2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include <trace/events/block.h>
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
76 * Number of guaranteed r10bios in case of extreme VM load:
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
98 static int max_queued_requests = 1024;
100 static void allow_barrier(struct r10conf *conf);
101 static void lower_barrier(struct r10conf *conf);
102 static int _enough(struct r10conf *conf, int previous, int ignore);
103 static int enough(struct r10conf *conf, int ignore);
104 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
106 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107 static void end_reshape_write(struct bio *bio);
108 static void end_reshape(struct r10conf *conf);
110 #define raid10_log(md, fmt, args...) \
111 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
113 #include "raid1-10.c"
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
119 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
121 return get_resync_pages(bio)->raid_bio;
124 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
126 struct r10conf *conf = data;
127 int size = offsetof(struct r10bio, devs[conf->copies]);
129 /* allocate a r10bio with room for raid_disks entries in the
131 return kzalloc(size, gfp_flags);
134 static void r10bio_pool_free(void *r10_bio, void *data)
139 /* amount of memory to reserve for resync requests */
140 #define RESYNC_WINDOW (1024*1024)
141 /* maximum number of concurrent requests, memory permitting */
142 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
145 * When performing a resync, we need to read and compare, so
146 * we need as many pages are there are copies.
147 * When performing a recovery, we need 2 bios, one for read,
148 * one for write (we recover only one drive per r10buf)
151 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
153 struct r10conf *conf = data;
154 struct r10bio *r10_bio;
157 int nalloc, nalloc_rp;
158 struct resync_pages *rps;
160 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
164 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
165 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
166 nalloc = conf->copies; /* resync */
168 nalloc = 2; /* recovery */
170 /* allocate once for all bios */
171 if (!conf->have_replacement)
174 nalloc_rp = nalloc * 2;
175 rps = kmalloc(sizeof(struct resync_pages) * nalloc_rp, gfp_flags);
177 goto out_free_r10bio;
182 for (j = nalloc ; j-- ; ) {
183 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
186 r10_bio->devs[j].bio = bio;
187 if (!conf->have_replacement)
189 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
192 r10_bio->devs[j].repl_bio = bio;
195 * Allocate RESYNC_PAGES data pages and attach them
198 for (j = 0; j < nalloc; j++) {
199 struct bio *rbio = r10_bio->devs[j].repl_bio;
200 struct resync_pages *rp, *rp_repl;
204 rp_repl = &rps[nalloc + j];
206 bio = r10_bio->devs[j].bio;
208 if (!j || test_bit(MD_RECOVERY_SYNC,
209 &conf->mddev->recovery)) {
210 if (resync_alloc_pages(rp, gfp_flags))
213 memcpy(rp, &rps[0], sizeof(*rp));
214 resync_get_all_pages(rp);
217 rp->raid_bio = r10_bio;
218 bio->bi_private = rp;
220 memcpy(rp_repl, rp, sizeof(*rp));
221 rbio->bi_private = rp_repl;
229 resync_free_pages(&rps[j * 2]);
233 for ( ; j < nalloc; j++) {
234 if (r10_bio->devs[j].bio)
235 bio_put(r10_bio->devs[j].bio);
236 if (r10_bio->devs[j].repl_bio)
237 bio_put(r10_bio->devs[j].repl_bio);
241 r10bio_pool_free(r10_bio, conf);
245 static void r10buf_pool_free(void *__r10_bio, void *data)
247 struct r10conf *conf = data;
248 struct r10bio *r10bio = __r10_bio;
250 struct resync_pages *rp = NULL;
252 for (j = conf->copies; j--; ) {
253 struct bio *bio = r10bio->devs[j].bio;
255 rp = get_resync_pages(bio);
256 resync_free_pages(rp);
259 bio = r10bio->devs[j].repl_bio;
264 /* resync pages array stored in the 1st bio's .bi_private */
267 r10bio_pool_free(r10bio, conf);
270 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
274 for (i = 0; i < conf->copies; i++) {
275 struct bio **bio = & r10_bio->devs[i].bio;
276 if (!BIO_SPECIAL(*bio))
279 bio = &r10_bio->devs[i].repl_bio;
280 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
286 static void free_r10bio(struct r10bio *r10_bio)
288 struct r10conf *conf = r10_bio->mddev->private;
290 put_all_bios(conf, r10_bio);
291 mempool_free(r10_bio, conf->r10bio_pool);
294 static void put_buf(struct r10bio *r10_bio)
296 struct r10conf *conf = r10_bio->mddev->private;
298 mempool_free(r10_bio, conf->r10buf_pool);
303 static void reschedule_retry(struct r10bio *r10_bio)
306 struct mddev *mddev = r10_bio->mddev;
307 struct r10conf *conf = mddev->private;
309 spin_lock_irqsave(&conf->device_lock, flags);
310 list_add(&r10_bio->retry_list, &conf->retry_list);
312 spin_unlock_irqrestore(&conf->device_lock, flags);
314 /* wake up frozen array... */
315 wake_up(&conf->wait_barrier);
317 md_wakeup_thread(mddev->thread);
321 * raid_end_bio_io() is called when we have finished servicing a mirrored
322 * operation and are ready to return a success/failure code to the buffer
325 static void raid_end_bio_io(struct r10bio *r10_bio)
327 struct bio *bio = r10_bio->master_bio;
328 struct r10conf *conf = r10_bio->mddev->private;
330 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
331 bio->bi_status = BLK_STS_IOERR;
335 * Wake up any possible resync thread that waits for the device
340 free_r10bio(r10_bio);
344 * Update disk head position estimator based on IRQ completion info.
346 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
348 struct r10conf *conf = r10_bio->mddev->private;
350 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
351 r10_bio->devs[slot].addr + (r10_bio->sectors);
355 * Find the disk number which triggered given bio
357 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
358 struct bio *bio, int *slotp, int *replp)
363 for (slot = 0; slot < conf->copies; slot++) {
364 if (r10_bio->devs[slot].bio == bio)
366 if (r10_bio->devs[slot].repl_bio == bio) {
372 BUG_ON(slot == conf->copies);
373 update_head_pos(slot, r10_bio);
379 return r10_bio->devs[slot].devnum;
382 static void raid10_end_read_request(struct bio *bio)
384 int uptodate = !bio->bi_status;
385 struct r10bio *r10_bio = bio->bi_private;
387 struct md_rdev *rdev;
388 struct r10conf *conf = r10_bio->mddev->private;
390 slot = r10_bio->read_slot;
391 dev = r10_bio->devs[slot].devnum;
392 rdev = r10_bio->devs[slot].rdev;
394 * this branch is our 'one mirror IO has finished' event handler:
396 update_head_pos(slot, r10_bio);
400 * Set R10BIO_Uptodate in our master bio, so that
401 * we will return a good error code to the higher
402 * levels even if IO on some other mirrored buffer fails.
404 * The 'master' represents the composite IO operation to
405 * user-side. So if something waits for IO, then it will
406 * wait for the 'master' bio.
408 set_bit(R10BIO_Uptodate, &r10_bio->state);
410 /* If all other devices that store this block have
411 * failed, we want to return the error upwards rather
412 * than fail the last device. Here we redefine
413 * "uptodate" to mean "Don't want to retry"
415 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
420 raid_end_bio_io(r10_bio);
421 rdev_dec_pending(rdev, conf->mddev);
424 * oops, read error - keep the refcount on the rdev
426 char b[BDEVNAME_SIZE];
427 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
429 bdevname(rdev->bdev, b),
430 (unsigned long long)r10_bio->sector);
431 set_bit(R10BIO_ReadError, &r10_bio->state);
432 reschedule_retry(r10_bio);
436 static void close_write(struct r10bio *r10_bio)
438 /* clear the bitmap if all writes complete successfully */
439 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
441 !test_bit(R10BIO_Degraded, &r10_bio->state),
443 md_write_end(r10_bio->mddev);
446 static void one_write_done(struct r10bio *r10_bio)
448 if (atomic_dec_and_test(&r10_bio->remaining)) {
449 if (test_bit(R10BIO_WriteError, &r10_bio->state))
450 reschedule_retry(r10_bio);
452 close_write(r10_bio);
453 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
454 reschedule_retry(r10_bio);
456 raid_end_bio_io(r10_bio);
461 static void raid10_end_write_request(struct bio *bio)
463 struct r10bio *r10_bio = bio->bi_private;
466 struct r10conf *conf = r10_bio->mddev->private;
468 struct md_rdev *rdev = NULL;
469 struct bio *to_put = NULL;
472 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
474 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
477 rdev = conf->mirrors[dev].replacement;
481 rdev = conf->mirrors[dev].rdev;
484 * this branch is our 'one mirror IO has finished' event handler:
486 if (bio->bi_status && !discard_error) {
488 /* Never record new bad blocks to replacement,
491 md_error(rdev->mddev, rdev);
493 set_bit(WriteErrorSeen, &rdev->flags);
494 if (!test_and_set_bit(WantReplacement, &rdev->flags))
495 set_bit(MD_RECOVERY_NEEDED,
496 &rdev->mddev->recovery);
499 if (test_bit(FailFast, &rdev->flags) &&
500 (bio->bi_opf & MD_FAILFAST)) {
501 md_error(rdev->mddev, rdev);
502 if (!test_bit(Faulty, &rdev->flags))
503 /* This is the only remaining device,
504 * We need to retry the write without
507 set_bit(R10BIO_WriteError, &r10_bio->state);
509 r10_bio->devs[slot].bio = NULL;
514 set_bit(R10BIO_WriteError, &r10_bio->state);
518 * Set R10BIO_Uptodate in our master bio, so that
519 * we will return a good error code for to the higher
520 * levels even if IO on some other mirrored buffer fails.
522 * The 'master' represents the composite IO operation to
523 * user-side. So if something waits for IO, then it will
524 * wait for the 'master' bio.
530 * Do not set R10BIO_Uptodate if the current device is
531 * rebuilding or Faulty. This is because we cannot use
532 * such device for properly reading the data back (we could
533 * potentially use it, if the current write would have felt
534 * before rdev->recovery_offset, but for simplicity we don't
537 if (test_bit(In_sync, &rdev->flags) &&
538 !test_bit(Faulty, &rdev->flags))
539 set_bit(R10BIO_Uptodate, &r10_bio->state);
541 /* Maybe we can clear some bad blocks. */
542 if (is_badblock(rdev,
543 r10_bio->devs[slot].addr,
545 &first_bad, &bad_sectors) && !discard_error) {
548 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
550 r10_bio->devs[slot].bio = IO_MADE_GOOD;
552 set_bit(R10BIO_MadeGood, &r10_bio->state);
558 * Let's see if all mirrored write operations have finished
561 one_write_done(r10_bio);
563 rdev_dec_pending(rdev, conf->mddev);
569 * RAID10 layout manager
570 * As well as the chunksize and raid_disks count, there are two
571 * parameters: near_copies and far_copies.
572 * near_copies * far_copies must be <= raid_disks.
573 * Normally one of these will be 1.
574 * If both are 1, we get raid0.
575 * If near_copies == raid_disks, we get raid1.
577 * Chunks are laid out in raid0 style with near_copies copies of the
578 * first chunk, followed by near_copies copies of the next chunk and
580 * If far_copies > 1, then after 1/far_copies of the array has been assigned
581 * as described above, we start again with a device offset of near_copies.
582 * So we effectively have another copy of the whole array further down all
583 * the drives, but with blocks on different drives.
584 * With this layout, and block is never stored twice on the one device.
586 * raid10_find_phys finds the sector offset of a given virtual sector
587 * on each device that it is on.
589 * raid10_find_virt does the reverse mapping, from a device and a
590 * sector offset to a virtual address
593 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
601 int last_far_set_start, last_far_set_size;
603 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
604 last_far_set_start *= geo->far_set_size;
606 last_far_set_size = geo->far_set_size;
607 last_far_set_size += (geo->raid_disks % geo->far_set_size);
609 /* now calculate first sector/dev */
610 chunk = r10bio->sector >> geo->chunk_shift;
611 sector = r10bio->sector & geo->chunk_mask;
613 chunk *= geo->near_copies;
615 dev = sector_div(stripe, geo->raid_disks);
617 stripe *= geo->far_copies;
619 sector += stripe << geo->chunk_shift;
621 /* and calculate all the others */
622 for (n = 0; n < geo->near_copies; n++) {
626 r10bio->devs[slot].devnum = d;
627 r10bio->devs[slot].addr = s;
630 for (f = 1; f < geo->far_copies; f++) {
631 set = d / geo->far_set_size;
632 d += geo->near_copies;
634 if ((geo->raid_disks % geo->far_set_size) &&
635 (d > last_far_set_start)) {
636 d -= last_far_set_start;
637 d %= last_far_set_size;
638 d += last_far_set_start;
640 d %= geo->far_set_size;
641 d += geo->far_set_size * set;
644 r10bio->devs[slot].devnum = d;
645 r10bio->devs[slot].addr = s;
649 if (dev >= geo->raid_disks) {
651 sector += (geo->chunk_mask + 1);
656 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
658 struct geom *geo = &conf->geo;
660 if (conf->reshape_progress != MaxSector &&
661 ((r10bio->sector >= conf->reshape_progress) !=
662 conf->mddev->reshape_backwards)) {
663 set_bit(R10BIO_Previous, &r10bio->state);
666 clear_bit(R10BIO_Previous, &r10bio->state);
668 __raid10_find_phys(geo, r10bio);
671 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
673 sector_t offset, chunk, vchunk;
674 /* Never use conf->prev as this is only called during resync
675 * or recovery, so reshape isn't happening
677 struct geom *geo = &conf->geo;
678 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
679 int far_set_size = geo->far_set_size;
680 int last_far_set_start;
682 if (geo->raid_disks % geo->far_set_size) {
683 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
684 last_far_set_start *= geo->far_set_size;
686 if (dev >= last_far_set_start) {
687 far_set_size = geo->far_set_size;
688 far_set_size += (geo->raid_disks % geo->far_set_size);
689 far_set_start = last_far_set_start;
693 offset = sector & geo->chunk_mask;
694 if (geo->far_offset) {
696 chunk = sector >> geo->chunk_shift;
697 fc = sector_div(chunk, geo->far_copies);
698 dev -= fc * geo->near_copies;
699 if (dev < far_set_start)
702 while (sector >= geo->stride) {
703 sector -= geo->stride;
704 if (dev < (geo->near_copies + far_set_start))
705 dev += far_set_size - geo->near_copies;
707 dev -= geo->near_copies;
709 chunk = sector >> geo->chunk_shift;
711 vchunk = chunk * geo->raid_disks + dev;
712 sector_div(vchunk, geo->near_copies);
713 return (vchunk << geo->chunk_shift) + offset;
717 * This routine returns the disk from which the requested read should
718 * be done. There is a per-array 'next expected sequential IO' sector
719 * number - if this matches on the next IO then we use the last disk.
