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)
114 * 'strct resync_pages' stores actual pages used for doing the resync
115 * IO, and it is per-bio, so make .bi_private points to it.
117 static inline struct resync_pages *get_resync_pages(struct bio *bio)
119 return bio->bi_private;
123 * for resync bio, r10bio pointer can be retrieved from the per-bio
124 * 'struct resync_pages'.
126 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
128 return get_resync_pages(bio)->raid_bio;
131 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
133 struct r10conf *conf = data;
134 int size = offsetof(struct r10bio, devs[conf->copies]);
136 /* allocate a r10bio with room for raid_disks entries in the
138 return kzalloc(size, gfp_flags);
141 static void r10bio_pool_free(void *r10_bio, void *data)
146 /* amount of memory to reserve for resync requests */
147 #define RESYNC_WINDOW (1024*1024)
148 /* maximum number of concurrent requests, memory permitting */
149 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
152 * When performing a resync, we need to read and compare, so
153 * we need as many pages are there are copies.
154 * When performing a recovery, we need 2 bios, one for read,
155 * one for write (we recover only one drive per r10buf)
158 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
160 struct r10conf *conf = data;
161 struct r10bio *r10_bio;
164 int nalloc, nalloc_rp;
165 struct resync_pages *rps;
167 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
171 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
172 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
173 nalloc = conf->copies; /* resync */
175 nalloc = 2; /* recovery */
177 /* allocate once for all bios */
178 if (!conf->have_replacement)
181 nalloc_rp = nalloc * 2;
182 rps = kmalloc(sizeof(struct resync_pages) * nalloc_rp, gfp_flags);
184 goto out_free_r10bio;
189 for (j = nalloc ; j-- ; ) {
190 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
193 r10_bio->devs[j].bio = bio;
194 if (!conf->have_replacement)
196 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
199 r10_bio->devs[j].repl_bio = bio;
202 * Allocate RESYNC_PAGES data pages and attach them
205 for (j = 0; j < nalloc; j++) {
206 struct bio *rbio = r10_bio->devs[j].repl_bio;
207 struct resync_pages *rp, *rp_repl;
211 rp_repl = &rps[nalloc + j];
213 bio = r10_bio->devs[j].bio;
215 if (!j || test_bit(MD_RECOVERY_SYNC,
216 &conf->mddev->recovery)) {
217 if (resync_alloc_pages(rp, gfp_flags))
220 memcpy(rp, &rps[0], sizeof(*rp));
221 resync_get_all_pages(rp);
225 rp->raid_bio = r10_bio;
226 bio->bi_private = rp;
228 memcpy(rp_repl, rp, sizeof(*rp));
229 rbio->bi_private = rp_repl;
237 resync_free_pages(&rps[j * 2]);
241 for ( ; j < nalloc; j++) {
242 if (r10_bio->devs[j].bio)
243 bio_put(r10_bio->devs[j].bio);
244 if (r10_bio->devs[j].repl_bio)
245 bio_put(r10_bio->devs[j].repl_bio);
249 r10bio_pool_free(r10_bio, conf);
253 static void r10buf_pool_free(void *__r10_bio, void *data)
255 struct r10conf *conf = data;
256 struct r10bio *r10bio = __r10_bio;
258 struct resync_pages *rp = NULL;
260 for (j = conf->copies; j--; ) {
261 struct bio *bio = r10bio->devs[j].bio;
263 rp = get_resync_pages(bio);
264 resync_free_pages(rp);
267 bio = r10bio->devs[j].repl_bio;
272 /* resync pages array stored in the 1st bio's .bi_private */
275 r10bio_pool_free(r10bio, conf);
278 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
282 for (i = 0; i < conf->copies; i++) {
283 struct bio **bio = & r10_bio->devs[i].bio;
284 if (!BIO_SPECIAL(*bio))
287 bio = &r10_bio->devs[i].repl_bio;
288 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
294 static void free_r10bio(struct r10bio *r10_bio)
296 struct r10conf *conf = r10_bio->mddev->private;
298 put_all_bios(conf, r10_bio);
299 mempool_free(r10_bio, conf->r10bio_pool);
302 static void put_buf(struct r10bio *r10_bio)
304 struct r10conf *conf = r10_bio->mddev->private;
306 mempool_free(r10_bio, conf->r10buf_pool);
311 static void reschedule_retry(struct r10bio *r10_bio)
314 struct mddev *mddev = r10_bio->mddev;
315 struct r10conf *conf = mddev->private;
317 spin_lock_irqsave(&conf->device_lock, flags);
318 list_add(&r10_bio->retry_list, &conf->retry_list);
320 spin_unlock_irqrestore(&conf->device_lock, flags);
322 /* wake up frozen array... */
323 wake_up(&conf->wait_barrier);
325 md_wakeup_thread(mddev->thread);
329 * raid_end_bio_io() is called when we have finished servicing a mirrored
330 * operation and are ready to return a success/failure code to the buffer
333 static void raid_end_bio_io(struct r10bio *r10_bio)
335 struct bio *bio = r10_bio->master_bio;
336 struct r10conf *conf = r10_bio->mddev->private;
338 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
339 bio->bi_error = -EIO;
343 * Wake up any possible resync thread that waits for the device
348 free_r10bio(r10_bio);
352 * Update disk head position estimator based on IRQ completion info.
354 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
356 struct r10conf *conf = r10_bio->mddev->private;
358 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
359 r10_bio->devs[slot].addr + (r10_bio->sectors);
363 * Find the disk number which triggered given bio
365 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
366 struct bio *bio, int *slotp, int *replp)
371 for (slot = 0; slot < conf->copies; slot++) {
372 if (r10_bio->devs[slot].bio == bio)
374 if (r10_bio->devs[slot].repl_bio == bio) {
380 BUG_ON(slot == conf->copies);
381 update_head_pos(slot, r10_bio);
387 return r10_bio->devs[slot].devnum;
390 static void raid10_end_read_request(struct bio *bio)
392 int uptodate = !bio->bi_error;
393 struct r10bio *r10_bio = bio->bi_private;
395 struct md_rdev *rdev;
396 struct r10conf *conf = r10_bio->mddev->private;
398 slot = r10_bio->read_slot;
399 dev = r10_bio->devs[slot].devnum;
400 rdev = r10_bio->devs[slot].rdev;
402 * this branch is our 'one mirror IO has finished' event handler:
404 update_head_pos(slot, r10_bio);
408 * Set R10BIO_Uptodate in our master bio, so that
409 * we will return a good error code to the higher
410 * levels even if IO on some other mirrored buffer fails.
412 * The 'master' represents the composite IO operation to
413 * user-side. So if something waits for IO, then it will
414 * wait for the 'master' bio.
416 set_bit(R10BIO_Uptodate, &r10_bio->state);
418 /* If all other devices that store this block have
419 * failed, we want to return the error upwards rather
420 * than fail the last device. Here we redefine
421 * "uptodate" to mean "Don't want to retry"
423 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
428 raid_end_bio_io(r10_bio);
429 rdev_dec_pending(rdev, conf->mddev);
432 * oops, read error - keep the refcount on the rdev
434 char b[BDEVNAME_SIZE];
435 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
437 bdevname(rdev->bdev, b),
438 (unsigned long long)r10_bio->sector);
439 set_bit(R10BIO_ReadError, &r10_bio->state);
440 reschedule_retry(r10_bio);
444 static void close_write(struct r10bio *r10_bio)
446 /* clear the bitmap if all writes complete successfully */
447 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
449 !test_bit(R10BIO_Degraded, &r10_bio->state),
451 md_write_end(r10_bio->mddev);
454 static void one_write_done(struct r10bio *r10_bio)
456 if (atomic_dec_and_test(&r10_bio->remaining)) {
457 if (test_bit(R10BIO_WriteError, &r10_bio->state))
458 reschedule_retry(r10_bio);
460 close_write(r10_bio);
461 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
462 reschedule_retry(r10_bio);
464 raid_end_bio_io(r10_bio);
469 static void raid10_end_write_request(struct bio *bio)
471 struct r10bio *r10_bio = bio->bi_private;
474 struct r10conf *conf = r10_bio->mddev->private;
476 struct md_rdev *rdev = NULL;
477 struct bio *to_put = NULL;
480 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
482 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
485 rdev = conf->mirrors[dev].replacement;
489 rdev = conf->mirrors[dev].rdev;
492 * this branch is our 'one mirror IO has finished' event handler:
494 if (bio->bi_error && !discard_error) {
496 /* Never record new bad blocks to replacement,
499 md_error(rdev->mddev, rdev);
501 set_bit(WriteErrorSeen, &rdev->flags);
502 if (!test_and_set_bit(WantReplacement, &rdev->flags))
503 set_bit(MD_RECOVERY_NEEDED,
504 &rdev->mddev->recovery);
507 if (test_bit(FailFast, &rdev->flags) &&
508 (bio->bi_opf & MD_FAILFAST)) {
509 md_error(rdev->mddev, rdev);
510 if (!test_bit(Faulty, &rdev->flags))
511 /* This is the only remaining device,
512 * We need to retry the write without
515 set_bit(R10BIO_WriteError, &r10_bio->state);
517 r10_bio->devs[slot].bio = NULL;
522 set_bit(R10BIO_WriteError, &r10_bio->state);
526 * Set R10BIO_Uptodate in our master bio, so that
527 * we will return a good error code for to the higher
528 * levels even if IO on some other mirrored buffer fails.
530 * The 'master' represents the composite IO operation to
531 * user-side. So if something waits for IO, then it will
532 * wait for the 'master' bio.
538 * Do not set R10BIO_Uptodate if the current device is
539 * rebuilding or Faulty. This is because we cannot use
540 * such device for properly reading the data back (we could
541 * potentially use it, if the current write would have felt
542 * before rdev->recovery_offset, but for simplicity we don't
545 if (test_bit(In_sync, &rdev->flags) &&
546 !test_bit(Faulty, &rdev->flags))
547 set_bit(R10BIO_Uptodate, &r10_bio->state);
549 /* Maybe we can clear some bad blocks. */
550 if (is_badblock(rdev,
551 r10_bio->devs[slot].addr,
553 &first_bad, &bad_sectors) && !discard_error) {
556 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
558 r10_bio->devs[slot].bio = IO_MADE_GOOD;
560 set_bit(R10BIO_MadeGood, &r10_bio->state);
566 * Let's see if all mirrored write operations have finished
569 one_write_done(r10_bio);
571 rdev_dec_pending(rdev, conf->mddev);
577 * RAID10 layout manager
578 * As well as the chunksize and raid_disks count, there are two
579 * parameters: near_copies and far_copies.
580 * near_copies * far_copies must be <= raid_disks.
581 * Normally one of these will be 1.
582 * If both are 1, we get raid0.
583 * If near_copies == raid_disks, we get raid1.
585 * Chunks are laid out in raid0 style with near_copies copies of the
586 * first chunk, followed by near_copies copies of the next chunk and
588 * If far_copies > 1, then after 1/far_copies of the array has been assigned
589 * as described above, we start again with a device offset of near_copies.
590 * So we effectively have another copy of the whole array further down all
591 * the drives, but with blocks on different drives.
592 * With this layout, and block is never stored twice on the one device.
594 * raid10_find_phys finds the sector offset of a given virtual sector
595 * on each device that it is on.
597 * raid10_find_virt does the reverse mapping, from a device and a
598 * sector offset to a virtual address
601 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
609 int last_far_set_start, last_far_set_size;
611 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
612 last_far_set_start *= geo->far_set_size;
614 last_far_set_size = geo->far_set_size;
615 last_far_set_size += (geo->raid_disks % geo->far_set_size);
617 /* now calculate first sector/dev */
618 chunk = r10bio->sector >> geo->chunk_shift;
619 sector = r10bio->sector & geo->chunk_mask;
621 chunk *= geo->near_copies;
623 dev = sector_div(stripe, geo->raid_disks);
625 stripe *= geo->far_copies;
627 sector += stripe << geo->chunk_shift;
629 /* and calculate all the others */
630 for (n = 0; n < geo->near_copies; n++) {
634 r10bio->devs[slot].devnum = d;
635 r10bio->devs[slot].addr = s;
638 for (f = 1; f < geo->far_copies; f++) {
639 set = d / geo->far_set_size;
640 d += geo->near_copies;
642 if ((geo->raid_disks % geo->far_set_size) &&
643 (d > last_far_set_start)) {
644 d -= last_far_set_start;
645 d %= last_far_set_size;
646 d += last_far_set_start;
648 d %= geo->far_set_size;
649 d += geo->far_set_size * set;
652 r10bio->devs[slot].devnum = d;
653 r10bio->devs[slot].addr = s;
657 if (dev >= geo->raid_disks) {
659 sector += (geo->chunk_mask + 1);
664 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
666 struct geom *geo = &conf->geo;
668 if (conf->reshape_progress != MaxSector &&
669 ((r10bio->sector >= conf->reshape_progress) !=
670 conf->mddev->reshape_backwards)) {
671 set_bit(R10BIO_Previous, &r10bio->state);
674 clear_bit(R10BIO_Previous, &r10bio->state);
676 __raid10_find_phys(geo, r10bio);
679 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
681 sector_t offset, chunk, vchunk;
682 /* Never use conf->prev as this is only called during resync
683 * or recovery, so reshape isn't happening
685 struct geom *geo = &conf->geo;
686 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
687 int far_set_size = geo->far_set_size;
688 int last_far_set_start;
690 if (geo->raid_disks % geo->far_set_size) {
691 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
692 last_far_set_start *= geo->far_set_size;
694 if (dev >= last_far_set_start) {
695 far_set_size = geo->far_set_size;
696 far_set_size += (geo->raid_disks % geo->far_set_size);
697 far_set_start = last_far_set_start;
701 offset = sector & geo->chunk_mask;
702 if (geo->far_offset) {
704 chunk = sector >> geo->chunk_shift;
705 fc = sector_div(chunk, geo->far_copies);
706 dev -= fc * geo->near_copies;
707 if (dev < far_set_start)
710 while (sector >= geo->stride) {
711 sector -= geo->stride;
712 if (dev < (geo->near_copies + far_set_start))
713 dev += far_set_size - geo->near_copies;
715 dev -= geo->near_copies;
717 chunk = sector >> geo->chunk_shift;
719 vchunk = chunk * geo->raid_disks + dev;
720 sector_div(vchunk, geo->near_copies);
721 return (vchunk << geo->chunk_shift) + offset;
725 * This routine returns the disk from which the requested read should
726 * be done. There is a per-array 'next expected sequential IO' sector
727 * number - if this matches on the next IO then we use the last disk.
