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->geo.raid_disks]);
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(RESYNC_PAGES, gfp_flags);
151 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
152 r10_bio->devs[j].bio = bio;
153 if (!conf->have_replacement)
155 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
158 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
159 r10_bio->devs[j].repl_bio = bio;
162 * Allocate RESYNC_PAGES data pages and attach them
165 for (j = 0; j < nalloc; j++) {
166 struct bio *rbio = r10_bio->devs[j].repl_bio;
167 struct resync_pages *rp, *rp_repl;
171 rp_repl = &rps[nalloc + j];
173 bio = r10_bio->devs[j].bio;
175 if (!j || test_bit(MD_RECOVERY_SYNC,
176 &conf->mddev->recovery)) {
177 if (resync_alloc_pages(rp, gfp_flags))
180 memcpy(rp, &rps[0], sizeof(*rp));
181 resync_get_all_pages(rp);
184 rp->raid_bio = r10_bio;
185 bio->bi_private = rp;
187 memcpy(rp_repl, rp, sizeof(*rp));
188 rbio->bi_private = rp_repl;
196 resync_free_pages(&rps[j]);
200 for ( ; j < nalloc; j++) {
201 if (r10_bio->devs[j].bio)
202 bio_uninit(r10_bio->devs[j].bio);
203 kfree(r10_bio->devs[j].bio);
204 if (r10_bio->devs[j].repl_bio)
205 bio_uninit(r10_bio->devs[j].repl_bio);
206 kfree(r10_bio->devs[j].repl_bio);
210 rbio_pool_free(r10_bio, conf);
214 static void r10buf_pool_free(void *__r10_bio, void *data)
216 struct r10conf *conf = data;
217 struct r10bio *r10bio = __r10_bio;
219 struct resync_pages *rp = NULL;
221 for (j = conf->copies; j--; ) {
222 struct bio *bio = r10bio->devs[j].bio;
225 rp = get_resync_pages(bio);
226 resync_free_pages(rp);
231 bio = r10bio->devs[j].repl_bio;
238 /* resync pages array stored in the 1st bio's .bi_private */
241 rbio_pool_free(r10bio, conf);
244 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
248 for (i = 0; i < conf->geo.raid_disks; i++) {
249 struct bio **bio = & r10_bio->devs[i].bio;
250 if (!BIO_SPECIAL(*bio))
253 bio = &r10_bio->devs[i].repl_bio;
254 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
260 static void free_r10bio(struct r10bio *r10_bio)
262 struct r10conf *conf = r10_bio->mddev->private;
264 put_all_bios(conf, r10_bio);
265 mempool_free(r10_bio, &conf->r10bio_pool);
268 static void put_buf(struct r10bio *r10_bio)
270 struct r10conf *conf = r10_bio->mddev->private;
272 mempool_free(r10_bio, &conf->r10buf_pool);
277 static void reschedule_retry(struct r10bio *r10_bio)
280 struct mddev *mddev = r10_bio->mddev;
281 struct r10conf *conf = mddev->private;
283 spin_lock_irqsave(&conf->device_lock, flags);
284 list_add(&r10_bio->retry_list, &conf->retry_list);
286 spin_unlock_irqrestore(&conf->device_lock, flags);
288 /* wake up frozen array... */
289 wake_up(&conf->wait_barrier);
291 md_wakeup_thread(mddev->thread);
295 * raid_end_bio_io() is called when we have finished servicing a mirrored
296 * operation and are ready to return a success/failure code to the buffer
299 static void raid_end_bio_io(struct r10bio *r10_bio)
301 struct bio *bio = r10_bio->master_bio;
302 struct r10conf *conf = r10_bio->mddev->private;
304 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
305 bio->bi_status = BLK_STS_IOERR;
307 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
308 bio_end_io_acct(bio, r10_bio->start_time);
311 * Wake up any possible resync thread that waits for the device
316 free_r10bio(r10_bio);
320 * Update disk head position estimator based on IRQ completion info.
322 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
324 struct r10conf *conf = r10_bio->mddev->private;
326 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
327 r10_bio->devs[slot].addr + (r10_bio->sectors);
331 * Find the disk number which triggered given bio
333 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
334 struct bio *bio, int *slotp, int *replp)
339 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
340 if (r10_bio->devs[slot].bio == bio)
342 if (r10_bio->devs[slot].repl_bio == bio) {
348 update_head_pos(slot, r10_bio);
354 return r10_bio->devs[slot].devnum;
357 static void raid10_end_read_request(struct bio *bio)
359 int uptodate = !bio->bi_status;
360 struct r10bio *r10_bio = bio->bi_private;
362 struct md_rdev *rdev;
363 struct r10conf *conf = r10_bio->mddev->private;
365 slot = r10_bio->read_slot;
366 rdev = r10_bio->devs[slot].rdev;
368 * this branch is our 'one mirror IO has finished' event handler:
370 update_head_pos(slot, r10_bio);
374 * Set R10BIO_Uptodate in our master bio, so that
375 * we will return a good error code to the higher
376 * levels even if IO on some other mirrored buffer fails.
378 * The 'master' represents the composite IO operation to
379 * user-side. So if something waits for IO, then it will
380 * wait for the 'master' bio.
382 set_bit(R10BIO_Uptodate, &r10_bio->state);
384 /* If all other devices that store this block have
385 * failed, we want to return the error upwards rather
386 * than fail the last device. Here we redefine
387 * "uptodate" to mean "Don't want to retry"
389 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
394 raid_end_bio_io(r10_bio);
395 rdev_dec_pending(rdev, conf->mddev);
398 * oops, read error - keep the refcount on the rdev
400 char b[BDEVNAME_SIZE];
401 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
403 bdevname(rdev->bdev, b),
404 (unsigned long long)r10_bio->sector);
405 set_bit(R10BIO_ReadError, &r10_bio->state);
406 reschedule_retry(r10_bio);
410 static void close_write(struct r10bio *r10_bio)
412 /* clear the bitmap if all writes complete successfully */
413 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
415 !test_bit(R10BIO_Degraded, &r10_bio->state),
417 md_write_end(r10_bio->mddev);
420 static void one_write_done(struct r10bio *r10_bio)
422 if (atomic_dec_and_test(&r10_bio->remaining)) {
423 if (test_bit(R10BIO_WriteError, &r10_bio->state))
424 reschedule_retry(r10_bio);
426 close_write(r10_bio);
427 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
428 reschedule_retry(r10_bio);
430 raid_end_bio_io(r10_bio);
435 static void raid10_end_write_request(struct bio *bio)
437 struct r10bio *r10_bio = bio->bi_private;
440 struct r10conf *conf = r10_bio->mddev->private;
442 struct md_rdev *rdev = NULL;
443 struct bio *to_put = NULL;
446 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
448 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
451 rdev = conf->mirrors[dev].replacement;
455 rdev = conf->mirrors[dev].rdev;
458 * this branch is our 'one mirror IO has finished' event handler:
460 if (bio->bi_status && !discard_error) {
462 /* Never record new bad blocks to replacement,
465 md_error(rdev->mddev, rdev);
467 set_bit(WriteErrorSeen, &rdev->flags);
468 if (!test_and_set_bit(WantReplacement, &rdev->flags))
469 set_bit(MD_RECOVERY_NEEDED,
470 &rdev->mddev->recovery);
473 if (test_bit(FailFast, &rdev->flags) &&
474 (bio->bi_opf & MD_FAILFAST)) {
475 md_error(rdev->mddev, rdev);
479 * When the device is faulty, it is not necessary to
480 * handle write error.
482 if (!test_bit(Faulty, &rdev->flags))
483 set_bit(R10BIO_WriteError, &r10_bio->state);
485 /* Fail the request */
486 set_bit(R10BIO_Degraded, &r10_bio->state);
487 r10_bio->devs[slot].bio = NULL;
494 * Set R10BIO_Uptodate in our master bio, so that
495 * we will return a good error code for to the higher
496 * levels even if IO on some other mirrored buffer fails.
498 * The 'master' represents the composite IO operation to
499 * user-side. So if something waits for IO, then it will
500 * wait for the 'master' bio.
506 * Do not set R10BIO_Uptodate if the current device is
507 * rebuilding or Faulty. This is because we cannot use
508 * such device for properly reading the data back (we could
509 * potentially use it, if the current write would have felt
510 * before rdev->recovery_offset, but for simplicity we don't
513 if (test_bit(In_sync, &rdev->flags) &&
514 !test_bit(Faulty, &rdev->flags))
515 set_bit(R10BIO_Uptodate, &r10_bio->state);
517 /* Maybe we can clear some bad blocks. */
518 if (is_badblock(rdev,
519 r10_bio->devs[slot].addr,
521 &first_bad, &bad_sectors) && !discard_error) {
524 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
526 r10_bio->devs[slot].bio = IO_MADE_GOOD;
528 set_bit(R10BIO_MadeGood, &r10_bio->state);
534 * Let's see if all mirrored write operations have finished
537 one_write_done(r10_bio);
539 rdev_dec_pending(rdev, conf->mddev);
545 * RAID10 layout manager
546 * As well as the chunksize and raid_disks count, there are two
547 * parameters: near_copies and far_copies.
548 * near_copies * far_copies must be <= raid_disks.
549 * Normally one of these will be 1.
550 * If both are 1, we get raid0.
551 * If near_copies == raid_disks, we get raid1.
553 * Chunks are laid out in raid0 style with near_copies copies of the
554 * first chunk, followed by near_copies copies of the next chunk and
556 * If far_copies > 1, then after 1/far_copies of the array has been assigned
557 * as described above, we start again with a device offset of near_copies.
558 * So we effectively have another copy of the whole array further down all
559 * the drives, but with blocks on different drives.
560 * With this layout, and block is never stored twice on the one device.
562 * raid10_find_phys finds the sector offset of a given virtual sector
563 * on each device that it is on.
565 * raid10_find_virt does the reverse mapping, from a device and a
566 * sector offset to a virtual address
569 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
577 int last_far_set_start, last_far_set_size;
579 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
580 last_far_set_start *= geo->far_set_size;
582 last_far_set_size = geo->far_set_size;
583 last_far_set_size += (geo->raid_disks % geo->far_set_size);
585 /* now calculate first sector/dev */
586 chunk = r10bio->sector >> geo->chunk_shift;
587 sector = r10bio->sector & geo->chunk_mask;
589 chunk *= geo->near_copies;
591 dev = sector_div(stripe, geo->raid_disks);
593 stripe *= geo->far_copies;
595 sector += stripe << geo->chunk_shift;
597 /* and calculate all the others */
598 for (n = 0; n < geo->near_copies; n++) {
602 r10bio->devs[slot].devnum = d;
603 r10bio->devs[slot].addr = s;
606 for (f = 1; f < geo->far_copies; f++) {
607 set = d / geo->far_set_size;
608 d += geo->near_copies;
610 if ((geo->raid_disks % geo->far_set_size) &&
611 (d > last_far_set_start)) {
612 d -= last_far_set_start;
613 d %= last_far_set_size;
614 d += last_far_set_start;
616 d %= geo->far_set_size;
617 d += geo->far_set_size * set;
620 r10bio->devs[slot].devnum = d;
621 r10bio->devs[slot].addr = s;
625 if (dev >= geo->raid_disks) {
627 sector += (geo->chunk_mask + 1);
632 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
634 struct geom *geo = &conf->geo;
636 if (conf->reshape_progress != MaxSector &&
637 ((r10bio->sector >= conf->reshape_progress) !=
638 conf->mddev->reshape_backwards)) {
639 set_bit(R10BIO_Previous, &r10bio->state);
642 clear_bit(R10BIO_Previous, &r10bio->state);
644 __raid10_find_phys(geo, r10bio);
647 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
649 sector_t offset, chunk, vchunk;
650 /* Never use conf->prev as this is only called during resync
651 * or recovery, so reshape isn't happening
653 struct geom *geo = &conf->geo;
654 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
655 int far_set_size = geo->far_set_size;
656 int last_far_set_start;
658 if (geo->raid_disks % geo->far_set_size) {
659 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
660 last_far_set_start *= geo->far_set_size;
662 if (dev >= last_far_set_start) {
663 far_set_size = geo->far_set_size;
664 far_set_size += (geo->raid_disks % geo->far_set_size);
665 far_set_start = last_far_set_start;
669 offset = sector & geo->chunk_mask;
670 if (geo->far_offset) {
672 chunk = sector >> geo->chunk_shift;
673 fc = sector_div(chunk, geo->far_copies);
674 dev -= fc * geo->near_copies;
675 if (dev < far_set_start)
678 while (sector >= geo->stride) {
679 sector -= geo->stride;
680 if (dev < (geo->near_copies + far_set_start))
681 dev += far_set_size - geo->near_copies;
683 dev -= geo->near_copies;
685 chunk = sector >> geo->chunk_shift;
687 vchunk = chunk * geo->raid_disks + dev;
688 sector_div(vchunk, geo->near_copies);
689 return (vchunk << geo->chunk_shift) + offset;
693 * This routine returns the disk from which the requested read should
694 * be done. There is a per-array 'next expected sequential IO' sector
695 * number - if this matches on the next IO then we use the last disk.
696 * There is also a per-disk 'last know head position' sector that is
697 * maintained from IRQ contexts, both the normal and the resync IO
698 * completion handlers update this position correctly. If there is no
699 * perfect sequential match then we pick the disk whose head is closest.
701 * If there are 2 mirrors in the same 2 devices, performance degrades
702 * because position is mirror, not device based.
704 * The rdev for the device selected will have nr_pending incremented.
708 * FIXME: possibly should rethink readbalancing and do it differently
709 * depending on near_copies / far_copies geometry.
711 static struct md_rdev *read_balance(struct r10conf *conf,
712 struct r10bio *r10_bio,
715 const sector_t this_sector = r10_bio->sector;
717 int sectors = r10_bio->sectors;
718 int best_good_sectors;
719 sector_t new_distance, best_dist;
720 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
722 int best_dist_slot, best_pending_slot;
723 bool has_nonrot_disk = false;
724 unsigned int min_pending;
725 struct geom *geo = &conf->geo;
727 raid10_find_phys(conf, r10_bio);
730 min_pending = UINT_MAX;
731 best_dist_rdev = NULL;
732 best_pending_rdev = NULL;
733 best_dist = MaxSector;
734 best_good_sectors = 0;
736 clear_bit(R10BIO_FailFast, &r10_bio->state);
738 * Check if we can balance. We can balance on the whole
739 * device if no resync is going on (recovery is ok), or below
740 * the resync window. We take the first readable disk when
741 * above the resync window.
743 if ((conf->mddev->recovery_cp < MaxSector
744 && (this_sector + sectors >= conf->next_resync)) ||
745 (mddev_is_clustered(conf->mddev) &&
746 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
747 this_sector + sectors)))
750 for (slot = 0; slot < conf->copies ; slot++) {
754 unsigned int pending;
757 if (r10_bio->devs[slot].bio == IO_BLOCKED)
759 disk = r10_bio->devs[slot].devnum;
760 rdev = rcu_dereference(conf->mirrors[disk].replacement);
761 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
762 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
763 rdev = rcu_dereference(conf->mirrors[disk].rdev);
765 test_bit(Faulty, &rdev->flags))
767 if (!test_bit(In_sync, &rdev->flags) &&
768 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
771 dev_sector = r10_bio->devs[slot].addr;
772 if (is_badblock(rdev, dev_sector, sectors,
773 &first_bad, &bad_sectors)) {
774 if (best_dist < MaxSector)
775 /* Already have a better slot */
777 if (first_bad <= dev_sector) {
778 /* Cannot read here. If this is the
779 * 'primary' device, then we must not read
780 * beyond 'bad_sectors' from another device.