720 * There is also a per-disk 'last know head position' sector that is
721 * maintained from IRQ contexts, both the normal and the resync IO
722 * completion handlers update this position correctly. If there is no
723 * perfect sequential match then we pick the disk whose head is closest.
725 * If there are 2 mirrors in the same 2 devices, performance degrades
726 * because position is mirror, not device based.
728 * The rdev for the device selected will have nr_pending incremented.
732 * FIXME: possibly should rethink readbalancing and do it differently
733 * depending on near_copies / far_copies geometry.
735 static struct md_rdev *read_balance(struct r10conf *conf,
736 struct r10bio *r10_bio,
739 const sector_t this_sector = r10_bio->sector;
741 int sectors = r10_bio->sectors;
742 int best_good_sectors;
743 sector_t new_distance, best_dist;
744 struct md_rdev *best_rdev, *rdev = NULL;
747 struct geom *geo = &conf->geo;
749 raid10_find_phys(conf, r10_bio);
751 sectors = r10_bio->sectors;
754 best_dist = MaxSector;
755 best_good_sectors = 0;
757 clear_bit(R10BIO_FailFast, &r10_bio->state);
759 * Check if we can balance. We can balance on the whole
760 * device if no resync is going on (recovery is ok), or below
761 * the resync window. We take the first readable disk when
762 * above the resync window.
764 if (conf->mddev->recovery_cp < MaxSector
765 && (this_sector + sectors >= conf->next_resync))
768 for (slot = 0; slot < conf->copies ; slot++) {
773 if (r10_bio->devs[slot].bio == IO_BLOCKED)
775 disk = r10_bio->devs[slot].devnum;
776 rdev = rcu_dereference(conf->mirrors[disk].replacement);
777 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
778 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
779 rdev = rcu_dereference(conf->mirrors[disk].rdev);
781 test_bit(Faulty, &rdev->flags))
783 if (!test_bit(In_sync, &rdev->flags) &&
784 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
787 dev_sector = r10_bio->devs[slot].addr;
788 if (is_badblock(rdev, dev_sector, sectors,
789 &first_bad, &bad_sectors)) {
790 if (best_dist < MaxSector)
791 /* Already have a better slot */
793 if (first_bad <= dev_sector) {
794 /* Cannot read here. If this is the
795 * 'primary' device, then we must not read
796 * beyond 'bad_sectors' from another device.
798 bad_sectors -= (dev_sector - first_bad);
799 if (!do_balance && sectors > bad_sectors)
800 sectors = bad_sectors;
801 if (best_good_sectors > sectors)
802 best_good_sectors = sectors;
804 sector_t good_sectors =
805 first_bad - dev_sector;
806 if (good_sectors > best_good_sectors) {
807 best_good_sectors = good_sectors;
812 /* Must read from here */
817 best_good_sectors = sectors;
823 /* At least 2 disks to choose from so failfast is OK */
824 set_bit(R10BIO_FailFast, &r10_bio->state);
825 /* This optimisation is debatable, and completely destroys
826 * sequential read speed for 'far copies' arrays. So only
827 * keep it for 'near' arrays, and review those later.
829 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
832 /* for far > 1 always use the lowest address */
833 else if (geo->far_copies > 1)
834 new_distance = r10_bio->devs[slot].addr;
836 new_distance = abs(r10_bio->devs[slot].addr -
837 conf->mirrors[disk].head_position);
838 if (new_distance < best_dist) {
839 best_dist = new_distance;
844 if (slot >= conf->copies) {
850 atomic_inc(&rdev->nr_pending);
851 r10_bio->read_slot = slot;
855 *max_sectors = best_good_sectors;
860 static int raid10_congested(struct mddev *mddev, int bits)
862 struct r10conf *conf = mddev->private;
865 if ((bits & (1 << WB_async_congested)) &&
866 conf->pending_count >= max_queued_requests)
871 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
874 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
875 if (rdev && !test_bit(Faulty, &rdev->flags)) {
876 struct request_queue *q = bdev_get_queue(rdev->bdev);
878 ret |= bdi_congested(q->backing_dev_info, bits);
885 static void flush_pending_writes(struct r10conf *conf)
887 /* Any writes that have been queued but are awaiting
888 * bitmap updates get flushed here.
890 spin_lock_irq(&conf->device_lock);
892 if (conf->pending_bio_list.head) {
894 bio = bio_list_get(&conf->pending_bio_list);
895 conf->pending_count = 0;
896 spin_unlock_irq(&conf->device_lock);
897 /* flush any pending bitmap writes to disk
898 * before proceeding w/ I/O */
899 bitmap_unplug(conf->mddev->bitmap);
900 wake_up(&conf->wait_barrier);
902 while (bio) { /* submit pending writes */
903 struct bio *next = bio->bi_next;
904 struct md_rdev *rdev = (void*)bio->bi_disk;
906 bio_set_dev(bio, rdev->bdev);
907 if (test_bit(Faulty, &rdev->flags)) {
909 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
910 !blk_queue_discard(bio->bi_disk->queue)))
914 generic_make_request(bio);
918 spin_unlock_irq(&conf->device_lock);
922 * Sometimes we need to suspend IO while we do something else,
923 * either some resync/recovery, or reconfigure the array.
924 * To do this we raise a 'barrier'.
925 * The 'barrier' is a counter that can be raised multiple times
926 * to count how many activities are happening which preclude
928 * We can only raise the barrier if there is no pending IO.
929 * i.e. if nr_pending == 0.
930 * We choose only to raise the barrier if no-one is waiting for the
931 * barrier to go down. This means that as soon as an IO request
932 * is ready, no other operations which require a barrier will start
933 * until the IO request has had a chance.
935 * So: regular IO calls 'wait_barrier'. When that returns there
936 * is no backgroup IO happening, It must arrange to call
937 * allow_barrier when it has finished its IO.
938 * backgroup IO calls must call raise_barrier. Once that returns
939 * there is no normal IO happeing. It must arrange to call
940 * lower_barrier when the particular background IO completes.
943 static void raise_barrier(struct r10conf *conf, int force)
945 BUG_ON(force && !conf->barrier);
946 spin_lock_irq(&conf->resync_lock);
948 /* Wait until no block IO is waiting (unless 'force') */
949 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
952 /* block any new IO from starting */
955 /* Now wait for all pending IO to complete */
956 wait_event_lock_irq(conf->wait_barrier,
957 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
960 spin_unlock_irq(&conf->resync_lock);
963 static void lower_barrier(struct r10conf *conf)
966 spin_lock_irqsave(&conf->resync_lock, flags);
968 spin_unlock_irqrestore(&conf->resync_lock, flags);
969 wake_up(&conf->wait_barrier);
972 static void wait_barrier(struct r10conf *conf)
974 spin_lock_irq(&conf->resync_lock);
977 /* Wait for the barrier to drop.
978 * However if there are already pending
979 * requests (preventing the barrier from
980 * rising completely), and the
981 * pre-process bio queue isn't empty,
982 * then don't wait, as we need to empty
983 * that queue to get the nr_pending
986 raid10_log(conf->mddev, "wait barrier");
987 wait_event_lock_irq(conf->wait_barrier,
989 (atomic_read(&conf->nr_pending) &&
991 (!bio_list_empty(¤t->bio_list[0]) ||
992 !bio_list_empty(¤t->bio_list[1]))),
995 if (!conf->nr_waiting)
996 wake_up(&conf->wait_barrier);
998 atomic_inc(&conf->nr_pending);
999 spin_unlock_irq(&conf->resync_lock);
1002 static void allow_barrier(struct r10conf *conf)
1004 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1005 (conf->array_freeze_pending))
1006 wake_up(&conf->wait_barrier);
1009 static void freeze_array(struct r10conf *conf, int extra)
1011 /* stop syncio and normal IO and wait for everything to
1013 * We increment barrier and nr_waiting, and then
1014 * wait until nr_pending match nr_queued+extra
1015 * This is called in the context of one normal IO request
1016 * that has failed. Thus any sync request that might be pending
1017 * will be blocked by nr_pending, and we need to wait for
1018 * pending IO requests to complete or be queued for re-try.
1019 * Thus the number queued (nr_queued) plus this request (extra)
1020 * must match the number of pending IOs (nr_pending) before
1023 spin_lock_irq(&conf->resync_lock);
1024 conf->array_freeze_pending++;
1027 wait_event_lock_irq_cmd(conf->wait_barrier,
1028 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1030 flush_pending_writes(conf));
1032 conf->array_freeze_pending--;
1033 spin_unlock_irq(&conf->resync_lock);
1036 static void unfreeze_array(struct r10conf *conf)
1038 /* reverse the effect of the freeze */
1039 spin_lock_irq(&conf->resync_lock);
1042 wake_up(&conf->wait_barrier);
1043 spin_unlock_irq(&conf->resync_lock);
1046 static sector_t choose_data_offset(struct r10bio *r10_bio,
1047 struct md_rdev *rdev)
1049 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1050 test_bit(R10BIO_Previous, &r10_bio->state))
1051 return rdev->data_offset;
1053 return rdev->new_data_offset;
1056 struct raid10_plug_cb {
1057 struct blk_plug_cb cb;
1058 struct bio_list pending;
1062 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1064 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_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 conf->pending_count += plug->pending_cnt;
1074 spin_unlock_irq(&conf->device_lock);
1075 wake_up(&conf->wait_barrier);
1076 md_wakeup_thread(mddev->thread);
1081 /* we aren't scheduling, so we can do the write-out directly. */
1082 bio = bio_list_get(&plug->pending);
1083 bitmap_unplug(mddev->bitmap);
1084 wake_up(&conf->wait_barrier);
1086 while (bio) { /* submit pending writes */
1087 struct bio *next = bio->bi_next;
1088 struct md_rdev *rdev = (void*)bio->bi_disk;
1089 bio->bi_next = NULL;
1090 bio_set_dev(bio, rdev->bdev);
1091 if (test_bit(Faulty, &rdev->flags)) {
1093 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1094 !blk_queue_discard(bio->bi_disk->queue)))
1095 /* Just ignore it */
1098 generic_make_request(bio);
1104 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1105 struct r10bio *r10_bio)
1107 struct r10conf *conf = mddev->private;
1108 struct bio *read_bio;
1109 const int op = bio_op(bio);
1110 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1113 struct md_rdev *rdev;
1114 char b[BDEVNAME_SIZE];
1115 int slot = r10_bio->read_slot;
1116 struct md_rdev *err_rdev = NULL;
1117 gfp_t gfp = GFP_NOIO;
1119 if (r10_bio->devs[slot].rdev) {
1121 * This is an error retry, but we cannot
1122 * safely dereference the rdev in the r10_bio,
1123 * we must use the one in conf.
1124 * If it has already been disconnected (unlikely)
1125 * we lose the device name in error messages.
1129 * As we are blocking raid10, it is a little safer to
1132 gfp = GFP_NOIO | __GFP_HIGH;
1135 disk = r10_bio->devs[slot].devnum;
1136 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1138 bdevname(err_rdev->bdev, b);
1141 /* This never gets dereferenced */
1142 err_rdev = r10_bio->devs[slot].rdev;
1147 * Register the new request and wait if the reconstruction
1148 * thread has put up a bar for new requests.
1149 * Continue immediately if no resync is active currently.
1153 sectors = r10_bio->sectors;
1154 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1155 bio->bi_iter.bi_sector < conf->reshape_progress &&
1156 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1158 * IO spans the reshape position. Need to wait for reshape to
1161 raid10_log(conf->mddev, "wait reshape");
1162 allow_barrier(conf);
1163 wait_event(conf->wait_barrier,
1164 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1165 conf->reshape_progress >= bio->bi_iter.bi_sector +
1170 rdev = read_balance(conf, r10_bio, &max_sectors);
1173 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1175 (unsigned long long)r10_bio->sector);
1177 raid_end_bio_io(r10_bio);
1181 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1183 bdevname(rdev->bdev, b),
1184 (unsigned long long)r10_bio->sector);
1185 if (max_sectors < bio_sectors(bio)) {
1186 struct bio *split = bio_split(bio, max_sectors,
1187 gfp, conf->bio_split);
1188 bio_chain(split, bio);
1189 generic_make_request(bio);
1191 r10_bio->master_bio = bio;
1192 r10_bio->sectors = max_sectors;
1194 slot = r10_bio->read_slot;
1196 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1198 r10_bio->devs[slot].bio = read_bio;
1199 r10_bio->devs[slot].rdev = rdev;
1201 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1202 choose_data_offset(r10_bio, rdev);
1203 bio_set_dev(read_bio, rdev->bdev);
1204 read_bio->bi_end_io = raid10_end_read_request;
1205 bio_set_op_attrs(read_bio, op, do_sync);
1206 if (test_bit(FailFast, &rdev->flags) &&
1207 test_bit(R10BIO_FailFast, &r10_bio->state))
1208 read_bio->bi_opf |= MD_FAILFAST;
1209 read_bio->bi_private = r10_bio;
1212 trace_block_bio_remap(read_bio->bi_disk->queue,
1213 read_bio, disk_devt(mddev->gendisk),
1215 generic_make_request(read_bio);
1219 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1220 struct bio *bio, bool replacement,
1223 const int op = bio_op(bio);
1224 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1225 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1226 unsigned long flags;
1227 struct blk_plug_cb *cb;
1228 struct raid10_plug_cb *plug = NULL;
1229 struct r10conf *conf = mddev->private;
1230 struct md_rdev *rdev;
1231 int devnum = r10_bio->devs[n_copy].devnum;
1235 rdev = conf->mirrors[devnum].replacement;
1237 /* Replacement just got moved to main 'rdev' */
1239 rdev = conf->mirrors[devnum].rdev;
1242 rdev = conf->mirrors[devnum].rdev;
1244 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1246 r10_bio->devs[n_copy].repl_bio = mbio;
1248 r10_bio->devs[n_copy].bio = mbio;
1250 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1251 choose_data_offset(r10_bio, rdev));
1252 bio_set_dev(mbio, rdev->bdev);
1253 mbio->bi_end_io = raid10_end_write_request;
1254 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1255 if (!replacement && test_bit(FailFast,
1256 &conf->mirrors[devnum].rdev->flags)
1257 && enough(conf, devnum))
1258 mbio->bi_opf |= MD_FAILFAST;
1259 mbio->bi_private = r10_bio;
1261 if (conf->mddev->gendisk)
1262 trace_block_bio_remap(mbio->bi_disk->queue,
1263 mbio, disk_devt(conf->mddev->gendisk),
1265 /* flush_pending_writes() needs access to the rdev so...*/
1266 mbio->bi_disk = (void *)rdev;
1268 atomic_inc(&r10_bio->remaining);
1270 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1272 plug = container_of(cb, struct raid10_plug_cb, cb);
1276 bio_list_add(&plug->pending, mbio);
1277 plug->pending_cnt++;
1279 spin_lock_irqsave(&conf->device_lock, flags);
1280 bio_list_add(&conf->pending_bio_list, mbio);
1281 conf->pending_count++;
1282 spin_unlock_irqrestore(&conf->device_lock, flags);
1283 md_wakeup_thread(mddev->thread);
1287 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1288 struct r10bio *r10_bio)
1290 struct r10conf *conf = mddev->private;
1292 struct md_rdev *blocked_rdev;
1297 * Register the new request and wait if the reconstruction
1298 * thread has put up a bar for new requests.