728 * There is also a per-disk 'last know head position' sector that is
729 * maintained from IRQ contexts, both the normal and the resync IO
730 * completion handlers update this position correctly. If there is no
731 * perfect sequential match then we pick the disk whose head is closest.
733 * If there are 2 mirrors in the same 2 devices, performance degrades
734 * because position is mirror, not device based.
736 * The rdev for the device selected will have nr_pending incremented.
740 * FIXME: possibly should rethink readbalancing and do it differently
741 * depending on near_copies / far_copies geometry.
743 static struct md_rdev *read_balance(struct r10conf *conf,
744 struct r10bio *r10_bio,
747 const sector_t this_sector = r10_bio->sector;
749 int sectors = r10_bio->sectors;
750 int best_good_sectors;
751 sector_t new_distance, best_dist;
752 struct md_rdev *best_rdev, *rdev = NULL;
755 struct geom *geo = &conf->geo;
757 raid10_find_phys(conf, r10_bio);
759 sectors = r10_bio->sectors;
762 best_dist = MaxSector;
763 best_good_sectors = 0;
765 clear_bit(R10BIO_FailFast, &r10_bio->state);
767 * Check if we can balance. We can balance on the whole
768 * device if no resync is going on (recovery is ok), or below
769 * the resync window. We take the first readable disk when
770 * above the resync window.
772 if (conf->mddev->recovery_cp < MaxSector
773 && (this_sector + sectors >= conf->next_resync))
776 for (slot = 0; slot < conf->copies ; slot++) {
781 if (r10_bio->devs[slot].bio == IO_BLOCKED)
783 disk = r10_bio->devs[slot].devnum;
784 rdev = rcu_dereference(conf->mirrors[disk].replacement);
785 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
786 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
787 rdev = rcu_dereference(conf->mirrors[disk].rdev);
789 test_bit(Faulty, &rdev->flags))
791 if (!test_bit(In_sync, &rdev->flags) &&
792 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
795 dev_sector = r10_bio->devs[slot].addr;
796 if (is_badblock(rdev, dev_sector, sectors,
797 &first_bad, &bad_sectors)) {
798 if (best_dist < MaxSector)
799 /* Already have a better slot */
801 if (first_bad <= dev_sector) {
802 /* Cannot read here. If this is the
803 * 'primary' device, then we must not read
804 * beyond 'bad_sectors' from another device.
806 bad_sectors -= (dev_sector - first_bad);
807 if (!do_balance && sectors > bad_sectors)
808 sectors = bad_sectors;
809 if (best_good_sectors > sectors)
810 best_good_sectors = sectors;
812 sector_t good_sectors =
813 first_bad - dev_sector;
814 if (good_sectors > best_good_sectors) {
815 best_good_sectors = good_sectors;
820 /* Must read from here */
825 best_good_sectors = sectors;
831 /* At least 2 disks to choose from so failfast is OK */
832 set_bit(R10BIO_FailFast, &r10_bio->state);
833 /* This optimisation is debatable, and completely destroys
834 * sequential read speed for 'far copies' arrays. So only
835 * keep it for 'near' arrays, and review those later.
837 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
840 /* for far > 1 always use the lowest address */
841 else if (geo->far_copies > 1)
842 new_distance = r10_bio->devs[slot].addr;
844 new_distance = abs(r10_bio->devs[slot].addr -
845 conf->mirrors[disk].head_position);
846 if (new_distance < best_dist) {
847 best_dist = new_distance;
852 if (slot >= conf->copies) {
858 atomic_inc(&rdev->nr_pending);
859 r10_bio->read_slot = slot;
863 *max_sectors = best_good_sectors;
868 static int raid10_congested(struct mddev *mddev, int bits)
870 struct r10conf *conf = mddev->private;
873 if ((bits & (1 << WB_async_congested)) &&
874 conf->pending_count >= max_queued_requests)
879 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
882 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
883 if (rdev && !test_bit(Faulty, &rdev->flags)) {
884 struct request_queue *q = bdev_get_queue(rdev->bdev);
886 ret |= bdi_congested(q->backing_dev_info, bits);
893 static void flush_pending_writes(struct r10conf *conf)
895 /* Any writes that have been queued but are awaiting
896 * bitmap updates get flushed here.
898 spin_lock_irq(&conf->device_lock);
900 if (conf->pending_bio_list.head) {
902 bio = bio_list_get(&conf->pending_bio_list);
903 conf->pending_count = 0;
904 spin_unlock_irq(&conf->device_lock);
905 /* flush any pending bitmap writes to disk
906 * before proceeding w/ I/O */
907 bitmap_unplug(conf->mddev->bitmap);
908 wake_up(&conf->wait_barrier);
910 while (bio) { /* submit pending writes */
911 struct bio *next = bio->bi_next;
912 struct md_rdev *rdev = (void*)bio->bi_bdev;
914 bio->bi_bdev = rdev->bdev;
915 if (test_bit(Faulty, &rdev->flags)) {
916 bio->bi_error = -EIO;
918 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
919 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
923 generic_make_request(bio);
927 spin_unlock_irq(&conf->device_lock);
931 * Sometimes we need to suspend IO while we do something else,
932 * either some resync/recovery, or reconfigure the array.
933 * To do this we raise a 'barrier'.
934 * The 'barrier' is a counter that can be raised multiple times
935 * to count how many activities are happening which preclude
937 * We can only raise the barrier if there is no pending IO.
938 * i.e. if nr_pending == 0.
939 * We choose only to raise the barrier if no-one is waiting for the
940 * barrier to go down. This means that as soon as an IO request
941 * is ready, no other operations which require a barrier will start
942 * until the IO request has had a chance.
944 * So: regular IO calls 'wait_barrier'. When that returns there
945 * is no backgroup IO happening, It must arrange to call
946 * allow_barrier when it has finished its IO.
947 * backgroup IO calls must call raise_barrier. Once that returns
948 * there is no normal IO happeing. It must arrange to call
949 * lower_barrier when the particular background IO completes.
952 static void raise_barrier(struct r10conf *conf, int force)
954 BUG_ON(force && !conf->barrier);
955 spin_lock_irq(&conf->resync_lock);
957 /* Wait until no block IO is waiting (unless 'force') */
958 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
961 /* block any new IO from starting */
964 /* Now wait for all pending IO to complete */
965 wait_event_lock_irq(conf->wait_barrier,
966 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
969 spin_unlock_irq(&conf->resync_lock);
972 static void lower_barrier(struct r10conf *conf)
975 spin_lock_irqsave(&conf->resync_lock, flags);
977 spin_unlock_irqrestore(&conf->resync_lock, flags);
978 wake_up(&conf->wait_barrier);
981 static void wait_barrier(struct r10conf *conf)
983 spin_lock_irq(&conf->resync_lock);
986 /* Wait for the barrier to drop.
987 * However if there are already pending
988 * requests (preventing the barrier from
989 * rising completely), and the
990 * pre-process bio queue isn't empty,
991 * then don't wait, as we need to empty
992 * that queue to get the nr_pending
995 raid10_log(conf->mddev, "wait barrier");
996 wait_event_lock_irq(conf->wait_barrier,
998 (atomic_read(&conf->nr_pending) &&
1000 (!bio_list_empty(¤t->bio_list[0]) ||
1001 !bio_list_empty(¤t->bio_list[1]))),
1004 if (!conf->nr_waiting)
1005 wake_up(&conf->wait_barrier);
1007 atomic_inc(&conf->nr_pending);
1008 spin_unlock_irq(&conf->resync_lock);
1011 static void allow_barrier(struct r10conf *conf)
1013 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1014 (conf->array_freeze_pending))
1015 wake_up(&conf->wait_barrier);
1018 static void freeze_array(struct r10conf *conf, int extra)
1020 /* stop syncio and normal IO and wait for everything to
1022 * We increment barrier and nr_waiting, and then
1023 * wait until nr_pending match nr_queued+extra
1024 * This is called in the context of one normal IO request
1025 * that has failed. Thus any sync request that might be pending
1026 * will be blocked by nr_pending, and we need to wait for
1027 * pending IO requests to complete or be queued for re-try.
1028 * Thus the number queued (nr_queued) plus this request (extra)
1029 * must match the number of pending IOs (nr_pending) before
1032 spin_lock_irq(&conf->resync_lock);
1033 conf->array_freeze_pending++;
1036 wait_event_lock_irq_cmd(conf->wait_barrier,
1037 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1039 flush_pending_writes(conf));
1041 conf->array_freeze_pending--;
1042 spin_unlock_irq(&conf->resync_lock);
1045 static void unfreeze_array(struct r10conf *conf)
1047 /* reverse the effect of the freeze */
1048 spin_lock_irq(&conf->resync_lock);
1051 wake_up(&conf->wait_barrier);
1052 spin_unlock_irq(&conf->resync_lock);
1055 static sector_t choose_data_offset(struct r10bio *r10_bio,
1056 struct md_rdev *rdev)
1058 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1059 test_bit(R10BIO_Previous, &r10_bio->state))
1060 return rdev->data_offset;
1062 return rdev->new_data_offset;
1065 struct raid10_plug_cb {
1066 struct blk_plug_cb cb;
1067 struct bio_list pending;
1071 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1073 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1075 struct mddev *mddev = plug->cb.data;
1076 struct r10conf *conf = mddev->private;
1079 if (from_schedule || current->bio_list) {
1080 spin_lock_irq(&conf->device_lock);
1081 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1082 conf->pending_count += plug->pending_cnt;
1083 spin_unlock_irq(&conf->device_lock);
1084 wake_up(&conf->wait_barrier);
1085 md_wakeup_thread(mddev->thread);
1090 /* we aren't scheduling, so we can do the write-out directly. */
1091 bio = bio_list_get(&plug->pending);
1092 bitmap_unplug(mddev->bitmap);
1093 wake_up(&conf->wait_barrier);
1095 while (bio) { /* submit pending writes */
1096 struct bio *next = bio->bi_next;
1097 struct md_rdev *rdev = (void*)bio->bi_bdev;
1098 bio->bi_next = NULL;
1099 bio->bi_bdev = rdev->bdev;
1100 if (test_bit(Faulty, &rdev->flags)) {
1101 bio->bi_error = -EIO;
1103 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1104 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1105 /* Just ignore it */
1108 generic_make_request(bio);
1114 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1115 struct r10bio *r10_bio)
1117 struct r10conf *conf = mddev->private;
1118 struct bio *read_bio;
1119 const int op = bio_op(bio);
1120 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1123 struct md_rdev *rdev;
1124 char b[BDEVNAME_SIZE];
1125 int slot = r10_bio->read_slot;
1126 struct md_rdev *err_rdev = NULL;
1127 gfp_t gfp = GFP_NOIO;
1129 if (r10_bio->devs[slot].rdev) {
1131 * This is an error retry, but we cannot
1132 * safely dereference the rdev in the r10_bio,
1133 * we must use the one in conf.
1134 * If it has already been disconnected (unlikely)
1135 * we lose the device name in error messages.
1139 * As we are blocking raid10, it is a little safer to
1142 gfp = GFP_NOIO | __GFP_HIGH;
1145 disk = r10_bio->devs[slot].devnum;
1146 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1148 bdevname(err_rdev->bdev, b);
1151 /* This never gets dereferenced */
1152 err_rdev = r10_bio->devs[slot].rdev;
1157 * Register the new request and wait if the reconstruction
1158 * thread has put up a bar for new requests.
1159 * Continue immediately if no resync is active currently.
1163 sectors = r10_bio->sectors;
1164 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1165 bio->bi_iter.bi_sector < conf->reshape_progress &&
1166 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1168 * IO spans the reshape position. Need to wait for reshape to
1171 raid10_log(conf->mddev, "wait reshape");
1172 allow_barrier(conf);
1173 wait_event(conf->wait_barrier,
1174 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1175 conf->reshape_progress >= bio->bi_iter.bi_sector +
1180 rdev = read_balance(conf, r10_bio, &max_sectors);
1183 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1185 (unsigned long long)r10_bio->sector);
1187 raid_end_bio_io(r10_bio);
1191 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1193 bdevname(rdev->bdev, b),
1194 (unsigned long long)r10_bio->sector);
1195 if (max_sectors < bio_sectors(bio)) {
1196 struct bio *split = bio_split(bio, max_sectors,
1197 gfp, conf->bio_split);
1198 bio_chain(split, bio);
1199 generic_make_request(bio);
1201 r10_bio->master_bio = bio;
1202 r10_bio->sectors = max_sectors;
1204 slot = r10_bio->read_slot;
1206 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1208 r10_bio->devs[slot].bio = read_bio;
1209 r10_bio->devs[slot].rdev = rdev;
1211 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1212 choose_data_offset(r10_bio, rdev);
1213 read_bio->bi_bdev = rdev->bdev;
1214 read_bio->bi_end_io = raid10_end_read_request;
1215 bio_set_op_attrs(read_bio, op, do_sync);
1216 if (test_bit(FailFast, &rdev->flags) &&
1217 test_bit(R10BIO_FailFast, &r10_bio->state))
1218 read_bio->bi_opf |= MD_FAILFAST;
1219 read_bio->bi_private = r10_bio;
1222 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1223 read_bio, disk_devt(mddev->gendisk),
1225 generic_make_request(read_bio);
1229 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1230 struct bio *bio, bool replacement,
1233 const int op = bio_op(bio);
1234 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1235 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1236 unsigned long flags;
1237 struct blk_plug_cb *cb;
1238 struct raid10_plug_cb *plug = NULL;
1239 struct r10conf *conf = mddev->private;
1240 struct md_rdev *rdev;
1241 int devnum = r10_bio->devs[n_copy].devnum;
1245 rdev = conf->mirrors[devnum].replacement;
1247 /* Replacement just got moved to main 'rdev' */
1249 rdev = conf->mirrors[devnum].rdev;
1252 rdev = conf->mirrors[devnum].rdev;
1254 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1256 r10_bio->devs[n_copy].repl_bio = mbio;
1258 r10_bio->devs[n_copy].bio = mbio;
1260 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1261 choose_data_offset(r10_bio, rdev));
1262 mbio->bi_bdev = rdev->bdev;
1263 mbio->bi_end_io = raid10_end_write_request;
1264 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1265 if (!replacement && test_bit(FailFast,
1266 &conf->mirrors[devnum].rdev->flags)
1267 && enough(conf, devnum))
1268 mbio->bi_opf |= MD_FAILFAST;
1269 mbio->bi_private = r10_bio;
1271 if (conf->mddev->gendisk)
1272 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1273 mbio, disk_devt(conf->mddev->gendisk),
1275 /* flush_pending_writes() needs access to the rdev so...*/
1276 mbio->bi_bdev = (void *)rdev;
1278 atomic_inc(&r10_bio->remaining);
1280 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1282 plug = container_of(cb, struct raid10_plug_cb, cb);
1286 bio_list_add(&plug->pending, mbio);
1287 plug->pending_cnt++;
1289 spin_lock_irqsave(&conf->device_lock, flags);
1290 bio_list_add(&conf->pending_bio_list, mbio);
1291 conf->pending_count++;
1292 spin_unlock_irqrestore(&conf->device_lock, flags);
1293 md_wakeup_thread(mddev->thread);
1297 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1298 struct r10bio *r10_bio)
1300 struct r10conf *conf = mddev->private;
1302 struct md_rdev *blocked_rdev;
1306 md_write_start(mddev, bio);
1309 * Register the new request and wait if the reconstruction
1310 * thread has put up a bar for new requests.
