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 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"
396 spin_lock_irqsave(&conf->device_lock, flags);
397 if (!enough(conf, rdev->raid_disk))
399 spin_unlock_irqrestore(&conf->device_lock, flags);
402 raid_end_bio_io(r10_bio);
403 rdev_dec_pending(rdev, conf->mddev);
406 * oops, read error - keep the refcount on the rdev
408 char b[BDEVNAME_SIZE];
409 printk_ratelimited(KERN_ERR
410 "md/raid10:%s: %s: rescheduling sector %llu\n",
412 bdevname(rdev->bdev, b),
413 (unsigned long long)r10_bio->sector);
414 set_bit(R10BIO_ReadError, &r10_bio->state);
415 reschedule_retry(r10_bio);
419 static void close_write(struct r10bio *r10_bio)
421 /* clear the bitmap if all writes complete successfully */
422 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
424 !test_bit(R10BIO_Degraded, &r10_bio->state),
426 md_write_end(r10_bio->mddev);
429 static void one_write_done(struct r10bio *r10_bio)
431 if (atomic_dec_and_test(&r10_bio->remaining)) {
432 if (test_bit(R10BIO_WriteError, &r10_bio->state))
433 reschedule_retry(r10_bio);
435 close_write(r10_bio);
436 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437 reschedule_retry(r10_bio);
439 raid_end_bio_io(r10_bio);
444 static void raid10_end_write_request(struct bio *bio, int error)
446 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
447 struct r10bio *r10_bio = bio->bi_private;
450 struct r10conf *conf = r10_bio->mddev->private;
452 struct md_rdev *rdev = NULL;
454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
457 rdev = conf->mirrors[dev].replacement;
461 rdev = conf->mirrors[dev].rdev;
464 * this branch is our 'one mirror IO has finished' event handler:
468 /* Never record new bad blocks to replacement,
471 md_error(rdev->mddev, rdev);
473 set_bit(WriteErrorSeen, &rdev->flags);
474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
477 set_bit(R10BIO_WriteError, &r10_bio->state);
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
494 * Do not set R10BIO_Uptodate if the current device is
495 * rebuilding or Faulty. This is because we cannot use
496 * such device for properly reading the data back (we could
497 * potentially use it, if the current write would have felt
498 * before rdev->recovery_offset, but for simplicity we don't
501 if (test_bit(In_sync, &rdev->flags) &&
502 !test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_Uptodate, &r10_bio->state);
505 /* Maybe we can clear some bad blocks. */
506 if (is_badblock(rdev,
507 r10_bio->devs[slot].addr,
509 &first_bad, &bad_sectors)) {
512 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
514 r10_bio->devs[slot].bio = IO_MADE_GOOD;
516 set_bit(R10BIO_MadeGood, &r10_bio->state);
522 * Let's see if all mirrored write operations have finished
525 one_write_done(r10_bio);
527 rdev_dec_pending(rdev, conf->mddev);
531 * RAID10 layout manager
532 * As well as the chunksize and raid_disks count, there are two
533 * parameters: near_copies and far_copies.
534 * near_copies * far_copies must be <= raid_disks.
535 * Normally one of these will be 1.
536 * If both are 1, we get raid0.
537 * If near_copies == raid_disks, we get raid1.
539 * Chunks are laid out in raid0 style with near_copies copies of the
540 * first chunk, followed by near_copies copies of the next chunk and
542 * If far_copies > 1, then after 1/far_copies of the array has been assigned
543 * as described above, we start again with a device offset of near_copies.
544 * So we effectively have another copy of the whole array further down all
545 * the drives, but with blocks on different drives.
546 * With this layout, and block is never stored twice on the one device.
548 * raid10_find_phys finds the sector offset of a given virtual sector
549 * on each device that it is on.
