mddev->recovery);
2376 /* need to record an error - either for the block or the device */
2377 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2378 md_error(rdev->mddev, rdev);
2383 * This is a kernel thread which:
2385 * 1. Retries failed read operations on working mirrors.
2386 * 2. Updates the raid superblock when problems encounter.
2387 * 3. Performs writes following reads for array synchronising.
2390 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2392 int sect = 0; /* Offset from r10_bio->sector */
2393 int sectors = r10_bio->sectors;
2394 struct md_rdev *rdev;
2395 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2396 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2398 /* still own a reference to this rdev, so it cannot
2399 * have been cleared recently.
2401 rdev = conf->mirrors[d].rdev;
2403 if (test_bit(Faulty, &rdev->flags))
2404 /* drive has already been failed, just ignore any
2405 more fix_read_error() attempts */
2408 check_decay_read_errors(mddev, rdev);
2409 atomic_inc(&rdev->read_errors);
2410 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2411 char b[BDEVNAME_SIZE];
2412 bdevname(rdev->bdev, b);
2414 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2416 atomic_read(&rdev->read_errors), max_read_errors);
2417 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2419 md_error(mddev, rdev);
2420 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2426 int sl = r10_bio->read_slot;
2430 if (s > (PAGE_SIZE>>9))
2438 d = r10_bio->devs[sl].devnum;
2439 rdev = rcu_dereference(conf->mirrors[d].rdev);
2441 test_bit(In_sync, &rdev->flags) &&
2442 !test_bit(Faulty, &rdev->flags) &&
2443 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2444 &first_bad, &bad_sectors) == 0) {
2445 atomic_inc(&rdev->nr_pending);
2447 success = sync_page_io(rdev,
2448 r10_bio->devs[sl].addr +
2452 REQ_OP_READ, 0, false);
2453 rdev_dec_pending(rdev, mddev);
2459 if (sl == conf->copies)
2461 } while (!success && sl != r10_bio->read_slot);
2465 /* Cannot read from anywhere, just mark the block
2466 * as bad on the first device to discourage future
2469 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2470 rdev = conf->mirrors[dn].rdev;
2472 if (!rdev_set_badblocks(
2474 r10_bio->devs[r10_bio->read_slot].addr
2477 md_error(mddev, rdev);
2478 r10_bio->devs[r10_bio->read_slot].bio
2485 /* write it back and re-read */
2487 while (sl != r10_bio->read_slot) {
2488 char b[BDEVNAME_SIZE];
2493 d = r10_bio->devs[sl].devnum;
2494 rdev = rcu_dereference(conf->mirrors[d].rdev);
2496 test_bit(Faulty, &rdev->flags) ||
2497 !test_bit(In_sync, &rdev->flags))
2500 atomic_inc(&rdev->nr_pending);
2502 if (r10_sync_page_io(rdev,
2503 r10_bio->devs[sl].addr +
2505 s, conf->tmppage, WRITE)
2507 /* Well, this device is dead */
2508 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2510 (unsigned long long)(
2512 choose_data_offset(r10_bio,
2514 bdevname(rdev->bdev, b));
2515 pr_notice("md/raid10:%s: %s: failing drive\n",
2517 bdevname(rdev->bdev, b));
2519 rdev_dec_pending(rdev, mddev);
2523 while (sl != r10_bio->read_slot) {
2524 char b[BDEVNAME_SIZE];
2529 d = r10_bio->devs[sl].devnum;
2530 rdev = rcu_dereference(conf->mirrors[d].rdev);
2532 test_bit(Faulty, &rdev->flags) ||
2533 !test_bit(In_sync, &rdev->flags))
2536 atomic_inc(&rdev->nr_pending);
2538 switch (r10_sync_page_io(rdev,
2539 r10_bio->devs[sl].addr +
2544 /* Well, this device is dead */
2545 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2547 (unsigned long long)(
2549 choose_data_offset(r10_bio, rdev)),
2550 bdevname(rdev->bdev, b));
2551 pr_notice("md/raid10:%s: %s: failing drive\n",
2553 bdevname(rdev->bdev, b));
2556 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2558 (unsigned long long)(
2560 choose_data_offset(r10_bio, rdev)),
2561 bdevname(rdev->bdev, b));
2562 atomic_add(s, &rdev->corrected_errors);
2565 rdev_dec_pending(rdev, mddev);
2575 static int narrow_write_error(struct r10bio *r10_bio, int i)
2577 struct bio *bio = r10_bio->master_bio;
2578 struct mddev *mddev = r10_bio->mddev;
2579 struct r10conf *conf = mddev->private;
2580 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2581 /* bio has the data to be written to slot 'i' where
2582 * we just recently had a write error.
2583 * We repeatedly clone the bio and trim down to one block,
2584 * then try the write. Where the write fails we record
2586 * It is conceivable that the bio doesn't exactly align with
2587 * blocks. We must handle this.
2589 * We currently own a reference to the rdev.
2595 int sect_to_write = r10_bio->sectors;
2598 if (rdev->badblocks.shift < 0)
2601 block_sectors = roundup(1 << rdev->badblocks.shift,
2602 bdev_logical_block_size(rdev->bdev) >> 9);
2603 sector = r10_bio->sector;
2604 sectors = ((r10_bio->sector + block_sectors)
2605 & ~(sector_t)(block_sectors - 1))
2608 while (sect_to_write) {
2611 if (sectors > sect_to_write)
2612 sectors = sect_to_write;
2613 /* Write at 'sector' for 'sectors' */
2614 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2615 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2616 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2617 wbio->bi_iter.bi_sector = wsector +
2618 choose_data_offset(r10_bio, rdev);
2619 bio_set_dev(wbio, rdev->bdev);
2620 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2622 if (submit_bio_wait(wbio) < 0)
2624 ok = rdev_set_badblocks(rdev, wsector,
2629 sect_to_write -= sectors;
2631 sectors = block_sectors;
2636 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2638 int slot = r10_bio->read_slot;
2640 struct r10conf *conf = mddev->private;
2641 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2643 /* we got a read error. Maybe the drive is bad. Maybe just
2644 * the block and we can fix it.
2645 * We freeze all other IO, and try reading the block from
2646 * other devices. When we find one, we re-write
2647 * and check it that fixes the read error.
2648 * This is all done synchronously while the array is
2651 bio = r10_bio->devs[slot].bio;
2653 r10_bio->devs[slot].bio = NULL;
2656 r10_bio->devs[slot].bio = IO_BLOCKED;
2657 else if (!test_bit(FailFast, &rdev->flags)) {
2658 freeze_array(conf, 1);
2659 fix_read_error(conf, mddev, r10_bio);
2660 unfreeze_array(conf);
2662 md_error(mddev, rdev);
2664 rdev_dec_pending(rdev, mddev);
2665 allow_barrier(conf);
2667 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2670 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2672 /* Some sort of write request has finished and it
2673 * succeeded in writing where we thought there was a
2674 * bad block. So forget the bad block.
2675 * Or possibly if failed and we need to record
2679 struct md_rdev *rdev;
2681 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2682 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2683 for (m = 0; m < conf->copies; m++) {
2684 int dev = r10_bio->devs[m].devnum;
2685 rdev = conf->mirrors[dev].rdev;
2686 if (r10_bio->devs[m].bio == NULL ||
2687 r10_bio->devs[m].bio->bi_end_io == NULL)
2689 if (!r10_bio->devs[m].bio->bi_status) {
2690 rdev_clear_badblocks(
2692 r10_bio->devs[m].addr,
2693 r10_bio->sectors, 0);
2695 if (!rdev_set_badblocks(
2697 r10_bio->devs[m].addr,
2698 r10_bio->sectors, 0))
2699 md_error(conf->mddev, rdev);
2701 rdev = conf->mirrors[dev].replacement;
2702 if (r10_bio->devs[m].repl_bio == NULL ||
2703 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2706 if (!r10_bio->devs[m].repl_bio->bi_status) {
2707 rdev_clear_badblocks(
2709 r10_bio->devs[m].addr,
2710 r10_bio->sectors, 0);
2712 if (!rdev_set_badblocks(
2714 r10_bio->devs[m].addr,
2715 r10_bio->sectors, 0))
2716 md_error(conf->mddev, rdev);
2722 for (m = 0; m < conf->copies; m++) {
2723 int dev = r10_bio->devs[m].devnum;
2724 struct bio *bio = r10_bio->devs[m].bio;
2725 rdev = conf->mirrors[dev].rdev;
2726 if (bio == IO_MADE_GOOD) {
2727 rdev_clear_badblocks(
2729 r10_bio->devs[m].addr,
2730 r10_bio->sectors, 0);
2731 rdev_dec_pending(rdev, conf->mddev);
2732 } else if (bio != NULL && bio->bi_status) {
2734 if (!narrow_write_error(r10_bio, m)) {
2735 md_error(conf->mddev, rdev);
2736 set_bit(R10BIO_Degraded,
2739 rdev_dec_pending(rdev, conf->mddev);
2741 bio = r10_bio->devs[m].repl_bio;
2742 rdev = conf->mirrors[dev].replacement;
2743 if (rdev && bio == IO_MADE_GOOD) {
2744 rdev_clear_badblocks(
2746 r10_bio->devs[m].addr,
2747 r10_bio->sectors, 0);
2748 rdev_dec_pending(rdev, conf->mddev);
2752 spin_lock_irq(&conf->device_lock);
2753 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2755 spin_unlock_irq(&conf->device_lock);
2757 * In case freeze_array() is waiting for condition
2758 * nr_pending == nr_queued + extra to be true.
