1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid10.c : Multiple Devices driver for Linux
5 * Copyright (C) 2000-2004 Neil Brown
7 * RAID-10 support for md.
9 * Base on code in raid1.c. See raid1.c for further copyright information.
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
24 #include "md-bitmap.h"
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69 static int _enough(struct r10conf *conf, int previous, int ignore);
70 static int enough(struct r10conf *conf, int ignore);
71 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
73 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74 static void end_reshape_write(struct bio *bio);
75 static void end_reshape(struct r10conf *conf);
77 #define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
86 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
88 return get_resync_pages(bio)->raid_bio;
91 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->copies]);
96 /* allocate a r10bio with room for raid_disks entries in the
98 return kzalloc(size, gfp_flags);
101 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102 /* amount of memory to reserve for resync requests */
103 #define RESYNC_WINDOW (1024*1024)
104 /* maximum number of concurrent requests, memory permitting */
105 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
116 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
133 nalloc = 2; /* recovery */
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
142 goto out_free_r10bio;
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
151 r10_bio->devs[j].bio = bio;
152 if (!conf->have_replacement)
154 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
157 r10_bio->devs[j].repl_bio = bio;
160 * Allocate RESYNC_PAGES data pages and attach them
163 for (j = 0; j < nalloc; j++) {
164 struct bio *rbio = r10_bio->devs[j].repl_bio;
165 struct resync_pages *rp, *rp_repl;
169 rp_repl = &rps[nalloc + j];
171 bio = r10_bio->devs[j].bio;
173 if (!j || test_bit(MD_RECOVERY_SYNC,
174 &conf->mddev->recovery)) {
175 if (resync_alloc_pages(rp, gfp_flags))
178 memcpy(rp, &rps[0], sizeof(*rp));
179 resync_get_all_pages(rp);
182 rp->raid_bio = r10_bio;
183 bio->bi_private = rp;
185 memcpy(rp_repl, rp, sizeof(*rp));
186 rbio->bi_private = rp_repl;
194 resync_free_pages(&rps[j * 2]);
198 for ( ; j < nalloc; j++) {
199 if (r10_bio->devs[j].bio)
200 bio_put(r10_bio->devs[j].bio);
201 if (r10_bio->devs[j].repl_bio)
202 bio_put(r10_bio->devs[j].repl_bio);
206 rbio_pool_free(r10_bio, conf);
210 static void r10buf_pool_free(void *__r10_bio, void *data)
212 struct r10conf *conf = data;
213 struct r10bio *r10bio = __r10_bio;
215 struct resync_pages *rp = NULL;
217 for (j = conf->copies; j--; ) {
218 struct bio *bio = r10bio->devs[j].bio;
221 rp = get_resync_pages(bio);
222 resync_free_pages(rp);
226 bio = r10bio->devs[j].repl_bio;
231 /* resync pages array stored in the 1st bio's .bi_private */
234 rbio_pool_free(r10bio, conf);
237 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241 for (i = 0; i < conf->copies; i++) {
242 struct bio **bio = & r10_bio->devs[i].bio;
243 if (!BIO_SPECIAL(*bio))
246 bio = &r10_bio->devs[i].repl_bio;
247 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
253 static void free_r10bio(struct r10bio *r10_bio)
255 struct r10conf *conf = r10_bio->mddev->private;
257 put_all_bios(conf, r10_bio);
258 mempool_free(r10_bio, &conf->r10bio_pool);
261 static void put_buf(struct r10bio *r10_bio)
263 struct r10conf *conf = r10_bio->mddev->private;
265 mempool_free(r10_bio, &conf->r10buf_pool);
270 static void reschedule_retry(struct r10bio *r10_bio)
273 struct mddev *mddev = r10_bio->mddev;
274 struct r10conf *conf = mddev->private;
276 spin_lock_irqsave(&conf->device_lock, flags);
277 list_add(&r10_bio->retry_list, &conf->retry_list);
279 spin_unlock_irqrestore(&conf->device_lock, flags);
281 /* wake up frozen array... */
282 wake_up(&conf->wait_barrier);
284 md_wakeup_thread(mddev->thread);
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
292 static void raid_end_bio_io(struct r10bio *r10_bio)
294 struct bio *bio = r10_bio->master_bio;
295 struct r10conf *conf = r10_bio->mddev->private;
297 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
298 bio->bi_status = BLK_STS_IOERR;
302 * Wake up any possible resync thread that waits for the device
307 free_r10bio(r10_bio);
311 * Update disk head position estimator based on IRQ completion info.
313 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
315 struct r10conf *conf = r10_bio->mddev->private;
317 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
318 r10_bio->devs[slot].addr + (r10_bio->sectors);
322 * Find the disk number which triggered given bio
324 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
325 struct bio *bio, int *slotp, int *replp)
330 for (slot = 0; slot < conf->copies; slot++) {
331 if (r10_bio->devs[slot].bio == bio)
333 if (r10_bio->devs[slot].repl_bio == bio) {
339 BUG_ON(slot == conf->copies);
340 update_head_pos(slot, r10_bio);
346 return r10_bio->devs[slot].devnum;
349 static void raid10_end_read_request(struct bio *bio)
351 int uptodate = !bio->bi_status;
352 struct r10bio *r10_bio = bio->bi_private;
354 struct md_rdev *rdev;
355 struct r10conf *conf = r10_bio->mddev->private;
357 slot = r10_bio->read_slot;
358 rdev = r10_bio->devs[slot].rdev;
360 * this branch is our 'one mirror IO has finished' event handler:
362 update_head_pos(slot, r10_bio);
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
374 set_bit(R10BIO_Uptodate, &r10_bio->state);
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
381 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
386 raid_end_bio_io(r10_bio);
387 rdev_dec_pending(rdev, conf->mddev);
390 * oops, read error - keep the refcount on the rdev
392 char b[BDEVNAME_SIZE];
393 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
395 bdevname(rdev->bdev, b),
396 (unsigned long long)r10_bio->sector);
397 set_bit(R10BIO_ReadError, &r10_bio->state);
398 reschedule_retry(r10_bio);
402 static void close_write(struct r10bio *r10_bio)
404 /* clear the bitmap if all writes complete successfully */
405 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
407 !test_bit(R10BIO_Degraded, &r10_bio->state),
409 md_write_end(r10_bio->mddev);
412 static void one_write_done(struct r10bio *r10_bio)
414 if (atomic_dec_and_test(&r10_bio->remaining)) {
415 if (test_bit(R10BIO_WriteError, &r10_bio->state))
416 reschedule_retry(r10_bio);
418 close_write(r10_bio);
419 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
420 reschedule_retry(r10_bio);
422 raid_end_bio_io(r10_bio);
427 static void raid10_end_write_request(struct bio *bio)
429 struct r10bio *r10_bio = bio->bi_private;
432 struct r10conf *conf = r10_bio->mddev->private;
434 struct md_rdev *rdev = NULL;
435 struct bio *to_put = NULL;
438 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
440 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
443 rdev = conf->mirrors[dev].replacement;
447 rdev = conf->mirrors[dev].rdev;
450 * this branch is our 'one mirror IO has finished' event handler:
452 if (bio->bi_status && !discard_error) {
454 /* Never record new bad blocks to replacement,
457 md_error(rdev->mddev, rdev);
459 set_bit(WriteErrorSeen, &rdev->flags);
460 if (!test_and_set_bit(WantReplacement, &rdev->flags))
461 set_bit(MD_RECOVERY_NEEDED,
462 &rdev->mddev->recovery);
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST)) {
467 md_error(rdev->mddev, rdev);
468 if (!test_bit(Faulty, &rdev->flags))
469 /* This is the only remaining device,
470 * We need to retry the write without
473 set_bit(R10BIO_WriteError, &r10_bio->state);
475 r10_bio->devs[slot].bio = NULL;
480 set_bit(R10BIO_WriteError, &r10_bio->state);
484 * Set R10BIO_Uptodate in our master bio, so that
485 * we will return a good error code for to the higher
486 * levels even if IO on some other mirrored buffer fails.
488 * The 'master' represents the composite IO operation to
489 * user-side. So if something waits for IO, then it will
490 * wait for the 'master' bio.
496 * Do not set R10BIO_Uptodate if the current device is
497 * rebuilding or Faulty. This is because we cannot use
498 * such device for properly reading the data back (we could
499 * potentially use it, if the current write would have felt
500 * before rdev->recovery_offset, but for simplicity we don't
503 if (test_bit(In_sync, &rdev->flags) &&
504 !test_bit(Faulty, &rdev->flags))
505 set_bit(R10BIO_Uptodate, &r10_bio->state);
507 /* Maybe we can clear some bad blocks. */
508 if (is_badblock(rdev,
509 r10_bio->devs[slot].addr,
511 &first_bad, &bad_sectors) && !discard_error) {
514 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
516 r10_bio->devs[slot].bio = IO_MADE_GOOD;
518 set_bit(R10BIO_MadeGood, &r10_bio->state);
524 * Let's see if all mirrored write operations have finished
527 one_write_done(r10_bio);
529 rdev_dec_pending(rdev, conf->mddev);
535 * RAID10 layout manager
536 * As well as the chunksize and raid_disks count, there are two
537 * parameters: near_copies and far_copies.
538 * near_copies * far_copies must be <= raid_disks.
539 * Normally one of these will be 1.
540 * If both are 1, we get raid0.
541 * If near_copies == raid_disks, we get raid1.
543 * Chunks are laid out in raid0 style with near_copies copies of the
544 * first chunk, followed by near_copies copies of the next chunk and
546 * If far_copies > 1, then after 1/far_copies of the array has been assigned
547 * as described above, we start again with a device offset of near_copies.
548 * So we effectively have another copy of the whole array further down all
549 * the drives, but with blocks on different drives.
550 * With this layout, and block is never stored twice on the one device.
552 * raid10_find_phys finds the sector offset of a given virtual sector
553 * on each device that it is on.
555 * raid10_find_virt does the reverse mapping, from a device and a
556 * sector offset to a virtual address
559 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
567 int last_far_set_start, last_far_set_size;
569 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
570 last_far_set_start *= geo->far_set_size;
572 last_far_set_size = geo->far_set_size;
573 last_far_set_size += (geo->raid_disks % geo->far_set_size);
575 /* now calculate first sector/dev */
576 chunk = r10bio->sector >> geo->chunk_shift;
577 sector = r10bio->sector & geo->chunk_mask;
579 chunk *= geo->near_copies;
581 dev = sector_div(stripe, geo->raid_disks);
583 stripe *= geo->far_copies;
585 sector += stripe << geo->chunk_shift;
587 /* and calculate all the others */
588 for (n = 0; n < geo->near_copies; n++) {
592 r10bio->devs[slot].devnum = d;
593 r10bio->devs[slot].addr = s;
596 for (f = 1; f < geo->far_copies; f++) {
597 set = d / geo->far_set_size;
598 d += geo->near_copies;
600 if ((geo->raid_disks % geo->far_set_size) &&
601 (d > last_far_set_start)) {
602 d -= last_far_set_start;
603 d %= last_far_set_size;
604 d += last_far_set_start;
606 d %= geo->far_set_size;
607 d += geo->far_set_size * set;
610 r10bio->devs[slot].devnum = d;
611 r10bio->devs[slot].addr = s;
615 if (dev >= geo->raid_disks) {
617 sector += (geo->chunk_mask + 1);
622 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
624 struct geom *geo = &conf->geo;
626 if (conf->reshape_progress != MaxSector &&
627 ((r10bio->sector >= conf->reshape_progress) !=
628 conf->mddev->reshape_backwards)) {
629 set_bit(R10BIO_Previous, &r10bio->state);
632 clear_bit(R10BIO_Previous, &r10bio->state);
634 __raid10_find_phys(geo, r10bio);
637 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
639 sector_t offset, chunk, vchunk;
640 /* Never use conf->prev as this is only called during resync
641 * or recovery, so reshape isn't happening
643 struct geom *geo = &conf->geo;
644 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
645 int far_set_size = geo->far_set_size;
646 int last_far_set_start;
648 if (geo->raid_disks % geo->far_set_size) {
649 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
650 last_far_set_start *= geo->far_set_size;
652 if (dev >= last_far_set_start) {
653 far_set_size = geo->far_set_size;
654 far_set_size += (geo->raid_disks % geo->far_set_size);
655 far_set_start = last_far_set_start;
659 offset = sector & geo->chunk_mask;
660 if (geo->far_offset) {
662 chunk = sector >> geo->chunk_shift;
663 fc = sector_div(chunk, geo->far_copies);
664 dev -= fc * geo->near_copies;
665 if (dev < far_set_start)
668 while (sector >= geo->stride) {
669 sector -= geo->stride;
670 if (dev < (geo->near_copies + far_set_start))
671 dev += far_set_size - geo->near_copies;
673 dev -= geo->near_copies;
675 chunk = sector >> geo->chunk_shift;
677 vchunk = chunk * geo->raid_disks + dev;
678 sector_div(vchunk, geo->near_copies);
679 return (vchunk << geo->chunk_shift) + offset;
683 * This routine returns the disk from which the requested read should
684 * be done. There is a per-array 'next expected sequential IO' sector
685 * number - if this matches on the next IO then we use the last disk.
686 * There is also a per-disk 'last know head position' sector that is
687 * maintained from IRQ contexts, both the normal and the resync IO
688 * completion handlers update this position correctly. If there is no
689 * perfect sequential match then we pick the disk whose head is closest.
691 * If there are 2 mirrors in the same 2 devices, performance degrades
692 * because position is mirror, not device based.
694 * The rdev for the device selected will have nr_pending incremented.
698 * FIXME: possibly should rethink readbalancing and do it differently
699 * depending on near_copies / far_copies geometry.
701 static struct md_rdev *read_balance(struct r10conf *conf,
702 struct r10bio *r10_bio,
705 const sector_t this_sector = r10_bio->sector;
707 int sectors = r10_bio->sectors;
708 int best_good_sectors;
709 sector_t new_distance, best_dist;
710 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
712 int best_dist_slot, best_pending_slot;
713 bool has_nonrot_disk = false;
714 unsigned int min_pending;
715 struct geom *geo = &conf->geo;
717 raid10_find_phys(conf, r10_bio);
720 min_pending = UINT_MAX;
721 best_dist_rdev = NULL;
722 best_pending_rdev = NULL;
723 best_dist = MaxSector;
724 best_good_sectors = 0;
726 clear_bit(R10BIO_FailFast, &r10_bio->state);
728 * Check if we can balance. We can balance on the whole
729 * device if no resync is going on (recovery is ok), or below
730 * the resync window. We take the first readable disk when
731 * above the resync window.