551 * raid10_find_virt does the reverse mapping, from a device and a
552 * sector offset to a virtual address
555 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
563 int last_far_set_start, last_far_set_size;
565 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
566 last_far_set_start *= geo->far_set_size;
568 last_far_set_size = geo->far_set_size;
569 last_far_set_size += (geo->raid_disks % geo->far_set_size);
571 /* now calculate first sector/dev */
572 chunk = r10bio->sector >> geo->chunk_shift;
573 sector = r10bio->sector & geo->chunk_mask;
575 chunk *= geo->near_copies;
577 dev = sector_div(stripe, geo->raid_disks);
579 stripe *= geo->far_copies;
581 sector += stripe << geo->chunk_shift;
583 /* and calculate all the others */
584 for (n = 0; n < geo->near_copies; n++) {
588 r10bio->devs[slot].devnum = d;
589 r10bio->devs[slot].addr = s;
592 for (f = 1; f < geo->far_copies; f++) {
593 set = d / geo->far_set_size;
594 d += geo->near_copies;
596 if ((geo->raid_disks % geo->far_set_size) &&
597 (d > last_far_set_start)) {
598 d -= last_far_set_start;
599 d %= last_far_set_size;
600 d += last_far_set_start;
602 d %= geo->far_set_size;
603 d += geo->far_set_size * set;
606 r10bio->devs[slot].devnum = d;
607 r10bio->devs[slot].addr = s;
611 if (dev >= geo->raid_disks) {
613 sector += (geo->chunk_mask + 1);
618 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
620 struct geom *geo = &conf->geo;
622 if (conf->reshape_progress != MaxSector &&
623 ((r10bio->sector >= conf->reshape_progress) !=
624 conf->mddev->reshape_backwards)) {
625 set_bit(R10BIO_Previous, &r10bio->state);
628 clear_bit(R10BIO_Previous, &r10bio->state);
630 __raid10_find_phys(geo, r10bio);
633 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
635 sector_t offset, chunk, vchunk;
636 /* Never use conf->prev as this is only called during resync
637 * or recovery, so reshape isn't happening
639 struct geom *geo = &conf->geo;
640 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
641 int far_set_size = geo->far_set_size;
642 int last_far_set_start;
644 if (geo->raid_disks % geo->far_set_size) {
645 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
646 last_far_set_start *= geo->far_set_size;
648 if (dev >= last_far_set_start) {
649 far_set_size = geo->far_set_size;
650 far_set_size += (geo->raid_disks % geo->far_set_size);
651 far_set_start = last_far_set_start;
655 offset = sector & geo->chunk_mask;
656 if (geo->far_offset) {
658 chunk = sector >> geo->chunk_shift;
659 fc = sector_div(chunk, geo->far_copies);
660 dev -= fc * geo->near_copies;
661 if (dev < far_set_start)
664 while (sector >= geo->stride) {
665 sector -= geo->stride;
666 if (dev < (geo->near_copies + far_set_start))
667 dev += far_set_size - geo->near_copies;
669 dev -= geo->near_copies;
671 chunk = sector >> geo->chunk_shift;
673 vchunk = chunk * geo->raid_disks + dev;
674 sector_div(vchunk, geo->near_copies);
675 return (vchunk << geo->chunk_shift) + offset;
679 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
681 * @bvm: properties of new bio
682 * @biovec: the request that could be merged to it.
684 * Return amount of bytes we can accept at this offset
685 * This requires checking for end-of-chunk if near_copies != raid_disks,
686 * and for subordinate merge_bvec_fns if merge_check_needed.
688 static int raid10_mergeable_bvec(struct request_queue *q,
689 struct bvec_merge_data *bvm,
690 struct bio_vec *biovec)
692 struct mddev *mddev = q->queuedata;
693 struct r10conf *conf = mddev->private;
694 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
696 unsigned int chunk_sectors;
697 unsigned int bio_sectors = bvm->bi_size >> 9;
698 struct geom *geo = &conf->geo;
700 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
701 if (conf->reshape_progress != MaxSector &&
702 ((sector >= conf->reshape_progress) !=
703 conf->mddev->reshape_backwards))
706 if (geo->near_copies < geo->raid_disks) {
707 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
708 + bio_sectors)) << 9;
710 /* bio_add cannot handle a negative return */
712 if (max <= biovec->bv_len && bio_sectors == 0)
713 return biovec->bv_len;
715 max = biovec->bv_len;
717 if (mddev->merge_check_needed) {
719 struct r10bio r10_bio;
720 struct r10dev devs[conf->copies];
722 struct r10bio *r10_bio = &on_stack.r10_bio;
724 if (conf->reshape_progress != MaxSector) {
725 /* Cannot give any guidance during reshape */
726 if (max <= biovec->bv_len && bio_sectors == 0)
727 return biovec->bv_len;
730 r10_bio->sector = sector;
731 raid10_find_phys(conf, r10_bio);
733 for (s = 0; s < conf->copies; s++) {
734 int disk = r10_bio->devs[s].devnum;
735 struct md_rdev *rdev = rcu_dereference(
736 conf->mirrors[disk].rdev);
737 if (rdev && !test_bit(Faulty, &rdev->flags)) {
738 struct request_queue *q =
739 bdev_get_queue(rdev->bdev);
740 if (q->merge_bvec_fn) {
741 bvm->bi_sector = r10_bio->devs[s].addr
743 bvm->bi_bdev = rdev->bdev;
744 max = min(max, q->merge_bvec_fn(
748 rdev = rcu_dereference(conf->mirrors[disk].replacement);
749 if (rdev && !test_bit(Faulty, &rdev->flags)) {
750 struct request_queue *q =
751 bdev_get_queue(rdev->bdev);
752 if (q->merge_bvec_fn) {
753 bvm->bi_sector = r10_bio->devs[s].addr
755 bvm->bi_bdev = rdev->bdev;
756 max = min(max, q->merge_bvec_fn(
767 * This routine returns the disk from which the requested read should
768 * be done. There is a per-array 'next expected sequential IO' sector
769 * number - if this matches on the next IO then we use the last disk.
