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>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
74 * Number of guaranteed r10bios in case of extreme VM load:
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
96 static int max_queued_requests = 1024;
98 static void allow_barrier(struct r10conf *conf);
99 static void lower_barrier(struct r10conf *conf);
100 static int _enough(struct r10conf *conf, int previous, int ignore);
101 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104 static void end_reshape_write(struct bio *bio);
105 static void end_reshape(struct r10conf *conf);
107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109 struct r10conf *conf = data;
110 int size = offsetof(struct r10bio, devs[conf->copies]);
112 /* allocate a r10bio with room for raid_disks entries in the
114 return kzalloc(size, gfp_flags);
117 static void r10bio_pool_free(void *r10_bio, void *data)
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139 struct r10conf *conf = data;
141 struct r10bio *r10_bio;
146 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
150 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152 nalloc = conf->copies; /* resync */
154 nalloc = 2; /* recovery */
159 for (j = nalloc ; j-- ; ) {
160 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
163 r10_bio->devs[j].bio = bio;
164 if (!conf->have_replacement)
166 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
169 r10_bio->devs[j].repl_bio = bio;
172 * Allocate RESYNC_PAGES data pages and attach them
175 for (j = 0 ; j < nalloc; j++) {
176 struct bio *rbio = r10_bio->devs[j].repl_bio;
177 bio = r10_bio->devs[j].bio;
178 for (i = 0; i < RESYNC_PAGES; i++) {
179 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180 &conf->mddev->recovery)) {
181 /* we can share bv_page's during recovery
183 struct bio *rbio = r10_bio->devs[0].bio;
184 page = rbio->bi_io_vec[i].bv_page;
187 page = alloc_page(gfp_flags);
191 bio->bi_io_vec[i].bv_page = page;
193 rbio->bi_io_vec[i].bv_page = page;
201 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203 for (i = 0; i < RESYNC_PAGES ; i++)
204 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
207 for ( ; j < nalloc; j++) {
208 if (r10_bio->devs[j].bio)
209 bio_put(r10_bio->devs[j].bio);
210 if (r10_bio->devs[j].repl_bio)
211 bio_put(r10_bio->devs[j].repl_bio);
213 r10bio_pool_free(r10_bio, conf);
217 static void r10buf_pool_free(void *__r10_bio, void *data)
220 struct r10conf *conf = data;
221 struct r10bio *r10bio = __r10_bio;
224 for (j=0; j < conf->copies; j++) {
225 struct bio *bio = r10bio->devs[j].bio;
227 for (i = 0; i < RESYNC_PAGES; i++) {
228 safe_put_page(bio->bi_io_vec[i].bv_page);
229 bio->bi_io_vec[i].bv_page = NULL;
233 bio = r10bio->devs[j].repl_bio;
237 r10bio_pool_free(r10bio, conf);
240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
244 for (i = 0; i < conf->copies; i++) {
245 struct bio **bio = & r10_bio->devs[i].bio;
246 if (!BIO_SPECIAL(*bio))
249 bio = &r10_bio->devs[i].repl_bio;
250 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
256 static void free_r10bio(struct r10bio *r10_bio)
258 struct r10conf *conf = r10_bio->mddev->private;
260 put_all_bios(conf, r10_bio);
261 mempool_free(r10_bio, conf->r10bio_pool);
264 static void put_buf(struct r10bio *r10_bio)
266 struct r10conf *conf = r10_bio->mddev->private;
268 mempool_free(r10_bio, conf->r10buf_pool);
273 static void reschedule_retry(struct r10bio *r10_bio)
276 struct mddev *mddev = r10_bio->mddev;
277 struct r10conf *conf = mddev->private;
279 spin_lock_irqsave(&conf->device_lock, flags);
280 list_add(&r10_bio->retry_list, &conf->retry_list);
282 spin_unlock_irqrestore(&conf->device_lock, flags);
284 /* wake up frozen array... */
285 wake_up(&conf->wait_barrier);
287 md_wakeup_thread(mddev->thread);
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
295 static void raid_end_bio_io(struct r10bio *r10_bio)
297 struct bio *bio = r10_bio->master_bio;
299 struct r10conf *conf = r10_bio->mddev->private;
301 if (bio->bi_phys_segments) {
303 spin_lock_irqsave(&conf->device_lock, flags);
304 bio->bi_phys_segments--;
305 done = (bio->bi_phys_segments == 0);
306 spin_unlock_irqrestore(&conf->device_lock, flags);
309 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310 bio->bi_error = -EIO;
314 * Wake up any possible resync thread that waits for the device
319 free_r10bio(r10_bio);
323 * Update disk head position estimator based on IRQ completion info.
325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327 struct r10conf *conf = r10_bio->mddev->private;
329 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330 r10_bio->devs[slot].addr + (r10_bio->sectors);
334 * Find the disk number which triggered given bio
336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337 struct bio *bio, int *slotp, int *replp)
342 for (slot = 0; slot < conf->copies; slot++) {
343 if (r10_bio->devs[slot].bio == bio)
345 if (r10_bio->devs[slot].repl_bio == bio) {
351 BUG_ON(slot == conf->copies);
352 update_head_pos(slot, r10_bio);
358 return r10_bio->devs[slot].devnum;
361 static void raid10_end_read_request(struct bio *bio)
363 int uptodate = !bio->bi_error;
364 struct r10bio *r10_bio = bio->bi_private;
366 struct md_rdev *rdev;
367 struct r10conf *conf = r10_bio->mddev->private;
369 slot = r10_bio->read_slot;
370 dev = r10_bio->devs[slot].devnum;
371 rdev = r10_bio->devs[slot].rdev;
373 * this branch is our 'one mirror IO has finished' event handler:
375 update_head_pos(slot, r10_bio);
379 * Set R10BIO_Uptodate in our master bio, so that
380 * we will return a good error code to the higher
381 * levels even if IO on some other mirrored buffer fails.
383 * The 'master' represents the composite IO operation to
384 * user-side. So if something waits for IO, then it will
385 * wait for the 'master' bio.
387 set_bit(R10BIO_Uptodate, &r10_bio->state);
389 /* If all other devices that store this block have
390 * failed, we want to return the error upwards rather
391 * than fail the last device. Here we redefine
392 * "uptodate" to mean "Don't want to retry"
394 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
399 raid_end_bio_io(r10_bio);
400 rdev_dec_pending(rdev, conf->mddev);
403 * oops, read error - keep the refcount on the rdev
405 char b[BDEVNAME_SIZE];
406 printk_ratelimited(KERN_ERR
407 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 bdevname(rdev->bdev, b),
410 (unsigned long long)r10_bio->sector);
411 set_bit(R10BIO_ReadError, &r10_bio->state);
412 reschedule_retry(r10_bio);
416 static void close_write(struct r10bio *r10_bio)
418 /* clear the bitmap if all writes complete successfully */
419 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 md_write_end(r10_bio->mddev);
426 static void one_write_done(struct r10bio *r10_bio)
428 if (atomic_dec_and_test(&r10_bio->remaining)) {
429 if (test_bit(R10BIO_WriteError, &r10_bio->state))
430 reschedule_retry(r10_bio);
432 close_write(r10_bio);
433 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
434 reschedule_retry(r10_bio);
436 raid_end_bio_io(r10_bio);
441 static void raid10_end_write_request(struct bio *bio)
443 struct r10bio *r10_bio = bio->bi_private;
446 struct r10conf *conf = r10_bio->mddev->private;
448 struct md_rdev *rdev = NULL;
450 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
453 rdev = conf->mirrors[dev].replacement;
457 rdev = conf->mirrors[dev].rdev;
460 * this branch is our 'one mirror IO has finished' event handler:
464 /* Never record new bad blocks to replacement,
467 md_error(rdev->mddev, rdev);
469 set_bit(WriteErrorSeen, &rdev->flags);
470 if (!test_and_set_bit(WantReplacement, &rdev->flags))
471 set_bit(MD_RECOVERY_NEEDED,
472 &rdev->mddev->recovery);
473 set_bit(R10BIO_WriteError, &r10_bio->state);
478 * Set R10BIO_Uptodate in our master bio, so that
479 * we will return a good error code for to the higher
480 * levels even if IO on some other mirrored buffer fails.
482 * The 'master' represents the composite IO operation to
483 * user-side. So if something waits for IO, then it will
484 * wait for the 'master' bio.
490 * Do not set R10BIO_Uptodate if the current device is
491 * rebuilding or Faulty. This is because we cannot use
492 * such device for properly reading the data back (we could
493 * potentially use it, if the current write would have felt
494 * before rdev->recovery_offset, but for simplicity we don't
497 if (test_bit(In_sync, &rdev->flags) &&
498 !test_bit(Faulty, &rdev->flags))
499 set_bit(R10BIO_Uptodate, &r10_bio->state);
501 /* Maybe we can clear some bad blocks. */
502 if (is_badblock(rdev,
503 r10_bio->devs[slot].addr,
505 &first_bad, &bad_sectors)) {
508 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
510 r10_bio->devs[slot].bio = IO_MADE_GOOD;
512 set_bit(R10BIO_MadeGood, &r10_bio->state);
518 * Let's see if all mirrored write operations have finished
521 one_write_done(r10_bio);
523 rdev_dec_pending(rdev, conf->mddev);
527 * RAID10 layout manager
528 * As well as the chunksize and raid_disks count, there are two
529 * parameters: near_copies and far_copies.
530 * near_copies * far_copies must be <= raid_disks.
531 * Normally one of these will be 1.
532 * If both are 1, we get raid0.
533 * If near_copies == raid_disks, we get raid1.
535 * Chunks are laid out in raid0 style with near_copies copies of the
536 * first chunk, followed by near_copies copies of the next chunk and
538 * If far_copies > 1, then after 1/far_copies of the array has been assigned
539 * as described above, we start again with a device offset of near_copies.
540 * So we effectively have another copy of the whole array further down all
541 * the drives, but with blocks on different drives.
542 * With this layout, and block is never stored twice on the one device.
544 * raid10_find_phys finds the sector offset of a given virtual sector
545 * on each device that it is on.
547 * raid10_find_virt does the reverse mapping, from a device and a
548 * sector offset to a virtual address
551 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
559 int last_far_set_start, last_far_set_size;
561 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
562 last_far_set_start *= geo->far_set_size;
564 last_far_set_size = geo->far_set_size;
565 last_far_set_size += (geo->raid_disks % geo->far_set_size);
567 /* now calculate first sector/dev */
568 chunk = r10bio->sector >> geo->chunk_shift;
569 sector = r10bio->sector & geo->chunk_mask;
571 chunk *= geo->near_copies;
573 dev = sector_div(stripe, geo->raid_disks);
575 stripe *= geo->far_copies;
577 sector += stripe << geo->chunk_shift;
579 /* and calculate all the others */
580 for (n = 0; n < geo->near_copies; n++) {
584 r10bio->devs[slot].devnum = d;
585 r10bio->devs[slot].addr = s;
588 for (f = 1; f < geo->far_copies; f++) {
589 set = d / geo->far_set_size;
590 d += geo->near_copies;
592 if ((geo->raid_disks % geo->far_set_size) &&
593 (d > last_far_set_start)) {
594 d -= last_far_set_start;
595 d %= last_far_set_size;
596 d += last_far_set_start;
598 d %= geo->far_set_size;
599 d += geo->far_set_size * set;
602 r10bio->devs[slot].devnum = d;
603 r10bio->devs[slot].addr = s;
607 if (dev >= geo->raid_disks) {
609 sector += (geo->chunk_mask + 1);
614 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
616 struct geom *geo = &conf->geo;
618 if (conf->reshape_progress != MaxSector &&
619 ((r10bio->sector >= conf->reshape_progress) !=
620 conf->mddev->reshape_backwards)) {
621 set_bit(R10BIO_Previous, &r10bio->state);
624 clear_bit(R10BIO_Previous, &r10bio->state);
626 __raid10_find_phys(geo, r10bio);
629 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
631 sector_t offset, chunk, vchunk;
632 /* Never use conf->prev as this is only called during resync
633 * or recovery, so reshape isn't happening
635 struct geom *geo = &conf->geo;
636 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
637 int far_set_size = geo->far_set_size;
638 int last_far_set_start;
640 if (geo->raid_disks % geo->far_set_size) {
641 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
642 last_far_set_start *= geo->far_set_size;
644 if (dev >= last_far_set_start) {
645 far_set_size = geo->far_set_size;
646 far_set_size += (geo->raid_disks % geo->far_set_size);
647 far_set_start = last_far_set_start;
651 offset = sector & geo->chunk_mask;
652 if (geo->far_offset) {
654 chunk = sector >> geo->chunk_shift;
655 fc = sector_div(chunk, geo->far_copies);
656 dev -= fc * geo->near_copies;
657 if (dev < far_set_start)
660 while (sector >= geo->stride) {
661 sector -= geo->stride;
662 if (dev < (geo->near_copies + far_set_start))
663 dev += far_set_size - geo->near_copies;
665 dev -= geo->near_copies;
667 chunk = sector >> geo->chunk_shift;
669 vchunk = chunk * geo->raid_disks + dev;
670 sector_div(vchunk, geo->near_copies);
671 return (vchunk << geo->chunk_shift) + offset;
675 * This routine returns the disk from which the requested read should
676 * be done. There is a per-array 'next expected sequential IO' sector
677 * number - if this matches on the next IO then we use the last disk.
678 * There is also a per-disk 'last know head position' sector that is
679 * maintained from IRQ contexts, both the normal and the resync IO
680 * completion handlers update this position correctly. If there is no
681 * perfect sequential match then we pick the disk whose head is closest.
683 * If there are 2 mirrors in the same 2 devices, performance degrades
684 * because position is mirror, not device based.
686 * The rdev for the device selected will have nr_pending incremented.
690 * FIXME: possibly should rethink readbalancing and do it differently
691 * depending on near_copies / far_copies geometry.
693 static struct md_rdev *read_balance(struct r10conf *conf,
694 struct r10bio *r10_bio,
697 const sector_t this_sector = r10_bio->sector;
699 int sectors = r10_bio->sectors;
700 int best_good_sectors;
701 sector_t new_distance, best_dist;
702 struct md_rdev *best_rdev, *rdev = NULL;
705 struct geom *geo = &conf->geo;
707 raid10_find_phys(conf, r10_bio);
710 sectors = r10_bio->sectors;
713 best_dist = MaxSector;
714 best_good_sectors = 0;
717 * Check if we can balance. We can balance on the whole
718 * device if no resync is going on (recovery is ok), or below
719 * the resync window. We take the first readable disk when
720 * above the resync window.