782 bad_sectors -= (dev_sector - first_bad);
783 if (!do_balance && sectors > bad_sectors)
784 sectors = bad_sectors;
785 if (best_good_sectors > sectors)
786 best_good_sectors = sectors;
788 sector_t good_sectors =
789 first_bad - dev_sector;
790 if (good_sectors > best_good_sectors) {
791 best_good_sectors = good_sectors;
792 best_dist_slot = slot;
793 best_dist_rdev = rdev;
796 /* Must read from here */
801 best_good_sectors = sectors;
806 nonrot = bdev_nonrot(rdev->bdev);
807 has_nonrot_disk |= nonrot;
808 pending = atomic_read(&rdev->nr_pending);
809 if (min_pending > pending && nonrot) {
810 min_pending = pending;
811 best_pending_slot = slot;
812 best_pending_rdev = rdev;
815 if (best_dist_slot >= 0)
816 /* At least 2 disks to choose from so failfast is OK */
817 set_bit(R10BIO_FailFast, &r10_bio->state);
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
822 if (geo->near_copies > 1 && !pending)
825 /* for far > 1 always use the lowest address */
826 else if (geo->far_copies > 1)
827 new_distance = r10_bio->devs[slot].addr;
829 new_distance = abs(r10_bio->devs[slot].addr -
830 conf->mirrors[disk].head_position);
832 if (new_distance < best_dist) {
833 best_dist = new_distance;
834 best_dist_slot = slot;
835 best_dist_rdev = rdev;
838 if (slot >= conf->copies) {
839 if (has_nonrot_disk) {
840 slot = best_pending_slot;
841 rdev = best_pending_rdev;
843 slot = best_dist_slot;
844 rdev = best_dist_rdev;
849 atomic_inc(&rdev->nr_pending);
850 r10_bio->read_slot = slot;
854 *max_sectors = best_good_sectors;
859 static void flush_pending_writes(struct r10conf *conf)
861 /* Any writes that have been queued but are awaiting
862 * bitmap updates get flushed here.
864 spin_lock_irq(&conf->device_lock);
866 if (conf->pending_bio_list.head) {
867 struct blk_plug plug;
870 bio = bio_list_get(&conf->pending_bio_list);
871 spin_unlock_irq(&conf->device_lock);
874 * As this is called in a wait_event() loop (see freeze_array),
875 * current->state might be TASK_UNINTERRUPTIBLE which will
876 * cause a warning when we prepare to wait again. As it is
877 * rare that this path is taken, it is perfectly safe to force
878 * us to go around the wait_event() loop again, so the warning
879 * is a false-positive. Silence the warning by resetting
882 __set_current_state(TASK_RUNNING);
884 blk_start_plug(&plug);
885 /* flush any pending bitmap writes to disk
886 * before proceeding w/ I/O */
887 md_bitmap_unplug(conf->mddev->bitmap);
888 wake_up(&conf->wait_barrier);
890 while (bio) { /* submit pending writes */
891 struct bio *next = bio->bi_next;
892 struct md_rdev *rdev = (void*)bio->bi_bdev;
894 bio_set_dev(bio, rdev->bdev);
895 if (test_bit(Faulty, &rdev->flags)) {
897 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
898 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
902 submit_bio_noacct(bio);
905 blk_finish_plug(&plug);
907 spin_unlock_irq(&conf->device_lock);
911 * Sometimes we need to suspend IO while we do something else,
912 * either some resync/recovery, or reconfigure the array.
913 * To do this we raise a 'barrier'.
914 * The 'barrier' is a counter that can be raised multiple times
915 * to count how many activities are happening which preclude
917 * We can only raise the barrier if there is no pending IO.
918 * i.e. if nr_pending == 0.
919 * We choose only to raise the barrier if no-one is waiting for the
920 * barrier to go down. This means that as soon as an IO request
921 * is ready, no other operations which require a barrier will start
922 * until the IO request has had a chance.
924 * So: regular IO calls 'wait_barrier'. When that returns there
925 * is no backgroup IO happening, It must arrange to call
926 * allow_barrier when it has finished its IO.
927 * backgroup IO calls must call raise_barrier. Once that returns
928 * there is no normal IO happeing. It must arrange to call
929 * lower_barrier when the particular background IO completes.
932 static void raise_barrier(struct r10conf *conf, int force)
934 BUG_ON(force && !conf->barrier);
935 spin_lock_irq(&conf->resync_lock);
937 /* Wait until no block IO is waiting (unless 'force') */
938 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
941 /* block any new IO from starting */
944 /* Now wait for all pending IO to complete */
945 wait_event_lock_irq(conf->wait_barrier,
946 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
949 spin_unlock_irq(&conf->resync_lock);
952 static void lower_barrier(struct r10conf *conf)
955 spin_lock_irqsave(&conf->resync_lock, flags);
957 spin_unlock_irqrestore(&conf->resync_lock, flags);
958 wake_up(&conf->wait_barrier);
961 static bool wait_barrier(struct r10conf *conf, bool nowait)
965 spin_lock_irq(&conf->resync_lock);
967 struct bio_list *bio_list = current->bio_list;
969 /* Wait for the barrier to drop.
970 * However if there are already pending
971 * requests (preventing the barrier from
972 * rising completely), and the
973 * pre-process bio queue isn't empty,
974 * then don't wait, as we need to empty
975 * that queue to get the nr_pending
978 /* Return false when nowait flag is set */
982 raid10_log(conf->mddev, "wait barrier");
983 wait_event_lock_irq(conf->wait_barrier,
985 (atomic_read(&conf->nr_pending) &&
987 (!bio_list_empty(&bio_list[0]) ||
988 !bio_list_empty(&bio_list[1]))) ||
989 /* move on if recovery thread is
992 (conf->mddev->thread->tsk == current &&
993 test_bit(MD_RECOVERY_RUNNING,
994 &conf->mddev->recovery) &&
995 conf->nr_queued > 0),
999 if (!conf->nr_waiting)
1000 wake_up(&conf->wait_barrier);
1002 /* Only increment nr_pending when we wait */
1004 atomic_inc(&conf->nr_pending);
1005 spin_unlock_irq(&conf->resync_lock);
1009 static void allow_barrier(struct r10conf *conf)
1011 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1012 (conf->array_freeze_pending))
1013 wake_up(&conf->wait_barrier);
1016 static void freeze_array(struct r10conf *conf, int extra)
1018 /* stop syncio and normal IO and wait for everything to
1020 * We increment barrier and nr_waiting, and then
1021 * wait until nr_pending match nr_queued+extra
1022 * This is called in the context of one normal IO request
1023 * that has failed. Thus any sync request that might be pending
1024 * will be blocked by nr_pending, and we need to wait for
1025 * pending IO requests to complete or be queued for re-try.
1026 * Thus the number queued (nr_queued) plus this request (extra)
1027 * must match the number of pending IOs (nr_pending) before
1030 spin_lock_irq(&conf->resync_lock);
1031 conf->array_freeze_pending++;
1034 wait_event_lock_irq_cmd(conf->wait_barrier,
1035 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1037 flush_pending_writes(conf));
1039 conf->array_freeze_pending--;
1040 spin_unlock_irq(&conf->resync_lock);
1043 static void unfreeze_array(struct r10conf *conf)
1045 /* reverse the effect of the freeze */
1046 spin_lock_irq(&conf->resync_lock);
1049 wake_up(&conf->wait_barrier);
1050 spin_unlock_irq(&conf->resync_lock);
1053 static sector_t choose_data_offset(struct r10bio *r10_bio,
1054 struct md_rdev *rdev)
1056 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1057 test_bit(R10BIO_Previous, &r10_bio->state))
1058 return rdev->data_offset;
1060 return rdev->new_data_offset;
1063 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1065 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1066 struct mddev *mddev = plug->cb.data;
1067 struct r10conf *conf = mddev->private;
1070 if (from_schedule || current->bio_list) {
1071 spin_lock_irq(&conf->device_lock);
1072 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1073 spin_unlock_irq(&conf->device_lock);
1074 wake_up(&conf->wait_barrier);
1075 md_wakeup_thread(mddev->thread);
1080 /* we aren't scheduling, so we can do the write-out directly. */
1081 bio = bio_list_get(&plug->pending);
1082 md_bitmap_unplug(mddev->bitmap);
1083 wake_up(&conf->wait_barrier);
1085 while (bio) { /* submit pending writes */
1086 struct bio *next = bio->bi_next;
1087 struct md_rdev *rdev = (void*)bio->bi_bdev;
1088 bio->bi_next = NULL;
1089 bio_set_dev(bio, rdev->bdev);
1090 if (test_bit(Faulty, &rdev->flags)) {
1092 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1093 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
1094 /* Just ignore it */
1097 submit_bio_noacct(bio);
1104 * 1. Register the new request and wait if the reconstruction thread has put
1105 * up a bar for new requests. Continue immediately if no resync is active
1107 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1109 static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1110 struct bio *bio, sector_t sectors)
1112 /* Bail out if REQ_NOWAIT is set for the bio */
1113 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1114 bio_wouldblock_error(bio);
1117 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1118 bio->bi_iter.bi_sector < conf->reshape_progress &&
1119 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1120 allow_barrier(conf);
1121 if (bio->bi_opf & REQ_NOWAIT) {
1122 bio_wouldblock_error(bio);
1125 raid10_log(conf->mddev, "wait reshape");
1126 wait_event(conf->wait_barrier,
1127 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1128 conf->reshape_progress >= bio->bi_iter.bi_sector +
1130 wait_barrier(conf, false);
1135 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1136 struct r10bio *r10_bio)
1138 struct r10conf *conf = mddev->private;
1139 struct bio *read_bio;
1140 const int op = bio_op(bio);
1141 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1143 struct md_rdev *rdev;
1144 char b[BDEVNAME_SIZE];
1145 int slot = r10_bio->read_slot;
1146 struct md_rdev *err_rdev = NULL;
1147 gfp_t gfp = GFP_NOIO;
1149 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1151 * This is an error retry, but we cannot
1152 * safely dereference the rdev in the r10_bio,
1153 * we must use the one in conf.
1154 * If it has already been disconnected (unlikely)
1155 * we lose the device name in error messages.
1159 * As we are blocking raid10, it is a little safer to
1162 gfp = GFP_NOIO | __GFP_HIGH;
1165 disk = r10_bio->devs[slot].devnum;
1166 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1168 bdevname(err_rdev->bdev, b);
1171 /* This never gets dereferenced */
1172 err_rdev = r10_bio->devs[slot].rdev;
1177 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1179 rdev = read_balance(conf, r10_bio, &max_sectors);
1182 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1184 (unsigned long long)r10_bio->sector);
1186 raid_end_bio_io(r10_bio);
1190 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1192 bdevname(rdev->bdev, b),
1193 (unsigned long long)r10_bio->sector);
1194 if (max_sectors < bio_sectors(bio)) {
1195 struct bio *split = bio_split(bio, max_sectors,
1196 gfp, &conf->bio_split);
1197 bio_chain(split, bio);
1198 allow_barrier(conf);
1199 submit_bio_noacct(bio);
1200 wait_barrier(conf, false);
1202 r10_bio->master_bio = bio;
1203 r10_bio->sectors = max_sectors;
1205 slot = r10_bio->read_slot;
1207 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1208 r10_bio->start_time = bio_start_io_acct(bio);
1209 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1211 r10_bio->devs[slot].bio = read_bio;
1212 r10_bio->devs[slot].rdev = rdev;
1214 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1215 choose_data_offset(r10_bio, rdev);
1216 read_bio->bi_end_io = raid10_end_read_request;
1217 bio_set_op_attrs(read_bio, op, do_sync);
1218 if (test_bit(FailFast, &rdev->flags) &&
1219 test_bit(R10BIO_FailFast, &r10_bio->state))
1220 read_bio->bi_opf |= MD_FAILFAST;
1221 read_bio->bi_private = r10_bio;
1224 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1226 submit_bio_noacct(read_bio);
1230 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1231 struct bio *bio, bool replacement,
1234 const int op = bio_op(bio);
1235 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1236 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1237 unsigned long flags;
1238 struct blk_plug_cb *cb;
1239 struct raid1_plug_cb *plug = NULL;
1240 struct r10conf *conf = mddev->private;
1241 struct md_rdev *rdev;
1242 int devnum = r10_bio->devs[n_copy].devnum;
1246 rdev = conf->mirrors[devnum].replacement;
1248 /* Replacement just got moved to main 'rdev' */
1250 rdev = conf->mirrors[devnum].rdev;
1253 rdev = conf->mirrors[devnum].rdev;
1255 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1257 r10_bio->devs[n_copy].repl_bio = mbio;
1259 r10_bio->devs[n_copy].bio = mbio;
1261 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1262 choose_data_offset(r10_bio, rdev));
1263 mbio->bi_end_io = raid10_end_write_request;
1264 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1265 if (!replacement && test_bit(FailFast,
1266 &conf->mirrors[devnum].rdev->flags)
1267 && enough(conf, devnum))
1268 mbio->bi_opf |= MD_FAILFAST;
1269 mbio->bi_private = r10_bio;
1271 if (conf->mddev->gendisk)
1272 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1274 /* flush_pending_writes() needs access to the rdev so...*/
1275 mbio->bi_bdev = (void *)rdev;
1277 atomic_inc(&r10_bio->remaining);
1279 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1281 plug = container_of(cb, struct raid1_plug_cb, cb);
1285 bio_list_add(&plug->pending, mbio);
1287 spin_lock_irqsave(&conf->device_lock, flags);
1288 bio_list_add(&conf->pending_bio_list, mbio);
1289 spin_unlock_irqrestore(&conf->device_lock, flags);
1290 md_wakeup_thread(mddev->thread);
1294 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1297 struct r10conf *conf = mddev->private;
1298 struct md_rdev *blocked_rdev;
1301 blocked_rdev = NULL;
1303 for (i = 0; i < conf->copies; i++) {
1304 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1305 struct md_rdev *rrdev = rcu_dereference(
1306 conf->mirrors[i].replacement);
1309 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1310 atomic_inc(&rdev->nr_pending);
1311 blocked_rdev = rdev;
1314 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1315 atomic_inc(&rrdev->nr_pending);
1316 blocked_rdev = rrdev;
1320 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1322 sector_t dev_sector = r10_bio->devs[i].addr;
1327 * Discard request doesn't care the write result
1328 * so it doesn't need to wait blocked disk here.
1330 if (!r10_bio->sectors)
1333 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1334 &first_bad, &bad_sectors);
1337 * Mustn't write here until the bad block
1340 atomic_inc(&rdev->nr_pending);
1341 set_bit(BlockedBadBlocks, &rdev->flags);
1342 blocked_rdev = rdev;
1349 if (unlikely(blocked_rdev)) {
1350 /* Have to wait for this device to get unblocked, then retry */
1351 allow_barrier(conf);
1352 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1353 __func__, blocked_rdev->raid_disk);
1354 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1355 wait_barrier(conf, false);
1360 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1361 struct r10bio *r10_bio)
1363 struct r10conf *conf = mddev->private;
1368 if ((mddev_is_clustered(mddev) &&
1369 md_cluster_ops->area_resyncing(mddev, WRITE,
1370 bio->bi_iter.bi_sector,
1371 bio_end_sector(bio)))) {
1373 /* Bail out if REQ_NOWAIT is set for the bio */
1374 if (bio->bi_opf & REQ_NOWAIT) {
1375 bio_wouldblock_error(bio);
1379 prepare_to_wait(&conf->wait_barrier,
1381 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1382 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1386 finish_wait(&conf->wait_barrier, &w);
1389 sectors = r10_bio->sectors;
1390 if (!regular_request_wait(mddev, conf, bio, sectors))
1392 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1393 (mddev->reshape_backwards
1394 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1395 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1396 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1397 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1398 /* Need to update reshape_position in metadata */
1399 mddev->reshape_position = conf->reshape_progress;
1400 set_mask_bits(&mddev->sb_flags, 0,
1401 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1402 md_wakeup_thread(mddev->thread);
1403 if (bio->bi_opf & REQ_NOWAIT) {
1404 allow_barrier(conf);
1405 bio_wouldblock_error(bio);
1408 raid10_log(conf->mddev, "wait reshape metadata");
1409 wait_event(mddev->sb_wait,
1410 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1412 conf->reshape_safe = mddev->reshape_position;
1415 /* first select target devices under rcu_lock and
1416 * inc refcount on their rdev. Record them by setting
1418 * If there are known/acknowledged bad blocks on any device
1419 * on which we have seen a write error, we want to avoid
1420 * writing to those blocks. This potentially requires several
1421 * writes to write around the bad blocks. Each set of writes
1422 * gets its own r10_bio with a set of bios attached.