733 if ((conf->mddev->recovery_cp < MaxSector
734 && (this_sector + sectors >= conf->next_resync)) ||
735 (mddev_is_clustered(conf->mddev) &&
736 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
737 this_sector + sectors)))
740 for (slot = 0; slot < conf->copies ; slot++) {
744 unsigned int pending;
747 if (r10_bio->devs[slot].bio == IO_BLOCKED)
749 disk = r10_bio->devs[slot].devnum;
750 rdev = rcu_dereference(conf->mirrors[disk].replacement);
751 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
752 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
753 rdev = rcu_dereference(conf->mirrors[disk].rdev);
755 test_bit(Faulty, &rdev->flags))
757 if (!test_bit(In_sync, &rdev->flags) &&
758 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
761 dev_sector = r10_bio->devs[slot].addr;
762 if (is_badblock(rdev, dev_sector, sectors,
763 &first_bad, &bad_sectors)) {
764 if (best_dist < MaxSector)
765 /* Already have a better slot */
767 if (first_bad <= dev_sector) {
768 /* Cannot read here. If this is the
769 * 'primary' device, then we must not read
770 * beyond 'bad_sectors' from another device.
772 bad_sectors -= (dev_sector - first_bad);
773 if (!do_balance && sectors > bad_sectors)
774 sectors = bad_sectors;
775 if (best_good_sectors > sectors)
776 best_good_sectors = sectors;
778 sector_t good_sectors =
779 first_bad - dev_sector;
780 if (good_sectors > best_good_sectors) {
781 best_good_sectors = good_sectors;
782 best_dist_slot = slot;
783 best_dist_rdev = rdev;
786 /* Must read from here */
791 best_good_sectors = sectors;
796 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
797 has_nonrot_disk |= nonrot;
798 pending = atomic_read(&rdev->nr_pending);
799 if (min_pending > pending && nonrot) {
800 min_pending = pending;
801 best_pending_slot = slot;
802 best_pending_rdev = rdev;
805 if (best_dist_slot >= 0)
806 /* At least 2 disks to choose from so failfast is OK */
807 set_bit(R10BIO_FailFast, &r10_bio->state);
808 /* This optimisation is debatable, and completely destroys
809 * sequential read speed for 'far copies' arrays. So only
810 * keep it for 'near' arrays, and review those later.
812 if (geo->near_copies > 1 && !pending)
815 /* for far > 1 always use the lowest address */
816 else if (geo->far_copies > 1)
817 new_distance = r10_bio->devs[slot].addr;
819 new_distance = abs(r10_bio->devs[slot].addr -
820 conf->mirrors[disk].head_position);
822 if (new_distance < best_dist) {
823 best_dist = new_distance;
824 best_dist_slot = slot;
825 best_dist_rdev = rdev;
828 if (slot >= conf->copies) {
829 if (has_nonrot_disk) {
830 slot = best_pending_slot;
831 rdev = best_pending_rdev;
833 slot = best_dist_slot;
834 rdev = best_dist_rdev;
839 atomic_inc(&rdev->nr_pending);
840 r10_bio->read_slot = slot;
844 *max_sectors = best_good_sectors;
849 static int raid10_congested(struct mddev *mddev, int bits)
851 struct r10conf *conf = mddev->private;
854 if ((bits & (1 << WB_async_congested)) &&
855 conf->pending_count >= max_queued_requests)
860 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
863 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
864 if (rdev && !test_bit(Faulty, &rdev->flags)) {
865 struct request_queue *q = bdev_get_queue(rdev->bdev);
867 ret |= bdi_congested(q->backing_dev_info, bits);
874 static void flush_pending_writes(struct r10conf *conf)
876 /* Any writes that have been queued but are awaiting
877 * bitmap updates get flushed here.
879 spin_lock_irq(&conf->device_lock);
881 if (conf->pending_bio_list.head) {
882 struct blk_plug plug;
885 bio = bio_list_get(&conf->pending_bio_list);
886 conf->pending_count = 0;
887 spin_unlock_irq(&conf->device_lock);
890 * As this is called in a wait_event() loop (see freeze_array),
891 * current->state might be TASK_UNINTERRUPTIBLE which will
892 * cause a warning when we prepare to wait again. As it is
893 * rare that this path is taken, it is perfectly safe to force
894 * us to go around the wait_event() loop again, so the warning
895 * is a false-positive. Silence the warning by resetting
898 __set_current_state(TASK_RUNNING);
900 blk_start_plug(&plug);
901 /* flush any pending bitmap writes to disk
902 * before proceeding w/ I/O */
903 md_bitmap_unplug(conf->mddev->bitmap);
904 wake_up(&conf->wait_barrier);
906 while (bio) { /* submit pending writes */
907 struct bio *next = bio->bi_next;
908 struct md_rdev *rdev = (void*)bio->bi_disk;
910 bio_set_dev(bio, rdev->bdev);
911 if (test_bit(Faulty, &rdev->flags)) {
913 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
914 !blk_queue_discard(bio->bi_disk->queue)))
918 generic_make_request(bio);
921 blk_finish_plug(&plug);
923 spin_unlock_irq(&conf->device_lock);
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
948 static void raise_barrier(struct r10conf *conf, int force)
950 BUG_ON(force && !conf->barrier);
951 spin_lock_irq(&conf->resync_lock);
953 /* Wait until no block IO is waiting (unless 'force') */
954 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
957 /* block any new IO from starting */
960 /* Now wait for all pending IO to complete */
961 wait_event_lock_irq(conf->wait_barrier,
962 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
965 spin_unlock_irq(&conf->resync_lock);
968 static void lower_barrier(struct r10conf *conf)
971 spin_lock_irqsave(&conf->resync_lock, flags);
973 spin_unlock_irqrestore(&conf->resync_lock, flags);
974 wake_up(&conf->wait_barrier);
977 static void wait_barrier(struct r10conf *conf)
979 spin_lock_irq(&conf->resync_lock);
982 /* Wait for the barrier to drop.
983 * However if there are already pending
984 * requests (preventing the barrier from
985 * rising completely), and the
986 * pre-process bio queue isn't empty,
987 * then don't wait, as we need to empty
988 * that queue to get the nr_pending
991 raid10_log(conf->mddev, "wait barrier");
992 wait_event_lock_irq(conf->wait_barrier,
994 (atomic_read(&conf->nr_pending) &&
996 (!bio_list_empty(¤t->bio_list[0]) ||
997 !bio_list_empty(¤t->bio_list[1]))),
1000 if (!conf->nr_waiting)
1001 wake_up(&conf->wait_barrier);
1003 atomic_inc(&conf->nr_pending);
1004 spin_unlock_irq(&conf->resync_lock);
1007 static void allow_barrier(struct r10conf *conf)
1009 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1010 (conf->array_freeze_pending))
1011 wake_up(&conf->wait_barrier);
1014 static void freeze_array(struct r10conf *conf, int extra)
1016 /* stop syncio and normal IO and wait for everything to
1018 * We increment barrier and nr_waiting, and then
1019 * wait until nr_pending match nr_queued+extra
1020 * This is called in the context of one normal IO request
1021 * that has failed. Thus any sync request that might be pending
1022 * will be blocked by nr_pending, and we need to wait for
1023 * pending IO requests to complete or be queued for re-try.
1024 * Thus the number queued (nr_queued) plus this request (extra)
1025 * must match the number of pending IOs (nr_pending) before
1028 spin_lock_irq(&conf->resync_lock);
1029 conf->array_freeze_pending++;
1032 wait_event_lock_irq_cmd(conf->wait_barrier,
1033 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1035 flush_pending_writes(conf));
1037 conf->array_freeze_pending--;
1038 spin_unlock_irq(&conf->resync_lock);
1041 static void unfreeze_array(struct r10conf *conf)
1043 /* reverse the effect of the freeze */
1044 spin_lock_irq(&conf->resync_lock);
1047 wake_up(&conf->wait_barrier);
1048 spin_unlock_irq(&conf->resync_lock);
1051 static sector_t choose_data_offset(struct r10bio *r10_bio,
1052 struct md_rdev *rdev)
1054 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1055 test_bit(R10BIO_Previous, &r10_bio->state))
1056 return rdev->data_offset;
1058 return rdev->new_data_offset;
1061 struct raid10_plug_cb {
1062 struct blk_plug_cb cb;
1063 struct bio_list pending;
1067 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1069 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1071 struct mddev *mddev = plug->cb.data;
1072 struct r10conf *conf = mddev->private;
1075 if (from_schedule || current->bio_list) {
1076 spin_lock_irq(&conf->device_lock);
1077 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1078 conf->pending_count += plug->pending_cnt;
1079 spin_unlock_irq(&conf->device_lock);
1080 wake_up(&conf->wait_barrier);
1081 md_wakeup_thread(mddev->thread);
1086 /* we aren't scheduling, so we can do the write-out directly. */
1087 bio = bio_list_get(&plug->pending);
1088 md_bitmap_unplug(mddev->bitmap);
1089 wake_up(&conf->wait_barrier);
1091 while (bio) { /* submit pending writes */
1092 struct bio *next = bio->bi_next;
1093 struct md_rdev *rdev = (void*)bio->bi_disk;
1094 bio->bi_next = NULL;
1095 bio_set_dev(bio, rdev->bdev);
1096 if (test_bit(Faulty, &rdev->flags)) {
1098 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1099 !blk_queue_discard(bio->bi_disk->queue)))
1100 /* Just ignore it */
1103 generic_make_request(bio);
1110 * 1. Register the new request and wait if the reconstruction thread has put
1111 * up a bar for new requests. Continue immediately if no resync is active
1113 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1115 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1116 struct bio *bio, sector_t sectors)
1119 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1120 bio->bi_iter.bi_sector < conf->reshape_progress &&
1121 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1122 raid10_log(conf->mddev, "wait reshape");
1123 allow_barrier(conf);
1124 wait_event(conf->wait_barrier,
1125 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1126 conf->reshape_progress >= bio->bi_iter.bi_sector +
1132 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1133 struct r10bio *r10_bio)
1135 struct r10conf *conf = mddev->private;
1136 struct bio *read_bio;
1137 const int op = bio_op(bio);
1138 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1140 struct md_rdev *rdev;
1141 char b[BDEVNAME_SIZE];
1142 int slot = r10_bio->read_slot;
1143 struct md_rdev *err_rdev = NULL;
1144 gfp_t gfp = GFP_NOIO;
1146 if (r10_bio->devs[slot].rdev) {
1148 * This is an error retry, but we cannot
1149 * safely dereference the rdev in the r10_bio,
1150 * we must use the one in conf.
1151 * If it has already been disconnected (unlikely)
1152 * we lose the device name in error messages.
1156 * As we are blocking raid10, it is a little safer to
1159 gfp = GFP_NOIO | __GFP_HIGH;
1162 disk = r10_bio->devs[slot].devnum;
1163 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1165 bdevname(err_rdev->bdev, b);
1168 /* This never gets dereferenced */
1169 err_rdev = r10_bio->devs[slot].rdev;
1174 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1175 rdev = read_balance(conf, r10_bio, &max_sectors);
1178 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1180 (unsigned long long)r10_bio->sector);
1182 raid_end_bio_io(r10_bio);
1186 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1188 bdevname(rdev->bdev, b),
1189 (unsigned long long)r10_bio->sector);
1190 if (max_sectors < bio_sectors(bio)) {
1191 struct bio *split = bio_split(bio, max_sectors,
1192 gfp, &conf->bio_split);
1193 bio_chain(split, bio);
1194 allow_barrier(conf);
1195 generic_make_request(bio);
1198 r10_bio->master_bio = bio;
1199 r10_bio->sectors = max_sectors;
1201 slot = r10_bio->read_slot;
1203 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1205 r10_bio->devs[slot].bio = read_bio;
1206 r10_bio->devs[slot].rdev = rdev;
1208 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1209 choose_data_offset(r10_bio, rdev);
1210 bio_set_dev(read_bio, rdev->bdev);
1211 read_bio->bi_end_io = raid10_end_read_request;
1212 bio_set_op_attrs(read_bio, op, do_sync);
1213 if (test_bit(FailFast, &rdev->flags) &&
1214 test_bit(R10BIO_FailFast, &r10_bio->state))
1215 read_bio->bi_opf |= MD_FAILFAST;
1216 read_bio->bi_private = r10_bio;
1219 trace_block_bio_remap(read_bio->bi_disk->queue,
1220 read_bio, disk_devt(mddev->gendisk),
1222 generic_make_request(read_bio);
1226 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1227 struct bio *bio, bool replacement,
1230 const int op = bio_op(bio);
1231 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1232 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1233 unsigned long flags;
1234 struct blk_plug_cb *cb;
1235 struct raid10_plug_cb *plug = NULL;
1236 struct r10conf *conf = mddev->private;
1237 struct md_rdev *rdev;
1238 int devnum = r10_bio->devs[n_copy].devnum;
1242 rdev = conf->mirrors[devnum].replacement;
1244 /* Replacement just got moved to main 'rdev' */
1246 rdev = conf->mirrors[devnum].rdev;
1249 rdev = conf->mirrors[devnum].rdev;
1251 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1253 r10_bio->devs[n_copy].repl_bio = mbio;
1255 r10_bio->devs[n_copy].bio = mbio;
1257 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1258 choose_data_offset(r10_bio, rdev));
1259 bio_set_dev(mbio, rdev->bdev);
1260 mbio->bi_end_io = raid10_end_write_request;
1261 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1262 if (!replacement && test_bit(FailFast,
1263 &conf->mirrors[devnum].rdev->flags)
1264 && enough(conf, devnum))
1265 mbio->bi_opf |= MD_FAILFAST;
1266 mbio->bi_private = r10_bio;
1268 if (conf->mddev->gendisk)
1269 trace_block_bio_remap(mbio->bi_disk->queue,
1270 mbio, disk_devt(conf->mddev->gendisk),
1272 /* flush_pending_writes() needs access to the rdev so...*/
1273 mbio->bi_disk = (void *)rdev;
1275 atomic_inc(&r10_bio->remaining);
1277 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1279 plug = container_of(cb, struct raid10_plug_cb, cb);
1283 bio_list_add(&plug->pending, mbio);
1284 plug->pending_cnt++;
1286 spin_lock_irqsave(&conf->device_lock, flags);
1287 bio_list_add(&conf->pending_bio_list, mbio);
1288 conf->pending_count++;
1289 spin_unlock_irqrestore(&conf->device_lock, flags);
1290 md_wakeup_thread(mddev->thread);
1294 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1295 struct r10bio *r10_bio)
1297 struct r10conf *conf = mddev->private;
1299 struct md_rdev *blocked_rdev;
1303 if ((mddev_is_clustered(mddev) &&
1304 md_cluster_ops->area_resyncing(mddev, WRITE,
1305 bio->bi_iter.bi_sector,
1306 bio_end_sector(bio)))) {
1309 prepare_to_wait(&conf->wait_barrier,
1311 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1312 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1316 finish_wait(&conf->wait_barrier, &w);
1319 sectors = r10_bio->sectors;
1320 regular_request_wait(mddev, conf, bio, sectors);
1321 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1322 (mddev->reshape_backwards
1323 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1324 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1325 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1326 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1327 /* Need to update reshape_position in metadata */
1328 mddev->reshape_position = conf->reshape_progress;
1329 set_mask_bits(&mddev->sb_flags, 0,
1330 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1331 md_wakeup_thread(mddev->thread);
1332 raid10_log(conf->mddev, "wait reshape metadata");
1333 wait_event(mddev->sb_wait,
1334 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1336 conf->reshape_safe = mddev->reshape_position;
1339 if (conf->pending_count >= max_queued_requests) {
1340 md_wakeup_thread(mddev->thread);
1341 raid10_log(mddev, "wait queued");
1342 wait_event(conf->wait_barrier,
1343 conf->pending_count < max_queued_requests);
1345 /* first select target devices under rcu_lock and
1346 * inc refcount on their rdev. Record them by setting
1348 * If there are known/acknowledged bad blocks on any device
1349 * on which we have seen a write error, we want to avoid
1350 * writing to those blocks. This potentially requires several
1351 * writes to write around the bad blocks. Each set of writes
1352 * gets its own r10_bio with a set of bios attached.
