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, int error);
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 clear_bit(BIO_UPTODATE, &bio->bi_flags);
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, int error)
363 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364 struct r10bio *r10_bio = bio->bi_private;
366 struct md_rdev *rdev;
367 struct r10conf *conf = r10_bio->mddev->private;
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot, r10_bio);
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
395 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
400 raid_end_bio_io(r10_bio);
401 rdev_dec_pending(rdev, conf->mddev);
404 * oops, read error - keep the refcount on the rdev
406 char b[BDEVNAME_SIZE];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
410 bdevname(rdev->bdev, b),
411 (unsigned long long)r10_bio->sector);
412 set_bit(R10BIO_ReadError, &r10_bio->state);
413 reschedule_retry(r10_bio);
417 static void close_write(struct r10bio *r10_bio)
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
424 md_write_end(r10_bio->mddev);
427 static void one_write_done(struct r10bio *r10_bio)
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
437 raid_end_bio_io(r10_bio);
442 static void raid10_end_write_request(struct bio *bio, int error)
444 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
445 struct r10bio *r10_bio = bio->bi_private;
448 struct r10conf *conf = r10_bio->mddev->private;
450 struct md_rdev *rdev = NULL;
452 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
455 rdev = conf->mirrors[dev].replacement;
459 rdev = conf->mirrors[dev].rdev;
462 * this branch is our 'one mirror IO has finished' event handler:
466 /* Never record new bad blocks to replacement,
469 md_error(rdev->mddev, rdev);
471 set_bit(WriteErrorSeen, &rdev->flags);
472 if (!test_and_set_bit(WantReplacement, &rdev->flags))
473 set_bit(MD_RECOVERY_NEEDED,
474 &rdev->mddev->recovery);
475 set_bit(R10BIO_WriteError, &r10_bio->state);
480 * Set R10BIO_Uptodate in our master bio, so that
481 * we will return a good error code for to the higher
482 * levels even if IO on some other mirrored buffer fails.
484 * The 'master' represents the composite IO operation to
485 * user-side. So if something waits for IO, then it will
486 * wait for the 'master' bio.
492 * Do not set R10BIO_Uptodate if the current device is
493 * rebuilding or Faulty. This is because we cannot use
494 * such device for properly reading the data back (we could
495 * potentially use it, if the current write would have felt
496 * before rdev->recovery_offset, but for simplicity we don't
499 if (test_bit(In_sync, &rdev->flags) &&
500 !test_bit(Faulty, &rdev->flags))
501 set_bit(R10BIO_Uptodate, &r10_bio->state);
503 /* Maybe we can clear some bad blocks. */
504 if (is_badblock(rdev,
505 r10_bio->devs[slot].addr,
507 &first_bad, &bad_sectors)) {
510 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
512 r10_bio->devs[slot].bio = IO_MADE_GOOD;
514 set_bit(R10BIO_MadeGood, &r10_bio->state);
520 * Let's see if all mirrored write operations have finished
523 one_write_done(r10_bio);
525 rdev_dec_pending(rdev, conf->mddev);
529 * RAID10 layout manager
530 * As well as the chunksize and raid_disks count, there are two
531 * parameters: near_copies and far_copies.
532 * near_copies * far_copies must be <= raid_disks.
533 * Normally one of these will be 1.
534 * If both are 1, we get raid0.
535 * If near_copies == raid_disks, we get raid1.
537 * Chunks are laid out in raid0 style with near_copies copies of the
538 * first chunk, followed by near_copies copies of the next chunk and
540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
541 * as described above, we start again with a device offset of near_copies.
542 * So we effectively have another copy of the whole array further down all
543 * the drives, but with blocks on different drives.
544 * With this layout, and block is never stored twice on the one device.
546 * raid10_find_phys finds the sector offset of a given virtual sector
547 * on each device that it is on.
549 * raid10_find_virt does the reverse mapping, from a device and a
550 * sector offset to a virtual address
553 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
561 int last_far_set_start, last_far_set_size;
563 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
564 last_far_set_start *= geo->far_set_size;
566 last_far_set_size = geo->far_set_size;
567 last_far_set_size += (geo->raid_disks % geo->far_set_size);
569 /* now calculate first sector/dev */
570 chunk = r10bio->sector >> geo->chunk_shift;
571 sector = r10bio->sector & geo->chunk_mask;
573 chunk *= geo->near_copies;
575 dev = sector_div(stripe, geo->raid_disks);
577 stripe *= geo->far_copies;
579 sector += stripe << geo->chunk_shift;
581 /* and calculate all the others */
582 for (n = 0; n < geo->near_copies; n++) {
586 r10bio->devs[slot].devnum = d;
587 r10bio->devs[slot].addr = s;
590 for (f = 1; f < geo->far_copies; f++) {
591 set = d / geo->far_set_size;
592 d += geo->near_copies;
594 if ((geo->raid_disks % geo->far_set_size) &&
595 (d > last_far_set_start)) {
596 d -= last_far_set_start;
597 d %= last_far_set_size;
598 d += last_far_set_start;
600 d %= geo->far_set_size;
601 d += geo->far_set_size * set;
604 r10bio->devs[slot].devnum = d;
605 r10bio->devs[slot].addr = s;
609 if (dev >= geo->raid_disks) {
611 sector += (geo->chunk_mask + 1);
616 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
618 struct geom *geo = &conf->geo;
620 if (conf->reshape_progress != MaxSector &&
621 ((r10bio->sector >= conf->reshape_progress) !=
622 conf->mddev->reshape_backwards)) {
623 set_bit(R10BIO_Previous, &r10bio->state);
626 clear_bit(R10BIO_Previous, &r10bio->state);
628 __raid10_find_phys(geo, r10bio);
631 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
633 sector_t offset, chunk, vchunk;
634 /* Never use conf->prev as this is only called during resync
635 * or recovery, so reshape isn't happening
637 struct geom *geo = &conf->geo;
638 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
639 int far_set_size = geo->far_set_size;
640 int last_far_set_start;
642 if (geo->raid_disks % geo->far_set_size) {
643 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
644 last_far_set_start *= geo->far_set_size;
646 if (dev >= last_far_set_start) {
647 far_set_size = geo->far_set_size;
648 far_set_size += (geo->raid_disks % geo->far_set_size);
649 far_set_start = last_far_set_start;
653 offset = sector & geo->chunk_mask;
654 if (geo->far_offset) {
656 chunk = sector >> geo->chunk_shift;
657 fc = sector_div(chunk, geo->far_copies);
658 dev -= fc * geo->near_copies;
659 if (dev < far_set_start)
662 while (sector >= geo->stride) {
663 sector -= geo->stride;
664 if (dev < (geo->near_copies + far_set_start))
665 dev += far_set_size - geo->near_copies;
667 dev -= geo->near_copies;
669 chunk = sector >> geo->chunk_shift;
671 vchunk = chunk * geo->raid_disks + dev;
672 sector_div(vchunk, geo->near_copies);
673 return (vchunk << geo->chunk_shift) + offset;
677 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
679 * @bvm: properties of new bio
680 * @biovec: the request that could be merged to it.
682 * Return amount of bytes we can accept at this offset
683 * This requires checking for end-of-chunk if near_copies != raid_disks,
684 * and for subordinate merge_bvec_fns if merge_check_needed.
686 static int raid10_mergeable_bvec(struct request_queue *q,
687 struct bvec_merge_data *bvm,
688 struct bio_vec *biovec)
690 struct mddev *mddev = q->queuedata;
691 struct r10conf *conf = mddev->private;
692 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
694 unsigned int chunk_sectors;
695 unsigned int bio_sectors = bvm->bi_size >> 9;
696 struct geom *geo = &conf->geo;
698 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
699 if (conf->reshape_progress != MaxSector &&
700 ((sector >= conf->reshape_progress) !=
701 conf->mddev->reshape_backwards))
704 if (geo->near_copies < geo->raid_disks) {
705 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
706 + bio_sectors)) << 9;
708 /* bio_add cannot handle a negative return */
710 if (max <= biovec->bv_len && bio_sectors == 0)
711 return biovec->bv_len;
713 max = biovec->bv_len;
715 if (mddev->merge_check_needed) {
717 struct r10bio r10_bio;
718 struct r10dev devs[conf->copies];
720 struct r10bio *r10_bio = &on_stack.r10_bio;
722 if (conf->reshape_progress != MaxSector) {
723 /* Cannot give any guidance during reshape */
724 if (max <= biovec->bv_len && bio_sectors == 0)
725 return biovec->bv_len;
728 r10_bio->sector = sector;
729 raid10_find_phys(conf, r10_bio);
731 for (s = 0; s < conf->copies; s++) {
732 int disk = r10_bio->devs[s].devnum;
733 struct md_rdev *rdev = rcu_dereference(
734 conf->mirrors[disk].rdev);
735 if (rdev && !test_bit(Faulty, &rdev->flags)) {
736 struct request_queue *q =
737 bdev_get_queue(rdev->bdev);
738 if (q->merge_bvec_fn) {
739 bvm->bi_sector = r10_bio->devs[s].addr
741 bvm->bi_bdev = rdev->bdev;
742 max = min(max, q->merge_bvec_fn(
746 rdev = rcu_dereference(conf->mirrors[disk].replacement);
747 if (rdev && !test_bit(Faulty, &rdev->flags)) {
748 struct request_queue *q =
749 bdev_get_queue(rdev->bdev);
750 if (q->merge_bvec_fn) {
751 bvm->bi_sector = r10_bio->devs[s].addr
753 bvm->bi_bdev = rdev->bdev;
754 max = min(max, q->merge_bvec_fn(
765 * This routine returns the disk from which the requested read should
766 * be done. There is a per-array 'next expected sequential IO' sector
767 * number - if this matches on the next IO then we use the last disk.
768 * There is also a per-disk 'last know head position' sector that is
769 * maintained from IRQ contexts, both the normal and the resync IO
770 * completion handlers update this position correctly. If there is no
771 * perfect sequential match then we pick the disk whose head is closest.
773 * If there are 2 mirrors in the same 2 devices, performance degrades
774 * because position is mirror, not device based.
776 * The rdev for the device selected will have nr_pending incremented.
780 * FIXME: possibly should rethink readbalancing and do it differently
781 * depending on near_copies / far_copies geometry.
783 static struct md_rdev *read_balance(struct r10conf *conf,
784 struct r10bio *r10_bio,
787 const sector_t this_sector = r10_bio->sector;
789 int sectors = r10_bio->sectors;
790 int best_good_sectors;
791 sector_t new_distance, best_dist;
792 struct md_rdev *best_rdev, *rdev = NULL;
795 struct geom *geo = &conf->geo;
797 raid10_find_phys(conf, r10_bio);
800 sectors = r10_bio->sectors;
803 best_dist = MaxSector;
804 best_good_sectors = 0;
807 * Check if we can balance. We can balance on the whole
808 * device if no resync is going on (recovery is ok), or below
809 * the resync window. We take the first readable disk when
810 * above the resync window.
812 if (conf->mddev->recovery_cp < MaxSector
813 && (this_sector + sectors >= conf->next_resync))
816 for (slot = 0; slot < conf->copies ; slot++) {
821 if (r10_bio->devs[slot].bio == IO_BLOCKED)
823 disk = r10_bio->devs[slot].devnum;
824 rdev = rcu_dereference(conf->mirrors[disk].replacement);
825 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
826 test_bit(Unmerged, &rdev->flags) ||
827 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
828 rdev = rcu_dereference(conf->mirrors[disk].rdev);
830 test_bit(Faulty, &rdev->flags) ||
831 test_bit(Unmerged, &rdev->flags))
833 if (!test_bit(In_sync, &rdev->flags) &&
834 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
837 dev_sector = r10_bio->devs[slot].addr;
838 if (is_badblock(rdev, dev_sector, sectors,
839 &first_bad, &bad_sectors)) {
840 if (best_dist < MaxSector)
841 /* Already have a better slot */
843 if (first_bad <= dev_sector) {
844 /* Cannot read here. If this is the
845 * 'primary' device, then we must not read
846 * beyond 'bad_sectors' from another device.
848 bad_sectors -= (dev_sector - first_bad);
849 if (!do_balance && sectors > bad_sectors)
850 sectors = bad_sectors;
851 if (best_good_sectors > sectors)
852 best_good_sectors = sectors;
854 sector_t good_sectors =
855 first_bad - dev_sector;
856 if (good_sectors > best_good_sectors) {
857 best_good_sectors = good_sectors;
862 /* Must read from here */
867 best_good_sectors = sectors;
872 /* This optimisation is debatable, and completely destroys
873 * sequential read speed for 'far copies' arrays. So only
874 * keep it for 'near' arrays, and review those later.
876 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
879 /* for far > 1 always use the lowest address */
880 if (geo->far_copies > 1)
881 new_distance = r10_bio->devs[slot].addr;
883 new_distance = abs(r10_bio->devs[slot].addr -
884 conf->mirrors[disk].head_position);
885 if (new_distance < best_dist) {
886 best_dist = new_distance;
891 if (slot >= conf->copies) {
897 atomic_inc(&rdev->nr_pending);
898 if (test_bit(Faulty, &rdev->flags)) {
899 /* Cannot risk returning a device that failed
900 * before we inc'ed nr_pending
902 rdev_dec_pending(rdev, conf->mddev);
905 r10_bio->read_slot = slot;
909 *max_sectors = best_good_sectors;
914 int md_raid10_congested(struct mddev *mddev, int bits)
916 struct r10conf *conf = mddev->private;
919 if ((bits & (1 << BDI_async_congested)) &&
920 conf->pending_count >= max_queued_requests)
925 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
928 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
929 if (rdev && !test_bit(Faulty, &rdev->flags)) {
930 struct request_queue *q = bdev_get_queue(rdev->bdev);
932 ret |= bdi_congested(&q->backing_dev_info, bits);
938 EXPORT_SYMBOL_GPL(md_raid10_congested);
940 static int raid10_congested(void *data, int bits)
942 struct mddev *mddev = data;
944 return mddev_congested(mddev, bits) ||
945 md_raid10_congested(mddev, bits);
948 static void flush_pending_writes(struct r10conf *conf)
950 /* Any writes that have been queued but are awaiting
951 * bitmap updates get flushed here.
