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 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1593 spin_unlock_irqrestore(&conf->device_lock, flags);
1595 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1596 "md/raid10:%s: Operation continuing on %d devices.\n",
1597 mdname(mddev), bdevname(rdev->bdev, b),
1598 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1601 static void print_conf(struct r10conf *conf)
1604 struct raid10_info *tmp;
1606 printk(KERN_DEBUG "RAID10 conf printout:\n");
1608 printk(KERN_DEBUG "(!conf)\n");
1611 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1612 conf->geo.raid_disks);
1614 for (i = 0; i < conf->geo.raid_disks; i++) {
1615 char b[BDEVNAME_SIZE];
1616 tmp = conf->mirrors + i;
1618 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1619 i, !test_bit(In_sync, &tmp->rdev->flags),
1620 !test_bit(Faulty, &tmp->rdev->flags),
1621 bdevname(tmp->rdev->bdev,b));
1625 static void close_sync(struct r10conf *conf)
1628 allow_barrier(conf);
1630 mempool_destroy(conf->r10buf_pool);
1631 conf->r10buf_pool = NULL;
1634 static int raid10_spare_active(struct mddev *mddev)
1637 struct r10conf *conf = mddev->private;
1638 struct raid10_info *tmp;
1640 unsigned long flags;
1643 * Find all non-in_sync disks within the RAID10 configuration
1644 * and mark them in_sync
1646 for (i = 0; i < conf->geo.raid_disks; i++) {
1647 tmp = conf->mirrors + i;
1648 if (tmp->replacement
1649 && tmp->replacement->recovery_offset == MaxSector
1650 && !test_bit(Faulty, &tmp->replacement->flags)
1651 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1652 /* Replacement has just become active */
1654 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1657 /* Replaced device not technically faulty,
1658 * but we need to be sure it gets removed
1659 * and never re-added.
1661 set_bit(Faulty, &tmp->rdev->flags);
1662 sysfs_notify_dirent_safe(
1663 tmp->rdev->sysfs_state);
1665 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1666 } else if (tmp->rdev
1667 && tmp->rdev->recovery_offset == MaxSector
1668 && !test_bit(Faulty, &tmp->rdev->flags)
1669 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1671 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1674 spin_lock_irqsave(&conf->device_lock, flags);
1675 mddev->degraded -= count;
1676 spin_unlock_irqrestore(&conf->device_lock, flags);
1682 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1684 struct r10conf *conf = mddev->private;
1688 int last = conf->geo.raid_disks - 1;
1690 if (mddev->recovery_cp < MaxSector)
1691 /* only hot-add to in-sync arrays, as recovery is
1692 * very different from resync
1695 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1698 if (rdev->raid_disk >= 0)
1699 first = last = rdev->raid_disk;
1701 if (rdev->saved_raid_disk >= first &&
1702 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1703 mirror = rdev->saved_raid_disk;
1706 for ( ; mirror <= last ; mirror++) {
1707 struct raid10_info *p = &conf->mirrors[mirror];
1708 if (p->recovery_disabled == mddev->recovery_disabled)
1711 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1712 p->replacement != NULL)
1714 clear_bit(In_sync, &rdev->flags);
1715 set_bit(Replacement, &rdev->flags);
1716 rdev->raid_disk = mirror;
1719 disk_stack_limits(mddev->gendisk, rdev->bdev,
1720 rdev->data_offset << 9);
1722 rcu_assign_pointer(p->replacement, rdev);
1727 disk_stack_limits(mddev->gendisk, rdev->bdev,
1728 rdev->data_offset << 9);
1730 p->head_position = 0;
1731 p->recovery_disabled = mddev->recovery_disabled - 1;
1732 rdev->raid_disk = mirror;
1734 if (rdev->saved_raid_disk != mirror)
1736 rcu_assign_pointer(p->rdev, rdev);
1739 md_integrity_add_rdev(rdev, mddev);
1740 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1741 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1747 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1749 struct r10conf *conf = mddev->private;
1751 int number = rdev->raid_disk;
1752 struct md_rdev **rdevp;
1753 struct raid10_info *p = conf->mirrors + number;
1756 if (rdev == p->rdev)
1758 else if (rdev == p->replacement)
1759 rdevp = &p->replacement;
1763 if (test_bit(In_sync, &rdev->flags) ||
1764 atomic_read(&rdev->nr_pending)) {
1768 /* Only remove faulty devices if recovery
1771 if (!test_bit(Faulty, &rdev->flags) &&
1772 mddev->recovery_disabled != p->recovery_disabled &&
1773 (!p->replacement || p->replacement == rdev) &&
1774 number < conf->geo.raid_disks &&
1781 if (atomic_read(&rdev->nr_pending)) {
1782 /* lost the race, try later */
1786 } else if (p->replacement) {
1787 /* We must have just cleared 'rdev' */
1788 p->rdev = p->replacement;
1789 clear_bit(Replacement, &p->replacement->flags);
1790 smp_mb(); /* Make sure other CPUs may see both as identical
1791 * but will never see neither -- if they are careful.
1793 p->replacement = NULL;
1794 clear_bit(WantReplacement, &rdev->flags);
1796 /* We might have just remove the Replacement as faulty
1797 * Clear the flag just in case
1799 clear_bit(WantReplacement, &rdev->flags);
1801 err = md_integrity_register(mddev);
1809 static void end_sync_read(struct bio *bio)
1811 struct r10bio *r10_bio = bio->bi_private;
1812 struct r10conf *conf = r10_bio->mddev->private;
1815 if (bio == r10_bio->master_bio) {
1816 /* this is a reshape read */
1817 d = r10_bio->read_slot; /* really the read dev */
1819 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1822 set_bit(R10BIO_Uptodate, &r10_bio->state);
1824 /* The write handler will notice the lack of
1825 * R10BIO_Uptodate and record any errors etc
1827 atomic_add(r10_bio->sectors,
1828 &conf->mirrors[d].rdev->corrected_errors);
1830 /* for reconstruct, we always reschedule after a read.
1831 * for resync, only after all reads
1833 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1834 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1835 atomic_dec_and_test(&r10_bio->remaining)) {
1836 /* we have read all the blocks,
1837 * do the comparison in process context in raid10d
1839 reschedule_retry(r10_bio);
1843 static void end_sync_request(struct r10bio *r10_bio)
1845 struct mddev *mddev = r10_bio->mddev;
1847 while (atomic_dec_and_test(&r10_bio->remaining)) {
1848 if (r10_bio->master_bio == NULL) {
1849 /* the primary of several recovery bios */
1850 sector_t s = r10_bio->sectors;
1851 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1852 test_bit(R10BIO_WriteError, &r10_bio->state))
1853 reschedule_retry(r10_bio);
1856 md_done_sync(mddev, s, 1);
1859 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1860 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1861 test_bit(R10BIO_WriteError, &r10_bio->state))
1862 reschedule_retry(r10_bio);
1870 static void end_sync_write(struct bio *bio)
1872 struct r10bio *r10_bio = bio->bi_private;
1873 struct mddev *mddev = r10_bio->mddev;
1874 struct r10conf *conf = mddev->private;
1880 struct md_rdev *rdev = NULL;
1882 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1884 rdev = conf->mirrors[d].replacement;
1886 rdev = conf->mirrors[d].rdev;
1888 if (bio->bi_error) {
1890 md_error(mddev, rdev);
1892 set_bit(WriteErrorSeen, &rdev->flags);
1893 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1894 set_bit(MD_RECOVERY_NEEDED,
1895 &rdev->mddev->recovery);
1896 set_bit(R10BIO_WriteError, &r10_bio->state);
1898 } else if (is_badblock(rdev,
1899 r10_bio->devs[slot].addr,
1901 &first_bad, &bad_sectors))
1902 set_bit(R10BIO_MadeGood, &r10_bio->state);
1904 rdev_dec_pending(rdev, mddev);
1906 end_sync_request(r10_bio);
1910 * Note: sync and recover and handled very differently for raid10
1911 * This code is for resync.
1912 * For resync, we read through virtual addresses and read all blocks.
1913 * If there is any error, we schedule a write. The lowest numbered
1914 * drive is authoritative.
1915 * However requests come for physical address, so we need to map.
1916 * For every physical address there are raid_disks/copies virtual addresses,
1917 * which is always are least one, but is not necessarly an integer.
1918 * This means that a physical address can span multiple chunks, so we may
1919 * have to submit multiple io requests for a single sync request.
1922 * We check if all blocks are in-sync and only write to blocks that
1925 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1927 struct r10conf *conf = mddev->private;
1929 struct bio *tbio, *fbio;
1932 atomic_set(&r10_bio->remaining, 1);
1934 /* find the first device with a block */
1935 for (i=0; i<conf->copies; i++)
1936 if (!r10_bio->devs[i].bio->bi_error)
1939 if (i == conf->copies)
1943 fbio = r10_bio->devs[i].bio;
1945 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1946 /* now find blocks with errors */
1947 for (i=0 ; i < conf->copies ; i++) {
1950 tbio = r10_bio->devs[i].bio;
1952 if (tbio->bi_end_io != end_sync_read)
1956 if (!r10_bio->devs[i].bio->bi_error) {
1957 /* We know that the bi_io_vec layout is the same for
1958 * both 'first' and 'i', so we just compare them.
1959 * All vec entries are PAGE_SIZE;
1961 int sectors = r10_bio->sectors;
1962 for (j = 0; j < vcnt; j++) {
1963 int len = PAGE_SIZE;
1964 if (sectors < (len / 512))
1965 len = sectors * 512;
1966 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1967 page_address(tbio->bi_io_vec[j].bv_page),
1974 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1975 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1976 /* Don't fix anything. */
1979 /* Ok, we need to write this bio, either to correct an
1980 * inconsistency or to correct an unreadable block.