1311 * Continue immediately if no resync is active currently.
1315 sectors = r10_bio->sectors;
1316 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1317 bio->bi_iter.bi_sector < conf->reshape_progress &&
1318 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1320 * IO spans the reshape position. Need to wait for reshape to
1323 raid10_log(conf->mddev, "wait reshape");
1324 allow_barrier(conf);
1325 wait_event(conf->wait_barrier,
1326 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1327 conf->reshape_progress >= bio->bi_iter.bi_sector +
1332 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1333 (mddev->reshape_backwards
1334 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1335 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1336 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1337 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1338 /* Need to update reshape_position in metadata */
1339 mddev->reshape_position = conf->reshape_progress;
1340 set_mask_bits(&mddev->sb_flags, 0,
1341 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1342 md_wakeup_thread(mddev->thread);
1343 raid10_log(conf->mddev, "wait reshape metadata");
1344 wait_event(mddev->sb_wait,
1345 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1347 conf->reshape_safe = mddev->reshape_position;
1350 if (conf->pending_count >= max_queued_requests) {
1351 md_wakeup_thread(mddev->thread);
1352 raid10_log(mddev, "wait queued");
1353 wait_event(conf->wait_barrier,
1354 conf->pending_count < max_queued_requests);
1356 /* first select target devices under rcu_lock and
1357 * inc refcount on their rdev. Record them by setting
1359 * If there are known/acknowledged bad blocks on any device
1360 * on which we have seen a write error, we want to avoid
1361 * writing to those blocks. This potentially requires several
1362 * writes to write around the bad blocks. Each set of writes
1363 * gets its own r10_bio with a set of bios attached.
1366 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1367 raid10_find_phys(conf, r10_bio);
1369 blocked_rdev = NULL;
1371 max_sectors = r10_bio->sectors;
1373 for (i = 0; i < conf->copies; i++) {
1374 int d = r10_bio->devs[i].devnum;
1375 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1376 struct md_rdev *rrdev = rcu_dereference(
1377 conf->mirrors[d].replacement);
1380 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1381 atomic_inc(&rdev->nr_pending);
1382 blocked_rdev = rdev;
1385 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1386 atomic_inc(&rrdev->nr_pending);
1387 blocked_rdev = rrdev;
1390 if (rdev && (test_bit(Faulty, &rdev->flags)))
1392 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1395 r10_bio->devs[i].bio = NULL;
1396 r10_bio->devs[i].repl_bio = NULL;
1398 if (!rdev && !rrdev) {
1399 set_bit(R10BIO_Degraded, &r10_bio->state);
1402 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1404 sector_t dev_sector = r10_bio->devs[i].addr;
1408 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1409 &first_bad, &bad_sectors);
1411 /* Mustn't write here until the bad block
1414 atomic_inc(&rdev->nr_pending);
1415 set_bit(BlockedBadBlocks, &rdev->flags);
1416 blocked_rdev = rdev;
1419 if (is_bad && first_bad <= dev_sector) {
1420 /* Cannot write here at all */
1421 bad_sectors -= (dev_sector - first_bad);
1422 if (bad_sectors < max_sectors)
1423 /* Mustn't write more than bad_sectors
1424 * to other devices yet
1426 max_sectors = bad_sectors;
1427 /* We don't set R10BIO_Degraded as that
1428 * only applies if the disk is missing,
1429 * so it might be re-added, and we want to
1430 * know to recover this chunk.
1431 * In this case the device is here, and the
1432 * fact that this chunk is not in-sync is
1433 * recorded in the bad block log.
1438 int good_sectors = first_bad - dev_sector;
1439 if (good_sectors < max_sectors)
1440 max_sectors = good_sectors;
1444 r10_bio->devs[i].bio = bio;
1445 atomic_inc(&rdev->nr_pending);
1448 r10_bio->devs[i].repl_bio = bio;
1449 atomic_inc(&rrdev->nr_pending);
1454 if (unlikely(blocked_rdev)) {
1455 /* Have to wait for this device to get unblocked, then retry */
1459 for (j = 0; j < i; j++) {
1460 if (r10_bio->devs[j].bio) {
1461 d = r10_bio->devs[j].devnum;
1462 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1464 if (r10_bio->devs[j].repl_bio) {
1465 struct md_rdev *rdev;
1466 d = r10_bio->devs[j].devnum;
1467 rdev = conf->mirrors[d].replacement;
1469 /* Race with remove_disk */
1471 rdev = conf->mirrors[d].rdev;
1473 rdev_dec_pending(rdev, mddev);
1476 allow_barrier(conf);
1477 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1478 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1483 if (max_sectors < r10_bio->sectors)
1484 r10_bio->sectors = max_sectors;
1486 if (r10_bio->sectors < bio_sectors(bio)) {
1487 struct bio *split = bio_split(bio, r10_bio->sectors,
1488 GFP_NOIO, conf->bio_split);
1489 bio_chain(split, bio);
1490 generic_make_request(bio);
1492 r10_bio->master_bio = bio;
1495 atomic_set(&r10_bio->remaining, 1);
1496 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1498 for (i = 0; i < conf->copies; i++) {
1499 if (r10_bio->devs[i].bio)
1500 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1501 if (r10_bio->devs[i].repl_bio)
1502 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1504 one_write_done(r10_bio);
1507 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1509 struct r10conf *conf = mddev->private;
1510 struct r10bio *r10_bio;
1512 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1514 r10_bio->master_bio = bio;
1515 r10_bio->sectors = sectors;
1517 r10_bio->mddev = mddev;
1518 r10_bio->sector = bio->bi_iter.bi_sector;
1520 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1522 if (bio_data_dir(bio) == READ)
1523 raid10_read_request(mddev, bio, r10_bio);
1525 raid10_write_request(mddev, bio, r10_bio);
1528 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1530 struct r10conf *conf = mddev->private;
1531 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1532 int chunk_sects = chunk_mask + 1;
1533 int sectors = bio_sectors(bio);
1535 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1536 md_flush_request(mddev, bio);
1541 * If this request crosses a chunk boundary, we need to split
1544 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1545 sectors > chunk_sects
1546 && (conf->geo.near_copies < conf->geo.raid_disks
1547 || conf->prev.near_copies <
1548 conf->prev.raid_disks)))
1549 sectors = chunk_sects -
1550 (bio->bi_iter.bi_sector &
1552 __make_request(mddev, bio, sectors);
1554 /* In case raid10d snuck in to freeze_array */
1555 wake_up(&conf->wait_barrier);
1558 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1560 struct r10conf *conf = mddev->private;
1563 if (conf->geo.near_copies < conf->geo.raid_disks)
1564 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1565 if (conf->geo.near_copies > 1)
1566 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1567 if (conf->geo.far_copies > 1) {
1568 if (conf->geo.far_offset)
1569 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1571 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1572 if (conf->geo.far_set_size != conf->geo.raid_disks)
1573 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1575 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1576 conf->geo.raid_disks - mddev->degraded);
1578 for (i = 0; i < conf->geo.raid_disks; i++) {
1579 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1580 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1583 seq_printf(seq, "]");
1586 /* check if there are enough drives for
1587 * every block to appear on atleast one.
1588 * Don't consider the device numbered 'ignore'
1589 * as we might be about to remove it.
1591 static int _enough(struct r10conf *conf, int previous, int ignore)
1597 disks = conf->prev.raid_disks;
1598 ncopies = conf->prev.near_copies;
1600 disks = conf->geo.raid_disks;
1601 ncopies = conf->geo.near_copies;
1606 int n = conf->copies;
1610 struct md_rdev *rdev;
1611 if (this != ignore &&
1612 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1613 test_bit(In_sync, &rdev->flags))
1615 this = (this+1) % disks;
1619 first = (first + ncopies) % disks;
1620 } while (first != 0);
1627 static int enough(struct r10conf *conf, int ignore)
1629 /* when calling 'enough', both 'prev' and 'geo' must
1631 * This is ensured if ->reconfig_mutex or ->device_lock
1634 return _enough(conf, 0, ignore) &&
1635 _enough(conf, 1, ignore);
1638 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1640 char b[BDEVNAME_SIZE];
1641 struct r10conf *conf = mddev->private;
1642 unsigned long flags;
1645 * If it is not operational, then we have already marked it as dead
1646 * else if it is the last working disks, ignore the error, let the
1647 * next level up know.
1648 * else mark the drive as failed
1650 spin_lock_irqsave(&conf->device_lock, flags);
1651 if (test_bit(In_sync, &rdev->flags)
1652 && !enough(conf, rdev->raid_disk)) {
1654 * Don't fail the drive, just return an IO error.
1656 spin_unlock_irqrestore(&conf->device_lock, flags);
1659 if (test_and_clear_bit(In_sync, &rdev->flags))
1662 * If recovery is running, make sure it aborts.
1664 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1665 set_bit(Blocked, &rdev->flags);
1666 set_bit(Faulty, &rdev->flags);
1667 set_mask_bits(&mddev->sb_flags, 0,
1668 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1669 spin_unlock_irqrestore(&conf->device_lock, flags);
1670 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1671 "md/raid10:%s: Operation continuing on %d devices.\n",
1672 mdname(mddev), bdevname(rdev->bdev, b),
1673 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1676 static void print_conf(struct r10conf *conf)
1679 struct md_rdev *rdev;
1681 pr_debug("RAID10 conf printout:\n");
1683 pr_debug("(!conf)\n");
1686 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1687 conf->geo.raid_disks);
1689 /* This is only called with ->reconfix_mutex held, so
1690 * rcu protection of rdev is not needed */
1691 for (i = 0; i < conf->geo.raid_disks; i++) {
1692 char b[BDEVNAME_SIZE];
1693 rdev = conf->mirrors[i].rdev;
1695 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1696 i, !test_bit(In_sync, &rdev->flags),
1697 !test_bit(Faulty, &rdev->flags),
1698 bdevname(rdev->bdev,b));
1702 static void close_sync(struct r10conf *conf)
1705 allow_barrier(conf);
1707 mempool_destroy(conf->r10buf_pool);
1708 conf->r10buf_pool = NULL;
1711 static int raid10_spare_active(struct mddev *mddev)
1714 struct r10conf *conf = mddev->private;
1715 struct raid10_info *tmp;
1717 unsigned long flags;
1720 * Find all non-in_sync disks within the RAID10 configuration
1721 * and mark them in_sync
1723 for (i = 0; i < conf->geo.raid_disks; i++) {
1724 tmp = conf->mirrors + i;
1725 if (tmp->replacement
1726 && tmp->replacement->recovery_offset == MaxSector
1727 && !test_bit(Faulty, &tmp->replacement->flags)
1728 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1729 /* Replacement has just become active */
1731 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1734 /* Replaced device not technically faulty,
1735 * but we need to be sure it gets removed
1736 * and never re-added.
1738 set_bit(Faulty, &tmp->rdev->flags);
1739 sysfs_notify_dirent_safe(
1740 tmp->rdev->sysfs_state);
1742 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1743 } else if (tmp->rdev
1744 && tmp->rdev->recovery_offset == MaxSector
1745 && !test_bit(Faulty, &tmp->rdev->flags)
1746 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1748 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1751 spin_lock_irqsave(&conf->device_lock, flags);
1752 mddev->degraded -= count;
1753 spin_unlock_irqrestore(&conf->device_lock, flags);
1759 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1761 struct r10conf *conf = mddev->private;
1765 int last = conf->geo.raid_disks - 1;
1767 if (mddev->recovery_cp < MaxSector)
1768 /* only hot-add to in-sync arrays, as recovery is
1769 * very different from resync
1772 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1775 if (md_integrity_add_rdev(rdev, mddev))
1778 if (rdev->raid_disk >= 0)
1779 first = last = rdev->raid_disk;
1781 if (rdev->saved_raid_disk >= first &&
1782 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1783 mirror = rdev->saved_raid_disk;
1786 for ( ; mirror <= last ; mirror++) {
1787 struct raid10_info *p = &conf->mirrors[mirror];
1788 if (p->recovery_disabled == mddev->recovery_disabled)
1791 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1792 p->replacement != NULL)
1794 clear_bit(In_sync, &rdev->flags);
1795 set_bit(Replacement, &rdev->flags);
1796 rdev->raid_disk = mirror;
1799 disk_stack_limits(mddev->gendisk, rdev->bdev,
1800 rdev->data_offset << 9);
1802 rcu_assign_pointer(p->replacement, rdev);
1807 disk_stack_limits(mddev->gendisk, rdev->bdev,
1808 rdev->data_offset << 9);
1810 p->head_position = 0;
1811 p->recovery_disabled = mddev->recovery_disabled - 1;
1812 rdev->raid_disk = mirror;
1814 if (rdev->saved_raid_disk != mirror)
1816 rcu_assign_pointer(p->rdev, rdev);
1819 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1820 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1826 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1828 struct r10conf *conf = mddev->private;
1830 int number = rdev->raid_disk;
1831 struct md_rdev **rdevp;
1832 struct raid10_info *p = conf->mirrors + number;
1835 if (rdev == p->rdev)
1837 else if (rdev == p->replacement)
1838 rdevp = &p->replacement;
1842 if (test_bit(In_sync, &rdev->flags) ||
1843 atomic_read(&rdev->nr_pending)) {
1847 /* Only remove non-faulty devices if recovery
1850 if (!test_bit(Faulty, &rdev->flags) &&
1851 mddev->recovery_disabled != p->recovery_disabled &&
1852 (!p->replacement || p->replacement == rdev) &&
1853 number < conf->geo.raid_disks &&
1859 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1861 if (atomic_read(&rdev->nr_pending)) {
1862 /* lost the race, try later */
1868 if (p->replacement) {
1869 /* We must have just cleared 'rdev' */
1870 p->rdev = p->replacement;
1871 clear_bit(Replacement, &p->replacement->flags);
1872 smp_mb(); /* Make sure other CPUs may see both as identical
1873 * but will never see neither -- if they are careful.