770 * There is also a per-disk 'last know head position' sector that is
771 * maintained from IRQ contexts, both the normal and the resync IO
772 * completion handlers update this position correctly. If there is no
773 * perfect sequential match then we pick the disk whose head is closest.
775 * If there are 2 mirrors in the same 2 devices, performance degrades
776 * because position is mirror, not device based.
778 * The rdev for the device selected will have nr_pending incremented.
782 * FIXME: possibly should rethink readbalancing and do it differently
783 * depending on near_copies / far_copies geometry.
785 static struct md_rdev *read_balance(struct r10conf *conf,
786 struct r10bio *r10_bio,
789 const sector_t this_sector = r10_bio->sector;
791 int sectors = r10_bio->sectors;
792 int best_good_sectors;
793 sector_t new_distance, best_dist;
794 struct md_rdev *best_rdev, *rdev = NULL;
797 struct geom *geo = &conf->geo;
799 raid10_find_phys(conf, r10_bio);
802 sectors = r10_bio->sectors;
805 best_dist = MaxSector;
806 best_good_sectors = 0;
809 * Check if we can balance. We can balance on the whole
810 * device if no resync is going on (recovery is ok), or below
811 * the resync window. We take the first readable disk when
812 * above the resync window.
814 if (conf->mddev->recovery_cp < MaxSector
815 && (this_sector + sectors >= conf->next_resync))
818 for (slot = 0; slot < conf->copies ; slot++) {
823 if (r10_bio->devs[slot].bio == IO_BLOCKED)
825 disk = r10_bio->devs[slot].devnum;
826 rdev = rcu_dereference(conf->mirrors[disk].replacement);
827 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
828 test_bit(Unmerged, &rdev->flags) ||
829 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
830 rdev = rcu_dereference(conf->mirrors[disk].rdev);
832 test_bit(Faulty, &rdev->flags) ||
833 test_bit(Unmerged, &rdev->flags))
835 if (!test_bit(In_sync, &rdev->flags) &&
836 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
839 dev_sector = r10_bio->devs[slot].addr;
840 if (is_badblock(rdev, dev_sector, sectors,
841 &first_bad, &bad_sectors)) {
842 if (best_dist < MaxSector)
843 /* Already have a better slot */
845 if (first_bad <= dev_sector) {
846 /* Cannot read here. If this is the
847 * 'primary' device, then we must not read
848 * beyond 'bad_sectors' from another device.
850 bad_sectors -= (dev_sector - first_bad);
851 if (!do_balance && sectors > bad_sectors)
852 sectors = bad_sectors;
853 if (best_good_sectors > sectors)
854 best_good_sectors = sectors;
856 sector_t good_sectors =
857 first_bad - dev_sector;
858 if (good_sectors > best_good_sectors) {
859 best_good_sectors = good_sectors;
864 /* Must read from here */
869 best_good_sectors = sectors;
874 /* This optimisation is debatable, and completely destroys
875 * sequential read speed for 'far copies' arrays. So only
876 * keep it for 'near' arrays, and review those later.