1981 * First we need to fixup bv_offset, bv_len and
1982 * bi_vecs, as the read request might have corrupted these
1986 tbio->bi_vcnt = vcnt;
1987 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
1988 tbio->bi_rw = WRITE;
1989 tbio->bi_private = r10_bio;
1990 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1991 tbio->bi_end_io = end_sync_write;
1993 bio_copy_data(tbio, fbio);
1995 d = r10_bio->devs[i].devnum;
1996 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1997 atomic_inc(&r10_bio->remaining);
1998 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2000 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2001 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2002 generic_make_request(tbio);
2005 /* Now write out to any replacement devices
2008 for (i = 0; i < conf->copies; i++) {
2011 tbio = r10_bio->devs[i].repl_bio;
2012 if (!tbio || !tbio->bi_end_io)
2014 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2015 && r10_bio->devs[i].bio != fbio)
2016 bio_copy_data(tbio, fbio);
2017 d = r10_bio->devs[i].devnum;
2018 atomic_inc(&r10_bio->remaining);
2019 md_sync_acct(conf->mirrors[d].replacement->bdev,
2021 generic_make_request(tbio);
2025 if (atomic_dec_and_test(&r10_bio->remaining)) {
2026 md_done_sync(mddev, r10_bio->sectors, 1);
2032 * Now for the recovery code.
2033 * Recovery happens across physical sectors.
2034 * We recover all non-is_sync drives by finding the virtual address of
2035 * each, and then choose a working drive that also has that virt address.
2036 * There is a separate r10_bio for each non-in_sync drive.
2037 * Only the first two slots are in use. The first for reading,
2038 * The second for writing.
2041 static void fix_recovery_read_error(struct r10bio *r10_bio)
2043 /* We got a read error during recovery.
2044 * We repeat the read in smaller page-sized sections.
2045 * If a read succeeds, write it to the new device or record
2046 * a bad block if we cannot.
2047 * If a read fails, record a bad block on both old and
2050 struct mddev *mddev = r10_bio->mddev;
2051 struct r10conf *conf = mddev->private;
2052 struct bio *bio = r10_bio->devs[0].bio;
2054 int sectors = r10_bio->sectors;
2056 int dr = r10_bio->devs[0].devnum;
2057 int dw = r10_bio->devs[1].devnum;
2061 struct md_rdev *rdev;
2065 if (s > (PAGE_SIZE>>9))
2068 rdev = conf->mirrors[dr].rdev;
2069 addr = r10_bio->devs[0].addr + sect,
2070 ok = sync_page_io(rdev,
2073 bio->bi_io_vec[idx].bv_page,
2076 rdev = conf->mirrors[dw].rdev;
2077 addr = r10_bio->devs[1].addr + sect;
2078 ok = sync_page_io(rdev,
2081 bio->bi_io_vec[idx].bv_page,
2084 set_bit(WriteErrorSeen, &rdev->flags);
2085 if (!test_and_set_bit(WantReplacement,
2087 set_bit(MD_RECOVERY_NEEDED,
2088 &rdev->mddev->recovery);
2092 /* We don't worry if we cannot set a bad block -
2093 * it really is bad so there is no loss in not
2096 rdev_set_badblocks(rdev, addr, s, 0);
2098 if (rdev != conf->mirrors[dw].rdev) {
2099 /* need bad block on destination too */
2100 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2101 addr = r10_bio->devs[1].addr + sect;
2102 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2104 /* just abort the recovery */
2106 "md/raid10:%s: recovery aborted"
2107 " due to read error\n",
2110 conf->mirrors[dw].recovery_disabled
2111 = mddev->recovery_disabled;
2112 set_bit(MD_RECOVERY_INTR,
2125 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2127 struct r10conf *conf = mddev->private;
2129 struct bio *wbio, *wbio2;
2131 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2132 fix_recovery_read_error(r10_bio);
2133 end_sync_request(r10_bio);
2138 * share the pages with the first bio
2139 * and submit the write request
2141 d = r10_bio->devs[1].devnum;
2142 wbio = r10_bio->devs[1].bio;
2143 wbio2 = r10_bio->devs[1].repl_bio;
2144 /* Need to test wbio2->bi_end_io before we call
2145 * generic_make_request as if the former is NULL,
2146 * the latter is free to free wbio2.
2148 if (wbio2 && !wbio2->bi_end_io)
2150 if (wbio->bi_end_io) {
2151 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2152 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2153 generic_make_request(wbio);
2156 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2157 md_sync_acct(conf->mirrors[d].replacement->bdev,
2158 bio_sectors(wbio2));
2159 generic_make_request(wbio2);
2164 * Used by fix_read_error() to decay the per rdev read_errors.
2165 * We halve the read error count for every hour that has elapsed
2166 * since the last recorded read error.
2169 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2171 struct timespec cur_time_mon;
2172 unsigned long hours_since_last;
2173 unsigned int read_errors = atomic_read(&rdev->read_errors);
2175 ktime_get_ts(&cur_time_mon);
2177 if (rdev->last_read_error.tv_sec == 0 &&
2178 rdev->last_read_error.tv_nsec == 0) {
2179 /* first time we've seen a read error */
2180 rdev->last_read_error = cur_time_mon;
2184 hours_since_last = (cur_time_mon.tv_sec -
2185 rdev->last_read_error.tv_sec) / 3600;
2187 rdev->last_read_error = cur_time_mon;
2190 * if hours_since_last is > the number of bits in read_errors
2191 * just set read errors to 0. We do this to avoid
2192 * overflowing the shift of read_errors by hours_since_last.
2194 if (hours_since_last >= 8 * sizeof(read_errors))
2195 atomic_set(&rdev->read_errors, 0);
2197 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2200 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2201 int sectors, struct page *page, int rw)
2206 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2207 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2209 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2213 set_bit(WriteErrorSeen, &rdev->flags);
2214 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2215 set_bit(MD_RECOVERY_NEEDED,
2216 &rdev->mddev->recovery);
2218 /* need to record an error - either for the block or the device */
2219 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2220 md_error(rdev->mddev, rdev);
2225 * This is a kernel thread which:
2227 * 1. Retries failed read operations on working mirrors.
2228 * 2. Updates the raid superblock when problems encounter.
2229 * 3. Performs writes following reads for array synchronising.
2232 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2234 int sect = 0; /* Offset from r10_bio->sector */
2235 int sectors = r10_bio->sectors;
2236 struct md_rdev*rdev;
2237 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2238 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2240 /* still own a reference to this rdev, so it cannot
2241 * have been cleared recently.
2243 rdev = conf->mirrors[d].rdev;
2245 if (test_bit(Faulty, &rdev->flags))
2246 /* drive has already been failed, just ignore any
2247 more fix_read_error() attempts */
2250 check_decay_read_errors(mddev, rdev);
2251 atomic_inc(&rdev->read_errors);
2252 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2253 char b[BDEVNAME_SIZE];
2254 bdevname(rdev->bdev, b);
2257 "md/raid10:%s: %s: Raid device exceeded "
2258 "read_error threshold [cur %d:max %d]\n",
2260 atomic_read(&rdev->read_errors), max_read_errors);
2262 "md/raid10:%s: %s: Failing raid device\n",
2264 md_error(mddev, conf->mirrors[d].rdev);
2265 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2271 int sl = r10_bio->read_slot;
2275 if (s > (PAGE_SIZE>>9))
2283 d = r10_bio->devs[sl].devnum;
2284 rdev = rcu_dereference(conf->mirrors[d].rdev);
2286 test_bit(In_sync, &rdev->flags) &&
2287 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2288 &first_bad, &bad_sectors) == 0) {
2289 atomic_inc(&rdev->nr_pending);
2291 success = sync_page_io(rdev,
2292 r10_bio->devs[sl].addr +
2295 conf->tmppage, READ, false);
2296 rdev_dec_pending(rdev, mddev);
2302 if (sl == conf->copies)
2304 } while (!success && sl != r10_bio->read_slot);
2308 /* Cannot read from anywhere, just mark the block
2309 * as bad on the first device to discourage future
2312 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2313 rdev = conf->mirrors[dn].rdev;
2315 if (!rdev_set_badblocks(
2317 r10_bio->devs[r10_bio->read_slot].addr
2320 md_error(mddev, rdev);
2321 r10_bio->devs[r10_bio->read_slot].bio
2328 /* write it back and re-read */
2330 while (sl != r10_bio->read_slot) {
2331 char b[BDEVNAME_SIZE];
2336 d = r10_bio->devs[sl].devnum;
2337 rdev = rcu_dereference(conf->mirrors[d].rdev);
2339 !test_bit(In_sync, &rdev->flags))
2342 atomic_inc(&rdev->nr_pending);
2344 if (r10_sync_page_io(rdev,
2345 r10_bio->devs[sl].addr +
2347 s, conf->tmppage, WRITE)
2349 /* Well, this device is dead */
2351 "md/raid10:%s: read correction "
2353 " (%d sectors at %llu on %s)\n",
2355 (unsigned long long)(
2357 choose_data_offset(r10_bio,
2359 bdevname(rdev->bdev, b));
2360 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2363 bdevname(rdev->bdev, b));
2365 rdev_dec_pending(rdev, mddev);
2369 while (sl != r10_bio->read_slot) {
2370 char b[BDEVNAME_SIZE];
2375 d = r10_bio->devs[sl].devnum;
2376 rdev = rcu_dereference(conf->mirrors[d].rdev);
2378 !test_bit(In_sync, &rdev->flags))
2381 atomic_inc(&rdev->nr_pending);
2383 switch (r10_sync_page_io(rdev,
2384 r10_bio->devs[sl].addr +
2389 /* Well, this device is dead */
2391 "md/raid10:%s: unable to read back "
2393 " (%d sectors at %llu on %s)\n",
2395 (unsigned long long)(
2397 choose_data_offset(r10_bio, rdev)),
2398 bdevname(rdev->bdev, b));
2399 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2402 bdevname(rdev->bdev, b));
2406 "md/raid10:%s: read error corrected"
2407 " (%d sectors at %llu on %s)\n",
2409 (unsigned long long)(
2411 choose_data_offset(r10_bio, rdev)),
2412 bdevname(rdev->bdev, b));
2413 atomic_add(s, &rdev->corrected_errors);
2416 rdev_dec_pending(rdev, mddev);
2426 static int narrow_write_error(struct r10bio *r10_bio, int i)
2428 struct bio *bio = r10_bio->master_bio;
2429 struct mddev *mddev = r10_bio->mddev;
2430 struct r10conf *conf = mddev->private;
2431 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2432 /* bio has the data to be written to slot 'i' where
2433 * we just recently had a write error.