2760 wake_up(&conf->wait_barrier);
2761 md_wakeup_thread(conf->mddev->thread);
2763 if (test_bit(R10BIO_WriteError,
2765 close_write(r10_bio);
2766 raid_end_bio_io(r10_bio);
2771 static void raid10d(struct md_thread *thread)
2773 struct mddev *mddev = thread->mddev;
2774 struct r10bio *r10_bio;
2775 unsigned long flags;
2776 struct r10conf *conf = mddev->private;
2777 struct list_head *head = &conf->retry_list;
2778 struct blk_plug plug;
2780 md_check_recovery(mddev);
2782 if (!list_empty_careful(&conf->bio_end_io_list) &&
2783 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2785 spin_lock_irqsave(&conf->device_lock, flags);
2786 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2787 while (!list_empty(&conf->bio_end_io_list)) {
2788 list_move(conf->bio_end_io_list.prev, &tmp);
2792 spin_unlock_irqrestore(&conf->device_lock, flags);
2793 while (!list_empty(&tmp)) {
2794 r10_bio = list_first_entry(&tmp, struct r10bio,
2796 list_del(&r10_bio->retry_list);
2797 if (mddev->degraded)
2798 set_bit(R10BIO_Degraded, &r10_bio->state);
2800 if (test_bit(R10BIO_WriteError,
2802 close_write(r10_bio);
2803 raid_end_bio_io(r10_bio);
2807 blk_start_plug(&plug);
2810 flush_pending_writes(conf);
2812 spin_lock_irqsave(&conf->device_lock, flags);
2813 if (list_empty(head)) {
2814 spin_unlock_irqrestore(&conf->device_lock, flags);
2817 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2818 list_del(head->prev);
2820 spin_unlock_irqrestore(&conf->device_lock, flags);
2822 mddev = r10_bio->mddev;
2823 conf = mddev->private;
2824 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2825 test_bit(R10BIO_WriteError, &r10_bio->state))
2826 handle_write_completed(conf, r10_bio);
2827 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2828 reshape_request_write(mddev, r10_bio);
2829 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2830 sync_request_write(mddev, r10_bio);
2831 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2832 recovery_request_write(mddev, r10_bio);
2833 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2834 handle_read_error(mddev, r10_bio);
2839 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2840 md_check_recovery(mddev);
2842 blk_finish_plug(&plug);
2845 static int init_resync(struct r10conf *conf)
2849 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2850 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2851 conf->have_replacement = 0;
2852 for (i = 0; i < conf->geo.raid_disks; i++)
2853 if (conf->mirrors[i].replacement)
2854 conf->have_replacement = 1;
2855 ret = mempool_init(&conf->r10buf_pool, buffs,
2856 r10buf_pool_alloc, r10buf_pool_free, conf);
2859 conf->next_resync = 0;
2863 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2865 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2866 struct rsync_pages *rp;
2871 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2872 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2873 nalloc = conf->copies; /* resync */
2875 nalloc = 2; /* recovery */
2877 for (i = 0; i < nalloc; i++) {
2878 bio = r10bio->devs[i].bio;
2879 rp = bio->bi_private;
2881 bio->bi_private = rp;
2882 bio = r10bio->devs[i].repl_bio;
2884 rp = bio->bi_private;
2886 bio->bi_private = rp;
2893 * Set cluster_sync_high since we need other nodes to add the
2894 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2896 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2898 sector_t window_size;
2899 int extra_chunk, chunks;
2902 * First, here we define "stripe" as a unit which across
2903 * all member devices one time, so we get chunks by use
2904 * raid_disks / near_copies. Otherwise, if near_copies is
2905 * close to raid_disks, then resync window could increases
2906 * linearly with the increase of raid_disks, which means
2907 * we will suspend a really large IO window while it is not
2908 * necessary. If raid_disks is not divisible by near_copies,
2909 * an extra chunk is needed to ensure the whole "stripe" is
2913 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2914 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2918 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2921 * At least use a 32M window to align with raid1's resync window
2923 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2924 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2926 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2930 * perform a "sync" on one "block"
2932 * We need to make sure that no normal I/O request - particularly write
2933 * requests - conflict with active sync requests.
2935 * This is achieved by tracking pending requests and a 'barrier' concept
2936 * that can be installed to exclude normal IO requests.
2938 * Resync and recovery are handled very differently.
2939 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2941 * For resync, we iterate over virtual addresses, read all copies,
2942 * and update if there are differences. If only one copy is live,
2944 * For recovery, we iterate over physical addresses, read a good
2945 * value for each non-in_sync drive, and over-write.
2947 * So, for recovery we may have several outstanding complex requests for a
2948 * given address, one for each out-of-sync device. We model this by allocating
2949 * a number of r10_bio structures, one for each out-of-sync device.
2950 * As we setup these structures, we collect all bio's together into a list
2951 * which we then process collectively to add pages, and then process again
2952 * to pass to generic_make_request.
2954 * The r10_bio structures are linked using a borrowed master_bio pointer.
2955 * This link is counted in ->remaining. When the r10_bio that points to NULL
2956 * has its remaining count decremented to 0, the whole complex operation
2961 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2964 struct r10conf *conf = mddev->private;
2965 struct r10bio *r10_bio;
2966 struct bio *biolist = NULL, *bio;
2967 sector_t max_sector, nr_sectors;
2970 sector_t sync_blocks;
2971 sector_t sectors_skipped = 0;
2972 int chunks_skipped = 0;
2973 sector_t chunk_mask = conf->geo.chunk_mask;
2977 * Allow skipping a full rebuild for incremental assembly
2978 * of a clean array, like RAID1 does.
2980 if (mddev->bitmap == NULL &&
2981 mddev->recovery_cp == MaxSector &&
2982 mddev->reshape_position == MaxSector &&
2983 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2984 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2985 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2986 conf->fullsync == 0) {
2988 return mddev->dev_sectors - sector_nr;
2991 if (!mempool_initialized(&conf->r10buf_pool))
2992 if (init_resync(conf))
2996 max_sector = mddev->dev_sectors;
2997 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2998 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2999 max_sector = mddev->resync_max_sectors;
3000 if (sector_nr >= max_sector) {
3001 conf->cluster_sync_low = 0;
3002 conf->cluster_sync_high = 0;
3004 /* If we aborted, we need to abort the
3005 * sync on the 'current' bitmap chucks (there can
3006 * be several when recovering multiple devices).
3007 * as we may have started syncing it but not finished.
3008 * We can find the current address in
3009 * mddev->curr_resync, but for recovery,
3010 * we need to convert that to several
3011 * virtual addresses.
3013 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3019 if (mddev->curr_resync < max_sector) { /* aborted */
3020 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3021 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3023 else for (i = 0; i < conf->geo.raid_disks; i++) {
3025 raid10_find_virt(conf, mddev->curr_resync, i);
3026 md_bitmap_end_sync(mddev->bitmap, sect,
3030 /* completed sync */
3031 if ((!mddev->bitmap || conf->fullsync)
3032 && conf->have_replacement
3033 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3034 /* Completed a full sync so the replacements
3035 * are now fully recovered.