953 spin_lock_irq(&conf->device_lock);
955 if (conf->pending_bio_list.head) {
957 bio = bio_list_get(&conf->pending_bio_list);
958 conf->pending_count = 0;
959 spin_unlock_irq(&conf->device_lock);
960 /* flush any pending bitmap writes to disk
961 * before proceeding w/ I/O */
962 bitmap_unplug(conf->mddev->bitmap);
963 wake_up(&conf->wait_barrier);
965 while (bio) { /* submit pending writes */
966 struct bio *next = bio->bi_next;
968 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
969 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
973 generic_make_request(bio);
977 spin_unlock_irq(&conf->device_lock);
981 * Sometimes we need to suspend IO while we do something else,
982 * either some resync/recovery, or reconfigure the array.
983 * To do this we raise a 'barrier'.
984 * The 'barrier' is a counter that can be raised multiple times
985 * to count how many activities are happening which preclude
987 * We can only raise the barrier if there is no pending IO.
988 * i.e. if nr_pending == 0.
989 * We choose only to raise the barrier if no-one is waiting for the
990 * barrier to go down. This means that as soon as an IO request
991 * is ready, no other operations which require a barrier will start
992 * until the IO request has had a chance.
994 * So: regular IO calls 'wait_barrier'. When that returns there
995 * is no backgroup IO happening, It must arrange to call
996 * allow_barrier when it has finished its IO.
997 * backgroup IO calls must call raise_barrier. Once that returns
998 * there is no normal IO happeing. It must arrange to call
999 * lower_barrier when the particular background IO completes.
1002 static void raise_barrier(struct r10conf *conf, int force)
1004 BUG_ON(force && !conf->barrier);
1005 spin_lock_irq(&conf->resync_lock);
1007 /* Wait until no block IO is waiting (unless 'force') */
1008 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1011 /* block any new IO from starting */
1014 /* Now wait for all pending IO to complete */
1015 wait_event_lock_irq(conf->wait_barrier,
1016 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1019 spin_unlock_irq(&conf->resync_lock);
1022 static void lower_barrier(struct r10conf *conf)
1024 unsigned long flags;
1025 spin_lock_irqsave(&conf->resync_lock, flags);
1027 spin_unlock_irqrestore(&conf->resync_lock, flags);
1028 wake_up(&conf->wait_barrier);
1031 static void wait_barrier(struct r10conf *conf)
1033 spin_lock_irq(&conf->resync_lock);
1034 if (conf->barrier) {
1036 /* Wait for the barrier to drop.
1037 * However if there are already pending
1038 * requests (preventing the barrier from
1039 * rising completely), and the
1040 * pre-process bio queue isn't empty,
1041 * then don't wait, as we need to empty
1042 * that queue to get the nr_pending
1045 wait_event_lock_irq(conf->wait_barrier,
1047 (conf->nr_pending &&
1048 current->bio_list &&
1049 !bio_list_empty(current->bio_list)),
1054 spin_unlock_irq(&conf->resync_lock);
1057 static void allow_barrier(struct r10conf *conf)
1059 unsigned long flags;
1060 spin_lock_irqsave(&conf->resync_lock, flags);
1062 spin_unlock_irqrestore(&conf->resync_lock, flags);
1063 wake_up(&conf->wait_barrier);
1066 static void freeze_array(struct r10conf *conf, int extra)
1068 /* stop syncio and normal IO and wait for everything to
1070 * We increment barrier and nr_waiting, and then
1071 * wait until nr_pending match nr_queued+extra
1072 * This is called in the context of one normal IO request
1073 * that has failed. Thus any sync request that might be pending
1074 * will be blocked by nr_pending, and we need to wait for
1075 * pending IO requests to complete or be queued for re-try.
1076 * Thus the number queued (nr_queued) plus this request (extra)
1077 * must match the number of pending IOs (nr_pending) before
1080 spin_lock_irq(&conf->resync_lock);
1083 wait_event_lock_irq_cmd(conf->wait_barrier,
1084 conf->nr_pending == conf->nr_queued+extra,
1086 flush_pending_writes(conf));
1088 spin_unlock_irq(&conf->resync_lock);
1091 static void unfreeze_array(struct r10conf *conf)
1093 /* reverse the effect of the freeze */
1094 spin_lock_irq(&conf->resync_lock);
1097 wake_up(&conf->wait_barrier);
1098 spin_unlock_irq(&conf->resync_lock);
1101 static sector_t choose_data_offset(struct r10bio *r10_bio,
1102 struct md_rdev *rdev)
1104 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1105 test_bit(R10BIO_Previous, &r10_bio->state))
1106 return rdev->data_offset;
1108 return rdev->new_data_offset;
1111 struct raid10_plug_cb {
1112 struct blk_plug_cb cb;
1113 struct bio_list pending;
1117 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1119 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1121 struct mddev *mddev = plug->cb.data;
1122 struct r10conf *conf = mddev->private;
1125 if (from_schedule || current->bio_list) {
1126 spin_lock_irq(&conf->device_lock);
1127 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1128 conf->pending_count += plug->pending_cnt;
1129 spin_unlock_irq(&conf->device_lock);
1130 wake_up(&conf->wait_barrier);
1131 md_wakeup_thread(mddev->thread);
1136 /* we aren't scheduling, so we can do the write-out directly. */
1137 bio = bio_list_get(&plug->pending);
1138 bitmap_unplug(mddev->bitmap);
1139 wake_up(&conf->wait_barrier);
1141 while (bio) { /* submit pending writes */
1142 struct bio *next = bio->bi_next;
1143 bio->bi_next = NULL;
1144 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1145 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1146 /* Just ignore it */
1149 generic_make_request(bio);
1155 static void make_request(struct mddev *mddev, struct bio * bio)
1157 struct r10conf *conf = mddev->private;
1158 struct r10bio *r10_bio;
1159 struct bio *read_bio;
1161 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1162 int chunk_sects = chunk_mask + 1;
1163 const int rw = bio_data_dir(bio);
1164 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1165 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1166 const unsigned long do_discard = (bio->bi_rw
1167 & (REQ_DISCARD | REQ_SECURE));
1168 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1169 unsigned long flags;
1170 struct md_rdev *blocked_rdev;
1171 struct blk_plug_cb *cb;
1172 struct raid10_plug_cb *plug = NULL;
1173 int sectors_handled;
1177 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1178 md_flush_request(mddev, bio);
1182 /* If this request crosses a chunk boundary, we need to
1183 * split it. This will only happen for 1 PAGE (or less) requests.
1185 if (unlikely((bio->bi_sector & chunk_mask) + bio_sectors(bio)
1187 && (conf->geo.near_copies < conf->geo.raid_disks
1188 || conf->prev.near_copies < conf->prev.raid_disks))) {
1189 struct bio_pair *bp;
1190 /* Sanity check -- queue functions should prevent this happening */
1191 if (bio_segments(bio) > 1)
1193 /* This is a one page bio that upper layers
1194 * refuse to split for us, so we need to split it.
1197 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1199 /* Each of these 'make_request' calls will call 'wait_barrier'.
1200 * If the first succeeds but the second blocks due to the resync
1201 * thread raising the barrier, we will deadlock because the
1202 * IO to the underlying device will be queued in generic_make_request
1203 * and will never complete, so will never reduce nr_pending.
1204 * So increment nr_waiting here so no new raise_barriers will
1205 * succeed, and so the second wait_barrier cannot block.
1207 spin_lock_irq(&conf->resync_lock);
1209 spin_unlock_irq(&conf->resync_lock);
1211 make_request(mddev, &bp->bio1);
1212 make_request(mddev, &bp->bio2);
1214 spin_lock_irq(&conf->resync_lock);
1216 wake_up(&conf->wait_barrier);
1217 spin_unlock_irq(&conf->resync_lock);
1219 bio_pair_release(bp);
1222 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1223 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1224 (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1230 md_write_start(mddev, bio);
1233 * Register the new request and wait if the reconstruction
1234 * thread has put up a bar for new requests.
1235 * Continue immediately if no resync is active currently.
1239 sectors = bio_sectors(bio);
1240 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1241 bio->bi_sector < conf->reshape_progress &&
1242 bio->bi_sector + sectors > conf->reshape_progress) {
1243 /* IO spans the reshape position. Need to wait for
1246 allow_barrier(conf);
1247 wait_event(conf->wait_barrier,
1248 conf->reshape_progress <= bio->bi_sector ||
1249 conf->reshape_progress >= bio->bi_sector + sectors);
1252 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1253 bio_data_dir(bio) == WRITE &&
1254 (mddev->reshape_backwards
1255 ? (bio->bi_sector < conf->reshape_safe &&
1256 bio->bi_sector + sectors > conf->reshape_progress)
1257 : (bio->bi_sector + sectors > conf->reshape_safe &&
1258 bio->bi_sector < conf->reshape_progress))) {
1259 /* Need to update reshape_position in metadata */
1260 mddev->reshape_position = conf->reshape_progress;
1261 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1262 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1263 md_wakeup_thread(mddev->thread);
1264 wait_event(mddev->sb_wait,
1265 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1267 conf->reshape_safe = mddev->reshape_position;
1270 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1272 r10_bio->master_bio = bio;
1273 r10_bio->sectors = sectors;
1275 r10_bio->mddev = mddev;
1276 r10_bio->sector = bio->bi_sector;
1279 /* We might need to issue multiple reads to different
1280 * devices if there are bad blocks around, so we keep
1281 * track of the number of reads in bio->bi_phys_segments.
1282 * If this is 0, there is only one r10_bio and no locking
1283 * will be needed when the request completes. If it is
1284 * non-zero, then it is the number of not-completed requests.
1286 bio->bi_phys_segments = 0;
1287 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1291 * read balancing logic:
1293 struct md_rdev *rdev;
1297 rdev = read_balance(conf, r10_bio, &max_sectors);
1299 raid_end_bio_io(r10_bio);
1302 slot = r10_bio->read_slot;
1304 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1305 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1308 r10_bio->devs[slot].bio = read_bio;
1309 r10_bio->devs[slot].rdev = rdev;
1311 read_bio->bi_sector = r10_bio->devs[slot].addr +
1312 choose_data_offset(r10_bio, rdev);
1313 read_bio->bi_bdev = rdev->bdev;
1314 read_bio->bi_end_io = raid10_end_read_request;
1315 read_bio->bi_rw = READ | do_sync;
1316 read_bio->bi_private = r10_bio;
1318 if (max_sectors < r10_bio->sectors) {
1319 /* Could not read all from this device, so we will
1320 * need another r10_bio.
1322 sectors_handled = (r10_bio->sectors + max_sectors
1324 r10_bio->sectors = max_sectors;
1325 spin_lock_irq(&conf->device_lock);
1326 if (bio->bi_phys_segments == 0)
1327 bio->bi_phys_segments = 2;
1329 bio->bi_phys_segments++;
1330 spin_unlock(&conf->device_lock);
1331 /* Cannot call generic_make_request directly
1332 * as that will be queued in __generic_make_request
1333 * and subsequent mempool_alloc might block
1334 * waiting for it. so hand bio over to raid10d.
1336 reschedule_retry(r10_bio);
1338 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1340 r10_bio->master_bio = bio;
1341 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1343 r10_bio->mddev = mddev;
1344 r10_bio->sector = bio->bi_sector + sectors_handled;
1347 generic_make_request(read_bio);
1354 if (conf->pending_count >= max_queued_requests) {
1355 md_wakeup_thread(mddev->thread);
1356 wait_event(conf->wait_barrier,
1357 conf->pending_count < max_queued_requests);
1359 /* first select target devices under rcu_lock and
1360 * inc refcount on their rdev. Record them by setting
1362 * If there are known/acknowledged bad blocks on any device
1363 * on which we have seen a write error, we want to avoid
1364 * writing to those blocks. This potentially requires several
1365 * writes to write around the bad blocks. Each set of writes
1366 * gets its own r10_bio with a set of bios attached. The number
1367 * of r10_bios is recored in bio->bi_phys_segments just as with
1371 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1372 raid10_find_phys(conf, r10_bio);
1374 blocked_rdev = NULL;
1376 max_sectors = r10_bio->sectors;
1378 for (i = 0; i < conf->copies; i++) {
1379 int d = r10_bio->devs[i].devnum;
1380 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1381 struct md_rdev *rrdev = rcu_dereference(
1382 conf->mirrors[d].replacement);
1385 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1386 atomic_inc(&rdev->nr_pending);
1387 blocked_rdev = rdev;
1390 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1391 atomic_inc(&rrdev->nr_pending);
1392 blocked_rdev = rrdev;
1395 if (rdev && (test_bit(Faulty, &rdev->flags)
1396 || test_bit(Unmerged, &rdev->flags)))
1398 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1399 || test_bit(Unmerged, &rrdev->flags)))
1402 r10_bio->devs[i].bio = NULL;
1403 r10_bio->devs[i].repl_bio = NULL;
1405 if (!rdev && !rrdev) {
1406 set_bit(R10BIO_Degraded, &r10_bio->state);
1409 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1411 sector_t dev_sector = r10_bio->devs[i].addr;
1415 is_bad = is_badblock(rdev, dev_sector,
1417 &first_bad, &bad_sectors);
1419 /* Mustn't write here until the bad block
1422 atomic_inc(&rdev->nr_pending);
1423 set_bit(BlockedBadBlocks, &rdev->flags);
1424 blocked_rdev = rdev;
1427 if (is_bad && first_bad <= dev_sector) {
1428 /* Cannot write here at all */
1429 bad_sectors -= (dev_sector - first_bad);
1430 if (bad_sectors < max_sectors)
1431 /* Mustn't write more than bad_sectors
1432 * to other devices yet
1434 max_sectors = bad_sectors;
1435 /* We don't set R10BIO_Degraded as that
1436 * only applies if the disk is missing,
1437 * so it might be re-added, and we want to
1438 * know to recover this chunk.
1439 * In this case the device is here, and the
1440 * fact that this chunk is not in-sync is
1441 * recorded in the bad block log.