1299 * Continue immediately if no resync is active currently.
1303 sectors = r10_bio->sectors;
1304 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1305 bio->bi_iter.bi_sector < conf->reshape_progress &&
1306 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1308 * IO spans the reshape position. Need to wait for reshape to
1311 raid10_log(conf->mddev, "wait reshape");
1312 allow_barrier(conf);
1313 wait_event(conf->wait_barrier,
1314 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1315 conf->reshape_progress >= bio->bi_iter.bi_sector +
1320 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1321 (mddev->reshape_backwards
1322 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1323 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1324 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1325 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1326 /* Need to update reshape_position in metadata */
1327 mddev->reshape_position = conf->reshape_progress;
1328 set_mask_bits(&mddev->sb_flags, 0,
1329 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1330 md_wakeup_thread(mddev->thread);
1331 raid10_log(conf->mddev, "wait reshape metadata");
1332 wait_event(mddev->sb_wait,
1333 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1335 conf->reshape_safe = mddev->reshape_position;
1338 if (conf->pending_count >= max_queued_requests) {
1339 md_wakeup_thread(mddev->thread);
1340 raid10_log(mddev, "wait queued");
1341 wait_event(conf->wait_barrier,
1342 conf->pending_count < max_queued_requests);
1344 /* first select target devices under rcu_lock and
1345 * inc refcount on their rdev. Record them by setting
1347 * If there are known/acknowledged bad blocks on any device
1348 * on which we have seen a write error, we want to avoid
1349 * writing to those blocks. This potentially requires several
1350 * writes to write around the bad blocks. Each set of writes
1351 * gets its own r10_bio with a set of bios attached.
1354 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1355 raid10_find_phys(conf, r10_bio);
1357 blocked_rdev = NULL;
1359 max_sectors = r10_bio->sectors;
1361 for (i = 0; i < conf->copies; i++) {
1362 int d = r10_bio->devs[i].devnum;
1363 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1364 struct md_rdev *rrdev = rcu_dereference(
1365 conf->mirrors[d].replacement);
1368 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1369 atomic_inc(&rdev->nr_pending);
1370 blocked_rdev = rdev;
1373 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1374 atomic_inc(&rrdev->nr_pending);
1375 blocked_rdev = rrdev;
1378 if (rdev && (test_bit(Faulty, &rdev->flags)))
1380 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1383 r10_bio->devs[i].bio = NULL;
1384 r10_bio->devs[i].repl_bio = NULL;
1386 if (!rdev && !rrdev) {
1387 set_bit(R10BIO_Degraded, &r10_bio->state);
1390 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1392 sector_t dev_sector = r10_bio->devs[i].addr;
1396 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1397 &first_bad, &bad_sectors);
1399 /* Mustn't write here until the bad block
1402 atomic_inc(&rdev->nr_pending);
1403 set_bit(BlockedBadBlocks, &rdev->flags);
1404 blocked_rdev = rdev;
1407 if (is_bad && first_bad <= dev_sector) {
1408 /* Cannot write here at all */
1409 bad_sectors -= (dev_sector - first_bad);
1410 if (bad_sectors < max_sectors)
1411 /* Mustn't write more than bad_sectors
1412 * to other devices yet
1414 max_sectors = bad_sectors;
1415 /* We don't set R10BIO_Degraded as that
1416 * only applies if the disk is missing,
1417 * so it might be re-added, and we want to
1418 * know to recover this chunk.
1419 * In this case the device is here, and the
1420 * fact that this chunk is not in-sync is
1421 * recorded in the bad block log.
1426 int good_sectors = first_bad - dev_sector;
1427 if (good_sectors < max_sectors)
1428 max_sectors = good_sectors;
1432 r10_bio->devs[i].bio = bio;
1433 atomic_inc(&rdev->nr_pending);
1436 r10_bio->devs[i].repl_bio = bio;
1437 atomic_inc(&rrdev->nr_pending);
1442 if (unlikely(blocked_rdev)) {
1443 /* Have to wait for this device to get unblocked, then retry */
1447 for (j = 0; j < i; j++) {
1448 if (r10_bio->devs[j].bio) {
1449 d = r10_bio->devs[j].devnum;
1450 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1452 if (r10_bio->devs[j].repl_bio) {
1453 struct md_rdev *rdev;
1454 d = r10_bio->devs[j].devnum;
1455 rdev = conf->mirrors[d].replacement;
1457 /* Race with remove_disk */
1459 rdev = conf->mirrors[d].rdev;
1461 rdev_dec_pending(rdev, mddev);
1464 allow_barrier(conf);
1465 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1466 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1471 if (max_sectors < r10_bio->sectors)
1472 r10_bio->sectors = max_sectors;
1474 if (r10_bio->sectors < bio_sectors(bio)) {
1475 struct bio *split = bio_split(bio, r10_bio->sectors,
1476 GFP_NOIO, conf->bio_split);
1477 bio_chain(split, bio);
1478 generic_make_request(bio);
1480 r10_bio->master_bio = bio;
1483 atomic_set(&r10_bio->remaining, 1);
1484 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1486 for (i = 0; i < conf->copies; i++) {
1487 if (r10_bio->devs[i].bio)
1488 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1489 if (r10_bio->devs[i].repl_bio)
1490 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1492 one_write_done(r10_bio);
1495 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1497 struct r10conf *conf = mddev->private;
1498 struct r10bio *r10_bio;
1500 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1502 r10_bio->master_bio = bio;
1503 r10_bio->sectors = sectors;
1505 r10_bio->mddev = mddev;
1506 r10_bio->sector = bio->bi_iter.bi_sector;
1508 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1510 if (bio_data_dir(bio) == READ)
1511 raid10_read_request(mddev, bio, r10_bio);
1513 raid10_write_request(mddev, bio, r10_bio);
1516 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1518 struct r10conf *conf = mddev->private;
1519 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1520 int chunk_sects = chunk_mask + 1;
1521 int sectors = bio_sectors(bio);
1523 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1524 md_flush_request(mddev, bio);
1528 if (!md_write_start(mddev, bio))
1532 * If this request crosses a chunk boundary, we need to split
1535 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1536 sectors > chunk_sects
1537 && (conf->geo.near_copies < conf->geo.raid_disks
1538 || conf->prev.near_copies <
1539 conf->prev.raid_disks)))
1540 sectors = chunk_sects -
1541 (bio->bi_iter.bi_sector &
1543 __make_request(mddev, bio, sectors);
1545 /* In case raid10d snuck in to freeze_array */
1546 wake_up(&conf->wait_barrier);
1550 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1552 struct r10conf *conf = mddev->private;
1555 if (conf->geo.near_copies < conf->geo.raid_disks)
1556 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1557 if (conf->geo.near_copies > 1)
1558 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1559 if (conf->geo.far_copies > 1) {
1560 if (conf->geo.far_offset)
1561 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1563 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1564 if (conf->geo.far_set_size != conf->geo.raid_disks)
1565 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1567 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1568 conf->geo.raid_disks - mddev->degraded);
1570 for (i = 0; i < conf->geo.raid_disks; i++) {
1571 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1572 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1575 seq_printf(seq, "]");
1578 /* check if there are enough drives for
1579 * every block to appear on atleast one.
1580 * Don't consider the device numbered 'ignore'
1581 * as we might be about to remove it.
1583 static int _enough(struct r10conf *conf, int previous, int ignore)
1589 disks = conf->prev.raid_disks;
1590 ncopies = conf->prev.near_copies;
1592 disks = conf->geo.raid_disks;
1593 ncopies = conf->geo.near_copies;
1598 int n = conf->copies;
1602 struct md_rdev *rdev;
1603 if (this != ignore &&
1604 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1605 test_bit(In_sync, &rdev->flags))
1607 this = (this+1) % disks;
1611 first = (first + ncopies) % disks;
1612 } while (first != 0);
1619 static int enough(struct r10conf *conf, int ignore)
1621 /* when calling 'enough', both 'prev' and 'geo' must
1623 * This is ensured if ->reconfig_mutex or ->device_lock
1626 return _enough(conf, 0, ignore) &&
1627 _enough(conf, 1, ignore);
1630 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1632 char b[BDEVNAME_SIZE];
1633 struct r10conf *conf = mddev->private;
1634 unsigned long flags;
1637 * If it is not operational, then we have already marked it as dead
1638 * else if it is the last working disks, ignore the error, let the
1639 * next level up know.
1640 * else mark the drive as failed
1642 spin_lock_irqsave(&conf->device_lock, flags);
1643 if (test_bit(In_sync, &rdev->flags)
1644 && !enough(conf, rdev->raid_disk)) {
1646 * Don't fail the drive, just return an IO error.
1648 spin_unlock_irqrestore(&conf->device_lock, flags);
1651 if (test_and_clear_bit(In_sync, &rdev->flags))
1654 * If recovery is running, make sure it aborts.
1656 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1657 set_bit(Blocked, &rdev->flags);
1658 set_bit(Faulty, &rdev->flags);
1659 set_mask_bits(&mddev->sb_flags, 0,
1660 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1661 spin_unlock_irqrestore(&conf->device_lock, flags);
1662 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1663 "md/raid10:%s: Operation continuing on %d devices.\n",
1664 mdname(mddev), bdevname(rdev->bdev, b),
1665 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1668 static void print_conf(struct r10conf *conf)
1671 struct md_rdev *rdev;
1673 pr_debug("RAID10 conf printout:\n");
1675 pr_debug("(!conf)\n");
1678 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1679 conf->geo.raid_disks);
1681 /* This is only called with ->reconfix_mutex held, so
1682 * rcu protection of rdev is not needed */
1683 for (i = 0; i < conf->geo.raid_disks; i++) {
1684 char b[BDEVNAME_SIZE];
1685 rdev = conf->mirrors[i].rdev;
1687 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1688 i, !test_bit(In_sync, &rdev->flags),
1689 !test_bit(Faulty, &rdev->flags),
1690 bdevname(rdev->bdev,b));
1694 static void close_sync(struct r10conf *conf)
1697 allow_barrier(conf);
1699 mempool_destroy(conf->r10buf_pool);
1700 conf->r10buf_pool = NULL;
1703 static int raid10_spare_active(struct mddev *mddev)
1706 struct r10conf *conf = mddev->private;
1707 struct raid10_info *tmp;
1709 unsigned long flags;
1712 * Find all non-in_sync disks within the RAID10 configuration
1713 * and mark them in_sync
1715 for (i = 0; i < conf->geo.raid_disks; i++) {
1716 tmp = conf->mirrors + i;
1717 if (tmp->replacement
1718 && tmp->replacement->recovery_offset == MaxSector
1719 && !test_bit(Faulty, &tmp->replacement->flags)
1720 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1721 /* Replacement has just become active */
1723 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1726 /* Replaced device not technically faulty,
1727 * but we need to be sure it gets removed
1728 * and never re-added.
1730 set_bit(Faulty, &tmp->rdev->flags);
1731 sysfs_notify_dirent_safe(
1732 tmp->rdev->sysfs_state);
1734 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1735 } else if (tmp->rdev
1736 && tmp->rdev->recovery_offset == MaxSector
1737 && !test_bit(Faulty, &tmp->rdev->flags)
1738 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1740 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1743 spin_lock_irqsave(&conf->device_lock, flags);
1744 mddev->degraded -= count;
1745 spin_unlock_irqrestore(&conf->device_lock, flags);
1751 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1753 struct r10conf *conf = mddev->private;
1757 int last = conf->geo.raid_disks - 1;
1759 if (mddev->recovery_cp < MaxSector)
1760 /* only hot-add to in-sync arrays, as recovery is
1761 * very different from resync
1764 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1767 if (md_integrity_add_rdev(rdev, mddev))
1770 if (rdev->raid_disk >= 0)
1771 first = last = rdev->raid_disk;
1773 if (rdev->saved_raid_disk >= first &&
1774 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1775 mirror = rdev->saved_raid_disk;
1778 for ( ; mirror <= last ; mirror++) {
1779 struct raid10_info *p = &conf->mirrors[mirror];
1780 if (p->recovery_disabled == mddev->recovery_disabled)
1783 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1784 p->replacement != NULL)
1786 clear_bit(In_sync, &rdev->flags);
1787 set_bit(Replacement, &rdev->flags);
1788 rdev->raid_disk = mirror;
1791 disk_stack_limits(mddev->gendisk, rdev->bdev,
1792 rdev->data_offset << 9);
1794 rcu_assign_pointer(p->replacement, rdev);
1799 disk_stack_limits(mddev->gendisk, rdev->bdev,
1800 rdev->data_offset << 9);
1802 p->head_position = 0;
1803 p->recovery_disabled = mddev->recovery_disabled - 1;
1804 rdev->raid_disk = mirror;
1806 if (rdev->saved_raid_disk != mirror)
1808 rcu_assign_pointer(p->rdev, rdev);
1811 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1812 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1818 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1820 struct r10conf *conf = mddev->private;
1822 int number = rdev->raid_disk;
1823 struct md_rdev **rdevp;
1824 struct raid10_info *p = conf->mirrors + number;
1827 if (rdev == p->rdev)
1829 else if (rdev == p->replacement)
1830 rdevp = &p->replacement;
1834 if (test_bit(In_sync, &rdev->flags) ||
1835 atomic_read(&rdev->nr_pending)) {
1839 /* Only remove non-faulty devices if recovery
1842 if (!test_bit(Faulty, &rdev->flags) &&
1843 mddev->recovery_disabled != p->recovery_disabled &&
1844 (!p->replacement || p->replacement == rdev) &&
1845 number < conf->geo.raid_disks &&
1851 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1853 if (atomic_read(&rdev->nr_pending)) {
1854 /* lost the race, try later */
1860 if (p->replacement) {
1861 /* We must have just cleared 'rdev' */
1862 p->rdev = p->replacement;
1863 clear_bit(Replacement, &p->replacement->flags);
1864 smp_mb(); /* Make sure other CPUs may see both as identical
1865 * but will never see neither -- if they are careful.