2434 * We repeatedly clone the bio and trim down to one block,
2435 * then try the write. Where the write fails we record
2437 * It is conceivable that the bio doesn't exactly align with
2438 * blocks. We must handle this.
2440 * We currently own a reference to the rdev.
2446 int sect_to_write = r10_bio->sectors;
2449 if (rdev->badblocks.shift < 0)
2452 block_sectors = roundup(1 << rdev->badblocks.shift,
2453 bdev_logical_block_size(rdev->bdev) >> 9);
2454 sector = r10_bio->sector;
2455 sectors = ((r10_bio->sector + block_sectors)
2456 & ~(sector_t)(block_sectors - 1))
2459 while (sect_to_write) {
2461 if (sectors > sect_to_write)
2462 sectors = sect_to_write;
2463 /* Write at 'sector' for 'sectors' */
2464 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2465 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2466 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2467 choose_data_offset(r10_bio, rdev) +
2468 (sector - r10_bio->sector));
2469 wbio->bi_bdev = rdev->bdev;
2470 if (submit_bio_wait(WRITE, wbio) < 0)
2472 ok = rdev_set_badblocks(rdev, sector,
2477 sect_to_write -= sectors;
2479 sectors = block_sectors;
2484 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2486 int slot = r10_bio->read_slot;
2488 struct r10conf *conf = mddev->private;
2489 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2490 char b[BDEVNAME_SIZE];
2491 unsigned long do_sync;
2494 /* we got a read error. Maybe the drive is bad. Maybe just
2495 * the block and we can fix it.
2496 * We freeze all other IO, and try reading the block from
2497 * other devices. When we find one, we re-write
2498 * and check it that fixes the read error.
2499 * This is all done synchronously while the array is
2502 bio = r10_bio->devs[slot].bio;
2503 bdevname(bio->bi_bdev, b);
2505 r10_bio->devs[slot].bio = NULL;
2507 if (mddev->ro == 0) {
2508 freeze_array(conf, 1);
2509 fix_read_error(conf, mddev, r10_bio);
2510 unfreeze_array(conf);
2512 r10_bio->devs[slot].bio = IO_BLOCKED;
2514 rdev_dec_pending(rdev, mddev);
2517 rdev = read_balance(conf, r10_bio, &max_sectors);
2519 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2520 " read error for block %llu\n",
2522 (unsigned long long)r10_bio->sector);
2523 raid_end_bio_io(r10_bio);
2527 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2528 slot = r10_bio->read_slot;
2531 "md/raid10:%s: %s: redirecting "
2532 "sector %llu to another mirror\n",
2534 bdevname(rdev->bdev, b),
2535 (unsigned long long)r10_bio->sector);
2536 bio = bio_clone_mddev(r10_bio->master_bio,
2538 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2539 r10_bio->devs[slot].bio = bio;
2540 r10_bio->devs[slot].rdev = rdev;
2541 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2542 + choose_data_offset(r10_bio, rdev);
2543 bio->bi_bdev = rdev->bdev;
2544 bio->bi_rw = READ | do_sync;
2545 bio->bi_private = r10_bio;
2546 bio->bi_end_io = raid10_end_read_request;
2547 if (max_sectors < r10_bio->sectors) {
2548 /* Drat - have to split this up more */
2549 struct bio *mbio = r10_bio->master_bio;
2550 int sectors_handled =
2551 r10_bio->sector + max_sectors
2552 - mbio->bi_iter.bi_sector;
2553 r10_bio->sectors = max_sectors;
2554 spin_lock_irq(&conf->device_lock);
2555 if (mbio->bi_phys_segments == 0)
2556 mbio->bi_phys_segments = 2;
2558 mbio->bi_phys_segments++;
2559 spin_unlock_irq(&conf->device_lock);
2560 generic_make_request(bio);
2562 r10_bio = mempool_alloc(conf->r10bio_pool,
2564 r10_bio->master_bio = mbio;
2565 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2567 set_bit(R10BIO_ReadError,
2569 r10_bio->mddev = mddev;
2570 r10_bio->sector = mbio->bi_iter.bi_sector
2575 generic_make_request(bio);
2578 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2580 /* Some sort of write request has finished and it
2581 * succeeded in writing where we thought there was a
2582 * bad block. So forget the bad block.
2583 * Or possibly if failed and we need to record
2587 struct md_rdev *rdev;
2589 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2590 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2591 for (m = 0; m < conf->copies; m++) {
2592 int dev = r10_bio->devs[m].devnum;
2593 rdev = conf->mirrors[dev].rdev;
2594 if (r10_bio->devs[m].bio == NULL)
2596 if (!r10_bio->devs[m].bio->bi_error) {
2597 rdev_clear_badblocks(
2599 r10_bio->devs[m].addr,
2600 r10_bio->sectors, 0);
2602 if (!rdev_set_badblocks(
2604 r10_bio->devs[m].addr,
2605 r10_bio->sectors, 0))
2606 md_error(conf->mddev, rdev);
2608 rdev = conf->mirrors[dev].replacement;
2609 if (r10_bio->devs[m].repl_bio == NULL)
2612 if (!r10_bio->devs[m].repl_bio->bi_error) {
2613 rdev_clear_badblocks(
2615 r10_bio->devs[m].addr,
2616 r10_bio->sectors, 0);
2618 if (!rdev_set_badblocks(
2620 r10_bio->devs[m].addr,
2621 r10_bio->sectors, 0))
2622 md_error(conf->mddev, rdev);
2628 for (m = 0; m < conf->copies; m++) {
2629 int dev = r10_bio->devs[m].devnum;
2630 struct bio *bio = r10_bio->devs[m].bio;
2631 rdev = conf->mirrors[dev].rdev;
2632 if (bio == IO_MADE_GOOD) {
2633 rdev_clear_badblocks(
2635 r10_bio->devs[m].addr,
2636 r10_bio->sectors, 0);
2637 rdev_dec_pending(rdev, conf->mddev);
2638 } else if (bio != NULL && bio->bi_error) {
2640 if (!narrow_write_error(r10_bio, m)) {
2641 md_error(conf->mddev, rdev);
2642 set_bit(R10BIO_Degraded,
2645 rdev_dec_pending(rdev, conf->mddev);
2647 bio = r10_bio->devs[m].repl_bio;
2648 rdev = conf->mirrors[dev].replacement;
2649 if (rdev && bio == IO_MADE_GOOD) {
2650 rdev_clear_badblocks(
2652 r10_bio->devs[m].addr,
2653 r10_bio->sectors, 0);
2654 rdev_dec_pending(rdev, conf->mddev);
2658 spin_lock_irq(&conf->device_lock);
2659 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2660 spin_unlock_irq(&conf->device_lock);
2661 md_wakeup_thread(conf->mddev->thread);
2663 if (test_bit(R10BIO_WriteError,
2665 close_write(r10_bio);
2666 raid_end_bio_io(r10_bio);
2671 static void raid10d(struct md_thread *thread)
2673 struct mddev *mddev = thread->mddev;
2674 struct r10bio *r10_bio;
2675 unsigned long flags;
2676 struct r10conf *conf = mddev->private;
2677 struct list_head *head = &conf->retry_list;
2678 struct blk_plug plug;
2680 md_check_recovery(mddev);
2682 if (!list_empty_careful(&conf->bio_end_io_list) &&
2683 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2685 spin_lock_irqsave(&conf->device_lock, flags);
2686 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2687 list_add(&tmp, &conf->bio_end_io_list);
2688 list_del_init(&conf->bio_end_io_list);
2690 spin_unlock_irqrestore(&conf->device_lock, flags);
2691 while (!list_empty(&tmp)) {
2692 r10_bio = list_first_entry(&tmp, struct r10bio,
2694 list_del(&r10_bio->retry_list);
2695 if (mddev->degraded)
2696 set_bit(R10BIO_Degraded, &r10_bio->state);
2698 if (test_bit(R10BIO_WriteError,
2700 close_write(r10_bio);
2701 raid_end_bio_io(r10_bio);
2705 blk_start_plug(&plug);
2708 flush_pending_writes(conf);
2710 spin_lock_irqsave(&conf->device_lock, flags);
2711 if (list_empty(head)) {
2712 spin_unlock_irqrestore(&conf->device_lock, flags);
2715 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2716 list_del(head->prev);
2718 spin_unlock_irqrestore(&conf->device_lock, flags);
2720 mddev = r10_bio->mddev;
2721 conf = mddev->private;
2722 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2723 test_bit(R10BIO_WriteError, &r10_bio->state))
2724 handle_write_completed(conf, r10_bio);
2725 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2726 reshape_request_write(mddev, r10_bio);
2727 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2728 sync_request_write(mddev, r10_bio);
2729 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2730 recovery_request_write(mddev, r10_bio);
2731 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2732 handle_read_error(mddev, r10_bio);
2734 /* just a partial read to be scheduled from a
2737 int slot = r10_bio->read_slot;
2738 generic_make_request(r10_bio->devs[slot].bio);
2742 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2743 md_check_recovery(mddev);
2745 blk_finish_plug(&plug);
2748 static int init_resync(struct r10conf *conf)
2753 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2754 BUG_ON(conf->r10buf_pool);
2755 conf->have_replacement = 0;
2756 for (i = 0; i < conf->geo.raid_disks; i++)
2757 if (conf->mirrors[i].replacement)
2758 conf->have_replacement = 1;
2759 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2760 if (!conf->r10buf_pool)
2762 conf->next_resync = 0;
2767 * perform a "sync" on one "block"
2769 * We need to make sure that no normal I/O request - particularly write
2770 * requests - conflict with active sync requests.