1446 int good_sectors = first_bad - dev_sector;
1447 if (good_sectors < max_sectors)
1448 max_sectors = good_sectors;
1452 r10_bio->devs[i].bio = bio;
1453 atomic_inc(&rdev->nr_pending);
1456 r10_bio->devs[i].repl_bio = bio;
1457 atomic_inc(&rrdev->nr_pending);
1462 if (unlikely(blocked_rdev)) {
1463 /* Have to wait for this device to get unblocked, then retry */
1467 for (j = 0; j < i; j++) {
1468 if (r10_bio->devs[j].bio) {
1469 d = r10_bio->devs[j].devnum;
1470 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1472 if (r10_bio->devs[j].repl_bio) {
1473 struct md_rdev *rdev;
1474 d = r10_bio->devs[j].devnum;
1475 rdev = conf->mirrors[d].replacement;
1477 /* Race with remove_disk */
1479 rdev = conf->mirrors[d].rdev;
1481 rdev_dec_pending(rdev, mddev);
1484 allow_barrier(conf);
1485 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1490 if (max_sectors < r10_bio->sectors) {
1491 /* We are splitting this into multiple parts, so
1492 * we need to prepare for allocating another r10_bio.
1494 r10_bio->sectors = max_sectors;
1495 spin_lock_irq(&conf->device_lock);
1496 if (bio->bi_phys_segments == 0)
1497 bio->bi_phys_segments = 2;
1499 bio->bi_phys_segments++;
1500 spin_unlock_irq(&conf->device_lock);
1502 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1504 atomic_set(&r10_bio->remaining, 1);
1505 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1507 for (i = 0; i < conf->copies; i++) {
1509 int d = r10_bio->devs[i].devnum;
1510 if (r10_bio->devs[i].bio) {
1511 struct md_rdev *rdev = conf->mirrors[d].rdev;
1512 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1513 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1515 r10_bio->devs[i].bio = mbio;
1517 mbio->bi_sector = (r10_bio->devs[i].addr+
1518 choose_data_offset(r10_bio,
1520 mbio->bi_bdev = rdev->bdev;
1521 mbio->bi_end_io = raid10_end_write_request;
1523 WRITE | do_sync | do_fua | do_discard | do_same;
1524 mbio->bi_private = r10_bio;
1526 atomic_inc(&r10_bio->remaining);
1528 cb = blk_check_plugged(raid10_unplug, mddev,
1531 plug = container_of(cb, struct raid10_plug_cb,
1535 spin_lock_irqsave(&conf->device_lock, flags);
1537 bio_list_add(&plug->pending, mbio);
1538 plug->pending_cnt++;
1540 bio_list_add(&conf->pending_bio_list, mbio);
1541 conf->pending_count++;
1543 spin_unlock_irqrestore(&conf->device_lock, flags);
1545 md_wakeup_thread(mddev->thread);
1548 if (r10_bio->devs[i].repl_bio) {
1549 struct md_rdev *rdev = conf->mirrors[d].replacement;
1551 /* Replacement just got moved to main 'rdev' */
1553 rdev = conf->mirrors[d].rdev;
1555 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1556 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1558 r10_bio->devs[i].repl_bio = mbio;
1560 mbio->bi_sector = (r10_bio->devs[i].addr +
1563 mbio->bi_bdev = rdev->bdev;
1564 mbio->bi_end_io = raid10_end_write_request;
1566 WRITE | do_sync | do_fua | do_discard | do_same;
1567 mbio->bi_private = r10_bio;
1569 atomic_inc(&r10_bio->remaining);
1570 spin_lock_irqsave(&conf->device_lock, flags);
1571 bio_list_add(&conf->pending_bio_list, mbio);
1572 conf->pending_count++;
1573 spin_unlock_irqrestore(&conf->device_lock, flags);
1574 if (!mddev_check_plugged(mddev))
1575 md_wakeup_thread(mddev->thread);
1579 /* Don't remove the bias on 'remaining' (one_write_done) until
1580 * after checking if we need to go around again.
1583 if (sectors_handled < bio_sectors(bio)) {
1584 one_write_done(r10_bio);
1585 /* We need another r10_bio. It has already been counted
1586 * in bio->bi_phys_segments.
1588 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1590 r10_bio->master_bio = bio;
1591 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1593 r10_bio->mddev = mddev;
1594 r10_bio->sector = bio->bi_sector + sectors_handled;
1598 one_write_done(r10_bio);
1600 /* In case raid10d snuck in to freeze_array */
1601 wake_up(&conf->wait_barrier);
1604 static void status(struct seq_file *seq, struct mddev *mddev)
1606 struct r10conf *conf = mddev->private;
1609 if (conf->geo.near_copies < conf->geo.raid_disks)
1610 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1611 if (conf->geo.near_copies > 1)
1612 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1613 if (conf->geo.far_copies > 1) {
1614 if (conf->geo.far_offset)
1615 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1617 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1619 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1620 conf->geo.raid_disks - mddev->degraded);
1621 for (i = 0; i < conf->geo.raid_disks; i++)
1622 seq_printf(seq, "%s",
1623 conf->mirrors[i].rdev &&
1624 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1625 seq_printf(seq, "]");
1628 /* check if there are enough drives for
1629 * every block to appear on atleast one.
1630 * Don't consider the device numbered 'ignore'
1631 * as we might be about to remove it.
1633 static int _enough(struct r10conf *conf, int previous, int ignore)
1639 disks = conf->prev.raid_disks;
1640 ncopies = conf->prev.near_copies;
1642 disks = conf->geo.raid_disks;
1643 ncopies = conf->geo.near_copies;
1648 int n = conf->copies;
1652 struct md_rdev *rdev;
1653 if (this != ignore &&
1654 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1655 test_bit(In_sync, &rdev->flags))
1657 this = (this+1) % disks;
1661 first = (first + ncopies) % disks;
1662 } while (first != 0);
1669 static int enough(struct r10conf *conf, int ignore)
1671 /* when calling 'enough', both 'prev' and 'geo' must
1673 * This is ensured if ->reconfig_mutex or ->device_lock
1676 return _enough(conf, 0, ignore) &&
1677 _enough(conf, 1, ignore);
1680 static void error(struct mddev *mddev, struct md_rdev *rdev)
1682 char b[BDEVNAME_SIZE];
1683 struct r10conf *conf = mddev->private;
1684 unsigned long flags;
1687 * If it is not operational, then we have already marked it as dead
1688 * else if it is the last working disks, ignore the error, let the
1689 * next level up know.
1690 * else mark the drive as failed
1692 spin_lock_irqsave(&conf->device_lock, flags);
1693 if (test_bit(In_sync, &rdev->flags)
1694 && !enough(conf, rdev->raid_disk)) {
1696 * Don't fail the drive, just return an IO error.
1698 spin_unlock_irqrestore(&conf->device_lock, flags);
1701 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1704 * if recovery is running, make sure it aborts.
1706 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1708 set_bit(Blocked, &rdev->flags);
1709 set_bit(Faulty, &rdev->flags);
1710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1711 spin_unlock_irqrestore(&conf->device_lock, flags);
1713 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1714 "md/raid10:%s: Operation continuing on %d devices.\n",
1715 mdname(mddev), bdevname(rdev->bdev, b),
1716 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1719 static void print_conf(struct r10conf *conf)
1722 struct raid10_info *tmp;
1724 printk(KERN_DEBUG "RAID10 conf printout:\n");
1726 printk(KERN_DEBUG "(!conf)\n");
1729 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1730 conf->geo.raid_disks);
1732 for (i = 0; i < conf->geo.raid_disks; i++) {
1733 char b[BDEVNAME_SIZE];
1734 tmp = conf->mirrors + i;
1736 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1737 i, !test_bit(In_sync, &tmp->rdev->flags),
1738 !test_bit(Faulty, &tmp->rdev->flags),
1739 bdevname(tmp->rdev->bdev,b));
1743 static void close_sync(struct r10conf *conf)
1746 allow_barrier(conf);
1748 mempool_destroy(conf->r10buf_pool);
1749 conf->r10buf_pool = NULL;
1752 static int raid10_spare_active(struct mddev *mddev)
1755 struct r10conf *conf = mddev->private;
1756 struct raid10_info *tmp;
1758 unsigned long flags;
1761 * Find all non-in_sync disks within the RAID10 configuration
1762 * and mark them in_sync
1764 for (i = 0; i < conf->geo.raid_disks; i++) {
1765 tmp = conf->mirrors + i;
1766 if (tmp->replacement
1767 && tmp->replacement->recovery_offset == MaxSector
1768 && !test_bit(Faulty, &tmp->replacement->flags)
1769 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1770 /* Replacement has just become active */
1772 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1775 /* Replaced device not technically faulty,
1776 * but we need to be sure it gets removed
1777 * and never re-added.
1779 set_bit(Faulty, &tmp->rdev->flags);
1780 sysfs_notify_dirent_safe(
1781 tmp->rdev->sysfs_state);
1783 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1784 } else if (tmp->rdev
1785 && !test_bit(Faulty, &tmp->rdev->flags)
1786 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1788 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1791 spin_lock_irqsave(&conf->device_lock, flags);
1792 mddev->degraded -= count;
1793 spin_unlock_irqrestore(&conf->device_lock, flags);
1800 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1802 struct r10conf *conf = mddev->private;
1806 int last = conf->geo.raid_disks - 1;
1807 struct request_queue *q = bdev_get_queue(rdev->bdev);
1809 if (mddev->recovery_cp < MaxSector)
1810 /* only hot-add to in-sync arrays, as recovery is
1811 * very different from resync
1814 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1817 if (rdev->raid_disk >= 0)
1818 first = last = rdev->raid_disk;
1820 if (q->merge_bvec_fn) {
1821 set_bit(Unmerged, &rdev->flags);
1822 mddev->merge_check_needed = 1;
1825 if (rdev->saved_raid_disk >= first &&
1826 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1827 mirror = rdev->saved_raid_disk;
1830 for ( ; mirror <= last ; mirror++) {
1831 struct raid10_info *p = &conf->mirrors[mirror];
1832 if (p->recovery_disabled == mddev->recovery_disabled)
1835 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1836 p->replacement != NULL)
1838 clear_bit(In_sync, &rdev->flags);
1839 set_bit(Replacement, &rdev->flags);
1840 rdev->raid_disk = mirror;
1843 disk_stack_limits(mddev->gendisk, rdev->bdev,
1844 rdev->data_offset << 9);
1846 rcu_assign_pointer(p->replacement, rdev);
1851 disk_stack_limits(mddev->gendisk, rdev->bdev,
1852 rdev->data_offset << 9);
1854 p->head_position = 0;
1855 p->recovery_disabled = mddev->recovery_disabled - 1;
1856 rdev->raid_disk = mirror;
1858 if (rdev->saved_raid_disk != mirror)
1860 rcu_assign_pointer(p->rdev, rdev);
1863 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1864 /* Some requests might not have seen this new
1865 * merge_bvec_fn. We must wait for them to complete
1866 * before merging the device fully.
1867 * First we make sure any code which has tested
1868 * our function has submitted the request, then
1869 * we wait for all outstanding requests to complete.
1871 synchronize_sched();
1872 freeze_array(conf, 0);
1873 unfreeze_array(conf);
1874 clear_bit(Unmerged, &rdev->flags);
1876 md_integrity_add_rdev(rdev, mddev);
1877 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1878 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1884 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1886 struct r10conf *conf = mddev->private;
1888 int number = rdev->raid_disk;
1889 struct md_rdev **rdevp;
1890 struct raid10_info *p = conf->mirrors + number;
1893 if (rdev == p->rdev)
1895 else if (rdev == p->replacement)
1896 rdevp = &p->replacement;
1900 if (test_bit(In_sync, &rdev->flags) ||
1901 atomic_read(&rdev->nr_pending)) {
1905 /* Only remove faulty devices if recovery
1908 if (!test_bit(Faulty, &rdev->flags) &&
1909 mddev->recovery_disabled != p->recovery_disabled &&
1910 (!p->replacement || p->replacement == rdev) &&
1911 number < conf->geo.raid_disks &&
1918 if (atomic_read(&rdev->nr_pending)) {
1919 /* lost the race, try later */
1923 } else if (p->replacement) {
1924 /* We must have just cleared 'rdev' */
1925 p->rdev = p->replacement;
1926 clear_bit(Replacement, &p->replacement->flags);
1927 smp_mb(); /* Make sure other CPUs may see both as identical
1928 * but will never see neither -- if they are careful.
1930 p->replacement = NULL;
1931 clear_bit(WantReplacement, &rdev->flags);
1933 /* We might have just remove the Replacement as faulty
1934 * Clear the flag just in case
1936 clear_bit(WantReplacement, &rdev->flags);
1938 err = md_integrity_register(mddev);
1947 static void end_sync_read(struct bio *bio, int error)
1949 struct r10bio *r10_bio = bio->bi_private;
1950 struct r10conf *conf = r10_bio->mddev->private;
1953 if (bio == r10_bio->master_bio) {
1954 /* this is a reshape read */
1955 d = r10_bio->read_slot; /* really the read dev */
1957 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1959 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1960 set_bit(R10BIO_Uptodate, &r10_bio->state);
1962 /* The write handler will notice the lack of
1963 * R10BIO_Uptodate and record any errors etc
1965 atomic_add(r10_bio->sectors,
1966 &conf->mirrors[d].rdev->corrected_errors);
1968 /* for reconstruct, we always reschedule after a read.