1867 p->replacement = NULL;
1870 clear_bit(WantReplacement, &rdev->flags);
1871 err = md_integrity_register(mddev);
1879 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1881 struct r10conf *conf = r10_bio->mddev->private;
1883 if (!bio->bi_status)
1884 set_bit(R10BIO_Uptodate, &r10_bio->state);
1886 /* The write handler will notice the lack of
1887 * R10BIO_Uptodate and record any errors etc
1889 atomic_add(r10_bio->sectors,
1890 &conf->mirrors[d].rdev->corrected_errors);
1892 /* for reconstruct, we always reschedule after a read.
1893 * for resync, only after all reads
1895 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1896 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1897 atomic_dec_and_test(&r10_bio->remaining)) {
1898 /* we have read all the blocks,
1899 * do the comparison in process context in raid10d
1901 reschedule_retry(r10_bio);
1905 static void end_sync_read(struct bio *bio)
1907 struct r10bio *r10_bio = get_resync_r10bio(bio);
1908 struct r10conf *conf = r10_bio->mddev->private;
1909 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1911 __end_sync_read(r10_bio, bio, d);
1914 static void end_reshape_read(struct bio *bio)
1916 /* reshape read bio isn't allocated from r10buf_pool */
1917 struct r10bio *r10_bio = bio->bi_private;
1919 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1922 static void end_sync_request(struct r10bio *r10_bio)
1924 struct mddev *mddev = r10_bio->mddev;
1926 while (atomic_dec_and_test(&r10_bio->remaining)) {
1927 if (r10_bio->master_bio == NULL) {
1928 /* the primary of several recovery bios */
1929 sector_t s = r10_bio->sectors;
1930 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1931 test_bit(R10BIO_WriteError, &r10_bio->state))
1932 reschedule_retry(r10_bio);
1935 md_done_sync(mddev, s, 1);
1938 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1939 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1940 test_bit(R10BIO_WriteError, &r10_bio->state))
1941 reschedule_retry(r10_bio);
1949 static void end_sync_write(struct bio *bio)
1951 struct r10bio *r10_bio = get_resync_r10bio(bio);
1952 struct mddev *mddev = r10_bio->mddev;
1953 struct r10conf *conf = mddev->private;
1959 struct md_rdev *rdev = NULL;
1961 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1963 rdev = conf->mirrors[d].replacement;
1965 rdev = conf->mirrors[d].rdev;
1967 if (bio->bi_status) {
1969 md_error(mddev, rdev);
1971 set_bit(WriteErrorSeen, &rdev->flags);
1972 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1973 set_bit(MD_RECOVERY_NEEDED,
1974 &rdev->mddev->recovery);
1975 set_bit(R10BIO_WriteError, &r10_bio->state);
1977 } else if (is_badblock(rdev,
1978 r10_bio->devs[slot].addr,
1980 &first_bad, &bad_sectors))
1981 set_bit(R10BIO_MadeGood, &r10_bio->state);
1983 rdev_dec_pending(rdev, mddev);
1985 end_sync_request(r10_bio);
1989 * Note: sync and recover and handled very differently for raid10
1990 * This code is for resync.
1991 * For resync, we read through virtual addresses and read all blocks.
1992 * If there is any error, we schedule a write. The lowest numbered
1993 * drive is authoritative.
1994 * However requests come for physical address, so we need to map.
1995 * For every physical address there are raid_disks/copies virtual addresses,
1996 * which is always are least one, but is not necessarly an integer.
1997 * This means that a physical address can span multiple chunks, so we may
1998 * have to submit multiple io requests for a single sync request.
2001 * We check if all blocks are in-sync and only write to blocks that
2004 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2006 struct r10conf *conf = mddev->private;
2008 struct bio *tbio, *fbio;
2010 struct page **tpages, **fpages;
2012 atomic_set(&r10_bio->remaining, 1);
2014 /* find the first device with a block */
2015 for (i=0; i<conf->copies; i++)
2016 if (!r10_bio->devs[i].bio->bi_status)
2019 if (i == conf->copies)
2023 fbio = r10_bio->devs[i].bio;
2024 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2025 fbio->bi_iter.bi_idx = 0;
2026 fpages = get_resync_pages(fbio)->pages;
2028 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2029 /* now find blocks with errors */
2030 for (i=0 ; i < conf->copies ; i++) {
2032 struct md_rdev *rdev;
2033 struct resync_pages *rp;
2035 tbio = r10_bio->devs[i].bio;
2037 if (tbio->bi_end_io != end_sync_read)
2042 tpages = get_resync_pages(tbio)->pages;
2043 d = r10_bio->devs[i].devnum;
2044 rdev = conf->mirrors[d].rdev;
2045 if (!r10_bio->devs[i].bio->bi_status) {
2046 /* We know that the bi_io_vec layout is the same for
2047 * both 'first' and 'i', so we just compare them.
2048 * All vec entries are PAGE_SIZE;
2050 int sectors = r10_bio->sectors;
2051 for (j = 0; j < vcnt; j++) {
2052 int len = PAGE_SIZE;
2053 if (sectors < (len / 512))
2054 len = sectors * 512;
2055 if (memcmp(page_address(fpages[j]),
2056 page_address(tpages[j]),
2063 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2064 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2065 /* Don't fix anything. */
2067 } else if (test_bit(FailFast, &rdev->flags)) {
2068 /* Just give up on this device */
2069 md_error(rdev->mddev, rdev);
2072 /* Ok, we need to write this bio, either to correct an
2073 * inconsistency or to correct an unreadable block.
2074 * First we need to fixup bv_offset, bv_len and
2075 * bi_vecs, as the read request might have corrupted these
2077 rp = get_resync_pages(tbio);
2080 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2082 rp->raid_bio = r10_bio;
2083 tbio->bi_private = rp;
2084 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2085 tbio->bi_end_io = end_sync_write;
2086 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2088 bio_copy_data(tbio, fbio);
2090 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2091 atomic_inc(&r10_bio->remaining);
2092 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2094 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2095 tbio->bi_opf |= MD_FAILFAST;
2096 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2097 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2098 generic_make_request(tbio);
2101 /* Now write out to any replacement devices
2104 for (i = 0; i < conf->copies; i++) {
2107 tbio = r10_bio->devs[i].repl_bio;
2108 if (!tbio || !tbio->bi_end_io)
2110 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2111 && r10_bio->devs[i].bio != fbio)
2112 bio_copy_data(tbio, fbio);
2113 d = r10_bio->devs[i].devnum;
2114 atomic_inc(&r10_bio->remaining);
2115 md_sync_acct(conf->mirrors[d].replacement->bdev,
2117 generic_make_request(tbio);
2121 if (atomic_dec_and_test(&r10_bio->remaining)) {
2122 md_done_sync(mddev, r10_bio->sectors, 1);
2128 * Now for the recovery code.
2129 * Recovery happens across physical sectors.
2130 * We recover all non-is_sync drives by finding the virtual address of
2131 * each, and then choose a working drive that also has that virt address.
2132 * There is a separate r10_bio for each non-in_sync drive.
2133 * Only the first two slots are in use. The first for reading,
2134 * The second for writing.
2137 static void fix_recovery_read_error(struct r10bio *r10_bio)
2139 /* We got a read error during recovery.
2140 * We repeat the read in smaller page-sized sections.
2141 * If a read succeeds, write it to the new device or record
2142 * a bad block if we cannot.
2143 * If a read fails, record a bad block on both old and
2146 struct mddev *mddev = r10_bio->mddev;
2147 struct r10conf *conf = mddev->private;
2148 struct bio *bio = r10_bio->devs[0].bio;
2150 int sectors = r10_bio->sectors;
2152 int dr = r10_bio->devs[0].devnum;
2153 int dw = r10_bio->devs[1].devnum;
2154 struct page **pages = get_resync_pages(bio)->pages;
2158 struct md_rdev *rdev;
2162 if (s > (PAGE_SIZE>>9))
2165 rdev = conf->mirrors[dr].rdev;
2166 addr = r10_bio->devs[0].addr + sect,
2167 ok = sync_page_io(rdev,
2171 REQ_OP_READ, 0, false);
2173 rdev = conf->mirrors[dw].rdev;
2174 addr = r10_bio->devs[1].addr + sect;
2175 ok = sync_page_io(rdev,
2179 REQ_OP_WRITE, 0, false);
2181 set_bit(WriteErrorSeen, &rdev->flags);
2182 if (!test_and_set_bit(WantReplacement,
2184 set_bit(MD_RECOVERY_NEEDED,
2185 &rdev->mddev->recovery);
2189 /* We don't worry if we cannot set a bad block -
2190 * it really is bad so there is no loss in not
2193 rdev_set_badblocks(rdev, addr, s, 0);
2195 if (rdev != conf->mirrors[dw].rdev) {
2196 /* need bad block on destination too */
2197 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2198 addr = r10_bio->devs[1].addr + sect;
2199 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2201 /* just abort the recovery */
2202 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2205 conf->mirrors[dw].recovery_disabled
2206 = mddev->recovery_disabled;
2207 set_bit(MD_RECOVERY_INTR,
2220 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2222 struct r10conf *conf = mddev->private;
2224 struct bio *wbio, *wbio2;
2226 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2227 fix_recovery_read_error(r10_bio);
2228 end_sync_request(r10_bio);
2233 * share the pages with the first bio
2234 * and submit the write request
2236 d = r10_bio->devs[1].devnum;
2237 wbio = r10_bio->devs[1].bio;
2238 wbio2 = r10_bio->devs[1].repl_bio;
2239 /* Need to test wbio2->bi_end_io before we call
2240 * generic_make_request as if the former is NULL,
2241 * the latter is free to free wbio2.
2243 if (wbio2 && !wbio2->bi_end_io)
2245 if (wbio->bi_end_io) {
2246 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2247 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2248 generic_make_request(wbio);
2251 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2252 md_sync_acct(conf->mirrors[d].replacement->bdev,
2253 bio_sectors(wbio2));
2254 generic_make_request(wbio2);
2259 * Used by fix_read_error() to decay the per rdev read_errors.
2260 * We halve the read error count for every hour that has elapsed
2261 * since the last recorded read error.
2264 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2267 unsigned long hours_since_last;
2268 unsigned int read_errors = atomic_read(&rdev->read_errors);
2270 cur_time_mon = ktime_get_seconds();
2272 if (rdev->last_read_error == 0) {
2273 /* first time we've seen a read error */
2274 rdev->last_read_error = cur_time_mon;
2278 hours_since_last = (long)(cur_time_mon -
2279 rdev->last_read_error) / 3600;
2281 rdev->last_read_error = cur_time_mon;
2284 * if hours_since_last is > the number of bits in read_errors
2285 * just set read errors to 0. We do this to avoid
2286 * overflowing the shift of read_errors by hours_since_last.
2288 if (hours_since_last >= 8 * sizeof(read_errors))
2289 atomic_set(&rdev->read_errors, 0);
2291 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2294 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2295 int sectors, struct page *page, int rw)
2300 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2301 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2303 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2307 set_bit(WriteErrorSeen, &rdev->flags);
2308 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2309 set_bit(MD_RECOVERY_NEEDED,
2310 &rdev->mddev->recovery);
2312 /* need to record an error - either for the block or the device */
2313 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2314 md_error(rdev->mddev, rdev);
2319 * This is a kernel thread which:
2321 * 1. Retries failed read operations on working mirrors.
2322 * 2. Updates the raid superblock when problems encounter.
2323 * 3. Performs writes following reads for array synchronising.
2326 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2328 int sect = 0; /* Offset from r10_bio->sector */
2329 int sectors = r10_bio->sectors;
2330 struct md_rdev*rdev;
2331 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2332 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2334 /* still own a reference to this rdev, so it cannot
2335 * have been cleared recently.
2337 rdev = conf->mirrors[d].rdev;
2339 if (test_bit(Faulty, &rdev->flags))
2340 /* drive has already been failed, just ignore any
2341 more fix_read_error() attempts */
2344 check_decay_read_errors(mddev, rdev);
2345 atomic_inc(&rdev->read_errors);
2346 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2347 char b[BDEVNAME_SIZE];
2348 bdevname(rdev->bdev, b);
2350 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2352 atomic_read(&rdev->read_errors), max_read_errors);
2353 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2355 md_error(mddev, rdev);
2356 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2362 int sl = r10_bio->read_slot;
2366 if (s > (PAGE_SIZE>>9))
2374 d = r10_bio->devs[sl].devnum;
2375 rdev = rcu_dereference(conf->mirrors[d].rdev);
2377 test_bit(In_sync, &rdev->flags) &&
2378 !test_bit(Faulty, &rdev->flags) &&
2379 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2380 &first_bad, &bad_sectors) == 0) {
2381 atomic_inc(&rdev->nr_pending);
2383 success = sync_page_io(rdev,
2384 r10_bio->devs[sl].addr +
2388 REQ_OP_READ, 0, false);
2389 rdev_dec_pending(rdev, mddev);
2395 if (sl == conf->copies)
2397 } while (!success && sl != r10_bio->read_slot);
2401 /* Cannot read from anywhere, just mark the block
2402 * as bad on the first device to discourage future
2405 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2406 rdev = conf->mirrors[dn].rdev;
2408 if (!rdev_set_badblocks(
2410 r10_bio->devs[r10_bio->read_slot].addr
2413 md_error(mddev, rdev);
2414 r10_bio->devs[r10_bio->read_slot].bio
2421 /* write it back and re-read */
2423 while (sl != r10_bio->read_slot) {
2424 char b[BDEVNAME_SIZE];
2429 d = r10_bio->devs[sl].devnum;
2430 rdev = rcu_dereference(conf->mirrors[d].rdev);
2432 test_bit(Faulty, &rdev->flags) ||
2433 !test_bit(In_sync, &rdev->flags))
2436 atomic_inc(&rdev->nr_pending);
2438 if (r10_sync_page_io(rdev,
2439 r10_bio->devs[sl].addr +
2441 s, conf->tmppage, WRITE)
2443 /* Well, this device is dead */
2444 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2446 (unsigned long long)(
2448 choose_data_offset(r10_bio,
2450 bdevname(rdev->bdev, b));
2451 pr_notice("md/raid10:%s: %s: failing drive\n",
2453 bdevname(rdev->bdev, b));
2455 rdev_dec_pending(rdev, mddev);
2459 while (sl != r10_bio->read_slot) {
2460 char b[BDEVNAME_SIZE];
2465 d = r10_bio->devs[sl].devnum;
2466 rdev = rcu_dereference(conf->mirrors[d].rdev);
2468 test_bit(Faulty, &rdev->flags) ||
2469 !test_bit(In_sync, &rdev->flags))
2472 atomic_inc(&rdev->nr_pending);
2474 switch (r10_sync_page_io(rdev,
2475 r10_bio->devs[sl].addr +
2480 /* Well, this device is dead */
2481 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2483 (unsigned long long)(
2485 choose_data_offset(r10_bio, rdev)),
2486 bdevname(rdev->bdev, b));
2487 pr_notice("md/raid10:%s: %s: failing drive\n",
2489 bdevname(rdev->bdev, b));
2492 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2494 (unsigned long long)(
2496 choose_data_offset(r10_bio, rdev)),
2497 bdevname(rdev->bdev, b));
2498 atomic_add(s, &rdev->corrected_errors);
2501 rdev_dec_pending(rdev, mddev);
2511 static int narrow_write_error(struct r10bio *r10_bio, int i)
2513 struct bio *bio = r10_bio->master_bio;
2514 struct mddev *mddev = r10_bio->mddev;
2515 struct r10conf *conf = mddev->private;
2516 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2517 /* bio has the data to be written to slot 'i' where
2518 * we just recently had a write error.