878 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
881 /* for far > 1 always use the lowest address */
882 if (geo->far_copies > 1)
883 new_distance = r10_bio->devs[slot].addr;
885 new_distance = abs(r10_bio->devs[slot].addr -
886 conf->mirrors[disk].head_position);
887 if (new_distance < best_dist) {
888 best_dist = new_distance;
893 if (slot >= conf->copies) {
899 atomic_inc(&rdev->nr_pending);
900 if (test_bit(Faulty, &rdev->flags)) {
901 /* Cannot risk returning a device that failed
902 * before we inc'ed nr_pending
904 rdev_dec_pending(rdev, conf->mddev);
907 r10_bio->read_slot = slot;
911 *max_sectors = best_good_sectors;
916 int md_raid10_congested(struct mddev *mddev, int bits)
918 struct r10conf *conf = mddev->private;
921 if ((bits & (1 << BDI_async_congested)) &&
922 conf->pending_count >= max_queued_requests)
927 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
930 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
931 if (rdev && !test_bit(Faulty, &rdev->flags)) {
932 struct request_queue *q = bdev_get_queue(rdev->bdev);
934 ret |= bdi_congested(&q->backing_dev_info, bits);
940 EXPORT_SYMBOL_GPL(md_raid10_congested);
942 static int raid10_congested(void *data, int bits)
944 struct mddev *mddev = data;
946 return mddev_congested(mddev, bits) ||
947 md_raid10_congested(mddev, bits);
950 static void flush_pending_writes(struct r10conf *conf)
952 /* Any writes that have been queued but are awaiting
953 * bitmap updates get flushed here.
955 spin_lock_irq(&conf->device_lock);
957 if (conf->pending_bio_list.head) {
959 bio = bio_list_get(&conf->pending_bio_list);
960 conf->pending_count = 0;
961 spin_unlock_irq(&conf->device_lock);
962 /* flush any pending bitmap writes to disk
963 * before proceeding w/ I/O */
964 bitmap_unplug(conf->mddev->bitmap);
965 wake_up(&conf->wait_barrier);
967 while (bio) { /* submit pending writes */
968 struct bio *next = bio->bi_next;
970 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
971 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
975 generic_make_request(bio);
979 spin_unlock_irq(&conf->device_lock);
983 * Sometimes we need to suspend IO while we do something else,
984 * either some resync/recovery, or reconfigure the array.
985 * To do this we raise a 'barrier'.
986 * The 'barrier' is a counter that can be raised multiple times
987 * to count how many activities are happening which preclude
989 * We can only raise the barrier if there is no pending IO.
990 * i.e. if nr_pending == 0.
991 * We choose only to raise the barrier if no-one is waiting for the
992 * barrier to go down. This means that as soon as an IO request
993 * is ready, no other operations which require a barrier will start
994 * until the IO request has had a chance.
996 * So: regular IO calls 'wait_barrier'. When that returns there
997 * is no backgroup IO happening, It must arrange to call
998 * allow_barrier when it has finished its IO.
999 * backgroup IO calls must call raise_barrier. Once that returns
1000 * there is no normal IO happeing. It must arrange to call
1001 * lower_barrier when the particular background IO completes.
1004 static void raise_barrier(struct r10conf *conf, int force)
1006 BUG_ON(force && !conf->barrier);
1007 spin_lock_irq(&conf->resync_lock);
1009 /* Wait until no block IO is waiting (unless 'force') */
1010 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1013 /* block any new IO from starting */
1016 /* Now wait for all pending IO to complete */
1017 wait_event_lock_irq(conf->wait_barrier,
1018 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1021 spin_unlock_irq(&conf->resync_lock);
1024 static void lower_barrier(struct r10conf *conf)
1026 unsigned long flags;
1027 spin_lock_irqsave(&conf->resync_lock, flags);
1029 spin_unlock_irqrestore(&conf->resync_lock, flags);
1030 wake_up(&conf->wait_barrier);
1033 static void wait_barrier(struct r10conf *conf)
1035 spin_lock_irq(&conf->resync_lock);
1036 if (conf->barrier) {
1038 /* Wait for the barrier to drop.
1039 * However if there are already pending
1040 * requests (preventing the barrier from
1041 * rising completely), and the
1042 * pre-process bio queue isn't empty,
1043 * then don't wait, as we need to empty
1044 * that queue to get the nr_pending
1047 wait_event_lock_irq(conf->wait_barrier,
1049 (conf->nr_pending &&
1050 current->bio_list &&
1051 !bio_list_empty(current->bio_list)),
1056 spin_unlock_irq(&conf->resync_lock);
1059 static void allow_barrier(struct r10conf *conf)
1061 unsigned long flags;
1062 spin_lock_irqsave(&conf->resync_lock, flags);
1064 spin_unlock_irqrestore(&conf->resync_lock, flags);
1065 wake_up(&conf->wait_barrier);
1068 static void freeze_array(struct r10conf *conf)
1070 /* stop syncio and normal IO and wait for everything to
1072 * We increment barrier and nr_waiting, and then
1073 * wait until nr_pending match nr_queued+1
1074 * This is called in the context of one normal IO request
1075 * that has failed. Thus any sync request that might be pending
1076 * will be blocked by nr_pending, and we need to wait for
1077 * pending IO requests to complete or be queued for re-try.