1969 * for resync, only after all reads
1971 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1972 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1973 atomic_dec_and_test(&r10_bio->remaining)) {
1974 /* we have read all the blocks,
1975 * do the comparison in process context in raid10d
1977 reschedule_retry(r10_bio);
1981 static void end_sync_request(struct r10bio *r10_bio)
1983 struct mddev *mddev = r10_bio->mddev;
1985 while (atomic_dec_and_test(&r10_bio->remaining)) {
1986 if (r10_bio->master_bio == NULL) {
1987 /* the primary of several recovery bios */
1988 sector_t s = r10_bio->sectors;
1989 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1990 test_bit(R10BIO_WriteError, &r10_bio->state))
1991 reschedule_retry(r10_bio);
1994 md_done_sync(mddev, s, 1);
1997 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1998 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1999 test_bit(R10BIO_WriteError, &r10_bio->state))
2000 reschedule_retry(r10_bio);
2008 static void end_sync_write(struct bio *bio, int error)
2010 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2011 struct r10bio *r10_bio = bio->bi_private;
2012 struct mddev *mddev = r10_bio->mddev;
2013 struct r10conf *conf = mddev->private;
2019 struct md_rdev *rdev = NULL;
2021 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2023 rdev = conf->mirrors[d].replacement;
2025 rdev = conf->mirrors[d].rdev;
2029 md_error(mddev, rdev);
2031 set_bit(WriteErrorSeen, &rdev->flags);
2032 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2033 set_bit(MD_RECOVERY_NEEDED,
2034 &rdev->mddev->recovery);
2035 set_bit(R10BIO_WriteError, &r10_bio->state);
2037 } else if (is_badblock(rdev,
2038 r10_bio->devs[slot].addr,
2040 &first_bad, &bad_sectors))
2041 set_bit(R10BIO_MadeGood, &r10_bio->state);
2043 rdev_dec_pending(rdev, mddev);
2045 end_sync_request(r10_bio);
2049 * Note: sync and recover and handled very differently for raid10
2050 * This code is for resync.
2051 * For resync, we read through virtual addresses and read all blocks.
2052 * If there is any error, we schedule a write. The lowest numbered
2053 * drive is authoritative.
2054 * However requests come for physical address, so we need to map.
2055 * For every physical address there are raid_disks/copies virtual addresses,
2056 * which is always are least one, but is not necessarly an integer.
2057 * This means that a physical address can span multiple chunks, so we may
2058 * have to submit multiple io requests for a single sync request.
2061 * We check if all blocks are in-sync and only write to blocks that
2064 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2066 struct r10conf *conf = mddev->private;
2068 struct bio *tbio, *fbio;
2071 atomic_set(&r10_bio->remaining, 1);
2073 /* find the first device with a block */
2074 for (i=0; i<conf->copies; i++)
2075 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2078 if (i == conf->copies)
2082 fbio = r10_bio->devs[i].bio;
2084 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2085 /* now find blocks with errors */
2086 for (i=0 ; i < conf->copies ; i++) {
2089 tbio = r10_bio->devs[i].bio;
2091 if (tbio->bi_end_io != end_sync_read)
2095 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2096 /* We know that the bi_io_vec layout is the same for
2097 * both 'first' and 'i', so we just compare them.
2098 * All vec entries are PAGE_SIZE;
2100 int sectors = r10_bio->sectors;
2101 for (j = 0; j < vcnt; j++) {
2102 int len = PAGE_SIZE;
2103 if (sectors < (len / 512))
2104 len = sectors * 512;
2105 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2106 page_address(tbio->bi_io_vec[j].bv_page),
2113 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2114 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2115 /* Don't fix anything. */
2118 /* Ok, we need to write this bio, either to correct an
2119 * inconsistency or to correct an unreadable block.
2120 * First we need to fixup bv_offset, bv_len and
2121 * bi_vecs, as the read request might have corrupted these
2125 tbio->bi_vcnt = vcnt;
2126 tbio->bi_size = r10_bio->sectors << 9;
2127 tbio->bi_rw = WRITE;
2128 tbio->bi_private = r10_bio;
2129 tbio->bi_sector = r10_bio->devs[i].addr;
2131 for (j=0; j < vcnt ; j++) {
2132 tbio->bi_io_vec[j].bv_offset = 0;
2133 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2135 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2136 page_address(fbio->bi_io_vec[j].bv_page),
2139 tbio->bi_end_io = end_sync_write;
2141 d = r10_bio->devs[i].devnum;
2142 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2143 atomic_inc(&r10_bio->remaining);
2144 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2146 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2147 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2148 generic_make_request(tbio);
2151 /* Now write out to any replacement devices
2154 for (i = 0; i < conf->copies; i++) {
2157 tbio = r10_bio->devs[i].repl_bio;
2158 if (!tbio || !tbio->bi_end_io)
2160 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2161 && r10_bio->devs[i].bio != fbio)
2162 for (j = 0; j < vcnt; j++)
2163 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2164 page_address(fbio->bi_io_vec[j].bv_page),
2166 d = r10_bio->devs[i].devnum;
2167 atomic_inc(&r10_bio->remaining);
2168 md_sync_acct(conf->mirrors[d].replacement->bdev,
2170 generic_make_request(tbio);
2174 if (atomic_dec_and_test(&r10_bio->remaining)) {
2175 md_done_sync(mddev, r10_bio->sectors, 1);
2181 * Now for the recovery code.
2182 * Recovery happens across physical sectors.
2183 * We recover all non-is_sync drives by finding the virtual address of
2184 * each, and then choose a working drive that also has that virt address.
2185 * There is a separate r10_bio for each non-in_sync drive.
2186 * Only the first two slots are in use. The first for reading,
2187 * The second for writing.
2190 static void fix_recovery_read_error(struct r10bio *r10_bio)
2192 /* We got a read error during recovery.
2193 * We repeat the read in smaller page-sized sections.
2194 * If a read succeeds, write it to the new device or record
2195 * a bad block if we cannot.
2196 * If a read fails, record a bad block on both old and
2199 struct mddev *mddev = r10_bio->mddev;
2200 struct r10conf *conf = mddev->private;
2201 struct bio *bio = r10_bio->devs[0].bio;
2203 int sectors = r10_bio->sectors;
2205 int dr = r10_bio->devs[0].devnum;
2206 int dw = r10_bio->devs[1].devnum;
2210 struct md_rdev *rdev;
2214 if (s > (PAGE_SIZE>>9))
2217 rdev = conf->mirrors[dr].rdev;
2218 addr = r10_bio->devs[0].addr + sect,
2219 ok = sync_page_io(rdev,
2222 bio->bi_io_vec[idx].bv_page,
2225 rdev = conf->mirrors[dw].rdev;
2226 addr = r10_bio->devs[1].addr + sect;
2227 ok = sync_page_io(rdev,
2230 bio->bi_io_vec[idx].bv_page,
2233 set_bit(WriteErrorSeen, &rdev->flags);
2234 if (!test_and_set_bit(WantReplacement,
2236 set_bit(MD_RECOVERY_NEEDED,
2237 &rdev->mddev->recovery);
2241 /* We don't worry if we cannot set a bad block -
2242 * it really is bad so there is no loss in not
2245 rdev_set_badblocks(rdev, addr, s, 0);
2247 if (rdev != conf->mirrors[dw].rdev) {
2248 /* need bad block on destination too */
2249 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2250 addr = r10_bio->devs[1].addr + sect;
2251 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2253 /* just abort the recovery */
2255 "md/raid10:%s: recovery aborted"
2256 " due to read error\n",
2259 conf->mirrors[dw].recovery_disabled
2260 = mddev->recovery_disabled;
2261 set_bit(MD_RECOVERY_INTR,
2274 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2276 struct r10conf *conf = mddev->private;
2278 struct bio *wbio, *wbio2;
2280 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2281 fix_recovery_read_error(r10_bio);
2282 end_sync_request(r10_bio);
2287 * share the pages with the first bio
2288 * and submit the write request
2290 d = r10_bio->devs[1].devnum;
2291 wbio = r10_bio->devs[1].bio;
2292 wbio2 = r10_bio->devs[1].repl_bio;
2293 if (wbio->bi_end_io) {
2294 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2295 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2296 generic_make_request(wbio);
2298 if (wbio2 && wbio2->bi_end_io) {
2299 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2300 md_sync_acct(conf->mirrors[d].replacement->bdev,
2301 bio_sectors(wbio2));
2302 generic_make_request(wbio2);
2308 * Used by fix_read_error() to decay the per rdev read_errors.
2309 * We halve the read error count for every hour that has elapsed
2310 * since the last recorded read error.
2313 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2315 struct timespec cur_time_mon;
2316 unsigned long hours_since_last;
2317 unsigned int read_errors = atomic_read(&rdev->read_errors);
2319 ktime_get_ts(&cur_time_mon);
2321 if (rdev->last_read_error.tv_sec == 0 &&
2322 rdev->last_read_error.tv_nsec == 0) {
2323 /* first time we've seen a read error */
2324 rdev->last_read_error = cur_time_mon;
2328 hours_since_last = (cur_time_mon.tv_sec -
2329 rdev->last_read_error.tv_sec) / 3600;
2331 rdev->last_read_error = cur_time_mon;
2334 * if hours_since_last is > the number of bits in read_errors
2335 * just set read errors to 0. We do this to avoid
2336 * overflowing the shift of read_errors by hours_since_last.
2338 if (hours_since_last >= 8 * sizeof(read_errors))
2339 atomic_set(&rdev->read_errors, 0);
2341 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2344 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2345 int sectors, struct page *page, int rw)
2350 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2351 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2353 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2357 set_bit(WriteErrorSeen, &rdev->flags);
2358 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2359 set_bit(MD_RECOVERY_NEEDED,
2360 &rdev->mddev->recovery);
2362 /* need to record an error - either for the block or the device */
2363 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2364 md_error(rdev->mddev, rdev);
2369 * This is a kernel thread which:
2371 * 1. Retries failed read operations on working mirrors.
2372 * 2. Updates the raid superblock when problems encounter.
2373 * 3. Performs writes following reads for array synchronising.
2376 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2378 int sect = 0; /* Offset from r10_bio->sector */
2379 int sectors = r10_bio->sectors;
2380 struct md_rdev*rdev;
2381 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2382 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2384 /* still own a reference to this rdev, so it cannot
2385 * have been cleared recently.
2387 rdev = conf->mirrors[d].rdev;
2389 if (test_bit(Faulty, &rdev->flags))
2390 /* drive has already been failed, just ignore any
2391 more fix_read_error() attempts */
2394 check_decay_read_errors(mddev, rdev);
2395 atomic_inc(&rdev->read_errors);
2396 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2397 char b[BDEVNAME_SIZE];
2398 bdevname(rdev->bdev, b);
2401 "md/raid10:%s: %s: Raid device exceeded "
2402 "read_error threshold [cur %d:max %d]\n",
2404 atomic_read(&rdev->read_errors), max_read_errors);
2406 "md/raid10:%s: %s: Failing raid device\n",
2408 md_error(mddev, conf->mirrors[d].rdev);
2409 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2415 int sl = r10_bio->read_slot;
2419 if (s > (PAGE_SIZE>>9))
2427 d = r10_bio->devs[sl].devnum;
2428 rdev = rcu_dereference(conf->mirrors[d].rdev);
2430 !test_bit(Unmerged, &rdev->flags) &&
2431 test_bit(In_sync, &rdev->flags) &&
2432 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2433 &first_bad, &bad_sectors) == 0) {
2434 atomic_inc(&rdev->nr_pending);
2436 success = sync_page_io(rdev,
2437 r10_bio->devs[sl].addr +
2440 conf->tmppage, READ, false);
2441 rdev_dec_pending(rdev, mddev);
2447 if (sl == conf->copies)
2449 } while (!success && sl != r10_bio->read_slot);
2453 /* Cannot read from anywhere, just mark the block
2454 * as bad on the first device to discourage future
2457 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2458 rdev = conf->mirrors[dn].rdev;
2460 if (!rdev_set_badblocks(
2462 r10_bio->devs[r10_bio->read_slot].addr
2465 md_error(mddev, rdev);
2466 r10_bio->devs[r10_bio->read_slot].bio
2473 /* write it back and re-read */
2475 while (sl != r10_bio->read_slot) {
2476 char b[BDEVNAME_SIZE];
2481 d = r10_bio->devs[sl].devnum;
2482 rdev = rcu_dereference(conf->mirrors[d].rdev);
2484 test_bit(Unmerged, &rdev->flags) ||
2485 !test_bit(In_sync, &rdev->flags))
2488 atomic_inc(&rdev->nr_pending);
2490 if (r10_sync_page_io(rdev,
2491 r10_bio->devs[sl].addr +
2493 s, conf->tmppage, WRITE)
2495 /* Well, this device is dead */
2497 "md/raid10:%s: read correction "
2499 " (%d sectors at %llu on %s)\n",
2501 (unsigned long long)(
2503 choose_data_offset(r10_bio,
2505 bdevname(rdev->bdev, b));
2506 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2509 bdevname(rdev->bdev, b));
2511 rdev_dec_pending(rdev, mddev);
2515 while (sl != r10_bio->read_slot) {
2516 char b[BDEVNAME_SIZE];
2521 d = r10_bio->devs[sl].devnum;
2522 rdev = rcu_dereference(conf->mirrors[d].rdev);
2524 !test_bit(In_sync, &rdev->flags))
2527 atomic_inc(&rdev->nr_pending);
2529 switch (r10_sync_page_io(rdev,
2530 r10_bio->devs[sl].addr +
2535 /* Well, this device is dead */
2537 "md/raid10:%s: unable to read back "
2539 " (%d sectors at %llu on %s)\n",
2541 (unsigned long long)(
2543 choose_data_offset(r10_bio, rdev)),
2544 bdevname(rdev->bdev, b));
2545 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2548 bdevname(rdev->bdev, b));
2552 "md/raid10:%s: read error corrected"
2553 " (%d sectors at %llu on %s)\n",
2555 (unsigned long long)(
2557 choose_data_offset(r10_bio, rdev)),
2558 bdevname(rdev->bdev, b));
2559 atomic_add(s, &rdev->corrected_errors);
2562 rdev_dec_pending(rdev, mddev);
2572 static int narrow_write_error(struct r10bio *r10_bio, int i)
2574 struct bio *bio = r10_bio->master_bio;
2575 struct mddev *mddev = r10_bio->mddev;
2576 struct r10conf *conf = mddev->private;
2577 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2578 /* bio has the data to be written to slot 'i' where
2579 * we just recently had a write error.
2580 * We repeatedly clone the bio and trim down to one block,
2581 * then try the write. Where the write fails we record
2583 * It is conceivable that the bio doesn't exactly align with
2584 * blocks. We must handle this.
2586 * We currently own a reference to the rdev.