3038 for (i = 0; i < conf->geo.raid_disks; i++) {
3039 struct md_rdev *rdev =
3040 rcu_dereference(conf->mirrors[i].replacement);
3042 rdev->recovery_offset = MaxSector;
3048 md_bitmap_close_sync(mddev->bitmap);
3051 return sectors_skipped;
3054 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3055 return reshape_request(mddev, sector_nr, skipped);
3057 if (chunks_skipped >= conf->geo.raid_disks) {
3058 /* if there has been nothing to do on any drive,
3059 * then there is nothing to do at all..
3062 return (max_sector - sector_nr) + sectors_skipped;
3065 if (max_sector > mddev->resync_max)
3066 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3068 /* make sure whole request will fit in a chunk - if chunks
3071 if (conf->geo.near_copies < conf->geo.raid_disks &&
3072 max_sector > (sector_nr | chunk_mask))
3073 max_sector = (sector_nr | chunk_mask) + 1;
3076 * If there is non-resync activity waiting for a turn, then let it
3077 * though before starting on this new sync request.
3079 if (conf->nr_waiting)
3080 schedule_timeout_uninterruptible(1);
3082 /* Again, very different code for resync and recovery.
3083 * Both must result in an r10bio with a list of bios that
3084 * have bi_end_io, bi_sector, bi_disk set,
3085 * and bi_private set to the r10bio.
3086 * For recovery, we may actually create several r10bios
3087 * with 2 bios in each, that correspond to the bios in the main one.
3088 * In this case, the subordinate r10bios link back through a
3089 * borrowed master_bio pointer, and the counter in the master
3090 * includes a ref from each subordinate.
3092 /* First, we decide what to do and set ->bi_end_io
3093 * To end_sync_read if we want to read, and
3094 * end_sync_write if we will want to write.
3097 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3098 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3099 /* recovery... the complicated one */
3103 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3109 struct raid10_info *mirror = &conf->mirrors[i];
3110 struct md_rdev *mrdev, *mreplace;
3113 mrdev = rcu_dereference(mirror->rdev);
3114 mreplace = rcu_dereference(mirror->replacement);
3116 if ((mrdev == NULL ||
3117 test_bit(Faulty, &mrdev->flags) ||
3118 test_bit(In_sync, &mrdev->flags)) &&
3119 (mreplace == NULL ||
3120 test_bit(Faulty, &mreplace->flags))) {
3126 /* want to reconstruct this device */
3128 sect = raid10_find_virt(conf, sector_nr, i);
3129 if (sect >= mddev->resync_max_sectors) {
3130 /* last stripe is not complete - don't
3131 * try to recover this sector.
3136 if (mreplace && test_bit(Faulty, &mreplace->flags))
3138 /* Unless we are doing a full sync, or a replacement
3139 * we only need to recover the block if it is set in
3142 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3144 if (sync_blocks < max_sync)
3145 max_sync = sync_blocks;
3149 /* yep, skip the sync_blocks here, but don't assume
3150 * that there will never be anything to do here
3152 chunks_skipped = -1;
3156 atomic_inc(&mrdev->nr_pending);
3158 atomic_inc(&mreplace->nr_pending);
3161 r10_bio = raid10_alloc_init_r10buf(conf);
3163 raise_barrier(conf, rb2 != NULL);
3164 atomic_set(&r10_bio->remaining, 0);
3166 r10_bio->master_bio = (struct bio*)rb2;
3168 atomic_inc(&rb2->remaining);
3169 r10_bio->mddev = mddev;
3170 set_bit(R10BIO_IsRecover, &r10_bio->state);
3171 r10_bio->sector = sect;
3173 raid10_find_phys(conf, r10_bio);
3175 /* Need to check if the array will still be
3179 for (j = 0; j < conf->geo.raid_disks; j++) {
3180 struct md_rdev *rdev = rcu_dereference(
3181 conf->mirrors[j].rdev);
3182 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3188 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3189 &sync_blocks, still_degraded);
3192 for (j=0; j<conf->copies;j++) {
3194 int d = r10_bio->devs[j].devnum;
3195 sector_t from_addr, to_addr;
3196 struct md_rdev *rdev =
3197 rcu_dereference(conf->mirrors[d].rdev);
3198 sector_t sector, first_bad;
3201 !test_bit(In_sync, &rdev->flags))
3203 /* This is where we read from */
3205 sector = r10_bio->devs[j].addr;
3207 if (is_badblock(rdev, sector, max_sync,
3208 &first_bad, &bad_sectors)) {
3209 if (first_bad > sector)
3210 max_sync = first_bad - sector;
3212 bad_sectors -= (sector
3214 if (max_sync > bad_sectors)
3215 max_sync = bad_sectors;
3219 bio = r10_bio->devs[0].bio;
3220 bio->bi_next = biolist;
3222 bio->bi_end_io = end_sync_read;
3223 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3224 if (test_bit(FailFast, &rdev->flags))
3225 bio->bi_opf |= MD_FAILFAST;
3226 from_addr = r10_bio->devs[j].addr;
3227 bio->bi_iter.bi_sector = from_addr +
3229 bio_set_dev(bio, rdev->bdev);
3230 atomic_inc(&rdev->nr_pending);
3231 /* and we write to 'i' (if not in_sync) */
3233 for (k=0; k<conf->copies; k++)
3234 if (r10_bio->devs[k].devnum == i)
3236 BUG_ON(k == conf->copies);
3237 to_addr = r10_bio->devs[k].addr;
3238 r10_bio->devs[0].devnum = d;
3239 r10_bio->devs[0].addr = from_addr;
3240 r10_bio->devs[1].devnum = i;
3241 r10_bio->devs[1].addr = to_addr;
3243 if (!test_bit(In_sync, &mrdev->flags)) {
3244 bio = r10_bio->devs[1].bio;
3245 bio->bi_next = biolist;
3247 bio->bi_end_io = end_sync_write;
3248 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3249 bio->bi_iter.bi_sector = to_addr
3250 + mrdev->data_offset;
3251 bio_set_dev(bio, mrdev->bdev);
3252 atomic_inc(&r10_bio->remaining);
3254 r10_bio->devs[1].bio->bi_end_io = NULL;
3256 /* and maybe write to replacement */
3257 bio = r10_bio->devs[1].repl_bio;
3259 bio->bi_end_io = NULL;
3260 /* Note: if mreplace != NULL, then bio
3261 * cannot be NULL as r10buf_pool_alloc will
3262 * have allocated it.
3263 * So the second test here is pointless.
3264 * But it keeps semantic-checkers happy, and
3265 * this comment keeps human reviewers
3268 if (mreplace == NULL || bio == NULL ||
3269 test_bit(Faulty, &mreplace->flags))
3271 bio->bi_next = biolist;
3273 bio->bi_end_io = end_sync_write;
3274 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3275 bio->bi_iter.bi_sector = to_addr +
3276 mreplace->data_offset;
3277 bio_set_dev(bio, mreplace->bdev);
3278 atomic_inc(&r10_bio->remaining);
3282 if (j == conf->copies) {
3283 /* Cannot recover, so abort the recovery or
3284 * record a bad block */
3286 /* problem is that there are bad blocks
3287 * on other device(s)
3290 for (k = 0; k < conf->copies; k++)
3291 if (r10_bio->devs[k].devnum == i)
3293 if (!test_bit(In_sync,
3295 && !rdev_set_badblocks(
3297 r10_bio->devs[k].addr,
3301 !rdev_set_badblocks(
3303 r10_bio->devs[k].addr,
3308 if (!test_and_set_bit(MD_RECOVERY_INTR,
3310 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3312 mirror->recovery_disabled
3313 = mddev->recovery_disabled;
3317 atomic_dec(&rb2->remaining);
3319 rdev_dec_pending(mrdev, mddev);
3321 rdev_dec_pending(mreplace, mddev);
3324 rdev_dec_pending(mrdev, mddev);
3326 rdev_dec_pending(mreplace, mddev);
3327 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3328 /* Only want this if there is elsewhere to
3329 * read from. 'j' is currently the first
3333 for (; j < conf->copies; j++) {
3334 int d = r10_bio->devs[j].devnum;
3335 if (conf->mirrors[d].rdev &&
3337 &conf->mirrors[d].rdev->flags))
3341 r10_bio->devs[0].bio->bi_opf
3345 if (biolist == NULL) {
3347 struct r10bio *rb2 = r10_bio;
3348 r10_bio = (struct r10bio*) rb2->master_bio;
3349 rb2->master_bio = NULL;
3355 /* resync. Schedule a read for every block at this virt offset */
3359 * Since curr_resync_completed could probably not update in
3360 * time, and we will set cluster_sync_low based on it.