1875 p->replacement = NULL;
1878 clear_bit(WantReplacement, &rdev->flags);
1879 err = md_integrity_register(mddev);
1887 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1889 struct r10conf *conf = r10_bio->mddev->private;
1892 set_bit(R10BIO_Uptodate, &r10_bio->state);
1894 /* The write handler will notice the lack of
1895 * R10BIO_Uptodate and record any errors etc
1897 atomic_add(r10_bio->sectors,
1898 &conf->mirrors[d].rdev->corrected_errors);
1900 /* for reconstruct, we always reschedule after a read.
1901 * for resync, only after all reads
1903 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1904 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1905 atomic_dec_and_test(&r10_bio->remaining)) {
1906 /* we have read all the blocks,
1907 * do the comparison in process context in raid10d
1909 reschedule_retry(r10_bio);
1913 static void end_sync_read(struct bio *bio)
1915 struct r10bio *r10_bio = get_resync_r10bio(bio);
1916 struct r10conf *conf = r10_bio->mddev->private;
1917 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1919 __end_sync_read(r10_bio, bio, d);
1922 static void end_reshape_read(struct bio *bio)
1924 /* reshape read bio isn't allocated from r10buf_pool */
1925 struct r10bio *r10_bio = bio->bi_private;
1927 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1930 static void end_sync_request(struct r10bio *r10_bio)
1932 struct mddev *mddev = r10_bio->mddev;
1934 while (atomic_dec_and_test(&r10_bio->remaining)) {
1935 if (r10_bio->master_bio == NULL) {
1936 /* the primary of several recovery bios */
1937 sector_t s = r10_bio->sectors;
1938 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1939 test_bit(R10BIO_WriteError, &r10_bio->state))
1940 reschedule_retry(r10_bio);
1943 md_done_sync(mddev, s, 1);
1946 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1947 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1948 test_bit(R10BIO_WriteError, &r10_bio->state))
1949 reschedule_retry(r10_bio);
1957 static void end_sync_write(struct bio *bio)
1959 struct r10bio *r10_bio = get_resync_r10bio(bio);
1960 struct mddev *mddev = r10_bio->mddev;
1961 struct r10conf *conf = mddev->private;
1967 struct md_rdev *rdev = NULL;
1969 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1971 rdev = conf->mirrors[d].replacement;
1973 rdev = conf->mirrors[d].rdev;
1975 if (bio->bi_error) {
1977 md_error(mddev, rdev);
1979 set_bit(WriteErrorSeen, &rdev->flags);
1980 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1981 set_bit(MD_RECOVERY_NEEDED,
1982 &rdev->mddev->recovery);
1983 set_bit(R10BIO_WriteError, &r10_bio->state);
1985 } else if (is_badblock(rdev,
1986 r10_bio->devs[slot].addr,
1988 &first_bad, &bad_sectors))
1989 set_bit(R10BIO_MadeGood, &r10_bio->state);
1991 rdev_dec_pending(rdev, mddev);
1993 end_sync_request(r10_bio);
1997 * Note: sync and recover and handled very differently for raid10
1998 * This code is for resync.
1999 * For resync, we read through virtual addresses and read all blocks.
2000 * If there is any error, we schedule a write. The lowest numbered
2001 * drive is authoritative.
2002 * However requests come for physical address, so we need to map.
2003 * For every physical address there are raid_disks/copies virtual addresses,
2004 * which is always are least one, but is not necessarly an integer.
2005 * This means that a physical address can span multiple chunks, so we may
2006 * have to submit multiple io requests for a single sync request.
2009 * We check if all blocks are in-sync and only write to blocks that
2012 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2014 struct r10conf *conf = mddev->private;
2016 struct bio *tbio, *fbio;
2018 struct page **tpages, **fpages;
2020 atomic_set(&r10_bio->remaining, 1);
2022 /* find the first device with a block */
2023 for (i=0; i<conf->copies; i++)
2024 if (!r10_bio->devs[i].bio->bi_error)
2027 if (i == conf->copies)
2031 fbio = r10_bio->devs[i].bio;
2032 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2033 fbio->bi_iter.bi_idx = 0;
2034 fpages = get_resync_pages(fbio)->pages;
2036 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2037 /* now find blocks with errors */
2038 for (i=0 ; i < conf->copies ; i++) {
2040 struct md_rdev *rdev;
2041 struct resync_pages *rp;
2043 tbio = r10_bio->devs[i].bio;
2045 if (tbio->bi_end_io != end_sync_read)
2050 tpages = get_resync_pages(tbio)->pages;
2051 d = r10_bio->devs[i].devnum;
2052 rdev = conf->mirrors[d].rdev;
2053 if (!r10_bio->devs[i].bio->bi_error) {
2054 /* We know that the bi_io_vec layout is the same for
2055 * both 'first' and 'i', so we just compare them.
2056 * All vec entries are PAGE_SIZE;
2058 int sectors = r10_bio->sectors;
2059 for (j = 0; j < vcnt; j++) {
2060 int len = PAGE_SIZE;
2061 if (sectors < (len / 512))
2062 len = sectors * 512;
2063 if (memcmp(page_address(fpages[j]),
2064 page_address(tpages[j]),
2071 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2072 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2073 /* Don't fix anything. */
2075 } else if (test_bit(FailFast, &rdev->flags)) {
2076 /* Just give up on this device */
2077 md_error(rdev->mddev, rdev);
2080 /* Ok, we need to write this bio, either to correct an
2081 * inconsistency or to correct an unreadable block.
2082 * First we need to fixup bv_offset, bv_len and
2083 * bi_vecs, as the read request might have corrupted these
2085 rp = get_resync_pages(tbio);
2088 tbio->bi_vcnt = vcnt;
2089 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
2090 rp->raid_bio = r10_bio;
2091 tbio->bi_private = rp;
2092 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2093 tbio->bi_end_io = end_sync_write;
2094 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2096 bio_copy_data(tbio, fbio);
2098 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2099 atomic_inc(&r10_bio->remaining);
2100 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2102 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2103 tbio->bi_opf |= MD_FAILFAST;
2104 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2105 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2106 generic_make_request(tbio);
2109 /* Now write out to any replacement devices
2112 for (i = 0; i < conf->copies; i++) {
2115 tbio = r10_bio->devs[i].repl_bio;
2116 if (!tbio || !tbio->bi_end_io)
2118 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2119 && r10_bio->devs[i].bio != fbio)
2120 bio_copy_data(tbio, fbio);
2121 d = r10_bio->devs[i].devnum;
2122 atomic_inc(&r10_bio->remaining);
2123 md_sync_acct(conf->mirrors[d].replacement->bdev,
2125 generic_make_request(tbio);
2129 if (atomic_dec_and_test(&r10_bio->remaining)) {
2130 md_done_sync(mddev, r10_bio->sectors, 1);
2136 * Now for the recovery code.
2137 * Recovery happens across physical sectors.
2138 * We recover all non-is_sync drives by finding the virtual address of
2139 * each, and then choose a working drive that also has that virt address.
2140 * There is a separate r10_bio for each non-in_sync drive.
2141 * Only the first two slots are in use. The first for reading,
2142 * The second for writing.
2145 static void fix_recovery_read_error(struct r10bio *r10_bio)
2147 /* We got a read error during recovery.
2148 * We repeat the read in smaller page-sized sections.
2149 * If a read succeeds, write it to the new device or record
2150 * a bad block if we cannot.
2151 * If a read fails, record a bad block on both old and
2154 struct mddev *mddev = r10_bio->mddev;
2155 struct r10conf *conf = mddev->private;
2156 struct bio *bio = r10_bio->devs[0].bio;
2158 int sectors = r10_bio->sectors;
2160 int dr = r10_bio->devs[0].devnum;
2161 int dw = r10_bio->devs[1].devnum;
2162 struct page **pages = get_resync_pages(bio)->pages;
2166 struct md_rdev *rdev;
2170 if (s > (PAGE_SIZE>>9))
2173 rdev = conf->mirrors[dr].rdev;
2174 addr = r10_bio->devs[0].addr + sect,
2175 ok = sync_page_io(rdev,
2179 REQ_OP_READ, 0, false);
2181 rdev = conf->mirrors[dw].rdev;
2182 addr = r10_bio->devs[1].addr + sect;
2183 ok = sync_page_io(rdev,
2187 REQ_OP_WRITE, 0, false);
2189 set_bit(WriteErrorSeen, &rdev->flags);
2190 if (!test_and_set_bit(WantReplacement,
2192 set_bit(MD_RECOVERY_NEEDED,
2193 &rdev->mddev->recovery);
2197 /* We don't worry if we cannot set a bad block -
2198 * it really is bad so there is no loss in not
2201 rdev_set_badblocks(rdev, addr, s, 0);
2203 if (rdev != conf->mirrors[dw].rdev) {
2204 /* need bad block on destination too */
2205 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2206 addr = r10_bio->devs[1].addr + sect;
2207 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2209 /* just abort the recovery */
2210 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2213 conf->mirrors[dw].recovery_disabled
2214 = mddev->recovery_disabled;
2215 set_bit(MD_RECOVERY_INTR,
2228 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2230 struct r10conf *conf = mddev->private;
2232 struct bio *wbio, *wbio2;
2234 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2235 fix_recovery_read_error(r10_bio);
2236 end_sync_request(r10_bio);
2241 * share the pages with the first bio
2242 * and submit the write request
2244 d = r10_bio->devs[1].devnum;
2245 wbio = r10_bio->devs[1].bio;
2246 wbio2 = r10_bio->devs[1].repl_bio;
2247 /* Need to test wbio2->bi_end_io before we call
2248 * generic_make_request as if the former is NULL,
2249 * the latter is free to free wbio2.
2251 if (wbio2 && !wbio2->bi_end_io)
2253 if (wbio->bi_end_io) {
2254 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2255 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2256 generic_make_request(wbio);
2259 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2260 md_sync_acct(conf->mirrors[d].replacement->bdev,
2261 bio_sectors(wbio2));
2262 generic_make_request(wbio2);
2267 * Used by fix_read_error() to decay the per rdev read_errors.
2268 * We halve the read error count for every hour that has elapsed
2269 * since the last recorded read error.
2272 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2275 unsigned long hours_since_last;
2276 unsigned int read_errors = atomic_read(&rdev->read_errors);
2278 cur_time_mon = ktime_get_seconds();
2280 if (rdev->last_read_error == 0) {
2281 /* first time we've seen a read error */
2282 rdev->last_read_error = cur_time_mon;
2286 hours_since_last = (long)(cur_time_mon -
2287 rdev->last_read_error) / 3600;
2289 rdev->last_read_error = cur_time_mon;
2292 * if hours_since_last is > the number of bits in read_errors
2293 * just set read errors to 0. We do this to avoid
2294 * overflowing the shift of read_errors by hours_since_last.
2296 if (hours_since_last >= 8 * sizeof(read_errors))
2297 atomic_set(&rdev->read_errors, 0);
2299 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2302 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2303 int sectors, struct page *page, int rw)
2308 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2309 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2311 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2315 set_bit(WriteErrorSeen, &rdev->flags);
2316 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2317 set_bit(MD_RECOVERY_NEEDED,
2318 &rdev->mddev->recovery);
2320 /* need to record an error - either for the block or the device */
2321 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2322 md_error(rdev->mddev, rdev);
2327 * This is a kernel thread which:
2329 * 1. Retries failed read operations on working mirrors.
2330 * 2. Updates the raid superblock when problems encounter.
2331 * 3. Performs writes following reads for array synchronising.
2334 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2336 int sect = 0; /* Offset from r10_bio->sector */
2337 int sectors = r10_bio->sectors;
2338 struct md_rdev*rdev;
2339 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2340 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2342 /* still own a reference to this rdev, so it cannot
2343 * have been cleared recently.
2345 rdev = conf->mirrors[d].rdev;
2347 if (test_bit(Faulty, &rdev->flags))
2348 /* drive has already been failed, just ignore any
2349 more fix_read_error() attempts */
2352 check_decay_read_errors(mddev, rdev);
2353 atomic_inc(&rdev->read_errors);
2354 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2355 char b[BDEVNAME_SIZE];
2356 bdevname(rdev->bdev, b);
2358 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2360 atomic_read(&rdev->read_errors), max_read_errors);
2361 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2363 md_error(mddev, rdev);
2364 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2370 int sl = r10_bio->read_slot;
2374 if (s > (PAGE_SIZE>>9))
2382 d = r10_bio->devs[sl].devnum;
2383 rdev = rcu_dereference(conf->mirrors[d].rdev);
2385 test_bit(In_sync, &rdev->flags) &&
2386 !test_bit(Faulty, &rdev->flags) &&
2387 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2388 &first_bad, &bad_sectors) == 0) {
2389 atomic_inc(&rdev->nr_pending);
2391 success = sync_page_io(rdev,
2392 r10_bio->devs[sl].addr +
2396 REQ_OP_READ, 0, false);
2397 rdev_dec_pending(rdev, mddev);
2403 if (sl == conf->copies)
2405 } while (!success && sl != r10_bio->read_slot);
2409 /* Cannot read from anywhere, just mark the block
2410 * as bad on the first device to discourage future
2413 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2414 rdev = conf->mirrors[dn].rdev;
2416 if (!rdev_set_badblocks(
2418 r10_bio->devs[r10_bio->read_slot].addr
2421 md_error(mddev, rdev);
2422 r10_bio->devs[r10_bio->read_slot].bio
2429 /* write it back and re-read */
2431 while (sl != r10_bio->read_slot) {
2432 char b[BDEVNAME_SIZE];
2437 d = r10_bio->devs[sl].devnum;
2438 rdev = rcu_dereference(conf->mirrors[d].rdev);
2440 test_bit(Faulty, &rdev->flags) ||
2441 !test_bit(In_sync, &rdev->flags))
2444 atomic_inc(&rdev->nr_pending);
2446 if (r10_sync_page_io(rdev,
2447 r10_bio->devs[sl].addr +
2449 s, conf->tmppage, WRITE)
2451 /* Well, this device is dead */
2452 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2454 (unsigned long long)(
2456 choose_data_offset(r10_bio,
2458 bdevname(rdev->bdev, b));
2459 pr_notice("md/raid10:%s: %s: failing drive\n",
2461 bdevname(rdev->bdev, b));
2463 rdev_dec_pending(rdev, mddev);
2467 while (sl != r10_bio->read_slot) {
2468 char b[BDEVNAME_SIZE];
2473 d = r10_bio->devs[sl].devnum;
2474 rdev = rcu_dereference(conf->mirrors[d].rdev);
2476 test_bit(Faulty, &rdev->flags) ||
2477 !test_bit(In_sync, &rdev->flags))
2480 atomic_inc(&rdev->nr_pending);
2482 switch (r10_sync_page_io(rdev,
2483 r10_bio->devs[sl].addr +
2488 /* Well, this device is dead */
2489 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2491 (unsigned long long)(
2493 choose_data_offset(r10_bio, rdev)),
2494 bdevname(rdev->bdev, b));
2495 pr_notice("md/raid10:%s: %s: failing drive\n",
2497 bdevname(rdev->bdev, b));
2500 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2502 (unsigned long long)(
2504 choose_data_offset(r10_bio, rdev)),
2505 bdevname(rdev->bdev, b));
2506 atomic_add(s, &rdev->corrected_errors);
2509 rdev_dec_pending(rdev, mddev);
2519 static int narrow_write_error(struct r10bio *r10_bio, int i)
2521 struct bio *bio = r10_bio->master_bio;
2522 struct mddev *mddev = r10_bio->mddev;
2523 struct r10conf *conf = mddev->private;
2524 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2525 /* bio has the data to be written to slot 'i' where
2526 * we just recently had a write error.