1425 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1426 raid10_find_phys(conf, r10_bio);
1428 wait_blocked_dev(mddev, r10_bio);
1431 max_sectors = r10_bio->sectors;
1433 for (i = 0; i < conf->copies; i++) {
1434 int d = r10_bio->devs[i].devnum;
1435 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1436 struct md_rdev *rrdev = rcu_dereference(
1437 conf->mirrors[d].replacement);
1440 if (rdev && (test_bit(Faulty, &rdev->flags)))
1442 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1445 r10_bio->devs[i].bio = NULL;
1446 r10_bio->devs[i].repl_bio = NULL;
1448 if (!rdev && !rrdev) {
1449 set_bit(R10BIO_Degraded, &r10_bio->state);
1452 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1454 sector_t dev_sector = r10_bio->devs[i].addr;
1458 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1459 &first_bad, &bad_sectors);
1460 if (is_bad && first_bad <= dev_sector) {
1461 /* Cannot write here at all */
1462 bad_sectors -= (dev_sector - first_bad);
1463 if (bad_sectors < max_sectors)
1464 /* Mustn't write more than bad_sectors
1465 * to other devices yet
1467 max_sectors = bad_sectors;
1468 /* We don't set R10BIO_Degraded as that
1469 * only applies if the disk is missing,
1470 * so it might be re-added, and we want to
1471 * know to recover this chunk.
1472 * In this case the device is here, and the
1473 * fact that this chunk is not in-sync is
1474 * recorded in the bad block log.
1479 int good_sectors = first_bad - dev_sector;
1480 if (good_sectors < max_sectors)
1481 max_sectors = good_sectors;
1485 r10_bio->devs[i].bio = bio;
1486 atomic_inc(&rdev->nr_pending);
1489 r10_bio->devs[i].repl_bio = bio;
1490 atomic_inc(&rrdev->nr_pending);
1495 if (max_sectors < r10_bio->sectors)
1496 r10_bio->sectors = max_sectors;
1498 if (r10_bio->sectors < bio_sectors(bio)) {
1499 struct bio *split = bio_split(bio, r10_bio->sectors,
1500 GFP_NOIO, &conf->bio_split);
1501 bio_chain(split, bio);
1502 allow_barrier(conf);
1503 submit_bio_noacct(bio);
1504 wait_barrier(conf, false);
1506 r10_bio->master_bio = bio;
1509 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1510 r10_bio->start_time = bio_start_io_acct(bio);
1511 atomic_set(&r10_bio->remaining, 1);
1512 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1514 for (i = 0; i < conf->copies; i++) {
1515 if (r10_bio->devs[i].bio)
1516 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1517 if (r10_bio->devs[i].repl_bio)
1518 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1520 one_write_done(r10_bio);
1523 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1525 struct r10conf *conf = mddev->private;
1526 struct r10bio *r10_bio;
1528 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1530 r10_bio->master_bio = bio;
1531 r10_bio->sectors = sectors;
1533 r10_bio->mddev = mddev;
1534 r10_bio->sector = bio->bi_iter.bi_sector;
1536 r10_bio->read_slot = -1;
1537 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1538 conf->geo.raid_disks);
1540 if (bio_data_dir(bio) == READ)
1541 raid10_read_request(mddev, bio, r10_bio);
1543 raid10_write_request(mddev, bio, r10_bio);
1546 static void raid_end_discard_bio(struct r10bio *r10bio)
1548 struct r10conf *conf = r10bio->mddev->private;
1549 struct r10bio *first_r10bio;
1551 while (atomic_dec_and_test(&r10bio->remaining)) {
1553 allow_barrier(conf);
1555 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1556 first_r10bio = (struct r10bio *)r10bio->master_bio;
1557 free_r10bio(r10bio);
1558 r10bio = first_r10bio;
1560 md_write_end(r10bio->mddev);
1561 bio_endio(r10bio->master_bio);
1562 free_r10bio(r10bio);
1568 static void raid10_end_discard_request(struct bio *bio)
1570 struct r10bio *r10_bio = bio->bi_private;
1571 struct r10conf *conf = r10_bio->mddev->private;
1572 struct md_rdev *rdev = NULL;
1577 * We don't care the return value of discard bio
1579 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1580 set_bit(R10BIO_Uptodate, &r10_bio->state);
1582 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1584 rdev = conf->mirrors[dev].replacement;
1587 * raid10_remove_disk uses smp_mb to make sure rdev is set to
1588 * replacement before setting replacement to NULL. It can read
1589 * rdev first without barrier protect even replacment is NULL
1592 rdev = conf->mirrors[dev].rdev;
1595 raid_end_discard_bio(r10_bio);
1596 rdev_dec_pending(rdev, conf->mddev);
1600 * There are some limitations to handle discard bio
1601 * 1st, the discard size is bigger than stripe_size*2.
1602 * 2st, if the discard bio spans reshape progress, we use the old way to
1603 * handle discard bio
1605 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1607 struct r10conf *conf = mddev->private;
1608 struct geom *geo = &conf->geo;
1609 int far_copies = geo->far_copies;
1610 bool first_copy = true;
1611 struct r10bio *r10_bio, *first_r10bio;
1615 unsigned int stripe_size;
1616 unsigned int stripe_data_disks;
1617 sector_t split_size;
1618 sector_t bio_start, bio_end;
1619 sector_t first_stripe_index, last_stripe_index;
1620 sector_t start_disk_offset;
1621 unsigned int start_disk_index;
1622 sector_t end_disk_offset;
1623 unsigned int end_disk_index;
1624 unsigned int remainder;
1626 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1629 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1630 bio_wouldblock_error(bio);
1633 wait_barrier(conf, false);
1636 * Check reshape again to avoid reshape happens after checking
1637 * MD_RECOVERY_RESHAPE and before wait_barrier
1639 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1642 if (geo->near_copies)
1643 stripe_data_disks = geo->raid_disks / geo->near_copies +
1644 geo->raid_disks % geo->near_copies;
1646 stripe_data_disks = geo->raid_disks;
1648 stripe_size = stripe_data_disks << geo->chunk_shift;
1650 bio_start = bio->bi_iter.bi_sector;
1651 bio_end = bio_end_sector(bio);
1654 * Maybe one discard bio is smaller than strip size or across one
1655 * stripe and discard region is larger than one stripe size. For far
1656 * offset layout, if the discard region is not aligned with stripe
1657 * size, there is hole when we submit discard bio to member disk.
1658 * For simplicity, we only handle discard bio which discard region
1659 * is bigger than stripe_size * 2
1661 if (bio_sectors(bio) < stripe_size*2)
1665 * Keep bio aligned with strip size.
1667 div_u64_rem(bio_start, stripe_size, &remainder);
1669 split_size = stripe_size - remainder;
1670 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1671 bio_chain(split, bio);
1672 allow_barrier(conf);
1673 /* Resend the fist split part */
1674 submit_bio_noacct(split);
1675 wait_barrier(conf, false);
1677 div_u64_rem(bio_end, stripe_size, &remainder);
1679 split_size = bio_sectors(bio) - remainder;
1680 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1681 bio_chain(split, bio);
1682 allow_barrier(conf);
1683 /* Resend the second split part */
1684 submit_bio_noacct(bio);
1686 wait_barrier(conf, false);
1689 bio_start = bio->bi_iter.bi_sector;
1690 bio_end = bio_end_sector(bio);
1693 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1694 * One stripe contains the chunks from all member disk (one chunk from
1695 * one disk at the same HBA address). For layout detail, see 'man md 4'
1697 chunk = bio_start >> geo->chunk_shift;
1698 chunk *= geo->near_copies;
1699 first_stripe_index = chunk;
1700 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1701 if (geo->far_offset)
1702 first_stripe_index *= geo->far_copies;
1703 start_disk_offset = (bio_start & geo->chunk_mask) +
1704 (first_stripe_index << geo->chunk_shift);
1706 chunk = bio_end >> geo->chunk_shift;
1707 chunk *= geo->near_copies;
1708 last_stripe_index = chunk;
1709 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1710 if (geo->far_offset)
1711 last_stripe_index *= geo->far_copies;
1712 end_disk_offset = (bio_end & geo->chunk_mask) +
1713 (last_stripe_index << geo->chunk_shift);
1716 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1717 r10_bio->mddev = mddev;
1719 r10_bio->sectors = 0;
1720 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1721 wait_blocked_dev(mddev, r10_bio);
1724 * For far layout it needs more than one r10bio to cover all regions.
1725 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1726 * to record the discard bio. Other r10bio->master_bio record the first
1727 * r10bio. The first r10bio only release after all other r10bios finish.
1728 * The discard bio returns only first r10bio finishes
1731 r10_bio->master_bio = bio;
1732 set_bit(R10BIO_Discard, &r10_bio->state);
1734 first_r10bio = r10_bio;
1736 r10_bio->master_bio = (struct bio *)first_r10bio;
1739 * first select target devices under rcu_lock and
1740 * inc refcount on their rdev. Record them by setting
1744 for (disk = 0; disk < geo->raid_disks; disk++) {
1745 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1746 struct md_rdev *rrdev = rcu_dereference(
1747 conf->mirrors[disk].replacement);
1749 r10_bio->devs[disk].bio = NULL;
1750 r10_bio->devs[disk].repl_bio = NULL;
1752 if (rdev && (test_bit(Faulty, &rdev->flags)))
1754 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1756 if (!rdev && !rrdev)
1760 r10_bio->devs[disk].bio = bio;
1761 atomic_inc(&rdev->nr_pending);
1764 r10_bio->devs[disk].repl_bio = bio;
1765 atomic_inc(&rrdev->nr_pending);
1770 atomic_set(&r10_bio->remaining, 1);
1771 for (disk = 0; disk < geo->raid_disks; disk++) {
1772 sector_t dev_start, dev_end;
1773 struct bio *mbio, *rbio = NULL;
1776 * Now start to calculate the start and end address for each disk.
1777 * The space between dev_start and dev_end is the discard region.
1779 * For dev_start, it needs to consider three conditions:
1780 * 1st, the disk is before start_disk, you can imagine the disk in
1781 * the next stripe. So the dev_start is the start address of next
1783 * 2st, the disk is after start_disk, it means the disk is at the
1784 * same stripe of first disk
1785 * 3st, the first disk itself, we can use start_disk_offset directly
1787 if (disk < start_disk_index)
1788 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1789 else if (disk > start_disk_index)
1790 dev_start = first_stripe_index * mddev->chunk_sectors;
1792 dev_start = start_disk_offset;
1794 if (disk < end_disk_index)
1795 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1796 else if (disk > end_disk_index)
1797 dev_end = last_stripe_index * mddev->chunk_sectors;
1799 dev_end = end_disk_offset;
1802 * It only handles discard bio which size is >= stripe size, so
1803 * dev_end > dev_start all the time.
1804 * It doesn't need to use rcu lock to get rdev here. We already
1805 * add rdev->nr_pending in the first loop.
1807 if (r10_bio->devs[disk].bio) {
1808 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1809 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1811 mbio->bi_end_io = raid10_end_discard_request;
1812 mbio->bi_private = r10_bio;
1813 r10_bio->devs[disk].bio = mbio;
1814 r10_bio->devs[disk].devnum = disk;
1815 atomic_inc(&r10_bio->remaining);
1816 md_submit_discard_bio(mddev, rdev, mbio,
1817 dev_start + choose_data_offset(r10_bio, rdev),
1818 dev_end - dev_start);
1821 if (r10_bio->devs[disk].repl_bio) {
1822 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1823 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1825 rbio->bi_end_io = raid10_end_discard_request;
1826 rbio->bi_private = r10_bio;
1827 r10_bio->devs[disk].repl_bio = rbio;
1828 r10_bio->devs[disk].devnum = disk;
1829 atomic_inc(&r10_bio->remaining);
1830 md_submit_discard_bio(mddev, rrdev, rbio,
1831 dev_start + choose_data_offset(r10_bio, rrdev),
1832 dev_end - dev_start);
1837 if (!geo->far_offset && --far_copies) {
1838 first_stripe_index += geo->stride >> geo->chunk_shift;
1839 start_disk_offset += geo->stride;
1840 last_stripe_index += geo->stride >> geo->chunk_shift;
1841 end_disk_offset += geo->stride;
1842 atomic_inc(&first_r10bio->remaining);
1843 raid_end_discard_bio(r10_bio);
1844 wait_barrier(conf, false);
1848 raid_end_discard_bio(r10_bio);
1852 allow_barrier(conf);
1856 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1858 struct r10conf *conf = mddev->private;
1859 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1860 int chunk_sects = chunk_mask + 1;
1861 int sectors = bio_sectors(bio);
1863 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1864 && md_flush_request(mddev, bio))
1867 if (!md_write_start(mddev, bio))
1870 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1871 if (!raid10_handle_discard(mddev, bio))
1875 * If this request crosses a chunk boundary, we need to split
1878 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1879 sectors > chunk_sects
1880 && (conf->geo.near_copies < conf->geo.raid_disks
1881 || conf->prev.near_copies <
1882 conf->prev.raid_disks)))
1883 sectors = chunk_sects -
1884 (bio->bi_iter.bi_sector &
1886 __make_request(mddev, bio, sectors);
1888 /* In case raid10d snuck in to freeze_array */
1889 wake_up(&conf->wait_barrier);
1893 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1895 struct r10conf *conf = mddev->private;
1898 if (conf->geo.near_copies < conf->geo.raid_disks)
1899 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1900 if (conf->geo.near_copies > 1)
1901 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1902 if (conf->geo.far_copies > 1) {
1903 if (conf->geo.far_offset)
1904 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1906 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1907 if (conf->geo.far_set_size != conf->geo.raid_disks)
1908 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1910 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1911 conf->geo.raid_disks - mddev->degraded);
1913 for (i = 0; i < conf->geo.raid_disks; i++) {
1914 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1915 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1918 seq_printf(seq, "]");
1921 /* check if there are enough drives for
1922 * every block to appear on atleast one.
1923 * Don't consider the device numbered 'ignore'
1924 * as we might be about to remove it.
1926 static int _enough(struct r10conf *conf, int previous, int ignore)
1932 disks = conf->prev.raid_disks;
1933 ncopies = conf->prev.near_copies;
1935 disks = conf->geo.raid_disks;
1936 ncopies = conf->geo.near_copies;
1941 int n = conf->copies;
1945 struct md_rdev *rdev;
1946 if (this != ignore &&
1947 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1948 test_bit(In_sync, &rdev->flags))
1950 this = (this+1) % disks;
1954 first = (first + ncopies) % disks;
1955 } while (first != 0);
1962 static int enough(struct r10conf *conf, int ignore)
1964 /* when calling 'enough', both 'prev' and 'geo' must
1966 * This is ensured if ->reconfig_mutex or ->device_lock
1969 return _enough(conf, 0, ignore) &&
1970 _enough(conf, 1, ignore);
1973 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1975 char b[BDEVNAME_SIZE];
1976 struct r10conf *conf = mddev->private;
1977 unsigned long flags;
1980 * If it is not operational, then we have already marked it as dead
1981 * else if it is the last working disks with "fail_last_dev == false",
1982 * ignore the error, let the next level up know.
1983 * else mark the drive as failed
1985 spin_lock_irqsave(&conf->device_lock, flags);
1986 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1987 && !enough(conf, rdev->raid_disk)) {
1989 * Don't fail the drive, just return an IO error.
1991 spin_unlock_irqrestore(&conf->device_lock, flags);
1994 if (test_and_clear_bit(In_sync, &rdev->flags))
1997 * If recovery is running, make sure it aborts.
1999 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2000 set_bit(Blocked, &rdev->flags);
2001 set_bit(Faulty, &rdev->flags);
2002 set_mask_bits(&mddev->sb_flags, 0,
2003 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2004 spin_unlock_irqrestore(&conf->device_lock, flags);
2005 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
2006 "md/raid10:%s: Operation continuing on %d devices.\n",
2007 mdname(mddev), bdevname(rdev->bdev, b),
2008 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2011 static void print_conf(struct r10conf *conf)
2014 struct md_rdev *rdev;
2016 pr_debug("RAID10 conf printout:\n");
2018 pr_debug("(!conf)\n");
2021 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2022 conf->geo.raid_disks);
2024 /* This is only called with ->reconfix_mutex held, so
2025 * rcu protection of rdev is not needed */
2026 for (i = 0; i < conf->geo.raid_disks; i++) {
2027 char b[BDEVNAME_SIZE];
2028 rdev = conf->mirrors[i].rdev;
2030 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
2031 i, !test_bit(In_sync, &rdev->flags),
2032 !test_bit(Faulty, &rdev->flags),
2033 bdevname(rdev->bdev,b));
2037 static void close_sync(struct r10conf *conf)
2039 wait_barrier(conf, false);
2040 allow_barrier(conf);
2042 mempool_exit(&conf->r10buf_pool);
2045 static int raid10_spare_active(struct mddev *mddev)
2048 struct r10conf *conf = mddev->private;
2049 struct raid10_info *tmp;
2051 unsigned long flags;
2054 * Find all non-in_sync disks within the RAID10 configuration
2055 * and mark them in_sync
2057 for (i = 0; i < conf->geo.raid_disks; i++) {
2058 tmp = conf->mirrors + i;
2059 if (tmp->replacement
2060 && tmp->replacement->recovery_offset == MaxSector
2061 && !test_bit(Faulty, &tmp->replacement->flags)
2062 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2063 /* Replacement has just become active */
2065 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2068 /* Replaced device not technically faulty,
2069 * but we need to be sure it gets removed
2070 * and never re-added.