2772 * This is achieved by tracking pending requests and a 'barrier' concept
2773 * that can be installed to exclude normal IO requests.
2775 * Resync and recovery are handled very differently.
2776 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2778 * For resync, we iterate over virtual addresses, read all copies,
2779 * and update if there are differences. If only one copy is live,
2781 * For recovery, we iterate over physical addresses, read a good
2782 * value for each non-in_sync drive, and over-write.
2784 * So, for recovery we may have several outstanding complex requests for a
2785 * given address, one for each out-of-sync device. We model this by allocating
2786 * a number of r10_bio structures, one for each out-of-sync device.
2787 * As we setup these structures, we collect all bio's together into a list
2788 * which we then process collectively to add pages, and then process again
2789 * to pass to generic_make_request.
2791 * The r10_bio structures are linked using a borrowed master_bio pointer.
2792 * This link is counted in ->remaining. When the r10_bio that points to NULL
2793 * has its remaining count decremented to 0, the whole complex operation
2798 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2801 struct r10conf *conf = mddev->private;
2802 struct r10bio *r10_bio;
2803 struct bio *biolist = NULL, *bio;
2804 sector_t max_sector, nr_sectors;
2807 sector_t sync_blocks;
2808 sector_t sectors_skipped = 0;
2809 int chunks_skipped = 0;
2810 sector_t chunk_mask = conf->geo.chunk_mask;
2812 if (!conf->r10buf_pool)
2813 if (init_resync(conf))
2817 * Allow skipping a full rebuild for incremental assembly
2818 * of a clean array, like RAID1 does.
2820 if (mddev->bitmap == NULL &&
2821 mddev->recovery_cp == MaxSector &&
2822 mddev->reshape_position == MaxSector &&
2823 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2824 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2825 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2826 conf->fullsync == 0) {
2828 return mddev->dev_sectors - sector_nr;
2832 max_sector = mddev->dev_sectors;
2833 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2834 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2835 max_sector = mddev->resync_max_sectors;
2836 if (sector_nr >= max_sector) {
2837 /* If we aborted, we need to abort the
2838 * sync on the 'current' bitmap chucks (there can
2839 * be several when recovering multiple devices).
2840 * as we may have started syncing it but not finished.
2841 * We can find the current address in
2842 * mddev->curr_resync, but for recovery,
2843 * we need to convert that to several
2844 * virtual addresses.
2846 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2852 if (mddev->curr_resync < max_sector) { /* aborted */
2853 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2854 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2856 else for (i = 0; i < conf->geo.raid_disks; i++) {
2858 raid10_find_virt(conf, mddev->curr_resync, i);
2859 bitmap_end_sync(mddev->bitmap, sect,
2863 /* completed sync */
2864 if ((!mddev->bitmap || conf->fullsync)
2865 && conf->have_replacement
2866 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2867 /* Completed a full sync so the replacements
2868 * are now fully recovered.
2870 for (i = 0; i < conf->geo.raid_disks; i++)
2871 if (conf->mirrors[i].replacement)
2872 conf->mirrors[i].replacement
2878 bitmap_close_sync(mddev->bitmap);
2881 return sectors_skipped;
2884 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2885 return reshape_request(mddev, sector_nr, skipped);
2887 if (chunks_skipped >= conf->geo.raid_disks) {
2888 /* if there has been nothing to do on any drive,
2889 * then there is nothing to do at all..
2892 return (max_sector - sector_nr) + sectors_skipped;
2895 if (max_sector > mddev->resync_max)
2896 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2898 /* make sure whole request will fit in a chunk - if chunks
2901 if (conf->geo.near_copies < conf->geo.raid_disks &&
2902 max_sector > (sector_nr | chunk_mask))
2903 max_sector = (sector_nr | chunk_mask) + 1;
2905 /* Again, very different code for resync and recovery.
2906 * Both must result in an r10bio with a list of bios that
2907 * have bi_end_io, bi_sector, bi_bdev set,
2908 * and bi_private set to the r10bio.
2909 * For recovery, we may actually create several r10bios
2910 * with 2 bios in each, that correspond to the bios in the main one.
2911 * In this case, the subordinate r10bios link back through a
2912 * borrowed master_bio pointer, and the counter in the master
2913 * includes a ref from each subordinate.
2915 /* First, we decide what to do and set ->bi_end_io
2916 * To end_sync_read if we want to read, and
2917 * end_sync_write if we will want to write.
2920 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2921 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2922 /* recovery... the complicated one */
2926 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2932 struct raid10_info *mirror = &conf->mirrors[i];
2934 if ((mirror->rdev == NULL ||
2935 test_bit(In_sync, &mirror->rdev->flags))
2937 (mirror->replacement == NULL ||
2939 &mirror->replacement->flags)))
2943 /* want to reconstruct this device */
2945 sect = raid10_find_virt(conf, sector_nr, i);
2946 if (sect >= mddev->resync_max_sectors) {
2947 /* last stripe is not complete - don't
2948 * try to recover this sector.
2952 /* Unless we are doing a full sync, or a replacement
2953 * we only need to recover the block if it is set in
2956 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2958 if (sync_blocks < max_sync)
2959 max_sync = sync_blocks;
2961 mirror->replacement == NULL &&
2963 /* yep, skip the sync_blocks here, but don't assume
2964 * that there will never be anything to do here
2966 chunks_skipped = -1;
2970 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2972 raise_barrier(conf, rb2 != NULL);
2973 atomic_set(&r10_bio->remaining, 0);
2975 r10_bio->master_bio = (struct bio*)rb2;
2977 atomic_inc(&rb2->remaining);
2978 r10_bio->mddev = mddev;
2979 set_bit(R10BIO_IsRecover, &r10_bio->state);
2980 r10_bio->sector = sect;
2982 raid10_find_phys(conf, r10_bio);
2984 /* Need to check if the array will still be
2987 for (j = 0; j < conf->geo.raid_disks; j++)
2988 if (conf->mirrors[j].rdev == NULL ||
2989 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2994 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2995 &sync_blocks, still_degraded);
2998 for (j=0; j<conf->copies;j++) {
3000 int d = r10_bio->devs[j].devnum;
3001 sector_t from_addr, to_addr;
3002 struct md_rdev *rdev;
3003 sector_t sector, first_bad;
3005 if (!conf->mirrors[d].rdev ||
3006 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3008 /* This is where we read from */
3010 rdev = conf->mirrors[d].rdev;
3011 sector = r10_bio->devs[j].addr;
3013 if (is_badblock(rdev, sector, max_sync,
3014 &first_bad, &bad_sectors)) {
3015 if (first_bad > sector)
3016 max_sync = first_bad - sector;
3018 bad_sectors -= (sector
3020 if (max_sync > bad_sectors)
3021 max_sync = bad_sectors;
3025 bio = r10_bio->devs[0].bio;
3027 bio->bi_next = biolist;
3029 bio->bi_private = r10_bio;
3030 bio->bi_end_io = end_sync_read;
3032 from_addr = r10_bio->devs[j].addr;
3033 bio->bi_iter.bi_sector = from_addr +
3035 bio->bi_bdev = rdev->bdev;
3036 atomic_inc(&rdev->nr_pending);
3037 /* and we write to 'i' (if not in_sync) */
3039 for (k=0; k<conf->copies; k++)
3040 if (r10_bio->devs[k].devnum == i)
3042 BUG_ON(k == conf->copies);
3043 to_addr = r10_bio->devs[k].addr;
3044 r10_bio->devs[0].devnum = d;
3045 r10_bio->devs[0].addr = from_addr;
3046 r10_bio->devs[1].devnum = i;
3047 r10_bio->devs[1].addr = to_addr;
3049 rdev = mirror->rdev;
3050 if (!test_bit(In_sync, &rdev->flags)) {
3051 bio = r10_bio->devs[1].bio;
3053 bio->bi_next = biolist;
3055 bio->bi_private = r10_bio;
3056 bio->bi_end_io = end_sync_write;
3058 bio->bi_iter.bi_sector = to_addr
3059 + rdev->data_offset;
3060 bio->bi_bdev = rdev->bdev;
3061 atomic_inc(&r10_bio->remaining);
3063 r10_bio->devs[1].bio->bi_end_io = NULL;
3065 /* and maybe write to replacement */
3066 bio = r10_bio->devs[1].repl_bio;
3068 bio->bi_end_io = NULL;
3069 rdev = mirror->replacement;
3070 /* Note: if rdev != NULL, then bio
3071 * cannot be NULL as r10buf_pool_alloc will
3072 * have allocated it.
3073 * So the second test here is pointless.