1078 * Thus the number queued (nr_queued) plus this request (1)
1079 * must match the number of pending IOs (nr_pending) before
1082 spin_lock_irq(&conf->resync_lock);
1085 wait_event_lock_irq_cmd(conf->wait_barrier,
1086 conf->nr_pending == conf->nr_queued+1,
1088 flush_pending_writes(conf));
1090 spin_unlock_irq(&conf->resync_lock);
1093 static void unfreeze_array(struct r10conf *conf)
1095 /* reverse the effect of the freeze */
1096 spin_lock_irq(&conf->resync_lock);
1099 wake_up(&conf->wait_barrier);
1100 spin_unlock_irq(&conf->resync_lock);
1103 static sector_t choose_data_offset(struct r10bio *r10_bio,
1104 struct md_rdev *rdev)
1106 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1107 test_bit(R10BIO_Previous, &r10_bio->state))
1108 return rdev->data_offset;
1110 return rdev->new_data_offset;
1113 struct raid10_plug_cb {
1114 struct blk_plug_cb cb;
1115 struct bio_list pending;
1119 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1121 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1123 struct mddev *mddev = plug->cb.data;
1124 struct r10conf *conf = mddev->private;
1127 if (from_schedule || current->bio_list) {
1128 spin_lock_irq(&conf->device_lock);
1129 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1130 conf->pending_count += plug->pending_cnt;
1131 spin_unlock_irq(&conf->device_lock);
1132 wake_up(&conf->wait_barrier);
1133 md_wakeup_thread(mddev->thread);
1138 /* we aren't scheduling, so we can do the write-out directly. */
1139 bio = bio_list_get(&plug->pending);
1140 bitmap_unplug(mddev->bitmap);
1141 wake_up(&conf->wait_barrier);
1143 while (bio) { /* submit pending writes */
1144 struct bio *next = bio->bi_next;
1145 bio->bi_next = NULL;
1146 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1147 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1148 /* Just ignore it */
1151 generic_make_request(bio);
1157 static void make_request(struct mddev *mddev, struct bio * bio)
1159 struct r10conf *conf = mddev->private;
1160 struct r10bio *r10_bio;
1161 struct bio *read_bio;
1163 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1164 int chunk_sects = chunk_mask + 1;
1165 const int rw = bio_data_dir(bio);
1166 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1167 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1168 const unsigned long do_discard = (bio->bi_rw
1169 & (REQ_DISCARD | REQ_SECURE));
1170 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1171 unsigned long flags;
1172 struct md_rdev *blocked_rdev;
1173 struct blk_plug_cb *cb;
1174 struct raid10_plug_cb *plug = NULL;
1175 int sectors_handled;
1179 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1180 md_flush_request(mddev, bio);
1184 /* If this request crosses a chunk boundary, we need to
1185 * split it. This will only happen for 1 PAGE (or less) requests.
1187 if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1189 && (conf->geo.near_copies < conf->geo.raid_disks
1190 || conf->prev.near_copies < conf->prev.raid_disks))) {
1191 struct bio_pair *bp;
1192 /* Sanity check -- queue functions should prevent this happening */
1193 if ((bio->bi_vcnt != 1 && bio->bi_vcnt != 0) ||
1196 /* This is a one page bio that upper layers
1197 * refuse to split for us, so we need to split it.
1200 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1202 /* Each of these 'make_request' calls will call 'wait_barrier'.
1203 * If the first succeeds but the second blocks due to the resync
1204 * thread raising the barrier, we will deadlock because the
1205 * IO to the underlying device will be queued in generic_make_request
1206 * and will never complete, so will never reduce nr_pending.
1207 * So increment nr_waiting here so no new raise_barriers will
1208 * succeed, and so the second wait_barrier cannot block.
1210 spin_lock_irq(&conf->resync_lock);
1212 spin_unlock_irq(&conf->resync_lock);
1214 make_request(mddev, &bp->bio1);
1215 make_request(mddev, &bp->bio2);
1217 spin_lock_irq(&conf->resync_lock);
1219 wake_up(&conf->wait_barrier);
1220 spin_unlock_irq(&conf->resync_lock);
1222 bio_pair_release(bp);
1225 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1226 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1227 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1233 md_write_start(mddev, bio);
1236 * Register the new request and wait if the reconstruction
1237 * thread has put up a bar for new requests.