2072 set_bit(Faulty, &tmp->rdev->flags);
2073 sysfs_notify_dirent_safe(
2074 tmp->rdev->sysfs_state);
2076 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2077 } else if (tmp->rdev
2078 && tmp->rdev->recovery_offset == MaxSector
2079 && !test_bit(Faulty, &tmp->rdev->flags)
2080 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2082 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2085 spin_lock_irqsave(&conf->device_lock, flags);
2086 mddev->degraded -= count;
2087 spin_unlock_irqrestore(&conf->device_lock, flags);
2093 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2095 struct r10conf *conf = mddev->private;
2099 int last = conf->geo.raid_disks - 1;
2101 if (mddev->recovery_cp < MaxSector)
2102 /* only hot-add to in-sync arrays, as recovery is
2103 * very different from resync
2106 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2109 if (md_integrity_add_rdev(rdev, mddev))
2112 if (rdev->raid_disk >= 0)
2113 first = last = rdev->raid_disk;
2115 if (rdev->saved_raid_disk >= first &&
2116 rdev->saved_raid_disk < conf->geo.raid_disks &&
2117 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2118 mirror = rdev->saved_raid_disk;
2121 for ( ; mirror <= last ; mirror++) {
2122 struct raid10_info *p = &conf->mirrors[mirror];
2123 if (p->recovery_disabled == mddev->recovery_disabled)
2126 if (!test_bit(WantReplacement, &p->rdev->flags) ||
2127 p->replacement != NULL)
2129 clear_bit(In_sync, &rdev->flags);
2130 set_bit(Replacement, &rdev->flags);
2131 rdev->raid_disk = mirror;
2134 disk_stack_limits(mddev->gendisk, rdev->bdev,
2135 rdev->data_offset << 9);
2137 rcu_assign_pointer(p->replacement, rdev);
2142 disk_stack_limits(mddev->gendisk, rdev->bdev,
2143 rdev->data_offset << 9);
2145 p->head_position = 0;
2146 p->recovery_disabled = mddev->recovery_disabled - 1;
2147 rdev->raid_disk = mirror;
2149 if (rdev->saved_raid_disk != mirror)
2151 rcu_assign_pointer(p->rdev, rdev);
2154 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
2155 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
2161 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2163 struct r10conf *conf = mddev->private;
2165 int number = rdev->raid_disk;
2166 struct md_rdev **rdevp;
2167 struct raid10_info *p = conf->mirrors + number;
2170 if (rdev == p->rdev)
2172 else if (rdev == p->replacement)
2173 rdevp = &p->replacement;
2177 if (test_bit(In_sync, &rdev->flags) ||
2178 atomic_read(&rdev->nr_pending)) {
2182 /* Only remove non-faulty devices if recovery
2185 if (!test_bit(Faulty, &rdev->flags) &&
2186 mddev->recovery_disabled != p->recovery_disabled &&
2187 (!p->replacement || p->replacement == rdev) &&
2188 number < conf->geo.raid_disks &&
2194 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
2196 if (atomic_read(&rdev->nr_pending)) {
2197 /* lost the race, try later */
2203 if (p->replacement) {
2204 /* We must have just cleared 'rdev' */
2205 p->rdev = p->replacement;
2206 clear_bit(Replacement, &p->replacement->flags);
2207 smp_mb(); /* Make sure other CPUs may see both as identical
2208 * but will never see neither -- if they are careful.
2210 p->replacement = NULL;
2213 clear_bit(WantReplacement, &rdev->flags);
2214 err = md_integrity_register(mddev);
2222 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2224 struct r10conf *conf = r10_bio->mddev->private;
2226 if (!bio->bi_status)
2227 set_bit(R10BIO_Uptodate, &r10_bio->state);
2229 /* The write handler will notice the lack of
2230 * R10BIO_Uptodate and record any errors etc
2232 atomic_add(r10_bio->sectors,
2233 &conf->mirrors[d].rdev->corrected_errors);
2235 /* for reconstruct, we always reschedule after a read.
2236 * for resync, only after all reads
2238 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2239 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2240 atomic_dec_and_test(&r10_bio->remaining)) {
2241 /* we have read all the blocks,
2242 * do the comparison in process context in raid10d
2244 reschedule_retry(r10_bio);
2248 static void end_sync_read(struct bio *bio)
2250 struct r10bio *r10_bio = get_resync_r10bio(bio);
2251 struct r10conf *conf = r10_bio->mddev->private;
2252 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2254 __end_sync_read(r10_bio, bio, d);
2257 static void end_reshape_read(struct bio *bio)
2259 /* reshape read bio isn't allocated from r10buf_pool */
2260 struct r10bio *r10_bio = bio->bi_private;
2262 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2265 static void end_sync_request(struct r10bio *r10_bio)
2267 struct mddev *mddev = r10_bio->mddev;
2269 while (atomic_dec_and_test(&r10_bio->remaining)) {
2270 if (r10_bio->master_bio == NULL) {
2271 /* the primary of several recovery bios */
2272 sector_t s = r10_bio->sectors;
2273 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2274 test_bit(R10BIO_WriteError, &r10_bio->state))
2275 reschedule_retry(r10_bio);
2278 md_done_sync(mddev, s, 1);
2281 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2282 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2283 test_bit(R10BIO_WriteError, &r10_bio->state))
2284 reschedule_retry(r10_bio);
2292 static void end_sync_write(struct bio *bio)
2294 struct r10bio *r10_bio = get_resync_r10bio(bio);
2295 struct mddev *mddev = r10_bio->mddev;
2296 struct r10conf *conf = mddev->private;
2302 struct md_rdev *rdev = NULL;
2304 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2306 rdev = conf->mirrors[d].replacement;
2308 rdev = conf->mirrors[d].rdev;
2310 if (bio->bi_status) {
2312 md_error(mddev, rdev);
2314 set_bit(WriteErrorSeen, &rdev->flags);
2315 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2316 set_bit(MD_RECOVERY_NEEDED,
2317 &rdev->mddev->recovery);
2318 set_bit(R10BIO_WriteError, &r10_bio->state);
2320 } else if (is_badblock(rdev,
2321 r10_bio->devs[slot].addr,
2323 &first_bad, &bad_sectors))
2324 set_bit(R10BIO_MadeGood, &r10_bio->state);
2326 rdev_dec_pending(rdev, mddev);
2328 end_sync_request(r10_bio);
2332 * Note: sync and recover and handled very differently for raid10
2333 * This code is for resync.
2334 * For resync, we read through virtual addresses and read all blocks.
2335 * If there is any error, we schedule a write. The lowest numbered
2336 * drive is authoritative.
2337 * However requests come for physical address, so we need to map.
2338 * For every physical address there are raid_disks/copies virtual addresses,
2339 * which is always are least one, but is not necessarly an integer.
2340 * This means that a physical address can span multiple chunks, so we may
2341 * have to submit multiple io requests for a single sync request.
2344 * We check if all blocks are in-sync and only write to blocks that
2347 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2349 struct r10conf *conf = mddev->private;
2351 struct bio *tbio, *fbio;
2353 struct page **tpages, **fpages;
2355 atomic_set(&r10_bio->remaining, 1);
2357 /* find the first device with a block */
2358 for (i=0; i<conf->copies; i++)
2359 if (!r10_bio->devs[i].bio->bi_status)
2362 if (i == conf->copies)
2366 fbio = r10_bio->devs[i].bio;
2367 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2368 fbio->bi_iter.bi_idx = 0;
2369 fpages = get_resync_pages(fbio)->pages;
2371 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2372 /* now find blocks with errors */
2373 for (i=0 ; i < conf->copies ; i++) {
2375 struct md_rdev *rdev;
2376 struct resync_pages *rp;
2378 tbio = r10_bio->devs[i].bio;
2380 if (tbio->bi_end_io != end_sync_read)
2385 tpages = get_resync_pages(tbio)->pages;
2386 d = r10_bio->devs[i].devnum;
2387 rdev = conf->mirrors[d].rdev;
2388 if (!r10_bio->devs[i].bio->bi_status) {
2389 /* We know that the bi_io_vec layout is the same for
2390 * both 'first' and 'i', so we just compare them.
2391 * All vec entries are PAGE_SIZE;
2393 int sectors = r10_bio->sectors;
2394 for (j = 0; j < vcnt; j++) {
2395 int len = PAGE_SIZE;
2396 if (sectors < (len / 512))
2397 len = sectors * 512;
2398 if (memcmp(page_address(fpages[j]),
2399 page_address(tpages[j]),
2406 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2407 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2408 /* Don't fix anything. */
2410 } else if (test_bit(FailFast, &rdev->flags)) {
2411 /* Just give up on this device */
2412 md_error(rdev->mddev, rdev);
2415 /* Ok, we need to write this bio, either to correct an
2416 * inconsistency or to correct an unreadable block.
2417 * First we need to fixup bv_offset, bv_len and
2418 * bi_vecs, as the read request might have corrupted these
2420 rp = get_resync_pages(tbio);
2421 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2423 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2425 rp->raid_bio = r10_bio;
2426 tbio->bi_private = rp;
2427 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2428 tbio->bi_end_io = end_sync_write;
2430 bio_copy_data(tbio, fbio);
2432 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2433 atomic_inc(&r10_bio->remaining);
2434 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2436 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2437 tbio->bi_opf |= MD_FAILFAST;
2438 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2439 submit_bio_noacct(tbio);
2442 /* Now write out to any replacement devices
2445 for (i = 0; i < conf->copies; i++) {
2448 tbio = r10_bio->devs[i].repl_bio;
2449 if (!tbio || !tbio->bi_end_io)
2451 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2452 && r10_bio->devs[i].bio != fbio)
2453 bio_copy_data(tbio, fbio);
2454 d = r10_bio->devs[i].devnum;
2455 atomic_inc(&r10_bio->remaining);
2456 md_sync_acct(conf->mirrors[d].replacement->bdev,
2458 submit_bio_noacct(tbio);
2462 if (atomic_dec_and_test(&r10_bio->remaining)) {
2463 md_done_sync(mddev, r10_bio->sectors, 1);
2469 * Now for the recovery code.
2470 * Recovery happens across physical sectors.
2471 * We recover all non-is_sync drives by finding the virtual address of
2472 * each, and then choose a working drive that also has that virt address.
2473 * There is a separate r10_bio for each non-in_sync drive.
2474 * Only the first two slots are in use. The first for reading,
2475 * The second for writing.
2478 static void fix_recovery_read_error(struct r10bio *r10_bio)
2480 /* We got a read error during recovery.
2481 * We repeat the read in smaller page-sized sections.
2482 * If a read succeeds, write it to the new device or record
2483 * a bad block if we cannot.
2484 * If a read fails, record a bad block on both old and
2487 struct mddev *mddev = r10_bio->mddev;
2488 struct r10conf *conf = mddev->private;
2489 struct bio *bio = r10_bio->devs[0].bio;
2491 int sectors = r10_bio->sectors;
2493 int dr = r10_bio->devs[0].devnum;
2494 int dw = r10_bio->devs[1].devnum;
2495 struct page **pages = get_resync_pages(bio)->pages;
2499 struct md_rdev *rdev;
2503 if (s > (PAGE_SIZE>>9))
2506 rdev = conf->mirrors[dr].rdev;
2507 addr = r10_bio->devs[0].addr + sect,
2508 ok = sync_page_io(rdev,
2512 REQ_OP_READ, 0, false);
2514 rdev = conf->mirrors[dw].rdev;
2515 addr = r10_bio->devs[1].addr + sect;
2516 ok = sync_page_io(rdev,
2520 REQ_OP_WRITE, 0, false);
2522 set_bit(WriteErrorSeen, &rdev->flags);
2523 if (!test_and_set_bit(WantReplacement,
2525 set_bit(MD_RECOVERY_NEEDED,
2526 &rdev->mddev->recovery);
2530 /* We don't worry if we cannot set a bad block -
2531 * it really is bad so there is no loss in not
2534 rdev_set_badblocks(rdev, addr, s, 0);
2536 if (rdev != conf->mirrors[dw].rdev) {
2537 /* need bad block on destination too */
2538 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2539 addr = r10_bio->devs[1].addr + sect;
2540 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2542 /* just abort the recovery */
2543 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2546 conf->mirrors[dw].recovery_disabled
2547 = mddev->recovery_disabled;
2548 set_bit(MD_RECOVERY_INTR,
2561 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2563 struct r10conf *conf = mddev->private;
2565 struct bio *wbio, *wbio2;
2567 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2568 fix_recovery_read_error(r10_bio);
2569 end_sync_request(r10_bio);
2574 * share the pages with the first bio
2575 * and submit the write request
2577 d = r10_bio->devs[1].devnum;
2578 wbio = r10_bio->devs[1].bio;
2579 wbio2 = r10_bio->devs[1].repl_bio;
2580 /* Need to test wbio2->bi_end_io before we call
2581 * submit_bio_noacct as if the former is NULL,
2582 * the latter is free to free wbio2.
2584 if (wbio2 && !wbio2->bi_end_io)
2586 if (wbio->bi_end_io) {
2587 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2588 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2589 submit_bio_noacct(wbio);
2592 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2593 md_sync_acct(conf->mirrors[d].replacement->bdev,
2594 bio_sectors(wbio2));
2595 submit_bio_noacct(wbio2);
2600 * Used by fix_read_error() to decay the per rdev read_errors.
2601 * We halve the read error count for every hour that has elapsed
2602 * since the last recorded read error.
2605 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2608 unsigned long hours_since_last;
2609 unsigned int read_errors = atomic_read(&rdev->read_errors);
2611 cur_time_mon = ktime_get_seconds();
2613 if (rdev->last_read_error == 0) {
2614 /* first time we've seen a read error */
2615 rdev->last_read_error = cur_time_mon;
2619 hours_since_last = (long)(cur_time_mon -
2620 rdev->last_read_error) / 3600;
2622 rdev->last_read_error = cur_time_mon;
2625 * if hours_since_last is > the number of bits in read_errors
2626 * just set read errors to 0. We do this to avoid
2627 * overflowing the shift of read_errors by hours_since_last.
2629 if (hours_since_last >= 8 * sizeof(read_errors))
2630 atomic_set(&rdev->read_errors, 0);
2632 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2635 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2636 int sectors, struct page *page, int rw)
2641 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2642 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2644 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2648 set_bit(WriteErrorSeen, &rdev->flags);
2649 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2650 set_bit(MD_RECOVERY_NEEDED,
2651 &rdev->mddev->recovery);
2653 /* need to record an error - either for the block or the device */
2654 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2655 md_error(rdev->mddev, rdev);
2660 * This is a kernel thread which:
2662 * 1. Retries failed read operations on working mirrors.
2663 * 2. Updates the raid superblock when problems encounter.
2664 * 3. Performs writes following reads for array synchronising.
2667 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2669 int sect = 0; /* Offset from r10_bio->sector */
2670 int sectors = r10_bio->sectors;
2671 struct md_rdev *rdev;
2672 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2673 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2675 /* still own a reference to this rdev, so it cannot
2676 * have been cleared recently.