722 if (conf->mddev->recovery_cp < MaxSector
723 && (this_sector + sectors >= conf->next_resync))
726 for (slot = 0; slot < conf->copies ; slot++) {
731 if (r10_bio->devs[slot].bio == IO_BLOCKED)
733 disk = r10_bio->devs[slot].devnum;
734 rdev = rcu_dereference(conf->mirrors[disk].replacement);
735 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
736 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
737 rdev = rcu_dereference(conf->mirrors[disk].rdev);
739 test_bit(Faulty, &rdev->flags))
741 if (!test_bit(In_sync, &rdev->flags) &&
742 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
745 dev_sector = r10_bio->devs[slot].addr;
746 if (is_badblock(rdev, dev_sector, sectors,
747 &first_bad, &bad_sectors)) {
748 if (best_dist < MaxSector)
749 /* Already have a better slot */
751 if (first_bad <= dev_sector) {
752 /* Cannot read here. If this is the
753 * 'primary' device, then we must not read
754 * beyond 'bad_sectors' from another device.
756 bad_sectors -= (dev_sector - first_bad);
757 if (!do_balance && sectors > bad_sectors)
758 sectors = bad_sectors;
759 if (best_good_sectors > sectors)
760 best_good_sectors = sectors;
762 sector_t good_sectors =
763 first_bad - dev_sector;
764 if (good_sectors > best_good_sectors) {
765 best_good_sectors = good_sectors;
770 /* Must read from here */
775 best_good_sectors = sectors;
780 /* This optimisation is debatable, and completely destroys
781 * sequential read speed for 'far copies' arrays. So only
782 * keep it for 'near' arrays, and review those later.
784 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
787 /* for far > 1 always use the lowest address */
788 if (geo->far_copies > 1)
789 new_distance = r10_bio->devs[slot].addr;
791 new_distance = abs(r10_bio->devs[slot].addr -
792 conf->mirrors[disk].head_position);
793 if (new_distance < best_dist) {
794 best_dist = new_distance;
799 if (slot >= conf->copies) {
805 atomic_inc(&rdev->nr_pending);
806 if (test_bit(Faulty, &rdev->flags)) {
807 /* Cannot risk returning a device that failed
808 * before we inc'ed nr_pending
810 rdev_dec_pending(rdev, conf->mddev);
813 r10_bio->read_slot = slot;
817 *max_sectors = best_good_sectors;
822 static int raid10_congested(struct mddev *mddev, int bits)
824 struct r10conf *conf = mddev->private;
827 if ((bits & (1 << WB_async_congested)) &&
828 conf->pending_count >= max_queued_requests)
833 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
836 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
837 if (rdev && !test_bit(Faulty, &rdev->flags)) {
838 struct request_queue *q = bdev_get_queue(rdev->bdev);
840 ret |= bdi_congested(&q->backing_dev_info, bits);
847 static void flush_pending_writes(struct r10conf *conf)
849 /* Any writes that have been queued but are awaiting
850 * bitmap updates get flushed here.
852 spin_lock_irq(&conf->device_lock);
854 if (conf->pending_bio_list.head) {
856 bio = bio_list_get(&conf->pending_bio_list);
857 conf->pending_count = 0;
858 spin_unlock_irq(&conf->device_lock);
859 /* flush any pending bitmap writes to disk
860 * before proceeding w/ I/O */
861 bitmap_unplug(conf->mddev->bitmap);
862 wake_up(&conf->wait_barrier);
864 while (bio) { /* submit pending writes */
865 struct bio *next = bio->bi_next;
867 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
868 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
872 generic_make_request(bio);
876 spin_unlock_irq(&conf->device_lock);
880 * Sometimes we need to suspend IO while we do something else,
881 * either some resync/recovery, or reconfigure the array.
882 * To do this we raise a 'barrier'.
883 * The 'barrier' is a counter that can be raised multiple times
884 * to count how many activities are happening which preclude
886 * We can only raise the barrier if there is no pending IO.
887 * i.e. if nr_pending == 0.
888 * We choose only to raise the barrier if no-one is waiting for the
889 * barrier to go down. This means that as soon as an IO request
890 * is ready, no other operations which require a barrier will start
891 * until the IO request has had a chance.
893 * So: regular IO calls 'wait_barrier'. When that returns there
894 * is no backgroup IO happening, It must arrange to call
895 * allow_barrier when it has finished its IO.
896 * backgroup IO calls must call raise_barrier. Once that returns
897 * there is no normal IO happeing. It must arrange to call
898 * lower_barrier when the particular background IO completes.
901 static void raise_barrier(struct r10conf *conf, int force)
903 BUG_ON(force && !conf->barrier);
904 spin_lock_irq(&conf->resync_lock);
906 /* Wait until no block IO is waiting (unless 'force') */
907 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
910 /* block any new IO from starting */
913 /* Now wait for all pending IO to complete */
914 wait_event_lock_irq(conf->wait_barrier,
915 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
918 spin_unlock_irq(&conf->resync_lock);
921 static void lower_barrier(struct r10conf *conf)
924 spin_lock_irqsave(&conf->resync_lock, flags);
926 spin_unlock_irqrestore(&conf->resync_lock, flags);
927 wake_up(&conf->wait_barrier);
930 static void wait_barrier(struct r10conf *conf)
932 spin_lock_irq(&conf->resync_lock);
935 /* Wait for the barrier to drop.
936 * However if there are already pending
937 * requests (preventing the barrier from
938 * rising completely), and the
939 * pre-process bio queue isn't empty,
940 * then don't wait, as we need to empty
941 * that queue to get the nr_pending
944 wait_event_lock_irq(conf->wait_barrier,
948 !bio_list_empty(current->bio_list)),
953 spin_unlock_irq(&conf->resync_lock);
956 static void allow_barrier(struct r10conf *conf)
959 spin_lock_irqsave(&conf->resync_lock, flags);
961 spin_unlock_irqrestore(&conf->resync_lock, flags);
962 wake_up(&conf->wait_barrier);
965 static void freeze_array(struct r10conf *conf, int extra)
967 /* stop syncio and normal IO and wait for everything to
969 * We increment barrier and nr_waiting, and then
970 * wait until nr_pending match nr_queued+extra
971 * This is called in the context of one normal IO request
972 * that has failed. Thus any sync request that might be pending
973 * will be blocked by nr_pending, and we need to wait for
974 * pending IO requests to complete or be queued for re-try.
975 * Thus the number queued (nr_queued) plus this request (extra)
976 * must match the number of pending IOs (nr_pending) before
979 spin_lock_irq(&conf->resync_lock);
982 wait_event_lock_irq_cmd(conf->wait_barrier,
983 conf->nr_pending == conf->nr_queued+extra,
985 flush_pending_writes(conf));
987 spin_unlock_irq(&conf->resync_lock);
990 static void unfreeze_array(struct r10conf *conf)
992 /* reverse the effect of the freeze */
993 spin_lock_irq(&conf->resync_lock);
996 wake_up(&conf->wait_barrier);
997 spin_unlock_irq(&conf->resync_lock);
1000 static sector_t choose_data_offset(struct r10bio *r10_bio,
1001 struct md_rdev *rdev)
1003 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1004 test_bit(R10BIO_Previous, &r10_bio->state))
1005 return rdev->data_offset;
1007 return rdev->new_data_offset;
1010 struct raid10_plug_cb {
1011 struct blk_plug_cb cb;
1012 struct bio_list pending;
1016 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1018 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1020 struct mddev *mddev = plug->cb.data;
1021 struct r10conf *conf = mddev->private;
1024 if (from_schedule || current->bio_list) {
1025 spin_lock_irq(&conf->device_lock);
1026 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1027 conf->pending_count += plug->pending_cnt;
1028 spin_unlock_irq(&conf->device_lock);
1029 wake_up(&conf->wait_barrier);
1030 md_wakeup_thread(mddev->thread);
1035 /* we aren't scheduling, so we can do the write-out directly. */
1036 bio = bio_list_get(&plug->pending);
1037 bitmap_unplug(mddev->bitmap);
1038 wake_up(&conf->wait_barrier);
1040 while (bio) { /* submit pending writes */
1041 struct bio *next = bio->bi_next;
1042 bio->bi_next = NULL;
1043 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1044 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1045 /* Just ignore it */
1048 generic_make_request(bio);
1054 static void __make_request(struct mddev *mddev, struct bio *bio)
1056 struct r10conf *conf = mddev->private;
1057 struct r10bio *r10_bio;
1058 struct bio *read_bio;
1060 const int rw = bio_data_dir(bio);
1061 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1062 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1063 const unsigned long do_discard = (bio->bi_rw
1064 & (REQ_DISCARD | REQ_SECURE));
1065 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1066 unsigned long flags;
1067 struct md_rdev *blocked_rdev;
1068 struct blk_plug_cb *cb;
1069 struct raid10_plug_cb *plug = NULL;
1070 int sectors_handled;
1075 * Register the new request and wait if the reconstruction
1076 * thread has put up a bar for new requests.
1077 * Continue immediately if no resync is active currently.
1081 sectors = bio_sectors(bio);
1082 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1083 bio->bi_iter.bi_sector < conf->reshape_progress &&
1084 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1085 /* IO spans the reshape position. Need to wait for
1088 allow_barrier(conf);
1089 wait_event(conf->wait_barrier,
1090 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1091 conf->reshape_progress >= bio->bi_iter.bi_sector +
1095 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1096 bio_data_dir(bio) == WRITE &&
1097 (mddev->reshape_backwards
1098 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1099 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1100 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1101 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1102 /* Need to update reshape_position in metadata */
1103 mddev->reshape_position = conf->reshape_progress;
1104 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1105 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1106 md_wakeup_thread(mddev->thread);
1107 wait_event(mddev->sb_wait,
1108 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1110 conf->reshape_safe = mddev->reshape_position;
1113 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1115 r10_bio->master_bio = bio;
1116 r10_bio->sectors = sectors;
1118 r10_bio->mddev = mddev;
1119 r10_bio->sector = bio->bi_iter.bi_sector;
1122 /* We might need to issue multiple reads to different
1123 * devices if there are bad blocks around, so we keep
1124 * track of the number of reads in bio->bi_phys_segments.
1125 * If this is 0, there is only one r10_bio and no locking
1126 * will be needed when the request completes. If it is
1127 * non-zero, then it is the number of not-completed requests.
1129 bio->bi_phys_segments = 0;
1130 bio_clear_flag(bio, BIO_SEG_VALID);
1134 * read balancing logic:
1136 struct md_rdev *rdev;
1140 rdev = read_balance(conf, r10_bio, &max_sectors);
1142 raid_end_bio_io(r10_bio);
1145 slot = r10_bio->read_slot;
1147 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1148 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1151 r10_bio->devs[slot].bio = read_bio;
1152 r10_bio->devs[slot].rdev = rdev;
1154 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1155 choose_data_offset(r10_bio, rdev);
1156 read_bio->bi_bdev = rdev->bdev;
1157 read_bio->bi_end_io = raid10_end_read_request;
1158 read_bio->bi_rw = READ | do_sync;
1159 read_bio->bi_private = r10_bio;
1161 if (max_sectors < r10_bio->sectors) {
1162 /* Could not read all from this device, so we will
1163 * need another r10_bio.
1165 sectors_handled = (r10_bio->sector + max_sectors
1166 - bio->bi_iter.bi_sector);
1167 r10_bio->sectors = max_sectors;
1168 spin_lock_irq(&conf->device_lock);
1169 if (bio->bi_phys_segments == 0)
1170 bio->bi_phys_segments = 2;
1172 bio->bi_phys_segments++;
1173 spin_unlock_irq(&conf->device_lock);
1174 /* Cannot call generic_make_request directly
1175 * as that will be queued in __generic_make_request
1176 * and subsequent mempool_alloc might block
1177 * waiting for it. so hand bio over to raid10d.
1179 reschedule_retry(r10_bio);
1181 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1183 r10_bio->master_bio = bio;
1184 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1186 r10_bio->mddev = mddev;
1187 r10_bio->sector = bio->bi_iter.bi_sector +
1191 generic_make_request(read_bio);
1198 if (conf->pending_count >= max_queued_requests) {
1199 md_wakeup_thread(mddev->thread);
1200 wait_event(conf->wait_barrier,
1201 conf->pending_count < max_queued_requests);
1203 /* first select target devices under rcu_lock and
1204 * inc refcount on their rdev. Record them by setting
1206 * If there are known/acknowledged bad blocks on any device
1207 * on which we have seen a write error, we want to avoid
1208 * writing to those blocks. This potentially requires several
1209 * writes to write around the bad blocks. Each set of writes
1210 * gets its own r10_bio with a set of bios attached. The number
1211 * of r10_bios is recored in bio->bi_phys_segments just as with
1215 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1216 raid10_find_phys(conf, r10_bio);
1218 blocked_rdev = NULL;
1220 max_sectors = r10_bio->sectors;
1222 for (i = 0; i < conf->copies; i++) {
1223 int d = r10_bio->devs[i].devnum;
1224 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1225 struct md_rdev *rrdev = rcu_dereference(
1226 conf->mirrors[d].replacement);
1229 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1230 atomic_inc(&rdev->nr_pending);
1231 blocked_rdev = rdev;
1234 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1235 atomic_inc(&rrdev->nr_pending);
1236 blocked_rdev = rrdev;
1239 if (rdev && (test_bit(Faulty, &rdev->flags)))
1241 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1244 r10_bio->devs[i].bio = NULL;
1245 r10_bio->devs[i].repl_bio = NULL;
1247 if (!rdev && !rrdev) {
1248 set_bit(R10BIO_Degraded, &r10_bio->state);
1251 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1253 sector_t dev_sector = r10_bio->devs[i].addr;
1257 is_bad = is_badblock(rdev, dev_sector,
1259 &first_bad, &bad_sectors);
1261 /* Mustn't write here until the bad block
1264 atomic_inc(&rdev->nr_pending);
1265 set_bit(BlockedBadBlocks, &rdev->flags);
1266 blocked_rdev = rdev;
1269 if (is_bad && first_bad <= dev_sector) {
1270 /* Cannot write here at all */
1271 bad_sectors -= (dev_sector - first_bad);
1272 if (bad_sectors < max_sectors)
1273 /* Mustn't write more than bad_sectors
1274 * to other devices yet
1276 max_sectors = bad_sectors;
1277 /* We don't set R10BIO_Degraded as that
1278 * only applies if the disk is missing,
1279 * so it might be re-added, and we want to
1280 * know to recover this chunk.