3361 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3362 * safety reason, which ensures curr_resync_completed is
3363 * updated in bitmap_cond_end_sync.
3365 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3366 mddev_is_clustered(mddev) &&
3367 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3369 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3370 &sync_blocks, mddev->degraded) &&
3371 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3372 &mddev->recovery)) {
3373 /* We can skip this block */
3375 return sync_blocks + sectors_skipped;
3377 if (sync_blocks < max_sync)
3378 max_sync = sync_blocks;
3379 r10_bio = raid10_alloc_init_r10buf(conf);
3382 r10_bio->mddev = mddev;
3383 atomic_set(&r10_bio->remaining, 0);
3384 raise_barrier(conf, 0);
3385 conf->next_resync = sector_nr;
3387 r10_bio->master_bio = NULL;
3388 r10_bio->sector = sector_nr;
3389 set_bit(R10BIO_IsSync, &r10_bio->state);
3390 raid10_find_phys(conf, r10_bio);
3391 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3393 for (i = 0; i < conf->copies; i++) {
3394 int d = r10_bio->devs[i].devnum;
3395 sector_t first_bad, sector;
3397 struct md_rdev *rdev;
3399 if (r10_bio->devs[i].repl_bio)
3400 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3402 bio = r10_bio->devs[i].bio;
3403 bio->bi_status = BLK_STS_IOERR;
3405 rdev = rcu_dereference(conf->mirrors[d].rdev);
3406 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3410 sector = r10_bio->devs[i].addr;
3411 if (is_badblock(rdev, sector, max_sync,
3412 &first_bad, &bad_sectors)) {
3413 if (first_bad > sector)
3414 max_sync = first_bad - sector;
3416 bad_sectors -= (sector - first_bad);
3417 if (max_sync > bad_sectors)
3418 max_sync = bad_sectors;
3423 atomic_inc(&rdev->nr_pending);
3424 atomic_inc(&r10_bio->remaining);
3425 bio->bi_next = biolist;
3427 bio->bi_end_io = end_sync_read;
3428 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3429 if (test_bit(FailFast, &rdev->flags))
3430 bio->bi_opf |= MD_FAILFAST;
3431 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3432 bio_set_dev(bio, rdev->bdev);
3435 rdev = rcu_dereference(conf->mirrors[d].replacement);
3436 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3440 atomic_inc(&rdev->nr_pending);
3442 /* Need to set up for writing to the replacement */
3443 bio = r10_bio->devs[i].repl_bio;
3444 bio->bi_status = BLK_STS_IOERR;
3446 sector = r10_bio->devs[i].addr;
3447 bio->bi_next = biolist;
3449 bio->bi_end_io = end_sync_write;
3450 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3451 if (test_bit(FailFast, &rdev->flags))
3452 bio->bi_opf |= MD_FAILFAST;
3453 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3454 bio_set_dev(bio, rdev->bdev);
3460 for (i=0; i<conf->copies; i++) {
3461 int d = r10_bio->devs[i].devnum;
3462 if (r10_bio->devs[i].bio->bi_end_io)
3463 rdev_dec_pending(conf->mirrors[d].rdev,
3465 if (r10_bio->devs[i].repl_bio &&
3466 r10_bio->devs[i].repl_bio->bi_end_io)
3468 conf->mirrors[d].replacement,
3478 if (sector_nr + max_sync < max_sector)
3479 max_sector = sector_nr + max_sync;
3482 int len = PAGE_SIZE;
3483 if (sector_nr + (len>>9) > max_sector)
3484 len = (max_sector - sector_nr) << 9;
3487 for (bio= biolist ; bio ; bio=bio->bi_next) {
3488 struct resync_pages *rp = get_resync_pages(bio);
3489 page = resync_fetch_page(rp, page_idx);
3491 * won't fail because the vec table is big enough
3492 * to hold all these pages
3494 bio_add_page(bio, page, len, 0);
3496 nr_sectors += len>>9;
3497 sector_nr += len>>9;
3498 } while (++page_idx < RESYNC_PAGES);
3499 r10_bio->sectors = nr_sectors;
3501 if (mddev_is_clustered(mddev) &&
3502 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3503 /* It is resync not recovery */
3504 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3505 conf->cluster_sync_low = mddev->curr_resync_completed;
3506 raid10_set_cluster_sync_high(conf);
3507 /* Send resync message */
3508 md_cluster_ops->resync_info_update(mddev,
3509 conf->cluster_sync_low,
3510 conf->cluster_sync_high);
3512 } else if (mddev_is_clustered(mddev)) {
3513 /* This is recovery not resync */
3514 sector_t sect_va1, sect_va2;
3515 bool broadcast_msg = false;
3517 for (i = 0; i < conf->geo.raid_disks; i++) {
3519 * sector_nr is a device address for recovery, so we
3520 * need translate it to array address before compare
3521 * with cluster_sync_high.
3523 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3525 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3526 broadcast_msg = true;
3528 * curr_resync_completed is similar as
3529 * sector_nr, so make the translation too.
3531 sect_va2 = raid10_find_virt(conf,
3532 mddev->curr_resync_completed, i);
3534 if (conf->cluster_sync_low == 0 ||
3535 conf->cluster_sync_low > sect_va2)
3536 conf->cluster_sync_low = sect_va2;
3539 if (broadcast_msg) {
3540 raid10_set_cluster_sync_high(conf);
3541 md_cluster_ops->resync_info_update(mddev,
3542 conf->cluster_sync_low,
3543 conf->cluster_sync_high);
3549 biolist = biolist->bi_next;
3551 bio->bi_next = NULL;
3552 r10_bio = get_resync_r10bio(bio);
3553 r10_bio->sectors = nr_sectors;
3555 if (bio->bi_end_io == end_sync_read) {
3556 md_sync_acct_bio(bio, nr_sectors);
3558 generic_make_request(bio);
3562 if (sectors_skipped)
3563 /* pretend they weren't skipped, it makes
3564 * no important difference in this case
3566 md_done_sync(mddev, sectors_skipped, 1);
3568 return sectors_skipped + nr_sectors;
3570 /* There is nowhere to write, so all non-sync
3571 * drives must be failed or in resync, all drives
3572 * have a bad block, so try the next chunk...
3574 if (sector_nr + max_sync < max_sector)
3575 max_sector = sector_nr + max_sync;
3577 sectors_skipped += (max_sector - sector_nr);
3579 sector_nr = max_sector;
3584 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3587 struct r10conf *conf = mddev->private;
3590 raid_disks = min(conf->geo.raid_disks,
3591 conf->prev.raid_disks);
3593 sectors = conf->dev_sectors;
3595 size = sectors >> conf->geo.chunk_shift;
3596 sector_div(size, conf->geo.far_copies);
3597 size = size * raid_disks;
3598 sector_div(size, conf->geo.near_copies);
3600 return size << conf->geo.chunk_shift;
3603 static void calc_sectors(struct r10conf *conf, sector_t size)
3605 /* Calculate the number of sectors-per-device that will
3606 * actually be used, and set conf->dev_sectors and
3610 size = size >> conf->geo.chunk_shift;
3611 sector_div(size, conf->geo.far_copies);
3612 size = size * conf->geo.raid_disks;
3613 sector_div(size, conf->geo.near_copies);
3614 /* 'size' is now the number of chunks in the array */
3615 /* calculate "used chunks per device" */
3616 size = size * conf->copies;
3618 /* We need to round up when dividing by raid_disks to
3619 * get the stride size.