2519 * We repeatedly clone the bio and trim down to one block,
2520 * then try the write. Where the write fails we record
2522 * It is conceivable that the bio doesn't exactly align with
2523 * blocks. We must handle this.
2525 * We currently own a reference to the rdev.
2531 int sect_to_write = r10_bio->sectors;
2534 if (rdev->badblocks.shift < 0)
2537 block_sectors = roundup(1 << rdev->badblocks.shift,
2538 bdev_logical_block_size(rdev->bdev) >> 9);
2539 sector = r10_bio->sector;
2540 sectors = ((r10_bio->sector + block_sectors)
2541 & ~(sector_t)(block_sectors - 1))
2544 while (sect_to_write) {
2547 if (sectors > sect_to_write)
2548 sectors = sect_to_write;
2549 /* Write at 'sector' for 'sectors' */
2550 wbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
2551 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2552 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2553 wbio->bi_iter.bi_sector = wsector +
2554 choose_data_offset(r10_bio, rdev);
2555 bio_set_dev(wbio, rdev->bdev);
2556 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2558 if (submit_bio_wait(wbio) < 0)
2560 ok = rdev_set_badblocks(rdev, wsector,
2565 sect_to_write -= sectors;
2567 sectors = block_sectors;
2572 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2574 int slot = r10_bio->read_slot;
2576 struct r10conf *conf = mddev->private;
2577 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2578 sector_t bio_last_sector;
2580 /* we got a read error. Maybe the drive is bad. Maybe just
2581 * the block and we can fix it.
2582 * We freeze all other IO, and try reading the block from
2583 * other devices. When we find one, we re-write
2584 * and check it that fixes the read error.
2585 * This is all done synchronously while the array is
2588 bio = r10_bio->devs[slot].bio;
2589 bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
2591 r10_bio->devs[slot].bio = NULL;
2594 r10_bio->devs[slot].bio = IO_BLOCKED;
2595 else if (!test_bit(FailFast, &rdev->flags)) {
2596 freeze_array(conf, 1);
2597 fix_read_error(conf, mddev, r10_bio);
2598 unfreeze_array(conf);
2600 md_error(mddev, rdev);
2602 rdev_dec_pending(rdev, mddev);
2603 allow_barrier(conf);
2605 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2608 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2610 /* Some sort of write request has finished and it
2611 * succeeded in writing where we thought there was a
2612 * bad block. So forget the bad block.
2613 * Or possibly if failed and we need to record
2617 struct md_rdev *rdev;
2619 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2620 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2621 for (m = 0; m < conf->copies; m++) {
2622 int dev = r10_bio->devs[m].devnum;
2623 rdev = conf->mirrors[dev].rdev;
2624 if (r10_bio->devs[m].bio == NULL)
2626 if (!r10_bio->devs[m].bio->bi_status) {
2627 rdev_clear_badblocks(
2629 r10_bio->devs[m].addr,
2630 r10_bio->sectors, 0);
2632 if (!rdev_set_badblocks(
2634 r10_bio->devs[m].addr,
2635 r10_bio->sectors, 0))
2636 md_error(conf->mddev, rdev);
2638 rdev = conf->mirrors[dev].replacement;
2639 if (r10_bio->devs[m].repl_bio == NULL)
2642 if (!r10_bio->devs[m].repl_bio->bi_status) {
2643 rdev_clear_badblocks(
2645 r10_bio->devs[m].addr,
2646 r10_bio->sectors, 0);
2648 if (!rdev_set_badblocks(
2650 r10_bio->devs[m].addr,
2651 r10_bio->sectors, 0))
2652 md_error(conf->mddev, rdev);
2658 for (m = 0; m < conf->copies; m++) {
2659 int dev = r10_bio->devs[m].devnum;
2660 struct bio *bio = r10_bio->devs[m].bio;
2661 rdev = conf->mirrors[dev].rdev;
2662 if (bio == IO_MADE_GOOD) {
2663 rdev_clear_badblocks(
2665 r10_bio->devs[m].addr,
2666 r10_bio->sectors, 0);
2667 rdev_dec_pending(rdev, conf->mddev);
2668 } else if (bio != NULL && bio->bi_status) {
2670 if (!narrow_write_error(r10_bio, m)) {
2671 md_error(conf->mddev, rdev);
2672 set_bit(R10BIO_Degraded,
2675 rdev_dec_pending(rdev, conf->mddev);
2677 bio = r10_bio->devs[m].repl_bio;
2678 rdev = conf->mirrors[dev].replacement;
2679 if (rdev && bio == IO_MADE_GOOD) {
2680 rdev_clear_badblocks(
2682 r10_bio->devs[m].addr,
2683 r10_bio->sectors, 0);
2684 rdev_dec_pending(rdev, conf->mddev);
2688 spin_lock_irq(&conf->device_lock);
2689 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2691 spin_unlock_irq(&conf->device_lock);
2693 * In case freeze_array() is waiting for condition
2694 * nr_pending == nr_queued + extra to be true.
2696 wake_up(&conf->wait_barrier);
2697 md_wakeup_thread(conf->mddev->thread);
2699 if (test_bit(R10BIO_WriteError,
2701 close_write(r10_bio);
2702 raid_end_bio_io(r10_bio);
2707 static void raid10d(struct md_thread *thread)
2709 struct mddev *mddev = thread->mddev;
2710 struct r10bio *r10_bio;
2711 unsigned long flags;
2712 struct r10conf *conf = mddev->private;
2713 struct list_head *head = &conf->retry_list;
2714 struct blk_plug plug;
2716 md_check_recovery(mddev);
2718 if (!list_empty_careful(&conf->bio_end_io_list) &&
2719 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2721 spin_lock_irqsave(&conf->device_lock, flags);
2722 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2723 while (!list_empty(&conf->bio_end_io_list)) {
2724 list_move(conf->bio_end_io_list.prev, &tmp);
2728 spin_unlock_irqrestore(&conf->device_lock, flags);
2729 while (!list_empty(&tmp)) {
2730 r10_bio = list_first_entry(&tmp, struct r10bio,
2732 list_del(&r10_bio->retry_list);
2733 if (mddev->degraded)
2734 set_bit(R10BIO_Degraded, &r10_bio->state);
2736 if (test_bit(R10BIO_WriteError,
2738 close_write(r10_bio);
2739 raid_end_bio_io(r10_bio);
2743 blk_start_plug(&plug);
2746 flush_pending_writes(conf);
2748 spin_lock_irqsave(&conf->device_lock, flags);
2749 if (list_empty(head)) {
2750 spin_unlock_irqrestore(&conf->device_lock, flags);
2753 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2754 list_del(head->prev);
2756 spin_unlock_irqrestore(&conf->device_lock, flags);
2758 mddev = r10_bio->mddev;
2759 conf = mddev->private;
2760 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2761 test_bit(R10BIO_WriteError, &r10_bio->state))
2762 handle_write_completed(conf, r10_bio);
2763 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2764 reshape_request_write(mddev, r10_bio);
2765 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2766 sync_request_write(mddev, r10_bio);
2767 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2768 recovery_request_write(mddev, r10_bio);
2769 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2770 handle_read_error(mddev, r10_bio);
2775 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2776 md_check_recovery(mddev);
2778 blk_finish_plug(&plug);
2781 static int init_resync(struct r10conf *conf)
2786 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2787 BUG_ON(conf->r10buf_pool);
2788 conf->have_replacement = 0;
2789 for (i = 0; i < conf->geo.raid_disks; i++)
2790 if (conf->mirrors[i].replacement)
2791 conf->have_replacement = 1;
2792 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2793 if (!conf->r10buf_pool)
2795 conf->next_resync = 0;
2800 * perform a "sync" on one "block"
2802 * We need to make sure that no normal I/O request - particularly write
2803 * requests - conflict with active sync requests.
2805 * This is achieved by tracking pending requests and a 'barrier' concept
2806 * that can be installed to exclude normal IO requests.
2808 * Resync and recovery are handled very differently.
2809 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2811 * For resync, we iterate over virtual addresses, read all copies,
2812 * and update if there are differences. If only one copy is live,
2814 * For recovery, we iterate over physical addresses, read a good
2815 * value for each non-in_sync drive, and over-write.
2817 * So, for recovery we may have several outstanding complex requests for a
2818 * given address, one for each out-of-sync device. We model this by allocating
2819 * a number of r10_bio structures, one for each out-of-sync device.
2820 * As we setup these structures, we collect all bio's together into a list
2821 * which we then process collectively to add pages, and then process again
2822 * to pass to generic_make_request.
2824 * The r10_bio structures are linked using a borrowed master_bio pointer.
2825 * This link is counted in ->remaining. When the r10_bio that points to NULL
2826 * has its remaining count decremented to 0, the whole complex operation
2831 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2834 struct r10conf *conf = mddev->private;
2835 struct r10bio *r10_bio;
2836 struct bio *biolist = NULL, *bio;
2837 sector_t max_sector, nr_sectors;
2840 sector_t sync_blocks;
2841 sector_t sectors_skipped = 0;
2842 int chunks_skipped = 0;
2843 sector_t chunk_mask = conf->geo.chunk_mask;
2846 if (!conf->r10buf_pool)
2847 if (init_resync(conf))
2851 * Allow skipping a full rebuild for incremental assembly
2852 * of a clean array, like RAID1 does.
2854 if (mddev->bitmap == NULL &&
2855 mddev->recovery_cp == MaxSector &&
2856 mddev->reshape_position == MaxSector &&
2857 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2858 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2859 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2860 conf->fullsync == 0) {
2862 return mddev->dev_sectors - sector_nr;
2866 max_sector = mddev->dev_sectors;
2867 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2868 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2869 max_sector = mddev->resync_max_sectors;
2870 if (sector_nr >= max_sector) {
2871 /* If we aborted, we need to abort the
2872 * sync on the 'current' bitmap chucks (there can
2873 * be several when recovering multiple devices).
2874 * as we may have started syncing it but not finished.
2875 * We can find the current address in
2876 * mddev->curr_resync, but for recovery,
2877 * we need to convert that to several
2878 * virtual addresses.
2880 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2886 if (mddev->curr_resync < max_sector) { /* aborted */
2887 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2888 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2890 else for (i = 0; i < conf->geo.raid_disks; i++) {
2892 raid10_find_virt(conf, mddev->curr_resync, i);
2893 bitmap_end_sync(mddev->bitmap, sect,
2897 /* completed sync */
2898 if ((!mddev->bitmap || conf->fullsync)
2899 && conf->have_replacement
2900 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2901 /* Completed a full sync so the replacements
2902 * are now fully recovered.
2905 for (i = 0; i < conf->geo.raid_disks; i++) {
2906 struct md_rdev *rdev =
2907 rcu_dereference(conf->mirrors[i].replacement);
2909 rdev->recovery_offset = MaxSector;
2915 bitmap_close_sync(mddev->bitmap);
2918 return sectors_skipped;
2921 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2922 return reshape_request(mddev, sector_nr, skipped);
2924 if (chunks_skipped >= conf->geo.raid_disks) {
2925 /* if there has been nothing to do on any drive,
2926 * then there is nothing to do at all..
2929 return (max_sector - sector_nr) + sectors_skipped;
2932 if (max_sector > mddev->resync_max)
2933 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2935 /* make sure whole request will fit in a chunk - if chunks
2938 if (conf->geo.near_copies < conf->geo.raid_disks &&
2939 max_sector > (sector_nr | chunk_mask))
2940 max_sector = (sector_nr | chunk_mask) + 1;
2943 * If there is non-resync activity waiting for a turn, then let it
2944 * though before starting on this new sync request.
2946 if (conf->nr_waiting)
2947 schedule_timeout_uninterruptible(1);
2949 /* Again, very different code for resync and recovery.
2950 * Both must result in an r10bio with a list of bios that
2951 * have bi_end_io, bi_sector, bi_disk set,
2952 * and bi_private set to the r10bio.
2953 * For recovery, we may actually create several r10bios
2954 * with 2 bios in each, that correspond to the bios in the main one.
2955 * In this case, the subordinate r10bios link back through a
2956 * borrowed master_bio pointer, and the counter in the master
2957 * includes a ref from each subordinate.
2959 /* First, we decide what to do and set ->bi_end_io
2960 * To end_sync_read if we want to read, and
2961 * end_sync_write if we will want to write.
2964 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2965 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2966 /* recovery... the complicated one */
2970 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2976 struct raid10_info *mirror = &conf->mirrors[i];
2977 struct md_rdev *mrdev, *mreplace;
2980 mrdev = rcu_dereference(mirror->rdev);
2981 mreplace = rcu_dereference(mirror->replacement);
2983 if ((mrdev == NULL ||
2984 test_bit(Faulty, &mrdev->flags) ||
2985 test_bit(In_sync, &mrdev->flags)) &&
2986 (mreplace == NULL ||
2987 test_bit(Faulty, &mreplace->flags))) {
2993 /* want to reconstruct this device */
2995 sect = raid10_find_virt(conf, sector_nr, i);
2996 if (sect >= mddev->resync_max_sectors) {
2997 /* last stripe is not complete - don't
2998 * try to recover this sector.