1238 * Continue immediately if no resync is active currently.
1242 sectors = bio->bi_size >> 9;
1243 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1244 bio->bi_sector < conf->reshape_progress &&
1245 bio->bi_sector + sectors > conf->reshape_progress) {
1246 /* IO spans the reshape position. Need to wait for
1249 allow_barrier(conf);
1250 wait_event(conf->wait_barrier,
1251 conf->reshape_progress <= bio->bi_sector ||
1252 conf->reshape_progress >= bio->bi_sector + sectors);
1255 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1256 bio_data_dir(bio) == WRITE &&
1257 (mddev->reshape_backwards
1258 ? (bio->bi_sector < conf->reshape_safe &&
1259 bio->bi_sector + sectors > conf->reshape_progress)
1260 : (bio->bi_sector + sectors > conf->reshape_safe &&
1261 bio->bi_sector < conf->reshape_progress))) {
1262 /* Need to update reshape_position in metadata */
1263 mddev->reshape_position = conf->reshape_progress;
1264 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1265 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1266 md_wakeup_thread(mddev->thread);
1267 wait_event(mddev->sb_wait,
1268 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1270 conf->reshape_safe = mddev->reshape_position;
1273 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1275 r10_bio->master_bio = bio;
1276 r10_bio->sectors = sectors;
1278 r10_bio->mddev = mddev;
1279 r10_bio->sector = bio->bi_sector;
1282 /* We might need to issue multiple reads to different
1283 * devices if there are bad blocks around, so we keep
1284 * track of the number of reads in bio->bi_phys_segments.
1285 * If this is 0, there is only one r10_bio and no locking
1286 * will be needed when the request completes. If it is
1287 * non-zero, then it is the number of not-completed requests.
1289 bio->bi_phys_segments = 0;
1290 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1294 * read balancing logic:
1296 struct md_rdev *rdev;
1300 rdev = read_balance(conf, r10_bio, &max_sectors);
1302 raid_end_bio_io(r10_bio);
1305 slot = r10_bio->read_slot;
1307 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1308 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1311 r10_bio->devs[slot].bio = read_bio;
1312 r10_bio->devs[slot].rdev = rdev;
1314 read_bio->bi_sector = r10_bio->devs[slot].addr +
1315 choose_data_offset(r10_bio, rdev);
1316 read_bio->bi_bdev = rdev->bdev;
1317 read_bio->bi_end_io = raid10_end_read_request;
1318 read_bio->bi_rw = READ | do_sync;
1319 read_bio->bi_private = r10_bio;
1321 if (max_sectors < r10_bio->sectors) {
1322 /* Could not read all from this device, so we will
1323 * need another r10_bio.
1325 sectors_handled = (r10_bio->sectors + max_sectors
1327 r10_bio->sectors = max_sectors;
1328 spin_lock_irq(&conf->device_lock);
1329 if (bio->bi_phys_segments == 0)
1330 bio->bi_phys_segments = 2;
1332 bio->bi_phys_segments++;
1333 spin_unlock(&conf->device_lock);
1334 /* Cannot call generic_make_request directly
1335 * as that will be queued in __generic_make_request
1336 * and subsequent mempool_alloc might block
1337 * waiting for it. so hand bio over to raid10d.