3621 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3623 conf->dev_sectors = size << conf->geo.chunk_shift;
3625 if (conf->geo.far_offset)
3626 conf->geo.stride = 1 << conf->geo.chunk_shift;
3628 sector_div(size, conf->geo.far_copies);
3629 conf->geo.stride = size << conf->geo.chunk_shift;
3633 enum geo_type {geo_new, geo_old, geo_start};
3634 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3637 int layout, chunk, disks;
3640 layout = mddev->layout;
3641 chunk = mddev->chunk_sectors;
3642 disks = mddev->raid_disks - mddev->delta_disks;
3645 layout = mddev->new_layout;
3646 chunk = mddev->new_chunk_sectors;
3647 disks = mddev->raid_disks;
3649 default: /* avoid 'may be unused' warnings */
3650 case geo_start: /* new when starting reshape - raid_disks not
3652 layout = mddev->new_layout;
3653 chunk = mddev->new_chunk_sectors;
3654 disks = mddev->raid_disks + mddev->delta_disks;
3659 if (chunk < (PAGE_SIZE >> 9) ||
3660 !is_power_of_2(chunk))
3663 fc = (layout >> 8) & 255;
3664 fo = layout & (1<<16);
3665 geo->raid_disks = disks;
3666 geo->near_copies = nc;
3667 geo->far_copies = fc;
3668 geo->far_offset = fo;
3669 switch (layout >> 17) {
3670 case 0: /* original layout. simple but not always optimal */
3671 geo->far_set_size = disks;
3673 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3674 * actually using this, but leave code here just in case.*/
3675 geo->far_set_size = disks/fc;
3676 WARN(geo->far_set_size < fc,
3677 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3679 case 2: /* "improved" layout fixed to match documentation */
3680 geo->far_set_size = fc * nc;
3682 default: /* Not a valid layout */
3685 geo->chunk_mask = chunk - 1;
3686 geo->chunk_shift = ffz(~chunk);
3690 static void raid10_free_conf(struct r10conf *conf)
3695 mempool_exit(&conf->r10bio_pool);
3696 kfree(conf->mirrors);
3697 kfree(conf->mirrors_old);
3698 kfree(conf->mirrors_new);
3699 safe_put_page(conf->tmppage);
3700 bioset_exit(&conf->bio_split);
3704 static struct r10conf *setup_conf(struct mddev *mddev)
3706 struct r10conf *conf = NULL;
3711 copies = setup_geo(&geo, mddev, geo_new);
3714 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3715 mdname(mddev), PAGE_SIZE);
3719 if (copies < 2 || copies > mddev->raid_disks) {
3720 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3721 mdname(mddev), mddev->new_layout);
3726 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3730 /* FIXME calc properly */
3731 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3732 sizeof(struct raid10_info),
3737 conf->tmppage = alloc_page(GFP_KERNEL);
3742 conf->copies = copies;
3743 err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3744 r10bio_pool_free, conf);
3748 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3752 calc_sectors(conf, mddev->dev_sectors);
3753 if (mddev->reshape_position == MaxSector) {
3754 conf->prev = conf->geo;
3755 conf->reshape_progress = MaxSector;
3757 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3761 conf->reshape_progress = mddev->reshape_position;
3762 if (conf->prev.far_offset)
3763 conf->prev.stride = 1 << conf->prev.chunk_shift;
3765 /* far_copies must be 1 */
3766 conf->prev.stride = conf->dev_sectors;
3768 conf->reshape_safe = conf->reshape_progress;
3769 spin_lock_init(&conf->device_lock);
3770 INIT_LIST_HEAD(&conf->retry_list);
3771 INIT_LIST_HEAD(&conf->bio_end_io_list);
3773 spin_lock_init(&conf->resync_lock);
3774 init_waitqueue_head(&conf->wait_barrier);
3775 atomic_set(&conf->nr_pending, 0);
3778 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3782 conf->mddev = mddev;
3786 raid10_free_conf(conf);
3787 return ERR_PTR(err);
3790 static void raid10_set_io_opt(struct r10conf *conf)
3792 int raid_disks = conf->geo.raid_disks;
3794 if (!(conf->geo.raid_disks % conf->geo.near_copies))
3795 raid_disks /= conf->geo.near_copies;
3796 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
3800 static int raid10_run(struct mddev *mddev)
3802 struct r10conf *conf;
3804 struct raid10_info *disk;
3805 struct md_rdev *rdev;
3807 sector_t min_offset_diff = 0;
3809 bool discard_supported = false;
3811 if (mddev_init_writes_pending(mddev) < 0)
3814 if (mddev->private == NULL) {
3815 conf = setup_conf(mddev);
3817 return PTR_ERR(conf);
3818 mddev->private = conf;
3820 conf = mddev->private;
3824 mddev->thread = conf->thread;
3825 conf->thread = NULL;
3827 if (mddev_is_clustered(conf->mddev)) {
3830 fc = (mddev->layout >> 8) & 255;
3831 fo = mddev->layout & (1<<16);
3832 if (fc > 1 || fo > 0) {
3833 pr_err("only near layout is supported by clustered"
3840 blk_queue_max_discard_sectors(mddev->queue,
3841 mddev->chunk_sectors);
3842 blk_queue_max_write_same_sectors(mddev->queue, 0);
3843 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3844 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
3845 raid10_set_io_opt(conf);
3848 rdev_for_each(rdev, mddev) {
3851 disk_idx = rdev->raid_disk;
3854 if (disk_idx >= conf->geo.raid_disks &&
3855 disk_idx >= conf->prev.raid_disks)
3857 disk = conf->mirrors + disk_idx;
3859 if (test_bit(Replacement, &rdev->flags)) {
3860 if (disk->replacement)
3862 disk->replacement = rdev;
3868 diff = (rdev->new_data_offset - rdev->data_offset);
3869 if (!mddev->reshape_backwards)
3873 if (first || diff < min_offset_diff)
3874 min_offset_diff = diff;
3877 disk_stack_limits(mddev->gendisk, rdev->bdev,
3878 rdev->data_offset << 9);
3880 disk->head_position = 0;
3882 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3883 discard_supported = true;
3888 if (discard_supported)
3889 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3892 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3895 /* need to check that every block has at least one working mirror */
3896 if (!enough(conf, -1)) {
3897 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3902 if (conf->reshape_progress != MaxSector) {
3903 /* must ensure that shape change is supported */
3904 if (conf->geo.far_copies != 1 &&
3905 conf->geo.far_offset == 0)
3907 if (conf->prev.far_copies != 1 &&
3908 conf->prev.far_offset == 0)
3912 mddev->degraded = 0;
3914 i < conf->geo.raid_disks
3915 || i < conf->prev.raid_disks;
3918 disk = conf->mirrors + i;
3920 if (!disk->rdev && disk->replacement) {
3921 /* The replacement is all we have - use it */
3922 disk->rdev = disk->replacement;
3923 disk->replacement = NULL;
3924 clear_bit(Replacement, &disk->rdev->flags);
3928 !test_bit(In_sync, &disk->rdev->flags)) {
3929 disk->head_position = 0;
3932 disk->rdev->saved_raid_disk < 0)
3936 if (disk->replacement &&
3937 !test_bit(In_sync, &disk->replacement->flags) &&
3938 disk->replacement->saved_raid_disk < 0) {
3942 disk->recovery_disabled = mddev->recovery_disabled - 1;
3945 if (mddev->recovery_cp != MaxSector)
3946 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3948 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3949 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3950 conf->geo.raid_disks);
3952 * Ok, everything is just fine now
3954 mddev->dev_sectors = conf->dev_sectors;
3955 size = raid10_size(mddev, 0, 0);
3956 md_set_array_sectors(mddev, size);
3957 mddev->resync_max_sectors = size;
3958 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3961 int stripe = conf->geo.raid_disks *
3962 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3964 /* Calculate max read-ahead size.
3965 * We need to readahead at least twice a whole stripe....
3968 stripe /= conf->geo.near_copies;
3969 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3970 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3973 if (md_integrity_register(mddev))
3976 if (conf->reshape_progress != MaxSector) {
3977 unsigned long before_length, after_length;
3979 before_length = ((1 << conf->prev.chunk_shift) *
3980 conf->prev.far_copies);
3981 after_length = ((1 << conf->geo.chunk_shift) *
3982 conf->geo.far_copies);
3984 if (max(before_length, after_length) > min_offset_diff) {
3985 /* This cannot work */
3986 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3989 conf->offset_diff = min_offset_diff;
3991 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3992 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3993 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3994 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3995 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3997 if (!mddev->sync_thread)
4004 md_unregister_thread(&mddev->thread);
4005 raid10_free_conf(conf);
4006 mddev->private = NULL;
4011 static void raid10_free(struct mddev *mddev, void *priv)
4013 raid10_free_conf(priv);
4016 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4018 struct r10conf *conf = mddev->private;
4021 raise_barrier(conf, 0);
4023 lower_barrier(conf);
4026 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4028 /* Resize of 'far' arrays is not supported.