3003 if (mreplace && test_bit(Faulty, &mreplace->flags))
3005 /* Unless we are doing a full sync, or a replacement
3006 * we only need to recover the block if it is set in
3009 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3011 if (sync_blocks < max_sync)
3012 max_sync = sync_blocks;
3016 /* yep, skip the sync_blocks here, but don't assume
3017 * that there will never be anything to do here
3019 chunks_skipped = -1;
3023 atomic_inc(&mrdev->nr_pending);
3025 atomic_inc(&mreplace->nr_pending);
3028 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3030 raise_barrier(conf, rb2 != NULL);
3031 atomic_set(&r10_bio->remaining, 0);
3033 r10_bio->master_bio = (struct bio*)rb2;
3035 atomic_inc(&rb2->remaining);
3036 r10_bio->mddev = mddev;
3037 set_bit(R10BIO_IsRecover, &r10_bio->state);
3038 r10_bio->sector = sect;
3040 raid10_find_phys(conf, r10_bio);
3042 /* Need to check if the array will still be
3046 for (j = 0; j < conf->geo.raid_disks; j++) {
3047 struct md_rdev *rdev = rcu_dereference(
3048 conf->mirrors[j].rdev);
3049 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3055 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3056 &sync_blocks, still_degraded);
3059 for (j=0; j<conf->copies;j++) {
3061 int d = r10_bio->devs[j].devnum;
3062 sector_t from_addr, to_addr;
3063 struct md_rdev *rdev =
3064 rcu_dereference(conf->mirrors[d].rdev);
3065 sector_t sector, first_bad;
3068 !test_bit(In_sync, &rdev->flags))
3070 /* This is where we read from */
3072 sector = r10_bio->devs[j].addr;
3074 if (is_badblock(rdev, sector, max_sync,
3075 &first_bad, &bad_sectors)) {
3076 if (first_bad > sector)
3077 max_sync = first_bad - sector;
3079 bad_sectors -= (sector
3081 if (max_sync > bad_sectors)
3082 max_sync = bad_sectors;
3086 bio = r10_bio->devs[0].bio;
3087 bio->bi_next = biolist;
3089 bio->bi_end_io = end_sync_read;
3090 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3091 if (test_bit(FailFast, &rdev->flags))
3092 bio->bi_opf |= MD_FAILFAST;
3093 from_addr = r10_bio->devs[j].addr;
3094 bio->bi_iter.bi_sector = from_addr +
3096 bio_set_dev(bio, rdev->bdev);
3097 atomic_inc(&rdev->nr_pending);
3098 /* and we write to 'i' (if not in_sync) */
3100 for (k=0; k<conf->copies; k++)
3101 if (r10_bio->devs[k].devnum == i)
3103 BUG_ON(k == conf->copies);
3104 to_addr = r10_bio->devs[k].addr;
3105 r10_bio->devs[0].devnum = d;
3106 r10_bio->devs[0].addr = from_addr;
3107 r10_bio->devs[1].devnum = i;
3108 r10_bio->devs[1].addr = to_addr;
3110 if (!test_bit(In_sync, &mrdev->flags)) {
3111 bio = r10_bio->devs[1].bio;
3112 bio->bi_next = biolist;
3114 bio->bi_end_io = end_sync_write;
3115 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3116 bio->bi_iter.bi_sector = to_addr
3117 + mrdev->data_offset;
3118 bio_set_dev(bio, mrdev->bdev);
3119 atomic_inc(&r10_bio->remaining);
3121 r10_bio->devs[1].bio->bi_end_io = NULL;
3123 /* and maybe write to replacement */
3124 bio = r10_bio->devs[1].repl_bio;
3126 bio->bi_end_io = NULL;
3127 /* Note: if mreplace != NULL, then bio
3128 * cannot be NULL as r10buf_pool_alloc will
3129 * have allocated it.
3130 * So the second test here is pointless.
3131 * But it keeps semantic-checkers happy, and
3132 * this comment keeps human reviewers
3135 if (mreplace == NULL || bio == NULL ||
3136 test_bit(Faulty, &mreplace->flags))
3138 bio->bi_next = biolist;
3140 bio->bi_end_io = end_sync_write;
3141 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3142 bio->bi_iter.bi_sector = to_addr +
3143 mreplace->data_offset;
3144 bio_set_dev(bio, mreplace->bdev);
3145 atomic_inc(&r10_bio->remaining);
3149 if (j == conf->copies) {
3150 /* Cannot recover, so abort the recovery or
3151 * record a bad block */
3153 /* problem is that there are bad blocks
3154 * on other device(s)
3157 for (k = 0; k < conf->copies; k++)
3158 if (r10_bio->devs[k].devnum == i)
3160 if (!test_bit(In_sync,
3162 && !rdev_set_badblocks(
3164 r10_bio->devs[k].addr,
3168 !rdev_set_badblocks(
3170 r10_bio->devs[k].addr,
3175 if (!test_and_set_bit(MD_RECOVERY_INTR,
3177 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3179 mirror->recovery_disabled
3180 = mddev->recovery_disabled;
3184 atomic_dec(&rb2->remaining);
3186 rdev_dec_pending(mrdev, mddev);
3188 rdev_dec_pending(mreplace, mddev);
3191 rdev_dec_pending(mrdev, mddev);
3193 rdev_dec_pending(mreplace, mddev);
3194 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3195 /* Only want this if there is elsewhere to
3196 * read from. 'j' is currently the first
3200 for (; j < conf->copies; j++) {
3201 int d = r10_bio->devs[j].devnum;
3202 if (conf->mirrors[d].rdev &&
3204 &conf->mirrors[d].rdev->flags))
3208 r10_bio->devs[0].bio->bi_opf
3212 if (biolist == NULL) {
3214 struct r10bio *rb2 = r10_bio;
3215 r10_bio = (struct r10bio*) rb2->master_bio;
3216 rb2->master_bio = NULL;
3222 /* resync. Schedule a read for every block at this virt offset */
3225 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3227 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3228 &sync_blocks, mddev->degraded) &&
3229 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3230 &mddev->recovery)) {
3231 /* We can skip this block */
3233 return sync_blocks + sectors_skipped;
3235 if (sync_blocks < max_sync)
3236 max_sync = sync_blocks;
3237 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3240 r10_bio->mddev = mddev;
3241 atomic_set(&r10_bio->remaining, 0);
3242 raise_barrier(conf, 0);
3243 conf->next_resync = sector_nr;
3245 r10_bio->master_bio = NULL;
3246 r10_bio->sector = sector_nr;
3247 set_bit(R10BIO_IsSync, &r10_bio->state);
3248 raid10_find_phys(conf, r10_bio);
3249 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3251 for (i = 0; i < conf->copies; i++) {
3252 int d = r10_bio->devs[i].devnum;
3253 sector_t first_bad, sector;
3255 struct md_rdev *rdev;
3257 if (r10_bio->devs[i].repl_bio)
3258 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3260 bio = r10_bio->devs[i].bio;
3261 bio->bi_status = BLK_STS_IOERR;
3263 rdev = rcu_dereference(conf->mirrors[d].rdev);
3264 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3268 sector = r10_bio->devs[i].addr;
3269 if (is_badblock(rdev, sector, max_sync,
3270 &first_bad, &bad_sectors)) {
3271 if (first_bad > sector)
3272 max_sync = first_bad - sector;
3274 bad_sectors -= (sector - first_bad);
3275 if (max_sync > bad_sectors)
3276 max_sync = bad_sectors;
3281 atomic_inc(&rdev->nr_pending);
3282 atomic_inc(&r10_bio->remaining);
3283 bio->bi_next = biolist;
3285 bio->bi_end_io = end_sync_read;
3286 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3287 if (test_bit(FailFast, &rdev->flags))
3288 bio->bi_opf |= MD_FAILFAST;
3289 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3290 bio_set_dev(bio, rdev->bdev);
3293 rdev = rcu_dereference(conf->mirrors[d].replacement);
3294 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3298 atomic_inc(&rdev->nr_pending);
3300 /* Need to set up for writing to the replacement */
3301 bio = r10_bio->devs[i].repl_bio;
3302 bio->bi_status = BLK_STS_IOERR;
3304 sector = r10_bio->devs[i].addr;
3305 bio->bi_next = biolist;
3307 bio->bi_end_io = end_sync_write;
3308 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3309 if (test_bit(FailFast, &rdev->flags))
3310 bio->bi_opf |= MD_FAILFAST;
3311 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3312 bio_set_dev(bio, rdev->bdev);
3318 for (i=0; i<conf->copies; i++) {
3319 int d = r10_bio->devs[i].devnum;
3320 if (r10_bio->devs[i].bio->bi_end_io)
3321 rdev_dec_pending(conf->mirrors[d].rdev,
3323 if (r10_bio->devs[i].repl_bio &&
3324 r10_bio->devs[i].repl_bio->bi_end_io)
3326 conf->mirrors[d].replacement,
3336 if (sector_nr + max_sync < max_sector)
3337 max_sector = sector_nr + max_sync;
3340 int len = PAGE_SIZE;
3341 if (sector_nr + (len>>9) > max_sector)
3342 len = (max_sector - sector_nr) << 9;
3345 for (bio= biolist ; bio ; bio=bio->bi_next) {
3346 struct resync_pages *rp = get_resync_pages(bio);
3347 page = resync_fetch_page(rp, page_idx);
3349 * won't fail because the vec table is big enough
3350 * to hold all these pages
3352 bio_add_page(bio, page, len, 0);
3354 nr_sectors += len>>9;
3355 sector_nr += len>>9;
3356 } while (++page_idx < RESYNC_PAGES);
3357 r10_bio->sectors = nr_sectors;
3361 biolist = biolist->bi_next;
3363 bio->bi_next = NULL;
3364 r10_bio = get_resync_r10bio(bio);
3365 r10_bio->sectors = nr_sectors;
3367 if (bio->bi_end_io == end_sync_read) {
3368 md_sync_acct_bio(bio, nr_sectors);
3370 generic_make_request(bio);
3374 if (sectors_skipped)
3375 /* pretend they weren't skipped, it makes
3376 * no important difference in this case
3378 md_done_sync(mddev, sectors_skipped, 1);
3380 return sectors_skipped + nr_sectors;
3382 /* There is nowhere to write, so all non-sync
3383 * drives must be failed or in resync, all drives
3384 * have a bad block, so try the next chunk...
3386 if (sector_nr + max_sync < max_sector)
3387 max_sector = sector_nr + max_sync;
3389 sectors_skipped += (max_sector - sector_nr);
3391 sector_nr = max_sector;
3396 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3399 struct r10conf *conf = mddev->private;
3402 raid_disks = min(conf->geo.raid_disks,
3403 conf->prev.raid_disks);
3405 sectors = conf->dev_sectors;
3407 size = sectors >> conf->geo.chunk_shift;
3408 sector_div(size, conf->geo.far_copies);
3409 size = size * raid_disks;
3410 sector_div(size, conf->geo.near_copies);
3412 return size << conf->geo.chunk_shift;
3415 static void calc_sectors(struct r10conf *conf, sector_t size)
3417 /* Calculate the number of sectors-per-device that will
3418 * actually be used, and set conf->dev_sectors and
3422 size = size >> conf->geo.chunk_shift;
3423 sector_div(size, conf->geo.far_copies);
3424 size = size * conf->geo.raid_disks;
3425 sector_div(size, conf->geo.near_copies);
3426 /* 'size' is now the number of chunks in the array */
3427 /* calculate "used chunks per device" */
3428 size = size * conf->copies;
3430 /* We need to round up when dividing by raid_disks to
3431 * get the stride size.
3433 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3435 conf->dev_sectors = size << conf->geo.chunk_shift;
3437 if (conf->geo.far_offset)
3438 conf->geo.stride = 1 << conf->geo.chunk_shift;
3440 sector_div(size, conf->geo.far_copies);
3441 conf->geo.stride = size << conf->geo.chunk_shift;
3445 enum geo_type {geo_new, geo_old, geo_start};
3446 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3449 int layout, chunk, disks;
3452 layout = mddev->layout;
3453 chunk = mddev->chunk_sectors;
3454 disks = mddev->raid_disks - mddev->delta_disks;
3457 layout = mddev->new_layout;
3458 chunk = mddev->new_chunk_sectors;
3459 disks = mddev->raid_disks;
3461 default: /* avoid 'may be unused' warnings */
3462 case geo_start: /* new when starting reshape - raid_disks not
3464 layout = mddev->new_layout;
3465 chunk = mddev->new_chunk_sectors;
3466 disks = mddev->raid_disks + mddev->delta_disks;
3471 if (chunk < (PAGE_SIZE >> 9) ||
3472 !is_power_of_2(chunk))
3475 fc = (layout >> 8) & 255;
3476 fo = layout & (1<<16);
3477 geo->raid_disks = disks;
3478 geo->near_copies = nc;
3479 geo->far_copies = fc;
3480 geo->far_offset = fo;
3481 switch (layout >> 17) {
3482 case 0: /* original layout. simple but not always optimal */
3483 geo->far_set_size = disks;
3485 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3486 * actually using this, but leave code here just in case.*/
3487 geo->far_set_size = disks/fc;
3488 WARN(geo->far_set_size < fc,
3489 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3491 case 2: /* "improved" layout fixed to match documentation */
3492 geo->far_set_size = fc * nc;
3494 default: /* Not a valid layout */
3497 geo->chunk_mask = chunk - 1;
3498 geo->chunk_shift = ffz(~chunk);
3502 static struct r10conf *setup_conf(struct mddev *mddev)
3504 struct r10conf *conf = NULL;
3509 copies = setup_geo(&geo, mddev, geo_new);
3512 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3513 mdname(mddev), PAGE_SIZE);
3517 if (copies < 2 || copies > mddev->raid_disks) {
3518 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3519 mdname(mddev), mddev->new_layout);
3524 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3528 /* FIXME calc properly */
3529 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3530 max(0,-mddev->delta_disks)),
3535 conf->tmppage = alloc_page(GFP_KERNEL);
3540 conf->copies = copies;
3541 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3542 r10bio_pool_free, conf);
3543 if (!conf->r10bio_pool)
3546 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
3547 if (!conf->bio_split)
3550 calc_sectors(conf, mddev->dev_sectors);
3551 if (mddev->reshape_position == MaxSector) {
3552 conf->prev = conf->geo;
3553 conf->reshape_progress = MaxSector;
3555 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3559 conf->reshape_progress = mddev->reshape_position;
3560 if (conf->prev.far_offset)
3561 conf->prev.stride = 1 << conf->prev.chunk_shift;
3563 /* far_copies must be 1 */
3564 conf->prev.stride = conf->dev_sectors;
3566 conf->reshape_safe = conf->reshape_progress;
3567 spin_lock_init(&conf->device_lock);
3568 INIT_LIST_HEAD(&conf->retry_list);
3569 INIT_LIST_HEAD(&conf->bio_end_io_list);
3571 spin_lock_init(&conf->resync_lock);
3572 init_waitqueue_head(&conf->wait_barrier);
3573 atomic_set(&conf->nr_pending, 0);
3575 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3579 conf->mddev = mddev;
3584 mempool_destroy(conf->r10bio_pool);
3585 kfree(conf->mirrors);
3586 safe_put_page(conf->tmppage);
3587 if (conf->bio_split)
3588 bioset_free(conf->bio_split);
3591 return ERR_PTR(err);
3594 static int raid10_run(struct mddev *mddev)
3596 struct r10conf *conf;
3597 int i, disk_idx, chunk_size;
3598 struct raid10_info *disk;
3599 struct md_rdev *rdev;
3601 sector_t min_offset_diff = 0;
3603 bool discard_supported = false;
3605 if (mddev_init_writes_pending(mddev) < 0)
3608 if (mddev->private == NULL) {
3609 conf = setup_conf(mddev);
3611 return PTR_ERR(conf);
3612 mddev->private = conf;
3614 conf = mddev->private;
3618 mddev->thread = conf->thread;
3619 conf->thread = NULL;
3621 chunk_size = mddev->chunk_sectors << 9;
3623 blk_queue_max_discard_sectors(mddev->queue,
3624 mddev->chunk_sectors);
3625 blk_queue_max_write_same_sectors(mddev->queue, 0);
3626 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3627 blk_queue_io_min(mddev->queue, chunk_size);
3628 if (conf->geo.raid_disks % conf->geo.near_copies)
3629 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3631 blk_queue_io_opt(mddev->queue, chunk_size *
3632 (conf->geo.raid_disks / conf->geo.near_copies));
3635 rdev_for_each(rdev, mddev) {
3638 disk_idx = rdev->raid_disk;
3641 if (disk_idx >= conf->geo.raid_disks &&
3642 disk_idx >= conf->prev.raid_disks)
3644 disk = conf->mirrors + disk_idx;
3646 if (test_bit(Replacement, &rdev->flags)) {
3647 if (disk->replacement)
3649 disk->replacement = rdev;
3655 diff = (rdev->new_data_offset - rdev->data_offset);
3656 if (!mddev->reshape_backwards)
3660 if (first || diff < min_offset_diff)
3661 min_offset_diff = diff;
3664 disk_stack_limits(mddev->gendisk, rdev->bdev,
3665 rdev->data_offset << 9);
3667 disk->head_position = 0;
3669 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3670 discard_supported = true;
3675 if (discard_supported)
3676 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3679 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3682 /* need to check that every block has at least one working mirror */
3683 if (!enough(conf, -1)) {
3684 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3689 if (conf->reshape_progress != MaxSector) {
3690 /* must ensure that shape change is supported */
3691 if (conf->geo.far_copies != 1 &&
3692 conf->geo.far_offset == 0)
3694 if (conf->prev.far_copies != 1 &&
3695 conf->prev.far_offset == 0)
3699 mddev->degraded = 0;
3701 i < conf->geo.raid_disks
3702 || i < conf->prev.raid_disks;
3705 disk = conf->mirrors + i;
3707 if (!disk->rdev && disk->replacement) {
3708 /* The replacement is all we have - use it */
3709 disk->rdev = disk->replacement;
3710 disk->replacement = NULL;
3711 clear_bit(Replacement, &disk->rdev->flags);
3715 !test_bit(In_sync, &disk->rdev->flags)) {
3716 disk->head_position = 0;
3719 disk->rdev->saved_raid_disk < 0)
3722 disk->recovery_disabled = mddev->recovery_disabled - 1;
3725 if (mddev->recovery_cp != MaxSector)
3726 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3728 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3729 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3730 conf->geo.raid_disks);
3732 * Ok, everything is just fine now
3734 mddev->dev_sectors = conf->dev_sectors;
3735 size = raid10_size(mddev, 0, 0);
3736 md_set_array_sectors(mddev, size);
3737 mddev->resync_max_sectors = size;
3738 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3741 int stripe = conf->geo.raid_disks *
3742 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3744 /* Calculate max read-ahead size.