2527 * We repeatedly clone the bio and trim down to one block,
2528 * then try the write. Where the write fails we record
2530 * It is conceivable that the bio doesn't exactly align with
2531 * blocks. We must handle this.
2533 * We currently own a reference to the rdev.
2539 int sect_to_write = r10_bio->sectors;
2542 if (rdev->badblocks.shift < 0)
2545 block_sectors = roundup(1 << rdev->badblocks.shift,
2546 bdev_logical_block_size(rdev->bdev) >> 9);
2547 sector = r10_bio->sector;
2548 sectors = ((r10_bio->sector + block_sectors)
2549 & ~(sector_t)(block_sectors - 1))
2552 while (sect_to_write) {
2555 if (sectors > sect_to_write)
2556 sectors = sect_to_write;
2557 /* Write at 'sector' for 'sectors' */
2558 wbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
2559 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2560 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2561 wbio->bi_iter.bi_sector = wsector +
2562 choose_data_offset(r10_bio, rdev);
2563 wbio->bi_bdev = rdev->bdev;
2564 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2566 if (submit_bio_wait(wbio) < 0)
2568 ok = rdev_set_badblocks(rdev, wsector,
2573 sect_to_write -= sectors;
2575 sectors = block_sectors;
2580 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2582 int slot = r10_bio->read_slot;
2584 struct r10conf *conf = mddev->private;
2585 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2587 sector_t bio_last_sector;
2589 /* we got a read error. Maybe the drive is bad. Maybe just
2590 * the block and we can fix it.
2591 * We freeze all other IO, and try reading the block from
2592 * other devices. When we find one, we re-write
2593 * and check it that fixes the read error.
2594 * This is all done synchronously while the array is
2597 bio = r10_bio->devs[slot].bio;
2598 bio_dev = bio->bi_bdev->bd_dev;
2599 bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
2601 r10_bio->devs[slot].bio = NULL;
2604 r10_bio->devs[slot].bio = IO_BLOCKED;
2605 else if (!test_bit(FailFast, &rdev->flags)) {
2606 freeze_array(conf, 1);
2607 fix_read_error(conf, mddev, r10_bio);
2608 unfreeze_array(conf);
2610 md_error(mddev, rdev);
2612 rdev_dec_pending(rdev, mddev);
2613 allow_barrier(conf);
2615 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2618 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2620 /* Some sort of write request has finished and it
2621 * succeeded in writing where we thought there was a
2622 * bad block. So forget the bad block.
2623 * Or possibly if failed and we need to record
2627 struct md_rdev *rdev;
2629 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2630 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2631 for (m = 0; m < conf->copies; m++) {
2632 int dev = r10_bio->devs[m].devnum;
2633 rdev = conf->mirrors[dev].rdev;
2634 if (r10_bio->devs[m].bio == NULL)
2636 if (!r10_bio->devs[m].bio->bi_error) {
2637 rdev_clear_badblocks(
2639 r10_bio->devs[m].addr,
2640 r10_bio->sectors, 0);
2642 if (!rdev_set_badblocks(
2644 r10_bio->devs[m].addr,
2645 r10_bio->sectors, 0))
2646 md_error(conf->mddev, rdev);
2648 rdev = conf->mirrors[dev].replacement;
2649 if (r10_bio->devs[m].repl_bio == NULL)
2652 if (!r10_bio->devs[m].repl_bio->bi_error) {
2653 rdev_clear_badblocks(
2655 r10_bio->devs[m].addr,
2656 r10_bio->sectors, 0);
2658 if (!rdev_set_badblocks(
2660 r10_bio->devs[m].addr,
2661 r10_bio->sectors, 0))
2662 md_error(conf->mddev, rdev);
2668 for (m = 0; m < conf->copies; m++) {
2669 int dev = r10_bio->devs[m].devnum;
2670 struct bio *bio = r10_bio->devs[m].bio;
2671 rdev = conf->mirrors[dev].rdev;
2672 if (bio == IO_MADE_GOOD) {
2673 rdev_clear_badblocks(
2675 r10_bio->devs[m].addr,
2676 r10_bio->sectors, 0);
2677 rdev_dec_pending(rdev, conf->mddev);
2678 } else if (bio != NULL && bio->bi_error) {
2680 if (!narrow_write_error(r10_bio, m)) {
2681 md_error(conf->mddev, rdev);
2682 set_bit(R10BIO_Degraded,
2685 rdev_dec_pending(rdev, conf->mddev);
2687 bio = r10_bio->devs[m].repl_bio;
2688 rdev = conf->mirrors[dev].replacement;
2689 if (rdev && bio == IO_MADE_GOOD) {
2690 rdev_clear_badblocks(
2692 r10_bio->devs[m].addr,
2693 r10_bio->sectors, 0);
2694 rdev_dec_pending(rdev, conf->mddev);
2698 spin_lock_irq(&conf->device_lock);
2699 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2701 spin_unlock_irq(&conf->device_lock);
2703 * In case freeze_array() is waiting for condition
2704 * nr_pending == nr_queued + extra to be true.
2706 wake_up(&conf->wait_barrier);
2707 md_wakeup_thread(conf->mddev->thread);
2709 if (test_bit(R10BIO_WriteError,
2711 close_write(r10_bio);
2712 raid_end_bio_io(r10_bio);
2717 static void raid10d(struct md_thread *thread)
2719 struct mddev *mddev = thread->mddev;
2720 struct r10bio *r10_bio;
2721 unsigned long flags;
2722 struct r10conf *conf = mddev->private;
2723 struct list_head *head = &conf->retry_list;
2724 struct blk_plug plug;
2726 md_check_recovery(mddev);
2728 if (!list_empty_careful(&conf->bio_end_io_list) &&
2729 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2731 spin_lock_irqsave(&conf->device_lock, flags);
2732 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2733 while (!list_empty(&conf->bio_end_io_list)) {
2734 list_move(conf->bio_end_io_list.prev, &tmp);
2738 spin_unlock_irqrestore(&conf->device_lock, flags);
2739 while (!list_empty(&tmp)) {
2740 r10_bio = list_first_entry(&tmp, struct r10bio,
2742 list_del(&r10_bio->retry_list);
2743 if (mddev->degraded)
2744 set_bit(R10BIO_Degraded, &r10_bio->state);
2746 if (test_bit(R10BIO_WriteError,
2748 close_write(r10_bio);
2749 raid_end_bio_io(r10_bio);
2753 blk_start_plug(&plug);
2756 flush_pending_writes(conf);
2758 spin_lock_irqsave(&conf->device_lock, flags);
2759 if (list_empty(head)) {
2760 spin_unlock_irqrestore(&conf->device_lock, flags);
2763 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2764 list_del(head->prev);
2766 spin_unlock_irqrestore(&conf->device_lock, flags);
2768 mddev = r10_bio->mddev;
2769 conf = mddev->private;
2770 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2771 test_bit(R10BIO_WriteError, &r10_bio->state))
2772 handle_write_completed(conf, r10_bio);
2773 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2774 reshape_request_write(mddev, r10_bio);
2775 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2776 sync_request_write(mddev, r10_bio);
2777 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2778 recovery_request_write(mddev, r10_bio);
2779 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2780 handle_read_error(mddev, r10_bio);
2785 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2786 md_check_recovery(mddev);
2788 blk_finish_plug(&plug);
2791 static int init_resync(struct r10conf *conf)
2796 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2797 BUG_ON(conf->r10buf_pool);
2798 conf->have_replacement = 0;
2799 for (i = 0; i < conf->geo.raid_disks; i++)
2800 if (conf->mirrors[i].replacement)
2801 conf->have_replacement = 1;
2802 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2803 if (!conf->r10buf_pool)
2805 conf->next_resync = 0;
2810 * perform a "sync" on one "block"
2812 * We need to make sure that no normal I/O request - particularly write
2813 * requests - conflict with active sync requests.
2815 * This is achieved by tracking pending requests and a 'barrier' concept
2816 * that can be installed to exclude normal IO requests.
2818 * Resync and recovery are handled very differently.
2819 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2821 * For resync, we iterate over virtual addresses, read all copies,
2822 * and update if there are differences. If only one copy is live,
2824 * For recovery, we iterate over physical addresses, read a good
2825 * value for each non-in_sync drive, and over-write.
2827 * So, for recovery we may have several outstanding complex requests for a
2828 * given address, one for each out-of-sync device. We model this by allocating
2829 * a number of r10_bio structures, one for each out-of-sync device.
2830 * As we setup these structures, we collect all bio's together into a list
2831 * which we then process collectively to add pages, and then process again
2832 * to pass to generic_make_request.
2834 * The r10_bio structures are linked using a borrowed master_bio pointer.
2835 * This link is counted in ->remaining. When the r10_bio that points to NULL
2836 * has its remaining count decremented to 0, the whole complex operation
2841 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2844 struct r10conf *conf = mddev->private;
2845 struct r10bio *r10_bio;
2846 struct bio *biolist = NULL, *bio;
2847 sector_t max_sector, nr_sectors;
2850 sector_t sync_blocks;
2851 sector_t sectors_skipped = 0;
2852 int chunks_skipped = 0;
2853 sector_t chunk_mask = conf->geo.chunk_mask;
2855 if (!conf->r10buf_pool)
2856 if (init_resync(conf))
2860 * Allow skipping a full rebuild for incremental assembly
2861 * of a clean array, like RAID1 does.
2863 if (mddev->bitmap == NULL &&
2864 mddev->recovery_cp == MaxSector &&
2865 mddev->reshape_position == MaxSector &&
2866 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2867 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2868 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2869 conf->fullsync == 0) {
2871 return mddev->dev_sectors - sector_nr;
2875 max_sector = mddev->dev_sectors;
2876 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2877 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2878 max_sector = mddev->resync_max_sectors;
2879 if (sector_nr >= max_sector) {
2880 /* If we aborted, we need to abort the
2881 * sync on the 'current' bitmap chucks (there can
2882 * be several when recovering multiple devices).
2883 * as we may have started syncing it but not finished.
2884 * We can find the current address in
2885 * mddev->curr_resync, but for recovery,
2886 * we need to convert that to several
2887 * virtual addresses.
2889 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2895 if (mddev->curr_resync < max_sector) { /* aborted */
2896 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2897 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2899 else for (i = 0; i < conf->geo.raid_disks; i++) {
2901 raid10_find_virt(conf, mddev->curr_resync, i);
2902 bitmap_end_sync(mddev->bitmap, sect,
2906 /* completed sync */
2907 if ((!mddev->bitmap || conf->fullsync)
2908 && conf->have_replacement
2909 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2910 /* Completed a full sync so the replacements
2911 * are now fully recovered.
2914 for (i = 0; i < conf->geo.raid_disks; i++) {
2915 struct md_rdev *rdev =
2916 rcu_dereference(conf->mirrors[i].replacement);
2918 rdev->recovery_offset = MaxSector;
2924 bitmap_close_sync(mddev->bitmap);
2927 return sectors_skipped;
2930 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2931 return reshape_request(mddev, sector_nr, skipped);
2933 if (chunks_skipped >= conf->geo.raid_disks) {
2934 /* if there has been nothing to do on any drive,
2935 * then there is nothing to do at all..
2938 return (max_sector - sector_nr) + sectors_skipped;
2941 if (max_sector > mddev->resync_max)
2942 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2944 /* make sure whole request will fit in a chunk - if chunks
2947 if (conf->geo.near_copies < conf->geo.raid_disks &&
2948 max_sector > (sector_nr | chunk_mask))
2949 max_sector = (sector_nr | chunk_mask) + 1;
2952 * If there is non-resync activity waiting for a turn, then let it
2953 * though before starting on this new sync request.
2955 if (conf->nr_waiting)
2956 schedule_timeout_uninterruptible(1);
2958 /* Again, very different code for resync and recovery.
2959 * Both must result in an r10bio with a list of bios that
2960 * have bi_end_io, bi_sector, bi_bdev set,
2961 * and bi_private set to the r10bio.
2962 * For recovery, we may actually create several r10bios
2963 * with 2 bios in each, that correspond to the bios in the main one.
2964 * In this case, the subordinate r10bios link back through a
2965 * borrowed master_bio pointer, and the counter in the master
2966 * includes a ref from each subordinate.
2968 /* First, we decide what to do and set ->bi_end_io
2969 * To end_sync_read if we want to read, and
2970 * end_sync_write if we will want to write.
2973 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2974 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2975 /* recovery... the complicated one */
2979 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2985 struct raid10_info *mirror = &conf->mirrors[i];
2986 struct md_rdev *mrdev, *mreplace;
2989 mrdev = rcu_dereference(mirror->rdev);
2990 mreplace = rcu_dereference(mirror->replacement);
2992 if ((mrdev == NULL ||
2993 test_bit(Faulty, &mrdev->flags) ||
2994 test_bit(In_sync, &mrdev->flags)) &&
2995 (mreplace == NULL ||
2996 test_bit(Faulty, &mreplace->flags))) {
3002 /* want to reconstruct this device */
3004 sect = raid10_find_virt(conf, sector_nr, i);
3005 if (sect >= mddev->resync_max_sectors) {
3006 /* last stripe is not complete - don't
3007 * try to recover this sector.