1281 * In this case the device is here, and the
1282 * fact that this chunk is not in-sync is
1283 * recorded in the bad block log.
1288 int good_sectors = first_bad - dev_sector;
1289 if (good_sectors < max_sectors)
1290 max_sectors = good_sectors;
1294 r10_bio->devs[i].bio = bio;
1295 atomic_inc(&rdev->nr_pending);
1298 r10_bio->devs[i].repl_bio = bio;
1299 atomic_inc(&rrdev->nr_pending);
1304 if (unlikely(blocked_rdev)) {
1305 /* Have to wait for this device to get unblocked, then retry */
1309 for (j = 0; j < i; j++) {
1310 if (r10_bio->devs[j].bio) {
1311 d = r10_bio->devs[j].devnum;
1312 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1314 if (r10_bio->devs[j].repl_bio) {
1315 struct md_rdev *rdev;
1316 d = r10_bio->devs[j].devnum;
1317 rdev = conf->mirrors[d].replacement;
1319 /* Race with remove_disk */
1321 rdev = conf->mirrors[d].rdev;
1323 rdev_dec_pending(rdev, mddev);
1326 allow_barrier(conf);
1327 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1332 if (max_sectors < r10_bio->sectors) {
1333 /* We are splitting this into multiple parts, so
1334 * we need to prepare for allocating another r10_bio.
1336 r10_bio->sectors = max_sectors;
1337 spin_lock_irq(&conf->device_lock);
1338 if (bio->bi_phys_segments == 0)
1339 bio->bi_phys_segments = 2;
1341 bio->bi_phys_segments++;
1342 spin_unlock_irq(&conf->device_lock);
1344 sectors_handled = r10_bio->sector + max_sectors -
1345 bio->bi_iter.bi_sector;
1347 atomic_set(&r10_bio->remaining, 1);
1348 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1350 for (i = 0; i < conf->copies; i++) {
1352 int d = r10_bio->devs[i].devnum;
1353 if (r10_bio->devs[i].bio) {
1354 struct md_rdev *rdev = conf->mirrors[d].rdev;
1355 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1356 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1358 r10_bio->devs[i].bio = mbio;
1360 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1361 choose_data_offset(r10_bio,
1363 mbio->bi_bdev = rdev->bdev;
1364 mbio->bi_end_io = raid10_end_write_request;
1366 WRITE | do_sync | do_fua | do_discard | do_same;
1367 mbio->bi_private = r10_bio;
1369 atomic_inc(&r10_bio->remaining);
1371 cb = blk_check_plugged(raid10_unplug, mddev,
1374 plug = container_of(cb, struct raid10_plug_cb,
1378 spin_lock_irqsave(&conf->device_lock, flags);
1380 bio_list_add(&plug->pending, mbio);
1381 plug->pending_cnt++;
1383 bio_list_add(&conf->pending_bio_list, mbio);
1384 conf->pending_count++;
1386 spin_unlock_irqrestore(&conf->device_lock, flags);
1388 md_wakeup_thread(mddev->thread);
1391 if (r10_bio->devs[i].repl_bio) {
1392 struct md_rdev *rdev = conf->mirrors[d].replacement;
1394 /* Replacement just got moved to main 'rdev' */
1396 rdev = conf->mirrors[d].rdev;
1398 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1399 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1401 r10_bio->devs[i].repl_bio = mbio;
1403 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1406 mbio->bi_bdev = rdev->bdev;
1407 mbio->bi_end_io = raid10_end_write_request;
1409 WRITE | do_sync | do_fua | do_discard | do_same;
1410 mbio->bi_private = r10_bio;
1412 atomic_inc(&r10_bio->remaining);
1413 spin_lock_irqsave(&conf->device_lock, flags);
1414 bio_list_add(&conf->pending_bio_list, mbio);
1415 conf->pending_count++;
1416 spin_unlock_irqrestore(&conf->device_lock, flags);
1417 if (!mddev_check_plugged(mddev))
1418 md_wakeup_thread(mddev->thread);
1422 /* Don't remove the bias on 'remaining' (one_write_done) until
1423 * after checking if we need to go around again.
1426 if (sectors_handled < bio_sectors(bio)) {
1427 one_write_done(r10_bio);
1428 /* We need another r10_bio. It has already been counted
1429 * in bio->bi_phys_segments.
1431 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1433 r10_bio->master_bio = bio;
1434 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1436 r10_bio->mddev = mddev;
1437 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1441 one_write_done(r10_bio);
1444 static void make_request(struct mddev *mddev, struct bio *bio)
1446 struct r10conf *conf = mddev->private;
1447 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1448 int chunk_sects = chunk_mask + 1;
1452 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1453 md_flush_request(mddev, bio);
1457 md_write_start(mddev, bio);
1462 * If this request crosses a chunk boundary, we need to split
1465 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1466 bio_sectors(bio) > chunk_sects
1467 && (conf->geo.near_copies < conf->geo.raid_disks
1468 || conf->prev.near_copies <
1469 conf->prev.raid_disks))) {
1470 split = bio_split(bio, chunk_sects -
1471 (bio->bi_iter.bi_sector &
1473 GFP_NOIO, fs_bio_set);
1474 bio_chain(split, bio);
1479 __make_request(mddev, split);
1480 } while (split != bio);
1482 /* In case raid10d snuck in to freeze_array */
1483 wake_up(&conf->wait_barrier);
1486 static void status(struct seq_file *seq, struct mddev *mddev)
1488 struct r10conf *conf = mddev->private;
1491 if (conf->geo.near_copies < conf->geo.raid_disks)
1492 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1493 if (conf->geo.near_copies > 1)
1494 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1495 if (conf->geo.far_copies > 1) {
1496 if (conf->geo.far_offset)
1497 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1499 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1501 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1502 conf->geo.raid_disks - mddev->degraded);
1503 for (i = 0; i < conf->geo.raid_disks; i++)
1504 seq_printf(seq, "%s",
1505 conf->mirrors[i].rdev &&
1506 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1507 seq_printf(seq, "]");
1510 /* check if there are enough drives for
1511 * every block to appear on atleast one.
1512 * Don't consider the device numbered 'ignore'
1513 * as we might be about to remove it.
1515 static int _enough(struct r10conf *conf, int previous, int ignore)
1521 disks = conf->prev.raid_disks;
1522 ncopies = conf->prev.near_copies;
1524 disks = conf->geo.raid_disks;
1525 ncopies = conf->geo.near_copies;
1530 int n = conf->copies;
1534 struct md_rdev *rdev;
1535 if (this != ignore &&
1536 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1537 test_bit(In_sync, &rdev->flags))
1539 this = (this+1) % disks;
1543 first = (first + ncopies) % disks;
1544 } while (first != 0);
1551 static int enough(struct r10conf *conf, int ignore)
1553 /* when calling 'enough', both 'prev' and 'geo' must
1555 * This is ensured if ->reconfig_mutex or ->device_lock
1558 return _enough(conf, 0, ignore) &&
1559 _enough(conf, 1, ignore);
1562 static void error(struct mddev *mddev, struct md_rdev *rdev)
1564 char b[BDEVNAME_SIZE];
1565 struct r10conf *conf = mddev->private;
1566 unsigned long flags;
1569 * If it is not operational, then we have already marked it as dead
1570 * else if it is the last working disks, ignore the error, let the
1571 * next level up know.
1572 * else mark the drive as failed
1574 spin_lock_irqsave(&conf->device_lock, flags);
1575 if (test_bit(In_sync, &rdev->flags)
1576 && !enough(conf, rdev->raid_disk)) {
1578 * Don't fail the drive, just return an IO error.
1580 spin_unlock_irqrestore(&conf->device_lock, flags);
1583 if (test_and_clear_bit(In_sync, &rdev->flags))
1586 * If recovery is running, make sure it aborts.
1588 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1589 set_bit(Blocked, &rdev->flags);
1590 set_bit(Faulty, &rdev->flags);
1591 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1592 spin_unlock_irqrestore(&conf->device_lock, flags);
1594 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1595 "md/raid10:%s: Operation continuing on %d devices.\n",
1596 mdname(mddev), bdevname(rdev->bdev, b),
1597 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1600 static void print_conf(struct r10conf *conf)
1603 struct raid10_info *tmp;
1605 printk(KERN_DEBUG "RAID10 conf printout:\n");
1607 printk(KERN_DEBUG "(!conf)\n");
1610 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1611 conf->geo.raid_disks);
1613 for (i = 0; i < conf->geo.raid_disks; i++) {
1614 char b[BDEVNAME_SIZE];
1615 tmp = conf->mirrors + i;
1617 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1618 i, !test_bit(In_sync, &tmp->rdev->flags),
1619 !test_bit(Faulty, &tmp->rdev->flags),
1620 bdevname(tmp->rdev->bdev,b));
1624 static void close_sync(struct r10conf *conf)
1627 allow_barrier(conf);
1629 mempool_destroy(conf->r10buf_pool);
1630 conf->r10buf_pool = NULL;
1633 static int raid10_spare_active(struct mddev *mddev)
1636 struct r10conf *conf = mddev->private;
1637 struct raid10_info *tmp;
1639 unsigned long flags;
1642 * Find all non-in_sync disks within the RAID10 configuration
1643 * and mark them in_sync
1645 for (i = 0; i < conf->geo.raid_disks; i++) {
1646 tmp = conf->mirrors + i;
1647 if (tmp->replacement
1648 && tmp->replacement->recovery_offset == MaxSector
1649 && !test_bit(Faulty, &tmp->replacement->flags)
1650 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1651 /* Replacement has just become active */
1653 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1656 /* Replaced device not technically faulty,
1657 * but we need to be sure it gets removed
1658 * and never re-added.
1660 set_bit(Faulty, &tmp->rdev->flags);
1661 sysfs_notify_dirent_safe(
1662 tmp->rdev->sysfs_state);
1664 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1665 } else if (tmp->rdev
1666 && tmp->rdev->recovery_offset == MaxSector
1667 && !test_bit(Faulty, &tmp->rdev->flags)
1668 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1670 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1673 spin_lock_irqsave(&conf->device_lock, flags);
1674 mddev->degraded -= count;
1675 spin_unlock_irqrestore(&conf->device_lock, flags);
1681 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1683 struct r10conf *conf = mddev->private;
1687 int last = conf->geo.raid_disks - 1;
1689 if (mddev->recovery_cp < MaxSector)
1690 /* only hot-add to in-sync arrays, as recovery is
1691 * very different from resync
1694 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1697 if (rdev->raid_disk >= 0)
1698 first = last = rdev->raid_disk;
1700 if (rdev->saved_raid_disk >= first &&
1701 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1702 mirror = rdev->saved_raid_disk;
1705 for ( ; mirror <= last ; mirror++) {
1706 struct raid10_info *p = &conf->mirrors[mirror];
1707 if (p->recovery_disabled == mddev->recovery_disabled)
1710 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1711 p->replacement != NULL)
1713 clear_bit(In_sync, &rdev->flags);
1714 set_bit(Replacement, &rdev->flags);
1715 rdev->raid_disk = mirror;
1718 disk_stack_limits(mddev->gendisk, rdev->bdev,
1719 rdev->data_offset << 9);
1721 rcu_assign_pointer(p->replacement, rdev);
1726 disk_stack_limits(mddev->gendisk, rdev->bdev,
1727 rdev->data_offset << 9);
1729 p->head_position = 0;
1730 p->recovery_disabled = mddev->recovery_disabled - 1;
1731 rdev->raid_disk = mirror;
1733 if (rdev->saved_raid_disk != mirror)
1735 rcu_assign_pointer(p->rdev, rdev);
1738 md_integrity_add_rdev(rdev, mddev);
1739 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1740 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1746 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1748 struct r10conf *conf = mddev->private;
1750 int number = rdev->raid_disk;
1751 struct md_rdev **rdevp;
1752 struct raid10_info *p = conf->mirrors + number;
1755 if (rdev == p->rdev)
1757 else if (rdev == p->replacement)
1758 rdevp = &p->replacement;
1762 if (test_bit(In_sync, &rdev->flags) ||
1763 atomic_read(&rdev->nr_pending)) {
1767 /* Only remove faulty devices if recovery
1770 if (!test_bit(Faulty, &rdev->flags) &&
1771 mddev->recovery_disabled != p->recovery_disabled &&
1772 (!p->replacement || p->replacement == rdev) &&
1773 number < conf->geo.raid_disks &&
1780 if (atomic_read(&rdev->nr_pending)) {
1781 /* lost the race, try later */
1785 } else if (p->replacement) {
1786 /* We must have just cleared 'rdev' */
1787 p->rdev = p->replacement;
1788 clear_bit(Replacement, &p->replacement->flags);
1789 smp_mb(); /* Make sure other CPUs may see both as identical
1790 * but will never see neither -- if they are careful.
1792 p->replacement = NULL;
1793 clear_bit(WantReplacement, &rdev->flags);
1795 /* We might have just remove the Replacement as faulty
1796 * Clear the flag just in case
1798 clear_bit(WantReplacement, &rdev->flags);
1800 err = md_integrity_register(mddev);
1808 static void end_sync_read(struct bio *bio)
1810 struct r10bio *r10_bio = bio->bi_private;
1811 struct r10conf *conf = r10_bio->mddev->private;
1814 if (bio == r10_bio->master_bio) {
1815 /* this is a reshape read */
1816 d = r10_bio->read_slot; /* really the read dev */
1818 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1821 set_bit(R10BIO_Uptodate, &r10_bio->state);
1823 /* The write handler will notice the lack of
1824 * R10BIO_Uptodate and record any errors etc
1826 atomic_add(r10_bio->sectors,
1827 &conf->mirrors[d].rdev->corrected_errors);
1829 /* for reconstruct, we always reschedule after a read.