1355 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1356 raid10_find_phys(conf, r10_bio);
1358 blocked_rdev = NULL;
1360 max_sectors = r10_bio->sectors;
1362 for (i = 0; i < conf->copies; i++) {
1363 int d = r10_bio->devs[i].devnum;
1364 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1365 struct md_rdev *rrdev = rcu_dereference(
1366 conf->mirrors[d].replacement);
1369 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1370 atomic_inc(&rdev->nr_pending);
1371 blocked_rdev = rdev;
1374 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1375 atomic_inc(&rrdev->nr_pending);
1376 blocked_rdev = rrdev;
1379 if (rdev && (test_bit(Faulty, &rdev->flags)))
1381 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1384 r10_bio->devs[i].bio = NULL;
1385 r10_bio->devs[i].repl_bio = NULL;
1387 if (!rdev && !rrdev) {
1388 set_bit(R10BIO_Degraded, &r10_bio->state);
1391 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1393 sector_t dev_sector = r10_bio->devs[i].addr;
1397 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1398 &first_bad, &bad_sectors);
1400 /* Mustn't write here until the bad block
1403 atomic_inc(&rdev->nr_pending);
1404 set_bit(BlockedBadBlocks, &rdev->flags);
1405 blocked_rdev = rdev;
1408 if (is_bad && first_bad <= dev_sector) {
1409 /* Cannot write here at all */
1410 bad_sectors -= (dev_sector - first_bad);
1411 if (bad_sectors < max_sectors)
1412 /* Mustn't write more than bad_sectors
1413 * to other devices yet
1415 max_sectors = bad_sectors;
1416 /* We don't set R10BIO_Degraded as that
1417 * only applies if the disk is missing,
1418 * so it might be re-added, and we want to
1419 * know to recover this chunk.
1420 * In this case the device is here, and the
1421 * fact that this chunk is not in-sync is
1422 * recorded in the bad block log.
1427 int good_sectors = first_bad - dev_sector;
1428 if (good_sectors < max_sectors)
1429 max_sectors = good_sectors;
1433 r10_bio->devs[i].bio = bio;
1434 atomic_inc(&rdev->nr_pending);
1437 r10_bio->devs[i].repl_bio = bio;
1438 atomic_inc(&rrdev->nr_pending);
1443 if (unlikely(blocked_rdev)) {
1444 /* Have to wait for this device to get unblocked, then retry */
1448 for (j = 0; j < i; j++) {
1449 if (r10_bio->devs[j].bio) {
1450 d = r10_bio->devs[j].devnum;
1451 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1453 if (r10_bio->devs[j].repl_bio) {
1454 struct md_rdev *rdev;
1455 d = r10_bio->devs[j].devnum;
1456 rdev = conf->mirrors[d].replacement;
1458 /* Race with remove_disk */
1460 rdev = conf->mirrors[d].rdev;
1462 rdev_dec_pending(rdev, mddev);
1465 allow_barrier(conf);
1466 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1467 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1472 if (max_sectors < r10_bio->sectors)
1473 r10_bio->sectors = max_sectors;
1475 if (r10_bio->sectors < bio_sectors(bio)) {
1476 struct bio *split = bio_split(bio, r10_bio->sectors,
1477 GFP_NOIO, &conf->bio_split);
1478 bio_chain(split, bio);
1479 allow_barrier(conf);
1480 generic_make_request(bio);
1483 r10_bio->master_bio = bio;
1486 atomic_set(&r10_bio->remaining, 1);
1487 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1489 for (i = 0; i < conf->copies; i++) {
1490 if (r10_bio->devs[i].bio)
1491 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1492 if (r10_bio->devs[i].repl_bio)
1493 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1495 one_write_done(r10_bio);
1498 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1500 struct r10conf *conf = mddev->private;
1501 struct r10bio *r10_bio;
1503 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1505 r10_bio->master_bio = bio;
1506 r10_bio->sectors = sectors;
1508 r10_bio->mddev = mddev;
1509 r10_bio->sector = bio->bi_iter.bi_sector;
1511 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1513 if (bio_data_dir(bio) == READ)
1514 raid10_read_request(mddev, bio, r10_bio);
1516 raid10_write_request(mddev, bio, r10_bio);
1519 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1521 struct r10conf *conf = mddev->private;
1522 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1523 int chunk_sects = chunk_mask + 1;
1524 int sectors = bio_sectors(bio);
1526 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1527 md_flush_request(mddev, bio);
1531 if (!md_write_start(mddev, bio))
1535 * If this request crosses a chunk boundary, we need to split
1538 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1539 sectors > chunk_sects
1540 && (conf->geo.near_copies < conf->geo.raid_disks
1541 || conf->prev.near_copies <
1542 conf->prev.raid_disks)))
1543 sectors = chunk_sects -
1544 (bio->bi_iter.bi_sector &
1546 __make_request(mddev, bio, sectors);
1548 /* In case raid10d snuck in to freeze_array */
1549 wake_up(&conf->wait_barrier);
1553 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1555 struct r10conf *conf = mddev->private;
1558 if (conf->geo.near_copies < conf->geo.raid_disks)
1559 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1560 if (conf->geo.near_copies > 1)
1561 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1562 if (conf->geo.far_copies > 1) {
1563 if (conf->geo.far_offset)
1564 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1566 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1567 if (conf->geo.far_set_size != conf->geo.raid_disks)
1568 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1570 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1571 conf->geo.raid_disks - mddev->degraded);
1573 for (i = 0; i < conf->geo.raid_disks; i++) {
1574 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1575 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1578 seq_printf(seq, "]");
1581 /* check if there are enough drives for
1582 * every block to appear on atleast one.
1583 * Don't consider the device numbered 'ignore'
1584 * as we might be about to remove it.
1586 static int _enough(struct r10conf *conf, int previous, int ignore)
1592 disks = conf->prev.raid_disks;
1593 ncopies = conf->prev.near_copies;
1595 disks = conf->geo.raid_disks;
1596 ncopies = conf->geo.near_copies;
1601 int n = conf->copies;
1605 struct md_rdev *rdev;
1606 if (this != ignore &&
1607 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1608 test_bit(In_sync, &rdev->flags))
1610 this = (this+1) % disks;
1614 first = (first + ncopies) % disks;
1615 } while (first != 0);
1622 static int enough(struct r10conf *conf, int ignore)
1624 /* when calling 'enough', both 'prev' and 'geo' must
1626 * This is ensured if ->reconfig_mutex or ->device_lock
1629 return _enough(conf, 0, ignore) &&
1630 _enough(conf, 1, ignore);
1633 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1635 char b[BDEVNAME_SIZE];
1636 struct r10conf *conf = mddev->private;
1637 unsigned long flags;
1640 * If it is not operational, then we have already marked it as dead
1641 * else if it is the last working disks, ignore the error, let the
1642 * next level up know.
1643 * else mark the drive as failed
1645 spin_lock_irqsave(&conf->device_lock, flags);
1646 if (test_bit(In_sync, &rdev->flags)
1647 && !enough(conf, rdev->raid_disk)) {
1649 * Don't fail the drive, just return an IO error.
1651 spin_unlock_irqrestore(&conf->device_lock, flags);
1654 if (test_and_clear_bit(In_sync, &rdev->flags))
1657 * If recovery is running, make sure it aborts.
1659 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1660 set_bit(Blocked, &rdev->flags);
1661 set_bit(Faulty, &rdev->flags);
1662 set_mask_bits(&mddev->sb_flags, 0,
1663 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1664 spin_unlock_irqrestore(&conf->device_lock, flags);
1665 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1666 "md/raid10:%s: Operation continuing on %d devices.\n",
1667 mdname(mddev), bdevname(rdev->bdev, b),
1668 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1671 static void print_conf(struct r10conf *conf)
1674 struct md_rdev *rdev;
1676 pr_debug("RAID10 conf printout:\n");
1678 pr_debug("(!conf)\n");
1681 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1682 conf->geo.raid_disks);
1684 /* This is only called with ->reconfix_mutex held, so
1685 * rcu protection of rdev is not needed */
1686 for (i = 0; i < conf->geo.raid_disks; i++) {
1687 char b[BDEVNAME_SIZE];
1688 rdev = conf->mirrors[i].rdev;
1690 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1691 i, !test_bit(In_sync, &rdev->flags),
1692 !test_bit(Faulty, &rdev->flags),
1693 bdevname(rdev->bdev,b));
1697 static void close_sync(struct r10conf *conf)
1700 allow_barrier(conf);
1702 mempool_exit(&conf->r10buf_pool);
1705 static int raid10_spare_active(struct mddev *mddev)
1708 struct r10conf *conf = mddev->private;
1709 struct raid10_info *tmp;
1711 unsigned long flags;
1714 * Find all non-in_sync disks within the RAID10 configuration
1715 * and mark them in_sync
1717 for (i = 0; i < conf->geo.raid_disks; i++) {
1718 tmp = conf->mirrors + i;
1719 if (tmp->replacement
1720 && tmp->replacement->recovery_offset == MaxSector
1721 && !test_bit(Faulty, &tmp->replacement->flags)
1722 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1723 /* Replacement has just become active */
1725 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1728 /* Replaced device not technically faulty,
1729 * but we need to be sure it gets removed
1730 * and never re-added.
1732 set_bit(Faulty, &tmp->rdev->flags);
1733 sysfs_notify_dirent_safe(
1734 tmp->rdev->sysfs_state);
1736 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1737 } else if (tmp->rdev
1738 && tmp->rdev->recovery_offset == MaxSector
1739 && !test_bit(Faulty, &tmp->rdev->flags)
1740 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1742 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1745 spin_lock_irqsave(&conf->device_lock, flags);
1746 mddev->degraded -= count;
1747 spin_unlock_irqrestore(&conf->device_lock, flags);
1753 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1755 struct r10conf *conf = mddev->private;
1759 int last = conf->geo.raid_disks - 1;
1761 if (mddev->recovery_cp < MaxSector)
1762 /* only hot-add to in-sync arrays, as recovery is
1763 * very different from resync
1766 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1769 if (md_integrity_add_rdev(rdev, mddev))
1772 if (rdev->raid_disk >= 0)
1773 first = last = rdev->raid_disk;
1775 if (rdev->saved_raid_disk >= first &&
1776 rdev->saved_raid_disk < conf->geo.raid_disks &&
1777 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1778 mirror = rdev->saved_raid_disk;
1781 for ( ; mirror <= last ; mirror++) {
1782 struct raid10_info *p = &conf->mirrors[mirror];
1783 if (p->recovery_disabled == mddev->recovery_disabled)
1786 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1787 p->replacement != NULL)
1789 clear_bit(In_sync, &rdev->flags);
1790 set_bit(Replacement, &rdev->flags);
1791 rdev->raid_disk = mirror;
1794 disk_stack_limits(mddev->gendisk, rdev->bdev,
1795 rdev->data_offset << 9);
1797 rcu_assign_pointer(p->replacement, rdev);
1802 disk_stack_limits(mddev->gendisk, rdev->bdev,
1803 rdev->data_offset << 9);
1805 p->head_position = 0;
1806 p->recovery_disabled = mddev->recovery_disabled - 1;
1807 rdev->raid_disk = mirror;
1809 if (rdev->saved_raid_disk != mirror)
1811 rcu_assign_pointer(p->rdev, rdev);
1814 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1815 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1821 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1823 struct r10conf *conf = mddev->private;
1825 int number = rdev->raid_disk;
1826 struct md_rdev **rdevp;
1827 struct raid10_info *p = conf->mirrors + number;
1830 if (rdev == p->rdev)
1832 else if (rdev == p->replacement)
1833 rdevp = &p->replacement;
1837 if (test_bit(In_sync, &rdev->flags) ||
1838 atomic_read(&rdev->nr_pending)) {
1842 /* Only remove non-faulty devices if recovery
1845 if (!test_bit(Faulty, &rdev->flags) &&
1846 mddev->recovery_disabled != p->recovery_disabled &&
1847 (!p->replacement || p->replacement == rdev) &&
1848 number < conf->geo.raid_disks &&
1854 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1856 if (atomic_read(&rdev->nr_pending)) {
1857 /* lost the race, try later */
1863 if (p->replacement) {
1864 /* We must have just cleared 'rdev' */
1865 p->rdev = p->replacement;
1866 clear_bit(Replacement, &p->replacement->flags);
1867 smp_mb(); /* Make sure other CPUs may see both as identical
1868 * but will never see neither -- if they are careful.
1870 p->replacement = NULL;
1873 clear_bit(WantReplacement, &rdev->flags);
1874 err = md_integrity_register(mddev);
1882 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1884 struct r10conf *conf = r10_bio->mddev->private;
1886 if (!bio->bi_status)
1887 set_bit(R10BIO_Uptodate, &r10_bio->state);
1889 /* The write handler will notice the lack of
1890 * R10BIO_Uptodate and record any errors etc
1892 atomic_add(r10_bio->sectors,
1893 &conf->mirrors[d].rdev->corrected_errors);
1895 /* for reconstruct, we always reschedule after a read.
1896 * for resync, only after all reads
1898 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1899 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1900 atomic_dec_and_test(&r10_bio->remaining)) {
1901 /* we have read all the blocks,
1902 * do the comparison in process context in raid10d
1904 reschedule_retry(r10_bio);
1908 static void end_sync_read(struct bio *bio)
1910 struct r10bio *r10_bio = get_resync_r10bio(bio);
1911 struct r10conf *conf = r10_bio->mddev->private;
1912 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1914 __end_sync_read(r10_bio, bio, d);
1917 static void end_reshape_read(struct bio *bio)
1919 /* reshape read bio isn't allocated from r10buf_pool */
1920 struct r10bio *r10_bio = bio->bi_private;
1922 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1925 static void end_sync_request(struct r10bio *r10_bio)
1927 struct mddev *mddev = r10_bio->mddev;
1929 while (atomic_dec_and_test(&r10_bio->remaining)) {
1930 if (r10_bio->master_bio == NULL) {
1931 /* the primary of several recovery bios */
1932 sector_t s = r10_bio->sectors;
1933 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1934 test_bit(R10BIO_WriteError, &r10_bio->state))
1935 reschedule_retry(r10_bio);
1938 md_done_sync(mddev, s, 1);
1941 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1942 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1943 test_bit(R10BIO_WriteError, &r10_bio->state))
1944 reschedule_retry(r10_bio);
1952 static void end_sync_write(struct bio *bio)
1954 struct r10bio *r10_bio = get_resync_r10bio(bio);
1955 struct mddev *mddev = r10_bio->mddev;
1956 struct r10conf *conf = mddev->private;
1962 struct md_rdev *rdev = NULL;
1964 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1966 rdev = conf->mirrors[d].replacement;
1968 rdev = conf->mirrors[d].rdev;
1970 if (bio->bi_status) {
1972 md_error(mddev, rdev);
1974 set_bit(WriteErrorSeen, &rdev->flags);
1975 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1976 set_bit(MD_RECOVERY_NEEDED,
1977 &rdev->mddev->recovery);
1978 set_bit(R10BIO_WriteError, &r10_bio->state);
1980 } else if (is_badblock(rdev,
1981 r10_bio->devs[slot].addr,
1983 &first_bad, &bad_sectors))
1984 set_bit(R10BIO_MadeGood, &r10_bio->state);
1986 rdev_dec_pending(rdev, mddev);
1988 end_sync_request(r10_bio);
1992 * Note: sync and recover and handled very differently for raid10
1993 * This code is for resync.