3074 * But it keeps semantic-checkers happy, and
3075 * this comment keeps human reviewers
3078 if (rdev == NULL || bio == NULL ||
3079 test_bit(Faulty, &rdev->flags))
3082 bio->bi_next = biolist;
3084 bio->bi_private = r10_bio;
3085 bio->bi_end_io = end_sync_write;
3087 bio->bi_iter.bi_sector = to_addr +
3089 bio->bi_bdev = rdev->bdev;
3090 atomic_inc(&r10_bio->remaining);
3093 if (j == conf->copies) {
3094 /* Cannot recover, so abort the recovery or
3095 * record a bad block */
3097 /* problem is that there are bad blocks
3098 * on other device(s)
3101 for (k = 0; k < conf->copies; k++)
3102 if (r10_bio->devs[k].devnum == i)
3104 if (!test_bit(In_sync,
3105 &mirror->rdev->flags)
3106 && !rdev_set_badblocks(
3108 r10_bio->devs[k].addr,
3111 if (mirror->replacement &&
3112 !rdev_set_badblocks(
3113 mirror->replacement,
3114 r10_bio->devs[k].addr,
3119 if (!test_and_set_bit(MD_RECOVERY_INTR,
3121 printk(KERN_INFO "md/raid10:%s: insufficient "
3122 "working devices for recovery.\n",
3124 mirror->recovery_disabled
3125 = mddev->recovery_disabled;
3129 atomic_dec(&rb2->remaining);
3134 if (biolist == NULL) {
3136 struct r10bio *rb2 = r10_bio;
3137 r10_bio = (struct r10bio*) rb2->master_bio;
3138 rb2->master_bio = NULL;
3144 /* resync. Schedule a read for every block at this virt offset */
3147 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3149 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3150 &sync_blocks, mddev->degraded) &&
3151 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3152 &mddev->recovery)) {
3153 /* We can skip this block */
3155 return sync_blocks + sectors_skipped;
3157 if (sync_blocks < max_sync)
3158 max_sync = sync_blocks;
3159 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3162 r10_bio->mddev = mddev;
3163 atomic_set(&r10_bio->remaining, 0);
3164 raise_barrier(conf, 0);
3165 conf->next_resync = sector_nr;
3167 r10_bio->master_bio = NULL;
3168 r10_bio->sector = sector_nr;
3169 set_bit(R10BIO_IsSync, &r10_bio->state);
3170 raid10_find_phys(conf, r10_bio);
3171 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3173 for (i = 0; i < conf->copies; i++) {
3174 int d = r10_bio->devs[i].devnum;
3175 sector_t first_bad, sector;
3178 if (r10_bio->devs[i].repl_bio)
3179 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3181 bio = r10_bio->devs[i].bio;
3183 bio->bi_error = -EIO;
3184 if (conf->mirrors[d].rdev == NULL ||
3185 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3187 sector = r10_bio->devs[i].addr;
3188 if (is_badblock(conf->mirrors[d].rdev,
3190 &first_bad, &bad_sectors)) {
3191 if (first_bad > sector)
3192 max_sync = first_bad - sector;
3194 bad_sectors -= (sector - first_bad);
3195 if (max_sync > bad_sectors)
3196 max_sync = bad_sectors;
3200 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3201 atomic_inc(&r10_bio->remaining);
3202 bio->bi_next = biolist;
3204 bio->bi_private = r10_bio;
3205 bio->bi_end_io = end_sync_read;
3207 bio->bi_iter.bi_sector = sector +
3208 conf->mirrors[d].rdev->data_offset;
3209 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3212 if (conf->mirrors[d].replacement == NULL ||
3214 &conf->mirrors[d].replacement->flags))
3217 /* Need to set up for writing to the replacement */
3218 bio = r10_bio->devs[i].repl_bio;
3220 bio->bi_error = -EIO;
3222 sector = r10_bio->devs[i].addr;
3223 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3224 bio->bi_next = biolist;
3226 bio->bi_private = r10_bio;
3227 bio->bi_end_io = end_sync_write;
3229 bio->bi_iter.bi_sector = sector +
3230 conf->mirrors[d].replacement->data_offset;
3231 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3236 for (i=0; i<conf->copies; i++) {
3237 int d = r10_bio->devs[i].devnum;
3238 if (r10_bio->devs[i].bio->bi_end_io)
3239 rdev_dec_pending(conf->mirrors[d].rdev,
3241 if (r10_bio->devs[i].repl_bio &&
3242 r10_bio->devs[i].repl_bio->bi_end_io)
3244 conf->mirrors[d].replacement,
3254 if (sector_nr + max_sync < max_sector)
3255 max_sector = sector_nr + max_sync;
3258 int len = PAGE_SIZE;
3259 if (sector_nr + (len>>9) > max_sector)
3260 len = (max_sector - sector_nr) << 9;
3263 for (bio= biolist ; bio ; bio=bio->bi_next) {
3265 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3266 if (bio_add_page(bio, page, len, 0))
3270 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3271 for (bio2 = biolist;
3272 bio2 && bio2 != bio;
3273 bio2 = bio2->bi_next) {
3274 /* remove last page from this bio */
3276 bio2->bi_iter.bi_size -= len;
3277 bio_clear_flag(bio2, BIO_SEG_VALID);
3281 nr_sectors += len>>9;
3282 sector_nr += len>>9;
3283 } while (biolist->bi_vcnt < RESYNC_PAGES);
3285 r10_bio->sectors = nr_sectors;
3289 biolist = biolist->bi_next;
3291 bio->bi_next = NULL;
3292 r10_bio = bio->bi_private;
3293 r10_bio->sectors = nr_sectors;
3295 if (bio->bi_end_io == end_sync_read) {
3296 md_sync_acct(bio->bi_bdev, nr_sectors);
3298 generic_make_request(bio);
3302 if (sectors_skipped)
3303 /* pretend they weren't skipped, it makes
3304 * no important difference in this case
3306 md_done_sync(mddev, sectors_skipped, 1);
3308 return sectors_skipped + nr_sectors;
3310 /* There is nowhere to write, so all non-sync
3311 * drives must be failed or in resync, all drives
3312 * have a bad block, so try the next chunk...
3314 if (sector_nr + max_sync < max_sector)
3315 max_sector = sector_nr + max_sync;
3317 sectors_skipped += (max_sector - sector_nr);
3319 sector_nr = max_sector;
3324 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3327 struct r10conf *conf = mddev->private;
3330 raid_disks = min(conf->geo.raid_disks,
3331 conf->prev.raid_disks);
3333 sectors = conf->dev_sectors;
3335 size = sectors >> conf->geo.chunk_shift;
3336 sector_div(size, conf->geo.far_copies);
3337 size = size * raid_disks;
3338 sector_div(size, conf->geo.near_copies);
3340 return size << conf->geo.chunk_shift;
3343 static void calc_sectors(struct r10conf *conf, sector_t size)
3345 /* Calculate the number of sectors-per-device that will
3346 * actually be used, and set conf->dev_sectors and
3350 size = size >> conf->geo.chunk_shift;
3351 sector_div(size, conf->geo.far_copies);
3352 size = size * conf->geo.raid_disks;
3353 sector_div(size, conf->geo.near_copies);
3354 /* 'size' is now the number of chunks in the array */
3355 /* calculate "used chunks per device" */
3356 size = size * conf->copies;
3358 /* We need to round up when dividing by raid_disks to
3359 * get the stride size.
3361 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3363 conf->dev_sectors = size << conf->geo.chunk_shift;
3365 if (conf->geo.far_offset)
3366 conf->geo.stride = 1 << conf->geo.chunk_shift;
3368 sector_div(size, conf->geo.far_copies);
3369 conf->geo.stride = size << conf->geo.chunk_shift;
3373 enum geo_type {geo_new, geo_old, geo_start};
3374 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3377 int layout, chunk, disks;
3380 layout = mddev->layout;
3381 chunk = mddev->chunk_sectors;
3382 disks = mddev->raid_disks - mddev->delta_disks;
3385 layout = mddev->new_layout;
3386 chunk = mddev->new_chunk_sectors;
3387 disks = mddev->raid_disks;
3389 default: /* avoid 'may be unused' warnings */
3390 case geo_start: /* new when starting reshape - raid_disks not
3392 layout = mddev->new_layout;
3393 chunk = mddev->new_chunk_sectors;
3394 disks = mddev->raid_disks + mddev->delta_disks;
3399 if (chunk < (PAGE_SIZE >> 9) ||
3400 !is_power_of_2(chunk))
3403 fc = (layout >> 8) & 255;
3404 fo = layout & (1<<16);
3405 geo->raid_disks = disks;
3406 geo->near_copies = nc;
3407 geo->far_copies = fc;
3408 geo->far_offset = fo;
3409 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3410 geo->chunk_mask = chunk - 1;
3411 geo->chunk_shift = ffz(~chunk);
3415 static struct r10conf *setup_conf(struct mddev *mddev)
3417 struct r10conf *conf = NULL;
3422 copies = setup_geo(&geo, mddev, geo_new);
3425 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3426 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3427 mdname(mddev), PAGE_SIZE);
3431 if (copies < 2 || copies > mddev->raid_disks) {
3432 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3433 mdname(mddev), mddev->new_layout);
3438 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3442 /* FIXME calc properly */
3443 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3444 max(0,-mddev->delta_disks)),
3449 conf->tmppage = alloc_page(GFP_KERNEL);
3454 conf->copies = copies;
3455 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3456 r10bio_pool_free, conf);
3457 if (!conf->r10bio_pool)
3460 calc_sectors(conf, mddev->dev_sectors);
3461 if (mddev->reshape_position == MaxSector) {
3462 conf->prev = conf->geo;
3463 conf->reshape_progress = MaxSector;
3465 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3469 conf->reshape_progress = mddev->reshape_position;
3470 if (conf->prev.far_offset)
3471 conf->prev.stride = 1 << conf->prev.chunk_shift;
3473 /* far_copies must be 1 */
3474 conf->prev.stride = conf->dev_sectors;
3476 conf->reshape_safe = conf->reshape_progress;
3477 spin_lock_init(&conf->device_lock);
3478 INIT_LIST_HEAD(&conf->retry_list);
3479 INIT_LIST_HEAD(&conf->bio_end_io_list);
3481 spin_lock_init(&conf->resync_lock);
3482 init_waitqueue_head(&conf->wait_barrier);
3484 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3488 conf->mddev = mddev;
3493 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3496 mempool_destroy(conf->r10bio_pool);
3497 kfree(conf->mirrors);
3498 safe_put_page(conf->tmppage);
3501 return ERR_PTR(err);
3504 static int run(struct mddev *mddev)