4029 * For 'near' and 'offset' arrays we can set the
4030 * number of sectors used to be an appropriate multiple
4031 * of the chunk size.
4032 * For 'offset', this is far_copies*chunksize.
4033 * For 'near' the multiplier is the LCM of
4034 * near_copies and raid_disks.
4035 * So if far_copies > 1 && !far_offset, fail.
4036 * Else find LCM(raid_disks, near_copy)*far_copies and
4037 * multiply by chunk_size. Then round to this number.
4038 * This is mostly done by raid10_size()
4040 struct r10conf *conf = mddev->private;
4041 sector_t oldsize, size;
4043 if (mddev->reshape_position != MaxSector)
4046 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4049 oldsize = raid10_size(mddev, 0, 0);
4050 size = raid10_size(mddev, sectors, 0);
4051 if (mddev->external_size &&
4052 mddev->array_sectors > size)
4054 if (mddev->bitmap) {
4055 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4059 md_set_array_sectors(mddev, size);
4060 if (sectors > mddev->dev_sectors &&
4061 mddev->recovery_cp > oldsize) {
4062 mddev->recovery_cp = oldsize;
4063 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4065 calc_sectors(conf, sectors);
4066 mddev->dev_sectors = conf->dev_sectors;
4067 mddev->resync_max_sectors = size;
4071 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4073 struct md_rdev *rdev;
4074 struct r10conf *conf;
4076 if (mddev->degraded > 0) {
4077 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4079 return ERR_PTR(-EINVAL);
4081 sector_div(size, devs);
4083 /* Set new parameters */
4084 mddev->new_level = 10;
4085 /* new layout: far_copies = 1, near_copies = 2 */
4086 mddev->new_layout = (1<<8) + 2;
4087 mddev->new_chunk_sectors = mddev->chunk_sectors;
4088 mddev->delta_disks = mddev->raid_disks;
4089 mddev->raid_disks *= 2;
4090 /* make sure it will be not marked as dirty */
4091 mddev->recovery_cp = MaxSector;
4092 mddev->dev_sectors = size;
4094 conf = setup_conf(mddev);
4095 if (!IS_ERR(conf)) {
4096 rdev_for_each(rdev, mddev)
4097 if (rdev->raid_disk >= 0) {
4098 rdev->new_raid_disk = rdev->raid_disk * 2;
4099 rdev->sectors = size;
4107 static void *raid10_takeover(struct mddev *mddev)
4109 struct r0conf *raid0_conf;
4111 /* raid10 can take over:
4112 * raid0 - providing it has only two drives
4114 if (mddev->level == 0) {
4115 /* for raid0 takeover only one zone is supported */
4116 raid0_conf = mddev->private;
4117 if (raid0_conf->nr_strip_zones > 1) {
4118 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4120 return ERR_PTR(-EINVAL);
4122 return raid10_takeover_raid0(mddev,
4123 raid0_conf->strip_zone->zone_end,
4124 raid0_conf->strip_zone->nb_dev);
4126 return ERR_PTR(-EINVAL);
4129 static int raid10_check_reshape(struct mddev *mddev)
4131 /* Called when there is a request to change
4132 * - layout (to ->new_layout)
4133 * - chunk size (to ->new_chunk_sectors)
4134 * - raid_disks (by delta_disks)
4135 * or when trying to restart a reshape that was ongoing.
4137 * We need to validate the request and possibly allocate
4138 * space if that might be an issue later.
4140 * Currently we reject any reshape of a 'far' mode array,
4141 * allow chunk size to change if new is generally acceptable,
4142 * allow raid_disks to increase, and allow
4143 * a switch between 'near' mode and 'offset' mode.
4145 struct r10conf *conf = mddev->private;
4148 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4151 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4152 /* mustn't change number of copies */
4154 if (geo.far_copies > 1 && !geo.far_offset)
4155 /* Cannot switch to 'far' mode */
4158 if (mddev->array_sectors & geo.chunk_mask)
4159 /* not factor of array size */
4162 if (!enough(conf, -1))
4165 kfree(conf->mirrors_new);
4166 conf->mirrors_new = NULL;
4167 if (mddev->delta_disks > 0) {
4168 /* allocate new 'mirrors' list */
4170 kcalloc(mddev->raid_disks + mddev->delta_disks,
4171 sizeof(struct raid10_info),
4173 if (!conf->mirrors_new)
4180 * Need to check if array has failed when deciding whether to:
4182 * - remove non-faulty devices
4185 * This determination is simple when no reshape is happening.
4186 * However if there is a reshape, we need to carefully check
4187 * both the before and after sections.
4188 * This is because some failed devices may only affect one
4189 * of the two sections, and some non-in_sync devices may
4190 * be insync in the section most affected by failed devices.
4192 static int calc_degraded(struct r10conf *conf)
4194 int degraded, degraded2;
4199 /* 'prev' section first */
4200 for (i = 0; i < conf->prev.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 /* When we can reduce the number of devices in
4206 * an array, this might not contribute to
4207 * 'degraded'. It does now.
4212 if (conf->geo.raid_disks == conf->prev.raid_disks)
4216 for (i = 0; i < conf->geo.raid_disks; i++) {
4217 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4218 if (!rdev || test_bit(Faulty, &rdev->flags))
4220 else if (!test_bit(In_sync, &rdev->flags)) {
4221 /* If reshape is increasing the number of devices,
4222 * this section has already been recovered, so
4223 * it doesn't contribute to degraded.
4226 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4231 if (degraded2 > degraded)
4236 static int raid10_start_reshape(struct mddev *mddev)
4238 /* A 'reshape' has been requested. This commits
4239 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4240 * This also checks if there are enough spares and adds them
4242 * We currently require enough spares to make the final
4243 * array non-degraded. We also require that the difference
4244 * between old and new data_offset - on each device - is
4245 * enough that we never risk over-writing.
4248 unsigned long before_length, after_length;
4249 sector_t min_offset_diff = 0;
4252 struct r10conf *conf = mddev->private;
4253 struct md_rdev *rdev;
4257 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4260 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4263 before_length = ((1 << conf->prev.chunk_shift) *
4264 conf->prev.far_copies);
4265 after_length = ((1 << conf->geo.chunk_shift) *
4266 conf->geo.far_copies);
4268 rdev_for_each(rdev, mddev) {
4269 if (!test_bit(In_sync, &rdev->flags)
4270 && !test_bit(Faulty, &rdev->flags))
4272 if (rdev->raid_disk >= 0) {
4273 long long diff = (rdev->new_data_offset
4274 - rdev->data_offset);
4275 if (!mddev->reshape_backwards)
4279 if (first || diff < min_offset_diff)
4280 min_offset_diff = diff;
4285 if (max(before_length, after_length) > min_offset_diff)
4288 if (spares < mddev->delta_disks)
4291 conf->offset_diff = min_offset_diff;
4292 spin_lock_irq(&conf->device_lock);
4293 if (conf->mirrors_new) {
4294 memcpy(conf->mirrors_new, conf->mirrors,
4295 sizeof(struct raid10_info)*conf->prev.raid_disks);
4297 kfree(conf->mirrors_old);
4298 conf->mirrors_old = conf->mirrors;
4299 conf->mirrors = conf->mirrors_new;
4300 conf->mirrors_new = NULL;
4302 setup_geo(&conf->geo, mddev, geo_start);
4304 if (mddev->reshape_backwards) {
4305 sector_t size = raid10_size(mddev, 0, 0);
4306 if (size < mddev->array_sectors) {
4307 spin_unlock_irq(&conf->device_lock);
4308 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4312 mddev->resync_max_sectors = size;
4313 conf->reshape_progress = size;
4315 conf->reshape_progress = 0;
4316 conf->reshape_safe = conf->reshape_progress;
4317 spin_unlock_irq(&conf->device_lock);
4319 if (mddev->delta_disks && mddev->bitmap) {
4320 ret = md_bitmap_resize(mddev->bitmap,
4321 raid10_size(mddev, 0, conf->geo.raid_disks),
4326 if (mddev->delta_disks > 0) {
4327 rdev_for_each(rdev, mddev)
4328 if (rdev->raid_disk < 0 &&
4329 !test_bit(Faulty, &rdev->flags)) {
4330 if (raid10_add_disk(mddev, rdev) == 0) {
4331 if (rdev->raid_disk >=
4332 conf->prev.raid_disks)
4333 set_bit(In_sync, &rdev->flags);
4335 rdev->recovery_offset = 0;
4337 if (sysfs_link_rdev(mddev, rdev))
4338 /* Failure here is OK */;
4340 } else if (rdev->raid_disk >= conf->prev.raid_disks
4341 && !test_bit(Faulty, &rdev->flags)) {
4342 /* This is a spare that was manually added */
4343 set_bit(In_sync, &rdev->flags);
4346 /* When a reshape changes the number of devices,
4347 * ->degraded is measured against the larger of the
4348 * pre and post numbers.