2678 rdev = conf->mirrors[d].rdev;
2680 if (test_bit(Faulty, &rdev->flags))
2681 /* drive has already been failed, just ignore any
2682 more fix_read_error() attempts */
2685 check_decay_read_errors(mddev, rdev);
2686 atomic_inc(&rdev->read_errors);
2687 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2688 char b[BDEVNAME_SIZE];
2689 bdevname(rdev->bdev, b);
2691 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2693 atomic_read(&rdev->read_errors), max_read_errors);
2694 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2696 md_error(mddev, rdev);
2697 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2703 int sl = r10_bio->read_slot;
2707 if (s > (PAGE_SIZE>>9))
2715 d = r10_bio->devs[sl].devnum;
2716 rdev = rcu_dereference(conf->mirrors[d].rdev);
2718 test_bit(In_sync, &rdev->flags) &&
2719 !test_bit(Faulty, &rdev->flags) &&
2720 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2721 &first_bad, &bad_sectors) == 0) {
2722 atomic_inc(&rdev->nr_pending);
2724 success = sync_page_io(rdev,
2725 r10_bio->devs[sl].addr +
2729 REQ_OP_READ, 0, false);
2730 rdev_dec_pending(rdev, mddev);
2736 if (sl == conf->copies)
2738 } while (!success && sl != r10_bio->read_slot);
2742 /* Cannot read from anywhere, just mark the block
2743 * as bad on the first device to discourage future
2746 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2747 rdev = conf->mirrors[dn].rdev;
2749 if (!rdev_set_badblocks(
2751 r10_bio->devs[r10_bio->read_slot].addr
2754 md_error(mddev, rdev);
2755 r10_bio->devs[r10_bio->read_slot].bio
2762 /* write it back and re-read */
2764 while (sl != r10_bio->read_slot) {
2765 char b[BDEVNAME_SIZE];
2770 d = r10_bio->devs[sl].devnum;
2771 rdev = rcu_dereference(conf->mirrors[d].rdev);
2773 test_bit(Faulty, &rdev->flags) ||
2774 !test_bit(In_sync, &rdev->flags))
2777 atomic_inc(&rdev->nr_pending);
2779 if (r10_sync_page_io(rdev,
2780 r10_bio->devs[sl].addr +
2782 s, conf->tmppage, WRITE)
2784 /* Well, this device is dead */
2785 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2787 (unsigned long long)(
2789 choose_data_offset(r10_bio,
2791 bdevname(rdev->bdev, b));
2792 pr_notice("md/raid10:%s: %s: failing drive\n",
2794 bdevname(rdev->bdev, b));
2796 rdev_dec_pending(rdev, mddev);
2800 while (sl != r10_bio->read_slot) {
2801 char b[BDEVNAME_SIZE];
2806 d = r10_bio->devs[sl].devnum;
2807 rdev = rcu_dereference(conf->mirrors[d].rdev);
2809 test_bit(Faulty, &rdev->flags) ||
2810 !test_bit(In_sync, &rdev->flags))
2813 atomic_inc(&rdev->nr_pending);
2815 switch (r10_sync_page_io(rdev,
2816 r10_bio->devs[sl].addr +
2821 /* Well, this device is dead */
2822 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2824 (unsigned long long)(
2826 choose_data_offset(r10_bio, rdev)),
2827 bdevname(rdev->bdev, b));
2828 pr_notice("md/raid10:%s: %s: failing drive\n",
2830 bdevname(rdev->bdev, b));
2833 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2835 (unsigned long long)(
2837 choose_data_offset(r10_bio, rdev)),
2838 bdevname(rdev->bdev, b));
2839 atomic_add(s, &rdev->corrected_errors);
2842 rdev_dec_pending(rdev, mddev);
2852 static int narrow_write_error(struct r10bio *r10_bio, int i)
2854 struct bio *bio = r10_bio->master_bio;
2855 struct mddev *mddev = r10_bio->mddev;
2856 struct r10conf *conf = mddev->private;
2857 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2858 /* bio has the data to be written to slot 'i' where
2859 * we just recently had a write error.
2860 * We repeatedly clone the bio and trim down to one block,
2861 * then try the write. Where the write fails we record
2863 * It is conceivable that the bio doesn't exactly align with
2864 * blocks. We must handle this.
2866 * We currently own a reference to the rdev.
2872 int sect_to_write = r10_bio->sectors;
2875 if (rdev->badblocks.shift < 0)
2878 block_sectors = roundup(1 << rdev->badblocks.shift,
2879 bdev_logical_block_size(rdev->bdev) >> 9);
2880 sector = r10_bio->sector;
2881 sectors = ((r10_bio->sector + block_sectors)
2882 & ~(sector_t)(block_sectors - 1))
2885 while (sect_to_write) {
2888 if (sectors > sect_to_write)
2889 sectors = sect_to_write;
2890 /* Write at 'sector' for 'sectors' */
2891 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2893 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2894 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2895 wbio->bi_iter.bi_sector = wsector +
2896 choose_data_offset(r10_bio, rdev);
2897 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2899 if (submit_bio_wait(wbio) < 0)
2901 ok = rdev_set_badblocks(rdev, wsector,
2906 sect_to_write -= sectors;
2908 sectors = block_sectors;
2913 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2915 int slot = r10_bio->read_slot;
2917 struct r10conf *conf = mddev->private;
2918 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2920 /* we got a read error. Maybe the drive is bad. Maybe just
2921 * the block and we can fix it.
2922 * We freeze all other IO, and try reading the block from
2923 * other devices. When we find one, we re-write
2924 * and check it that fixes the read error.
2925 * This is all done synchronously while the array is
2928 bio = r10_bio->devs[slot].bio;
2930 r10_bio->devs[slot].bio = NULL;
2933 r10_bio->devs[slot].bio = IO_BLOCKED;
2934 else if (!test_bit(FailFast, &rdev->flags)) {
2935 freeze_array(conf, 1);
2936 fix_read_error(conf, mddev, r10_bio);
2937 unfreeze_array(conf);
2939 md_error(mddev, rdev);
2941 rdev_dec_pending(rdev, mddev);
2942 allow_barrier(conf);
2944 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2947 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2949 /* Some sort of write request has finished and it
2950 * succeeded in writing where we thought there was a
2951 * bad block. So forget the bad block.
2952 * Or possibly if failed and we need to record
2956 struct md_rdev *rdev;
2958 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2959 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2960 for (m = 0; m < conf->copies; m++) {
2961 int dev = r10_bio->devs[m].devnum;
2962 rdev = conf->mirrors[dev].rdev;
2963 if (r10_bio->devs[m].bio == NULL ||
2964 r10_bio->devs[m].bio->bi_end_io == NULL)
2966 if (!r10_bio->devs[m].bio->bi_status) {
2967 rdev_clear_badblocks(
2969 r10_bio->devs[m].addr,
2970 r10_bio->sectors, 0);
2972 if (!rdev_set_badblocks(
2974 r10_bio->devs[m].addr,
2975 r10_bio->sectors, 0))
2976 md_error(conf->mddev, rdev);
2978 rdev = conf->mirrors[dev].replacement;
2979 if (r10_bio->devs[m].repl_bio == NULL ||
2980 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2983 if (!r10_bio->devs[m].repl_bio->bi_status) {
2984 rdev_clear_badblocks(
2986 r10_bio->devs[m].addr,
2987 r10_bio->sectors, 0);
2989 if (!rdev_set_badblocks(
2991 r10_bio->devs[m].addr,
2992 r10_bio->sectors, 0))
2993 md_error(conf->mddev, rdev);
2999 for (m = 0; m < conf->copies; m++) {
3000 int dev = r10_bio->devs[m].devnum;
3001 struct bio *bio = r10_bio->devs[m].bio;
3002 rdev = conf->mirrors[dev].rdev;
3003 if (bio == IO_MADE_GOOD) {
3004 rdev_clear_badblocks(
3006 r10_bio->devs[m].addr,
3007 r10_bio->sectors, 0);
3008 rdev_dec_pending(rdev, conf->mddev);
3009 } else if (bio != NULL && bio->bi_status) {
3011 if (!narrow_write_error(r10_bio, m)) {
3012 md_error(conf->mddev, rdev);
3013 set_bit(R10BIO_Degraded,
3016 rdev_dec_pending(rdev, conf->mddev);
3018 bio = r10_bio->devs[m].repl_bio;
3019 rdev = conf->mirrors[dev].replacement;
3020 if (rdev && bio == IO_MADE_GOOD) {
3021 rdev_clear_badblocks(
3023 r10_bio->devs[m].addr,
3024 r10_bio->sectors, 0);
3025 rdev_dec_pending(rdev, conf->mddev);
3029 spin_lock_irq(&conf->device_lock);
3030 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
3032 spin_unlock_irq(&conf->device_lock);
3034 * In case freeze_array() is waiting for condition
3035 * nr_pending == nr_queued + extra to be true.
3037 wake_up(&conf->wait_barrier);
3038 md_wakeup_thread(conf->mddev->thread);
3040 if (test_bit(R10BIO_WriteError,
3042 close_write(r10_bio);
3043 raid_end_bio_io(r10_bio);
3048 static void raid10d(struct md_thread *thread)
3050 struct mddev *mddev = thread->mddev;
3051 struct r10bio *r10_bio;
3052 unsigned long flags;
3053 struct r10conf *conf = mddev->private;
3054 struct list_head *head = &conf->retry_list;
3055 struct blk_plug plug;
3057 md_check_recovery(mddev);
3059 if (!list_empty_careful(&conf->bio_end_io_list) &&
3060 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3062 spin_lock_irqsave(&conf->device_lock, flags);
3063 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3064 while (!list_empty(&conf->bio_end_io_list)) {
3065 list_move(conf->bio_end_io_list.prev, &tmp);
3069 spin_unlock_irqrestore(&conf->device_lock, flags);
3070 while (!list_empty(&tmp)) {
3071 r10_bio = list_first_entry(&tmp, struct r10bio,
3073 list_del(&r10_bio->retry_list);
3074 if (mddev->degraded)
3075 set_bit(R10BIO_Degraded, &r10_bio->state);
3077 if (test_bit(R10BIO_WriteError,
3079 close_write(r10_bio);
3080 raid_end_bio_io(r10_bio);
3084 blk_start_plug(&plug);
3087 flush_pending_writes(conf);
3089 spin_lock_irqsave(&conf->device_lock, flags);
3090 if (list_empty(head)) {
3091 spin_unlock_irqrestore(&conf->device_lock, flags);
3094 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3095 list_del(head->prev);
3097 spin_unlock_irqrestore(&conf->device_lock, flags);
3099 mddev = r10_bio->mddev;
3100 conf = mddev->private;
3101 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3102 test_bit(R10BIO_WriteError, &r10_bio->state))
3103 handle_write_completed(conf, r10_bio);
3104 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3105 reshape_request_write(mddev, r10_bio);
3106 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3107 sync_request_write(mddev, r10_bio);
3108 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3109 recovery_request_write(mddev, r10_bio);
3110 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3111 handle_read_error(mddev, r10_bio);
3116 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3117 md_check_recovery(mddev);
3119 blk_finish_plug(&plug);
3122 static int init_resync(struct r10conf *conf)
3126 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3127 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3128 conf->have_replacement = 0;
3129 for (i = 0; i < conf->geo.raid_disks; i++)
3130 if (conf->mirrors[i].replacement)
3131 conf->have_replacement = 1;
3132 ret = mempool_init(&conf->r10buf_pool, buffs,
3133 r10buf_pool_alloc, r10buf_pool_free, conf);
3136 conf->next_resync = 0;
3140 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3142 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3143 struct rsync_pages *rp;
3148 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3149 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3150 nalloc = conf->copies; /* resync */
3152 nalloc = 2; /* recovery */
3154 for (i = 0; i < nalloc; i++) {
3155 bio = r10bio->devs[i].bio;
3156 rp = bio->bi_private;
3157 bio_reset(bio, NULL, 0);
3158 bio->bi_private = rp;
3159 bio = r10bio->devs[i].repl_bio;
3161 rp = bio->bi_private;
3162 bio_reset(bio, NULL, 0);
3163 bio->bi_private = rp;
3170 * Set cluster_sync_high since we need other nodes to add the
3171 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3173 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3175 sector_t window_size;
3176 int extra_chunk, chunks;
3179 * First, here we define "stripe" as a unit which across
3180 * all member devices one time, so we get chunks by use
3181 * raid_disks / near_copies. Otherwise, if near_copies is
3182 * close to raid_disks, then resync window could increases
3183 * linearly with the increase of raid_disks, which means
3184 * we will suspend a really large IO window while it is not
3185 * necessary. If raid_disks is not divisible by near_copies,
3186 * an extra chunk is needed to ensure the whole "stripe" is
3190 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3191 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3195 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3198 * At least use a 32M window to align with raid1's resync window
3200 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3201 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3203 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3207 * perform a "sync" on one "block"
3209 * We need to make sure that no normal I/O request - particularly write
3210 * requests - conflict with active sync requests.
3212 * This is achieved by tracking pending requests and a 'barrier' concept
3213 * that can be installed to exclude normal IO requests.
3215 * Resync and recovery are handled very differently.
3216 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3218 * For resync, we iterate over virtual addresses, read all copies,
3219 * and update if there are differences. If only one copy is live,
3221 * For recovery, we iterate over physical addresses, read a good
3222 * value for each non-in_sync drive, and over-write.
3224 * So, for recovery we may have several outstanding complex requests for a
3225 * given address, one for each out-of-sync device. We model this by allocating
3226 * a number of r10_bio structures, one for each out-of-sync device.
3227 * As we setup these structures, we collect all bio's together into a list
3228 * which we then process collectively to add pages, and then process again
3229 * to pass to submit_bio_noacct.
3231 * The r10_bio structures are linked using a borrowed master_bio pointer.
3232 * This link is counted in ->remaining. When the r10_bio that points to NULL
3233 * has its remaining count decremented to 0, the whole complex operation
3238 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3241 struct r10conf *conf = mddev->private;
3242 struct r10bio *r10_bio;
3243 struct bio *biolist = NULL, *bio;
3244 sector_t max_sector, nr_sectors;
3247 sector_t sync_blocks;
3248 sector_t sectors_skipped = 0;
3249 int chunks_skipped = 0;
3250 sector_t chunk_mask = conf->geo.chunk_mask;
3253 if (!mempool_initialized(&conf->r10buf_pool))
3254 if (init_resync(conf))
3258 * Allow skipping a full rebuild for incremental assembly
3259 * of a clean array, like RAID1 does.
3261 if (mddev->bitmap == NULL &&
3262 mddev->recovery_cp == MaxSector &&
3263 mddev->reshape_position == MaxSector &&
3264 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3265 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3266 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3267 conf->fullsync == 0) {
3269 return mddev->dev_sectors - sector_nr;
3273 max_sector = mddev->dev_sectors;
3274 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3275 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3276 max_sector = mddev->resync_max_sectors;
3277 if (sector_nr >= max_sector) {
3278 conf->cluster_sync_low = 0;
3279 conf->cluster_sync_high = 0;
3281 /* If we aborted, we need to abort the
3282 * sync on the 'current' bitmap chucks (there can
3283 * be several when recovering multiple devices).
3284 * as we may have started syncing it but not finished.
3285 * We can find the current address in
3286 * mddev->curr_resync, but for recovery,
3287 * we need to convert that to several
3288 * virtual addresses.
3290 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3296 if (mddev->curr_resync < max_sector) { /* aborted */
3297 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3298 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3300 else for (i = 0; i < conf->geo.raid_disks; i++) {
3302 raid10_find_virt(conf, mddev->curr_resync, i);
3303 md_bitmap_end_sync(mddev->bitmap, sect,
3307 /* completed sync */
3308 if ((!mddev->bitmap || conf->fullsync)
3309 && conf->have_replacement
3310 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3311 /* Completed a full sync so the replacements
3312 * are now fully recovered.
3315 for (i = 0; i < conf->geo.raid_disks; i++) {
3316 struct md_rdev *rdev =
3317 rcu_dereference(conf->mirrors[i].replacement);
3319 rdev->recovery_offset = MaxSector;
3325 md_bitmap_close_sync(mddev->bitmap);
3328 return sectors_skipped;
3331 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3332 return reshape_request(mddev, sector_nr, skipped);
3334 if (chunks_skipped >= conf->geo.raid_disks) {
3335 /* if there has been nothing to do on any drive,
3336 * then there is nothing to do at all..
3339 return (max_sector - sector_nr) + sectors_skipped;
3342 if (max_sector > mddev->resync_max)
3343 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3345 /* make sure whole request will fit in a chunk - if chunks
3348 if (conf->geo.near_copies < conf->geo.raid_disks &&
3349 max_sector > (sector_nr | chunk_mask))
3350 max_sector = (sector_nr | chunk_mask) + 1;
3353 * If there is non-resync activity waiting for a turn, then let it
3354 * though before starting on this new sync request.