2592 int sect_to_write = r10_bio->sectors;
2595 if (rdev->badblocks.shift < 0)
2598 block_sectors = 1 << rdev->badblocks.shift;
2599 sector = r10_bio->sector;
2600 sectors = ((r10_bio->sector + block_sectors)
2601 & ~(sector_t)(block_sectors - 1))
2604 while (sect_to_write) {
2606 if (sectors > sect_to_write)
2607 sectors = sect_to_write;
2608 /* Write at 'sector' for 'sectors' */
2609 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2610 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2611 wbio->bi_sector = (r10_bio->devs[i].addr+
2612 choose_data_offset(r10_bio, rdev) +
2613 (sector - r10_bio->sector));
2614 wbio->bi_bdev = rdev->bdev;
2615 if (submit_bio_wait(WRITE, wbio) == 0)
2617 ok = rdev_set_badblocks(rdev, sector,
2622 sect_to_write -= sectors;
2624 sectors = block_sectors;
2629 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2631 int slot = r10_bio->read_slot;
2633 struct r10conf *conf = mddev->private;
2634 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2635 char b[BDEVNAME_SIZE];
2636 unsigned long do_sync;
2639 /* we got a read error. Maybe the drive is bad. Maybe just
2640 * the block and we can fix it.
2641 * We freeze all other IO, and try reading the block from
2642 * other devices. When we find one, we re-write
2643 * and check it that fixes the read error.
2644 * This is all done synchronously while the array is
2647 bio = r10_bio->devs[slot].bio;
2648 bdevname(bio->bi_bdev, b);
2650 r10_bio->devs[slot].bio = NULL;
2652 if (mddev->ro == 0) {
2653 freeze_array(conf, 1);
2654 fix_read_error(conf, mddev, r10_bio);
2655 unfreeze_array(conf);
2657 r10_bio->devs[slot].bio = IO_BLOCKED;
2659 rdev_dec_pending(rdev, mddev);
2662 rdev = read_balance(conf, r10_bio, &max_sectors);
2664 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2665 " read error for block %llu\n",
2667 (unsigned long long)r10_bio->sector);
2668 raid_end_bio_io(r10_bio);
2672 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2673 slot = r10_bio->read_slot;
2676 "md/raid10:%s: %s: redirecting "
2677 "sector %llu to another mirror\n",
2679 bdevname(rdev->bdev, b),
2680 (unsigned long long)r10_bio->sector);
2681 bio = bio_clone_mddev(r10_bio->master_bio,
2684 r10_bio->sector - bio->bi_sector,
2686 r10_bio->devs[slot].bio = bio;
2687 r10_bio->devs[slot].rdev = rdev;
2688 bio->bi_sector = r10_bio->devs[slot].addr
2689 + choose_data_offset(r10_bio, rdev);
2690 bio->bi_bdev = rdev->bdev;
2691 bio->bi_rw = READ | do_sync;
2692 bio->bi_private = r10_bio;
2693 bio->bi_end_io = raid10_end_read_request;
2694 if (max_sectors < r10_bio->sectors) {
2695 /* Drat - have to split this up more */
2696 struct bio *mbio = r10_bio->master_bio;
2697 int sectors_handled =
2698 r10_bio->sector + max_sectors
2700 r10_bio->sectors = max_sectors;
2701 spin_lock_irq(&conf->device_lock);
2702 if (mbio->bi_phys_segments == 0)
2703 mbio->bi_phys_segments = 2;
2705 mbio->bi_phys_segments++;
2706 spin_unlock_irq(&conf->device_lock);
2707 generic_make_request(bio);
2709 r10_bio = mempool_alloc(conf->r10bio_pool,
2711 r10_bio->master_bio = mbio;
2712 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2714 set_bit(R10BIO_ReadError,
2716 r10_bio->mddev = mddev;
2717 r10_bio->sector = mbio->bi_sector
2722 generic_make_request(bio);
2725 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2727 /* Some sort of write request has finished and it
2728 * succeeded in writing where we thought there was a
2729 * bad block. So forget the bad block.
2730 * Or possibly if failed and we need to record
2734 struct md_rdev *rdev;
2736 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2737 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2738 for (m = 0; m < conf->copies; m++) {
2739 int dev = r10_bio->devs[m].devnum;
2740 rdev = conf->mirrors[dev].rdev;
2741 if (r10_bio->devs[m].bio == NULL)
2743 if (test_bit(BIO_UPTODATE,
2744 &r10_bio->devs[m].bio->bi_flags)) {
2745 rdev_clear_badblocks(
2747 r10_bio->devs[m].addr,
2748 r10_bio->sectors, 0);
2750 if (!rdev_set_badblocks(
2752 r10_bio->devs[m].addr,
2753 r10_bio->sectors, 0))
2754 md_error(conf->mddev, rdev);
2756 rdev = conf->mirrors[dev].replacement;
2757 if (r10_bio->devs[m].repl_bio == NULL)
2759 if (test_bit(BIO_UPTODATE,
2760 &r10_bio->devs[m].repl_bio->bi_flags)) {
2761 rdev_clear_badblocks(
2763 r10_bio->devs[m].addr,
2764 r10_bio->sectors, 0);
2766 if (!rdev_set_badblocks(
2768 r10_bio->devs[m].addr,
2769 r10_bio->sectors, 0))
2770 md_error(conf->mddev, rdev);
2775 for (m = 0; m < conf->copies; m++) {
2776 int dev = r10_bio->devs[m].devnum;
2777 struct bio *bio = r10_bio->devs[m].bio;
2778 rdev = conf->mirrors[dev].rdev;
2779 if (bio == IO_MADE_GOOD) {
2780 rdev_clear_badblocks(
2782 r10_bio->devs[m].addr,
2783 r10_bio->sectors, 0);
2784 rdev_dec_pending(rdev, conf->mddev);
2785 } else if (bio != NULL &&
2786 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2787 if (!narrow_write_error(r10_bio, m)) {
2788 md_error(conf->mddev, rdev);
2789 set_bit(R10BIO_Degraded,
2792 rdev_dec_pending(rdev, conf->mddev);
2794 bio = r10_bio->devs[m].repl_bio;
2795 rdev = conf->mirrors[dev].replacement;
2796 if (rdev && bio == IO_MADE_GOOD) {
2797 rdev_clear_badblocks(
2799 r10_bio->devs[m].addr,
2800 r10_bio->sectors, 0);
2801 rdev_dec_pending(rdev, conf->mddev);
2804 if (test_bit(R10BIO_WriteError,
2806 close_write(r10_bio);
2807 raid_end_bio_io(r10_bio);
2811 static void raid10d(struct md_thread *thread)
2813 struct mddev *mddev = thread->mddev;
2814 struct r10bio *r10_bio;
2815 unsigned long flags;
2816 struct r10conf *conf = mddev->private;
2817 struct list_head *head = &conf->retry_list;
2818 struct blk_plug plug;
2820 md_check_recovery(mddev);
2822 blk_start_plug(&plug);
2825 flush_pending_writes(conf);
2827 spin_lock_irqsave(&conf->device_lock, flags);
2828 if (list_empty(head)) {
2829 spin_unlock_irqrestore(&conf->device_lock, flags);
2832 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2833 list_del(head->prev);
2835 spin_unlock_irqrestore(&conf->device_lock, flags);
2837 mddev = r10_bio->mddev;
2838 conf = mddev->private;
2839 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2840 test_bit(R10BIO_WriteError, &r10_bio->state))
2841 handle_write_completed(conf, r10_bio);
2842 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2843 reshape_request_write(mddev, r10_bio);
2844 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2845 sync_request_write(mddev, r10_bio);
2846 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2847 recovery_request_write(mddev, r10_bio);
2848 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2849 handle_read_error(mddev, r10_bio);
2851 /* just a partial read to be scheduled from a
2854 int slot = r10_bio->read_slot;
2855 generic_make_request(r10_bio->devs[slot].bio);
2859 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2860 md_check_recovery(mddev);
2862 blk_finish_plug(&plug);
2866 static int init_resync(struct r10conf *conf)
2871 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2872 BUG_ON(conf->r10buf_pool);
2873 conf->have_replacement = 0;
2874 for (i = 0; i < conf->geo.raid_disks; i++)
2875 if (conf->mirrors[i].replacement)
2876 conf->have_replacement = 1;
2877 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2878 if (!conf->r10buf_pool)
2880 conf->next_resync = 0;
2885 * perform a "sync" on one "block"
2887 * We need to make sure that no normal I/O request - particularly write
2888 * requests - conflict with active sync requests.
2890 * This is achieved by tracking pending requests and a 'barrier' concept
2891 * that can be installed to exclude normal IO requests.
2893 * Resync and recovery are handled very differently.
2894 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2896 * For resync, we iterate over virtual addresses, read all copies,
2897 * and update if there are differences. If only one copy is live,
2899 * For recovery, we iterate over physical addresses, read a good
2900 * value for each non-in_sync drive, and over-write.
2902 * So, for recovery we may have several outstanding complex requests for a
2903 * given address, one for each out-of-sync device. We model this by allocating
2904 * a number of r10_bio structures, one for each out-of-sync device.
2905 * As we setup these structures, we collect all bio's together into a list
2906 * which we then process collectively to add pages, and then process again
2907 * to pass to generic_make_request.
2909 * The r10_bio structures are linked using a borrowed master_bio pointer.
2910 * This link is counted in ->remaining. When the r10_bio that points to NULL
2911 * has its remaining count decremented to 0, the whole complex operation
2916 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2917 int *skipped, int go_faster)
2919 struct r10conf *conf = mddev->private;
2920 struct r10bio *r10_bio;
2921 struct bio *biolist = NULL, *bio;
2922 sector_t max_sector, nr_sectors;
2925 sector_t sync_blocks;
2926 sector_t sectors_skipped = 0;
2927 int chunks_skipped = 0;
2928 sector_t chunk_mask = conf->geo.chunk_mask;
2930 if (!conf->r10buf_pool)
2931 if (init_resync(conf))
2935 * Allow skipping a full rebuild for incremental assembly
2936 * of a clean array, like RAID1 does.
2938 if (mddev->bitmap == NULL &&
2939 mddev->recovery_cp == MaxSector &&
2940 mddev->reshape_position == MaxSector &&
2941 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2942 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2943 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2944 conf->fullsync == 0) {
2946 return mddev->dev_sectors - sector_nr;
2950 max_sector = mddev->dev_sectors;
2951 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2952 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2953 max_sector = mddev->resync_max_sectors;
2954 if (sector_nr >= max_sector) {
2955 /* If we aborted, we need to abort the
2956 * sync on the 'current' bitmap chucks (there can
2957 * be several when recovering multiple devices).
2958 * as we may have started syncing it but not finished.
2959 * We can find the current address in
2960 * mddev->curr_resync, but for recovery,
2961 * we need to convert that to several
2962 * virtual addresses.
2964 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2969 if (mddev->curr_resync < max_sector) { /* aborted */
2970 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2971 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2973 else for (i = 0; i < conf->geo.raid_disks; i++) {
2975 raid10_find_virt(conf, mddev->curr_resync, i);
2976 bitmap_end_sync(mddev->bitmap, sect,
2980 /* completed sync */
2981 if ((!mddev->bitmap || conf->fullsync)
2982 && conf->have_replacement
2983 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2984 /* Completed a full sync so the replacements
2985 * are now fully recovered.
2987 for (i = 0; i < conf->geo.raid_disks; i++)
2988 if (conf->mirrors[i].replacement)
2989 conf->mirrors[i].replacement
2995 bitmap_close_sync(mddev->bitmap);
2998 return sectors_skipped;
3001 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3002 return reshape_request(mddev, sector_nr, skipped);
3004 if (chunks_skipped >= conf->geo.raid_disks) {
3005 /* if there has been nothing to do on any drive,
3006 * then there is nothing to do at all..
3009 return (max_sector - sector_nr) + sectors_skipped;
3012 if (max_sector > mddev->resync_max)
3013 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3015 /* make sure whole request will fit in a chunk - if chunks
3018 if (conf->geo.near_copies < conf->geo.raid_disks &&
3019 max_sector > (sector_nr | chunk_mask))
3020 max_sector = (sector_nr | chunk_mask) + 1;
3022 * If there is non-resync activity waiting for us then
3023 * put in a delay to throttle resync.
3025 if (!go_faster && conf->nr_waiting)
3026 msleep_interruptible(1000);
3028 /* Again, very different code for resync and recovery.
3029 * Both must result in an r10bio with a list of bios that
3030 * have bi_end_io, bi_sector, bi_bdev set,
3031 * and bi_private set to the r10bio.
3032 * For recovery, we may actually create several r10bios
3033 * with 2 bios in each, that correspond to the bios in the main one.
3034 * In this case, the subordinate r10bios link back through a
3035 * borrowed master_bio pointer, and the counter in the master
3036 * includes a ref from each subordinate.
3038 /* First, we decide what to do and set ->bi_end_io
3039 * To end_sync_read if we want to read, and
3040 * end_sync_write if we will want to write.
3043 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3044 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3045 /* recovery... the complicated one */
3049 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3055 struct raid10_info *mirror = &conf->mirrors[i];
3057 if ((mirror->rdev == NULL ||
3058 test_bit(In_sync, &mirror->rdev->flags))
3060 (mirror->replacement == NULL ||
3062 &mirror->replacement->flags)))
3066 /* want to reconstruct this device */
3068 sect = raid10_find_virt(conf, sector_nr, i);
3069 if (sect >= mddev->resync_max_sectors) {
3070 /* last stripe is not complete - don't
3071 * try to recover this sector.