4350 spin_lock_irq(&conf->device_lock);
4351 mddev->degraded = calc_degraded(conf);
4352 spin_unlock_irq(&conf->device_lock);
4353 mddev->raid_disks = conf->geo.raid_disks;
4354 mddev->reshape_position = conf->reshape_progress;
4355 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4357 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4358 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4359 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4360 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4361 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4363 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4365 if (!mddev->sync_thread) {
4369 conf->reshape_checkpoint = jiffies;
4370 md_wakeup_thread(mddev->sync_thread);
4371 md_new_event(mddev);
4375 mddev->recovery = 0;
4376 spin_lock_irq(&conf->device_lock);
4377 conf->geo = conf->prev;
4378 mddev->raid_disks = conf->geo.raid_disks;
4379 rdev_for_each(rdev, mddev)
4380 rdev->new_data_offset = rdev->data_offset;
4382 conf->reshape_progress = MaxSector;
4383 conf->reshape_safe = MaxSector;
4384 mddev->reshape_position = MaxSector;
4385 spin_unlock_irq(&conf->device_lock);
4389 /* Calculate the last device-address that could contain
4390 * any block from the chunk that includes the array-address 's'
4391 * and report the next address.
4392 * i.e. the address returned will be chunk-aligned and after
4393 * any data that is in the chunk containing 's'.
4395 static sector_t last_dev_address(sector_t s, struct geom *geo)
4397 s = (s | geo->chunk_mask) + 1;
4398 s >>= geo->chunk_shift;
4399 s *= geo->near_copies;
4400 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4401 s *= geo->far_copies;
4402 s <<= geo->chunk_shift;
4406 /* Calculate the first device-address that could contain
4407 * any block from the chunk that includes the array-address 's'.
4408 * This too will be the start of a chunk
4410 static sector_t first_dev_address(sector_t s, struct geom *geo)
4412 s >>= geo->chunk_shift;
4413 s *= geo->near_copies;
4414 sector_div(s, geo->raid_disks);
4415 s *= geo->far_copies;
4416 s <<= geo->chunk_shift;
4420 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4423 /* We simply copy at most one chunk (smallest of old and new)
4424 * at a time, possibly less if that exceeds RESYNC_PAGES,
4425 * or we hit a bad block or something.
4426 * This might mean we pause for normal IO in the middle of
4427 * a chunk, but that is not a problem as mddev->reshape_position
4428 * can record any location.
4430 * If we will want to write to a location that isn't
4431 * yet recorded as 'safe' (i.e. in metadata on disk) then
4432 * we need to flush all reshape requests and update the metadata.
4434 * When reshaping forwards (e.g. to more devices), we interpret
4435 * 'safe' as the earliest block which might not have been copied
4436 * down yet. We divide this by previous stripe size and multiply
4437 * by previous stripe length to get lowest device offset that we
4438 * cannot write to yet.
4439 * We interpret 'sector_nr' as an address that we want to write to.
4440 * From this we use last_device_address() to find where we might
4441 * write to, and first_device_address on the 'safe' position.
4442 * If this 'next' write position is after the 'safe' position,
4443 * we must update the metadata to increase the 'safe' position.
4445 * When reshaping backwards, we round in the opposite direction
4446 * and perform the reverse test: next write position must not be
4447 * less than current safe position.
4449 * In all this the minimum difference in data offsets
4450 * (conf->offset_diff - always positive) allows a bit of slack,
4451 * so next can be after 'safe', but not by more than offset_diff
4453 * We need to prepare all the bios here before we start any IO
4454 * to ensure the size we choose is acceptable to all devices.
4455 * The means one for each copy for write-out and an extra one for
4457 * We store the read-in bio in ->master_bio and the others in
4458 * ->devs[x].bio and ->devs[x].repl_bio.
4460 struct r10conf *conf = mddev->private;
4461 struct r10bio *r10_bio;
4462 sector_t next, safe, last;
4466 struct md_rdev *rdev;
4469 struct bio *bio, *read_bio;
4470 int sectors_done = 0;
4471 struct page **pages;
4473 if (sector_nr == 0) {
4474 /* If restarting in the middle, skip the initial sectors */
4475 if (mddev->reshape_backwards &&
4476 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4477 sector_nr = (raid10_size(mddev, 0, 0)
4478 - conf->reshape_progress);
4479 } else if (!mddev->reshape_backwards &&
4480 conf->reshape_progress > 0)
4481 sector_nr = conf->reshape_progress;
4483 mddev->curr_resync_completed = sector_nr;
4484 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4490 /* We don't use sector_nr to track where we are up to
4491 * as that doesn't work well for ->reshape_backwards.
4492 * So just use ->reshape_progress.
4494 if (mddev->reshape_backwards) {
4495 /* 'next' is the earliest device address that we might
4496 * write to for this chunk in the new layout
4498 next = first_dev_address(conf->reshape_progress - 1,
4501 /* 'safe' is the last device address that we might read from
4502 * in the old layout after a restart
4504 safe = last_dev_address(conf->reshape_safe - 1,
4507 if (next + conf->offset_diff < safe)
4510 last = conf->reshape_progress - 1;
4511 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4512 & conf->prev.chunk_mask);
4513 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4514 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4516 /* 'next' is after the last device address that we
4517 * might write to for this chunk in the new layout
4519 next = last_dev_address(conf->reshape_progress, &conf->geo);
4521 /* 'safe' is the earliest device address that we might
4522 * read from in the old layout after a restart
4524 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4526 /* Need to update metadata if 'next' might be beyond 'safe'
4527 * as that would possibly corrupt data
4529 if (next > safe + conf->offset_diff)
4532 sector_nr = conf->reshape_progress;
4533 last = sector_nr | (conf->geo.chunk_mask
4534 & conf->prev.chunk_mask);
4536 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4537 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4541 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4542 /* Need to update reshape_position in metadata */
4544 mddev->reshape_position = conf->reshape_progress;
4545 if (mddev->reshape_backwards)
4546 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4547 - conf->reshape_progress;
4549 mddev->curr_resync_completed = conf->reshape_progress;
4550 conf->reshape_checkpoint = jiffies;
4551 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4552 md_wakeup_thread(mddev->thread);
4553 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4554 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4555 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4556 allow_barrier(conf);
4557 return sectors_done;
4559 conf->reshape_safe = mddev->reshape_position;
4560 allow_barrier(conf);
4563 raise_barrier(conf, 0);
4565 /* Now schedule reads for blocks from sector_nr to last */
4566 r10_bio = raid10_alloc_init_r10buf(conf);
4568 raise_barrier(conf, 1);
4569 atomic_set(&r10_bio->remaining, 0);
4570 r10_bio->mddev = mddev;
4571 r10_bio->sector = sector_nr;
4572 set_bit(R10BIO_IsReshape, &r10_bio->state);
4573 r10_bio->sectors = last - sector_nr + 1;
4574 rdev = read_balance(conf, r10_bio, &max_sectors);
4575 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4578 /* Cannot read from here, so need to record bad blocks
4579 * on all the target devices.