1830 * for resync, only after all reads
1832 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1833 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1834 atomic_dec_and_test(&r10_bio->remaining)) {
1835 /* we have read all the blocks,
1836 * do the comparison in process context in raid10d
1838 reschedule_retry(r10_bio);
1842 static void end_sync_request(struct r10bio *r10_bio)
1844 struct mddev *mddev = r10_bio->mddev;
1846 while (atomic_dec_and_test(&r10_bio->remaining)) {
1847 if (r10_bio->master_bio == NULL) {
1848 /* the primary of several recovery bios */
1849 sector_t s = r10_bio->sectors;
1850 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1851 test_bit(R10BIO_WriteError, &r10_bio->state))
1852 reschedule_retry(r10_bio);
1855 md_done_sync(mddev, s, 1);
1858 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1859 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1860 test_bit(R10BIO_WriteError, &r10_bio->state))
1861 reschedule_retry(r10_bio);
1869 static void end_sync_write(struct bio *bio)
1871 struct r10bio *r10_bio = bio->bi_private;
1872 struct mddev *mddev = r10_bio->mddev;
1873 struct r10conf *conf = mddev->private;
1879 struct md_rdev *rdev = NULL;
1881 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1883 rdev = conf->mirrors[d].replacement;
1885 rdev = conf->mirrors[d].rdev;
1887 if (bio->bi_error) {
1889 md_error(mddev, rdev);
1891 set_bit(WriteErrorSeen, &rdev->flags);
1892 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1893 set_bit(MD_RECOVERY_NEEDED,
1894 &rdev->mddev->recovery);
1895 set_bit(R10BIO_WriteError, &r10_bio->state);
1897 } else if (is_badblock(rdev,
1898 r10_bio->devs[slot].addr,
1900 &first_bad, &bad_sectors))
1901 set_bit(R10BIO_MadeGood, &r10_bio->state);
1903 rdev_dec_pending(rdev, mddev);
1905 end_sync_request(r10_bio);
1909 * Note: sync and recover and handled very differently for raid10
1910 * This code is for resync.
1911 * For resync, we read through virtual addresses and read all blocks.
1912 * If there is any error, we schedule a write. The lowest numbered
1913 * drive is authoritative.
1914 * However requests come for physical address, so we need to map.
1915 * For every physical address there are raid_disks/copies virtual addresses,
1916 * which is always are least one, but is not necessarly an integer.
1917 * This means that a physical address can span multiple chunks, so we may
1918 * have to submit multiple io requests for a single sync request.
1921 * We check if all blocks are in-sync and only write to blocks that
1924 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1926 struct r10conf *conf = mddev->private;
1928 struct bio *tbio, *fbio;
1931 atomic_set(&r10_bio->remaining, 1);
1933 /* find the first device with a block */
1934 for (i=0; i<conf->copies; i++)
1935 if (!r10_bio->devs[i].bio->bi_error)
1938 if (i == conf->copies)
1942 fbio = r10_bio->devs[i].bio;
1944 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1945 /* now find blocks with errors */
1946 for (i=0 ; i < conf->copies ; i++) {
1949 tbio = r10_bio->devs[i].bio;
1951 if (tbio->bi_end_io != end_sync_read)
1955 if (!r10_bio->devs[i].bio->bi_error) {
1956 /* We know that the bi_io_vec layout is the same for
1957 * both 'first' and 'i', so we just compare them.
1958 * All vec entries are PAGE_SIZE;
1960 int sectors = r10_bio->sectors;
1961 for (j = 0; j < vcnt; j++) {
1962 int len = PAGE_SIZE;
1963 if (sectors < (len / 512))
1964 len = sectors * 512;
1965 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1966 page_address(tbio->bi_io_vec[j].bv_page),
1973 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1974 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1975 /* Don't fix anything. */
1978 /* Ok, we need to write this bio, either to correct an
1979 * inconsistency or to correct an unreadable block.
1980 * First we need to fixup bv_offset, bv_len and
1981 * bi_vecs, as the read request might have corrupted these
1985 tbio->bi_vcnt = vcnt;
1986 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
1987 tbio->bi_rw = WRITE;
1988 tbio->bi_private = r10_bio;
1989 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1990 tbio->bi_end_io = end_sync_write;
1992 bio_copy_data(tbio, fbio);
1994 d = r10_bio->devs[i].devnum;
1995 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1996 atomic_inc(&r10_bio->remaining);
1997 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
1999 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2000 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2001 generic_make_request(tbio);
2004 /* Now write out to any replacement devices
2007 for (i = 0; i < conf->copies; i++) {
2010 tbio = r10_bio->devs[i].repl_bio;
2011 if (!tbio || !tbio->bi_end_io)
2013 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2014 && r10_bio->devs[i].bio != fbio)
2015 bio_copy_data(tbio, fbio);
2016 d = r10_bio->devs[i].devnum;
2017 atomic_inc(&r10_bio->remaining);
2018 md_sync_acct(conf->mirrors[d].replacement->bdev,
2020 generic_make_request(tbio);
2024 if (atomic_dec_and_test(&r10_bio->remaining)) {
2025 md_done_sync(mddev, r10_bio->sectors, 1);
2031 * Now for the recovery code.
2032 * Recovery happens across physical sectors.
2033 * We recover all non-is_sync drives by finding the virtual address of
2034 * each, and then choose a working drive that also has that virt address.
2035 * There is a separate r10_bio for each non-in_sync drive.
2036 * Only the first two slots are in use. The first for reading,
2037 * The second for writing.
2040 static void fix_recovery_read_error(struct r10bio *r10_bio)
2042 /* We got a read error during recovery.
2043 * We repeat the read in smaller page-sized sections.
2044 * If a read succeeds, write it to the new device or record
2045 * a bad block if we cannot.
2046 * If a read fails, record a bad block on both old and
2049 struct mddev *mddev = r10_bio->mddev;
2050 struct r10conf *conf = mddev->private;
2051 struct bio *bio = r10_bio->devs[0].bio;
2053 int sectors = r10_bio->sectors;
2055 int dr = r10_bio->devs[0].devnum;
2056 int dw = r10_bio->devs[1].devnum;
2060 struct md_rdev *rdev;
2064 if (s > (PAGE_SIZE>>9))
2067 rdev = conf->mirrors[dr].rdev;
2068 addr = r10_bio->devs[0].addr + sect,
2069 ok = sync_page_io(rdev,
2072 bio->bi_io_vec[idx].bv_page,
2075 rdev = conf->mirrors[dw].rdev;
2076 addr = r10_bio->devs[1].addr + sect;
2077 ok = sync_page_io(rdev,
2080 bio->bi_io_vec[idx].bv_page,
2083 set_bit(WriteErrorSeen, &rdev->flags);
2084 if (!test_and_set_bit(WantReplacement,
2086 set_bit(MD_RECOVERY_NEEDED,
2087 &rdev->mddev->recovery);
2091 /* We don't worry if we cannot set a bad block -
2092 * it really is bad so there is no loss in not
2095 rdev_set_badblocks(rdev, addr, s, 0);
2097 if (rdev != conf->mirrors[dw].rdev) {
2098 /* need bad block on destination too */
2099 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2100 addr = r10_bio->devs[1].addr + sect;
2101 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2103 /* just abort the recovery */
2105 "md/raid10:%s: recovery aborted"
2106 " due to read error\n",
2109 conf->mirrors[dw].recovery_disabled
2110 = mddev->recovery_disabled;
2111 set_bit(MD_RECOVERY_INTR,
2124 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2126 struct r10conf *conf = mddev->private;
2128 struct bio *wbio, *wbio2;
2130 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2131 fix_recovery_read_error(r10_bio);
2132 end_sync_request(r10_bio);
2137 * share the pages with the first bio
2138 * and submit the write request
2140 d = r10_bio->devs[1].devnum;
2141 wbio = r10_bio->devs[1].bio;
2142 wbio2 = r10_bio->devs[1].repl_bio;
2143 /* Need to test wbio2->bi_end_io before we call
2144 * generic_make_request as if the former is NULL,
2145 * the latter is free to free wbio2.
2147 if (wbio2 && !wbio2->bi_end_io)
2149 if (wbio->bi_end_io) {
2150 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2151 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2152 generic_make_request(wbio);
2155 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2156 md_sync_acct(conf->mirrors[d].replacement->bdev,
2157 bio_sectors(wbio2));
2158 generic_make_request(wbio2);
2163 * Used by fix_read_error() to decay the per rdev read_errors.
2164 * We halve the read error count for every hour that has elapsed
2165 * since the last recorded read error.
2168 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2170 struct timespec cur_time_mon;
2171 unsigned long hours_since_last;
2172 unsigned int read_errors = atomic_read(&rdev->read_errors);
2174 ktime_get_ts(&cur_time_mon);
2176 if (rdev->last_read_error.tv_sec == 0 &&
2177 rdev->last_read_error.tv_nsec == 0) {
2178 /* first time we've seen a read error */
2179 rdev->last_read_error = cur_time_mon;
2183 hours_since_last = (cur_time_mon.tv_sec -
2184 rdev->last_read_error.tv_sec) / 3600;
2186 rdev->last_read_error = cur_time_mon;
2189 * if hours_since_last is > the number of bits in read_errors
2190 * just set read errors to 0. We do this to avoid
2191 * overflowing the shift of read_errors by hours_since_last.
2193 if (hours_since_last >= 8 * sizeof(read_errors))
2194 atomic_set(&rdev->read_errors, 0);
2196 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2199 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2200 int sectors, struct page *page, int rw)
2205 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2206 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2208 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2212 set_bit(WriteErrorSeen, &rdev->flags);
2213 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2214 set_bit(MD_RECOVERY_NEEDED,
2215 &rdev->mddev->recovery);
2217 /* need to record an error - either for the block or the device */
2218 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2219 md_error(rdev->mddev, rdev);
2224 * This is a kernel thread which:
2226 * 1. Retries failed read operations on working mirrors.
2227 * 2. Updates the raid superblock when problems encounter.
2228 * 3. Performs writes following reads for array synchronising.
2231 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2233 int sect = 0; /* Offset from r10_bio->sector */
2234 int sectors = r10_bio->sectors;
2235 struct md_rdev*rdev;
2236 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2237 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2239 /* still own a reference to this rdev, so it cannot
2240 * have been cleared recently.
2242 rdev = conf->mirrors[d].rdev;
2244 if (test_bit(Faulty, &rdev->flags))
2245 /* drive has already been failed, just ignore any
2246 more fix_read_error() attempts */
2249 check_decay_read_errors(mddev, rdev);
2250 atomic_inc(&rdev->read_errors);
2251 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2252 char b[BDEVNAME_SIZE];
2253 bdevname(rdev->bdev, b);
2256 "md/raid10:%s: %s: Raid device exceeded "
2257 "read_error threshold [cur %d:max %d]\n",
2259 atomic_read(&rdev->read_errors), max_read_errors);
2261 "md/raid10:%s: %s: Failing raid device\n",
2263 md_error(mddev, conf->mirrors[d].rdev);
2264 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2270 int sl = r10_bio->read_slot;
2274 if (s > (PAGE_SIZE>>9))
2282 d = r10_bio->devs[sl].devnum;
2283 rdev = rcu_dereference(conf->mirrors[d].rdev);
2285 test_bit(In_sync, &rdev->flags) &&
2286 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2287 &first_bad, &bad_sectors) == 0) {
2288 atomic_inc(&rdev->nr_pending);
2290 success = sync_page_io(rdev,
2291 r10_bio->devs[sl].addr +
2294 conf->tmppage, READ, false);
2295 rdev_dec_pending(rdev, mddev);
2301 if (sl == conf->copies)
2303 } while (!success && sl != r10_bio->read_slot);
2307 /* Cannot read from anywhere, just mark the block
2308 * as bad on the first device to discourage future
2311 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2312 rdev = conf->mirrors[dn].rdev;
2314 if (!rdev_set_badblocks(
2316 r10_bio->devs[r10_bio->read_slot].addr
2319 md_error(mddev, rdev);
2320 r10_bio->devs[r10_bio->read_slot].bio
2327 /* write it back and re-read */
2329 while (sl != r10_bio->read_slot) {
2330 char b[BDEVNAME_SIZE];
2335 d = r10_bio->devs[sl].devnum;
2336 rdev = rcu_dereference(conf->mirrors[d].rdev);
2338 !test_bit(In_sync, &rdev->flags))
2341 atomic_inc(&rdev->nr_pending);
2343 if (r10_sync_page_io(rdev,
2344 r10_bio->devs[sl].addr +
2346 s, conf->tmppage, WRITE)
2348 /* Well, this device is dead */
2350 "md/raid10:%s: read correction "
2352 " (%d sectors at %llu on %s)\n",
2354 (unsigned long long)(
2356 choose_data_offset(r10_bio,
2358 bdevname(rdev->bdev, b));
2359 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2362 bdevname(rdev->bdev, b));
2364 rdev_dec_pending(rdev, mddev);
2368 while (sl != r10_bio->read_slot) {
2369 char b[BDEVNAME_SIZE];
2374 d = r10_bio->devs[sl].devnum;
2375 rdev = rcu_dereference(conf->mirrors[d].rdev);
2377 !test_bit(In_sync, &rdev->flags))
2380 atomic_inc(&rdev->nr_pending);
2382 switch (r10_sync_page_io(rdev,
2383 r10_bio->devs[sl].addr +
2388 /* Well, this device is dead */
2390 "md/raid10:%s: unable to read back "
2392 " (%d sectors at %llu on %s)\n",
2394 (unsigned long long)(
2396 choose_data_offset(r10_bio, rdev)),
2397 bdevname(rdev->bdev, b));
2398 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2401 bdevname(rdev->bdev, b));
2405 "md/raid10:%s: read error corrected"
2406 " (%d sectors at %llu on %s)\n",
2408 (unsigned long long)(
2410 choose_data_offset(r10_bio, rdev)),
2411 bdevname(rdev->bdev, b));
2412 atomic_add(s, &rdev->corrected_errors);
2415 rdev_dec_pending(rdev, mddev);
2425 static int narrow_write_error(struct r10bio *r10_bio, int i)
2427 struct bio *bio = r10_bio->master_bio;
2428 struct mddev *mddev = r10_bio->mddev;
2429 struct r10conf *conf = mddev->private;
2430 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2431 /* bio has the data to be written to slot 'i' where
2432 * we just recently had a write error.
2433 * We repeatedly clone the bio and trim down to one block,
2434 * then try the write. Where the write fails we record
2436 * It is conceivable that the bio doesn't exactly align with
2437 * blocks. We must handle this.