1994 * For resync, we read through virtual addresses and read all blocks.
1995 * If there is any error, we schedule a write. The lowest numbered
1996 * drive is authoritative.
1997 * However requests come for physical address, so we need to map.
1998 * For every physical address there are raid_disks/copies virtual addresses,
1999 * which is always are least one, but is not necessarly an integer.
2000 * This means that a physical address can span multiple chunks, so we may
2001 * have to submit multiple io requests for a single sync request.
2004 * We check if all blocks are in-sync and only write to blocks that
2007 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2009 struct r10conf *conf = mddev->private;
2011 struct bio *tbio, *fbio;
2013 struct page **tpages, **fpages;
2015 atomic_set(&r10_bio->remaining, 1);
2017 /* find the first device with a block */
2018 for (i=0; i<conf->copies; i++)
2019 if (!r10_bio->devs[i].bio->bi_status)
2022 if (i == conf->copies)
2026 fbio = r10_bio->devs[i].bio;
2027 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2028 fbio->bi_iter.bi_idx = 0;
2029 fpages = get_resync_pages(fbio)->pages;
2031 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2032 /* now find blocks with errors */
2033 for (i=0 ; i < conf->copies ; i++) {
2035 struct md_rdev *rdev;
2036 struct resync_pages *rp;
2038 tbio = r10_bio->devs[i].bio;
2040 if (tbio->bi_end_io != end_sync_read)
2045 tpages = get_resync_pages(tbio)->pages;
2046 d = r10_bio->devs[i].devnum;
2047 rdev = conf->mirrors[d].rdev;
2048 if (!r10_bio->devs[i].bio->bi_status) {
2049 /* We know that the bi_io_vec layout is the same for
2050 * both 'first' and 'i', so we just compare them.
2051 * All vec entries are PAGE_SIZE;
2053 int sectors = r10_bio->sectors;
2054 for (j = 0; j < vcnt; j++) {
2055 int len = PAGE_SIZE;
2056 if (sectors < (len / 512))
2057 len = sectors * 512;
2058 if (memcmp(page_address(fpages[j]),
2059 page_address(tpages[j]),
2066 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2067 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2068 /* Don't fix anything. */
2070 } else if (test_bit(FailFast, &rdev->flags)) {
2071 /* Just give up on this device */
2072 md_error(rdev->mddev, rdev);
2075 /* Ok, we need to write this bio, either to correct an
2076 * inconsistency or to correct an unreadable block.
2077 * First we need to fixup bv_offset, bv_len and
2078 * bi_vecs, as the read request might have corrupted these
2080 rp = get_resync_pages(tbio);
2083 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2085 rp->raid_bio = r10_bio;
2086 tbio->bi_private = rp;
2087 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2088 tbio->bi_end_io = end_sync_write;
2089 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2091 bio_copy_data(tbio, fbio);
2093 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2094 atomic_inc(&r10_bio->remaining);
2095 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2097 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2098 tbio->bi_opf |= MD_FAILFAST;
2099 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2100 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2101 generic_make_request(tbio);
2104 /* Now write out to any replacement devices
2107 for (i = 0; i < conf->copies; i++) {
2110 tbio = r10_bio->devs[i].repl_bio;
2111 if (!tbio || !tbio->bi_end_io)
2113 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2114 && r10_bio->devs[i].bio != fbio)
2115 bio_copy_data(tbio, fbio);
2116 d = r10_bio->devs[i].devnum;
2117 atomic_inc(&r10_bio->remaining);
2118 md_sync_acct(conf->mirrors[d].replacement->bdev,
2120 generic_make_request(tbio);
2124 if (atomic_dec_and_test(&r10_bio->remaining)) {
2125 md_done_sync(mddev, r10_bio->sectors, 1);
2131 * Now for the recovery code.
2132 * Recovery happens across physical sectors.
2133 * We recover all non-is_sync drives by finding the virtual address of
2134 * each, and then choose a working drive that also has that virt address.
2135 * There is a separate r10_bio for each non-in_sync drive.
2136 * Only the first two slots are in use. The first for reading,
2137 * The second for writing.
2140 static void fix_recovery_read_error(struct r10bio *r10_bio)
2142 /* We got a read error during recovery.
2143 * We repeat the read in smaller page-sized sections.
2144 * If a read succeeds, write it to the new device or record
2145 * a bad block if we cannot.
2146 * If a read fails, record a bad block on both old and
2149 struct mddev *mddev = r10_bio->mddev;
2150 struct r10conf *conf = mddev->private;
2151 struct bio *bio = r10_bio->devs[0].bio;
2153 int sectors = r10_bio->sectors;
2155 int dr = r10_bio->devs[0].devnum;
2156 int dw = r10_bio->devs[1].devnum;
2157 struct page **pages = get_resync_pages(bio)->pages;
2161 struct md_rdev *rdev;
2165 if (s > (PAGE_SIZE>>9))
2168 rdev = conf->mirrors[dr].rdev;
2169 addr = r10_bio->devs[0].addr + sect,
2170 ok = sync_page_io(rdev,
2174 REQ_OP_READ, 0, false);
2176 rdev = conf->mirrors[dw].rdev;
2177 addr = r10_bio->devs[1].addr + sect;
2178 ok = sync_page_io(rdev,
2182 REQ_OP_WRITE, 0, false);
2184 set_bit(WriteErrorSeen, &rdev->flags);
2185 if (!test_and_set_bit(WantReplacement,
2187 set_bit(MD_RECOVERY_NEEDED,
2188 &rdev->mddev->recovery);
2192 /* We don't worry if we cannot set a bad block -
2193 * it really is bad so there is no loss in not
2196 rdev_set_badblocks(rdev, addr, s, 0);
2198 if (rdev != conf->mirrors[dw].rdev) {
2199 /* need bad block on destination too */
2200 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2201 addr = r10_bio->devs[1].addr + sect;
2202 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2204 /* just abort the recovery */
2205 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2208 conf->mirrors[dw].recovery_disabled
2209 = mddev->recovery_disabled;
2210 set_bit(MD_RECOVERY_INTR,
2223 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2225 struct r10conf *conf = mddev->private;
2227 struct bio *wbio, *wbio2;
2229 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2230 fix_recovery_read_error(r10_bio);
2231 end_sync_request(r10_bio);
2236 * share the pages with the first bio
2237 * and submit the write request
2239 d = r10_bio->devs[1].devnum;
2240 wbio = r10_bio->devs[1].bio;
2241 wbio2 = r10_bio->devs[1].repl_bio;
2242 /* Need to test wbio2->bi_end_io before we call
2243 * generic_make_request as if the former is NULL,
2244 * the latter is free to free wbio2.
2246 if (wbio2 && !wbio2->bi_end_io)
2248 if (wbio->bi_end_io) {
2249 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2250 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2251 generic_make_request(wbio);
2254 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2255 md_sync_acct(conf->mirrors[d].replacement->bdev,
2256 bio_sectors(wbio2));
2257 generic_make_request(wbio2);
2262 * Used by fix_read_error() to decay the per rdev read_errors.
2263 * We halve the read error count for every hour that has elapsed
2264 * since the last recorded read error.
2267 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2270 unsigned long hours_since_last;
2271 unsigned int read_errors = atomic_read(&rdev->read_errors);
2273 cur_time_mon = ktime_get_seconds();
2275 if (rdev->last_read_error == 0) {
2276 /* first time we've seen a read error */
2277 rdev->last_read_error = cur_time_mon;
2281 hours_since_last = (long)(cur_time_mon -
2282 rdev->last_read_error) / 3600;
2284 rdev->last_read_error = cur_time_mon;
2287 * if hours_since_last is > the number of bits in read_errors
2288 * just set read errors to 0. We do this to avoid
2289 * overflowing the shift of read_errors by hours_since_last.
2291 if (hours_since_last >= 8 * sizeof(read_errors))
2292 atomic_set(&rdev->read_errors, 0);
2294 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2297 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2298 int sectors, struct page *page, int rw)
2303 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2304 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2306 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2310 set_bit(WriteErrorSeen, &rdev->flags);
2311 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2312 set_bit(MD_RECOVERY_NEEDED,
2313 &rdev->mddev->recovery);
2315 /* need to record an error - either for the block or the device */
2316 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2317 md_error(rdev->mddev, rdev);
2322 * This is a kernel thread which:
2324 * 1. Retries failed read operations on working mirrors.
2325 * 2. Updates the raid superblock when problems encounter.
2326 * 3. Performs writes following reads for array synchronising.
2329 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2331 int sect = 0; /* Offset from r10_bio->sector */
2332 int sectors = r10_bio->sectors;
2333 struct md_rdev *rdev;
2334 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2335 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2337 /* still own a reference to this rdev, so it cannot
2338 * have been cleared recently.
2340 rdev = conf->mirrors[d].rdev;
2342 if (test_bit(Faulty, &rdev->flags))
2343 /* drive has already been failed, just ignore any
2344 more fix_read_error() attempts */
2347 check_decay_read_errors(mddev, rdev);
2348 atomic_inc(&rdev->read_errors);
2349 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2350 char b[BDEVNAME_SIZE];
2351 bdevname(rdev->bdev, b);
2353 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2355 atomic_read(&rdev->read_errors), max_read_errors);
2356 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2358 md_error(mddev, rdev);
2359 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2365 int sl = r10_bio->read_slot;
2369 if (s > (PAGE_SIZE>>9))
2377 d = r10_bio->devs[sl].devnum;
2378 rdev = rcu_dereference(conf->mirrors[d].rdev);
2380 test_bit(In_sync, &rdev->flags) &&
2381 !test_bit(Faulty, &rdev->flags) &&
2382 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2383 &first_bad, &bad_sectors) == 0) {
2384 atomic_inc(&rdev->nr_pending);
2386 success = sync_page_io(rdev,
2387 r10_bio->devs[sl].addr +
2391 REQ_OP_READ, 0, false);
2392 rdev_dec_pending(rdev, mddev);
2398 if (sl == conf->copies)
2400 } while (!success && sl != r10_bio->read_slot);
2404 /* Cannot read from anywhere, just mark the block
2405 * as bad on the first device to discourage future
2408 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2409 rdev = conf->mirrors[dn].rdev;
2411 if (!rdev_set_badblocks(
2413 r10_bio->devs[r10_bio->read_slot].addr
2416 md_error(mddev, rdev);
2417 r10_bio->devs[r10_bio->read_slot].bio
2424 /* write it back and re-read */
2426 while (sl != r10_bio->read_slot) {
2427 char b[BDEVNAME_SIZE];
2432 d = r10_bio->devs[sl].devnum;
2433 rdev = rcu_dereference(conf->mirrors[d].rdev);
2435 test_bit(Faulty, &rdev->flags) ||
2436 !test_bit(In_sync, &rdev->flags))
2439 atomic_inc(&rdev->nr_pending);
2441 if (r10_sync_page_io(rdev,
2442 r10_bio->devs[sl].addr +
2444 s, conf->tmppage, WRITE)
2446 /* Well, this device is dead */
2447 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2449 (unsigned long long)(
2451 choose_data_offset(r10_bio,
2453 bdevname(rdev->bdev, b));
2454 pr_notice("md/raid10:%s: %s: failing drive\n",
2456 bdevname(rdev->bdev, b));
2458 rdev_dec_pending(rdev, mddev);
2462 while (sl != r10_bio->read_slot) {
2463 char b[BDEVNAME_SIZE];
2468 d = r10_bio->devs[sl].devnum;
2469 rdev = rcu_dereference(conf->mirrors[d].rdev);
2471 test_bit(Faulty, &rdev->flags) ||
2472 !test_bit(In_sync, &rdev->flags))
2475 atomic_inc(&rdev->nr_pending);
2477 switch (r10_sync_page_io(rdev,
2478 r10_bio->devs[sl].addr +
2483 /* Well, this device is dead */
2484 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2486 (unsigned long long)(
2488 choose_data_offset(r10_bio, rdev)),
2489 bdevname(rdev->bdev, b));
2490 pr_notice("md/raid10:%s: %s: failing drive\n",
2492 bdevname(rdev->bdev, b));
2495 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2497 (unsigned long long)(
2499 choose_data_offset(r10_bio, rdev)),
2500 bdevname(rdev->bdev, b));
2501 atomic_add(s, &rdev->corrected_errors);
2504 rdev_dec_pending(rdev, mddev);
2514 static int narrow_write_error(struct r10bio *r10_bio, int i)
2516 struct bio *bio = r10_bio->master_bio;
2517 struct mddev *mddev = r10_bio->mddev;
2518 struct r10conf *conf = mddev->private;
2519 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2520 /* bio has the data to be written to slot 'i' where
2521 * we just recently had a write error.
2522 * We repeatedly clone the bio and trim down to one block,
2523 * then try the write. Where the write fails we record
2525 * It is conceivable that the bio doesn't exactly align with
2526 * blocks. We must handle this.
2528 * We currently own a reference to the rdev.
2534 int sect_to_write = r10_bio->sectors;
2537 if (rdev->badblocks.shift < 0)
2540 block_sectors = roundup(1 << rdev->badblocks.shift,
2541 bdev_logical_block_size(rdev->bdev) >> 9);
2542 sector = r10_bio->sector;
2543 sectors = ((r10_bio->sector + block_sectors)
2544 & ~(sector_t)(block_sectors - 1))
2547 while (sect_to_write) {
2550 if (sectors > sect_to_write)
2551 sectors = sect_to_write;
2552 /* Write at 'sector' for 'sectors' */
2553 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2554 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2555 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2556 wbio->bi_iter.bi_sector = wsector +
2557 choose_data_offset(r10_bio, rdev);
2558 bio_set_dev(wbio, rdev->bdev);
2559 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2561 if (submit_bio_wait(wbio) < 0)
2563 ok = rdev_set_badblocks(rdev, wsector,
2568 sect_to_write -= sectors;
2570 sectors = block_sectors;
2575 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2577 int slot = r10_bio->read_slot;
2579 struct r10conf *conf = mddev->private;
2580 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2582 /* we got a read error. Maybe the drive is bad. Maybe just
2583 * the block and we can fix it.
2584 * We freeze all other IO, and try reading the block from
2585 * other devices. When we find one, we re-write
2586 * and check it that fixes the read error.