4582 mempool_free(r10_bio, &conf->r10buf_pool);
4583 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4584 return sectors_done;
4587 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4589 bio_set_dev(read_bio, rdev->bdev);
4590 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4591 + rdev->data_offset);
4592 read_bio->bi_private = r10_bio;
4593 read_bio->bi_end_io = end_reshape_read;
4594 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4595 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4596 read_bio->bi_status = 0;
4597 read_bio->bi_vcnt = 0;
4598 read_bio->bi_iter.bi_size = 0;
4599 r10_bio->master_bio = read_bio;
4600 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4602 /* Now find the locations in the new layout */
4603 __raid10_find_phys(&conf->geo, r10_bio);
4606 read_bio->bi_next = NULL;
4609 for (s = 0; s < conf->copies*2; s++) {
4611 int d = r10_bio->devs[s/2].devnum;
4612 struct md_rdev *rdev2;
4614 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4615 b = r10_bio->devs[s/2].repl_bio;
4617 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4618 b = r10_bio->devs[s/2].bio;
4620 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4623 bio_set_dev(b, rdev2->bdev);
4624 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4625 rdev2->new_data_offset;
4626 b->bi_end_io = end_reshape_write;
4627 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4632 /* Now add as many pages as possible to all of these bios. */
4635 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4636 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4637 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4638 int len = (max_sectors - s) << 9;
4639 if (len > PAGE_SIZE)
4641 for (bio = blist; bio ; bio = bio->bi_next) {
4643 * won't fail because the vec table is big enough
4644 * to hold all these pages
4646 bio_add_page(bio, page, len, 0);
4648 sector_nr += len >> 9;
4649 nr_sectors += len >> 9;
4652 r10_bio->sectors = nr_sectors;
4654 /* Now submit the read */
4655 md_sync_acct_bio(read_bio, r10_bio->sectors);
4656 atomic_inc(&r10_bio->remaining);
4657 read_bio->bi_next = NULL;
4658 generic_make_request(read_bio);
4659 sectors_done += nr_sectors;
4660 if (sector_nr <= last)
4663 lower_barrier(conf);
4665 /* Now that we have done the whole section we can
4666 * update reshape_progress
4668 if (mddev->reshape_backwards)
4669 conf->reshape_progress -= sectors_done;
4671 conf->reshape_progress += sectors_done;
4673 return sectors_done;
4676 static void end_reshape_request(struct r10bio *r10_bio);
4677 static int handle_reshape_read_error(struct mddev *mddev,
4678 struct r10bio *r10_bio);
4679 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4681 /* Reshape read completed. Hopefully we have a block
4683 * If we got a read error then we do sync 1-page reads from
4684 * elsewhere until we find the data - or give up.
4686 struct r10conf *conf = mddev->private;
4689 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4690 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4691 /* Reshape has been aborted */
4692 md_done_sync(mddev, r10_bio->sectors, 0);
4696 /* We definitely have the data in the pages, schedule the
4699 atomic_set(&r10_bio->remaining, 1);
4700 for (s = 0; s < conf->copies*2; s++) {
4702 int d = r10_bio->devs[s/2].devnum;
4703 struct md_rdev *rdev;
4706 rdev = rcu_dereference(conf->mirrors[d].replacement);
4707 b = r10_bio->devs[s/2].repl_bio;
4709 rdev = rcu_dereference(conf->mirrors[d].rdev);
4710 b = r10_bio->devs[s/2].bio;
4712 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4716 atomic_inc(&rdev->nr_pending);
4718 md_sync_acct_bio(b, r10_bio->sectors);
4719 atomic_inc(&r10_bio->remaining);
4721 generic_make_request(b);
4723 end_reshape_request(r10_bio);
4726 static void end_reshape(struct r10conf *conf)
4728 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4731 spin_lock_irq(&conf->device_lock);
4732 conf->prev = conf->geo;
4733 md_finish_reshape(conf->mddev);
4735 conf->reshape_progress = MaxSector;
4736 conf->reshape_safe = MaxSector;
4737 spin_unlock_irq(&conf->device_lock);
4739 /* read-ahead size must cover two whole stripes, which is
4740 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4742 if (conf->mddev->queue) {
4743 int stripe = conf->geo.raid_disks *
4744 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4745 stripe /= conf->geo.near_copies;
4746 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4747 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4748 raid10_set_io_opt(conf);
4753 static int handle_reshape_read_error(struct mddev *mddev,
4754 struct r10bio *r10_bio)
4756 /* Use sync reads to get the blocks from somewhere else */
4757 int sectors = r10_bio->sectors;
4758 struct r10conf *conf = mddev->private;
4759 struct r10bio *r10b;
4762 struct page **pages;
4764 r10b = kmalloc(sizeof(*r10b) +
4765 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4767 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4771 /* reshape IOs share pages from .devs[0].bio */
4772 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4774 r10b->sector = r10_bio->sector;
4775 __raid10_find_phys(&conf->prev, r10b);
4780 int first_slot = slot;
4782 if (s > (PAGE_SIZE >> 9))
4787 int d = r10b->devs[slot].devnum;
4788 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4791 test_bit(Faulty, &rdev->flags) ||
4792 !test_bit(In_sync, &rdev->flags))
4795 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4796 atomic_inc(&rdev->nr_pending);
4798 success = sync_page_io(rdev,
4802 REQ_OP_READ, 0, false);
4803 rdev_dec_pending(rdev, mddev);
4809 if (slot >= conf->copies)
4811 if (slot == first_slot)
4816 /* couldn't read this block, must give up */
4817 set_bit(MD_RECOVERY_INTR,
4829 static void end_reshape_write(struct bio *bio)
4831 struct r10bio *r10_bio = get_resync_r10bio(bio);
4832 struct mddev *mddev = r10_bio->mddev;
4833 struct r10conf *conf = mddev->private;
4837 struct md_rdev *rdev = NULL;
4839 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4841 rdev = conf->mirrors[d].replacement;
4844 rdev = conf->mirrors[d].rdev;
4847 if (bio->bi_status) {
4848 /* FIXME should record badblock */
4849 md_error(mddev, rdev);
4852 rdev_dec_pending(rdev, mddev);
4853 end_reshape_request(r10_bio);
4856 static void end_reshape_request(struct r10bio *r10_bio)
4858 if (!atomic_dec_and_test(&r10_bio->remaining))
4860 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4861 bio_put(r10_bio->master_bio);
4865 static void raid10_finish_reshape(struct mddev *mddev)
4867 struct r10conf *conf = mddev->private;
4869 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4872 if (mddev->delta_disks > 0) {
4873 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4874 mddev->recovery_cp = mddev->resync_max_sectors;
4875 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4877 mddev->resync_max_sectors = mddev->array_sectors;
4881 for (d = conf->geo.raid_disks ;
4882 d < conf->geo.raid_disks - mddev->delta_disks;
4884 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4886 clear_bit(In_sync, &rdev->flags);
4887 rdev = rcu_dereference(conf->mirrors[d].replacement);
4889 clear_bit(In_sync, &rdev->flags);
4893 mddev->layout = mddev->new_layout;
4894 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4895 mddev->reshape_position = MaxSector;
4896 mddev->delta_disks = 0;
4897 mddev->reshape_backwards = 0;
4900 static struct md_personality raid10_personality =
4904 .owner = THIS_MODULE,
4905 .make_request = raid10_make_request,
4907 .free = raid10_free,
4908 .status = raid10_status,
4909 .error_handler = raid10_error,
4910 .hot_add_disk = raid10_add_disk,
4911 .hot_remove_disk= raid10_remove_disk,
4912 .spare_active = raid10_spare_active,
4913 .sync_request = raid10_sync_request,
4914 .quiesce = raid10_quiesce,
4915 .size = raid10_size,
4916 .resize = raid10_resize,
4917 .takeover = raid10_takeover,
4918 .check_reshape = raid10_check_reshape,
4919 .start_reshape = raid10_start_reshape,
4920 .finish_reshape = raid10_finish_reshape,
4921 .congested = raid10_congested,
4924 static int __init raid_init(void)
4926 return register_md_personality(&raid10_personality);
4929 static void raid_exit(void)
4931 unregister_md_personality(&raid10_personality);
4934 module_init(raid_init);
4935 module_exit(raid_exit);
4936 MODULE_LICENSE("GPL");
4937 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4938 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4939 MODULE_ALIAS("md-raid10");
4940 MODULE_ALIAS("md-level-10");
4942 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);