2439 * We currently own a reference to the rdev.
2445 int sect_to_write = r10_bio->sectors;
2448 if (rdev->badblocks.shift < 0)
2451 block_sectors = roundup(1 << rdev->badblocks.shift,
2452 bdev_logical_block_size(rdev->bdev) >> 9);
2453 sector = r10_bio->sector;
2454 sectors = ((r10_bio->sector + block_sectors)
2455 & ~(sector_t)(block_sectors - 1))
2458 while (sect_to_write) {
2460 if (sectors > sect_to_write)
2461 sectors = sect_to_write;
2462 /* Write at 'sector' for 'sectors' */
2463 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2464 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2465 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2466 choose_data_offset(r10_bio, rdev) +
2467 (sector - r10_bio->sector));
2468 wbio->bi_bdev = rdev->bdev;
2469 if (submit_bio_wait(WRITE, wbio) == 0)
2471 ok = rdev_set_badblocks(rdev, sector,
2476 sect_to_write -= sectors;
2478 sectors = block_sectors;
2483 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2485 int slot = r10_bio->read_slot;
2487 struct r10conf *conf = mddev->private;
2488 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2489 char b[BDEVNAME_SIZE];
2490 unsigned long do_sync;
2493 /* we got a read error. Maybe the drive is bad. Maybe just
2494 * the block and we can fix it.
2495 * We freeze all other IO, and try reading the block from
2496 * other devices. When we find one, we re-write
2497 * and check it that fixes the read error.
2498 * This is all done synchronously while the array is
2501 bio = r10_bio->devs[slot].bio;
2502 bdevname(bio->bi_bdev, b);
2504 r10_bio->devs[slot].bio = NULL;
2506 if (mddev->ro == 0) {
2507 freeze_array(conf, 1);
2508 fix_read_error(conf, mddev, r10_bio);
2509 unfreeze_array(conf);
2511 r10_bio->devs[slot].bio = IO_BLOCKED;
2513 rdev_dec_pending(rdev, mddev);
2516 rdev = read_balance(conf, r10_bio, &max_sectors);
2518 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2519 " read error for block %llu\n",
2521 (unsigned long long)r10_bio->sector);
2522 raid_end_bio_io(r10_bio);
2526 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2527 slot = r10_bio->read_slot;
2530 "md/raid10:%s: %s: redirecting "
2531 "sector %llu to another mirror\n",
2533 bdevname(rdev->bdev, b),
2534 (unsigned long long)r10_bio->sector);
2535 bio = bio_clone_mddev(r10_bio->master_bio,
2537 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2538 r10_bio->devs[slot].bio = bio;
2539 r10_bio->devs[slot].rdev = rdev;
2540 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2541 + choose_data_offset(r10_bio, rdev);
2542 bio->bi_bdev = rdev->bdev;
2543 bio->bi_rw = READ | do_sync;
2544 bio->bi_private = r10_bio;
2545 bio->bi_end_io = raid10_end_read_request;
2546 if (max_sectors < r10_bio->sectors) {
2547 /* Drat - have to split this up more */
2548 struct bio *mbio = r10_bio->master_bio;
2549 int sectors_handled =
2550 r10_bio->sector + max_sectors
2551 - mbio->bi_iter.bi_sector;
2552 r10_bio->sectors = max_sectors;
2553 spin_lock_irq(&conf->device_lock);
2554 if (mbio->bi_phys_segments == 0)
2555 mbio->bi_phys_segments = 2;
2557 mbio->bi_phys_segments++;
2558 spin_unlock_irq(&conf->device_lock);
2559 generic_make_request(bio);
2561 r10_bio = mempool_alloc(conf->r10bio_pool,
2563 r10_bio->master_bio = mbio;
2564 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2566 set_bit(R10BIO_ReadError,
2568 r10_bio->mddev = mddev;
2569 r10_bio->sector = mbio->bi_iter.bi_sector
2574 generic_make_request(bio);
2577 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2579 /* Some sort of write request has finished and it
2580 * succeeded in writing where we thought there was a
2581 * bad block. So forget the bad block.
2582 * Or possibly if failed and we need to record
2586 struct md_rdev *rdev;
2588 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2589 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2590 for (m = 0; m < conf->copies; m++) {
2591 int dev = r10_bio->devs[m].devnum;
2592 rdev = conf->mirrors[dev].rdev;
2593 if (r10_bio->devs[m].bio == NULL)
2595 if (!r10_bio->devs[m].bio->bi_error) {
2596 rdev_clear_badblocks(
2598 r10_bio->devs[m].addr,
2599 r10_bio->sectors, 0);
2601 if (!rdev_set_badblocks(
2603 r10_bio->devs[m].addr,
2604 r10_bio->sectors, 0))
2605 md_error(conf->mddev, rdev);
2607 rdev = conf->mirrors[dev].replacement;
2608 if (r10_bio->devs[m].repl_bio == NULL)
2611 if (!r10_bio->devs[m].repl_bio->bi_error) {
2612 rdev_clear_badblocks(
2614 r10_bio->devs[m].addr,
2615 r10_bio->sectors, 0);
2617 if (!rdev_set_badblocks(
2619 r10_bio->devs[m].addr,
2620 r10_bio->sectors, 0))
2621 md_error(conf->mddev, rdev);
2626 for (m = 0; m < conf->copies; m++) {
2627 int dev = r10_bio->devs[m].devnum;
2628 struct bio *bio = r10_bio->devs[m].bio;
2629 rdev = conf->mirrors[dev].rdev;
2630 if (bio == IO_MADE_GOOD) {
2631 rdev_clear_badblocks(
2633 r10_bio->devs[m].addr,
2634 r10_bio->sectors, 0);
2635 rdev_dec_pending(rdev, conf->mddev);
2636 } else if (bio != NULL && bio->bi_error) {
2637 if (!narrow_write_error(r10_bio, m)) {
2638 md_error(conf->mddev, rdev);
2639 set_bit(R10BIO_Degraded,
2642 rdev_dec_pending(rdev, conf->mddev);
2644 bio = r10_bio->devs[m].repl_bio;
2645 rdev = conf->mirrors[dev].replacement;
2646 if (rdev && bio == IO_MADE_GOOD) {
2647 rdev_clear_badblocks(
2649 r10_bio->devs[m].addr,
2650 r10_bio->sectors, 0);
2651 rdev_dec_pending(rdev, conf->mddev);
2654 if (test_bit(R10BIO_WriteError,
2656 close_write(r10_bio);
2657 raid_end_bio_io(r10_bio);
2661 static void raid10d(struct md_thread *thread)
2663 struct mddev *mddev = thread->mddev;
2664 struct r10bio *r10_bio;
2665 unsigned long flags;
2666 struct r10conf *conf = mddev->private;
2667 struct list_head *head = &conf->retry_list;
2668 struct blk_plug plug;
2670 md_check_recovery(mddev);
2672 blk_start_plug(&plug);
2675 flush_pending_writes(conf);
2677 spin_lock_irqsave(&conf->device_lock, flags);
2678 if (list_empty(head)) {
2679 spin_unlock_irqrestore(&conf->device_lock, flags);
2682 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2683 list_del(head->prev);
2685 spin_unlock_irqrestore(&conf->device_lock, flags);
2687 mddev = r10_bio->mddev;
2688 conf = mddev->private;
2689 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2690 test_bit(R10BIO_WriteError, &r10_bio->state))
2691 handle_write_completed(conf, r10_bio);
2692 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2693 reshape_request_write(mddev, r10_bio);
2694 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2695 sync_request_write(mddev, r10_bio);
2696 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2697 recovery_request_write(mddev, r10_bio);
2698 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2699 handle_read_error(mddev, r10_bio);
2701 /* just a partial read to be scheduled from a
2704 int slot = r10_bio->read_slot;
2705 generic_make_request(r10_bio->devs[slot].bio);
2709 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2710 md_check_recovery(mddev);
2712 blk_finish_plug(&plug);
2715 static int init_resync(struct r10conf *conf)
2720 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2721 BUG_ON(conf->r10buf_pool);
2722 conf->have_replacement = 0;
2723 for (i = 0; i < conf->geo.raid_disks; i++)
2724 if (conf->mirrors[i].replacement)
2725 conf->have_replacement = 1;
2726 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2727 if (!conf->r10buf_pool)
2729 conf->next_resync = 0;
2734 * perform a "sync" on one "block"
2736 * We need to make sure that no normal I/O request - particularly write
2737 * requests - conflict with active sync requests.
2739 * This is achieved by tracking pending requests and a 'barrier' concept
2740 * that can be installed to exclude normal IO requests.
2742 * Resync and recovery are handled very differently.
2743 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2745 * For resync, we iterate over virtual addresses, read all copies,
2746 * and update if there are differences. If only one copy is live,
2748 * For recovery, we iterate over physical addresses, read a good
2749 * value for each non-in_sync drive, and over-write.
2751 * So, for recovery we may have several outstanding complex requests for a
2752 * given address, one for each out-of-sync device. We model this by allocating
2753 * a number of r10_bio structures, one for each out-of-sync device.
2754 * As we setup these structures, we collect all bio's together into a list
2755 * which we then process collectively to add pages, and then process again
2756 * to pass to generic_make_request.
2758 * The r10_bio structures are linked using a borrowed master_bio pointer.
2759 * This link is counted in ->remaining. When the r10_bio that points to NULL
2760 * has its remaining count decremented to 0, the whole complex operation
2765 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2768 struct r10conf *conf = mddev->private;
2769 struct r10bio *r10_bio;
2770 struct bio *biolist = NULL, *bio;
2771 sector_t max_sector, nr_sectors;
2774 sector_t sync_blocks;
2775 sector_t sectors_skipped = 0;
2776 int chunks_skipped = 0;
2777 sector_t chunk_mask = conf->geo.chunk_mask;
2779 if (!conf->r10buf_pool)
2780 if (init_resync(conf))
2784 * Allow skipping a full rebuild for incremental assembly
2785 * of a clean array, like RAID1 does.
2787 if (mddev->bitmap == NULL &&
2788 mddev->recovery_cp == MaxSector &&
2789 mddev->reshape_position == MaxSector &&
2790 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2791 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2792 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2793 conf->fullsync == 0) {
2795 return mddev->dev_sectors - sector_nr;
2799 max_sector = mddev->dev_sectors;
2800 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2801 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2802 max_sector = mddev->resync_max_sectors;
2803 if (sector_nr >= max_sector) {
2804 /* If we aborted, we need to abort the
2805 * sync on the 'current' bitmap chucks (there can
2806 * be several when recovering multiple devices).
2807 * as we may have started syncing it but not finished.
2808 * We can find the current address in
2809 * mddev->curr_resync, but for recovery,
2810 * we need to convert that to several
2811 * virtual addresses.
2813 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2819 if (mddev->curr_resync < max_sector) { /* aborted */
2820 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2821 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2823 else for (i = 0; i < conf->geo.raid_disks; i++) {
2825 raid10_find_virt(conf, mddev->curr_resync, i);
2826 bitmap_end_sync(mddev->bitmap, sect,
2830 /* completed sync */
2831 if ((!mddev->bitmap || conf->fullsync)
2832 && conf->have_replacement
2833 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2834 /* Completed a full sync so the replacements
2835 * are now fully recovered.
2837 for (i = 0; i < conf->geo.raid_disks; i++)
2838 if (conf->mirrors[i].replacement)
2839 conf->mirrors[i].replacement
2845 bitmap_close_sync(mddev->bitmap);
2848 return sectors_skipped;
2851 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2852 return reshape_request(mddev, sector_nr, skipped);
2854 if (chunks_skipped >= conf->geo.raid_disks) {
2855 /* if there has been nothing to do on any drive,
2856 * then there is nothing to do at all..
2859 return (max_sector - sector_nr) + sectors_skipped;
2862 if (max_sector > mddev->resync_max)
2863 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2865 /* make sure whole request will fit in a chunk - if chunks
2868 if (conf->geo.near_copies < conf->geo.raid_disks &&
2869 max_sector > (sector_nr | chunk_mask))
2870 max_sector = (sector_nr | chunk_mask) + 1;
2872 /* Again, very different code for resync and recovery.
2873 * Both must result in an r10bio with a list of bios that
2874 * have bi_end_io, bi_sector, bi_bdev set,
2875 * and bi_private set to the r10bio.
2876 * For recovery, we may actually create several r10bios
2877 * with 2 bios in each, that correspond to the bios in the main one.
2878 * In this case, the subordinate r10bios link back through a
2879 * borrowed master_bio pointer, and the counter in the master
2880 * includes a ref from each subordinate.
2882 /* First, we decide what to do and set ->bi_end_io
2883 * To end_sync_read if we want to read, and
2884 * end_sync_write if we will want to write.
2887 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2888 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2889 /* recovery... the complicated one */
2893 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2899 struct raid10_info *mirror = &conf->mirrors[i];
2901 if ((mirror->rdev == NULL ||
2902 test_bit(In_sync, &mirror->rdev->flags))
2904 (mirror->replacement == NULL ||
2906 &mirror->replacement->flags)))
2910 /* want to reconstruct this device */
2912 sect = raid10_find_virt(conf, sector_nr, i);
2913 if (sect >= mddev->resync_max_sectors) {
2914 /* last stripe is not complete - don't
2915 * try to recover this sector.