3075 /* Unless we are doing a full sync, or a replacement
3076 * we only need to recover the block if it is set in
3079 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3081 if (sync_blocks < max_sync)
3082 max_sync = sync_blocks;
3084 mirror->replacement == NULL &&
3086 /* yep, skip the sync_blocks here, but don't assume
3087 * that there will never be anything to do here
3089 chunks_skipped = -1;
3093 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3094 raise_barrier(conf, rb2 != NULL);
3095 atomic_set(&r10_bio->remaining, 0);
3097 r10_bio->master_bio = (struct bio*)rb2;
3099 atomic_inc(&rb2->remaining);
3100 r10_bio->mddev = mddev;
3101 set_bit(R10BIO_IsRecover, &r10_bio->state);
3102 r10_bio->sector = sect;
3104 raid10_find_phys(conf, r10_bio);
3106 /* Need to check if the array will still be
3109 for (j = 0; j < conf->geo.raid_disks; j++)
3110 if (conf->mirrors[j].rdev == NULL ||
3111 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3116 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3117 &sync_blocks, still_degraded);
3120 for (j=0; j<conf->copies;j++) {
3122 int d = r10_bio->devs[j].devnum;
3123 sector_t from_addr, to_addr;
3124 struct md_rdev *rdev;
3125 sector_t sector, first_bad;
3127 if (!conf->mirrors[d].rdev ||
3128 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3130 /* This is where we read from */
3132 rdev = conf->mirrors[d].rdev;
3133 sector = r10_bio->devs[j].addr;
3135 if (is_badblock(rdev, sector, max_sync,
3136 &first_bad, &bad_sectors)) {
3137 if (first_bad > sector)
3138 max_sync = first_bad - sector;
3140 bad_sectors -= (sector
3142 if (max_sync > bad_sectors)
3143 max_sync = bad_sectors;
3147 bio = r10_bio->devs[0].bio;
3149 bio->bi_next = biolist;
3151 bio->bi_private = r10_bio;
3152 bio->bi_end_io = end_sync_read;
3154 from_addr = r10_bio->devs[j].addr;
3155 bio->bi_sector = from_addr + rdev->data_offset;
3156 bio->bi_bdev = rdev->bdev;
3157 atomic_inc(&rdev->nr_pending);
3158 /* and we write to 'i' (if not in_sync) */
3160 for (k=0; k<conf->copies; k++)
3161 if (r10_bio->devs[k].devnum == i)
3163 BUG_ON(k == conf->copies);
3164 to_addr = r10_bio->devs[k].addr;
3165 r10_bio->devs[0].devnum = d;
3166 r10_bio->devs[0].addr = from_addr;
3167 r10_bio->devs[1].devnum = i;
3168 r10_bio->devs[1].addr = to_addr;
3170 rdev = mirror->rdev;
3171 if (!test_bit(In_sync, &rdev->flags)) {
3172 bio = r10_bio->devs[1].bio;
3174 bio->bi_next = biolist;
3176 bio->bi_private = r10_bio;
3177 bio->bi_end_io = end_sync_write;
3179 bio->bi_sector = to_addr
3180 + rdev->data_offset;
3181 bio->bi_bdev = rdev->bdev;
3182 atomic_inc(&r10_bio->remaining);
3184 r10_bio->devs[1].bio->bi_end_io = NULL;
3186 /* and maybe write to replacement */
3187 bio = r10_bio->devs[1].repl_bio;
3189 bio->bi_end_io = NULL;
3190 rdev = mirror->replacement;
3191 /* Note: if rdev != NULL, then bio
3192 * cannot be NULL as r10buf_pool_alloc will
3193 * have allocated it.
3194 * So the second test here is pointless.
3195 * But it keeps semantic-checkers happy, and
3196 * this comment keeps human reviewers
3199 if (rdev == NULL || bio == NULL ||
3200 test_bit(Faulty, &rdev->flags))
3203 bio->bi_next = biolist;
3205 bio->bi_private = r10_bio;
3206 bio->bi_end_io = end_sync_write;
3208 bio->bi_sector = to_addr + rdev->data_offset;
3209 bio->bi_bdev = rdev->bdev;
3210 atomic_inc(&r10_bio->remaining);
3213 if (j == conf->copies) {
3214 /* Cannot recover, so abort the recovery or
3215 * record a bad block */
3218 atomic_dec(&rb2->remaining);
3221 /* problem is that there are bad blocks
3222 * on other device(s)
3225 for (k = 0; k < conf->copies; k++)
3226 if (r10_bio->devs[k].devnum == i)
3228 if (!test_bit(In_sync,
3229 &mirror->rdev->flags)
3230 && !rdev_set_badblocks(
3232 r10_bio->devs[k].addr,
3235 if (mirror->replacement &&
3236 !rdev_set_badblocks(
3237 mirror->replacement,
3238 r10_bio->devs[k].addr,
3243 if (!test_and_set_bit(MD_RECOVERY_INTR,
3245 printk(KERN_INFO "md/raid10:%s: insufficient "
3246 "working devices for recovery.\n",
3248 mirror->recovery_disabled
3249 = mddev->recovery_disabled;
3254 if (biolist == NULL) {
3256 struct r10bio *rb2 = r10_bio;
3257 r10_bio = (struct r10bio*) rb2->master_bio;
3258 rb2->master_bio = NULL;
3264 /* resync. Schedule a read for every block at this virt offset */
3267 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3269 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3270 &sync_blocks, mddev->degraded) &&
3271 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3272 &mddev->recovery)) {
3273 /* We can skip this block */
3275 return sync_blocks + sectors_skipped;
3277 if (sync_blocks < max_sync)
3278 max_sync = sync_blocks;
3279 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3281 r10_bio->mddev = mddev;
3282 atomic_set(&r10_bio->remaining, 0);
3283 raise_barrier(conf, 0);
3284 conf->next_resync = sector_nr;
3286 r10_bio->master_bio = NULL;
3287 r10_bio->sector = sector_nr;
3288 set_bit(R10BIO_IsSync, &r10_bio->state);
3289 raid10_find_phys(conf, r10_bio);
3290 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3292 for (i = 0; i < conf->copies; i++) {
3293 int d = r10_bio->devs[i].devnum;
3294 sector_t first_bad, sector;
3297 if (r10_bio->devs[i].repl_bio)
3298 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3300 bio = r10_bio->devs[i].bio;
3302 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3303 if (conf->mirrors[d].rdev == NULL ||
3304 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3306 sector = r10_bio->devs[i].addr;
3307 if (is_badblock(conf->mirrors[d].rdev,
3309 &first_bad, &bad_sectors)) {
3310 if (first_bad > sector)
3311 max_sync = first_bad - sector;
3313 bad_sectors -= (sector - first_bad);
3314 if (max_sync > bad_sectors)
3315 max_sync = bad_sectors;
3319 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3320 atomic_inc(&r10_bio->remaining);
3321 bio->bi_next = biolist;
3323 bio->bi_private = r10_bio;
3324 bio->bi_end_io = end_sync_read;
3326 bio->bi_sector = sector +
3327 conf->mirrors[d].rdev->data_offset;
3328 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3331 if (conf->mirrors[d].replacement == NULL ||
3333 &conf->mirrors[d].replacement->flags))
3336 /* Need to set up for writing to the replacement */
3337 bio = r10_bio->devs[i].repl_bio;
3339 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3341 sector = r10_bio->devs[i].addr;
3342 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3343 bio->bi_next = biolist;
3345 bio->bi_private = r10_bio;
3346 bio->bi_end_io = end_sync_write;
3348 bio->bi_sector = sector +
3349 conf->mirrors[d].replacement->data_offset;
3350 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3355 for (i=0; i<conf->copies; i++) {
3356 int d = r10_bio->devs[i].devnum;
3357 if (r10_bio->devs[i].bio->bi_end_io)
3358 rdev_dec_pending(conf->mirrors[d].rdev,
3360 if (r10_bio->devs[i].repl_bio &&
3361 r10_bio->devs[i].repl_bio->bi_end_io)
3363 conf->mirrors[d].replacement,
3373 if (sector_nr + max_sync < max_sector)
3374 max_sector = sector_nr + max_sync;
3377 int len = PAGE_SIZE;
3378 if (sector_nr + (len>>9) > max_sector)
3379 len = (max_sector - sector_nr) << 9;
3382 for (bio= biolist ; bio ; bio=bio->bi_next) {
3384 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3385 if (bio_add_page(bio, page, len, 0))
3389 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3390 for (bio2 = biolist;
3391 bio2 && bio2 != bio;
3392 bio2 = bio2->bi_next) {
3393 /* remove last page from this bio */
3395 bio2->bi_size -= len;
3396 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3400 nr_sectors += len>>9;
3401 sector_nr += len>>9;
3402 } while (biolist->bi_vcnt < RESYNC_PAGES);
3404 r10_bio->sectors = nr_sectors;
3408 biolist = biolist->bi_next;
3410 bio->bi_next = NULL;
3411 r10_bio = bio->bi_private;
3412 r10_bio->sectors = nr_sectors;
3414 if (bio->bi_end_io == end_sync_read) {
3415 md_sync_acct(bio->bi_bdev, nr_sectors);
3416 set_bit(BIO_UPTODATE, &bio->bi_flags);
3417 generic_make_request(bio);
3421 if (sectors_skipped)
3422 /* pretend they weren't skipped, it makes
3423 * no important difference in this case
3425 md_done_sync(mddev, sectors_skipped, 1);
3427 return sectors_skipped + nr_sectors;
3429 /* There is nowhere to write, so all non-sync
3430 * drives must be failed or in resync, all drives
3431 * have a bad block, so try the next chunk...
3433 if (sector_nr + max_sync < max_sector)
3434 max_sector = sector_nr + max_sync;
3436 sectors_skipped += (max_sector - sector_nr);
3438 sector_nr = max_sector;
3443 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3446 struct r10conf *conf = mddev->private;
3449 raid_disks = min(conf->geo.raid_disks,
3450 conf->prev.raid_disks);
3452 sectors = conf->dev_sectors;
3454 size = sectors >> conf->geo.chunk_shift;
3455 sector_div(size, conf->geo.far_copies);
3456 size = size * raid_disks;
3457 sector_div(size, conf->geo.near_copies);
3459 return size << conf->geo.chunk_shift;
3462 static void calc_sectors(struct r10conf *conf, sector_t size)
3464 /* Calculate the number of sectors-per-device that will
3465 * actually be used, and set conf->dev_sectors and
3469 size = size >> conf->geo.chunk_shift;
3470 sector_div(size, conf->geo.far_copies);
3471 size = size * conf->geo.raid_disks;
3472 sector_div(size, conf->geo.near_copies);
3473 /* 'size' is now the number of chunks in the array */
3474 /* calculate "used chunks per device" */
3475 size = size * conf->copies;
3477 /* We need to round up when dividing by raid_disks to
3478 * get the stride size.
3480 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3482 conf->dev_sectors = size << conf->geo.chunk_shift;
3484 if (conf->geo.far_offset)
3485 conf->geo.stride = 1 << conf->geo.chunk_shift;
3487 sector_div(size, conf->geo.far_copies);
3488 conf->geo.stride = size << conf->geo.chunk_shift;
3492 enum geo_type {geo_new, geo_old, geo_start};
3493 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3496 int layout, chunk, disks;
3499 layout = mddev->layout;
3500 chunk = mddev->chunk_sectors;
3501 disks = mddev->raid_disks - mddev->delta_disks;
3504 layout = mddev->new_layout;
3505 chunk = mddev->new_chunk_sectors;
3506 disks = mddev->raid_disks;
3508 default: /* avoid 'may be unused' warnings */
3509 case geo_start: /* new when starting reshape - raid_disks not
3511 layout = mddev->new_layout;
3512 chunk = mddev->new_chunk_sectors;
3513 disks = mddev->raid_disks + mddev->delta_disks;
3518 if (chunk < (PAGE_SIZE >> 9) ||
3519 !is_power_of_2(chunk))
3522 fc = (layout >> 8) & 255;
3523 fo = layout & (1<<16);
3524 geo->raid_disks = disks;
3525 geo->near_copies = nc;
3526 geo->far_copies = fc;
3527 geo->far_offset = fo;
3528 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3529 geo->chunk_mask = chunk - 1;
3530 geo->chunk_shift = ffz(~chunk);
3534 static struct r10conf *setup_conf(struct mddev *mddev)
3536 struct r10conf *conf = NULL;
3541 copies = setup_geo(&geo, mddev, geo_new);
3544 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3545 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3546 mdname(mddev), PAGE_SIZE);
3550 if (copies < 2 || copies > mddev->raid_disks) {
3551 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3552 mdname(mddev), mddev->new_layout);
3557 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3561 /* FIXME calc properly */
3562 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3563 max(0,-mddev->delta_disks)),
3568 conf->tmppage = alloc_page(GFP_KERNEL);
3573 conf->copies = copies;
3574 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3575 r10bio_pool_free, conf);
3576 if (!conf->r10bio_pool)
3579 calc_sectors(conf, mddev->dev_sectors);
3580 if (mddev->reshape_position == MaxSector) {
3581 conf->prev = conf->geo;
3582 conf->reshape_progress = MaxSector;
3584 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3588 conf->reshape_progress = mddev->reshape_position;
3589 if (conf->prev.far_offset)
3590 conf->prev.stride = 1 << conf->prev.chunk_shift;
3592 /* far_copies must be 1 */
3593 conf->prev.stride = conf->dev_sectors;
3595 spin_lock_init(&conf->device_lock);
3596 INIT_LIST_HEAD(&conf->retry_list);
3598 spin_lock_init(&conf->resync_lock);
3599 init_waitqueue_head(&conf->wait_barrier);
3601 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3605 conf->mddev = mddev;
3610 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3613 if (conf->r10bio_pool)
3614 mempool_destroy(conf->r10bio_pool);
3615 kfree(conf->mirrors);
3616 safe_put_page(conf->tmppage);
3619 return ERR_PTR(err);
3622 static int run(struct mddev *mddev)
3624 struct r10conf *conf;
3625 int i, disk_idx, chunk_size;
3626 struct raid10_info *disk;
3627 struct md_rdev *rdev;
3629 sector_t min_offset_diff = 0;
3631 bool discard_supported = false;
3633 if (mddev->private == NULL) {
3634 conf = setup_conf(mddev);
3636 return PTR_ERR(conf);
3637 mddev->private = conf;
3639 conf = mddev->private;
3643 mddev->thread = conf->thread;
3644 conf->thread = NULL;
3646 chunk_size = mddev->chunk_sectors << 9;
3648 blk_queue_max_discard_sectors(mddev->queue,
3649 mddev->chunk_sectors);
3650 blk_queue_max_write_same_sectors(mddev->queue, 0);
3651 blk_queue_io_min(mddev->queue, chunk_size);
3652 if (conf->geo.raid_disks % conf->geo.near_copies)
3653 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3655 blk_queue_io_opt(mddev->queue, chunk_size *
3656 (conf->geo.raid_disks / conf->geo.near_copies));
3659 rdev_for_each(rdev, mddev) {
3661 struct request_queue *q;
3663 disk_idx = rdev->raid_disk;
3666 if (disk_idx >= conf->geo.raid_disks &&
3667 disk_idx >= conf->prev.raid_disks)
3669 disk = conf->mirrors + disk_idx;
3671 if (test_bit(Replacement, &rdev->flags)) {
3672 if (disk->replacement)
3674 disk->replacement = rdev;
3680 q = bdev_get_queue(rdev->bdev);
3681 if (q->merge_bvec_fn)
3682 mddev->merge_check_needed = 1;
3683 diff = (rdev->new_data_offset - rdev->data_offset);
3684 if (!mddev->reshape_backwards)
3688 if (first || diff < min_offset_diff)
3689 min_offset_diff = diff;
3692 disk_stack_limits(mddev->gendisk, rdev->bdev,
3693 rdev->data_offset << 9);
3695 disk->head_position = 0;
3697 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3698 discard_supported = true;
3702 if (discard_supported)
3703 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3706 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3709 /* need to check that every block has at least one working mirror */
3710 if (!enough(conf, -1)) {
3711 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3716 if (conf->reshape_progress != MaxSector) {
3717 /* must ensure that shape change is supported */
3718 if (conf->geo.far_copies != 1 &&
3719 conf->geo.far_offset == 0)
3721 if (conf->prev.far_copies != 1 &&
3722 conf->prev.far_offset == 0)
3726 mddev->degraded = 0;
3728 i < conf->geo.raid_disks
3729 || i < conf->prev.raid_disks;
3732 disk = conf->mirrors + i;
3734 if (!disk->rdev && disk->replacement) {
3735 /* The replacement is all we have - use it */
3736 disk->rdev = disk->replacement;
3737 disk->replacement = NULL;
3738 clear_bit(Replacement, &disk->rdev->flags);
3742 !test_bit(In_sync, &disk->rdev->flags)) {
3743 disk->head_position = 0;
3748 disk->recovery_disabled = mddev->recovery_disabled - 1;
3751 if (mddev->recovery_cp != MaxSector)
3752 printk(KERN_NOTICE "md/raid10:%s: not clean"
3753 " -- starting background reconstruction\n",
3756 "md/raid10:%s: active with %d out of %d devices\n",
3757 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3758 conf->geo.raid_disks);
3760 * Ok, everything is just fine now
3762 mddev->dev_sectors = conf->dev_sectors;
3763 size = raid10_size(mddev, 0, 0);
3764 md_set_array_sectors(mddev, size);
3765 mddev->resync_max_sectors = size;
3768 int stripe = conf->geo.raid_disks *
3769 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3770 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3771 mddev->queue->backing_dev_info.congested_data = mddev;
3773 /* Calculate max read-ahead size.