2587 * This is all done synchronously while the array is
2590 bio = r10_bio->devs[slot].bio;
2592 r10_bio->devs[slot].bio = NULL;
2595 r10_bio->devs[slot].bio = IO_BLOCKED;
2596 else if (!test_bit(FailFast, &rdev->flags)) {
2597 freeze_array(conf, 1);
2598 fix_read_error(conf, mddev, r10_bio);
2599 unfreeze_array(conf);
2601 md_error(mddev, rdev);
2603 rdev_dec_pending(rdev, mddev);
2604 allow_barrier(conf);
2606 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2609 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2611 /* Some sort of write request has finished and it
2612 * succeeded in writing where we thought there was a
2613 * bad block. So forget the bad block.
2614 * Or possibly if failed and we need to record
2618 struct md_rdev *rdev;
2620 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2621 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2622 for (m = 0; m < conf->copies; m++) {
2623 int dev = r10_bio->devs[m].devnum;
2624 rdev = conf->mirrors[dev].rdev;
2625 if (r10_bio->devs[m].bio == NULL ||
2626 r10_bio->devs[m].bio->bi_end_io == NULL)
2628 if (!r10_bio->devs[m].bio->bi_status) {
2629 rdev_clear_badblocks(
2631 r10_bio->devs[m].addr,
2632 r10_bio->sectors, 0);
2634 if (!rdev_set_badblocks(
2636 r10_bio->devs[m].addr,
2637 r10_bio->sectors, 0))
2638 md_error(conf->mddev, rdev);
2640 rdev = conf->mirrors[dev].replacement;
2641 if (r10_bio->devs[m].repl_bio == NULL ||
2642 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2645 if (!r10_bio->devs[m].repl_bio->bi_status) {
2646 rdev_clear_badblocks(
2648 r10_bio->devs[m].addr,
2649 r10_bio->sectors, 0);
2651 if (!rdev_set_badblocks(
2653 r10_bio->devs[m].addr,
2654 r10_bio->sectors, 0))
2655 md_error(conf->mddev, rdev);
2661 for (m = 0; m < conf->copies; m++) {
2662 int dev = r10_bio->devs[m].devnum;
2663 struct bio *bio = r10_bio->devs[m].bio;
2664 rdev = conf->mirrors[dev].rdev;
2665 if (bio == IO_MADE_GOOD) {
2666 rdev_clear_badblocks(
2668 r10_bio->devs[m].addr,
2669 r10_bio->sectors, 0);
2670 rdev_dec_pending(rdev, conf->mddev);
2671 } else if (bio != NULL && bio->bi_status) {
2673 if (!narrow_write_error(r10_bio, m)) {
2674 md_error(conf->mddev, rdev);
2675 set_bit(R10BIO_Degraded,
2678 rdev_dec_pending(rdev, conf->mddev);
2680 bio = r10_bio->devs[m].repl_bio;
2681 rdev = conf->mirrors[dev].replacement;
2682 if (rdev && bio == IO_MADE_GOOD) {
2683 rdev_clear_badblocks(
2685 r10_bio->devs[m].addr,
2686 r10_bio->sectors, 0);
2687 rdev_dec_pending(rdev, conf->mddev);
2691 spin_lock_irq(&conf->device_lock);
2692 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2694 spin_unlock_irq(&conf->device_lock);
2696 * In case freeze_array() is waiting for condition
2697 * nr_pending == nr_queued + extra to be true.
2699 wake_up(&conf->wait_barrier);
2700 md_wakeup_thread(conf->mddev->thread);
2702 if (test_bit(R10BIO_WriteError,
2704 close_write(r10_bio);
2705 raid_end_bio_io(r10_bio);
2710 static void raid10d(struct md_thread *thread)
2712 struct mddev *mddev = thread->mddev;
2713 struct r10bio *r10_bio;
2714 unsigned long flags;
2715 struct r10conf *conf = mddev->private;
2716 struct list_head *head = &conf->retry_list;
2717 struct blk_plug plug;
2719 md_check_recovery(mddev);
2721 if (!list_empty_careful(&conf->bio_end_io_list) &&
2722 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2724 spin_lock_irqsave(&conf->device_lock, flags);
2725 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2726 while (!list_empty(&conf->bio_end_io_list)) {
2727 list_move(conf->bio_end_io_list.prev, &tmp);
2731 spin_unlock_irqrestore(&conf->device_lock, flags);
2732 while (!list_empty(&tmp)) {
2733 r10_bio = list_first_entry(&tmp, struct r10bio,
2735 list_del(&r10_bio->retry_list);
2736 if (mddev->degraded)
2737 set_bit(R10BIO_Degraded, &r10_bio->state);
2739 if (test_bit(R10BIO_WriteError,
2741 close_write(r10_bio);
2742 raid_end_bio_io(r10_bio);
2746 blk_start_plug(&plug);
2749 flush_pending_writes(conf);
2751 spin_lock_irqsave(&conf->device_lock, flags);
2752 if (list_empty(head)) {
2753 spin_unlock_irqrestore(&conf->device_lock, flags);
2756 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2757 list_del(head->prev);
2759 spin_unlock_irqrestore(&conf->device_lock, flags);
2761 mddev = r10_bio->mddev;
2762 conf = mddev->private;
2763 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2764 test_bit(R10BIO_WriteError, &r10_bio->state))
2765 handle_write_completed(conf, r10_bio);
2766 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2767 reshape_request_write(mddev, r10_bio);
2768 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2769 sync_request_write(mddev, r10_bio);
2770 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2771 recovery_request_write(mddev, r10_bio);
2772 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2773 handle_read_error(mddev, r10_bio);
2778 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2779 md_check_recovery(mddev);
2781 blk_finish_plug(&plug);
2784 static int init_resync(struct r10conf *conf)
2788 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2789 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2790 conf->have_replacement = 0;
2791 for (i = 0; i < conf->geo.raid_disks; i++)
2792 if (conf->mirrors[i].replacement)
2793 conf->have_replacement = 1;
2794 ret = mempool_init(&conf->r10buf_pool, buffs,
2795 r10buf_pool_alloc, r10buf_pool_free, conf);
2798 conf->next_resync = 0;
2802 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2804 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2805 struct rsync_pages *rp;
2810 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2811 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2812 nalloc = conf->copies; /* resync */
2814 nalloc = 2; /* recovery */
2816 for (i = 0; i < nalloc; i++) {
2817 bio = r10bio->devs[i].bio;
2818 rp = bio->bi_private;
2820 bio->bi_private = rp;
2821 bio = r10bio->devs[i].repl_bio;
2823 rp = bio->bi_private;
2825 bio->bi_private = rp;
2832 * Set cluster_sync_high since we need other nodes to add the
2833 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2835 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2837 sector_t window_size;
2838 int extra_chunk, chunks;
2841 * First, here we define "stripe" as a unit which across
2842 * all member devices one time, so we get chunks by use
2843 * raid_disks / near_copies. Otherwise, if near_copies is
2844 * close to raid_disks, then resync window could increases
2845 * linearly with the increase of raid_disks, which means
2846 * we will suspend a really large IO window while it is not
2847 * necessary. If raid_disks is not divisible by near_copies,
2848 * an extra chunk is needed to ensure the whole "stripe" is
2852 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2853 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2857 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2860 * At least use a 32M window to align with raid1's resync window
2862 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2863 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2865 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2869 * perform a "sync" on one "block"
2871 * We need to make sure that no normal I/O request - particularly write
2872 * requests - conflict with active sync requests.
2874 * This is achieved by tracking pending requests and a 'barrier' concept
2875 * that can be installed to exclude normal IO requests.
2877 * Resync and recovery are handled very differently.
2878 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2880 * For resync, we iterate over virtual addresses, read all copies,
2881 * and update if there are differences. If only one copy is live,
2883 * For recovery, we iterate over physical addresses, read a good
2884 * value for each non-in_sync drive, and over-write.
2886 * So, for recovery we may have several outstanding complex requests for a
2887 * given address, one for each out-of-sync device. We model this by allocating
2888 * a number of r10_bio structures, one for each out-of-sync device.
2889 * As we setup these structures, we collect all bio's together into a list
2890 * which we then process collectively to add pages, and then process again
2891 * to pass to generic_make_request.
2893 * The r10_bio structures are linked using a borrowed master_bio pointer.
2894 * This link is counted in ->remaining. When the r10_bio that points to NULL
2895 * has its remaining count decremented to 0, the whole complex operation
2900 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2903 struct r10conf *conf = mddev->private;
2904 struct r10bio *r10_bio;
2905 struct bio *biolist = NULL, *bio;
2906 sector_t max_sector, nr_sectors;
2909 sector_t sync_blocks;
2910 sector_t sectors_skipped = 0;
2911 int chunks_skipped = 0;
2912 sector_t chunk_mask = conf->geo.chunk_mask;
2915 if (!mempool_initialized(&conf->r10buf_pool))
2916 if (init_resync(conf))
2920 * Allow skipping a full rebuild for incremental assembly
2921 * of a clean array, like RAID1 does.
2923 if (mddev->bitmap == NULL &&
2924 mddev->recovery_cp == MaxSector &&
2925 mddev->reshape_position == MaxSector &&
2926 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2927 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2928 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2929 conf->fullsync == 0) {
2931 return mddev->dev_sectors - sector_nr;
2935 max_sector = mddev->dev_sectors;
2936 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2937 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2938 max_sector = mddev->resync_max_sectors;
2939 if (sector_nr >= max_sector) {
2940 conf->cluster_sync_low = 0;
2941 conf->cluster_sync_high = 0;
2943 /* If we aborted, we need to abort the
2944 * sync on the 'current' bitmap chucks (there can
2945 * be several when recovering multiple devices).
2946 * as we may have started syncing it but not finished.
2947 * We can find the current address in
2948 * mddev->curr_resync, but for recovery,
2949 * we need to convert that to several
2950 * virtual addresses.
2952 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2958 if (mddev->curr_resync < max_sector) { /* aborted */
2959 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2960 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2962 else for (i = 0; i < conf->geo.raid_disks; i++) {
2964 raid10_find_virt(conf, mddev->curr_resync, i);
2965 md_bitmap_end_sync(mddev->bitmap, sect,
2969 /* completed sync */
2970 if ((!mddev->bitmap || conf->fullsync)
2971 && conf->have_replacement
2972 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2973 /* Completed a full sync so the replacements
2974 * are now fully recovered.
2977 for (i = 0; i < conf->geo.raid_disks; i++) {
2978 struct md_rdev *rdev =
2979 rcu_dereference(conf->mirrors[i].replacement);
2981 rdev->recovery_offset = MaxSector;
2987 md_bitmap_close_sync(mddev->bitmap);
2990 return sectors_skipped;
2993 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2994 return reshape_request(mddev, sector_nr, skipped);
2996 if (chunks_skipped >= conf->geo.raid_disks) {
2997 /* if there has been nothing to do on any drive,
2998 * then there is nothing to do at all..
3001 return (max_sector - sector_nr) + sectors_skipped;
3004 if (max_sector > mddev->resync_max)
3005 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3007 /* make sure whole request will fit in a chunk - if chunks
3010 if (conf->geo.near_copies < conf->geo.raid_disks &&
3011 max_sector > (sector_nr | chunk_mask))
3012 max_sector = (sector_nr | chunk_mask) + 1;
3015 * If there is non-resync activity waiting for a turn, then let it
3016 * though before starting on this new sync request.
3018 if (conf->nr_waiting)
3019 schedule_timeout_uninterruptible(1);
3021 /* Again, very different code for resync and recovery.
3022 * Both must result in an r10bio with a list of bios that
3023 * have bi_end_io, bi_sector, bi_disk set,
3024 * and bi_private set to the r10bio.
3025 * For recovery, we may actually create several r10bios
3026 * with 2 bios in each, that correspond to the bios in the main one.
3027 * In this case, the subordinate r10bios link back through a
3028 * borrowed master_bio pointer, and the counter in the master
3029 * includes a ref from each subordinate.
3031 /* First, we decide what to do and set ->bi_end_io
3032 * To end_sync_read if we want to read, and
3033 * end_sync_write if we will want to write.
3036 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3037 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3038 /* recovery... the complicated one */
3042 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3048 int need_recover = 0;
3049 int need_replace = 0;
3050 struct raid10_info *mirror = &conf->mirrors[i];
3051 struct md_rdev *mrdev, *mreplace;
3054 mrdev = rcu_dereference(mirror->rdev);
3055 mreplace = rcu_dereference(mirror->replacement);
3057 if (mrdev != NULL &&
3058 !test_bit(Faulty, &mrdev->flags) &&
3059 !test_bit(In_sync, &mrdev->flags))
3061 if (mreplace != NULL &&
3062 !test_bit(Faulty, &mreplace->flags))
3065 if (!need_recover && !need_replace) {
3071 /* want to reconstruct this device */
3073 sect = raid10_find_virt(conf, sector_nr, i);
3074 if (sect >= mddev->resync_max_sectors) {
3075 /* last stripe is not complete - don't
3076 * try to recover this sector.
3081 if (mreplace && test_bit(Faulty, &mreplace->flags))
3083 /* Unless we are doing a full sync, or a replacement
3084 * we only need to recover the block if it is set in
3087 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3089 if (sync_blocks < max_sync)
3090 max_sync = sync_blocks;
3094 /* yep, skip the sync_blocks here, but don't assume
3095 * that there will never be anything to do here
3097 chunks_skipped = -1;
3101 atomic_inc(&mrdev->nr_pending);
3103 atomic_inc(&mreplace->nr_pending);
3106 r10_bio = raid10_alloc_init_r10buf(conf);
3108 raise_barrier(conf, rb2 != NULL);
3109 atomic_set(&r10_bio->remaining, 0);
3111 r10_bio->master_bio = (struct bio*)rb2;
3113 atomic_inc(&rb2->remaining);
3114 r10_bio->mddev = mddev;
3115 set_bit(R10BIO_IsRecover, &r10_bio->state);
3116 r10_bio->sector = sect;
3118 raid10_find_phys(conf, r10_bio);
3120 /* Need to check if the array will still be
3124 for (j = 0; j < conf->geo.raid_disks; j++) {
3125 struct md_rdev *rdev = rcu_dereference(
3126 conf->mirrors[j].rdev);
3127 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3133 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3134 &sync_blocks, still_degraded);
3137 for (j=0; j<conf->copies;j++) {
3139 int d = r10_bio->devs[j].devnum;
3140 sector_t from_addr, to_addr;
3141 struct md_rdev *rdev =
3142 rcu_dereference(conf->mirrors[d].rdev);
3143 sector_t sector, first_bad;
3146 !test_bit(In_sync, &rdev->flags))
3148 /* This is where we read from */
3150 sector = r10_bio->devs[j].addr;
3152 if (is_badblock(rdev, sector, max_sync,
3153 &first_bad, &bad_sectors)) {
3154 if (first_bad > sector)
3155 max_sync = first_bad - sector;
3157 bad_sectors -= (sector
3159 if (max_sync > bad_sectors)
3160 max_sync = bad_sectors;
3164 bio = r10_bio->devs[0].bio;
3165 bio->bi_next = biolist;
3167 bio->bi_end_io = end_sync_read;
3168 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3169 if (test_bit(FailFast, &rdev->flags))
3170 bio->bi_opf |= MD_FAILFAST;
3171 from_addr = r10_bio->devs[j].addr;
3172 bio->bi_iter.bi_sector = from_addr +
3174 bio_set_dev(bio, rdev->bdev);
3175 atomic_inc(&rdev->nr_pending);
3176 /* and we write to 'i' (if not in_sync) */
3178 for (k=0; k<conf->copies; k++)
3179 if (r10_bio->devs[k].devnum == i)
3181 BUG_ON(k == conf->copies);
3182 to_addr = r10_bio->devs[k].addr;
3183 r10_bio->devs[0].devnum = d;
3184 r10_bio->devs[0].addr = from_addr;
3185 r10_bio->devs[1].devnum = i;
3186 r10_bio->devs[1].addr = to_addr;
3189 bio = r10_bio->devs[1].bio;
3190 bio->bi_next = biolist;
3192 bio->bi_end_io = end_sync_write;
3193 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3194 bio->bi_iter.bi_sector = to_addr
3195 + mrdev->data_offset;
3196 bio_set_dev(bio, mrdev->bdev);
3197 atomic_inc(&r10_bio->remaining);
3199 r10_bio->devs[1].bio->bi_end_io = NULL;
3201 /* and maybe write to replacement */
3202 bio = r10_bio->devs[1].repl_bio;
3204 bio->bi_end_io = NULL;
3205 /* Note: if need_replace, then bio
3206 * cannot be NULL as r10buf_pool_alloc will
3207 * have allocated it.