3012 if (mreplace && test_bit(Faulty, &mreplace->flags))
3014 /* Unless we are doing a full sync, or a replacement
3015 * we only need to recover the block if it is set in
3018 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3020 if (sync_blocks < max_sync)
3021 max_sync = sync_blocks;
3025 /* yep, skip the sync_blocks here, but don't assume
3026 * that there will never be anything to do here
3028 chunks_skipped = -1;
3032 atomic_inc(&mrdev->nr_pending);
3034 atomic_inc(&mreplace->nr_pending);
3037 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3039 raise_barrier(conf, rb2 != NULL);
3040 atomic_set(&r10_bio->remaining, 0);
3042 r10_bio->master_bio = (struct bio*)rb2;
3044 atomic_inc(&rb2->remaining);
3045 r10_bio->mddev = mddev;
3046 set_bit(R10BIO_IsRecover, &r10_bio->state);
3047 r10_bio->sector = sect;
3049 raid10_find_phys(conf, r10_bio);
3051 /* Need to check if the array will still be
3055 for (j = 0; j < conf->geo.raid_disks; j++) {
3056 struct md_rdev *rdev = rcu_dereference(
3057 conf->mirrors[j].rdev);
3058 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3064 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3065 &sync_blocks, still_degraded);
3068 for (j=0; j<conf->copies;j++) {
3070 int d = r10_bio->devs[j].devnum;
3071 sector_t from_addr, to_addr;
3072 struct md_rdev *rdev =
3073 rcu_dereference(conf->mirrors[d].rdev);
3074 sector_t sector, first_bad;
3077 !test_bit(In_sync, &rdev->flags))
3079 /* This is where we read from */
3081 sector = r10_bio->devs[j].addr;
3083 if (is_badblock(rdev, sector, max_sync,
3084 &first_bad, &bad_sectors)) {
3085 if (first_bad > sector)
3086 max_sync = first_bad - sector;
3088 bad_sectors -= (sector
3090 if (max_sync > bad_sectors)
3091 max_sync = bad_sectors;
3095 bio = r10_bio->devs[0].bio;
3096 bio->bi_next = biolist;
3098 bio->bi_end_io = end_sync_read;
3099 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3100 if (test_bit(FailFast, &rdev->flags))
3101 bio->bi_opf |= MD_FAILFAST;
3102 from_addr = r10_bio->devs[j].addr;
3103 bio->bi_iter.bi_sector = from_addr +
3105 bio->bi_bdev = rdev->bdev;
3106 atomic_inc(&rdev->nr_pending);
3107 /* and we write to 'i' (if not in_sync) */
3109 for (k=0; k<conf->copies; k++)
3110 if (r10_bio->devs[k].devnum == i)
3112 BUG_ON(k == conf->copies);
3113 to_addr = r10_bio->devs[k].addr;
3114 r10_bio->devs[0].devnum = d;
3115 r10_bio->devs[0].addr = from_addr;
3116 r10_bio->devs[1].devnum = i;
3117 r10_bio->devs[1].addr = to_addr;
3119 if (!test_bit(In_sync, &mrdev->flags)) {
3120 bio = r10_bio->devs[1].bio;
3121 bio->bi_next = biolist;
3123 bio->bi_end_io = end_sync_write;
3124 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3125 bio->bi_iter.bi_sector = to_addr
3126 + mrdev->data_offset;
3127 bio->bi_bdev = mrdev->bdev;
3128 atomic_inc(&r10_bio->remaining);
3130 r10_bio->devs[1].bio->bi_end_io = NULL;
3132 /* and maybe write to replacement */
3133 bio = r10_bio->devs[1].repl_bio;
3135 bio->bi_end_io = NULL;
3136 /* Note: if mreplace != NULL, then bio
3137 * cannot be NULL as r10buf_pool_alloc will
3138 * have allocated it.
3139 * So the second test here is pointless.
3140 * But it keeps semantic-checkers happy, and
3141 * this comment keeps human reviewers
3144 if (mreplace == NULL || bio == NULL ||
3145 test_bit(Faulty, &mreplace->flags))
3147 bio->bi_next = biolist;
3149 bio->bi_end_io = end_sync_write;
3150 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3151 bio->bi_iter.bi_sector = to_addr +
3152 mreplace->data_offset;
3153 bio->bi_bdev = mreplace->bdev;
3154 atomic_inc(&r10_bio->remaining);
3158 if (j == conf->copies) {
3159 /* Cannot recover, so abort the recovery or
3160 * record a bad block */
3162 /* problem is that there are bad blocks
3163 * on other device(s)
3166 for (k = 0; k < conf->copies; k++)
3167 if (r10_bio->devs[k].devnum == i)
3169 if (!test_bit(In_sync,
3171 && !rdev_set_badblocks(
3173 r10_bio->devs[k].addr,
3177 !rdev_set_badblocks(
3179 r10_bio->devs[k].addr,
3184 if (!test_and_set_bit(MD_RECOVERY_INTR,
3186 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3188 mirror->recovery_disabled
3189 = mddev->recovery_disabled;
3193 atomic_dec(&rb2->remaining);
3195 rdev_dec_pending(mrdev, mddev);
3197 rdev_dec_pending(mreplace, mddev);
3200 rdev_dec_pending(mrdev, mddev);
3202 rdev_dec_pending(mreplace, mddev);
3203 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3204 /* Only want this if there is elsewhere to
3205 * read from. 'j' is currently the first
3209 for (; j < conf->copies; j++) {
3210 int d = r10_bio->devs[j].devnum;
3211 if (conf->mirrors[d].rdev &&
3213 &conf->mirrors[d].rdev->flags))
3217 r10_bio->devs[0].bio->bi_opf
3221 if (biolist == NULL) {
3223 struct r10bio *rb2 = r10_bio;
3224 r10_bio = (struct r10bio*) rb2->master_bio;
3225 rb2->master_bio = NULL;
3231 /* resync. Schedule a read for every block at this virt offset */
3234 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3236 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3237 &sync_blocks, mddev->degraded) &&
3238 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3239 &mddev->recovery)) {
3240 /* We can skip this block */
3242 return sync_blocks + sectors_skipped;
3244 if (sync_blocks < max_sync)
3245 max_sync = sync_blocks;
3246 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3249 r10_bio->mddev = mddev;
3250 atomic_set(&r10_bio->remaining, 0);
3251 raise_barrier(conf, 0);
3252 conf->next_resync = sector_nr;
3254 r10_bio->master_bio = NULL;
3255 r10_bio->sector = sector_nr;
3256 set_bit(R10BIO_IsSync, &r10_bio->state);
3257 raid10_find_phys(conf, r10_bio);
3258 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3260 for (i = 0; i < conf->copies; i++) {
3261 int d = r10_bio->devs[i].devnum;
3262 sector_t first_bad, sector;
3264 struct md_rdev *rdev;
3266 if (r10_bio->devs[i].repl_bio)
3267 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3269 bio = r10_bio->devs[i].bio;
3270 bio->bi_error = -EIO;
3272 rdev = rcu_dereference(conf->mirrors[d].rdev);
3273 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3277 sector = r10_bio->devs[i].addr;
3278 if (is_badblock(rdev, sector, max_sync,
3279 &first_bad, &bad_sectors)) {
3280 if (first_bad > sector)
3281 max_sync = first_bad - sector;
3283 bad_sectors -= (sector - first_bad);
3284 if (max_sync > bad_sectors)
3285 max_sync = bad_sectors;
3290 atomic_inc(&rdev->nr_pending);
3291 atomic_inc(&r10_bio->remaining);
3292 bio->bi_next = biolist;
3294 bio->bi_end_io = end_sync_read;
3295 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3296 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3297 bio->bi_opf |= MD_FAILFAST;
3298 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3299 bio->bi_bdev = rdev->bdev;
3302 rdev = rcu_dereference(conf->mirrors[d].replacement);
3303 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3307 atomic_inc(&rdev->nr_pending);
3310 /* Need to set up for writing to the replacement */
3311 bio = r10_bio->devs[i].repl_bio;
3312 bio->bi_error = -EIO;
3314 sector = r10_bio->devs[i].addr;
3315 bio->bi_next = biolist;
3317 bio->bi_end_io = end_sync_write;
3318 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3319 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3320 bio->bi_opf |= MD_FAILFAST;
3321 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3322 bio->bi_bdev = rdev->bdev;
3327 for (i=0; i<conf->copies; i++) {
3328 int d = r10_bio->devs[i].devnum;
3329 if (r10_bio->devs[i].bio->bi_end_io)
3330 rdev_dec_pending(conf->mirrors[d].rdev,
3332 if (r10_bio->devs[i].repl_bio &&
3333 r10_bio->devs[i].repl_bio->bi_end_io)
3335 conf->mirrors[d].replacement,
3345 if (sector_nr + max_sync < max_sector)
3346 max_sector = sector_nr + max_sync;
3349 int len = PAGE_SIZE;
3350 if (sector_nr + (len>>9) > max_sector)
3351 len = (max_sector - sector_nr) << 9;
3354 for (bio= biolist ; bio ; bio=bio->bi_next) {
3355 struct resync_pages *rp = get_resync_pages(bio);
3356 page = resync_fetch_page(rp, rp->idx++);
3358 * won't fail because the vec table is big enough
3359 * to hold all these pages
3361 bio_add_page(bio, page, len, 0);
3363 nr_sectors += len>>9;
3364 sector_nr += len>>9;
3365 } while (get_resync_pages(biolist)->idx < RESYNC_PAGES);
3366 r10_bio->sectors = nr_sectors;
3370 biolist = biolist->bi_next;
3372 bio->bi_next = NULL;
3373 r10_bio = get_resync_r10bio(bio);
3374 r10_bio->sectors = nr_sectors;
3376 if (bio->bi_end_io == end_sync_read) {
3377 md_sync_acct(bio->bi_bdev, nr_sectors);
3379 generic_make_request(bio);
3383 if (sectors_skipped)
3384 /* pretend they weren't skipped, it makes
3385 * no important difference in this case
3387 md_done_sync(mddev, sectors_skipped, 1);
3389 return sectors_skipped + nr_sectors;
3391 /* There is nowhere to write, so all non-sync
3392 * drives must be failed or in resync, all drives
3393 * have a bad block, so try the next chunk...
3395 if (sector_nr + max_sync < max_sector)
3396 max_sector = sector_nr + max_sync;
3398 sectors_skipped += (max_sector - sector_nr);
3400 sector_nr = max_sector;
3405 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3408 struct r10conf *conf = mddev->private;
3411 raid_disks = min(conf->geo.raid_disks,
3412 conf->prev.raid_disks);
3414 sectors = conf->dev_sectors;
3416 size = sectors >> conf->geo.chunk_shift;
3417 sector_div(size, conf->geo.far_copies);
3418 size = size * raid_disks;
3419 sector_div(size, conf->geo.near_copies);
3421 return size << conf->geo.chunk_shift;
3424 static void calc_sectors(struct r10conf *conf, sector_t size)
3426 /* Calculate the number of sectors-per-device that will
3427 * actually be used, and set conf->dev_sectors and
3431 size = size >> conf->geo.chunk_shift;
3432 sector_div(size, conf->geo.far_copies);
3433 size = size * conf->geo.raid_disks;
3434 sector_div(size, conf->geo.near_copies);
3435 /* 'size' is now the number of chunks in the array */
3436 /* calculate "used chunks per device" */
3437 size = size * conf->copies;
3439 /* We need to round up when dividing by raid_disks to
3440 * get the stride size.
3442 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3444 conf->dev_sectors = size << conf->geo.chunk_shift;
3446 if (conf->geo.far_offset)
3447 conf->geo.stride = 1 << conf->geo.chunk_shift;
3449 sector_div(size, conf->geo.far_copies);
3450 conf->geo.stride = size << conf->geo.chunk_shift;
3454 enum geo_type {geo_new, geo_old, geo_start};
3455 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3458 int layout, chunk, disks;
3461 layout = mddev->layout;
3462 chunk = mddev->chunk_sectors;
3463 disks = mddev->raid_disks - mddev->delta_disks;
3466 layout = mddev->new_layout;
3467 chunk = mddev->new_chunk_sectors;
3468 disks = mddev->raid_disks;
3470 default: /* avoid 'may be unused' warnings */
3471 case geo_start: /* new when starting reshape - raid_disks not
3473 layout = mddev->new_layout;
3474 chunk = mddev->new_chunk_sectors;
3475 disks = mddev->raid_disks + mddev->delta_disks;
3480 if (chunk < (PAGE_SIZE >> 9) ||
3481 !is_power_of_2(chunk))
3484 fc = (layout >> 8) & 255;
3485 fo = layout & (1<<16);
3486 geo->raid_disks = disks;
3487 geo->near_copies = nc;
3488 geo->far_copies = fc;
3489 geo->far_offset = fo;
3490 switch (layout >> 17) {
3491 case 0: /* original layout. simple but not always optimal */
3492 geo->far_set_size = disks;
3494 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3495 * actually using this, but leave code here just in case.*/
3496 geo->far_set_size = disks/fc;
3497 WARN(geo->far_set_size < fc,
3498 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3500 case 2: /* "improved" layout fixed to match documentation */
3501 geo->far_set_size = fc * nc;
3503 default: /* Not a valid layout */
3506 geo->chunk_mask = chunk - 1;
3507 geo->chunk_shift = ffz(~chunk);
3511 static struct r10conf *setup_conf(struct mddev *mddev)
3513 struct r10conf *conf = NULL;
3518 copies = setup_geo(&geo, mddev, geo_new);
3521 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3522 mdname(mddev), PAGE_SIZE);
3526 if (copies < 2 || copies > mddev->raid_disks) {
3527 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3528 mdname(mddev), mddev->new_layout);
3533 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3537 /* FIXME calc properly */
3538 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3539 max(0,-mddev->delta_disks)),
3544 conf->tmppage = alloc_page(GFP_KERNEL);
3549 conf->copies = copies;
3550 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3551 r10bio_pool_free, conf);
3552 if (!conf->r10bio_pool)
3555 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0);
3556 if (!conf->bio_split)
3559 calc_sectors(conf, mddev->dev_sectors);
3560 if (mddev->reshape_position == MaxSector) {
3561 conf->prev = conf->geo;
3562 conf->reshape_progress = MaxSector;
3564 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3568 conf->reshape_progress = mddev->reshape_position;
3569 if (conf->prev.far_offset)
3570 conf->prev.stride = 1 << conf->prev.chunk_shift;
3572 /* far_copies must be 1 */
3573 conf->prev.stride = conf->dev_sectors;
3575 conf->reshape_safe = conf->reshape_progress;
3576 spin_lock_init(&conf->device_lock);
3577 INIT_LIST_HEAD(&conf->retry_list);
3578 INIT_LIST_HEAD(&conf->bio_end_io_list);
3580 spin_lock_init(&conf->resync_lock);
3581 init_waitqueue_head(&conf->wait_barrier);
3582 atomic_set(&conf->nr_pending, 0);
3584 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3588 conf->mddev = mddev;
3593 mempool_destroy(conf->r10bio_pool);
3594 kfree(conf->mirrors);
3595 safe_put_page(conf->tmppage);
3596 if (conf->bio_split)
3597 bioset_free(conf->bio_split);
3600 return ERR_PTR(err);
3603 static int raid10_run(struct mddev *mddev)
3605 struct r10conf *conf;
3606 int i, disk_idx, chunk_size;
3607 struct raid10_info *disk;
3608 struct md_rdev *rdev;
3610 sector_t min_offset_diff = 0;
3612 bool discard_supported = false;
3614 if (mddev_init_writes_pending(mddev) < 0)
3617 if (mddev->private == NULL) {
3618 conf = setup_conf(mddev);
3620 return PTR_ERR(conf);
3621 mddev->private = conf;
3623 conf = mddev->private;
3627 mddev->thread = conf->thread;
3628 conf->thread = NULL;
3630 chunk_size = mddev->chunk_sectors << 9;
3632 blk_queue_max_discard_sectors(mddev->queue,
3633 mddev->chunk_sectors);
3634 blk_queue_max_write_same_sectors(mddev->queue, 0);
3635 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3636 blk_queue_io_min(mddev->queue, chunk_size);
3637 if (conf->geo.raid_disks % conf->geo.near_copies)
3638 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3640 blk_queue_io_opt(mddev->queue, chunk_size *
3641 (conf->geo.raid_disks / conf->geo.near_copies));
3644 rdev_for_each(rdev, mddev) {
3647 disk_idx = rdev->raid_disk;
3650 if (disk_idx >= conf->geo.raid_disks &&
3651 disk_idx >= conf->prev.raid_disks)
3653 disk = conf->mirrors + disk_idx;
3655 if (test_bit(Replacement, &rdev->flags)) {
3656 if (disk->replacement)
3658 disk->replacement = rdev;
3664 diff = (rdev->new_data_offset - rdev->data_offset);
3665 if (!mddev->reshape_backwards)
3669 if (first || diff < min_offset_diff)
3670 min_offset_diff = diff;
3673 disk_stack_limits(mddev->gendisk, rdev->bdev,
3674 rdev->data_offset << 9);
3676 disk->head_position = 0;
3678 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3679 discard_supported = true;
3684 if (discard_supported)
3685 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3688 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3691 /* need to check that every block has at least one working mirror */
3692 if (!enough(conf, -1)) {
3693 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3698 if (conf->reshape_progress != MaxSector) {
3699 /* must ensure that shape change is supported */
3700 if (conf->geo.far_copies != 1 &&
3701 conf->geo.far_offset == 0)
3703 if (conf->prev.far_copies != 1 &&
3704 conf->prev.far_offset == 0)
3708 mddev->degraded = 0;
3710 i < conf->geo.raid_disks
3711 || i < conf->prev.raid_disks;
3714 disk = conf->mirrors + i;
3716 if (!disk->rdev && disk->replacement) {
3717 /* The replacement is all we have - use it */
3718 disk->rdev = disk->replacement;
3719 disk->replacement = NULL;
3720 clear_bit(Replacement, &disk->rdev->flags);
3724 !test_bit(In_sync, &disk->rdev->flags)) {
3725 disk->head_position = 0;
3728 disk->rdev->saved_raid_disk < 0)
3731 disk->recovery_disabled = mddev->recovery_disabled - 1;
3734 if (mddev->recovery_cp != MaxSector)
3735 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3737 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3738 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3739 conf->geo.raid_disks);
3741 * Ok, everything is just fine now
3743 mddev->dev_sectors = conf->dev_sectors;
3744 size = raid10_size(mddev, 0, 0);
3745 md_set_array_sectors(mddev, size);
3746 mddev->resync_max_sectors = size;
3747 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3750 int stripe = conf->geo.raid_disks *
3751 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3753 /* Calculate max read-ahead size.