3356 if (conf->nr_waiting)
3357 schedule_timeout_uninterruptible(1);
3359 /* Again, very different code for resync and recovery.
3360 * Both must result in an r10bio with a list of bios that
3361 * have bi_end_io, bi_sector, bi_bdev set,
3362 * and bi_private set to the r10bio.
3363 * For recovery, we may actually create several r10bios
3364 * with 2 bios in each, that correspond to the bios in the main one.
3365 * In this case, the subordinate r10bios link back through a
3366 * borrowed master_bio pointer, and the counter in the master
3367 * includes a ref from each subordinate.
3369 /* First, we decide what to do and set ->bi_end_io
3370 * To end_sync_read if we want to read, and
3371 * end_sync_write if we will want to write.
3374 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3375 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3376 /* recovery... the complicated one */
3380 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3386 int need_recover = 0;
3387 int need_replace = 0;
3388 struct raid10_info *mirror = &conf->mirrors[i];
3389 struct md_rdev *mrdev, *mreplace;
3392 mrdev = rcu_dereference(mirror->rdev);
3393 mreplace = rcu_dereference(mirror->replacement);
3395 if (mrdev != NULL &&
3396 !test_bit(Faulty, &mrdev->flags) &&
3397 !test_bit(In_sync, &mrdev->flags))
3399 if (mreplace != NULL &&
3400 !test_bit(Faulty, &mreplace->flags))
3403 if (!need_recover && !need_replace) {
3409 /* want to reconstruct this device */
3411 sect = raid10_find_virt(conf, sector_nr, i);
3412 if (sect >= mddev->resync_max_sectors) {
3413 /* last stripe is not complete - don't
3414 * try to recover this sector.
3419 if (mreplace && test_bit(Faulty, &mreplace->flags))
3421 /* Unless we are doing a full sync, or a replacement
3422 * we only need to recover the block if it is set in
3425 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3427 if (sync_blocks < max_sync)
3428 max_sync = sync_blocks;
3432 /* yep, skip the sync_blocks here, but don't assume
3433 * that there will never be anything to do here
3435 chunks_skipped = -1;
3439 atomic_inc(&mrdev->nr_pending);
3441 atomic_inc(&mreplace->nr_pending);
3444 r10_bio = raid10_alloc_init_r10buf(conf);
3446 raise_barrier(conf, rb2 != NULL);
3447 atomic_set(&r10_bio->remaining, 0);
3449 r10_bio->master_bio = (struct bio*)rb2;
3451 atomic_inc(&rb2->remaining);
3452 r10_bio->mddev = mddev;
3453 set_bit(R10BIO_IsRecover, &r10_bio->state);
3454 r10_bio->sector = sect;
3456 raid10_find_phys(conf, r10_bio);
3458 /* Need to check if the array will still be
3462 for (j = 0; j < conf->geo.raid_disks; j++) {
3463 struct md_rdev *rdev = rcu_dereference(
3464 conf->mirrors[j].rdev);
3465 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3471 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3472 &sync_blocks, still_degraded);
3475 for (j=0; j<conf->copies;j++) {
3477 int d = r10_bio->devs[j].devnum;
3478 sector_t from_addr, to_addr;
3479 struct md_rdev *rdev =
3480 rcu_dereference(conf->mirrors[d].rdev);
3481 sector_t sector, first_bad;
3484 !test_bit(In_sync, &rdev->flags))
3486 /* This is where we read from */
3488 sector = r10_bio->devs[j].addr;
3490 if (is_badblock(rdev, sector, max_sync,
3491 &first_bad, &bad_sectors)) {
3492 if (first_bad > sector)
3493 max_sync = first_bad - sector;
3495 bad_sectors -= (sector
3497 if (max_sync > bad_sectors)
3498 max_sync = bad_sectors;
3502 bio = r10_bio->devs[0].bio;
3503 bio->bi_next = biolist;
3505 bio->bi_end_io = end_sync_read;
3506 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3507 if (test_bit(FailFast, &rdev->flags))
3508 bio->bi_opf |= MD_FAILFAST;
3509 from_addr = r10_bio->devs[j].addr;
3510 bio->bi_iter.bi_sector = from_addr +
3512 bio_set_dev(bio, rdev->bdev);
3513 atomic_inc(&rdev->nr_pending);
3514 /* and we write to 'i' (if not in_sync) */
3516 for (k=0; k<conf->copies; k++)
3517 if (r10_bio->devs[k].devnum == i)
3519 BUG_ON(k == conf->copies);
3520 to_addr = r10_bio->devs[k].addr;
3521 r10_bio->devs[0].devnum = d;
3522 r10_bio->devs[0].addr = from_addr;
3523 r10_bio->devs[1].devnum = i;
3524 r10_bio->devs[1].addr = to_addr;
3527 bio = r10_bio->devs[1].bio;
3528 bio->bi_next = biolist;
3530 bio->bi_end_io = end_sync_write;
3531 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3532 bio->bi_iter.bi_sector = to_addr
3533 + mrdev->data_offset;
3534 bio_set_dev(bio, mrdev->bdev);
3535 atomic_inc(&r10_bio->remaining);
3537 r10_bio->devs[1].bio->bi_end_io = NULL;
3539 /* and maybe write to replacement */
3540 bio = r10_bio->devs[1].repl_bio;
3542 bio->bi_end_io = NULL;
3543 /* Note: if need_replace, then bio
3544 * cannot be NULL as r10buf_pool_alloc will
3545 * have allocated it.
3549 bio->bi_next = biolist;
3551 bio->bi_end_io = end_sync_write;
3552 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3553 bio->bi_iter.bi_sector = to_addr +
3554 mreplace->data_offset;
3555 bio_set_dev(bio, mreplace->bdev);
3556 atomic_inc(&r10_bio->remaining);
3560 if (j == conf->copies) {
3561 /* Cannot recover, so abort the recovery or
3562 * record a bad block */
3564 /* problem is that there are bad blocks
3565 * on other device(s)
3568 for (k = 0; k < conf->copies; k++)
3569 if (r10_bio->devs[k].devnum == i)
3571 if (!test_bit(In_sync,
3573 && !rdev_set_badblocks(
3575 r10_bio->devs[k].addr,
3579 !rdev_set_badblocks(
3581 r10_bio->devs[k].addr,
3586 if (!test_and_set_bit(MD_RECOVERY_INTR,
3588 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3590 mirror->recovery_disabled
3591 = mddev->recovery_disabled;
3595 atomic_dec(&rb2->remaining);
3597 rdev_dec_pending(mrdev, mddev);
3599 rdev_dec_pending(mreplace, mddev);
3602 rdev_dec_pending(mrdev, mddev);
3604 rdev_dec_pending(mreplace, mddev);
3605 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3606 /* Only want this if there is elsewhere to
3607 * read from. 'j' is currently the first
3611 for (; j < conf->copies; j++) {
3612 int d = r10_bio->devs[j].devnum;
3613 if (conf->mirrors[d].rdev &&
3615 &conf->mirrors[d].rdev->flags))
3619 r10_bio->devs[0].bio->bi_opf
3623 if (biolist == NULL) {
3625 struct r10bio *rb2 = r10_bio;
3626 r10_bio = (struct r10bio*) rb2->master_bio;
3627 rb2->master_bio = NULL;
3633 /* resync. Schedule a read for every block at this virt offset */
3637 * Since curr_resync_completed could probably not update in
3638 * time, and we will set cluster_sync_low based on it.
3639 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3640 * safety reason, which ensures curr_resync_completed is
3641 * updated in bitmap_cond_end_sync.
3643 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3644 mddev_is_clustered(mddev) &&
3645 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3647 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3648 &sync_blocks, mddev->degraded) &&
3649 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3650 &mddev->recovery)) {
3651 /* We can skip this block */
3653 return sync_blocks + sectors_skipped;
3655 if (sync_blocks < max_sync)
3656 max_sync = sync_blocks;
3657 r10_bio = raid10_alloc_init_r10buf(conf);
3660 r10_bio->mddev = mddev;
3661 atomic_set(&r10_bio->remaining, 0);
3662 raise_barrier(conf, 0);
3663 conf->next_resync = sector_nr;
3665 r10_bio->master_bio = NULL;
3666 r10_bio->sector = sector_nr;
3667 set_bit(R10BIO_IsSync, &r10_bio->state);
3668 raid10_find_phys(conf, r10_bio);
3669 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3671 for (i = 0; i < conf->copies; i++) {
3672 int d = r10_bio->devs[i].devnum;
3673 sector_t first_bad, sector;
3675 struct md_rdev *rdev;
3677 if (r10_bio->devs[i].repl_bio)
3678 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3680 bio = r10_bio->devs[i].bio;
3681 bio->bi_status = BLK_STS_IOERR;
3683 rdev = rcu_dereference(conf->mirrors[d].rdev);
3684 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3688 sector = r10_bio->devs[i].addr;
3689 if (is_badblock(rdev, sector, max_sync,
3690 &first_bad, &bad_sectors)) {
3691 if (first_bad > sector)
3692 max_sync = first_bad - sector;
3694 bad_sectors -= (sector - first_bad);
3695 if (max_sync > bad_sectors)
3696 max_sync = bad_sectors;
3701 atomic_inc(&rdev->nr_pending);
3702 atomic_inc(&r10_bio->remaining);
3703 bio->bi_next = biolist;
3705 bio->bi_end_io = end_sync_read;
3706 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3707 if (test_bit(FailFast, &rdev->flags))
3708 bio->bi_opf |= MD_FAILFAST;
3709 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3710 bio_set_dev(bio, rdev->bdev);
3713 rdev = rcu_dereference(conf->mirrors[d].replacement);
3714 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3718 atomic_inc(&rdev->nr_pending);
3720 /* Need to set up for writing to the replacement */
3721 bio = r10_bio->devs[i].repl_bio;
3722 bio->bi_status = BLK_STS_IOERR;
3724 sector = r10_bio->devs[i].addr;
3725 bio->bi_next = biolist;
3727 bio->bi_end_io = end_sync_write;
3728 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3729 if (test_bit(FailFast, &rdev->flags))
3730 bio->bi_opf |= MD_FAILFAST;
3731 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3732 bio_set_dev(bio, rdev->bdev);
3738 for (i=0; i<conf->copies; i++) {
3739 int d = r10_bio->devs[i].devnum;
3740 if (r10_bio->devs[i].bio->bi_end_io)
3741 rdev_dec_pending(conf->mirrors[d].rdev,
3743 if (r10_bio->devs[i].repl_bio &&
3744 r10_bio->devs[i].repl_bio->bi_end_io)
3746 conf->mirrors[d].replacement,
3756 if (sector_nr + max_sync < max_sector)
3757 max_sector = sector_nr + max_sync;
3760 int len = PAGE_SIZE;
3761 if (sector_nr + (len>>9) > max_sector)
3762 len = (max_sector - sector_nr) << 9;
3765 for (bio= biolist ; bio ; bio=bio->bi_next) {
3766 struct resync_pages *rp = get_resync_pages(bio);
3767 page = resync_fetch_page(rp, page_idx);
3769 * won't fail because the vec table is big enough
3770 * to hold all these pages
3772 bio_add_page(bio, page, len, 0);
3774 nr_sectors += len>>9;
3775 sector_nr += len>>9;
3776 } while (++page_idx < RESYNC_PAGES);
3777 r10_bio->sectors = nr_sectors;
3779 if (mddev_is_clustered(mddev) &&
3780 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3781 /* It is resync not recovery */
3782 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3783 conf->cluster_sync_low = mddev->curr_resync_completed;
3784 raid10_set_cluster_sync_high(conf);
3785 /* Send resync message */
3786 md_cluster_ops->resync_info_update(mddev,
3787 conf->cluster_sync_low,
3788 conf->cluster_sync_high);
3790 } else if (mddev_is_clustered(mddev)) {
3791 /* This is recovery not resync */
3792 sector_t sect_va1, sect_va2;
3793 bool broadcast_msg = false;
3795 for (i = 0; i < conf->geo.raid_disks; i++) {
3797 * sector_nr is a device address for recovery, so we
3798 * need translate it to array address before compare
3799 * with cluster_sync_high.
3801 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3803 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3804 broadcast_msg = true;
3806 * curr_resync_completed is similar as
3807 * sector_nr, so make the translation too.
3809 sect_va2 = raid10_find_virt(conf,
3810 mddev->curr_resync_completed, i);
3812 if (conf->cluster_sync_low == 0 ||
3813 conf->cluster_sync_low > sect_va2)
3814 conf->cluster_sync_low = sect_va2;
3817 if (broadcast_msg) {
3818 raid10_set_cluster_sync_high(conf);
3819 md_cluster_ops->resync_info_update(mddev,
3820 conf->cluster_sync_low,
3821 conf->cluster_sync_high);
3827 biolist = biolist->bi_next;
3829 bio->bi_next = NULL;
3830 r10_bio = get_resync_r10bio(bio);
3831 r10_bio->sectors = nr_sectors;
3833 if (bio->bi_end_io == end_sync_read) {
3834 md_sync_acct_bio(bio, nr_sectors);
3836 submit_bio_noacct(bio);
3840 if (sectors_skipped)
3841 /* pretend they weren't skipped, it makes
3842 * no important difference in this case
3844 md_done_sync(mddev, sectors_skipped, 1);
3846 return sectors_skipped + nr_sectors;
3848 /* There is nowhere to write, so all non-sync
3849 * drives must be failed or in resync, all drives
3850 * have a bad block, so try the next chunk...
3852 if (sector_nr + max_sync < max_sector)
3853 max_sector = sector_nr + max_sync;
3855 sectors_skipped += (max_sector - sector_nr);
3857 sector_nr = max_sector;
3862 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3865 struct r10conf *conf = mddev->private;
3868 raid_disks = min(conf->geo.raid_disks,
3869 conf->prev.raid_disks);
3871 sectors = conf->dev_sectors;
3873 size = sectors >> conf->geo.chunk_shift;
3874 sector_div(size, conf->geo.far_copies);
3875 size = size * raid_disks;
3876 sector_div(size, conf->geo.near_copies);
3878 return size << conf->geo.chunk_shift;
3881 static void calc_sectors(struct r10conf *conf, sector_t size)
3883 /* Calculate the number of sectors-per-device that will
3884 * actually be used, and set conf->dev_sectors and
3888 size = size >> conf->geo.chunk_shift;
3889 sector_div(size, conf->geo.far_copies);
3890 size = size * conf->geo.raid_disks;
3891 sector_div(size, conf->geo.near_copies);
3892 /* 'size' is now the number of chunks in the array */
3893 /* calculate "used chunks per device" */
3894 size = size * conf->copies;
3896 /* We need to round up when dividing by raid_disks to
3897 * get the stride size.