3211 bio->bi_next = biolist;
3213 bio->bi_end_io = end_sync_write;
3214 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3215 bio->bi_iter.bi_sector = to_addr +
3216 mreplace->data_offset;
3217 bio_set_dev(bio, mreplace->bdev);
3218 atomic_inc(&r10_bio->remaining);
3222 if (j == conf->copies) {
3223 /* Cannot recover, so abort the recovery or
3224 * record a bad block */
3226 /* problem is that there are bad blocks
3227 * on other device(s)
3230 for (k = 0; k < conf->copies; k++)
3231 if (r10_bio->devs[k].devnum == i)
3233 if (!test_bit(In_sync,
3235 && !rdev_set_badblocks(
3237 r10_bio->devs[k].addr,
3241 !rdev_set_badblocks(
3243 r10_bio->devs[k].addr,
3248 if (!test_and_set_bit(MD_RECOVERY_INTR,
3250 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3252 mirror->recovery_disabled
3253 = mddev->recovery_disabled;
3257 atomic_dec(&rb2->remaining);
3259 rdev_dec_pending(mrdev, mddev);
3261 rdev_dec_pending(mreplace, mddev);
3264 rdev_dec_pending(mrdev, mddev);
3266 rdev_dec_pending(mreplace, mddev);
3267 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3268 /* Only want this if there is elsewhere to
3269 * read from. 'j' is currently the first
3273 for (; j < conf->copies; j++) {
3274 int d = r10_bio->devs[j].devnum;
3275 if (conf->mirrors[d].rdev &&
3277 &conf->mirrors[d].rdev->flags))
3281 r10_bio->devs[0].bio->bi_opf
3285 if (biolist == NULL) {
3287 struct r10bio *rb2 = r10_bio;
3288 r10_bio = (struct r10bio*) rb2->master_bio;
3289 rb2->master_bio = NULL;
3295 /* resync. Schedule a read for every block at this virt offset */
3299 * Since curr_resync_completed could probably not update in
3300 * time, and we will set cluster_sync_low based on it.
3301 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3302 * safety reason, which ensures curr_resync_completed is
3303 * updated in bitmap_cond_end_sync.
3305 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3306 mddev_is_clustered(mddev) &&
3307 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3309 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3310 &sync_blocks, mddev->degraded) &&
3311 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3312 &mddev->recovery)) {
3313 /* We can skip this block */
3315 return sync_blocks + sectors_skipped;
3317 if (sync_blocks < max_sync)
3318 max_sync = sync_blocks;
3319 r10_bio = raid10_alloc_init_r10buf(conf);
3322 r10_bio->mddev = mddev;
3323 atomic_set(&r10_bio->remaining, 0);
3324 raise_barrier(conf, 0);
3325 conf->next_resync = sector_nr;
3327 r10_bio->master_bio = NULL;
3328 r10_bio->sector = sector_nr;
3329 set_bit(R10BIO_IsSync, &r10_bio->state);
3330 raid10_find_phys(conf, r10_bio);
3331 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3333 for (i = 0; i < conf->copies; i++) {
3334 int d = r10_bio->devs[i].devnum;
3335 sector_t first_bad, sector;
3337 struct md_rdev *rdev;
3339 if (r10_bio->devs[i].repl_bio)
3340 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3342 bio = r10_bio->devs[i].bio;
3343 bio->bi_status = BLK_STS_IOERR;
3345 rdev = rcu_dereference(conf->mirrors[d].rdev);
3346 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3350 sector = r10_bio->devs[i].addr;
3351 if (is_badblock(rdev, sector, max_sync,
3352 &first_bad, &bad_sectors)) {
3353 if (first_bad > sector)
3354 max_sync = first_bad - sector;
3356 bad_sectors -= (sector - first_bad);
3357 if (max_sync > bad_sectors)
3358 max_sync = bad_sectors;
3363 atomic_inc(&rdev->nr_pending);
3364 atomic_inc(&r10_bio->remaining);
3365 bio->bi_next = biolist;
3367 bio->bi_end_io = end_sync_read;
3368 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3369 if (test_bit(FailFast, &rdev->flags))
3370 bio->bi_opf |= MD_FAILFAST;
3371 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3372 bio_set_dev(bio, rdev->bdev);
3375 rdev = rcu_dereference(conf->mirrors[d].replacement);
3376 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3380 atomic_inc(&rdev->nr_pending);
3382 /* Need to set up for writing to the replacement */
3383 bio = r10_bio->devs[i].repl_bio;
3384 bio->bi_status = BLK_STS_IOERR;
3386 sector = r10_bio->devs[i].addr;
3387 bio->bi_next = biolist;
3389 bio->bi_end_io = end_sync_write;
3390 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3391 if (test_bit(FailFast, &rdev->flags))
3392 bio->bi_opf |= MD_FAILFAST;
3393 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3394 bio_set_dev(bio, rdev->bdev);
3400 for (i=0; i<conf->copies; i++) {
3401 int d = r10_bio->devs[i].devnum;
3402 if (r10_bio->devs[i].bio->bi_end_io)
3403 rdev_dec_pending(conf->mirrors[d].rdev,
3405 if (r10_bio->devs[i].repl_bio &&
3406 r10_bio->devs[i].repl_bio->bi_end_io)
3408 conf->mirrors[d].replacement,
3418 if (sector_nr + max_sync < max_sector)
3419 max_sector = sector_nr + max_sync;
3422 int len = PAGE_SIZE;
3423 if (sector_nr + (len>>9) > max_sector)
3424 len = (max_sector - sector_nr) << 9;
3427 for (bio= biolist ; bio ; bio=bio->bi_next) {
3428 struct resync_pages *rp = get_resync_pages(bio);
3429 page = resync_fetch_page(rp, page_idx);
3431 * won't fail because the vec table is big enough
3432 * to hold all these pages
3434 bio_add_page(bio, page, len, 0);
3436 nr_sectors += len>>9;
3437 sector_nr += len>>9;
3438 } while (++page_idx < RESYNC_PAGES);
3439 r10_bio->sectors = nr_sectors;
3441 if (mddev_is_clustered(mddev) &&
3442 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3443 /* It is resync not recovery */
3444 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3445 conf->cluster_sync_low = mddev->curr_resync_completed;
3446 raid10_set_cluster_sync_high(conf);
3447 /* Send resync message */
3448 md_cluster_ops->resync_info_update(mddev,
3449 conf->cluster_sync_low,
3450 conf->cluster_sync_high);
3452 } else if (mddev_is_clustered(mddev)) {
3453 /* This is recovery not resync */
3454 sector_t sect_va1, sect_va2;
3455 bool broadcast_msg = false;
3457 for (i = 0; i < conf->geo.raid_disks; i++) {
3459 * sector_nr is a device address for recovery, so we
3460 * need translate it to array address before compare
3461 * with cluster_sync_high.
3463 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3465 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3466 broadcast_msg = true;
3468 * curr_resync_completed is similar as
3469 * sector_nr, so make the translation too.
3471 sect_va2 = raid10_find_virt(conf,
3472 mddev->curr_resync_completed, i);
3474 if (conf->cluster_sync_low == 0 ||
3475 conf->cluster_sync_low > sect_va2)
3476 conf->cluster_sync_low = sect_va2;
3479 if (broadcast_msg) {
3480 raid10_set_cluster_sync_high(conf);
3481 md_cluster_ops->resync_info_update(mddev,
3482 conf->cluster_sync_low,
3483 conf->cluster_sync_high);
3489 biolist = biolist->bi_next;
3491 bio->bi_next = NULL;
3492 r10_bio = get_resync_r10bio(bio);
3493 r10_bio->sectors = nr_sectors;
3495 if (bio->bi_end_io == end_sync_read) {
3496 md_sync_acct_bio(bio, nr_sectors);
3498 generic_make_request(bio);
3502 if (sectors_skipped)
3503 /* pretend they weren't skipped, it makes
3504 * no important difference in this case
3506 md_done_sync(mddev, sectors_skipped, 1);
3508 return sectors_skipped + nr_sectors;
3510 /* There is nowhere to write, so all non-sync
3511 * drives must be failed or in resync, all drives
3512 * have a bad block, so try the next chunk...
3514 if (sector_nr + max_sync < max_sector)
3515 max_sector = sector_nr + max_sync;
3517 sectors_skipped += (max_sector - sector_nr);
3519 sector_nr = max_sector;
3524 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3527 struct r10conf *conf = mddev->private;
3530 raid_disks = min(conf->geo.raid_disks,
3531 conf->prev.raid_disks);
3533 sectors = conf->dev_sectors;
3535 size = sectors >> conf->geo.chunk_shift;
3536 sector_div(size, conf->geo.far_copies);
3537 size = size * raid_disks;
3538 sector_div(size, conf->geo.near_copies);
3540 return size << conf->geo.chunk_shift;
3543 static void calc_sectors(struct r10conf *conf, sector_t size)
3545 /* Calculate the number of sectors-per-device that will
3546 * actually be used, and set conf->dev_sectors and
3550 size = size >> conf->geo.chunk_shift;
3551 sector_div(size, conf->geo.far_copies);
3552 size = size * conf->geo.raid_disks;
3553 sector_div(size, conf->geo.near_copies);
3554 /* 'size' is now the number of chunks in the array */
3555 /* calculate "used chunks per device" */
3556 size = size * conf->copies;
3558 /* We need to round up when dividing by raid_disks to
3559 * get the stride size.
3561 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3563 conf->dev_sectors = size << conf->geo.chunk_shift;
3565 if (conf->geo.far_offset)
3566 conf->geo.stride = 1 << conf->geo.chunk_shift;
3568 sector_div(size, conf->geo.far_copies);
3569 conf->geo.stride = size << conf->geo.chunk_shift;
3573 enum geo_type {geo_new, geo_old, geo_start};
3574 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3577 int layout, chunk, disks;
3580 layout = mddev->layout;
3581 chunk = mddev->chunk_sectors;
3582 disks = mddev->raid_disks - mddev->delta_disks;
3585 layout = mddev->new_layout;
3586 chunk = mddev->new_chunk_sectors;
3587 disks = mddev->raid_disks;
3589 default: /* avoid 'may be unused' warnings */
3590 case geo_start: /* new when starting reshape - raid_disks not
3592 layout = mddev->new_layout;
3593 chunk = mddev->new_chunk_sectors;
3594 disks = mddev->raid_disks + mddev->delta_disks;
3599 if (chunk < (PAGE_SIZE >> 9) ||
3600 !is_power_of_2(chunk))
3603 fc = (layout >> 8) & 255;
3604 fo = layout & (1<<16);
3605 geo->raid_disks = disks;
3606 geo->near_copies = nc;
3607 geo->far_copies = fc;
3608 geo->far_offset = fo;
3609 switch (layout >> 17) {
3610 case 0: /* original layout. simple but not always optimal */
3611 geo->far_set_size = disks;
3613 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3614 * actually using this, but leave code here just in case.*/
3615 geo->far_set_size = disks/fc;
3616 WARN(geo->far_set_size < fc,
3617 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3619 case 2: /* "improved" layout fixed to match documentation */
3620 geo->far_set_size = fc * nc;
3622 default: /* Not a valid layout */
3625 geo->chunk_mask = chunk - 1;
3626 geo->chunk_shift = ffz(~chunk);
3630 static struct r10conf *setup_conf(struct mddev *mddev)
3632 struct r10conf *conf = NULL;
3637 copies = setup_geo(&geo, mddev, geo_new);
3640 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3641 mdname(mddev), PAGE_SIZE);
3645 if (copies < 2 || copies > mddev->raid_disks) {
3646 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3647 mdname(mddev), mddev->new_layout);
3652 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3656 /* FIXME calc properly */
3657 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3658 sizeof(struct raid10_info),
3663 conf->tmppage = alloc_page(GFP_KERNEL);
3668 conf->copies = copies;
3669 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3670 rbio_pool_free, conf);
3674 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3678 calc_sectors(conf, mddev->dev_sectors);
3679 if (mddev->reshape_position == MaxSector) {
3680 conf->prev = conf->geo;
3681 conf->reshape_progress = MaxSector;
3683 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3687 conf->reshape_progress = mddev->reshape_position;
3688 if (conf->prev.far_offset)
3689 conf->prev.stride = 1 << conf->prev.chunk_shift;
3691 /* far_copies must be 1 */
3692 conf->prev.stride = conf->dev_sectors;
3694 conf->reshape_safe = conf->reshape_progress;
3695 spin_lock_init(&conf->device_lock);
3696 INIT_LIST_HEAD(&conf->retry_list);
3697 INIT_LIST_HEAD(&conf->bio_end_io_list);
3699 spin_lock_init(&conf->resync_lock);
3700 init_waitqueue_head(&conf->wait_barrier);
3701 atomic_set(&conf->nr_pending, 0);
3704 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3708 conf->mddev = mddev;
3713 mempool_exit(&conf->r10bio_pool);
3714 kfree(conf->mirrors);
3715 safe_put_page(conf->tmppage);
3716 bioset_exit(&conf->bio_split);
3719 return ERR_PTR(err);
3722 static int raid10_run(struct mddev *mddev)
3724 struct r10conf *conf;
3725 int i, disk_idx, chunk_size;
3726 struct raid10_info *disk;
3727 struct md_rdev *rdev;
3729 sector_t min_offset_diff = 0;
3731 bool discard_supported = false;
3733 if (mddev_init_writes_pending(mddev) < 0)
3736 if (mddev->private == NULL) {
3737 conf = setup_conf(mddev);
3739 return PTR_ERR(conf);
3740 mddev->private = conf;
3742 conf = mddev->private;
3746 if (mddev_is_clustered(conf->mddev)) {
3749 fc = (mddev->layout >> 8) & 255;
3750 fo = mddev->layout & (1<<16);
3751 if (fc > 1 || fo > 0) {
3752 pr_err("only near layout is supported by clustered"
3758 mddev->thread = conf->thread;
3759 conf->thread = NULL;
3761 chunk_size = mddev->chunk_sectors << 9;
3763 blk_queue_max_discard_sectors(mddev->queue,
3764 mddev->chunk_sectors);
3765 blk_queue_max_write_same_sectors(mddev->queue, 0);
3766 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3767 blk_queue_io_min(mddev->queue, chunk_size);
3768 if (conf->geo.raid_disks % conf->geo.near_copies)
3769 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3771 blk_queue_io_opt(mddev->queue, chunk_size *
3772 (conf->geo.raid_disks / conf->geo.near_copies));
3775 rdev_for_each(rdev, mddev) {
3778 disk_idx = rdev->raid_disk;
3781 if (disk_idx >= conf->geo.raid_disks &&
3782 disk_idx >= conf->prev.raid_disks)
3784 disk = conf->mirrors + disk_idx;
3786 if (test_bit(Replacement, &rdev->flags)) {
3787 if (disk->replacement)
3789 disk->replacement = rdev;
3795 diff = (rdev->new_data_offset - rdev->data_offset);
3796 if (!mddev->reshape_backwards)
3800 if (first || diff < min_offset_diff)
3801 min_offset_diff = diff;
3804 disk_stack_limits(mddev->gendisk, rdev->bdev,
3805 rdev->data_offset << 9);
3807 disk->head_position = 0;
3809 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3810 discard_supported = true;
3815 if (discard_supported)
3816 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3819 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3822 /* need to check that every block has at least one working mirror */
3823 if (!enough(conf, -1)) {
3824 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3829 if (conf->reshape_progress != MaxSector) {
3830 /* must ensure that shape change is supported */
3831 if (conf->geo.far_copies != 1 &&
3832 conf->geo.far_offset == 0)
3834 if (conf->prev.far_copies != 1 &&
3835 conf->prev.far_offset == 0)
3839 mddev->degraded = 0;
3841 i < conf->geo.raid_disks
3842 || i < conf->prev.raid_disks;
3845 disk = conf->mirrors + i;
3847 if (!disk->rdev && disk->replacement) {
3848 /* The replacement is all we have - use it */
3849 disk->rdev = disk->replacement;
3850 disk->replacement = NULL;
3851 clear_bit(Replacement, &disk->rdev->flags);
3855 !test_bit(In_sync, &disk->rdev->flags)) {
3856 disk->head_position = 0;
3859 disk->rdev->saved_raid_disk < 0)
3863 if (disk->replacement &&
3864 !test_bit(In_sync, &disk->replacement->flags) &&
3865 disk->replacement->saved_raid_disk < 0) {
3869 disk->recovery_disabled = mddev->recovery_disabled - 1;
3872 if (mddev->recovery_cp != MaxSector)
3873 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3875 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3876 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3877 conf->geo.raid_disks);
3879 * Ok, everything is just fine now
3881 mddev->dev_sectors = conf->dev_sectors;
3882 size = raid10_size(mddev, 0, 0);
3883 md_set_array_sectors(mddev, size);
3884 mddev->resync_max_sectors = size;
3885 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3888 int stripe = conf->geo.raid_disks *
3889 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3891 /* Calculate max read-ahead size.