3506 struct r10conf *conf;
3507 int i, disk_idx, chunk_size;
3508 struct raid10_info *disk;
3509 struct md_rdev *rdev;
3511 sector_t min_offset_diff = 0;
3513 bool discard_supported = false;
3515 if (mddev->private == NULL) {
3516 conf = setup_conf(mddev);
3518 return PTR_ERR(conf);
3519 mddev->private = conf;
3521 conf = mddev->private;
3525 mddev->thread = conf->thread;
3526 conf->thread = NULL;
3528 chunk_size = mddev->chunk_sectors << 9;
3530 blk_queue_max_discard_sectors(mddev->queue,
3531 mddev->chunk_sectors);
3532 blk_queue_max_write_same_sectors(mddev->queue, 0);
3533 blk_queue_io_min(mddev->queue, chunk_size);
3534 if (conf->geo.raid_disks % conf->geo.near_copies)
3535 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3537 blk_queue_io_opt(mddev->queue, chunk_size *
3538 (conf->geo.raid_disks / conf->geo.near_copies));
3541 rdev_for_each(rdev, mddev) {
3543 struct request_queue *q;
3545 disk_idx = rdev->raid_disk;
3548 if (disk_idx >= conf->geo.raid_disks &&
3549 disk_idx >= conf->prev.raid_disks)
3551 disk = conf->mirrors + disk_idx;
3553 if (test_bit(Replacement, &rdev->flags)) {
3554 if (disk->replacement)
3556 disk->replacement = rdev;
3562 q = bdev_get_queue(rdev->bdev);
3563 diff = (rdev->new_data_offset - rdev->data_offset);
3564 if (!mddev->reshape_backwards)
3568 if (first || diff < min_offset_diff)
3569 min_offset_diff = diff;
3572 disk_stack_limits(mddev->gendisk, rdev->bdev,
3573 rdev->data_offset << 9);
3575 disk->head_position = 0;
3577 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3578 discard_supported = true;
3582 if (discard_supported)
3583 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3586 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3589 /* need to check that every block has at least one working mirror */
3590 if (!enough(conf, -1)) {
3591 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3596 if (conf->reshape_progress != MaxSector) {
3597 /* must ensure that shape change is supported */
3598 if (conf->geo.far_copies != 1 &&
3599 conf->geo.far_offset == 0)
3601 if (conf->prev.far_copies != 1 &&
3602 conf->prev.far_offset == 0)
3606 mddev->degraded = 0;
3608 i < conf->geo.raid_disks
3609 || i < conf->prev.raid_disks;
3612 disk = conf->mirrors + i;
3614 if (!disk->rdev && disk->replacement) {
3615 /* The replacement is all we have - use it */
3616 disk->rdev = disk->replacement;
3617 disk->replacement = NULL;
3618 clear_bit(Replacement, &disk->rdev->flags);
3622 !test_bit(In_sync, &disk->rdev->flags)) {
3623 disk->head_position = 0;
3626 disk->rdev->saved_raid_disk < 0)
3629 disk->recovery_disabled = mddev->recovery_disabled - 1;
3632 if (mddev->recovery_cp != MaxSector)
3633 printk(KERN_NOTICE "md/raid10:%s: not clean"
3634 " -- starting background reconstruction\n",
3637 "md/raid10:%s: active with %d out of %d devices\n",
3638 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3639 conf->geo.raid_disks);
3641 * Ok, everything is just fine now
3643 mddev->dev_sectors = conf->dev_sectors;
3644 size = raid10_size(mddev, 0, 0);
3645 md_set_array_sectors(mddev, size);
3646 mddev->resync_max_sectors = size;
3649 int stripe = conf->geo.raid_disks *
3650 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3652 /* Calculate max read-ahead size.
3653 * We need to readahead at least twice a whole stripe....
3656 stripe /= conf->geo.near_copies;
3657 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3658 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3661 if (md_integrity_register(mddev))
3664 if (conf->reshape_progress != MaxSector) {
3665 unsigned long before_length, after_length;
3667 before_length = ((1 << conf->prev.chunk_shift) *
3668 conf->prev.far_copies);
3669 after_length = ((1 << conf->geo.chunk_shift) *
3670 conf->geo.far_copies);
3672 if (max(before_length, after_length) > min_offset_diff) {
3673 /* This cannot work */
3674 printk("md/raid10: offset difference not enough to continue reshape\n");
3677 conf->offset_diff = min_offset_diff;
3679 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3680 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3681 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3682 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3683 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3690 md_unregister_thread(&mddev->thread);
3691 mempool_destroy(conf->r10bio_pool);
3692 safe_put_page(conf->tmppage);
3693 kfree(conf->mirrors);
3695 mddev->private = NULL;
3700 static void raid10_free(struct mddev *mddev, void *priv)
3702 struct r10conf *conf = priv;
3704 mempool_destroy(conf->r10bio_pool);
3705 safe_put_page(conf->tmppage);
3706 kfree(conf->mirrors);
3707 kfree(conf->mirrors_old);
3708 kfree(conf->mirrors_new);
3712 static void raid10_quiesce(struct mddev *mddev, int state)
3714 struct r10conf *conf = mddev->private;
3718 raise_barrier(conf, 0);
3721 lower_barrier(conf);
3726 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3728 /* Resize of 'far' arrays is not supported.
3729 * For 'near' and 'offset' arrays we can set the
3730 * number of sectors used to be an appropriate multiple
3731 * of the chunk size.
3732 * For 'offset', this is far_copies*chunksize.
3733 * For 'near' the multiplier is the LCM of
3734 * near_copies and raid_disks.
3735 * So if far_copies > 1 && !far_offset, fail.
3736 * Else find LCM(raid_disks, near_copy)*far_copies and
3737 * multiply by chunk_size. Then round to this number.
3738 * This is mostly done by raid10_size()
3740 struct r10conf *conf = mddev->private;
3741 sector_t oldsize, size;
3743 if (mddev->reshape_position != MaxSector)
3746 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3749 oldsize = raid10_size(mddev, 0, 0);
3750 size = raid10_size(mddev, sectors, 0);
3751 if (mddev->external_size &&
3752 mddev->array_sectors > size)
3754 if (mddev->bitmap) {
3755 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3759 md_set_array_sectors(mddev, size);
3760 set_capacity(mddev->gendisk, mddev->array_sectors);
3761 revalidate_disk(mddev->gendisk);
3762 if (sectors > mddev->dev_sectors &&
3763 mddev->recovery_cp > oldsize) {
3764 mddev->recovery_cp = oldsize;
3765 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3767 calc_sectors(conf, sectors);
3768 mddev->dev_sectors = conf->dev_sectors;
3769 mddev->resync_max_sectors = size;
3773 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3775 struct md_rdev *rdev;
3776 struct r10conf *conf;
3778 if (mddev->degraded > 0) {
3779 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3781 return ERR_PTR(-EINVAL);
3783 sector_div(size, devs);
3785 /* Set new parameters */
3786 mddev->new_level = 10;
3787 /* new layout: far_copies = 1, near_copies = 2 */
3788 mddev->new_layout = (1<<8) + 2;
3789 mddev->new_chunk_sectors = mddev->chunk_sectors;
3790 mddev->delta_disks = mddev->raid_disks;
3791 mddev->raid_disks *= 2;
3792 /* make sure it will be not marked as dirty */
3793 mddev->recovery_cp = MaxSector;
3794 mddev->dev_sectors = size;
3796 conf = setup_conf(mddev);
3797 if (!IS_ERR(conf)) {
3798 rdev_for_each(rdev, mddev)
3799 if (rdev->raid_disk >= 0) {
3800 rdev->new_raid_disk = rdev->raid_disk * 2;
3801 rdev->sectors = size;
3809 static void *raid10_takeover(struct mddev *mddev)
3811 struct r0conf *raid0_conf;
3813 /* raid10 can take over:
3814 * raid0 - providing it has only two drives
3816 if (mddev->level == 0) {
3817 /* for raid0 takeover only one zone is supported */
3818 raid0_conf = mddev->private;
3819 if (raid0_conf->nr_strip_zones > 1) {
3820 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3821 " with more than one zone.\n",
3823 return ERR_PTR(-EINVAL);
3825 return raid10_takeover_raid0(mddev,
3826 raid0_conf->strip_zone->zone_end,
3827 raid0_conf->strip_zone->nb_dev);
3829 return ERR_PTR(-EINVAL);
3832 static int raid10_check_reshape(struct mddev *mddev)
3834 /* Called when there is a request to change
3835 * - layout (to ->new_layout)
3836 * - chunk size (to ->new_chunk_sectors)
3837 * - raid_disks (by delta_disks)
3838 * or when trying to restart a reshape that was ongoing.
3840 * We need to validate the request and possibly allocate
3841 * space if that might be an issue later.
3843 * Currently we reject any reshape of a 'far' mode array,
3844 * allow chunk size to change if new is generally acceptable,
3845 * allow raid_disks to increase, and allow
3846 * a switch between 'near' mode and 'offset' mode.
3848 struct r10conf *conf = mddev->private;
3851 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3854 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3855 /* mustn't change number of copies */
3857 if (geo.far_copies > 1 && !geo.far_offset)
3858 /* Cannot switch to 'far' mode */
3861 if (mddev->array_sectors & geo.chunk_mask)
3862 /* not factor of array size */
3865 if (!enough(conf, -1))
3868 kfree(conf->mirrors_new);
3869 conf->mirrors_new = NULL;
3870 if (mddev->delta_disks > 0) {
3871 /* allocate new 'mirrors' list */
3872 conf->mirrors_new = kzalloc(
3873 sizeof(struct raid10_info)
3874 *(mddev->raid_disks +
3875 mddev->delta_disks),
3877 if (!conf->mirrors_new)
3884 * Need to check if array has failed when deciding whether to:
3886 * - remove non-faulty devices
3889 * This determination is simple when no reshape is happening.
3890 * However if there is a reshape, we need to carefully check
3891 * both the before and after sections.
3892 * This is because some failed devices may only affect one
3893 * of the two sections, and some non-in_sync devices may
3894 * be insync in the section most affected by failed devices.