3754 * We need to readahead at least twice a whole stripe....
3757 stripe /= conf->geo.near_copies;
3758 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3759 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3762 if (md_integrity_register(mddev))
3765 if (conf->reshape_progress != MaxSector) {
3766 unsigned long before_length, after_length;
3768 before_length = ((1 << conf->prev.chunk_shift) *
3769 conf->prev.far_copies);
3770 after_length = ((1 << conf->geo.chunk_shift) *
3771 conf->geo.far_copies);
3773 if (max(before_length, after_length) > min_offset_diff) {
3774 /* This cannot work */
3775 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3778 conf->offset_diff = min_offset_diff;
3780 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3781 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3782 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3783 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3784 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3791 md_unregister_thread(&mddev->thread);
3792 mempool_destroy(conf->r10bio_pool);
3793 safe_put_page(conf->tmppage);
3794 kfree(conf->mirrors);
3796 mddev->private = NULL;
3801 static void raid10_free(struct mddev *mddev, void *priv)
3803 struct r10conf *conf = priv;
3805 mempool_destroy(conf->r10bio_pool);
3806 safe_put_page(conf->tmppage);
3807 kfree(conf->mirrors);
3808 kfree(conf->mirrors_old);
3809 kfree(conf->mirrors_new);
3810 if (conf->bio_split)
3811 bioset_free(conf->bio_split);
3815 static void raid10_quiesce(struct mddev *mddev, int state)
3817 struct r10conf *conf = mddev->private;
3821 raise_barrier(conf, 0);
3824 lower_barrier(conf);
3829 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3831 /* Resize of 'far' arrays is not supported.
3832 * For 'near' and 'offset' arrays we can set the
3833 * number of sectors used to be an appropriate multiple
3834 * of the chunk size.
3835 * For 'offset', this is far_copies*chunksize.
3836 * For 'near' the multiplier is the LCM of
3837 * near_copies and raid_disks.
3838 * So if far_copies > 1 && !far_offset, fail.
3839 * Else find LCM(raid_disks, near_copy)*far_copies and
3840 * multiply by chunk_size. Then round to this number.
3841 * This is mostly done by raid10_size()
3843 struct r10conf *conf = mddev->private;
3844 sector_t oldsize, size;
3846 if (mddev->reshape_position != MaxSector)
3849 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3852 oldsize = raid10_size(mddev, 0, 0);
3853 size = raid10_size(mddev, sectors, 0);
3854 if (mddev->external_size &&
3855 mddev->array_sectors > size)
3857 if (mddev->bitmap) {
3858 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3862 md_set_array_sectors(mddev, size);
3863 if (sectors > mddev->dev_sectors &&
3864 mddev->recovery_cp > oldsize) {
3865 mddev->recovery_cp = oldsize;
3866 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3868 calc_sectors(conf, sectors);
3869 mddev->dev_sectors = conf->dev_sectors;
3870 mddev->resync_max_sectors = size;
3874 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3876 struct md_rdev *rdev;
3877 struct r10conf *conf;
3879 if (mddev->degraded > 0) {
3880 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3882 return ERR_PTR(-EINVAL);
3884 sector_div(size, devs);
3886 /* Set new parameters */
3887 mddev->new_level = 10;
3888 /* new layout: far_copies = 1, near_copies = 2 */
3889 mddev->new_layout = (1<<8) + 2;
3890 mddev->new_chunk_sectors = mddev->chunk_sectors;
3891 mddev->delta_disks = mddev->raid_disks;
3892 mddev->raid_disks *= 2;
3893 /* make sure it will be not marked as dirty */
3894 mddev->recovery_cp = MaxSector;
3895 mddev->dev_sectors = size;
3897 conf = setup_conf(mddev);
3898 if (!IS_ERR(conf)) {
3899 rdev_for_each(rdev, mddev)
3900 if (rdev->raid_disk >= 0) {
3901 rdev->new_raid_disk = rdev->raid_disk * 2;
3902 rdev->sectors = size;
3910 static void *raid10_takeover(struct mddev *mddev)
3912 struct r0conf *raid0_conf;
3914 /* raid10 can take over:
3915 * raid0 - providing it has only two drives
3917 if (mddev->level == 0) {
3918 /* for raid0 takeover only one zone is supported */
3919 raid0_conf = mddev->private;
3920 if (raid0_conf->nr_strip_zones > 1) {
3921 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
3923 return ERR_PTR(-EINVAL);
3925 return raid10_takeover_raid0(mddev,
3926 raid0_conf->strip_zone->zone_end,
3927 raid0_conf->strip_zone->nb_dev);
3929 return ERR_PTR(-EINVAL);
3932 static int raid10_check_reshape(struct mddev *mddev)
3934 /* Called when there is a request to change
3935 * - layout (to ->new_layout)
3936 * - chunk size (to ->new_chunk_sectors)
3937 * - raid_disks (by delta_disks)
3938 * or when trying to restart a reshape that was ongoing.
3940 * We need to validate the request and possibly allocate
3941 * space if that might be an issue later.
3943 * Currently we reject any reshape of a 'far' mode array,
3944 * allow chunk size to change if new is generally acceptable,
3945 * allow raid_disks to increase, and allow
3946 * a switch between 'near' mode and 'offset' mode.
3948 struct r10conf *conf = mddev->private;
3951 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3954 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3955 /* mustn't change number of copies */
3957 if (geo.far_copies > 1 && !geo.far_offset)
3958 /* Cannot switch to 'far' mode */
3961 if (mddev->array_sectors & geo.chunk_mask)
3962 /* not factor of array size */
3965 if (!enough(conf, -1))
3968 kfree(conf->mirrors_new);
3969 conf->mirrors_new = NULL;
3970 if (mddev->delta_disks > 0) {
3971 /* allocate new 'mirrors' list */
3972 conf->mirrors_new = kzalloc(
3973 sizeof(struct raid10_info)
3974 *(mddev->raid_disks +
3975 mddev->delta_disks),
3977 if (!conf->mirrors_new)
3984 * Need to check if array has failed when deciding whether to:
3986 * - remove non-faulty devices
3989 * This determination is simple when no reshape is happening.
3990 * However if there is a reshape, we need to carefully check
3991 * both the before and after sections.
3992 * This is because some failed devices may only affect one
3993 * of the two sections, and some non-in_sync devices may
3994 * be insync in the section most affected by failed devices.
3996 static int calc_degraded(struct r10conf *conf)
3998 int degraded, degraded2;
4003 /* 'prev' section first */
4004 for (i = 0; i < conf->prev.raid_disks; i++) {
4005 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4006 if (!rdev || test_bit(Faulty, &rdev->flags))
4008 else if (!test_bit(In_sync, &rdev->flags))
4009 /* When we can reduce the number of devices in
4010 * an array, this might not contribute to
4011 * 'degraded'. It does now.
4016 if (conf->geo.raid_disks == conf->prev.raid_disks)
4020 for (i = 0; i < conf->geo.raid_disks; i++) {
4021 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4022 if (!rdev || test_bit(Faulty, &rdev->flags))
4024 else if (!test_bit(In_sync, &rdev->flags)) {
4025 /* If reshape is increasing the number of devices,
4026 * this section has already been recovered, so
4027 * it doesn't contribute to degraded.
4030 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4035 if (degraded2 > degraded)
4040 static int raid10_start_reshape(struct mddev *mddev)
4042 /* A 'reshape' has been requested. This commits
4043 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4044 * This also checks if there are enough spares and adds them
4046 * We currently require enough spares to make the final
4047 * array non-degraded. We also require that the difference
4048 * between old and new data_offset - on each device - is
4049 * enough that we never risk over-writing.
4052 unsigned long before_length, after_length;
4053 sector_t min_offset_diff = 0;
4056 struct r10conf *conf = mddev->private;
4057 struct md_rdev *rdev;
4061 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4064 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4067 before_length = ((1 << conf->prev.chunk_shift) *
4068 conf->prev.far_copies);
4069 after_length = ((1 << conf->geo.chunk_shift) *
4070 conf->geo.far_copies);
4072 rdev_for_each(rdev, mddev) {
4073 if (!test_bit(In_sync, &rdev->flags)
4074 && !test_bit(Faulty, &rdev->flags))
4076 if (rdev->raid_disk >= 0) {
4077 long long diff = (rdev->new_data_offset
4078 - rdev->data_offset);
4079 if (!mddev->reshape_backwards)
4083 if (first || diff < min_offset_diff)
4084 min_offset_diff = diff;
4089 if (max(before_length, after_length) > min_offset_diff)
4092 if (spares < mddev->delta_disks)
4095 conf->offset_diff = min_offset_diff;
4096 spin_lock_irq(&conf->device_lock);
4097 if (conf->mirrors_new) {
4098 memcpy(conf->mirrors_new, conf->mirrors,
4099 sizeof(struct raid10_info)*conf->prev.raid_disks);
4101 kfree(conf->mirrors_old);
4102 conf->mirrors_old = conf->mirrors;
4103 conf->mirrors = conf->mirrors_new;
4104 conf->mirrors_new = NULL;
4106 setup_geo(&conf->geo, mddev, geo_start);
4108 if (mddev->reshape_backwards) {
4109 sector_t size = raid10_size(mddev, 0, 0);
4110 if (size < mddev->array_sectors) {
4111 spin_unlock_irq(&conf->device_lock);
4112 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4116 mddev->resync_max_sectors = size;
4117 conf->reshape_progress = size;
4119 conf->reshape_progress = 0;
4120 conf->reshape_safe = conf->reshape_progress;
4121 spin_unlock_irq(&conf->device_lock);
4123 if (mddev->delta_disks && mddev->bitmap) {
4124 ret = bitmap_resize(mddev->bitmap,
4125 raid10_size(mddev, 0,
4126 conf->geo.raid_disks),
4131 if (mddev->delta_disks > 0) {
4132 rdev_for_each(rdev, mddev)
4133 if (rdev->raid_disk < 0 &&
4134 !test_bit(Faulty, &rdev->flags)) {
4135 if (raid10_add_disk(mddev, rdev) == 0) {
4136 if (rdev->raid_disk >=
4137 conf->prev.raid_disks)
4138 set_bit(In_sync, &rdev->flags);
4140 rdev->recovery_offset = 0;
4142 if (sysfs_link_rdev(mddev, rdev))
4143 /* Failure here is OK */;
4145 } else if (rdev->raid_disk >= conf->prev.raid_disks
4146 && !test_bit(Faulty, &rdev->flags)) {
4147 /* This is a spare that was manually added */
4148 set_bit(In_sync, &rdev->flags);
4151 /* When a reshape changes the number of devices,
4152 * ->degraded is measured against the larger of the
4153 * pre and post numbers.
4155 spin_lock_irq(&conf->device_lock);
4156 mddev->degraded = calc_degraded(conf);
4157 spin_unlock_irq(&conf->device_lock);
4158 mddev->raid_disks = conf->geo.raid_disks;
4159 mddev->reshape_position = conf->reshape_progress;
4160 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4162 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4163 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4164 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4165 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4166 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4168 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4170 if (!mddev->sync_thread) {
4174 conf->reshape_checkpoint = jiffies;
4175 md_wakeup_thread(mddev->sync_thread);
4176 md_new_event(mddev);
4180 mddev->recovery = 0;
4181 spin_lock_irq(&conf->device_lock);
4182 conf->geo = conf->prev;
4183 mddev->raid_disks = conf->geo.raid_disks;
4184 rdev_for_each(rdev, mddev)
4185 rdev->new_data_offset = rdev->data_offset;
4187 conf->reshape_progress = MaxSector;
4188 conf->reshape_safe = MaxSector;
4189 mddev->reshape_position = MaxSector;
4190 spin_unlock_irq(&conf->device_lock);
4194 /* Calculate the last device-address that could contain
4195 * any block from the chunk that includes the array-address 's'
4196 * and report the next address.