1339 reschedule_retry(r10_bio);
1341 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1343 r10_bio->master_bio = bio;
1344 r10_bio->sectors = ((bio->bi_size >> 9)
1347 r10_bio->mddev = mddev;
1348 r10_bio->sector = bio->bi_sector + sectors_handled;
1351 generic_make_request(read_bio);
1358 if (conf->pending_count >= max_queued_requests) {
1359 md_wakeup_thread(mddev->thread);
1360 wait_event(conf->wait_barrier,
1361 conf->pending_count < max_queued_requests);
1363 /* first select target devices under rcu_lock and
1364 * inc refcount on their rdev. Record them by setting
1366 * If there are known/acknowledged bad blocks on any device
1367 * on which we have seen a write error, we want to avoid
1368 * writing to those blocks. This potentially requires several
1369 * writes to write around the bad blocks. Each set of writes
1370 * gets its own r10_bio with a set of bios attached. The number
1371 * of r10_bios is recored in bio->bi_phys_segments just as with
1375 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1376 raid10_find_phys(conf, r10_bio);
1378 blocked_rdev = NULL;
1380 max_sectors = r10_bio->sectors;
1382 for (i = 0; i < conf->copies; i++) {
1383 int d = r10_bio->devs[i].devnum;
1384 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1385 struct md_rdev *rrdev = rcu_dereference(
1386 conf->mirrors[d].replacement);
1389 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1390 atomic_inc(&rdev->nr_pending);
1391 blocked_rdev = rdev;
1394 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1395 atomic_inc(&rrdev->nr_pending);
1396 blocked_rdev = rrdev;
1399 if (rdev && (test_bit(Faulty, &rdev->flags)
1400 || test_bit(Unmerged, &rdev->flags)))
1402 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1403 || test_bit(Unmerged, &rrdev->flags)))
1406 r10_bio->devs[i].bio = NULL;
1407 r10_bio->devs[i].repl_bio = NULL;
1409 if (!rdev && !rrdev) {
1410 set_bit(R10BIO_Degraded, &r10_bio->state);
1413 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1415 sector_t dev_sector = r10_bio->devs[i].addr;
1419 is_bad = is_badblock(rdev, dev_sector,
1421 &first_bad, &bad_sectors);
1423 /* Mustn't write here until the bad block
1426 atomic_inc(&rdev->nr_pending);
1427 set_bit(BlockedBadBlocks, &rdev->flags);
1428 blocked_rdev = rdev;
1431 if (is_bad && first_bad <= dev_sector) {
1432 /* Cannot write here at all */
1433 bad_sectors -= (dev_sector - first_bad);
1434 if (bad_sectors < max_sectors)
1435 /* Mustn't write more than bad_sectors
1436 * to other devices yet
1438 max_sectors = bad_sectors;
1439 /* We don't set R10BIO_Degraded as that
1440 * only applies if the disk is missing,
1441 * so it might be re-added, and we want to
1442 * know to recover this chunk.
1443 * In this case the device is here, and the
1444 * fact that this chunk is not in-sync is
1445 * recorded in the bad block log.
1450 int good_sectors = first_bad - dev_sector;
1451 if (good_sectors < max_sectors)
1452 max_sectors = good_sectors;
1456 r10_bio->devs[i].bio = bio;
1457 atomic_inc(&rdev->nr_pending);
1460 r10_bio->devs[i].repl_bio = bio;
1461 atomic_inc(&rrdev->nr_pending);
1466 if (unlikely(blocked_rdev)) {
1467 /* Have to wait for this device to get unblocked, then retry */
1471 for (j = 0; j < i; j++) {
1472 if (r10_bio->devs[j].bio) {
1473 d = r10_bio->devs[j].devnum;
1474 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1476 if (r10_bio->devs[j].repl_bio) {
1477 struct md_rdev *rdev;
1478 d = r10_bio->devs[j].devnum;
1479 rdev = conf->mirrors[d].replacement;
1481 /* Race with remove_disk */
1483 rdev = conf->mirrors[d].rdev;
1485 rdev_dec_pending(rdev, mddev);
1488 allow_barrier(conf);
1489 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1494 if (max_sectors < r10_bio->sectors) {
1495 /* We are splitting this into multiple parts, so
1496 * we need to prepare for allocating another r10_bio.