2919 /* Unless we are doing a full sync, or a replacement
2920 * we only need to recover the block if it is set in
2923 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2925 if (sync_blocks < max_sync)
2926 max_sync = sync_blocks;
2928 mirror->replacement == NULL &&
2930 /* yep, skip the sync_blocks here, but don't assume
2931 * that there will never be anything to do here
2933 chunks_skipped = -1;
2937 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2939 raise_barrier(conf, rb2 != NULL);
2940 atomic_set(&r10_bio->remaining, 0);
2942 r10_bio->master_bio = (struct bio*)rb2;
2944 atomic_inc(&rb2->remaining);
2945 r10_bio->mddev = mddev;
2946 set_bit(R10BIO_IsRecover, &r10_bio->state);
2947 r10_bio->sector = sect;
2949 raid10_find_phys(conf, r10_bio);
2951 /* Need to check if the array will still be
2954 for (j = 0; j < conf->geo.raid_disks; j++)
2955 if (conf->mirrors[j].rdev == NULL ||
2956 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2961 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2962 &sync_blocks, still_degraded);
2965 for (j=0; j<conf->copies;j++) {
2967 int d = r10_bio->devs[j].devnum;
2968 sector_t from_addr, to_addr;
2969 struct md_rdev *rdev;
2970 sector_t sector, first_bad;
2972 if (!conf->mirrors[d].rdev ||
2973 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2975 /* This is where we read from */
2977 rdev = conf->mirrors[d].rdev;
2978 sector = r10_bio->devs[j].addr;
2980 if (is_badblock(rdev, sector, max_sync,
2981 &first_bad, &bad_sectors)) {
2982 if (first_bad > sector)
2983 max_sync = first_bad - sector;
2985 bad_sectors -= (sector
2987 if (max_sync > bad_sectors)
2988 max_sync = bad_sectors;
2992 bio = r10_bio->devs[0].bio;
2994 bio->bi_next = biolist;
2996 bio->bi_private = r10_bio;
2997 bio->bi_end_io = end_sync_read;
2999 from_addr = r10_bio->devs[j].addr;
3000 bio->bi_iter.bi_sector = from_addr +
3002 bio->bi_bdev = rdev->bdev;
3003 atomic_inc(&rdev->nr_pending);
3004 /* and we write to 'i' (if not in_sync) */
3006 for (k=0; k<conf->copies; k++)
3007 if (r10_bio->devs[k].devnum == i)
3009 BUG_ON(k == conf->copies);
3010 to_addr = r10_bio->devs[k].addr;
3011 r10_bio->devs[0].devnum = d;
3012 r10_bio->devs[0].addr = from_addr;
3013 r10_bio->devs[1].devnum = i;
3014 r10_bio->devs[1].addr = to_addr;
3016 rdev = mirror->rdev;
3017 if (!test_bit(In_sync, &rdev->flags)) {
3018 bio = r10_bio->devs[1].bio;
3020 bio->bi_next = biolist;
3022 bio->bi_private = r10_bio;
3023 bio->bi_end_io = end_sync_write;
3025 bio->bi_iter.bi_sector = to_addr
3026 + rdev->data_offset;
3027 bio->bi_bdev = rdev->bdev;
3028 atomic_inc(&r10_bio->remaining);
3030 r10_bio->devs[1].bio->bi_end_io = NULL;
3032 /* and maybe write to replacement */
3033 bio = r10_bio->devs[1].repl_bio;
3035 bio->bi_end_io = NULL;
3036 rdev = mirror->replacement;
3037 /* Note: if rdev != NULL, then bio
3038 * cannot be NULL as r10buf_pool_alloc will
3039 * have allocated it.
3040 * So the second test here is pointless.
3041 * But it keeps semantic-checkers happy, and
3042 * this comment keeps human reviewers
3045 if (rdev == NULL || bio == NULL ||
3046 test_bit(Faulty, &rdev->flags))
3049 bio->bi_next = biolist;
3051 bio->bi_private = r10_bio;
3052 bio->bi_end_io = end_sync_write;
3054 bio->bi_iter.bi_sector = to_addr +
3056 bio->bi_bdev = rdev->bdev;
3057 atomic_inc(&r10_bio->remaining);
3060 if (j == conf->copies) {
3061 /* Cannot recover, so abort the recovery or
3062 * record a bad block */
3064 /* problem is that there are bad blocks
3065 * on other device(s)
3068 for (k = 0; k < conf->copies; k++)
3069 if (r10_bio->devs[k].devnum == i)
3071 if (!test_bit(In_sync,
3072 &mirror->rdev->flags)
3073 && !rdev_set_badblocks(
3075 r10_bio->devs[k].addr,
3078 if (mirror->replacement &&
3079 !rdev_set_badblocks(
3080 mirror->replacement,
3081 r10_bio->devs[k].addr,
3086 if (!test_and_set_bit(MD_RECOVERY_INTR,
3088 printk(KERN_INFO "md/raid10:%s: insufficient "
3089 "working devices for recovery.\n",
3091 mirror->recovery_disabled
3092 = mddev->recovery_disabled;
3096 atomic_dec(&rb2->remaining);
3101 if (biolist == NULL) {
3103 struct r10bio *rb2 = r10_bio;
3104 r10_bio = (struct r10bio*) rb2->master_bio;
3105 rb2->master_bio = NULL;
3111 /* resync. Schedule a read for every block at this virt offset */
3114 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3116 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3117 &sync_blocks, mddev->degraded) &&
3118 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3119 &mddev->recovery)) {
3120 /* We can skip this block */
3122 return sync_blocks + sectors_skipped;
3124 if (sync_blocks < max_sync)
3125 max_sync = sync_blocks;
3126 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3129 r10_bio->mddev = mddev;
3130 atomic_set(&r10_bio->remaining, 0);
3131 raise_barrier(conf, 0);
3132 conf->next_resync = sector_nr;
3134 r10_bio->master_bio = NULL;
3135 r10_bio->sector = sector_nr;
3136 set_bit(R10BIO_IsSync, &r10_bio->state);
3137 raid10_find_phys(conf, r10_bio);
3138 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3140 for (i = 0; i < conf->copies; i++) {
3141 int d = r10_bio->devs[i].devnum;
3142 sector_t first_bad, sector;
3145 if (r10_bio->devs[i].repl_bio)
3146 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3148 bio = r10_bio->devs[i].bio;
3150 bio->bi_error = -EIO;
3151 if (conf->mirrors[d].rdev == NULL ||
3152 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3154 sector = r10_bio->devs[i].addr;
3155 if (is_badblock(conf->mirrors[d].rdev,
3157 &first_bad, &bad_sectors)) {
3158 if (first_bad > sector)
3159 max_sync = first_bad - sector;
3161 bad_sectors -= (sector - first_bad);
3162 if (max_sync > bad_sectors)
3163 max_sync = bad_sectors;
3167 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3168 atomic_inc(&r10_bio->remaining);
3169 bio->bi_next = biolist;
3171 bio->bi_private = r10_bio;
3172 bio->bi_end_io = end_sync_read;
3174 bio->bi_iter.bi_sector = sector +
3175 conf->mirrors[d].rdev->data_offset;
3176 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3179 if (conf->mirrors[d].replacement == NULL ||
3181 &conf->mirrors[d].replacement->flags))
3184 /* Need to set up for writing to the replacement */
3185 bio = r10_bio->devs[i].repl_bio;
3187 bio->bi_error = -EIO;
3189 sector = r10_bio->devs[i].addr;
3190 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3191 bio->bi_next = biolist;
3193 bio->bi_private = r10_bio;
3194 bio->bi_end_io = end_sync_write;
3196 bio->bi_iter.bi_sector = sector +
3197 conf->mirrors[d].replacement->data_offset;
3198 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3203 for (i=0; i<conf->copies; i++) {
3204 int d = r10_bio->devs[i].devnum;
3205 if (r10_bio->devs[i].bio->bi_end_io)
3206 rdev_dec_pending(conf->mirrors[d].rdev,
3208 if (r10_bio->devs[i].repl_bio &&
3209 r10_bio->devs[i].repl_bio->bi_end_io)
3211 conf->mirrors[d].replacement,
3221 if (sector_nr + max_sync < max_sector)
3222 max_sector = sector_nr + max_sync;
3225 int len = PAGE_SIZE;
3226 if (sector_nr + (len>>9) > max_sector)
3227 len = (max_sector - sector_nr) << 9;
3230 for (bio= biolist ; bio ; bio=bio->bi_next) {
3232 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3233 if (bio_add_page(bio, page, len, 0))
3237 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3238 for (bio2 = biolist;
3239 bio2 && bio2 != bio;
3240 bio2 = bio2->bi_next) {
3241 /* remove last page from this bio */
3243 bio2->bi_iter.bi_size -= len;
3244 bio_clear_flag(bio2, BIO_SEG_VALID);
3248 nr_sectors += len>>9;
3249 sector_nr += len>>9;
3250 } while (biolist->bi_vcnt < RESYNC_PAGES);
3252 r10_bio->sectors = nr_sectors;
3256 biolist = biolist->bi_next;
3258 bio->bi_next = NULL;
3259 r10_bio = bio->bi_private;
3260 r10_bio->sectors = nr_sectors;
3262 if (bio->bi_end_io == end_sync_read) {
3263 md_sync_acct(bio->bi_bdev, nr_sectors);
3265 generic_make_request(bio);
3269 if (sectors_skipped)
3270 /* pretend they weren't skipped, it makes
3271 * no important difference in this case
3273 md_done_sync(mddev, sectors_skipped, 1);
3275 return sectors_skipped + nr_sectors;
3277 /* There is nowhere to write, so all non-sync
3278 * drives must be failed or in resync, all drives
3279 * have a bad block, so try the next chunk...
3281 if (sector_nr + max_sync < max_sector)
3282 max_sector = sector_nr + max_sync;
3284 sectors_skipped += (max_sector - sector_nr);
3286 sector_nr = max_sector;
3291 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3294 struct r10conf *conf = mddev->private;
3297 raid_disks = min(conf->geo.raid_disks,
3298 conf->prev.raid_disks);
3300 sectors = conf->dev_sectors;
3302 size = sectors >> conf->geo.chunk_shift;
3303 sector_div(size, conf->geo.far_copies);
3304 size = size * raid_disks;
3305 sector_div(size, conf->geo.near_copies);
3307 return size << conf->geo.chunk_shift;
3310 static void calc_sectors(struct r10conf *conf, sector_t size)
3312 /* Calculate the number of sectors-per-device that will
3313 * actually be used, and set conf->dev_sectors and
3317 size = size >> conf->geo.chunk_shift;
3318 sector_div(size, conf->geo.far_copies);
3319 size = size * conf->geo.raid_disks;
3320 sector_div(size, conf->geo.near_copies);
3321 /* 'size' is now the number of chunks in the array */
3322 /* calculate "used chunks per device" */
3323 size = size * conf->copies;
3325 /* We need to round up when dividing by raid_disks to
3326 * get the stride size.
3328 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3330 conf->dev_sectors = size << conf->geo.chunk_shift;
3332 if (conf->geo.far_offset)
3333 conf->geo.stride = 1 << conf->geo.chunk_shift;
3335 sector_div(size, conf->geo.far_copies);
3336 conf->geo.stride = size << conf->geo.chunk_shift;
3340 enum geo_type {geo_new, geo_old, geo_start};
3341 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3344 int layout, chunk, disks;
3347 layout = mddev->layout;
3348 chunk = mddev->chunk_sectors;
3349 disks = mddev->raid_disks - mddev->delta_disks;
3352 layout = mddev->new_layout;
3353 chunk = mddev->new_chunk_sectors;
3354 disks = mddev->raid_disks;
3356 default: /* avoid 'may be unused' warnings */
3357 case geo_start: /* new when starting reshape - raid_disks not
3359 layout = mddev->new_layout;
3360 chunk = mddev->new_chunk_sectors;
3361 disks = mddev->raid_disks + mddev->delta_disks;
3366 if (chunk < (PAGE_SIZE >> 9) ||
3367 !is_power_of_2(chunk))
3370 fc = (layout >> 8) & 255;
3371 fo = layout & (1<<16);
3372 geo->raid_disks = disks;
3373 geo->near_copies = nc;
3374 geo->far_copies = fc;
3375 geo->far_offset = fo;
3376 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3377 geo->chunk_mask = chunk - 1;
3378 geo->chunk_shift = ffz(~chunk);
3382 static struct r10conf *setup_conf(struct mddev *mddev)
3384 struct r10conf *conf = NULL;
3389 copies = setup_geo(&geo, mddev, geo_new);
3392 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3393 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3394 mdname(mddev), PAGE_SIZE);
3398 if (copies < 2 || copies > mddev->raid_disks) {
3399 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3400 mdname(mddev), mddev->new_layout);
3405 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3409 /* FIXME calc properly */
3410 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3411 max(0,-mddev->delta_disks)),
3416 conf->tmppage = alloc_page(GFP_KERNEL);
3421 conf->copies = copies;
3422 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3423 r10bio_pool_free, conf);
3424 if (!conf->r10bio_pool)
3427 calc_sectors(conf, mddev->dev_sectors);
3428 if (mddev->reshape_position == MaxSector) {
3429 conf->prev = conf->geo;
3430 conf->reshape_progress = MaxSector;
3432 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3436 conf->reshape_progress = mddev->reshape_position;
3437 if (conf->prev.far_offset)
3438 conf->prev.stride = 1 << conf->prev.chunk_shift;
3440 /* far_copies must be 1 */
3441 conf->prev.stride = conf->dev_sectors;
3443 conf->reshape_safe = conf->reshape_progress;
3444 spin_lock_init(&conf->device_lock);
3445 INIT_LIST_HEAD(&conf->retry_list);
3447 spin_lock_init(&conf->resync_lock);
3448 init_waitqueue_head(&conf->wait_barrier);
3450 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3454 conf->mddev = mddev;
3459 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3462 if (conf->r10bio_pool)
3463 mempool_destroy(conf->r10bio_pool);
3464 kfree(conf->mirrors);
3465 safe_put_page(conf->tmppage);
3468 return ERR_PTR(err);
3471 static int run(struct mddev *mddev)
3473 struct r10conf *conf;
3474 int i, disk_idx, chunk_size;
3475 struct raid10_info *disk;
3476 struct md_rdev *rdev;
3478 sector_t min_offset_diff = 0;
3480 bool discard_supported = false;
3482 if (mddev->private == NULL) {
3483 conf = setup_conf(mddev);
3485 return PTR_ERR(conf);
3486 mddev->private = conf;
3488 conf = mddev->private;
3492 mddev->thread = conf->thread;
3493 conf->thread = NULL;
3495 chunk_size = mddev->chunk_sectors << 9;
3497 blk_queue_max_discard_sectors(mddev->queue,
3498 mddev->chunk_sectors);
3499 blk_queue_max_write_same_sectors(mddev->queue, 0);
3500 blk_queue_io_min(mddev->queue, chunk_size);
3501 if (conf->geo.raid_disks % conf->geo.near_copies)
3502 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3504 blk_queue_io_opt(mddev->queue, chunk_size *
3505 (conf->geo.raid_disks / conf->geo.near_copies));
3508 rdev_for_each(rdev, mddev) {
3510 struct request_queue *q;
3512 disk_idx = rdev->raid_disk;
3515 if (disk_idx >= conf->geo.raid_disks &&
3516 disk_idx >= conf->prev.raid_disks)
3518 disk = conf->mirrors + disk_idx;
3520 if (test_bit(Replacement, &rdev->flags)) {
3521 if (disk->replacement)
3523 disk->replacement = rdev;
3529 q = bdev_get_queue(rdev->bdev);
3530 diff = (rdev->new_data_offset - rdev->data_offset);
3531 if (!mddev->reshape_backwards)
3535 if (first || diff < min_offset_diff)
3536 min_offset_diff = diff;
3539 disk_stack_limits(mddev->gendisk, rdev->bdev,
3540 rdev->data_offset << 9);
3542 disk->head_position = 0;
3544 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3545 discard_supported = true;
3549 if (discard_supported)
3550 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3553 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3556 /* need to check that every block has at least one working mirror */
3557 if (!enough(conf, -1)) {
3558 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3563 if (conf->reshape_progress != MaxSector) {
3564 /* must ensure that shape change is supported */
3565 if (conf->geo.far_copies != 1 &&
3566 conf->geo.far_offset == 0)
3568 if (conf->prev.far_copies != 1 &&
3569 conf->prev.far_offset == 0)
3573 mddev->degraded = 0;
3575 i < conf->geo.raid_disks
3576 || i < conf->prev.raid_disks;
3579 disk = conf->mirrors + i;
3581 if (!disk->rdev && disk->replacement) {
3582 /* The replacement is all we have - use it */
3583 disk->rdev = disk->replacement;
3584 disk->replacement = NULL;
3585 clear_bit(Replacement, &disk->rdev->flags);
3589 !test_bit(In_sync, &disk->rdev->flags)) {
3590 disk->head_position = 0;
3593 disk->rdev->saved_raid_disk < 0)
3596 disk->recovery_disabled = mddev->recovery_disabled - 1;
3599 if (mddev->recovery_cp != MaxSector)
3600 printk(KERN_NOTICE "md/raid10:%s: not clean"
3601 " -- starting background reconstruction\n",
3604 "md/raid10:%s: active with %d out of %d devices\n",
3605 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3606 conf->geo.raid_disks);
3608 * Ok, everything is just fine now
3610 mddev->dev_sectors = conf->dev_sectors;
3611 size = raid10_size(mddev, 0, 0);
3612 md_set_array_sectors(mddev, size);
3613 mddev->resync_max_sectors = size;
3616 int stripe = conf->geo.raid_disks *
3617 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3619 /* Calculate max read-ahead size.