3892 * We need to readahead at least twice a whole stripe....
3895 stripe /= conf->geo.near_copies;
3896 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3897 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3900 if (md_integrity_register(mddev))
3903 if (conf->reshape_progress != MaxSector) {
3904 unsigned long before_length, after_length;
3906 before_length = ((1 << conf->prev.chunk_shift) *
3907 conf->prev.far_copies);
3908 after_length = ((1 << conf->geo.chunk_shift) *
3909 conf->geo.far_copies);
3911 if (max(before_length, after_length) > min_offset_diff) {
3912 /* This cannot work */
3913 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3916 conf->offset_diff = min_offset_diff;
3918 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3919 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3920 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3921 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3922 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3924 if (!mddev->sync_thread)
3931 md_unregister_thread(&mddev->thread);
3932 mempool_exit(&conf->r10bio_pool);
3933 safe_put_page(conf->tmppage);
3934 kfree(conf->mirrors);
3936 mddev->private = NULL;
3941 static void raid10_free(struct mddev *mddev, void *priv)
3943 struct r10conf *conf = priv;
3945 mempool_exit(&conf->r10bio_pool);
3946 safe_put_page(conf->tmppage);
3947 kfree(conf->mirrors);
3948 kfree(conf->mirrors_old);
3949 kfree(conf->mirrors_new);
3950 bioset_exit(&conf->bio_split);
3954 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3956 struct r10conf *conf = mddev->private;
3959 raise_barrier(conf, 0);
3961 lower_barrier(conf);
3964 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3966 /* Resize of 'far' arrays is not supported.
3967 * For 'near' and 'offset' arrays we can set the
3968 * number of sectors used to be an appropriate multiple
3969 * of the chunk size.
3970 * For 'offset', this is far_copies*chunksize.
3971 * For 'near' the multiplier is the LCM of
3972 * near_copies and raid_disks.
3973 * So if far_copies > 1 && !far_offset, fail.
3974 * Else find LCM(raid_disks, near_copy)*far_copies and
3975 * multiply by chunk_size. Then round to this number.
3976 * This is mostly done by raid10_size()
3978 struct r10conf *conf = mddev->private;
3979 sector_t oldsize, size;
3981 if (mddev->reshape_position != MaxSector)
3984 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3987 oldsize = raid10_size(mddev, 0, 0);
3988 size = raid10_size(mddev, sectors, 0);
3989 if (mddev->external_size &&
3990 mddev->array_sectors > size)
3992 if (mddev->bitmap) {
3993 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3997 md_set_array_sectors(mddev, size);
3998 if (sectors > mddev->dev_sectors &&
3999 mddev->recovery_cp > oldsize) {
4000 mddev->recovery_cp = oldsize;
4001 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4003 calc_sectors(conf, sectors);
4004 mddev->dev_sectors = conf->dev_sectors;
4005 mddev->resync_max_sectors = size;
4009 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4011 struct md_rdev *rdev;
4012 struct r10conf *conf;
4014 if (mddev->degraded > 0) {
4015 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4017 return ERR_PTR(-EINVAL);
4019 sector_div(size, devs);
4021 /* Set new parameters */
4022 mddev->new_level = 10;
4023 /* new layout: far_copies = 1, near_copies = 2 */
4024 mddev->new_layout = (1<<8) + 2;
4025 mddev->new_chunk_sectors = mddev->chunk_sectors;
4026 mddev->delta_disks = mddev->raid_disks;
4027 mddev->raid_disks *= 2;
4028 /* make sure it will be not marked as dirty */
4029 mddev->recovery_cp = MaxSector;
4030 mddev->dev_sectors = size;
4032 conf = setup_conf(mddev);
4033 if (!IS_ERR(conf)) {
4034 rdev_for_each(rdev, mddev)
4035 if (rdev->raid_disk >= 0) {
4036 rdev->new_raid_disk = rdev->raid_disk * 2;
4037 rdev->sectors = size;
4045 static void *raid10_takeover(struct mddev *mddev)
4047 struct r0conf *raid0_conf;
4049 /* raid10 can take over:
4050 * raid0 - providing it has only two drives
4052 if (mddev->level == 0) {
4053 /* for raid0 takeover only one zone is supported */
4054 raid0_conf = mddev->private;
4055 if (raid0_conf->nr_strip_zones > 1) {
4056 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4058 return ERR_PTR(-EINVAL);
4060 return raid10_takeover_raid0(mddev,
4061 raid0_conf->strip_zone->zone_end,
4062 raid0_conf->strip_zone->nb_dev);
4064 return ERR_PTR(-EINVAL);
4067 static int raid10_check_reshape(struct mddev *mddev)
4069 /* Called when there is a request to change
4070 * - layout (to ->new_layout)
4071 * - chunk size (to ->new_chunk_sectors)
4072 * - raid_disks (by delta_disks)
4073 * or when trying to restart a reshape that was ongoing.
4075 * We need to validate the request and possibly allocate
4076 * space if that might be an issue later.
4078 * Currently we reject any reshape of a 'far' mode array,
4079 * allow chunk size to change if new is generally acceptable,
4080 * allow raid_disks to increase, and allow
4081 * a switch between 'near' mode and 'offset' mode.
4083 struct r10conf *conf = mddev->private;
4086 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4089 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4090 /* mustn't change number of copies */
4092 if (geo.far_copies > 1 && !geo.far_offset)
4093 /* Cannot switch to 'far' mode */
4096 if (mddev->array_sectors & geo.chunk_mask)
4097 /* not factor of array size */
4100 if (!enough(conf, -1))
4103 kfree(conf->mirrors_new);
4104 conf->mirrors_new = NULL;
4105 if (mddev->delta_disks > 0) {
4106 /* allocate new 'mirrors' list */
4108 kcalloc(mddev->raid_disks + mddev->delta_disks,
4109 sizeof(struct raid10_info),
4111 if (!conf->mirrors_new)
4118 * Need to check if array has failed when deciding whether to:
4120 * - remove non-faulty devices
4123 * This determination is simple when no reshape is happening.
4124 * However if there is a reshape, we need to carefully check
4125 * both the before and after sections.
4126 * This is because some failed devices may only affect one
4127 * of the two sections, and some non-in_sync devices may
4128 * be insync in the section most affected by failed devices.
4130 static int calc_degraded(struct r10conf *conf)
4132 int degraded, degraded2;
4137 /* 'prev' section first */
4138 for (i = 0; i < conf->prev.raid_disks; i++) {
4139 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4140 if (!rdev || test_bit(Faulty, &rdev->flags))
4142 else if (!test_bit(In_sync, &rdev->flags))
4143 /* When we can reduce the number of devices in
4144 * an array, this might not contribute to
4145 * 'degraded'. It does now.
4150 if (conf->geo.raid_disks == conf->prev.raid_disks)
4154 for (i = 0; i < conf->geo.raid_disks; i++) {
4155 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4156 if (!rdev || test_bit(Faulty, &rdev->flags))
4158 else if (!test_bit(In_sync, &rdev->flags)) {
4159 /* If reshape is increasing the number of devices,
4160 * this section has already been recovered, so
4161 * it doesn't contribute to degraded.
4164 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4169 if (degraded2 > degraded)
4174 static int raid10_start_reshape(struct mddev *mddev)
4176 /* A 'reshape' has been requested. This commits
4177 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4178 * This also checks if there are enough spares and adds them
4180 * We currently require enough spares to make the final
4181 * array non-degraded. We also require that the difference
4182 * between old and new data_offset - on each device - is
4183 * enough that we never risk over-writing.
4186 unsigned long before_length, after_length;
4187 sector_t min_offset_diff = 0;
4190 struct r10conf *conf = mddev->private;
4191 struct md_rdev *rdev;
4195 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4198 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4201 before_length = ((1 << conf->prev.chunk_shift) *
4202 conf->prev.far_copies);
4203 after_length = ((1 << conf->geo.chunk_shift) *
4204 conf->geo.far_copies);
4206 rdev_for_each(rdev, mddev) {
4207 if (!test_bit(In_sync, &rdev->flags)
4208 && !test_bit(Faulty, &rdev->flags))
4210 if (rdev->raid_disk >= 0) {
4211 long long diff = (rdev->new_data_offset
4212 - rdev->data_offset);
4213 if (!mddev->reshape_backwards)
4217 if (first || diff < min_offset_diff)
4218 min_offset_diff = diff;
4223 if (max(before_length, after_length) > min_offset_diff)
4226 if (spares < mddev->delta_disks)
4229 conf->offset_diff = min_offset_diff;
4230 spin_lock_irq(&conf->device_lock);
4231 if (conf->mirrors_new) {
4232 memcpy(conf->mirrors_new, conf->mirrors,
4233 sizeof(struct raid10_info)*conf->prev.raid_disks);
4235 kfree(conf->mirrors_old);
4236 conf->mirrors_old = conf->mirrors;
4237 conf->mirrors = conf->mirrors_new;
4238 conf->mirrors_new = NULL;
4240 setup_geo(&conf->geo, mddev, geo_start);
4242 if (mddev->reshape_backwards) {
4243 sector_t size = raid10_size(mddev, 0, 0);
4244 if (size < mddev->array_sectors) {
4245 spin_unlock_irq(&conf->device_lock);
4246 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4250 mddev->resync_max_sectors = size;
4251 conf->reshape_progress = size;
4253 conf->reshape_progress = 0;
4254 conf->reshape_safe = conf->reshape_progress;
4255 spin_unlock_irq(&conf->device_lock);
4257 if (mddev->delta_disks && mddev->bitmap) {
4258 struct mdp_superblock_1 *sb = NULL;
4259 sector_t oldsize, newsize;
4261 oldsize = raid10_size(mddev, 0, 0);
4262 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4264 if (!mddev_is_clustered(mddev)) {
4265 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4272 rdev_for_each(rdev, mddev) {
4273 if (rdev->raid_disk > -1 &&
4274 !test_bit(Faulty, &rdev->flags))
4275 sb = page_address(rdev->sb_page);
4279 * some node is already performing reshape, and no need to
4280 * call md_bitmap_resize again since it should be called when
4281 * receiving BITMAP_RESIZE msg
4283 if ((sb && (le32_to_cpu(sb->feature_map) &
4284 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4287 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4291 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4293 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4298 if (mddev->delta_disks > 0) {
4299 rdev_for_each(rdev, mddev)
4300 if (rdev->raid_disk < 0 &&
4301 !test_bit(Faulty, &rdev->flags)) {
4302 if (raid10_add_disk(mddev, rdev) == 0) {
4303 if (rdev->raid_disk >=
4304 conf->prev.raid_disks)
4305 set_bit(In_sync, &rdev->flags);
4307 rdev->recovery_offset = 0;
4309 if (sysfs_link_rdev(mddev, rdev))
4310 /* Failure here is OK */;
4312 } else if (rdev->raid_disk >= conf->prev.raid_disks
4313 && !test_bit(Faulty, &rdev->flags)) {
4314 /* This is a spare that was manually added */
4315 set_bit(In_sync, &rdev->flags);
4318 /* When a reshape changes the number of devices,
4319 * ->degraded is measured against the larger of the
4320 * pre and post numbers.
4322 spin_lock_irq(&conf->device_lock);
4323 mddev->degraded = calc_degraded(conf);
4324 spin_unlock_irq(&conf->device_lock);
4325 mddev->raid_disks = conf->geo.raid_disks;
4326 mddev->reshape_position = conf->reshape_progress;
4327 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4329 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4330 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4331 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4332 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4333 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4335 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4337 if (!mddev->sync_thread) {
4341 conf->reshape_checkpoint = jiffies;
4342 md_wakeup_thread(mddev->sync_thread);
4343 md_new_event(mddev);
4347 mddev->recovery = 0;
4348 spin_lock_irq(&conf->device_lock);
4349 conf->geo = conf->prev;
4350 mddev->raid_disks = conf->geo.raid_disks;
4351 rdev_for_each(rdev, mddev)
4352 rdev->new_data_offset = rdev->data_offset;
4354 conf->reshape_progress = MaxSector;
4355 conf->reshape_safe = MaxSector;
4356 mddev->reshape_position = MaxSector;
4357 spin_unlock_irq(&conf->device_lock);
4361 /* Calculate the last device-address that could contain
4362 * any block from the chunk that includes the array-address 's'
4363 * and report the next address.