3774 * We need to readahead at least twice a whole stripe....
3777 stripe /= conf->geo.near_copies;
3778 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3779 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3780 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3784 if (md_integrity_register(mddev))
3787 if (conf->reshape_progress != MaxSector) {
3788 unsigned long before_length, after_length;
3790 before_length = ((1 << conf->prev.chunk_shift) *
3791 conf->prev.far_copies);
3792 after_length = ((1 << conf->geo.chunk_shift) *
3793 conf->geo.far_copies);
3795 if (max(before_length, after_length) > min_offset_diff) {
3796 /* This cannot work */
3797 printk("md/raid10: offset difference not enough to continue reshape\n");
3800 conf->offset_diff = min_offset_diff;
3802 conf->reshape_safe = conf->reshape_progress;
3803 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3804 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3805 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3806 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3807 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3814 md_unregister_thread(&mddev->thread);
3815 if (conf->r10bio_pool)
3816 mempool_destroy(conf->r10bio_pool);
3817 safe_put_page(conf->tmppage);
3818 kfree(conf->mirrors);
3820 mddev->private = NULL;
3825 static int stop(struct mddev *mddev)
3827 struct r10conf *conf = mddev->private;
3829 raise_barrier(conf, 0);
3830 lower_barrier(conf);
3832 md_unregister_thread(&mddev->thread);
3834 /* the unplug fn references 'conf'*/
3835 blk_sync_queue(mddev->queue);
3837 if (conf->r10bio_pool)
3838 mempool_destroy(conf->r10bio_pool);
3839 safe_put_page(conf->tmppage);
3840 kfree(conf->mirrors);
3842 mddev->private = NULL;
3846 static void raid10_quiesce(struct mddev *mddev, int state)
3848 struct r10conf *conf = mddev->private;
3852 raise_barrier(conf, 0);
3855 lower_barrier(conf);
3860 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3862 /* Resize of 'far' arrays is not supported.
3863 * For 'near' and 'offset' arrays we can set the
3864 * number of sectors used to be an appropriate multiple
3865 * of the chunk size.
3866 * For 'offset', this is far_copies*chunksize.
3867 * For 'near' the multiplier is the LCM of
3868 * near_copies and raid_disks.
3869 * So if far_copies > 1 && !far_offset, fail.
3870 * Else find LCM(raid_disks, near_copy)*far_copies and
3871 * multiply by chunk_size. Then round to this number.
3872 * This is mostly done by raid10_size()
3874 struct r10conf *conf = mddev->private;
3875 sector_t oldsize, size;
3877 if (mddev->reshape_position != MaxSector)
3880 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3883 oldsize = raid10_size(mddev, 0, 0);
3884 size = raid10_size(mddev, sectors, 0);
3885 if (mddev->external_size &&
3886 mddev->array_sectors > size)
3888 if (mddev->bitmap) {
3889 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3893 md_set_array_sectors(mddev, size);
3894 set_capacity(mddev->gendisk, mddev->array_sectors);
3895 revalidate_disk(mddev->gendisk);
3896 if (sectors > mddev->dev_sectors &&
3897 mddev->recovery_cp > oldsize) {
3898 mddev->recovery_cp = oldsize;
3899 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3901 calc_sectors(conf, sectors);
3902 mddev->dev_sectors = conf->dev_sectors;
3903 mddev->resync_max_sectors = size;
3907 static void *raid10_takeover_raid0(struct mddev *mddev)
3909 struct md_rdev *rdev;
3910 struct r10conf *conf;
3912 if (mddev->degraded > 0) {
3913 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3915 return ERR_PTR(-EINVAL);
3918 /* Set new parameters */
3919 mddev->new_level = 10;
3920 /* new layout: far_copies = 1, near_copies = 2 */
3921 mddev->new_layout = (1<<8) + 2;
3922 mddev->new_chunk_sectors = mddev->chunk_sectors;
3923 mddev->delta_disks = mddev->raid_disks;
3924 mddev->raid_disks *= 2;
3925 /* make sure it will be not marked as dirty */
3926 mddev->recovery_cp = MaxSector;
3928 conf = setup_conf(mddev);
3929 if (!IS_ERR(conf)) {
3930 rdev_for_each(rdev, mddev)
3931 if (rdev->raid_disk >= 0)
3932 rdev->new_raid_disk = rdev->raid_disk * 2;
3939 static void *raid10_takeover(struct mddev *mddev)
3941 struct r0conf *raid0_conf;
3943 /* raid10 can take over:
3944 * raid0 - providing it has only two drives
3946 if (mddev->level == 0) {
3947 /* for raid0 takeover only one zone is supported */
3948 raid0_conf = mddev->private;
3949 if (raid0_conf->nr_strip_zones > 1) {
3950 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3951 " with more than one zone.\n",
3953 return ERR_PTR(-EINVAL);
3955 return raid10_takeover_raid0(mddev);
3957 return ERR_PTR(-EINVAL);
3960 static int raid10_check_reshape(struct mddev *mddev)
3962 /* Called when there is a request to change
3963 * - layout (to ->new_layout)
3964 * - chunk size (to ->new_chunk_sectors)
3965 * - raid_disks (by delta_disks)
3966 * or when trying to restart a reshape that was ongoing.
3968 * We need to validate the request and possibly allocate
3969 * space if that might be an issue later.
3971 * Currently we reject any reshape of a 'far' mode array,
3972 * allow chunk size to change if new is generally acceptable,
3973 * allow raid_disks to increase, and allow
3974 * a switch between 'near' mode and 'offset' mode.
3976 struct r10conf *conf = mddev->private;
3979 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3982 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3983 /* mustn't change number of copies */
3985 if (geo.far_copies > 1 && !geo.far_offset)
3986 /* Cannot switch to 'far' mode */
3989 if (mddev->array_sectors & geo.chunk_mask)
3990 /* not factor of array size */
3993 if (!enough(conf, -1))
3996 kfree(conf->mirrors_new);
3997 conf->mirrors_new = NULL;
3998 if (mddev->delta_disks > 0) {
3999 /* allocate new 'mirrors' list */
4000 conf->mirrors_new = kzalloc(
4001 sizeof(struct raid10_info)
4002 *(mddev->raid_disks +
4003 mddev->delta_disks),
4005 if (!conf->mirrors_new)
4012 * Need to check if array has failed when deciding whether to:
4014 * - remove non-faulty devices
4017 * This determination is simple when no reshape is happening.
4018 * However if there is a reshape, we need to carefully check
4019 * both the before and after sections.
4020 * This is because some failed devices may only affect one
4021 * of the two sections, and some non-in_sync devices may
4022 * be insync in the section most affected by failed devices.
4024 static int calc_degraded(struct r10conf *conf)
4026 int degraded, degraded2;
4031 /* 'prev' section first */
4032 for (i = 0; i < conf->prev.raid_disks; i++) {
4033 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4034 if (!rdev || test_bit(Faulty, &rdev->flags))
4036 else if (!test_bit(In_sync, &rdev->flags))
4037 /* When we can reduce the number of devices in
4038 * an array, this might not contribute to
4039 * 'degraded'. It does now.
4044 if (conf->geo.raid_disks == conf->prev.raid_disks)
4048 for (i = 0; i < conf->geo.raid_disks; i++) {
4049 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4050 if (!rdev || test_bit(Faulty, &rdev->flags))
4052 else if (!test_bit(In_sync, &rdev->flags)) {
4053 /* If reshape is increasing the number of devices,
4054 * this section has already been recovered, so
4055 * it doesn't contribute to degraded.
4058 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4063 if (degraded2 > degraded)
4068 static int raid10_start_reshape(struct mddev *mddev)
4070 /* A 'reshape' has been requested. This commits
4071 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4072 * This also checks if there are enough spares and adds them
4074 * We currently require enough spares to make the final
4075 * array non-degraded. We also require that the difference
4076 * between old and new data_offset - on each device - is
4077 * enough that we never risk over-writing.
4080 unsigned long before_length, after_length;
4081 sector_t min_offset_diff = 0;
4084 struct r10conf *conf = mddev->private;
4085 struct md_rdev *rdev;
4089 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4092 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4095 before_length = ((1 << conf->prev.chunk_shift) *
4096 conf->prev.far_copies);
4097 after_length = ((1 << conf->geo.chunk_shift) *
4098 conf->geo.far_copies);
4100 rdev_for_each(rdev, mddev) {
4101 if (!test_bit(In_sync, &rdev->flags)
4102 && !test_bit(Faulty, &rdev->flags))
4104 if (rdev->raid_disk >= 0) {
4105 long long diff = (rdev->new_data_offset
4106 - rdev->data_offset);
4107 if (!mddev->reshape_backwards)
4111 if (first || diff < min_offset_diff)
4112 min_offset_diff = diff;
4116 if (max(before_length, after_length) > min_offset_diff)
4119 if (spares < mddev->delta_disks)
4122 conf->offset_diff = min_offset_diff;
4123 spin_lock_irq(&conf->device_lock);
4124 if (conf->mirrors_new) {
4125 memcpy(conf->mirrors_new, conf->mirrors,
4126 sizeof(struct raid10_info)*conf->prev.raid_disks);
4128 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4129 conf->mirrors_old = conf->mirrors;
4130 conf->mirrors = conf->mirrors_new;
4131 conf->mirrors_new = NULL;
4133 setup_geo(&conf->geo, mddev, geo_start);
4135 if (mddev->reshape_backwards) {
4136 sector_t size = raid10_size(mddev, 0, 0);
4137 if (size < mddev->array_sectors) {
4138 spin_unlock_irq(&conf->device_lock);
4139 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4143 mddev->resync_max_sectors = size;
4144 conf->reshape_progress = size;
4146 conf->reshape_progress = 0;
4147 spin_unlock_irq(&conf->device_lock);
4149 if (mddev->delta_disks && mddev->bitmap) {
4150 ret = bitmap_resize(mddev->bitmap,
4151 raid10_size(mddev, 0,
4152 conf->geo.raid_disks),
4157 if (mddev->delta_disks > 0) {
4158 rdev_for_each(rdev, mddev)
4159 if (rdev->raid_disk < 0 &&
4160 !test_bit(Faulty, &rdev->flags)) {
4161 if (raid10_add_disk(mddev, rdev) == 0) {
4162 if (rdev->raid_disk >=
4163 conf->prev.raid_disks)
4164 set_bit(In_sync, &rdev->flags);
4166 rdev->recovery_offset = 0;
4168 if (sysfs_link_rdev(mddev, rdev))
4169 /* Failure here is OK */;
4171 } else if (rdev->raid_disk >= conf->prev.raid_disks
4172 && !test_bit(Faulty, &rdev->flags)) {
4173 /* This is a spare that was manually added */
4174 set_bit(In_sync, &rdev->flags);
4177 /* When a reshape changes the number of devices,
4178 * ->degraded is measured against the larger of the
4179 * pre and post numbers.
4181 spin_lock_irq(&conf->device_lock);
4182 mddev->degraded = calc_degraded(conf);
4183 spin_unlock_irq(&conf->device_lock);
4184 mddev->raid_disks = conf->geo.raid_disks;
4185 mddev->reshape_position = conf->reshape_progress;
4186 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4188 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4189 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4190 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4191 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4193 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4195 if (!mddev->sync_thread) {
4199 conf->reshape_checkpoint = jiffies;
4200 md_wakeup_thread(mddev->sync_thread);
4201 md_new_event(mddev);
4205 mddev->recovery = 0;
4206 spin_lock_irq(&conf->device_lock);
4207 conf->geo = conf->prev;
4208 mddev->raid_disks = conf->geo.raid_disks;
4209 rdev_for_each(rdev, mddev)
4210 rdev->new_data_offset = rdev->data_offset;
4212 conf->reshape_progress = MaxSector;
4213 mddev->reshape_position = MaxSector;
4214 spin_unlock_irq(&conf->device_lock);
4218 /* Calculate the last device-address that could contain
4219 * any block from the chunk that includes the array-address 's'
4220 * and report the next address.