3620 * We need to readahead at least twice a whole stripe....
3623 stripe /= conf->geo.near_copies;
3624 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3625 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3628 if (md_integrity_register(mddev))
3631 if (conf->reshape_progress != MaxSector) {
3632 unsigned long before_length, after_length;
3634 before_length = ((1 << conf->prev.chunk_shift) *
3635 conf->prev.far_copies);
3636 after_length = ((1 << conf->geo.chunk_shift) *
3637 conf->geo.far_copies);
3639 if (max(before_length, after_length) > min_offset_diff) {
3640 /* This cannot work */
3641 printk("md/raid10: offset difference not enough to continue reshape\n");
3644 conf->offset_diff = min_offset_diff;
3646 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3647 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3648 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3649 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3650 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3657 md_unregister_thread(&mddev->thread);
3658 if (conf->r10bio_pool)
3659 mempool_destroy(conf->r10bio_pool);
3660 safe_put_page(conf->tmppage);
3661 kfree(conf->mirrors);
3663 mddev->private = NULL;
3668 static void raid10_free(struct mddev *mddev, void *priv)
3670 struct r10conf *conf = priv;
3672 if (conf->r10bio_pool)
3673 mempool_destroy(conf->r10bio_pool);
3674 safe_put_page(conf->tmppage);
3675 kfree(conf->mirrors);
3676 kfree(conf->mirrors_old);
3677 kfree(conf->mirrors_new);
3681 static void raid10_quiesce(struct mddev *mddev, int state)
3683 struct r10conf *conf = mddev->private;
3687 raise_barrier(conf, 0);
3690 lower_barrier(conf);
3695 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3697 /* Resize of 'far' arrays is not supported.
3698 * For 'near' and 'offset' arrays we can set the
3699 * number of sectors used to be an appropriate multiple
3700 * of the chunk size.
3701 * For 'offset', this is far_copies*chunksize.
3702 * For 'near' the multiplier is the LCM of
3703 * near_copies and raid_disks.
3704 * So if far_copies > 1 && !far_offset, fail.
3705 * Else find LCM(raid_disks, near_copy)*far_copies and
3706 * multiply by chunk_size. Then round to this number.
3707 * This is mostly done by raid10_size()
3709 struct r10conf *conf = mddev->private;
3710 sector_t oldsize, size;
3712 if (mddev->reshape_position != MaxSector)
3715 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3718 oldsize = raid10_size(mddev, 0, 0);
3719 size = raid10_size(mddev, sectors, 0);
3720 if (mddev->external_size &&
3721 mddev->array_sectors > size)
3723 if (mddev->bitmap) {
3724 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3728 md_set_array_sectors(mddev, size);
3729 set_capacity(mddev->gendisk, mddev->array_sectors);
3730 revalidate_disk(mddev->gendisk);
3731 if (sectors > mddev->dev_sectors &&
3732 mddev->recovery_cp > oldsize) {
3733 mddev->recovery_cp = oldsize;
3734 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3736 calc_sectors(conf, sectors);
3737 mddev->dev_sectors = conf->dev_sectors;
3738 mddev->resync_max_sectors = size;
3742 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3744 struct md_rdev *rdev;
3745 struct r10conf *conf;
3747 if (mddev->degraded > 0) {
3748 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3750 return ERR_PTR(-EINVAL);
3752 sector_div(size, devs);
3754 /* Set new parameters */
3755 mddev->new_level = 10;
3756 /* new layout: far_copies = 1, near_copies = 2 */
3757 mddev->new_layout = (1<<8) + 2;
3758 mddev->new_chunk_sectors = mddev->chunk_sectors;
3759 mddev->delta_disks = mddev->raid_disks;
3760 mddev->raid_disks *= 2;
3761 /* make sure it will be not marked as dirty */
3762 mddev->recovery_cp = MaxSector;
3763 mddev->dev_sectors = size;
3765 conf = setup_conf(mddev);
3766 if (!IS_ERR(conf)) {
3767 rdev_for_each(rdev, mddev)
3768 if (rdev->raid_disk >= 0) {
3769 rdev->new_raid_disk = rdev->raid_disk * 2;
3770 rdev->sectors = size;
3778 static void *raid10_takeover(struct mddev *mddev)
3780 struct r0conf *raid0_conf;
3782 /* raid10 can take over:
3783 * raid0 - providing it has only two drives
3785 if (mddev->level == 0) {
3786 /* for raid0 takeover only one zone is supported */
3787 raid0_conf = mddev->private;
3788 if (raid0_conf->nr_strip_zones > 1) {
3789 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3790 " with more than one zone.\n",
3792 return ERR_PTR(-EINVAL);
3794 return raid10_takeover_raid0(mddev,
3795 raid0_conf->strip_zone->zone_end,
3796 raid0_conf->strip_zone->nb_dev);
3798 return ERR_PTR(-EINVAL);
3801 static int raid10_check_reshape(struct mddev *mddev)
3803 /* Called when there is a request to change
3804 * - layout (to ->new_layout)
3805 * - chunk size (to ->new_chunk_sectors)
3806 * - raid_disks (by delta_disks)
3807 * or when trying to restart a reshape that was ongoing.
3809 * We need to validate the request and possibly allocate
3810 * space if that might be an issue later.
3812 * Currently we reject any reshape of a 'far' mode array,
3813 * allow chunk size to change if new is generally acceptable,
3814 * allow raid_disks to increase, and allow
3815 * a switch between 'near' mode and 'offset' mode.
3817 struct r10conf *conf = mddev->private;
3820 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3823 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3824 /* mustn't change number of copies */
3826 if (geo.far_copies > 1 && !geo.far_offset)
3827 /* Cannot switch to 'far' mode */
3830 if (mddev->array_sectors & geo.chunk_mask)
3831 /* not factor of array size */
3834 if (!enough(conf, -1))
3837 kfree(conf->mirrors_new);
3838 conf->mirrors_new = NULL;
3839 if (mddev->delta_disks > 0) {
3840 /* allocate new 'mirrors' list */
3841 conf->mirrors_new = kzalloc(
3842 sizeof(struct raid10_info)
3843 *(mddev->raid_disks +
3844 mddev->delta_disks),
3846 if (!conf->mirrors_new)
3853 * Need to check if array has failed when deciding whether to:
3855 * - remove non-faulty devices
3858 * This determination is simple when no reshape is happening.
3859 * However if there is a reshape, we need to carefully check
3860 * both the before and after sections.
3861 * This is because some failed devices may only affect one
3862 * of the two sections, and some non-in_sync devices may
3863 * be insync in the section most affected by failed devices.
3865 static int calc_degraded(struct r10conf *conf)
3867 int degraded, degraded2;
3872 /* 'prev' section first */
3873 for (i = 0; i < conf->prev.raid_disks; i++) {
3874 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3875 if (!rdev || test_bit(Faulty, &rdev->flags))
3877 else if (!test_bit(In_sync, &rdev->flags))
3878 /* When we can reduce the number of devices in
3879 * an array, this might not contribute to
3880 * 'degraded'. It does now.
3885 if (conf->geo.raid_disks == conf->prev.raid_disks)
3889 for (i = 0; i < conf->geo.raid_disks; i++) {
3890 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3891 if (!rdev || test_bit(Faulty, &rdev->flags))
3893 else if (!test_bit(In_sync, &rdev->flags)) {
3894 /* If reshape is increasing the number of devices,
3895 * this section has already been recovered, so
3896 * it doesn't contribute to degraded.
3899 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3904 if (degraded2 > degraded)
3909 static int raid10_start_reshape(struct mddev *mddev)
3911 /* A 'reshape' has been requested. This commits
3912 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3913 * This also checks if there are enough spares and adds them
3915 * We currently require enough spares to make the final
3916 * array non-degraded. We also require that the difference
3917 * between old and new data_offset - on each device - is
3918 * enough that we never risk over-writing.
3921 unsigned long before_length, after_length;
3922 sector_t min_offset_diff = 0;
3925 struct r10conf *conf = mddev->private;
3926 struct md_rdev *rdev;
3930 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3933 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3936 before_length = ((1 << conf->prev.chunk_shift) *
3937 conf->prev.far_copies);
3938 after_length = ((1 << conf->geo.chunk_shift) *
3939 conf->geo.far_copies);
3941 rdev_for_each(rdev, mddev) {
3942 if (!test_bit(In_sync, &rdev->flags)
3943 && !test_bit(Faulty, &rdev->flags))
3945 if (rdev->raid_disk >= 0) {
3946 long long diff = (rdev->new_data_offset
3947 - rdev->data_offset);
3948 if (!mddev->reshape_backwards)
3952 if (first || diff < min_offset_diff)
3953 min_offset_diff = diff;
3957 if (max(before_length, after_length) > min_offset_diff)
3960 if (spares < mddev->delta_disks)
3963 conf->offset_diff = min_offset_diff;
3964 spin_lock_irq(&conf->device_lock);
3965 if (conf->mirrors_new) {
3966 memcpy(conf->mirrors_new, conf->mirrors,
3967 sizeof(struct raid10_info)*conf->prev.raid_disks);
3969 kfree(conf->mirrors_old);
3970 conf->mirrors_old = conf->mirrors;
3971 conf->mirrors = conf->mirrors_new;
3972 conf->mirrors_new = NULL;
3974 setup_geo(&conf->geo, mddev, geo_start);
3976 if (mddev->reshape_backwards) {
3977 sector_t size = raid10_size(mddev, 0, 0);
3978 if (size < mddev->array_sectors) {
3979 spin_unlock_irq(&conf->device_lock);
3980 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3984 mddev->resync_max_sectors = size;
3985 conf->reshape_progress = size;
3987 conf->reshape_progress = 0;
3988 conf->reshape_safe = conf->reshape_progress;
3989 spin_unlock_irq(&conf->device_lock);
3991 if (mddev->delta_disks && mddev->bitmap) {
3992 ret = bitmap_resize(mddev->bitmap,
3993 raid10_size(mddev, 0,
3994 conf->geo.raid_disks),
3999 if (mddev->delta_disks > 0) {
4000 rdev_for_each(rdev, mddev)
4001 if (rdev->raid_disk < 0 &&
4002 !test_bit(Faulty, &rdev->flags)) {
4003 if (raid10_add_disk(mddev, rdev) == 0) {
4004 if (rdev->raid_disk >=
4005 conf->prev.raid_disks)
4006 set_bit(In_sync, &rdev->flags);
4008 rdev->recovery_offset = 0;
4010 if (sysfs_link_rdev(mddev, rdev))
4011 /* Failure here is OK */;
4013 } else if (rdev->raid_disk >= conf->prev.raid_disks
4014 && !test_bit(Faulty, &rdev->flags)) {
4015 /* This is a spare that was manually added */
4016 set_bit(In_sync, &rdev->flags);
4019 /* When a reshape changes the number of devices,
4020 * ->degraded is measured against the larger of the
4021 * pre and post numbers.
4023 spin_lock_irq(&conf->device_lock);
4024 mddev->degraded = calc_degraded(conf);
4025 spin_unlock_irq(&conf->device_lock);
4026 mddev->raid_disks = conf->geo.raid_disks;
4027 mddev->reshape_position = conf->reshape_progress;
4028 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4030 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4031 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4032 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4033 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4034 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4036 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4038 if (!mddev->sync_thread) {
4042 conf->reshape_checkpoint = jiffies;
4043 md_wakeup_thread(mddev->sync_thread);
4044 md_new_event(mddev);
4048 mddev->recovery = 0;
4049 spin_lock_irq(&conf->device_lock);
4050 conf->geo = conf->prev;
4051 mddev->raid_disks = conf->geo.raid_disks;
4052 rdev_for_each(rdev, mddev)
4053 rdev->new_data_offset = rdev->data_offset;
4055 conf->reshape_progress = MaxSector;
4056 conf->reshape_safe = MaxSector;
4057 mddev->reshape_position = MaxSector;
4058 spin_unlock_irq(&conf->device_lock);
4062 /* Calculate the last device-address that could contain
4063 * any block from the chunk that includes the array-address 's'
4064 * and report the next address.