3896 static int calc_degraded(struct r10conf *conf)
3898 int degraded, degraded2;
3903 /* 'prev' section first */
3904 for (i = 0; i < conf->prev.raid_disks; i++) {
3905 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3906 if (!rdev || test_bit(Faulty, &rdev->flags))
3908 else if (!test_bit(In_sync, &rdev->flags))
3909 /* When we can reduce the number of devices in
3910 * an array, this might not contribute to
3911 * 'degraded'. It does now.
3916 if (conf->geo.raid_disks == conf->prev.raid_disks)
3920 for (i = 0; i < conf->geo.raid_disks; i++) {
3921 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3922 if (!rdev || test_bit(Faulty, &rdev->flags))
3924 else if (!test_bit(In_sync, &rdev->flags)) {
3925 /* If reshape is increasing the number of devices,
3926 * this section has already been recovered, so
3927 * it doesn't contribute to degraded.
3930 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3935 if (degraded2 > degraded)
3940 static int raid10_start_reshape(struct mddev *mddev)
3942 /* A 'reshape' has been requested. This commits
3943 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3944 * This also checks if there are enough spares and adds them
3946 * We currently require enough spares to make the final
3947 * array non-degraded. We also require that the difference
3948 * between old and new data_offset - on each device - is
3949 * enough that we never risk over-writing.
3952 unsigned long before_length, after_length;
3953 sector_t min_offset_diff = 0;
3956 struct r10conf *conf = mddev->private;
3957 struct md_rdev *rdev;
3961 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3964 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3967 before_length = ((1 << conf->prev.chunk_shift) *
3968 conf->prev.far_copies);
3969 after_length = ((1 << conf->geo.chunk_shift) *
3970 conf->geo.far_copies);
3972 rdev_for_each(rdev, mddev) {
3973 if (!test_bit(In_sync, &rdev->flags)
3974 && !test_bit(Faulty, &rdev->flags))
3976 if (rdev->raid_disk >= 0) {
3977 long long diff = (rdev->new_data_offset
3978 - rdev->data_offset);
3979 if (!mddev->reshape_backwards)
3983 if (first || diff < min_offset_diff)
3984 min_offset_diff = diff;
3988 if (max(before_length, after_length) > min_offset_diff)
3991 if (spares < mddev->delta_disks)
3994 conf->offset_diff = min_offset_diff;
3995 spin_lock_irq(&conf->device_lock);
3996 if (conf->mirrors_new) {
3997 memcpy(conf->mirrors_new, conf->mirrors,
3998 sizeof(struct raid10_info)*conf->prev.raid_disks);
4000 kfree(conf->mirrors_old);
4001 conf->mirrors_old = conf->mirrors;
4002 conf->mirrors = conf->mirrors_new;
4003 conf->mirrors_new = NULL;
4005 setup_geo(&conf->geo, mddev, geo_start);
4007 if (mddev->reshape_backwards) {
4008 sector_t size = raid10_size(mddev, 0, 0);
4009 if (size < mddev->array_sectors) {
4010 spin_unlock_irq(&conf->device_lock);
4011 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4015 mddev->resync_max_sectors = size;
4016 conf->reshape_progress = size;
4018 conf->reshape_progress = 0;
4019 conf->reshape_safe = conf->reshape_progress;
4020 spin_unlock_irq(&conf->device_lock);
4022 if (mddev->delta_disks && mddev->bitmap) {
4023 ret = bitmap_resize(mddev->bitmap,
4024 raid10_size(mddev, 0,
4025 conf->geo.raid_disks),
4030 if (mddev->delta_disks > 0) {
4031 rdev_for_each(rdev, mddev)
4032 if (rdev->raid_disk < 0 &&
4033 !test_bit(Faulty, &rdev->flags)) {
4034 if (raid10_add_disk(mddev, rdev) == 0) {
4035 if (rdev->raid_disk >=
4036 conf->prev.raid_disks)
4037 set_bit(In_sync, &rdev->flags);
4039 rdev->recovery_offset = 0;
4041 if (sysfs_link_rdev(mddev, rdev))
4042 /* Failure here is OK */;
4044 } else if (rdev->raid_disk >= conf->prev.raid_disks
4045 && !test_bit(Faulty, &rdev->flags)) {
4046 /* This is a spare that was manually added */
4047 set_bit(In_sync, &rdev->flags);
4050 /* When a reshape changes the number of devices,
4051 * ->degraded is measured against the larger of the
4052 * pre and post numbers.
4054 spin_lock_irq(&conf->device_lock);
4055 mddev->degraded = calc_degraded(conf);
4056 spin_unlock_irq(&conf->device_lock);
4057 mddev->raid_disks = conf->geo.raid_disks;
4058 mddev->reshape_position = conf->reshape_progress;
4059 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4061 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4062 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4063 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4064 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4065 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4067 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4069 if (!mddev->sync_thread) {
4073 conf->reshape_checkpoint = jiffies;
4074 md_wakeup_thread(mddev->sync_thread);
4075 md_new_event(mddev);
4079 mddev->recovery = 0;
4080 spin_lock_irq(&conf->device_lock);
4081 conf->geo = conf->prev;
4082 mddev->raid_disks = conf->geo.raid_disks;
4083 rdev_for_each(rdev, mddev)
4084 rdev->new_data_offset = rdev->data_offset;
4086 conf->reshape_progress = MaxSector;
4087 conf->reshape_safe = MaxSector;
4088 mddev->reshape_position = MaxSector;
4089 spin_unlock_irq(&conf->device_lock);
4093 /* Calculate the last device-address that could contain
4094 * any block from the chunk that includes the array-address 's'
4095 * and report the next address.
4096 * i.e. the address returned will be chunk-aligned and after
4097 * any data that is in the chunk containing 's'.
4099 static sector_t last_dev_address(sector_t s, struct geom *geo)
4101 s = (s | geo->chunk_mask) + 1;
4102 s >>= geo->chunk_shift;
4103 s *= geo->near_copies;
4104 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4105 s *= geo->far_copies;
4106 s <<= geo->chunk_shift;
4110 /* Calculate the first device-address that could contain
4111 * any block from the chunk that includes the array-address 's'.
4112 * This too will be the start of a chunk
4114 static sector_t first_dev_address(sector_t s, struct geom *geo)
4116 s >>= geo->chunk_shift;
4117 s *= geo->near_copies;
4118 sector_div(s, geo->raid_disks);
4119 s *= geo->far_copies;
4120 s <<= geo->chunk_shift;
4124 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4127 /* We simply copy at most one chunk (smallest of old and new)
4128 * at a time, possibly less if that exceeds RESYNC_PAGES,
4129 * or we hit a bad block or something.
4130 * This might mean we pause for normal IO in the middle of
4131 * a chunk, but that is not a problem as mddev->reshape_position
4132 * can record any location.
4134 * If we will want to write to a location that isn't
4135 * yet recorded as 'safe' (i.e. in metadata on disk) then
4136 * we need to flush all reshape requests and update the metadata.
4138 * When reshaping forwards (e.g. to more devices), we interpret
4139 * 'safe' as the earliest block which might not have been copied
4140 * down yet. We divide this by previous stripe size and multiply
4141 * by previous stripe length to get lowest device offset that we
4142 * cannot write to yet.
4143 * We interpret 'sector_nr' as an address that we want to write to.
4144 * From this we use last_device_address() to find where we might
4145 * write to, and first_device_address on the 'safe' position.
4146 * If this 'next' write position is after the 'safe' position,
4147 * we must update the metadata to increase the 'safe' position.
4149 * When reshaping backwards, we round in the opposite direction
4150 * and perform the reverse test: next write position must not be
4151 * less than current safe position.
4153 * In all this the minimum difference in data offsets
4154 * (conf->offset_diff - always positive) allows a bit of slack,
4155 * so next can be after 'safe', but not by more than offset_diff
4157 * We need to prepare all the bios here before we start any IO
4158 * to ensure the size we choose is acceptable to all devices.
4159 * The means one for each copy for write-out and an extra one for
4161 * We store the read-in bio in ->master_bio and the others in
4162 * ->devs[x].bio and ->devs[x].repl_bio.
4164 struct r10conf *conf = mddev->private;
4165 struct r10bio *r10_bio;
4166 sector_t next, safe, last;
4170 struct md_rdev *rdev;
4173 struct bio *bio, *read_bio;
4174 int sectors_done = 0;
4176 if (sector_nr == 0) {
4177 /* If restarting in the middle, skip the initial sectors */
4178 if (mddev->reshape_backwards &&
4179 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4180 sector_nr = (raid10_size(mddev, 0, 0)
4181 - conf->reshape_progress);
4182 } else if (!mddev->reshape_backwards &&
4183 conf->reshape_progress > 0)
4184 sector_nr = conf->reshape_progress;
4186 mddev->curr_resync_completed = sector_nr;
4187 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4193 /* We don't use sector_nr to track where we are up to
4194 * as that doesn't work well for ->reshape_backwards.
4195 * So just use ->reshape_progress.