4221 * i.e. the address returned will be chunk-aligned and after
4222 * any data that is in the chunk containing 's'.
4224 static sector_t last_dev_address(sector_t s, struct geom *geo)
4226 s = (s | geo->chunk_mask) + 1;
4227 s >>= geo->chunk_shift;
4228 s *= geo->near_copies;
4229 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4230 s *= geo->far_copies;
4231 s <<= geo->chunk_shift;
4235 /* Calculate the first device-address that could contain
4236 * any block from the chunk that includes the array-address 's'.
4237 * This too will be the start of a chunk
4239 static sector_t first_dev_address(sector_t s, struct geom *geo)
4241 s >>= geo->chunk_shift;
4242 s *= geo->near_copies;
4243 sector_div(s, geo->raid_disks);
4244 s *= geo->far_copies;
4245 s <<= geo->chunk_shift;
4249 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4252 /* We simply copy at most one chunk (smallest of old and new)
4253 * at a time, possibly less if that exceeds RESYNC_PAGES,
4254 * or we hit a bad block or something.
4255 * This might mean we pause for normal IO in the middle of
4256 * a chunk, but that is not a problem was mddev->reshape_position
4257 * can record any location.
4259 * If we will want to write to a location that isn't
4260 * yet recorded as 'safe' (i.e. in metadata on disk) then
4261 * we need to flush all reshape requests and update the metadata.
4263 * When reshaping forwards (e.g. to more devices), we interpret
4264 * 'safe' as the earliest block which might not have been copied
4265 * down yet. We divide this by previous stripe size and multiply
4266 * by previous stripe length to get lowest device offset that we
4267 * cannot write to yet.
4268 * We interpret 'sector_nr' as an address that we want to write to.
4269 * From this we use last_device_address() to find where we might
4270 * write to, and first_device_address on the 'safe' position.
4271 * If this 'next' write position is after the 'safe' position,
4272 * we must update the metadata to increase the 'safe' position.
4274 * When reshaping backwards, we round in the opposite direction
4275 * and perform the reverse test: next write position must not be
4276 * less than current safe position.
4278 * In all this the minimum difference in data offsets
4279 * (conf->offset_diff - always positive) allows a bit of slack,
4280 * so next can be after 'safe', but not by more than offset_disk
4282 * We need to prepare all the bios here before we start any IO
4283 * to ensure the size we choose is acceptable to all devices.
4284 * The means one for each copy for write-out and an extra one for
4286 * We store the read-in bio in ->master_bio and the others in
4287 * ->devs[x].bio and ->devs[x].repl_bio.
4289 struct r10conf *conf = mddev->private;
4290 struct r10bio *r10_bio;
4291 sector_t next, safe, last;
4295 struct md_rdev *rdev;
4298 struct bio *bio, *read_bio;
4299 int sectors_done = 0;
4301 if (sector_nr == 0) {
4302 /* If restarting in the middle, skip the initial sectors */
4303 if (mddev->reshape_backwards &&
4304 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4305 sector_nr = (raid10_size(mddev, 0, 0)
4306 - conf->reshape_progress);
4307 } else if (!mddev->reshape_backwards &&
4308 conf->reshape_progress > 0)
4309 sector_nr = conf->reshape_progress;
4311 mddev->curr_resync_completed = sector_nr;
4312 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4318 /* We don't use sector_nr to track where we are up to
4319 * as that doesn't work well for ->reshape_backwards.
4320 * So just use ->reshape_progress.
4322 if (mddev->reshape_backwards) {
4323 /* 'next' is the earliest device address that we might
4324 * write to for this chunk in the new layout
4326 next = first_dev_address(conf->reshape_progress - 1,
4329 /* 'safe' is the last device address that we might read from
4330 * in the old layout after a restart
4332 safe = last_dev_address(conf->reshape_safe - 1,
4335 if (next + conf->offset_diff < safe)
4338 last = conf->reshape_progress - 1;
4339 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4340 & conf->prev.chunk_mask);
4341 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4342 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4344 /* 'next' is after the last device address that we
4345 * might write to for this chunk in the new layout
4347 next = last_dev_address(conf->reshape_progress, &conf->geo);
4349 /* 'safe' is the earliest device address that we might
4350 * read from in the old layout after a restart
4352 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4354 /* Need to update metadata if 'next' might be beyond 'safe'
4355 * as that would possibly corrupt data
4357 if (next > safe + conf->offset_diff)
4360 sector_nr = conf->reshape_progress;
4361 last = sector_nr | (conf->geo.chunk_mask
4362 & conf->prev.chunk_mask);
4364 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4365 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4369 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4370 /* Need to update reshape_position in metadata */
4372 mddev->reshape_position = conf->reshape_progress;
4373 if (mddev->reshape_backwards)
4374 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4375 - conf->reshape_progress;
4377 mddev->curr_resync_completed = conf->reshape_progress;
4378 conf->reshape_checkpoint = jiffies;
4379 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4380 md_wakeup_thread(mddev->thread);
4381 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4382 kthread_should_stop());
4383 conf->reshape_safe = mddev->reshape_position;
4384 allow_barrier(conf);
4388 /* Now schedule reads for blocks from sector_nr to last */
4389 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4390 raise_barrier(conf, sectors_done != 0);
4391 atomic_set(&r10_bio->remaining, 0);
4392 r10_bio->mddev = mddev;
4393 r10_bio->sector = sector_nr;
4394 set_bit(R10BIO_IsReshape, &r10_bio->state);
4395 r10_bio->sectors = last - sector_nr + 1;
4396 rdev = read_balance(conf, r10_bio, &max_sectors);
4397 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4400 /* Cannot read from here, so need to record bad blocks
4401 * on all the target devices.
4404 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4405 return sectors_done;
4408 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4410 read_bio->bi_bdev = rdev->bdev;
4411 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4412 + rdev->data_offset);
4413 read_bio->bi_private = r10_bio;
4414 read_bio->bi_end_io = end_sync_read;
4415 read_bio->bi_rw = READ;
4416 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4417 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4418 read_bio->bi_vcnt = 0;
4419 read_bio->bi_size = 0;
4420 r10_bio->master_bio = read_bio;
4421 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4423 /* Now find the locations in the new layout */
4424 __raid10_find_phys(&conf->geo, r10_bio);
4427 read_bio->bi_next = NULL;
4429 for (s = 0; s < conf->copies*2; s++) {
4431 int d = r10_bio->devs[s/2].devnum;
4432 struct md_rdev *rdev2;
4434 rdev2 = conf->mirrors[d].replacement;
4435 b = r10_bio->devs[s/2].repl_bio;
4437 rdev2 = conf->mirrors[d].rdev;
4438 b = r10_bio->devs[s/2].bio;
4440 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4444 b->bi_bdev = rdev2->bdev;
4445 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4446 b->bi_private = r10_bio;
4447 b->bi_end_io = end_reshape_write;
4453 /* Now add as many pages as possible to all of these bios. */
4456 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4457 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4458 int len = (max_sectors - s) << 9;
4459 if (len > PAGE_SIZE)
4461 for (bio = blist; bio ; bio = bio->bi_next) {
4463 if (bio_add_page(bio, page, len, 0))
4466 /* Didn't fit, must stop */
4468 bio2 && bio2 != bio;
4469 bio2 = bio2->bi_next) {
4470 /* Remove last page from this bio */
4472 bio2->bi_size -= len;
4473 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4477 sector_nr += len >> 9;
4478 nr_sectors += len >> 9;
4481 r10_bio->sectors = nr_sectors;
4483 /* Now submit the read */
4484 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4485 atomic_inc(&r10_bio->remaining);
4486 read_bio->bi_next = NULL;
4487 generic_make_request(read_bio);
4488 sector_nr += nr_sectors;
4489 sectors_done += nr_sectors;
4490 if (sector_nr <= last)
4493 /* Now that we have done the whole section we can
4494 * update reshape_progress
4496 if (mddev->reshape_backwards)
4497 conf->reshape_progress -= sectors_done;
4499 conf->reshape_progress += sectors_done;
4501 return sectors_done;
4504 static void end_reshape_request(struct r10bio *r10_bio);
4505 static int handle_reshape_read_error(struct mddev *mddev,
4506 struct r10bio *r10_bio);
4507 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4509 /* Reshape read completed. Hopefully we have a block
4511 * If we got a read error then we do sync 1-page reads from
4512 * elsewhere until we find the data - or give up.
4514 struct r10conf *conf = mddev->private;
4517 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4518 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4519 /* Reshape has been aborted */
4520 md_done_sync(mddev, r10_bio->sectors, 0);
4524 /* We definitely have the data in the pages, schedule the
4527 atomic_set(&r10_bio->remaining, 1);
4528 for (s = 0; s < conf->copies*2; s++) {
4530 int d = r10_bio->devs[s/2].devnum;
4531 struct md_rdev *rdev;
4533 rdev = conf->mirrors[d].replacement;
4534 b = r10_bio->devs[s/2].repl_bio;
4536 rdev = conf->mirrors[d].rdev;
4537 b = r10_bio->devs[s/2].bio;
4539 if (!rdev || test_bit(Faulty, &rdev->flags))
4541 atomic_inc(&rdev->nr_pending);
4542 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4543 atomic_inc(&r10_bio->remaining);
4545 generic_make_request(b);
4547 end_reshape_request(r10_bio);
4550 static void end_reshape(struct r10conf *conf)
4552 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4555 spin_lock_irq(&conf->device_lock);
4556 conf->prev = conf->geo;
4557 md_finish_reshape(conf->mddev);
4559 conf->reshape_progress = MaxSector;
4560 spin_unlock_irq(&conf->device_lock);
4562 /* read-ahead size must cover two whole stripes, which is
4563 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4565 if (conf->mddev->queue) {
4566 int stripe = conf->geo.raid_disks *
4567 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4568 stripe /= conf->geo.near_copies;
4569 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4570 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4576 static int handle_reshape_read_error(struct mddev *mddev,
4577 struct r10bio *r10_bio)
4579 /* Use sync reads to get the blocks from somewhere else */
4580 int sectors = r10_bio->sectors;
4581 struct r10conf *conf = mddev->private;
4583 struct r10bio r10_bio;
4584 struct r10dev devs[conf->copies];
4586 struct r10bio *r10b = &on_stack.r10_bio;
4589 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4591 r10b->sector = r10_bio->sector;
4592 __raid10_find_phys(&conf->prev, r10b);
4597 int first_slot = slot;
4599 if (s > (PAGE_SIZE >> 9))
4603 int d = r10b->devs[slot].devnum;
4604 struct md_rdev *rdev = conf->mirrors[d].rdev;
4607 test_bit(Faulty, &rdev->flags) ||
4608 !test_bit(In_sync, &rdev->flags))
4611 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4612 success = sync_page_io(rdev,
4621 if (slot >= conf->copies)
4623 if (slot == first_slot)
4627 /* couldn't read this block, must give up */
4628 set_bit(MD_RECOVERY_INTR,
4638 static void end_reshape_write(struct bio *bio, int error)
4640 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4641 struct r10bio *r10_bio = bio->bi_private;
4642 struct mddev *mddev = r10_bio->mddev;
4643 struct r10conf *conf = mddev->private;
4647 struct md_rdev *rdev = NULL;
4649 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4651 rdev = conf->mirrors[d].replacement;
4654 rdev = conf->mirrors[d].rdev;
4658 /* FIXME should record badblock */
4659 md_error(mddev, rdev);
4662 rdev_dec_pending(rdev, mddev);
4663 end_reshape_request(r10_bio);
4666 static void end_reshape_request(struct r10bio *r10_bio)
4668 if (!atomic_dec_and_test(&r10_bio->remaining))
4670 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4671 bio_put(r10_bio->master_bio);
4675 static void raid10_finish_reshape(struct mddev *mddev)
4677 struct r10conf *conf = mddev->private;
4679 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4682 if (mddev->delta_disks > 0) {
4683 sector_t size = raid10_size(mddev, 0, 0);
4684 md_set_array_sectors(mddev, size);
4685 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4686 mddev->recovery_cp = mddev->resync_max_sectors;
4687 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4689 mddev->resync_max_sectors = size;
4690 set_capacity(mddev->gendisk, mddev->array_sectors);
4691 revalidate_disk(mddev->gendisk);
4694 for (d = conf->geo.raid_disks ;
4695 d < conf->geo.raid_disks - mddev->delta_disks;
4697 struct md_rdev *rdev = conf->mirrors[d].rdev;
4699 clear_bit(In_sync, &rdev->flags);
4700 rdev = conf->mirrors[d].replacement;
4702 clear_bit(In_sync, &rdev->flags);
4705 mddev->layout = mddev->new_layout;
4706 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4707 mddev->reshape_position = MaxSector;
4708 mddev->delta_disks = 0;
4709 mddev->reshape_backwards = 0;
4712 static struct md_personality raid10_personality =
4716 .owner = THIS_MODULE,
4717 .make_request = make_request,
4721 .error_handler = error,
4722 .hot_add_disk = raid10_add_disk,
4723 .hot_remove_disk= raid10_remove_disk,
4724 .spare_active = raid10_spare_active,
4725 .sync_request = sync_request,
4726 .quiesce = raid10_quiesce,
4727 .size = raid10_size,
4728 .resize = raid10_resize,
4729 .takeover = raid10_takeover,
4730 .check_reshape = raid10_check_reshape,
4731 .start_reshape = raid10_start_reshape,
4732 .finish_reshape = raid10_finish_reshape,
4735 static int __init raid_init(void)
4737 return register_md_personality(&raid10_personality);
4740 static void raid_exit(void)
4742 unregister_md_personality(&raid10_personality);
4745 module_init(raid_init);
4746 module_exit(raid_exit);
4747 MODULE_LICENSE("GPL");
4748 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4749 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4750 MODULE_ALIAS("md-raid10");
4751 MODULE_ALIAS("md-level-10");
4753 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);