3745 * We need to readahead at least twice a whole stripe....
3748 stripe /= conf->geo.near_copies;
3749 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3750 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3753 if (md_integrity_register(mddev))
3756 if (conf->reshape_progress != MaxSector) {
3757 unsigned long before_length, after_length;
3759 before_length = ((1 << conf->prev.chunk_shift) *
3760 conf->prev.far_copies);
3761 after_length = ((1 << conf->geo.chunk_shift) *
3762 conf->geo.far_copies);
3764 if (max(before_length, after_length) > min_offset_diff) {
3765 /* This cannot work */
3766 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3769 conf->offset_diff = min_offset_diff;
3771 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3772 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3773 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3774 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3775 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3782 md_unregister_thread(&mddev->thread);
3783 mempool_destroy(conf->r10bio_pool);
3784 safe_put_page(conf->tmppage);
3785 kfree(conf->mirrors);
3787 mddev->private = NULL;
3792 static void raid10_free(struct mddev *mddev, void *priv)
3794 struct r10conf *conf = priv;
3796 mempool_destroy(conf->r10bio_pool);
3797 safe_put_page(conf->tmppage);
3798 kfree(conf->mirrors);
3799 kfree(conf->mirrors_old);
3800 kfree(conf->mirrors_new);
3801 if (conf->bio_split)
3802 bioset_free(conf->bio_split);
3806 static void raid10_quiesce(struct mddev *mddev, int state)
3808 struct r10conf *conf = mddev->private;
3812 raise_barrier(conf, 0);
3815 lower_barrier(conf);
3820 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3822 /* Resize of 'far' arrays is not supported.
3823 * For 'near' and 'offset' arrays we can set the
3824 * number of sectors used to be an appropriate multiple
3825 * of the chunk size.
3826 * For 'offset', this is far_copies*chunksize.
3827 * For 'near' the multiplier is the LCM of
3828 * near_copies and raid_disks.
3829 * So if far_copies > 1 && !far_offset, fail.
3830 * Else find LCM(raid_disks, near_copy)*far_copies and
3831 * multiply by chunk_size. Then round to this number.
3832 * This is mostly done by raid10_size()
3834 struct r10conf *conf = mddev->private;
3835 sector_t oldsize, size;
3837 if (mddev->reshape_position != MaxSector)
3840 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3843 oldsize = raid10_size(mddev, 0, 0);
3844 size = raid10_size(mddev, sectors, 0);
3845 if (mddev->external_size &&
3846 mddev->array_sectors > size)
3848 if (mddev->bitmap) {
3849 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3853 md_set_array_sectors(mddev, size);
3854 if (sectors > mddev->dev_sectors &&
3855 mddev->recovery_cp > oldsize) {
3856 mddev->recovery_cp = oldsize;
3857 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3859 calc_sectors(conf, sectors);
3860 mddev->dev_sectors = conf->dev_sectors;
3861 mddev->resync_max_sectors = size;
3865 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3867 struct md_rdev *rdev;
3868 struct r10conf *conf;
3870 if (mddev->degraded > 0) {
3871 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3873 return ERR_PTR(-EINVAL);
3875 sector_div(size, devs);
3877 /* Set new parameters */
3878 mddev->new_level = 10;
3879 /* new layout: far_copies = 1, near_copies = 2 */
3880 mddev->new_layout = (1<<8) + 2;
3881 mddev->new_chunk_sectors = mddev->chunk_sectors;
3882 mddev->delta_disks = mddev->raid_disks;
3883 mddev->raid_disks *= 2;
3884 /* make sure it will be not marked as dirty */
3885 mddev->recovery_cp = MaxSector;
3886 mddev->dev_sectors = size;
3888 conf = setup_conf(mddev);
3889 if (!IS_ERR(conf)) {
3890 rdev_for_each(rdev, mddev)
3891 if (rdev->raid_disk >= 0) {
3892 rdev->new_raid_disk = rdev->raid_disk * 2;
3893 rdev->sectors = size;
3901 static void *raid10_takeover(struct mddev *mddev)
3903 struct r0conf *raid0_conf;
3905 /* raid10 can take over:
3906 * raid0 - providing it has only two drives
3908 if (mddev->level == 0) {
3909 /* for raid0 takeover only one zone is supported */
3910 raid0_conf = mddev->private;
3911 if (raid0_conf->nr_strip_zones > 1) {
3912 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
3914 return ERR_PTR(-EINVAL);
3916 return raid10_takeover_raid0(mddev,
3917 raid0_conf->strip_zone->zone_end,
3918 raid0_conf->strip_zone->nb_dev);
3920 return ERR_PTR(-EINVAL);
3923 static int raid10_check_reshape(struct mddev *mddev)
3925 /* Called when there is a request to change
3926 * - layout (to ->new_layout)
3927 * - chunk size (to ->new_chunk_sectors)
3928 * - raid_disks (by delta_disks)
3929 * or when trying to restart a reshape that was ongoing.
3931 * We need to validate the request and possibly allocate
3932 * space if that might be an issue later.
3934 * Currently we reject any reshape of a 'far' mode array,
3935 * allow chunk size to change if new is generally acceptable,
3936 * allow raid_disks to increase, and allow
3937 * a switch between 'near' mode and 'offset' mode.
3939 struct r10conf *conf = mddev->private;
3942 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3945 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3946 /* mustn't change number of copies */
3948 if (geo.far_copies > 1 && !geo.far_offset)
3949 /* Cannot switch to 'far' mode */
3952 if (mddev->array_sectors & geo.chunk_mask)
3953 /* not factor of array size */
3956 if (!enough(conf, -1))
3959 kfree(conf->mirrors_new);
3960 conf->mirrors_new = NULL;
3961 if (mddev->delta_disks > 0) {
3962 /* allocate new 'mirrors' list */
3963 conf->mirrors_new = kzalloc(
3964 sizeof(struct raid10_info)
3965 *(mddev->raid_disks +
3966 mddev->delta_disks),
3968 if (!conf->mirrors_new)
3975 * Need to check if array has failed when deciding whether to:
3977 * - remove non-faulty devices
3980 * This determination is simple when no reshape is happening.
3981 * However if there is a reshape, we need to carefully check
3982 * both the before and after sections.
3983 * This is because some failed devices may only affect one
3984 * of the two sections, and some non-in_sync devices may
3985 * be insync in the section most affected by failed devices.
3987 static int calc_degraded(struct r10conf *conf)
3989 int degraded, degraded2;
3994 /* 'prev' section first */
3995 for (i = 0; i < conf->prev.raid_disks; i++) {
3996 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3997 if (!rdev || test_bit(Faulty, &rdev->flags))
3999 else if (!test_bit(In_sync, &rdev->flags))
4000 /* When we can reduce the number of devices in
4001 * an array, this might not contribute to
4002 * 'degraded'. It does now.
4007 if (conf->geo.raid_disks == conf->prev.raid_disks)
4011 for (i = 0; i < conf->geo.raid_disks; i++) {
4012 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4013 if (!rdev || test_bit(Faulty, &rdev->flags))
4015 else if (!test_bit(In_sync, &rdev->flags)) {
4016 /* If reshape is increasing the number of devices,
4017 * this section has already been recovered, so
4018 * it doesn't contribute to degraded.
4021 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4026 if (degraded2 > degraded)
4031 static int raid10_start_reshape(struct mddev *mddev)
4033 /* A 'reshape' has been requested. This commits
4034 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4035 * This also checks if there are enough spares and adds them
4037 * We currently require enough spares to make the final
4038 * array non-degraded. We also require that the difference
4039 * between old and new data_offset - on each device - is
4040 * enough that we never risk over-writing.
4043 unsigned long before_length, after_length;
4044 sector_t min_offset_diff = 0;
4047 struct r10conf *conf = mddev->private;
4048 struct md_rdev *rdev;
4052 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4055 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4058 before_length = ((1 << conf->prev.chunk_shift) *
4059 conf->prev.far_copies);
4060 after_length = ((1 << conf->geo.chunk_shift) *
4061 conf->geo.far_copies);
4063 rdev_for_each(rdev, mddev) {
4064 if (!test_bit(In_sync, &rdev->flags)
4065 && !test_bit(Faulty, &rdev->flags))
4067 if (rdev->raid_disk >= 0) {
4068 long long diff = (rdev->new_data_offset
4069 - rdev->data_offset);
4070 if (!mddev->reshape_backwards)
4074 if (first || diff < min_offset_diff)
4075 min_offset_diff = diff;
4080 if (max(before_length, after_length) > min_offset_diff)
4083 if (spares < mddev->delta_disks)
4086 conf->offset_diff = min_offset_diff;
4087 spin_lock_irq(&conf->device_lock);
4088 if (conf->mirrors_new) {
4089 memcpy(conf->mirrors_new, conf->mirrors,
4090 sizeof(struct raid10_info)*conf->prev.raid_disks);
4092 kfree(conf->mirrors_old);
4093 conf->mirrors_old = conf->mirrors;
4094 conf->mirrors = conf->mirrors_new;
4095 conf->mirrors_new = NULL;
4097 setup_geo(&conf->geo, mddev, geo_start);
4099 if (mddev->reshape_backwards) {
4100 sector_t size = raid10_size(mddev, 0, 0);
4101 if (size < mddev->array_sectors) {
4102 spin_unlock_irq(&conf->device_lock);
4103 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4107 mddev->resync_max_sectors = size;
4108 conf->reshape_progress = size;
4110 conf->reshape_progress = 0;
4111 conf->reshape_safe = conf->reshape_progress;
4112 spin_unlock_irq(&conf->device_lock);
4114 if (mddev->delta_disks && mddev->bitmap) {
4115 ret = bitmap_resize(mddev->bitmap,
4116 raid10_size(mddev, 0,
4117 conf->geo.raid_disks),
4122 if (mddev->delta_disks > 0) {
4123 rdev_for_each(rdev, mddev)
4124 if (rdev->raid_disk < 0 &&
4125 !test_bit(Faulty, &rdev->flags)) {
4126 if (raid10_add_disk(mddev, rdev) == 0) {
4127 if (rdev->raid_disk >=
4128 conf->prev.raid_disks)
4129 set_bit(In_sync, &rdev->flags);
4131 rdev->recovery_offset = 0;
4133 if (sysfs_link_rdev(mddev, rdev))
4134 /* Failure here is OK */;
4136 } else if (rdev->raid_disk >= conf->prev.raid_disks
4137 && !test_bit(Faulty, &rdev->flags)) {
4138 /* This is a spare that was manually added */
4139 set_bit(In_sync, &rdev->flags);
4142 /* When a reshape changes the number of devices,
4143 * ->degraded is measured against the larger of the
4144 * pre and post numbers.
4146 spin_lock_irq(&conf->device_lock);
4147 mddev->degraded = calc_degraded(conf);
4148 spin_unlock_irq(&conf->device_lock);
4149 mddev->raid_disks = conf->geo.raid_disks;
4150 mddev->reshape_position = conf->reshape_progress;
4151 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4153 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4154 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4155 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4156 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4157 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4159 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4161 if (!mddev->sync_thread) {
4165 conf->reshape_checkpoint = jiffies;
4166 md_wakeup_thread(mddev->sync_thread);
4167 md_new_event(mddev);
4171 mddev->recovery = 0;
4172 spin_lock_irq(&conf->device_lock);
4173 conf->geo = conf->prev;
4174 mddev->raid_disks = conf->geo.raid_disks;
4175 rdev_for_each(rdev, mddev)
4176 rdev->new_data_offset = rdev->data_offset;
4178 conf->reshape_progress = MaxSector;
4179 conf->reshape_safe = MaxSector;
4180 mddev->reshape_position = MaxSector;
4181 spin_unlock_irq(&conf->device_lock);
4185 /* Calculate the last device-address that could contain
4186 * any block from the chunk that includes the array-address 's'
4187 * and report the next address.