4197 if (mddev->reshape_backwards) {
4198 /* 'next' is the earliest device address that we might
4199 * write to for this chunk in the new layout
4201 next = first_dev_address(conf->reshape_progress - 1,
4204 /* 'safe' is the last device address that we might read from
4205 * in the old layout after a restart
4207 safe = last_dev_address(conf->reshape_safe - 1,
4210 if (next + conf->offset_diff < safe)
4213 last = conf->reshape_progress - 1;
4214 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4215 & conf->prev.chunk_mask);
4216 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4217 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4219 /* 'next' is after the last device address that we
4220 * might write to for this chunk in the new layout
4222 next = last_dev_address(conf->reshape_progress, &conf->geo);
4224 /* 'safe' is the earliest device address that we might
4225 * read from in the old layout after a restart
4227 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4229 /* Need to update metadata if 'next' might be beyond 'safe'
4230 * as that would possibly corrupt data
4232 if (next > safe + conf->offset_diff)
4235 sector_nr = conf->reshape_progress;
4236 last = sector_nr | (conf->geo.chunk_mask
4237 & conf->prev.chunk_mask);
4239 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4240 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4244 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4245 /* Need to update reshape_position in metadata */
4247 mddev->reshape_position = conf->reshape_progress;
4248 if (mddev->reshape_backwards)
4249 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4250 - conf->reshape_progress;
4252 mddev->curr_resync_completed = conf->reshape_progress;
4253 conf->reshape_checkpoint = jiffies;
4254 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4255 md_wakeup_thread(mddev->thread);
4256 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4257 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4258 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4259 allow_barrier(conf);
4260 return sectors_done;
4262 conf->reshape_safe = mddev->reshape_position;
4263 allow_barrier(conf);
4267 /* Now schedule reads for blocks from sector_nr to last */
4268 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4270 raise_barrier(conf, sectors_done != 0);
4271 atomic_set(&r10_bio->remaining, 0);
4272 r10_bio->mddev = mddev;
4273 r10_bio->sector = sector_nr;
4274 set_bit(R10BIO_IsReshape, &r10_bio->state);
4275 r10_bio->sectors = last - sector_nr + 1;
4276 rdev = read_balance(conf, r10_bio, &max_sectors);
4277 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4280 /* Cannot read from here, so need to record bad blocks
4281 * on all the target devices.
4284 mempool_free(r10_bio, conf->r10buf_pool);
4285 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4286 return sectors_done;
4289 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4291 read_bio->bi_bdev = rdev->bdev;
4292 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4293 + rdev->data_offset);
4294 read_bio->bi_private = r10_bio;
4295 read_bio->bi_end_io = end_sync_read;
4296 read_bio->bi_rw = READ;
4297 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4298 read_bio->bi_error = 0;
4299 read_bio->bi_vcnt = 0;
4300 read_bio->bi_iter.bi_size = 0;
4301 r10_bio->master_bio = read_bio;
4302 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4304 /* Now find the locations in the new layout */
4305 __raid10_find_phys(&conf->geo, r10_bio);
4308 read_bio->bi_next = NULL;
4310 for (s = 0; s < conf->copies*2; s++) {
4312 int d = r10_bio->devs[s/2].devnum;
4313 struct md_rdev *rdev2;
4315 rdev2 = conf->mirrors[d].replacement;
4316 b = r10_bio->devs[s/2].repl_bio;
4318 rdev2 = conf->mirrors[d].rdev;
4319 b = r10_bio->devs[s/2].bio;
4321 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4325 b->bi_bdev = rdev2->bdev;
4326 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4327 rdev2->new_data_offset;
4328 b->bi_private = r10_bio;
4329 b->bi_end_io = end_reshape_write;
4335 /* Now add as many pages as possible to all of these bios. */
4338 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4339 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4340 int len = (max_sectors - s) << 9;
4341 if (len > PAGE_SIZE)
4343 for (bio = blist; bio ; bio = bio->bi_next) {
4345 if (bio_add_page(bio, page, len, 0))
4348 /* Didn't fit, must stop */
4350 bio2 && bio2 != bio;
4351 bio2 = bio2->bi_next) {
4352 /* Remove last page from this bio */
4354 bio2->bi_iter.bi_size -= len;
4355 bio_clear_flag(bio2, BIO_SEG_VALID);
4359 sector_nr += len >> 9;
4360 nr_sectors += len >> 9;
4363 r10_bio->sectors = nr_sectors;
4365 /* Now submit the read */
4366 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4367 atomic_inc(&r10_bio->remaining);
4368 read_bio->bi_next = NULL;
4369 generic_make_request(read_bio);
4370 sector_nr += nr_sectors;
4371 sectors_done += nr_sectors;
4372 if (sector_nr <= last)
4375 /* Now that we have done the whole section we can
4376 * update reshape_progress
4378 if (mddev->reshape_backwards)
4379 conf->reshape_progress -= sectors_done;
4381 conf->reshape_progress += sectors_done;
4383 return sectors_done;
4386 static void end_reshape_request(struct r10bio *r10_bio);
4387 static int handle_reshape_read_error(struct mddev *mddev,
4388 struct r10bio *r10_bio);
4389 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4391 /* Reshape read completed. Hopefully we have a block
4393 * If we got a read error then we do sync 1-page reads from
4394 * elsewhere until we find the data - or give up.
4396 struct r10conf *conf = mddev->private;
4399 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4400 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4401 /* Reshape has been aborted */
4402 md_done_sync(mddev, r10_bio->sectors, 0);
4406 /* We definitely have the data in the pages, schedule the
4409 atomic_set(&r10_bio->remaining, 1);
4410 for (s = 0; s < conf->copies*2; s++) {
4412 int d = r10_bio->devs[s/2].devnum;
4413 struct md_rdev *rdev;
4415 rdev = conf->mirrors[d].replacement;
4416 b = r10_bio->devs[s/2].repl_bio;
4418 rdev = conf->mirrors[d].rdev;
4419 b = r10_bio->devs[s/2].bio;
4421 if (!rdev || test_bit(Faulty, &rdev->flags))
4423 atomic_inc(&rdev->nr_pending);
4424 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4425 atomic_inc(&r10_bio->remaining);
4427 generic_make_request(b);
4429 end_reshape_request(r10_bio);
4432 static void end_reshape(struct r10conf *conf)
4434 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4437 spin_lock_irq(&conf->device_lock);
4438 conf->prev = conf->geo;
4439 md_finish_reshape(conf->mddev);
4441 conf->reshape_progress = MaxSector;
4442 conf->reshape_safe = MaxSector;
4443 spin_unlock_irq(&conf->device_lock);
4445 /* read-ahead size must cover two whole stripes, which is
4446 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4448 if (conf->mddev->queue) {
4449 int stripe = conf->geo.raid_disks *
4450 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4451 stripe /= conf->geo.near_copies;
4452 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4453 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4458 static int handle_reshape_read_error(struct mddev *mddev,
4459 struct r10bio *r10_bio)
4461 /* Use sync reads to get the blocks from somewhere else */
4462 int sectors = r10_bio->sectors;
4463 struct r10conf *conf = mddev->private;
4465 struct r10bio r10_bio;
4466 struct r10dev devs[conf->copies];
4468 struct r10bio *r10b = &on_stack.r10_bio;
4471 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4473 r10b->sector = r10_bio->sector;
4474 __raid10_find_phys(&conf->prev, r10b);
4479 int first_slot = slot;
4481 if (s > (PAGE_SIZE >> 9))
4485 int d = r10b->devs[slot].devnum;
4486 struct md_rdev *rdev = conf->mirrors[d].rdev;
4489 test_bit(Faulty, &rdev->flags) ||
4490 !test_bit(In_sync, &rdev->flags))
4493 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4494 success = sync_page_io(rdev,
4503 if (slot >= conf->copies)
4505 if (slot == first_slot)
4509 /* couldn't read this block, must give up */
4510 set_bit(MD_RECOVERY_INTR,
4520 static void end_reshape_write(struct bio *bio)
4522 struct r10bio *r10_bio = bio->bi_private;
4523 struct mddev *mddev = r10_bio->mddev;
4524 struct r10conf *conf = mddev->private;
4528 struct md_rdev *rdev = NULL;
4530 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4532 rdev = conf->mirrors[d].replacement;
4535 rdev = conf->mirrors[d].rdev;
4538 if (bio->bi_error) {
4539 /* FIXME should record badblock */
4540 md_error(mddev, rdev);
4543 rdev_dec_pending(rdev, mddev);
4544 end_reshape_request(r10_bio);
4547 static void end_reshape_request(struct r10bio *r10_bio)
4549 if (!atomic_dec_and_test(&r10_bio->remaining))
4551 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4552 bio_put(r10_bio->master_bio);
4556 static void raid10_finish_reshape(struct mddev *mddev)
4558 struct r10conf *conf = mddev->private;
4560 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4563 if (mddev->delta_disks > 0) {
4564 sector_t size = raid10_size(mddev, 0, 0);
4565 md_set_array_sectors(mddev, size);
4566 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4567 mddev->recovery_cp = mddev->resync_max_sectors;
4568 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4570 mddev->resync_max_sectors = size;
4571 set_capacity(mddev->gendisk, mddev->array_sectors);
4572 revalidate_disk(mddev->gendisk);
4575 for (d = conf->geo.raid_disks ;
4576 d < conf->geo.raid_disks - mddev->delta_disks;
4578 struct md_rdev *rdev = conf->mirrors[d].rdev;
4580 clear_bit(In_sync, &rdev->flags);
4581 rdev = conf->mirrors[d].replacement;
4583 clear_bit(In_sync, &rdev->flags);
4586 mddev->layout = mddev->new_layout;
4587 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4588 mddev->reshape_position = MaxSector;
4589 mddev->delta_disks = 0;
4590 mddev->reshape_backwards = 0;
4593 static struct md_personality raid10_personality =
4597 .owner = THIS_MODULE,
4598 .make_request = make_request,
4600 .free = raid10_free,
4602 .error_handler = error,
4603 .hot_add_disk = raid10_add_disk,
4604 .hot_remove_disk= raid10_remove_disk,
4605 .spare_active = raid10_spare_active,
4606 .sync_request = sync_request,
4607 .quiesce = raid10_quiesce,
4608 .size = raid10_size,
4609 .resize = raid10_resize,
4610 .takeover = raid10_takeover,
4611 .check_reshape = raid10_check_reshape,
4612 .start_reshape = raid10_start_reshape,
4613 .finish_reshape = raid10_finish_reshape,
4614 .congested = raid10_congested,
4617 static int __init raid_init(void)
4619 return register_md_personality(&raid10_personality);
4622 static void raid_exit(void)
4624 unregister_md_personality(&raid10_personality);
4627 module_init(raid_init);
4628 module_exit(raid_exit);
4629 MODULE_LICENSE("GPL");
4630 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4631 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4632 MODULE_ALIAS("md-raid10");
4633 MODULE_ALIAS("md-level-10");
4635 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);