3899 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3901 conf->dev_sectors = size << conf->geo.chunk_shift;
3903 if (conf->geo.far_offset)
3904 conf->geo.stride = 1 << conf->geo.chunk_shift;
3906 sector_div(size, conf->geo.far_copies);
3907 conf->geo.stride = size << conf->geo.chunk_shift;
3911 enum geo_type {geo_new, geo_old, geo_start};
3912 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3915 int layout, chunk, disks;
3918 layout = mddev->layout;
3919 chunk = mddev->chunk_sectors;
3920 disks = mddev->raid_disks - mddev->delta_disks;
3923 layout = mddev->new_layout;
3924 chunk = mddev->new_chunk_sectors;
3925 disks = mddev->raid_disks;
3927 default: /* avoid 'may be unused' warnings */
3928 case geo_start: /* new when starting reshape - raid_disks not
3930 layout = mddev->new_layout;
3931 chunk = mddev->new_chunk_sectors;
3932 disks = mddev->raid_disks + mddev->delta_disks;
3937 if (chunk < (PAGE_SIZE >> 9) ||
3938 !is_power_of_2(chunk))
3941 fc = (layout >> 8) & 255;
3942 fo = layout & (1<<16);
3943 geo->raid_disks = disks;
3944 geo->near_copies = nc;
3945 geo->far_copies = fc;
3946 geo->far_offset = fo;
3947 switch (layout >> 17) {
3948 case 0: /* original layout. simple but not always optimal */
3949 geo->far_set_size = disks;
3951 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3952 * actually using this, but leave code here just in case.*/
3953 geo->far_set_size = disks/fc;
3954 WARN(geo->far_set_size < fc,
3955 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3957 case 2: /* "improved" layout fixed to match documentation */
3958 geo->far_set_size = fc * nc;
3960 default: /* Not a valid layout */
3963 geo->chunk_mask = chunk - 1;
3964 geo->chunk_shift = ffz(~chunk);
3968 static struct r10conf *setup_conf(struct mddev *mddev)
3970 struct r10conf *conf = NULL;
3975 copies = setup_geo(&geo, mddev, geo_new);
3978 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3979 mdname(mddev), PAGE_SIZE);
3983 if (copies < 2 || copies > mddev->raid_disks) {
3984 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3985 mdname(mddev), mddev->new_layout);
3990 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3994 /* FIXME calc properly */
3995 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3996 sizeof(struct raid10_info),
4001 conf->tmppage = alloc_page(GFP_KERNEL);
4006 conf->copies = copies;
4007 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
4008 rbio_pool_free, conf);
4012 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
4016 calc_sectors(conf, mddev->dev_sectors);
4017 if (mddev->reshape_position == MaxSector) {
4018 conf->prev = conf->geo;
4019 conf->reshape_progress = MaxSector;
4021 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
4025 conf->reshape_progress = mddev->reshape_position;
4026 if (conf->prev.far_offset)
4027 conf->prev.stride = 1 << conf->prev.chunk_shift;
4029 /* far_copies must be 1 */
4030 conf->prev.stride = conf->dev_sectors;
4032 conf->reshape_safe = conf->reshape_progress;
4033 spin_lock_init(&conf->device_lock);
4034 INIT_LIST_HEAD(&conf->retry_list);
4035 INIT_LIST_HEAD(&conf->bio_end_io_list);
4037 spin_lock_init(&conf->resync_lock);
4038 init_waitqueue_head(&conf->wait_barrier);
4039 atomic_set(&conf->nr_pending, 0);
4042 conf->thread = md_register_thread(raid10d, mddev, "raid10");
4046 conf->mddev = mddev;
4051 mempool_exit(&conf->r10bio_pool);
4052 kfree(conf->mirrors);
4053 safe_put_page(conf->tmppage);
4054 bioset_exit(&conf->bio_split);
4057 return ERR_PTR(err);
4060 static void raid10_set_io_opt(struct r10conf *conf)
4062 int raid_disks = conf->geo.raid_disks;
4064 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4065 raid_disks /= conf->geo.near_copies;
4066 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4070 static int raid10_run(struct mddev *mddev)
4072 struct r10conf *conf;
4074 struct raid10_info *disk;
4075 struct md_rdev *rdev;
4077 sector_t min_offset_diff = 0;
4079 bool discard_supported = false;
4081 if (mddev_init_writes_pending(mddev) < 0)
4084 if (mddev->private == NULL) {
4085 conf = setup_conf(mddev);
4087 return PTR_ERR(conf);
4088 mddev->private = conf;
4090 conf = mddev->private;
4094 if (mddev_is_clustered(conf->mddev)) {
4097 fc = (mddev->layout >> 8) & 255;
4098 fo = mddev->layout & (1<<16);
4099 if (fc > 1 || fo > 0) {
4100 pr_err("only near layout is supported by clustered"
4106 mddev->thread = conf->thread;
4107 conf->thread = NULL;
4110 blk_queue_max_discard_sectors(mddev->queue,
4112 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4113 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4114 raid10_set_io_opt(conf);
4117 rdev_for_each(rdev, mddev) {
4120 disk_idx = rdev->raid_disk;
4123 if (disk_idx >= conf->geo.raid_disks &&
4124 disk_idx >= conf->prev.raid_disks)
4126 disk = conf->mirrors + disk_idx;
4128 if (test_bit(Replacement, &rdev->flags)) {
4129 if (disk->replacement)
4131 disk->replacement = rdev;
4137 diff = (rdev->new_data_offset - rdev->data_offset);
4138 if (!mddev->reshape_backwards)
4142 if (first || diff < min_offset_diff)
4143 min_offset_diff = diff;
4146 disk_stack_limits(mddev->gendisk, rdev->bdev,
4147 rdev->data_offset << 9);
4149 disk->head_position = 0;
4151 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
4152 discard_supported = true;
4157 if (discard_supported)
4158 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
4161 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
4164 /* need to check that every block has at least one working mirror */
4165 if (!enough(conf, -1)) {
4166 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4171 if (conf->reshape_progress != MaxSector) {
4172 /* must ensure that shape change is supported */
4173 if (conf->geo.far_copies != 1 &&
4174 conf->geo.far_offset == 0)
4176 if (conf->prev.far_copies != 1 &&
4177 conf->prev.far_offset == 0)
4181 mddev->degraded = 0;
4183 i < conf->geo.raid_disks
4184 || i < conf->prev.raid_disks;
4187 disk = conf->mirrors + i;
4189 if (!disk->rdev && disk->replacement) {
4190 /* The replacement is all we have - use it */
4191 disk->rdev = disk->replacement;
4192 disk->replacement = NULL;
4193 clear_bit(Replacement, &disk->rdev->flags);
4197 !test_bit(In_sync, &disk->rdev->flags)) {
4198 disk->head_position = 0;
4201 disk->rdev->saved_raid_disk < 0)
4205 if (disk->replacement &&
4206 !test_bit(In_sync, &disk->replacement->flags) &&
4207 disk->replacement->saved_raid_disk < 0) {
4211 disk->recovery_disabled = mddev->recovery_disabled - 1;
4214 if (mddev->recovery_cp != MaxSector)
4215 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4217 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4218 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4219 conf->geo.raid_disks);
4221 * Ok, everything is just fine now
4223 mddev->dev_sectors = conf->dev_sectors;
4224 size = raid10_size(mddev, 0, 0);
4225 md_set_array_sectors(mddev, size);
4226 mddev->resync_max_sectors = size;
4227 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4229 if (md_integrity_register(mddev))
4232 if (conf->reshape_progress != MaxSector) {
4233 unsigned long before_length, after_length;
4235 before_length = ((1 << conf->prev.chunk_shift) *
4236 conf->prev.far_copies);
4237 after_length = ((1 << conf->geo.chunk_shift) *
4238 conf->geo.far_copies);
4240 if (max(before_length, after_length) > min_offset_diff) {
4241 /* This cannot work */
4242 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4245 conf->offset_diff = min_offset_diff;
4247 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4248 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4249 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4250 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4251 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4253 if (!mddev->sync_thread)
4260 md_unregister_thread(&mddev->thread);
4261 mempool_exit(&conf->r10bio_pool);
4262 safe_put_page(conf->tmppage);
4263 kfree(conf->mirrors);
4265 mddev->private = NULL;
4270 static void raid10_free(struct mddev *mddev, void *priv)
4272 struct r10conf *conf = priv;
4274 mempool_exit(&conf->r10bio_pool);
4275 safe_put_page(conf->tmppage);
4276 kfree(conf->mirrors);
4277 kfree(conf->mirrors_old);
4278 kfree(conf->mirrors_new);
4279 bioset_exit(&conf->bio_split);
4283 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4285 struct r10conf *conf = mddev->private;
4288 raise_barrier(conf, 0);
4290 lower_barrier(conf);
4293 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4295 /* Resize of 'far' arrays is not supported.
4296 * For 'near' and 'offset' arrays we can set the
4297 * number of sectors used to be an appropriate multiple
4298 * of the chunk size.
4299 * For 'offset', this is far_copies*chunksize.
4300 * For 'near' the multiplier is the LCM of
4301 * near_copies and raid_disks.
4302 * So if far_copies > 1 && !far_offset, fail.
4303 * Else find LCM(raid_disks, near_copy)*far_copies and
4304 * multiply by chunk_size. Then round to this number.
4305 * This is mostly done by raid10_size()
4307 struct r10conf *conf = mddev->private;
4308 sector_t oldsize, size;
4310 if (mddev->reshape_position != MaxSector)
4313 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4316 oldsize = raid10_size(mddev, 0, 0);
4317 size = raid10_size(mddev, sectors, 0);
4318 if (mddev->external_size &&
4319 mddev->array_sectors > size)
4321 if (mddev->bitmap) {
4322 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4326 md_set_array_sectors(mddev, size);
4327 if (sectors > mddev->dev_sectors &&
4328 mddev->recovery_cp > oldsize) {
4329 mddev->recovery_cp = oldsize;
4330 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4332 calc_sectors(conf, sectors);
4333 mddev->dev_sectors = conf->dev_sectors;
4334 mddev->resync_max_sectors = size;
4338 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4340 struct md_rdev *rdev;
4341 struct r10conf *conf;
4343 if (mddev->degraded > 0) {
4344 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4346 return ERR_PTR(-EINVAL);
4348 sector_div(size, devs);
4350 /* Set new parameters */
4351 mddev->new_level = 10;
4352 /* new layout: far_copies = 1, near_copies = 2 */
4353 mddev->new_layout = (1<<8) + 2;
4354 mddev->new_chunk_sectors = mddev->chunk_sectors;
4355 mddev->delta_disks = mddev->raid_disks;
4356 mddev->raid_disks *= 2;
4357 /* make sure it will be not marked as dirty */
4358 mddev->recovery_cp = MaxSector;
4359 mddev->dev_sectors = size;
4361 conf = setup_conf(mddev);
4362 if (!IS_ERR(conf)) {
4363 rdev_for_each(rdev, mddev)
4364 if (rdev->raid_disk >= 0) {
4365 rdev->new_raid_disk = rdev->raid_disk * 2;
4366 rdev->sectors = size;
4374 static void *raid10_takeover(struct mddev *mddev)
4376 struct r0conf *raid0_conf;
4378 /* raid10 can take over:
4379 * raid0 - providing it has only two drives
4381 if (mddev->level == 0) {
4382 /* for raid0 takeover only one zone is supported */
4383 raid0_conf = mddev->private;
4384 if (raid0_conf->nr_strip_zones > 1) {
4385 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4387 return ERR_PTR(-EINVAL);
4389 return raid10_takeover_raid0(mddev,
4390 raid0_conf->strip_zone->zone_end,
4391 raid0_conf->strip_zone->nb_dev);
4393 return ERR_PTR(-EINVAL);
4396 static int raid10_check_reshape(struct mddev *mddev)
4398 /* Called when there is a request to change
4399 * - layout (to ->new_layout)
4400 * - chunk size (to ->new_chunk_sectors)
4401 * - raid_disks (by delta_disks)
4402 * or when trying to restart a reshape that was ongoing.
4404 * We need to validate the request and possibly allocate
4405 * space if that might be an issue later.
4407 * Currently we reject any reshape of a 'far' mode array,
4408 * allow chunk size to change if new is generally acceptable,
4409 * allow raid_disks to increase, and allow
4410 * a switch between 'near' mode and 'offset' mode.
4412 struct r10conf *conf = mddev->private;
4415 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4418 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4419 /* mustn't change number of copies */
4421 if (geo.far_copies > 1 && !geo.far_offset)
4422 /* Cannot switch to 'far' mode */
4425 if (mddev->array_sectors & geo.chunk_mask)
4426 /* not factor of array size */
4429 if (!enough(conf, -1))
4432 kfree(conf->mirrors_new);
4433 conf->mirrors_new = NULL;
4434 if (mddev->delta_disks > 0) {
4435 /* allocate new 'mirrors' list */
4437 kcalloc(mddev->raid_disks + mddev->delta_disks,
4438 sizeof(struct raid10_info),
4440 if (!conf->mirrors_new)
4447 * Need to check if array has failed when deciding whether to:
4449 * - remove non-faulty devices
4452 * This determination is simple when no reshape is happening.
4453 * However if there is a reshape, we need to carefully check
4454 * both the before and after sections.
4455 * This is because some failed devices may only affect one
4456 * of the two sections, and some non-in_sync devices may
4457 * be insync in the section most affected by failed devices.
4459 static int calc_degraded(struct r10conf *conf)
4461 int degraded, degraded2;
4466 /* 'prev' section first */
4467 for (i = 0; i < conf->prev.raid_disks; i++) {
4468 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4469 if (!rdev || test_bit(Faulty, &rdev->flags))
4471 else if (!test_bit(In_sync, &rdev->flags))
4472 /* When we can reduce the number of devices in
4473 * an array, this might not contribute to
4474 * 'degraded'. It does now.
4479 if (conf->geo.raid_disks == conf->prev.raid_disks)
4483 for (i = 0; i < conf->geo.raid_disks; i++) {
4484 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4485 if (!rdev || test_bit(Faulty, &rdev->flags))
4487 else if (!test_bit(In_sync, &rdev->flags)) {
4488 /* If reshape is increasing the number of devices,
4489 * this section has already been recovered, so
4490 * it doesn't contribute to degraded.
4493 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4498 if (degraded2 > degraded)
4503 static int raid10_start_reshape(struct mddev *mddev)
4505 /* A 'reshape' has been requested. This commits
4506 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4507 * This also checks if there are enough spares and adds them
4509 * We currently require enough spares to make the final
4510 * array non-degraded. We also require that the difference
4511 * between old and new data_offset - on each device - is
4512 * enough that we never risk over-writing.
4515 unsigned long before_length, after_length;
4516 sector_t min_offset_diff = 0;
4519 struct r10conf *conf = mddev->private;
4520 struct md_rdev *rdev;
4524 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4527 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4530 before_length = ((1 << conf->prev.chunk_shift) *
4531 conf->prev.far_copies);
4532 after_length = ((1 << conf->geo.chunk_shift) *
4533 conf->geo.far_copies);
4535 rdev_for_each(rdev, mddev) {
4536 if (!test_bit(In_sync, &rdev->flags)
4537 && !test_bit(Faulty, &rdev->flags))
4539 if (rdev->raid_disk >= 0) {
4540 long long diff = (rdev->new_data_offset
4541 - rdev->data_offset);
4542 if (!mddev->reshape_backwards)
4546 if (first || diff < min_offset_diff)
4547 min_offset_diff = diff;
4552 if (max(before_length, after_length) > min_offset_diff)
4555 if (spares < mddev->delta_disks)
4558 conf->offset_diff = min_offset_diff;
4559 spin_lock_irq(&conf->device_lock);
4560 if (conf->mirrors_new) {
4561 memcpy(conf->mirrors_new, conf->mirrors,
4562 sizeof(struct raid10_info)*conf->prev.raid_disks);
4564 kfree(conf->mirrors_old);
4565 conf->mirrors_old = conf->mirrors;
4566 conf->mirrors = conf->mirrors_new;
4567 conf->mirrors_new = NULL;
4569 setup_geo(&conf->geo, mddev, geo_start);
4571 if (mddev->reshape_backwards) {
4572 sector_t size = raid10_size(mddev, 0, 0);
4573 if (size < mddev->array_sectors) {
4574 spin_unlock_irq(&conf->device_lock);
4575 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4579 mddev->resync_max_sectors = size;
4580 conf->reshape_progress = size;
4582 conf->reshape_progress = 0;
4583 conf->reshape_safe = conf->reshape_progress;
4584 spin_unlock_irq(&conf->device_lock);
4586 if (mddev->delta_disks && mddev->bitmap) {
4587 struct mdp_superblock_1 *sb = NULL;
4588 sector_t oldsize, newsize;
4590 oldsize = raid10_size(mddev, 0, 0);
4591 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4593 if (!mddev_is_clustered(mddev)) {
4594 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4601 rdev_for_each(rdev, mddev) {
4602 if (rdev->raid_disk > -1 &&
4603 !test_bit(Faulty, &rdev->flags))
4604 sb = page_address(rdev->sb_page);
4608 * some node is already performing reshape, and no need to
4609 * call md_bitmap_resize again since it should be called when
4610 * receiving BITMAP_RESIZE msg
4612 if ((sb && (le32_to_cpu(sb->feature_map) &
4613 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4616 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4620 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4622 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4627 if (mddev->delta_disks > 0) {
4628 rdev_for_each(rdev, mddev)
4629 if (rdev->raid_disk < 0 &&
4630 !test_bit(Faulty, &rdev->flags)) {
4631 if (raid10_add_disk(mddev, rdev) == 0) {
4632 if (rdev->raid_disk >=
4633 conf->prev.raid_disks)
4634 set_bit(In_sync, &rdev->flags);
4636 rdev->recovery_offset = 0;
4638 /* Failure here is OK */
4639 sysfs_link_rdev(mddev, rdev);
4641 } else if (rdev->raid_disk >= conf->prev.raid_disks
4642 && !test_bit(Faulty, &rdev->flags)) {
4643 /* This is a spare that was manually added */
4644 set_bit(In_sync, &rdev->flags);
4647 /* When a reshape changes the number of devices,
4648 * ->degraded is measured against the larger of the
4649 * pre and post numbers.