4197 * i.e. the address returned will be chunk-aligned and after
4198 * any data that is in the chunk containing 's'.
4200 static sector_t last_dev_address(sector_t s, struct geom *geo)
4202 s = (s | geo->chunk_mask) + 1;
4203 s >>= geo->chunk_shift;
4204 s *= geo->near_copies;
4205 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4206 s *= geo->far_copies;
4207 s <<= geo->chunk_shift;
4211 /* Calculate the first device-address that could contain
4212 * any block from the chunk that includes the array-address 's'.
4213 * This too will be the start of a chunk
4215 static sector_t first_dev_address(sector_t s, struct geom *geo)
4217 s >>= geo->chunk_shift;
4218 s *= geo->near_copies;
4219 sector_div(s, geo->raid_disks);
4220 s *= geo->far_copies;
4221 s <<= geo->chunk_shift;
4225 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4228 /* We simply copy at most one chunk (smallest of old and new)
4229 * at a time, possibly less if that exceeds RESYNC_PAGES,
4230 * or we hit a bad block or something.
4231 * This might mean we pause for normal IO in the middle of
4232 * a chunk, but that is not a problem as mddev->reshape_position
4233 * can record any location.
4235 * If we will want to write to a location that isn't
4236 * yet recorded as 'safe' (i.e. in metadata on disk) then
4237 * we need to flush all reshape requests and update the metadata.
4239 * When reshaping forwards (e.g. to more devices), we interpret
4240 * 'safe' as the earliest block which might not have been copied
4241 * down yet. We divide this by previous stripe size and multiply
4242 * by previous stripe length to get lowest device offset that we
4243 * cannot write to yet.
4244 * We interpret 'sector_nr' as an address that we want to write to.
4245 * From this we use last_device_address() to find where we might
4246 * write to, and first_device_address on the 'safe' position.
4247 * If this 'next' write position is after the 'safe' position,
4248 * we must update the metadata to increase the 'safe' position.
4250 * When reshaping backwards, we round in the opposite direction
4251 * and perform the reverse test: next write position must not be
4252 * less than current safe position.
4254 * In all this the minimum difference in data offsets
4255 * (conf->offset_diff - always positive) allows a bit of slack,
4256 * so next can be after 'safe', but not by more than offset_diff
4258 * We need to prepare all the bios here before we start any IO
4259 * to ensure the size we choose is acceptable to all devices.
4260 * The means one for each copy for write-out and an extra one for
4262 * We store the read-in bio in ->master_bio and the others in
4263 * ->devs[x].bio and ->devs[x].repl_bio.
4265 struct r10conf *conf = mddev->private;
4266 struct r10bio *r10_bio;
4267 sector_t next, safe, last;
4271 struct md_rdev *rdev;
4274 struct bio *bio, *read_bio;
4275 int sectors_done = 0;
4276 struct page **pages;
4278 if (sector_nr == 0) {
4279 /* If restarting in the middle, skip the initial sectors */
4280 if (mddev->reshape_backwards &&
4281 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4282 sector_nr = (raid10_size(mddev, 0, 0)
4283 - conf->reshape_progress);
4284 } else if (!mddev->reshape_backwards &&
4285 conf->reshape_progress > 0)
4286 sector_nr = conf->reshape_progress;
4288 mddev->curr_resync_completed = sector_nr;
4289 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4295 /* We don't use sector_nr to track where we are up to
4296 * as that doesn't work well for ->reshape_backwards.
4297 * So just use ->reshape_progress.
4299 if (mddev->reshape_backwards) {
4300 /* 'next' is the earliest device address that we might
4301 * write to for this chunk in the new layout
4303 next = first_dev_address(conf->reshape_progress - 1,
4306 /* 'safe' is the last device address that we might read from
4307 * in the old layout after a restart
4309 safe = last_dev_address(conf->reshape_safe - 1,
4312 if (next + conf->offset_diff < safe)
4315 last = conf->reshape_progress - 1;
4316 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4317 & conf->prev.chunk_mask);
4318 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4319 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4321 /* 'next' is after the last device address that we
4322 * might write to for this chunk in the new layout
4324 next = last_dev_address(conf->reshape_progress, &conf->geo);
4326 /* 'safe' is the earliest device address that we might
4327 * read from in the old layout after a restart
4329 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4331 /* Need to update metadata if 'next' might be beyond 'safe'
4332 * as that would possibly corrupt data
4334 if (next > safe + conf->offset_diff)
4337 sector_nr = conf->reshape_progress;
4338 last = sector_nr | (conf->geo.chunk_mask
4339 & conf->prev.chunk_mask);
4341 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4342 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4346 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4347 /* Need to update reshape_position in metadata */
4349 mddev->reshape_position = conf->reshape_progress;
4350 if (mddev->reshape_backwards)
4351 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4352 - conf->reshape_progress;
4354 mddev->curr_resync_completed = conf->reshape_progress;
4355 conf->reshape_checkpoint = jiffies;
4356 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4357 md_wakeup_thread(mddev->thread);
4358 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4359 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4360 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4361 allow_barrier(conf);
4362 return sectors_done;
4364 conf->reshape_safe = mddev->reshape_position;
4365 allow_barrier(conf);
4369 /* Now schedule reads for blocks from sector_nr to last */
4370 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4372 raise_barrier(conf, sectors_done != 0);
4373 atomic_set(&r10_bio->remaining, 0);
4374 r10_bio->mddev = mddev;
4375 r10_bio->sector = sector_nr;
4376 set_bit(R10BIO_IsReshape, &r10_bio->state);
4377 r10_bio->sectors = last - sector_nr + 1;
4378 rdev = read_balance(conf, r10_bio, &max_sectors);
4379 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4382 /* Cannot read from here, so need to record bad blocks
4383 * on all the target devices.
4386 mempool_free(r10_bio, conf->r10buf_pool);
4387 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4388 return sectors_done;
4391 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4393 read_bio->bi_bdev = rdev->bdev;
4394 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4395 + rdev->data_offset);
4396 read_bio->bi_private = r10_bio;
4397 read_bio->bi_end_io = end_reshape_read;
4398 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4399 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4400 read_bio->bi_error = 0;
4401 read_bio->bi_vcnt = 0;
4402 read_bio->bi_iter.bi_size = 0;
4403 r10_bio->master_bio = read_bio;
4404 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4406 /* Now find the locations in the new layout */
4407 __raid10_find_phys(&conf->geo, r10_bio);
4410 read_bio->bi_next = NULL;
4413 for (s = 0; s < conf->copies*2; s++) {
4415 int d = r10_bio->devs[s/2].devnum;
4416 struct md_rdev *rdev2;
4418 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4419 b = r10_bio->devs[s/2].repl_bio;
4421 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4422 b = r10_bio->devs[s/2].bio;
4424 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4427 b->bi_bdev = rdev2->bdev;
4428 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4429 rdev2->new_data_offset;
4430 b->bi_end_io = end_reshape_write;
4431 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4436 /* Now add as many pages as possible to all of these bios. */
4439 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4440 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4441 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4442 int len = (max_sectors - s) << 9;
4443 if (len > PAGE_SIZE)
4445 for (bio = blist; bio ; bio = bio->bi_next) {
4447 * won't fail because the vec table is big enough
4448 * to hold all these pages
4450 bio_add_page(bio, page, len, 0);
4452 sector_nr += len >> 9;
4453 nr_sectors += len >> 9;
4456 r10_bio->sectors = nr_sectors;
4458 /* Now submit the read */
4459 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4460 atomic_inc(&r10_bio->remaining);
4461 read_bio->bi_next = NULL;
4462 generic_make_request(read_bio);
4463 sector_nr += nr_sectors;
4464 sectors_done += nr_sectors;
4465 if (sector_nr <= last)
4468 /* Now that we have done the whole section we can
4469 * update reshape_progress
4471 if (mddev->reshape_backwards)
4472 conf->reshape_progress -= sectors_done;
4474 conf->reshape_progress += sectors_done;
4476 return sectors_done;
4479 static void end_reshape_request(struct r10bio *r10_bio);
4480 static int handle_reshape_read_error(struct mddev *mddev,
4481 struct r10bio *r10_bio);
4482 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4484 /* Reshape read completed. Hopefully we have a block
4486 * If we got a read error then we do sync 1-page reads from
4487 * elsewhere until we find the data - or give up.
4489 struct r10conf *conf = mddev->private;
4492 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4493 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4494 /* Reshape has been aborted */
4495 md_done_sync(mddev, r10_bio->sectors, 0);
4499 /* We definitely have the data in the pages, schedule the
4502 atomic_set(&r10_bio->remaining, 1);
4503 for (s = 0; s < conf->copies*2; s++) {
4505 int d = r10_bio->devs[s/2].devnum;
4506 struct md_rdev *rdev;
4509 rdev = rcu_dereference(conf->mirrors[d].replacement);
4510 b = r10_bio->devs[s/2].repl_bio;
4512 rdev = rcu_dereference(conf->mirrors[d].rdev);
4513 b = r10_bio->devs[s/2].bio;
4515 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4519 atomic_inc(&rdev->nr_pending);
4521 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4522 atomic_inc(&r10_bio->remaining);
4524 generic_make_request(b);
4526 end_reshape_request(r10_bio);
4529 static void end_reshape(struct r10conf *conf)
4531 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4534 spin_lock_irq(&conf->device_lock);
4535 conf->prev = conf->geo;
4536 md_finish_reshape(conf->mddev);
4538 conf->reshape_progress = MaxSector;
4539 conf->reshape_safe = MaxSector;
4540 spin_unlock_irq(&conf->device_lock);
4542 /* read-ahead size must cover two whole stripes, which is
4543 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4545 if (conf->mddev->queue) {
4546 int stripe = conf->geo.raid_disks *
4547 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4548 stripe /= conf->geo.near_copies;
4549 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4550 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4555 static int handle_reshape_read_error(struct mddev *mddev,
4556 struct r10bio *r10_bio)
4558 /* Use sync reads to get the blocks from somewhere else */
4559 int sectors = r10_bio->sectors;
4560 struct r10conf *conf = mddev->private;
4562 struct r10bio r10_bio;
4563 struct r10dev devs[conf->copies];
4565 struct r10bio *r10b = &on_stack.r10_bio;
4568 struct page **pages;
4570 /* reshape IOs share pages from .devs[0].bio */
4571 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4573 r10b->sector = r10_bio->sector;
4574 __raid10_find_phys(&conf->prev, r10b);
4579 int first_slot = slot;
4581 if (s > (PAGE_SIZE >> 9))
4586 int d = r10b->devs[slot].devnum;
4587 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4590 test_bit(Faulty, &rdev->flags) ||
4591 !test_bit(In_sync, &rdev->flags))
4594 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4595 atomic_inc(&rdev->nr_pending);
4597 success = sync_page_io(rdev,
4601 REQ_OP_READ, 0, false);
4602 rdev_dec_pending(rdev, mddev);
4608 if (slot >= conf->copies)
4610 if (slot == first_slot)
4615 /* couldn't read this block, must give up */
4616 set_bit(MD_RECOVERY_INTR,
4626 static void end_reshape_write(struct bio *bio)
4628 struct r10bio *r10_bio = get_resync_r10bio(bio);
4629 struct mddev *mddev = r10_bio->mddev;
4630 struct r10conf *conf = mddev->private;
4634 struct md_rdev *rdev = NULL;
4636 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4638 rdev = conf->mirrors[d].replacement;
4641 rdev = conf->mirrors[d].rdev;
4644 if (bio->bi_error) {
4645 /* FIXME should record badblock */
4646 md_error(mddev, rdev);
4649 rdev_dec_pending(rdev, mddev);
4650 end_reshape_request(r10_bio);
4653 static void end_reshape_request(struct r10bio *r10_bio)
4655 if (!atomic_dec_and_test(&r10_bio->remaining))
4657 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4658 bio_put(r10_bio->master_bio);
4662 static void raid10_finish_reshape(struct mddev *mddev)
4664 struct r10conf *conf = mddev->private;
4666 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4669 if (mddev->delta_disks > 0) {
4670 sector_t size = raid10_size(mddev, 0, 0);
4671 md_set_array_sectors(mddev, size);
4672 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4673 mddev->recovery_cp = mddev->resync_max_sectors;
4674 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4676 mddev->resync_max_sectors = size;
4678 set_capacity(mddev->gendisk, mddev->array_sectors);
4679 revalidate_disk(mddev->gendisk);
4684 for (d = conf->geo.raid_disks ;
4685 d < conf->geo.raid_disks - mddev->delta_disks;
4687 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4689 clear_bit(In_sync, &rdev->flags);
4690 rdev = rcu_dereference(conf->mirrors[d].replacement);
4692 clear_bit(In_sync, &rdev->flags);
4696 mddev->layout = mddev->new_layout;
4697 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4698 mddev->reshape_position = MaxSector;
4699 mddev->delta_disks = 0;
4700 mddev->reshape_backwards = 0;
4703 static struct md_personality raid10_personality =
4707 .owner = THIS_MODULE,
4708 .make_request = raid10_make_request,
4710 .free = raid10_free,
4711 .status = raid10_status,
4712 .error_handler = raid10_error,
4713 .hot_add_disk = raid10_add_disk,
4714 .hot_remove_disk= raid10_remove_disk,
4715 .spare_active = raid10_spare_active,
4716 .sync_request = raid10_sync_request,
4717 .quiesce = raid10_quiesce,
4718 .size = raid10_size,
4719 .resize = raid10_resize,
4720 .takeover = raid10_takeover,
4721 .check_reshape = raid10_check_reshape,
4722 .start_reshape = raid10_start_reshape,
4723 .finish_reshape = raid10_finish_reshape,
4724 .congested = raid10_congested,
4727 static int __init raid_init(void)
4729 return register_md_personality(&raid10_personality);
4732 static void raid_exit(void)
4734 unregister_md_personality(&raid10_personality);
4737 module_init(raid_init);
4738 module_exit(raid_exit);
4739 MODULE_LICENSE("GPL");
4740 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4741 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4742 MODULE_ALIAS("md-raid10");
4743 MODULE_ALIAS("md-level-10");
4745 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);