4364 * i.e. the address returned will be chunk-aligned and after
4365 * any data that is in the chunk containing 's'.
4367 static sector_t last_dev_address(sector_t s, struct geom *geo)
4369 s = (s | geo->chunk_mask) + 1;
4370 s >>= geo->chunk_shift;
4371 s *= geo->near_copies;
4372 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4373 s *= geo->far_copies;
4374 s <<= geo->chunk_shift;
4378 /* Calculate the first device-address that could contain
4379 * any block from the chunk that includes the array-address 's'.
4380 * This too will be the start of a chunk
4382 static sector_t first_dev_address(sector_t s, struct geom *geo)
4384 s >>= geo->chunk_shift;
4385 s *= geo->near_copies;
4386 sector_div(s, geo->raid_disks);
4387 s *= geo->far_copies;
4388 s <<= geo->chunk_shift;
4392 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4395 /* We simply copy at most one chunk (smallest of old and new)
4396 * at a time, possibly less if that exceeds RESYNC_PAGES,
4397 * or we hit a bad block or something.
4398 * This might mean we pause for normal IO in the middle of
4399 * a chunk, but that is not a problem as mddev->reshape_position
4400 * can record any location.
4402 * If we will want to write to a location that isn't
4403 * yet recorded as 'safe' (i.e. in metadata on disk) then
4404 * we need to flush all reshape requests and update the metadata.
4406 * When reshaping forwards (e.g. to more devices), we interpret
4407 * 'safe' as the earliest block which might not have been copied
4408 * down yet. We divide this by previous stripe size and multiply
4409 * by previous stripe length to get lowest device offset that we
4410 * cannot write to yet.
4411 * We interpret 'sector_nr' as an address that we want to write to.
4412 * From this we use last_device_address() to find where we might
4413 * write to, and first_device_address on the 'safe' position.
4414 * If this 'next' write position is after the 'safe' position,
4415 * we must update the metadata to increase the 'safe' position.
4417 * When reshaping backwards, we round in the opposite direction
4418 * and perform the reverse test: next write position must not be
4419 * less than current safe position.
4421 * In all this the minimum difference in data offsets
4422 * (conf->offset_diff - always positive) allows a bit of slack,
4423 * so next can be after 'safe', but not by more than offset_diff
4425 * We need to prepare all the bios here before we start any IO
4426 * to ensure the size we choose is acceptable to all devices.
4427 * The means one for each copy for write-out and an extra one for
4429 * We store the read-in bio in ->master_bio and the others in
4430 * ->devs[x].bio and ->devs[x].repl_bio.
4432 struct r10conf *conf = mddev->private;
4433 struct r10bio *r10_bio;
4434 sector_t next, safe, last;
4438 struct md_rdev *rdev;
4441 struct bio *bio, *read_bio;
4442 int sectors_done = 0;
4443 struct page **pages;
4445 if (sector_nr == 0) {
4446 /* If restarting in the middle, skip the initial sectors */
4447 if (mddev->reshape_backwards &&
4448 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4449 sector_nr = (raid10_size(mddev, 0, 0)
4450 - conf->reshape_progress);
4451 } else if (!mddev->reshape_backwards &&
4452 conf->reshape_progress > 0)
4453 sector_nr = conf->reshape_progress;
4455 mddev->curr_resync_completed = sector_nr;
4456 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4462 /* We don't use sector_nr to track where we are up to
4463 * as that doesn't work well for ->reshape_backwards.
4464 * So just use ->reshape_progress.
4466 if (mddev->reshape_backwards) {
4467 /* 'next' is the earliest device address that we might
4468 * write to for this chunk in the new layout
4470 next = first_dev_address(conf->reshape_progress - 1,
4473 /* 'safe' is the last device address that we might read from
4474 * in the old layout after a restart
4476 safe = last_dev_address(conf->reshape_safe - 1,
4479 if (next + conf->offset_diff < safe)
4482 last = conf->reshape_progress - 1;
4483 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4484 & conf->prev.chunk_mask);
4485 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4486 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4488 /* 'next' is after the last device address that we
4489 * might write to for this chunk in the new layout
4491 next = last_dev_address(conf->reshape_progress, &conf->geo);
4493 /* 'safe' is the earliest device address that we might
4494 * read from in the old layout after a restart
4496 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4498 /* Need to update metadata if 'next' might be beyond 'safe'
4499 * as that would possibly corrupt data
4501 if (next > safe + conf->offset_diff)
4504 sector_nr = conf->reshape_progress;
4505 last = sector_nr | (conf->geo.chunk_mask
4506 & conf->prev.chunk_mask);
4508 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4509 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4513 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4514 /* Need to update reshape_position in metadata */
4516 mddev->reshape_position = conf->reshape_progress;
4517 if (mddev->reshape_backwards)
4518 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4519 - conf->reshape_progress;
4521 mddev->curr_resync_completed = conf->reshape_progress;
4522 conf->reshape_checkpoint = jiffies;
4523 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4524 md_wakeup_thread(mddev->thread);
4525 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4526 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4527 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4528 allow_barrier(conf);
4529 return sectors_done;
4531 conf->reshape_safe = mddev->reshape_position;
4532 allow_barrier(conf);
4535 raise_barrier(conf, 0);
4537 /* Now schedule reads for blocks from sector_nr to last */
4538 r10_bio = raid10_alloc_init_r10buf(conf);
4540 raise_barrier(conf, 1);
4541 atomic_set(&r10_bio->remaining, 0);
4542 r10_bio->mddev = mddev;
4543 r10_bio->sector = sector_nr;
4544 set_bit(R10BIO_IsReshape, &r10_bio->state);
4545 r10_bio->sectors = last - sector_nr + 1;
4546 rdev = read_balance(conf, r10_bio, &max_sectors);
4547 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4550 /* Cannot read from here, so need to record bad blocks
4551 * on all the target devices.
4554 mempool_free(r10_bio, &conf->r10buf_pool);
4555 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4556 return sectors_done;
4559 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4561 bio_set_dev(read_bio, rdev->bdev);
4562 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4563 + rdev->data_offset);
4564 read_bio->bi_private = r10_bio;
4565 read_bio->bi_end_io = end_reshape_read;
4566 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4567 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4568 read_bio->bi_status = 0;
4569 read_bio->bi_vcnt = 0;
4570 read_bio->bi_iter.bi_size = 0;
4571 r10_bio->master_bio = read_bio;
4572 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4575 * Broadcast RESYNC message to other nodes, so all nodes would not
4576 * write to the region to avoid conflict.
4578 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4579 struct mdp_superblock_1 *sb = NULL;
4580 int sb_reshape_pos = 0;
4582 conf->cluster_sync_low = sector_nr;
4583 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4584 sb = page_address(rdev->sb_page);
4586 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4588 * Set cluster_sync_low again if next address for array
4589 * reshape is less than cluster_sync_low. Since we can't
4590 * update cluster_sync_low until it has finished reshape.
4592 if (sb_reshape_pos < conf->cluster_sync_low)
4593 conf->cluster_sync_low = sb_reshape_pos;
4596 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4597 conf->cluster_sync_high);
4600 /* Now find the locations in the new layout */
4601 __raid10_find_phys(&conf->geo, r10_bio);
4604 read_bio->bi_next = NULL;
4607 for (s = 0; s < conf->copies*2; s++) {
4609 int d = r10_bio->devs[s/2].devnum;
4610 struct md_rdev *rdev2;
4612 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4613 b = r10_bio->devs[s/2].repl_bio;
4615 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4616 b = r10_bio->devs[s/2].bio;
4618 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4621 bio_set_dev(b, rdev2->bdev);
4622 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4623 rdev2->new_data_offset;
4624 b->bi_end_io = end_reshape_write;
4625 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4630 /* Now add as many pages as possible to all of these bios. */
4633 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4634 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4635 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4636 int len = (max_sectors - s) << 9;
4637 if (len > PAGE_SIZE)
4639 for (bio = blist; bio ; bio = bio->bi_next) {
4641 * won't fail because the vec table is big enough
4642 * to hold all these pages
4644 bio_add_page(bio, page, len, 0);
4646 sector_nr += len >> 9;
4647 nr_sectors += len >> 9;
4650 r10_bio->sectors = nr_sectors;
4652 /* Now submit the read */
4653 md_sync_acct_bio(read_bio, r10_bio->sectors);
4654 atomic_inc(&r10_bio->remaining);
4655 read_bio->bi_next = NULL;
4656 generic_make_request(read_bio);
4657 sectors_done += nr_sectors;
4658 if (sector_nr <= last)
4661 lower_barrier(conf);
4663 /* Now that we have done the whole section we can
4664 * update reshape_progress
4666 if (mddev->reshape_backwards)
4667 conf->reshape_progress -= sectors_done;
4669 conf->reshape_progress += sectors_done;
4671 return sectors_done;
4674 static void end_reshape_request(struct r10bio *r10_bio);
4675 static int handle_reshape_read_error(struct mddev *mddev,
4676 struct r10bio *r10_bio);
4677 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4679 /* Reshape read completed. Hopefully we have a block
4681 * If we got a read error then we do sync 1-page reads from
4682 * elsewhere until we find the data - or give up.
4684 struct r10conf *conf = mddev->private;
4687 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4688 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4689 /* Reshape has been aborted */
4690 md_done_sync(mddev, r10_bio->sectors, 0);
4694 /* We definitely have the data in the pages, schedule the
4697 atomic_set(&r10_bio->remaining, 1);
4698 for (s = 0; s < conf->copies*2; s++) {
4700 int d = r10_bio->devs[s/2].devnum;
4701 struct md_rdev *rdev;
4704 rdev = rcu_dereference(conf->mirrors[d].replacement);
4705 b = r10_bio->devs[s/2].repl_bio;
4707 rdev = rcu_dereference(conf->mirrors[d].rdev);
4708 b = r10_bio->devs[s/2].bio;
4710 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4714 atomic_inc(&rdev->nr_pending);
4716 md_sync_acct_bio(b, r10_bio->sectors);
4717 atomic_inc(&r10_bio->remaining);
4719 generic_make_request(b);
4721 end_reshape_request(r10_bio);
4724 static void end_reshape(struct r10conf *conf)
4726 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4729 spin_lock_irq(&conf->device_lock);
4730 conf->prev = conf->geo;
4731 md_finish_reshape(conf->mddev);
4733 conf->reshape_progress = MaxSector;
4734 conf->reshape_safe = MaxSector;
4735 spin_unlock_irq(&conf->device_lock);
4737 /* read-ahead size must cover two whole stripes, which is
4738 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4740 if (conf->mddev->queue) {
4741 int stripe = conf->geo.raid_disks *
4742 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4743 stripe /= conf->geo.near_copies;
4744 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4745 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4750 static void raid10_update_reshape_pos(struct mddev *mddev)
4752 struct r10conf *conf = mddev->private;
4755 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4756 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4757 || mddev->reshape_position == MaxSector)
4758 conf->reshape_progress = mddev->reshape_position;
4763 static int handle_reshape_read_error(struct mddev *mddev,
4764 struct r10bio *r10_bio)
4766 /* Use sync reads to get the blocks from somewhere else */
4767 int sectors = r10_bio->sectors;
4768 struct r10conf *conf = mddev->private;
4769 struct r10bio *r10b;
4772 struct page **pages;
4774 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4776 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4780 /* reshape IOs share pages from .devs[0].bio */
4781 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4783 r10b->sector = r10_bio->sector;
4784 __raid10_find_phys(&conf->prev, r10b);
4789 int first_slot = slot;
4791 if (s > (PAGE_SIZE >> 9))
4796 int d = r10b->devs[slot].devnum;
4797 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4800 test_bit(Faulty, &rdev->flags) ||
4801 !test_bit(In_sync, &rdev->flags))
4804 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4805 atomic_inc(&rdev->nr_pending);
4807 success = sync_page_io(rdev,
4811 REQ_OP_READ, 0, false);
4812 rdev_dec_pending(rdev, mddev);
4818 if (slot >= conf->copies)
4820 if (slot == first_slot)
4825 /* couldn't read this block, must give up */
4826 set_bit(MD_RECOVERY_INTR,
4838 static void end_reshape_write(struct bio *bio)
4840 struct r10bio *r10_bio = get_resync_r10bio(bio);
4841 struct mddev *mddev = r10_bio->mddev;
4842 struct r10conf *conf = mddev->private;
4846 struct md_rdev *rdev = NULL;
4848 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4850 rdev = conf->mirrors[d].replacement;
4853 rdev = conf->mirrors[d].rdev;
4856 if (bio->bi_status) {
4857 /* FIXME should record badblock */
4858 md_error(mddev, rdev);
4861 rdev_dec_pending(rdev, mddev);
4862 end_reshape_request(r10_bio);
4865 static void end_reshape_request(struct r10bio *r10_bio)
4867 if (!atomic_dec_and_test(&r10_bio->remaining))
4869 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4870 bio_put(r10_bio->master_bio);
4874 static void raid10_finish_reshape(struct mddev *mddev)
4876 struct r10conf *conf = mddev->private;
4878 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4881 if (mddev->delta_disks > 0) {
4882 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4883 mddev->recovery_cp = mddev->resync_max_sectors;
4884 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4886 mddev->resync_max_sectors = mddev->array_sectors;
4890 for (d = conf->geo.raid_disks ;
4891 d < conf->geo.raid_disks - mddev->delta_disks;
4893 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4895 clear_bit(In_sync, &rdev->flags);
4896 rdev = rcu_dereference(conf->mirrors[d].replacement);
4898 clear_bit(In_sync, &rdev->flags);
4902 mddev->layout = mddev->new_layout;
4903 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4904 mddev->reshape_position = MaxSector;
4905 mddev->delta_disks = 0;
4906 mddev->reshape_backwards = 0;
4909 static struct md_personality raid10_personality =
4913 .owner = THIS_MODULE,
4914 .make_request = raid10_make_request,
4916 .free = raid10_free,
4917 .status = raid10_status,
4918 .error_handler = raid10_error,
4919 .hot_add_disk = raid10_add_disk,
4920 .hot_remove_disk= raid10_remove_disk,
4921 .spare_active = raid10_spare_active,
4922 .sync_request = raid10_sync_request,
4923 .quiesce = raid10_quiesce,
4924 .size = raid10_size,
4925 .resize = raid10_resize,
4926 .takeover = raid10_takeover,
4927 .check_reshape = raid10_check_reshape,
4928 .start_reshape = raid10_start_reshape,
4929 .finish_reshape = raid10_finish_reshape,
4930 .update_reshape_pos = raid10_update_reshape_pos,
4931 .congested = raid10_congested,
4934 static int __init raid_init(void)
4936 return register_md_personality(&raid10_personality);
4939 static void raid_exit(void)
4941 unregister_md_personality(&raid10_personality);
4944 module_init(raid_init);
4945 module_exit(raid_exit);
4946 MODULE_LICENSE("GPL");
4947 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4948 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4949 MODULE_ALIAS("md-raid10");
4950 MODULE_ALIAS("md-level-10");
4952 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);