4188 * i.e. the address returned will be chunk-aligned and after
4189 * any data that is in the chunk containing 's'.
4191 static sector_t last_dev_address(sector_t s, struct geom *geo)
4193 s = (s | geo->chunk_mask) + 1;
4194 s >>= geo->chunk_shift;
4195 s *= geo->near_copies;
4196 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4197 s *= geo->far_copies;
4198 s <<= geo->chunk_shift;
4202 /* Calculate the first device-address that could contain
4203 * any block from the chunk that includes the array-address 's'.
4204 * This too will be the start of a chunk
4206 static sector_t first_dev_address(sector_t s, struct geom *geo)
4208 s >>= geo->chunk_shift;
4209 s *= geo->near_copies;
4210 sector_div(s, geo->raid_disks);
4211 s *= geo->far_copies;
4212 s <<= geo->chunk_shift;
4216 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4219 /* We simply copy at most one chunk (smallest of old and new)
4220 * at a time, possibly less if that exceeds RESYNC_PAGES,
4221 * or we hit a bad block or something.
4222 * This might mean we pause for normal IO in the middle of
4223 * a chunk, but that is not a problem as mddev->reshape_position
4224 * can record any location.
4226 * If we will want to write to a location that isn't
4227 * yet recorded as 'safe' (i.e. in metadata on disk) then
4228 * we need to flush all reshape requests and update the metadata.
4230 * When reshaping forwards (e.g. to more devices), we interpret
4231 * 'safe' as the earliest block which might not have been copied
4232 * down yet. We divide this by previous stripe size and multiply
4233 * by previous stripe length to get lowest device offset that we
4234 * cannot write to yet.
4235 * We interpret 'sector_nr' as an address that we want to write to.
4236 * From this we use last_device_address() to find where we might
4237 * write to, and first_device_address on the 'safe' position.
4238 * If this 'next' write position is after the 'safe' position,
4239 * we must update the metadata to increase the 'safe' position.
4241 * When reshaping backwards, we round in the opposite direction
4242 * and perform the reverse test: next write position must not be
4243 * less than current safe position.
4245 * In all this the minimum difference in data offsets
4246 * (conf->offset_diff - always positive) allows a bit of slack,
4247 * so next can be after 'safe', but not by more than offset_diff
4249 * We need to prepare all the bios here before we start any IO
4250 * to ensure the size we choose is acceptable to all devices.
4251 * The means one for each copy for write-out and an extra one for
4253 * We store the read-in bio in ->master_bio and the others in
4254 * ->devs[x].bio and ->devs[x].repl_bio.
4256 struct r10conf *conf = mddev->private;
4257 struct r10bio *r10_bio;
4258 sector_t next, safe, last;
4262 struct md_rdev *rdev;
4265 struct bio *bio, *read_bio;
4266 int sectors_done = 0;
4267 struct page **pages;
4269 if (sector_nr == 0) {
4270 /* If restarting in the middle, skip the initial sectors */
4271 if (mddev->reshape_backwards &&
4272 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4273 sector_nr = (raid10_size(mddev, 0, 0)
4274 - conf->reshape_progress);
4275 } else if (!mddev->reshape_backwards &&
4276 conf->reshape_progress > 0)
4277 sector_nr = conf->reshape_progress;
4279 mddev->curr_resync_completed = sector_nr;
4280 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4286 /* We don't use sector_nr to track where we are up to
4287 * as that doesn't work well for ->reshape_backwards.
4288 * So just use ->reshape_progress.
4290 if (mddev->reshape_backwards) {
4291 /* 'next' is the earliest device address that we might
4292 * write to for this chunk in the new layout
4294 next = first_dev_address(conf->reshape_progress - 1,
4297 /* 'safe' is the last device address that we might read from
4298 * in the old layout after a restart
4300 safe = last_dev_address(conf->reshape_safe - 1,
4303 if (next + conf->offset_diff < safe)
4306 last = conf->reshape_progress - 1;
4307 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4308 & conf->prev.chunk_mask);
4309 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4310 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4312 /* 'next' is after the last device address that we
4313 * might write to for this chunk in the new layout
4315 next = last_dev_address(conf->reshape_progress, &conf->geo);
4317 /* 'safe' is the earliest device address that we might
4318 * read from in the old layout after a restart
4320 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4322 /* Need to update metadata if 'next' might be beyond 'safe'
4323 * as that would possibly corrupt data
4325 if (next > safe + conf->offset_diff)
4328 sector_nr = conf->reshape_progress;
4329 last = sector_nr | (conf->geo.chunk_mask
4330 & conf->prev.chunk_mask);
4332 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4333 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4337 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4338 /* Need to update reshape_position in metadata */
4340 mddev->reshape_position = conf->reshape_progress;
4341 if (mddev->reshape_backwards)
4342 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4343 - conf->reshape_progress;
4345 mddev->curr_resync_completed = conf->reshape_progress;
4346 conf->reshape_checkpoint = jiffies;
4347 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4348 md_wakeup_thread(mddev->thread);
4349 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4350 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4351 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4352 allow_barrier(conf);
4353 return sectors_done;
4355 conf->reshape_safe = mddev->reshape_position;
4356 allow_barrier(conf);
4360 /* Now schedule reads for blocks from sector_nr to last */
4361 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4363 raise_barrier(conf, sectors_done != 0);
4364 atomic_set(&r10_bio->remaining, 0);
4365 r10_bio->mddev = mddev;
4366 r10_bio->sector = sector_nr;
4367 set_bit(R10BIO_IsReshape, &r10_bio->state);
4368 r10_bio->sectors = last - sector_nr + 1;
4369 rdev = read_balance(conf, r10_bio, &max_sectors);
4370 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4373 /* Cannot read from here, so need to record bad blocks
4374 * on all the target devices.
4377 mempool_free(r10_bio, conf->r10buf_pool);
4378 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4379 return sectors_done;
4382 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4384 bio_set_dev(read_bio, rdev->bdev);
4385 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4386 + rdev->data_offset);
4387 read_bio->bi_private = r10_bio;
4388 read_bio->bi_end_io = end_reshape_read;
4389 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4390 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4391 read_bio->bi_status = 0;
4392 read_bio->bi_vcnt = 0;
4393 read_bio->bi_iter.bi_size = 0;
4394 r10_bio->master_bio = read_bio;
4395 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4397 /* Now find the locations in the new layout */
4398 __raid10_find_phys(&conf->geo, r10_bio);
4401 read_bio->bi_next = NULL;
4404 for (s = 0; s < conf->copies*2; s++) {
4406 int d = r10_bio->devs[s/2].devnum;
4407 struct md_rdev *rdev2;
4409 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4410 b = r10_bio->devs[s/2].repl_bio;
4412 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4413 b = r10_bio->devs[s/2].bio;
4415 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4418 bio_set_dev(b, rdev2->bdev);
4419 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4420 rdev2->new_data_offset;
4421 b->bi_end_io = end_reshape_write;
4422 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4427 /* Now add as many pages as possible to all of these bios. */
4430 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4431 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4432 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4433 int len = (max_sectors - s) << 9;
4434 if (len > PAGE_SIZE)
4436 for (bio = blist; bio ; bio = bio->bi_next) {
4438 * won't fail because the vec table is big enough
4439 * to hold all these pages
4441 bio_add_page(bio, page, len, 0);
4443 sector_nr += len >> 9;
4444 nr_sectors += len >> 9;
4447 r10_bio->sectors = nr_sectors;
4449 /* Now submit the read */
4450 md_sync_acct_bio(read_bio, r10_bio->sectors);
4451 atomic_inc(&r10_bio->remaining);
4452 read_bio->bi_next = NULL;
4453 generic_make_request(read_bio);
4454 sector_nr += nr_sectors;
4455 sectors_done += nr_sectors;
4456 if (sector_nr <= last)
4459 /* Now that we have done the whole section we can
4460 * update reshape_progress
4462 if (mddev->reshape_backwards)
4463 conf->reshape_progress -= sectors_done;
4465 conf->reshape_progress += sectors_done;
4467 return sectors_done;
4470 static void end_reshape_request(struct r10bio *r10_bio);
4471 static int handle_reshape_read_error(struct mddev *mddev,
4472 struct r10bio *r10_bio);
4473 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4475 /* Reshape read completed. Hopefully we have a block
4477 * If we got a read error then we do sync 1-page reads from
4478 * elsewhere until we find the data - or give up.
4480 struct r10conf *conf = mddev->private;
4483 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4484 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4485 /* Reshape has been aborted */
4486 md_done_sync(mddev, r10_bio->sectors, 0);
4490 /* We definitely have the data in the pages, schedule the
4493 atomic_set(&r10_bio->remaining, 1);
4494 for (s = 0; s < conf->copies*2; s++) {
4496 int d = r10_bio->devs[s/2].devnum;
4497 struct md_rdev *rdev;
4500 rdev = rcu_dereference(conf->mirrors[d].replacement);
4501 b = r10_bio->devs[s/2].repl_bio;
4503 rdev = rcu_dereference(conf->mirrors[d].rdev);
4504 b = r10_bio->devs[s/2].bio;
4506 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4510 atomic_inc(&rdev->nr_pending);
4512 md_sync_acct_bio(b, r10_bio->sectors);
4513 atomic_inc(&r10_bio->remaining);
4515 generic_make_request(b);
4517 end_reshape_request(r10_bio);
4520 static void end_reshape(struct r10conf *conf)
4522 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4525 spin_lock_irq(&conf->device_lock);
4526 conf->prev = conf->geo;
4527 md_finish_reshape(conf->mddev);
4529 conf->reshape_progress = MaxSector;
4530 conf->reshape_safe = MaxSector;
4531 spin_unlock_irq(&conf->device_lock);
4533 /* read-ahead size must cover two whole stripes, which is
4534 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4536 if (conf->mddev->queue) {
4537 int stripe = conf->geo.raid_disks *
4538 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4539 stripe /= conf->geo.near_copies;
4540 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4541 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4546 static int handle_reshape_read_error(struct mddev *mddev,
4547 struct r10bio *r10_bio)
4549 /* Use sync reads to get the blocks from somewhere else */
4550 int sectors = r10_bio->sectors;
4551 struct r10conf *conf = mddev->private;
4553 struct r10bio r10_bio;
4554 struct r10dev devs[conf->copies];
4556 struct r10bio *r10b = &on_stack.r10_bio;
4559 struct page **pages;
4561 /* reshape IOs share pages from .devs[0].bio */
4562 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4564 r10b->sector = r10_bio->sector;
4565 __raid10_find_phys(&conf->prev, r10b);
4570 int first_slot = slot;
4572 if (s > (PAGE_SIZE >> 9))
4577 int d = r10b->devs[slot].devnum;
4578 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4581 test_bit(Faulty, &rdev->flags) ||
4582 !test_bit(In_sync, &rdev->flags))
4585 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4586 atomic_inc(&rdev->nr_pending);
4588 success = sync_page_io(rdev,
4592 REQ_OP_READ, 0, false);
4593 rdev_dec_pending(rdev, mddev);
4599 if (slot >= conf->copies)
4601 if (slot == first_slot)
4606 /* couldn't read this block, must give up */
4607 set_bit(MD_RECOVERY_INTR,
4617 static void end_reshape_write(struct bio *bio)
4619 struct r10bio *r10_bio = get_resync_r10bio(bio);
4620 struct mddev *mddev = r10_bio->mddev;
4621 struct r10conf *conf = mddev->private;
4625 struct md_rdev *rdev = NULL;
4627 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4629 rdev = conf->mirrors[d].replacement;
4632 rdev = conf->mirrors[d].rdev;
4635 if (bio->bi_status) {
4636 /* FIXME should record badblock */
4637 md_error(mddev, rdev);
4640 rdev_dec_pending(rdev, mddev);
4641 end_reshape_request(r10_bio);
4644 static void end_reshape_request(struct r10bio *r10_bio)
4646 if (!atomic_dec_and_test(&r10_bio->remaining))
4648 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4649 bio_put(r10_bio->master_bio);
4653 static void raid10_finish_reshape(struct mddev *mddev)
4655 struct r10conf *conf = mddev->private;
4657 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4660 if (mddev->delta_disks > 0) {
4661 sector_t size = raid10_size(mddev, 0, 0);
4662 md_set_array_sectors(mddev, size);
4663 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4664 mddev->recovery_cp = mddev->resync_max_sectors;
4665 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4667 mddev->resync_max_sectors = size;
4669 set_capacity(mddev->gendisk, mddev->array_sectors);
4670 revalidate_disk(mddev->gendisk);
4675 for (d = conf->geo.raid_disks ;
4676 d < conf->geo.raid_disks - mddev->delta_disks;
4678 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4680 clear_bit(In_sync, &rdev->flags);
4681 rdev = rcu_dereference(conf->mirrors[d].replacement);
4683 clear_bit(In_sync, &rdev->flags);
4687 mddev->layout = mddev->new_layout;
4688 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4689 mddev->reshape_position = MaxSector;
4690 mddev->delta_disks = 0;
4691 mddev->reshape_backwards = 0;
4694 static struct md_personality raid10_personality =
4698 .owner = THIS_MODULE,
4699 .make_request = raid10_make_request,
4701 .free = raid10_free,
4702 .status = raid10_status,
4703 .error_handler = raid10_error,
4704 .hot_add_disk = raid10_add_disk,
4705 .hot_remove_disk= raid10_remove_disk,
4706 .spare_active = raid10_spare_active,
4707 .sync_request = raid10_sync_request,
4708 .quiesce = raid10_quiesce,
4709 .size = raid10_size,
4710 .resize = raid10_resize,
4711 .takeover = raid10_takeover,
4712 .check_reshape = raid10_check_reshape,
4713 .start_reshape = raid10_start_reshape,
4714 .finish_reshape = raid10_finish_reshape,
4715 .congested = raid10_congested,
4718 static int __init raid_init(void)
4720 return register_md_personality(&raid10_personality);
4723 static void raid_exit(void)
4725 unregister_md_personality(&raid10_personality);
4728 module_init(raid_init);
4729 module_exit(raid_exit);
4730 MODULE_LICENSE("GPL");
4731 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4732 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4733 MODULE_ALIAS("md-raid10");
4734 MODULE_ALIAS("md-level-10");
4736 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);