4065 * i.e. the address returned will be chunk-aligned and after
4066 * any data that is in the chunk containing 's'.
4068 static sector_t last_dev_address(sector_t s, struct geom *geo)
4070 s = (s | geo->chunk_mask) + 1;
4071 s >>= geo->chunk_shift;
4072 s *= geo->near_copies;
4073 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4074 s *= geo->far_copies;
4075 s <<= geo->chunk_shift;
4079 /* Calculate the first device-address that could contain
4080 * any block from the chunk that includes the array-address 's'.
4081 * This too will be the start of a chunk
4083 static sector_t first_dev_address(sector_t s, struct geom *geo)
4085 s >>= geo->chunk_shift;
4086 s *= geo->near_copies;
4087 sector_div(s, geo->raid_disks);
4088 s *= geo->far_copies;
4089 s <<= geo->chunk_shift;
4093 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4096 /* We simply copy at most one chunk (smallest of old and new)
4097 * at a time, possibly less if that exceeds RESYNC_PAGES,
4098 * or we hit a bad block or something.
4099 * This might mean we pause for normal IO in the middle of
4100 * a chunk, but that is not a problem was mddev->reshape_position
4101 * can record any location.
4103 * If we will want to write to a location that isn't
4104 * yet recorded as 'safe' (i.e. in metadata on disk) then
4105 * we need to flush all reshape requests and update the metadata.
4107 * When reshaping forwards (e.g. to more devices), we interpret
4108 * 'safe' as the earliest block which might not have been copied
4109 * down yet. We divide this by previous stripe size and multiply
4110 * by previous stripe length to get lowest device offset that we
4111 * cannot write to yet.
4112 * We interpret 'sector_nr' as an address that we want to write to.
4113 * From this we use last_device_address() to find where we might
4114 * write to, and first_device_address on the 'safe' position.
4115 * If this 'next' write position is after the 'safe' position,
4116 * we must update the metadata to increase the 'safe' position.
4118 * When reshaping backwards, we round in the opposite direction
4119 * and perform the reverse test: next write position must not be
4120 * less than current safe position.
4122 * In all this the minimum difference in data offsets
4123 * (conf->offset_diff - always positive) allows a bit of slack,
4124 * so next can be after 'safe', but not by more than offset_disk
4126 * We need to prepare all the bios here before we start any IO
4127 * to ensure the size we choose is acceptable to all devices.
4128 * The means one for each copy for write-out and an extra one for
4130 * We store the read-in bio in ->master_bio and the others in
4131 * ->devs[x].bio and ->devs[x].repl_bio.
4133 struct r10conf *conf = mddev->private;
4134 struct r10bio *r10_bio;
4135 sector_t next, safe, last;
4139 struct md_rdev *rdev;
4142 struct bio *bio, *read_bio;
4143 int sectors_done = 0;
4145 if (sector_nr == 0) {
4146 /* If restarting in the middle, skip the initial sectors */
4147 if (mddev->reshape_backwards &&
4148 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4149 sector_nr = (raid10_size(mddev, 0, 0)
4150 - conf->reshape_progress);
4151 } else if (!mddev->reshape_backwards &&
4152 conf->reshape_progress > 0)
4153 sector_nr = conf->reshape_progress;
4155 mddev->curr_resync_completed = sector_nr;
4156 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4162 /* We don't use sector_nr to track where we are up to
4163 * as that doesn't work well for ->reshape_backwards.
4164 * So just use ->reshape_progress.
4166 if (mddev->reshape_backwards) {
4167 /* 'next' is the earliest device address that we might
4168 * write to for this chunk in the new layout
4170 next = first_dev_address(conf->reshape_progress - 1,
4173 /* 'safe' is the last device address that we might read from
4174 * in the old layout after a restart
4176 safe = last_dev_address(conf->reshape_safe - 1,
4179 if (next + conf->offset_diff < safe)
4182 last = conf->reshape_progress - 1;
4183 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4184 & conf->prev.chunk_mask);
4185 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4186 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4188 /* 'next' is after the last device address that we
4189 * might write to for this chunk in the new layout
4191 next = last_dev_address(conf->reshape_progress, &conf->geo);
4193 /* 'safe' is the earliest device address that we might
4194 * read from in the old layout after a restart
4196 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4198 /* Need to update metadata if 'next' might be beyond 'safe'
4199 * as that would possibly corrupt data
4201 if (next > safe + conf->offset_diff)
4204 sector_nr = conf->reshape_progress;
4205 last = sector_nr | (conf->geo.chunk_mask
4206 & conf->prev.chunk_mask);
4208 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4209 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4213 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4214 /* Need to update reshape_position in metadata */
4216 mddev->reshape_position = conf->reshape_progress;
4217 if (mddev->reshape_backwards)
4218 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4219 - conf->reshape_progress;
4221 mddev->curr_resync_completed = conf->reshape_progress;
4222 conf->reshape_checkpoint = jiffies;
4223 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4224 md_wakeup_thread(mddev->thread);
4225 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4226 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4227 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4228 allow_barrier(conf);
4229 return sectors_done;
4231 conf->reshape_safe = mddev->reshape_position;
4232 allow_barrier(conf);
4236 /* Now schedule reads for blocks from sector_nr to last */
4237 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4239 raise_barrier(conf, sectors_done != 0);
4240 atomic_set(&r10_bio->remaining, 0);
4241 r10_bio->mddev = mddev;
4242 r10_bio->sector = sector_nr;
4243 set_bit(R10BIO_IsReshape, &r10_bio->state);
4244 r10_bio->sectors = last - sector_nr + 1;
4245 rdev = read_balance(conf, r10_bio, &max_sectors);
4246 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4249 /* Cannot read from here, so need to record bad blocks
4250 * on all the target devices.
4253 mempool_free(r10_bio, conf->r10buf_pool);
4254 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4255 return sectors_done;
4258 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4260 read_bio->bi_bdev = rdev->bdev;
4261 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4262 + rdev->data_offset);
4263 read_bio->bi_private = r10_bio;
4264 read_bio->bi_end_io = end_sync_read;
4265 read_bio->bi_rw = READ;
4266 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4267 read_bio->bi_error = 0;
4268 read_bio->bi_vcnt = 0;
4269 read_bio->bi_iter.bi_size = 0;
4270 r10_bio->master_bio = read_bio;
4271 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4273 /* Now find the locations in the new layout */
4274 __raid10_find_phys(&conf->geo, r10_bio);
4277 read_bio->bi_next = NULL;
4279 for (s = 0; s < conf->copies*2; s++) {
4281 int d = r10_bio->devs[s/2].devnum;
4282 struct md_rdev *rdev2;
4284 rdev2 = conf->mirrors[d].replacement;
4285 b = r10_bio->devs[s/2].repl_bio;
4287 rdev2 = conf->mirrors[d].rdev;
4288 b = r10_bio->devs[s/2].bio;
4290 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4294 b->bi_bdev = rdev2->bdev;
4295 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4296 rdev2->new_data_offset;
4297 b->bi_private = r10_bio;
4298 b->bi_end_io = end_reshape_write;
4304 /* Now add as many pages as possible to all of these bios. */
4307 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4308 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4309 int len = (max_sectors - s) << 9;
4310 if (len > PAGE_SIZE)
4312 for (bio = blist; bio ; bio = bio->bi_next) {
4314 if (bio_add_page(bio, page, len, 0))
4317 /* Didn't fit, must stop */
4319 bio2 && bio2 != bio;
4320 bio2 = bio2->bi_next) {
4321 /* Remove last page from this bio */
4323 bio2->bi_iter.bi_size -= len;
4324 bio_clear_flag(bio2, BIO_SEG_VALID);
4328 sector_nr += len >> 9;
4329 nr_sectors += len >> 9;
4332 r10_bio->sectors = nr_sectors;
4334 /* Now submit the read */
4335 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4336 atomic_inc(&r10_bio->remaining);
4337 read_bio->bi_next = NULL;
4338 generic_make_request(read_bio);
4339 sector_nr += nr_sectors;
4340 sectors_done += nr_sectors;
4341 if (sector_nr <= last)
4344 /* Now that we have done the whole section we can
4345 * update reshape_progress
4347 if (mddev->reshape_backwards)
4348 conf->reshape_progress -= sectors_done;
4350 conf->reshape_progress += sectors_done;
4352 return sectors_done;
4355 static void end_reshape_request(struct r10bio *r10_bio);
4356 static int handle_reshape_read_error(struct mddev *mddev,
4357 struct r10bio *r10_bio);
4358 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4360 /* Reshape read completed. Hopefully we have a block
4362 * If we got a read error then we do sync 1-page reads from
4363 * elsewhere until we find the data - or give up.
4365 struct r10conf *conf = mddev->private;
4368 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4369 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4370 /* Reshape has been aborted */
4371 md_done_sync(mddev, r10_bio->sectors, 0);
4375 /* We definitely have the data in the pages, schedule the
4378 atomic_set(&r10_bio->remaining, 1);
4379 for (s = 0; s < conf->copies*2; s++) {
4381 int d = r10_bio->devs[s/2].devnum;
4382 struct md_rdev *rdev;
4384 rdev = conf->mirrors[d].replacement;
4385 b = r10_bio->devs[s/2].repl_bio;
4387 rdev = conf->mirrors[d].rdev;
4388 b = r10_bio->devs[s/2].bio;
4390 if (!rdev || test_bit(Faulty, &rdev->flags))
4392 atomic_inc(&rdev->nr_pending);
4393 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4394 atomic_inc(&r10_bio->remaining);
4396 generic_make_request(b);
4398 end_reshape_request(r10_bio);
4401 static void end_reshape(struct r10conf *conf)
4403 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4406 spin_lock_irq(&conf->device_lock);
4407 conf->prev = conf->geo;
4408 md_finish_reshape(conf->mddev);
4410 conf->reshape_progress = MaxSector;
4411 conf->reshape_safe = MaxSector;
4412 spin_unlock_irq(&conf->device_lock);
4414 /* read-ahead size must cover two whole stripes, which is
4415 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4417 if (conf->mddev->queue) {
4418 int stripe = conf->geo.raid_disks *
4419 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4420 stripe /= conf->geo.near_copies;
4421 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4422 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4427 static int handle_reshape_read_error(struct mddev *mddev,
4428 struct r10bio *r10_bio)
4430 /* Use sync reads to get the blocks from somewhere else */
4431 int sectors = r10_bio->sectors;
4432 struct r10conf *conf = mddev->private;
4434 struct r10bio r10_bio;
4435 struct r10dev devs[conf->copies];
4437 struct r10bio *r10b = &on_stack.r10_bio;
4440 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4442 r10b->sector = r10_bio->sector;
4443 __raid10_find_phys(&conf->prev, r10b);
4448 int first_slot = slot;
4450 if (s > (PAGE_SIZE >> 9))
4454 int d = r10b->devs[slot].devnum;
4455 struct md_rdev *rdev = conf->mirrors[d].rdev;
4458 test_bit(Faulty, &rdev->flags) ||
4459 !test_bit(In_sync, &rdev->flags))
4462 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4463 success = sync_page_io(rdev,
4472 if (slot >= conf->copies)
4474 if (slot == first_slot)
4478 /* couldn't read this block, must give up */
4479 set_bit(MD_RECOVERY_INTR,
4489 static void end_reshape_write(struct bio *bio)
4491 struct r10bio *r10_bio = bio->bi_private;
4492 struct mddev *mddev = r10_bio->mddev;
4493 struct r10conf *conf = mddev->private;
4497 struct md_rdev *rdev = NULL;
4499 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4501 rdev = conf->mirrors[d].replacement;
4504 rdev = conf->mirrors[d].rdev;
4507 if (bio->bi_error) {
4508 /* FIXME should record badblock */
4509 md_error(mddev, rdev);
4512 rdev_dec_pending(rdev, mddev);
4513 end_reshape_request(r10_bio);
4516 static void end_reshape_request(struct r10bio *r10_bio)
4518 if (!atomic_dec_and_test(&r10_bio->remaining))
4520 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4521 bio_put(r10_bio->master_bio);
4525 static void raid10_finish_reshape(struct mddev *mddev)
4527 struct r10conf *conf = mddev->private;
4529 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4532 if (mddev->delta_disks > 0) {
4533 sector_t size = raid10_size(mddev, 0, 0);
4534 md_set_array_sectors(mddev, size);
4535 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4536 mddev->recovery_cp = mddev->resync_max_sectors;
4537 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4539 mddev->resync_max_sectors = size;
4540 set_capacity(mddev->gendisk, mddev->array_sectors);
4541 revalidate_disk(mddev->gendisk);
4544 for (d = conf->geo.raid_disks ;
4545 d < conf->geo.raid_disks - mddev->delta_disks;
4547 struct md_rdev *rdev = conf->mirrors[d].rdev;
4549 clear_bit(In_sync, &rdev->flags);
4550 rdev = conf->mirrors[d].replacement;
4552 clear_bit(In_sync, &rdev->flags);
4555 mddev->layout = mddev->new_layout;
4556 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4557 mddev->reshape_position = MaxSector;
4558 mddev->delta_disks = 0;
4559 mddev->reshape_backwards = 0;
4562 static struct md_personality raid10_personality =
4566 .owner = THIS_MODULE,
4567 .make_request = make_request,
4569 .free = raid10_free,
4571 .error_handler = error,
4572 .hot_add_disk = raid10_add_disk,
4573 .hot_remove_disk= raid10_remove_disk,
4574 .spare_active = raid10_spare_active,
4575 .sync_request = sync_request,
4576 .quiesce = raid10_quiesce,
4577 .size = raid10_size,
4578 .resize = raid10_resize,
4579 .takeover = raid10_takeover,
4580 .check_reshape = raid10_check_reshape,
4581 .start_reshape = raid10_start_reshape,
4582 .finish_reshape = raid10_finish_reshape,
4583 .congested = raid10_congested,
4586 static int __init raid_init(void)
4588 return register_md_personality(&raid10_personality);
4591 static void raid_exit(void)
4593 unregister_md_personality(&raid10_personality);
4596 module_init(raid_init);
4597 module_exit(raid_exit);
4598 MODULE_LICENSE("GPL");
4599 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4600 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4601 MODULE_ALIAS("md-raid10");
4602 MODULE_ALIAS("md-level-10");
4604 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);