4651 spin_lock_irq(&conf->device_lock);
4652 mddev->degraded = calc_degraded(conf);
4653 spin_unlock_irq(&conf->device_lock);
4654 mddev->raid_disks = conf->geo.raid_disks;
4655 mddev->reshape_position = conf->reshape_progress;
4656 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4658 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4659 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4660 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4661 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4662 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4664 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4666 if (!mddev->sync_thread) {
4670 conf->reshape_checkpoint = jiffies;
4671 md_wakeup_thread(mddev->sync_thread);
4676 mddev->recovery = 0;
4677 spin_lock_irq(&conf->device_lock);
4678 conf->geo = conf->prev;
4679 mddev->raid_disks = conf->geo.raid_disks;
4680 rdev_for_each(rdev, mddev)
4681 rdev->new_data_offset = rdev->data_offset;
4683 conf->reshape_progress = MaxSector;
4684 conf->reshape_safe = MaxSector;
4685 mddev->reshape_position = MaxSector;
4686 spin_unlock_irq(&conf->device_lock);
4690 /* Calculate the last device-address that could contain
4691 * any block from the chunk that includes the array-address 's'
4692 * and report the next address.
4693 * i.e. the address returned will be chunk-aligned and after
4694 * any data that is in the chunk containing 's'.
4696 static sector_t last_dev_address(sector_t s, struct geom *geo)
4698 s = (s | geo->chunk_mask) + 1;
4699 s >>= geo->chunk_shift;
4700 s *= geo->near_copies;
4701 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4702 s *= geo->far_copies;
4703 s <<= geo->chunk_shift;
4707 /* Calculate the first device-address that could contain
4708 * any block from the chunk that includes the array-address 's'.
4709 * This too will be the start of a chunk
4711 static sector_t first_dev_address(sector_t s, struct geom *geo)
4713 s >>= geo->chunk_shift;
4714 s *= geo->near_copies;
4715 sector_div(s, geo->raid_disks);
4716 s *= geo->far_copies;
4717 s <<= geo->chunk_shift;
4721 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4724 /* We simply copy at most one chunk (smallest of old and new)
4725 * at a time, possibly less if that exceeds RESYNC_PAGES,
4726 * or we hit a bad block or something.
4727 * This might mean we pause for normal IO in the middle of
4728 * a chunk, but that is not a problem as mddev->reshape_position
4729 * can record any location.
4731 * If we will want to write to a location that isn't
4732 * yet recorded as 'safe' (i.e. in metadata on disk) then
4733 * we need to flush all reshape requests and update the metadata.
4735 * When reshaping forwards (e.g. to more devices), we interpret
4736 * 'safe' as the earliest block which might not have been copied
4737 * down yet. We divide this by previous stripe size and multiply
4738 * by previous stripe length to get lowest device offset that we
4739 * cannot write to yet.
4740 * We interpret 'sector_nr' as an address that we want to write to.
4741 * From this we use last_device_address() to find where we might
4742 * write to, and first_device_address on the 'safe' position.
4743 * If this 'next' write position is after the 'safe' position,
4744 * we must update the metadata to increase the 'safe' position.
4746 * When reshaping backwards, we round in the opposite direction
4747 * and perform the reverse test: next write position must not be
4748 * less than current safe position.
4750 * In all this the minimum difference in data offsets
4751 * (conf->offset_diff - always positive) allows a bit of slack,
4752 * so next can be after 'safe', but not by more than offset_diff
4754 * We need to prepare all the bios here before we start any IO
4755 * to ensure the size we choose is acceptable to all devices.
4756 * The means one for each copy for write-out and an extra one for
4758 * We store the read-in bio in ->master_bio and the others in
4759 * ->devs[x].bio and ->devs[x].repl_bio.
4761 struct r10conf *conf = mddev->private;
4762 struct r10bio *r10_bio;
4763 sector_t next, safe, last;
4767 struct md_rdev *rdev;
4770 struct bio *bio, *read_bio;
4771 int sectors_done = 0;
4772 struct page **pages;
4774 if (sector_nr == 0) {
4775 /* If restarting in the middle, skip the initial sectors */
4776 if (mddev->reshape_backwards &&
4777 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4778 sector_nr = (raid10_size(mddev, 0, 0)
4779 - conf->reshape_progress);
4780 } else if (!mddev->reshape_backwards &&
4781 conf->reshape_progress > 0)
4782 sector_nr = conf->reshape_progress;
4784 mddev->curr_resync_completed = sector_nr;
4785 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4791 /* We don't use sector_nr to track where we are up to
4792 * as that doesn't work well for ->reshape_backwards.
4793 * So just use ->reshape_progress.
4795 if (mddev->reshape_backwards) {
4796 /* 'next' is the earliest device address that we might
4797 * write to for this chunk in the new layout
4799 next = first_dev_address(conf->reshape_progress - 1,
4802 /* 'safe' is the last device address that we might read from
4803 * in the old layout after a restart
4805 safe = last_dev_address(conf->reshape_safe - 1,
4808 if (next + conf->offset_diff < safe)
4811 last = conf->reshape_progress - 1;
4812 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4813 & conf->prev.chunk_mask);
4814 if (sector_nr + RESYNC_SECTORS < last)
4815 sector_nr = last + 1 - RESYNC_SECTORS;
4817 /* 'next' is after the last device address that we
4818 * might write to for this chunk in the new layout
4820 next = last_dev_address(conf->reshape_progress, &conf->geo);
4822 /* 'safe' is the earliest device address that we might
4823 * read from in the old layout after a restart
4825 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4827 /* Need to update metadata if 'next' might be beyond 'safe'
4828 * as that would possibly corrupt data
4830 if (next > safe + conf->offset_diff)
4833 sector_nr = conf->reshape_progress;
4834 last = sector_nr | (conf->geo.chunk_mask
4835 & conf->prev.chunk_mask);
4837 if (sector_nr + RESYNC_SECTORS <= last)
4838 last = sector_nr + RESYNC_SECTORS - 1;
4842 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4843 /* Need to update reshape_position in metadata */
4844 wait_barrier(conf, false);
4845 mddev->reshape_position = conf->reshape_progress;
4846 if (mddev->reshape_backwards)
4847 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4848 - conf->reshape_progress;
4850 mddev->curr_resync_completed = conf->reshape_progress;
4851 conf->reshape_checkpoint = jiffies;
4852 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4853 md_wakeup_thread(mddev->thread);
4854 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4855 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4856 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4857 allow_barrier(conf);
4858 return sectors_done;
4860 conf->reshape_safe = mddev->reshape_position;
4861 allow_barrier(conf);
4864 raise_barrier(conf, 0);
4866 /* Now schedule reads for blocks from sector_nr to last */
4867 r10_bio = raid10_alloc_init_r10buf(conf);
4869 raise_barrier(conf, 1);
4870 atomic_set(&r10_bio->remaining, 0);
4871 r10_bio->mddev = mddev;
4872 r10_bio->sector = sector_nr;
4873 set_bit(R10BIO_IsReshape, &r10_bio->state);
4874 r10_bio->sectors = last - sector_nr + 1;
4875 rdev = read_balance(conf, r10_bio, &max_sectors);
4876 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4879 /* Cannot read from here, so need to record bad blocks
4880 * on all the target devices.
4883 mempool_free(r10_bio, &conf->r10buf_pool);
4884 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4885 return sectors_done;
4888 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4889 GFP_KERNEL, &mddev->bio_set);
4890 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4891 + rdev->data_offset);
4892 read_bio->bi_private = r10_bio;
4893 read_bio->bi_end_io = end_reshape_read;
4894 r10_bio->master_bio = read_bio;
4895 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4898 * Broadcast RESYNC message to other nodes, so all nodes would not
4899 * write to the region to avoid conflict.
4901 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4902 struct mdp_superblock_1 *sb = NULL;
4903 int sb_reshape_pos = 0;
4905 conf->cluster_sync_low = sector_nr;
4906 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4907 sb = page_address(rdev->sb_page);
4909 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4911 * Set cluster_sync_low again if next address for array
4912 * reshape is less than cluster_sync_low. Since we can't
4913 * update cluster_sync_low until it has finished reshape.
4915 if (sb_reshape_pos < conf->cluster_sync_low)
4916 conf->cluster_sync_low = sb_reshape_pos;
4919 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4920 conf->cluster_sync_high);
4923 /* Now find the locations in the new layout */
4924 __raid10_find_phys(&conf->geo, r10_bio);
4927 read_bio->bi_next = NULL;
4930 for (s = 0; s < conf->copies*2; s++) {
4932 int d = r10_bio->devs[s/2].devnum;
4933 struct md_rdev *rdev2;
4935 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4936 b = r10_bio->devs[s/2].repl_bio;
4938 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4939 b = r10_bio->devs[s/2].bio;
4941 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4944 bio_set_dev(b, rdev2->bdev);
4945 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4946 rdev2->new_data_offset;
4947 b->bi_end_io = end_reshape_write;
4948 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4953 /* Now add as many pages as possible to all of these bios. */
4956 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4957 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4958 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4959 int len = (max_sectors - s) << 9;
4960 if (len > PAGE_SIZE)
4962 for (bio = blist; bio ; bio = bio->bi_next) {
4964 * won't fail because the vec table is big enough
4965 * to hold all these pages
4967 bio_add_page(bio, page, len, 0);
4969 sector_nr += len >> 9;
4970 nr_sectors += len >> 9;
4973 r10_bio->sectors = nr_sectors;
4975 /* Now submit the read */
4976 md_sync_acct_bio(read_bio, r10_bio->sectors);
4977 atomic_inc(&r10_bio->remaining);
4978 read_bio->bi_next = NULL;
4979 submit_bio_noacct(read_bio);
4980 sectors_done += nr_sectors;
4981 if (sector_nr <= last)
4984 lower_barrier(conf);
4986 /* Now that we have done the whole section we can
4987 * update reshape_progress
4989 if (mddev->reshape_backwards)
4990 conf->reshape_progress -= sectors_done;
4992 conf->reshape_progress += sectors_done;
4994 return sectors_done;
4997 static void end_reshape_request(struct r10bio *r10_bio);
4998 static int handle_reshape_read_error(struct mddev *mddev,
4999 struct r10bio *r10_bio);
5000 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
5002 /* Reshape read completed. Hopefully we have a block
5004 * If we got a read error then we do sync 1-page reads from
5005 * elsewhere until we find the data - or give up.
5007 struct r10conf *conf = mddev->private;
5010 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
5011 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
5012 /* Reshape has been aborted */
5013 md_done_sync(mddev, r10_bio->sectors, 0);
5017 /* We definitely have the data in the pages, schedule the
5020 atomic_set(&r10_bio->remaining, 1);
5021 for (s = 0; s < conf->copies*2; s++) {
5023 int d = r10_bio->devs[s/2].devnum;
5024 struct md_rdev *rdev;
5027 rdev = rcu_dereference(conf->mirrors[d].replacement);
5028 b = r10_bio->devs[s/2].repl_bio;
5030 rdev = rcu_dereference(conf->mirrors[d].rdev);
5031 b = r10_bio->devs[s/2].bio;
5033 if (!rdev || test_bit(Faulty, &rdev->flags)) {
5037 atomic_inc(&rdev->nr_pending);
5039 md_sync_acct_bio(b, r10_bio->sectors);
5040 atomic_inc(&r10_bio->remaining);
5042 submit_bio_noacct(b);
5044 end_reshape_request(r10_bio);
5047 static void end_reshape(struct r10conf *conf)
5049 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
5052 spin_lock_irq(&conf->device_lock);
5053 conf->prev = conf->geo;
5054 md_finish_reshape(conf->mddev);
5056 conf->reshape_progress = MaxSector;
5057 conf->reshape_safe = MaxSector;
5058 spin_unlock_irq(&conf->device_lock);
5060 if (conf->mddev->queue)
5061 raid10_set_io_opt(conf);
5065 static void raid10_update_reshape_pos(struct mddev *mddev)
5067 struct r10conf *conf = mddev->private;
5070 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
5071 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
5072 || mddev->reshape_position == MaxSector)
5073 conf->reshape_progress = mddev->reshape_position;
5078 static int handle_reshape_read_error(struct mddev *mddev,
5079 struct r10bio *r10_bio)
5081 /* Use sync reads to get the blocks from somewhere else */
5082 int sectors = r10_bio->sectors;
5083 struct r10conf *conf = mddev->private;
5084 struct r10bio *r10b;
5087 struct page **pages;
5089 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5091 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5095 /* reshape IOs share pages from .devs[0].bio */
5096 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5098 r10b->sector = r10_bio->sector;
5099 __raid10_find_phys(&conf->prev, r10b);
5104 int first_slot = slot;
5106 if (s > (PAGE_SIZE >> 9))
5111 int d = r10b->devs[slot].devnum;
5112 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5115 test_bit(Faulty, &rdev->flags) ||
5116 !test_bit(In_sync, &rdev->flags))
5119 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5120 atomic_inc(&rdev->nr_pending);
5122 success = sync_page_io(rdev,
5126 REQ_OP_READ, 0, false);
5127 rdev_dec_pending(rdev, mddev);
5133 if (slot >= conf->copies)
5135 if (slot == first_slot)
5140 /* couldn't read this block, must give up */
5141 set_bit(MD_RECOVERY_INTR,
5153 static void end_reshape_write(struct bio *bio)
5155 struct r10bio *r10_bio = get_resync_r10bio(bio);
5156 struct mddev *mddev = r10_bio->mddev;
5157 struct r10conf *conf = mddev->private;
5161 struct md_rdev *rdev = NULL;
5163 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5165 rdev = conf->mirrors[d].replacement;
5168 rdev = conf->mirrors[d].rdev;
5171 if (bio->bi_status) {
5172 /* FIXME should record badblock */
5173 md_error(mddev, rdev);
5176 rdev_dec_pending(rdev, mddev);
5177 end_reshape_request(r10_bio);
5180 static void end_reshape_request(struct r10bio *r10_bio)
5182 if (!atomic_dec_and_test(&r10_bio->remaining))
5184 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5185 bio_put(r10_bio->master_bio);
5189 static void raid10_finish_reshape(struct mddev *mddev)
5191 struct r10conf *conf = mddev->private;
5193 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5196 if (mddev->delta_disks > 0) {
5197 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5198 mddev->recovery_cp = mddev->resync_max_sectors;
5199 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5201 mddev->resync_max_sectors = mddev->array_sectors;
5205 for (d = conf->geo.raid_disks ;
5206 d < conf->geo.raid_disks - mddev->delta_disks;
5208 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5210 clear_bit(In_sync, &rdev->flags);
5211 rdev = rcu_dereference(conf->mirrors[d].replacement);
5213 clear_bit(In_sync, &rdev->flags);
5217 mddev->layout = mddev->new_layout;
5218 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5219 mddev->reshape_position = MaxSector;
5220 mddev->delta_disks = 0;
5221 mddev->reshape_backwards = 0;
5224 static struct md_personality raid10_personality =
5228 .owner = THIS_MODULE,
5229 .make_request = raid10_make_request,
5231 .free = raid10_free,
5232 .status = raid10_status,
5233 .error_handler = raid10_error,
5234 .hot_add_disk = raid10_add_disk,
5235 .hot_remove_disk= raid10_remove_disk,
5236 .spare_active = raid10_spare_active,
5237 .sync_request = raid10_sync_request,
5238 .quiesce = raid10_quiesce,
5239 .size = raid10_size,
5240 .resize = raid10_resize,
5241 .takeover = raid10_takeover,
5242 .check_reshape = raid10_check_reshape,
5243 .start_reshape = raid10_start_reshape,
5244 .finish_reshape = raid10_finish_reshape,
5245 .update_reshape_pos = raid10_update_reshape_pos,
5248 static int __init raid_init(void)
5250 return register_md_personality(&raid10_personality);
5253 static void raid_exit(void)
5255 unregister_md_personality(&raid10_personality);
5258 module_init(raid_init);
5259 module_exit(raid_exit);
5260 MODULE_LICENSE("GPL");
5261 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5262 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5263 MODULE_ALIAS("md-raid10");
5264 MODULE_ALIAS("md-level-10");