1498 r10_bio->sectors = max_sectors;
1499 spin_lock_irq(&conf->device_lock);
1500 if (bio->bi_phys_segments == 0)
1501 bio->bi_phys_segments = 2;
1503 bio->bi_phys_segments++;
1504 spin_unlock_irq(&conf->device_lock);
1506 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1508 atomic_set(&r10_bio->remaining, 1);
1509 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1511 for (i = 0; i < conf->copies; i++) {
1513 int d = r10_bio->devs[i].devnum;
1514 if (r10_bio->devs[i].bio) {
1515 struct md_rdev *rdev = conf->mirrors[d].rdev;
1516 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1517 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1519 r10_bio->devs[i].bio = mbio;
1521 mbio->bi_sector = (r10_bio->devs[i].addr+
1522 choose_data_offset(r10_bio,
1524 mbio->bi_bdev = rdev->bdev;
1525 mbio->bi_end_io = raid10_end_write_request;
1527 WRITE | do_sync | do_fua | do_discard | do_same;
1528 mbio->bi_private = r10_bio;
1530 atomic_inc(&r10_bio->remaining);
1532 cb = blk_check_plugged(raid10_unplug, mddev,
1535 plug = container_of(cb, struct raid10_plug_cb,
1539 spin_lock_irqsave(&conf->device_lock, flags);
1541 bio_list_add(&plug->pending, mbio);
1542 plug->pending_cnt++;
1544 bio_list_add(&conf->pending_bio_list, mbio);
1545 conf->pending_count++;
1547 spin_unlock_irqrestore(&conf->device_lock, flags);
1549 md_wakeup_thread(mddev->thread);
1552 if (r10_bio->devs[i].repl_bio) {
1553 struct md_rdev *rdev = conf->mirrors[d].replacement;
1555 /* Replacement just got moved to main 'rdev' */
1557 rdev = conf->mirrors[d].rdev;
1559 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1560 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1562 r10_bio->devs[i].repl_bio = mbio;
1564 mbio->bi_sector = (r10_bio->devs[i].addr +
1567 mbio->bi_bdev = rdev->bdev;
1568 mbio->bi_end_io = raid10_end_write_request;
1570 WRITE | do_sync | do_fua | do_discard | do_same;
1571 mbio->bi_private = r10_bio;
1573 atomic_inc(&r10_bio->remaining);
1574 spin_lock_irqsave(&conf->device_lock, flags);
1575 bio_list_add(&conf->pending_bio_list, mbio);
1576 conf->pending_count++;
1577 spin_unlock_irqrestore(&conf->device_lock, flags);
1578 if (!mddev_check_plugged(mddev))
1579 md_wakeup_thread(mddev->thread);
1583 /* Don't remove the bias on 'remaining' (one_write_done) until
1584 * after checking if we need to go around again.
1587 if (sectors_handled < (bio->bi_size >> 9)) {
1588 one_write_done(r10_bio);
1589 /* We need another r10_bio. It has already been counted
1590 * in bio->bi_phys_segments.
1592 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1594 r10_bio->master_bio = bio;
1595 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1597 r10_bio->mddev = mddev;
1598 r10_bio->sector = bio->bi_sector + sectors_handled;
1602 one_write_done(r10_bio);
1604 /* In case raid10d snuck in to freeze_array */
1605 wake_up(&conf->wait_barrier);
1608 static void status(struct seq_file *seq, struct mddev *mddev)
1610 struct r10conf *conf = mddev->private;
1613 if (conf->geo.near_copies < conf->geo.raid_disks)
1614 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1615 if (conf->geo.near_copies > 1)
1616 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1617 if (conf->geo.far_copies > 1) {
1618 if (conf->geo.far_offset)
1619 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1621 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1623 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1624 conf->geo.raid_disks - mddev->degraded);
1625 for (i = 0; i < conf->geo.raid_disks; i++)
1626 seq_printf(seq, "%s",
1627 conf->mirrors[i].rdev &&
1628 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1629 seq_printf(seq, "]");
1632 /* check if there are enough drives for
1633 * every block to appear on atleast one.
1634 * Don't consider the device numbered 'ignore'
1635 * as we might be about to remove it.
1637 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1642 int n = conf->copies;
1646 if (conf->mirrors[this].rdev &&
1649 this = (this+1) % geo->raid_disks;
1653 first = (first + geo->near_copies) % geo->raid_disks;
1654 } while (first != 0);
1658 static int enough(struct r10conf *conf, int ignore)
1660 return _enough(conf, &conf->geo, ignore) &&
1661 _enough(conf, &conf->prev, ignore);
1664 static void error(struct mddev *mddev, struct md_rdev *rdev)
1666 char b[BDEVNAME_SIZE];
1667 struct r10conf *conf = mddev->private;
1670 * If it is not operational, then we have already marked it as dead
1671 * else if it is the last working disks, ignore the error, let the
1672 * next level up know.
1673 * else mark the drive as failed
1675 if (test_bit(In_sync, &rdev->flags)
1676 && !enough(conf, rdev->raid_disk))
1678 * Don't fail the drive, just return an IO error.
1681 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1682 unsigned long flags;
1683 spin_lock_irqsave(&conf->device_lock, flags);
1685 spin_unlock_irqrestore(&conf->device_lock, flags);
1687 * if recovery is running, make sure it aborts.
1689 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1691 set_bit(Blocked, &rdev->flags);
1692 set_bit(Faulty, &rdev->flags);
1693 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1695 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1696 "md/raid10:%s: Operation continuing on %d devices.\n",
1697 mdname(mddev), bdevname(rdev->bdev, b),
1698 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1701 static void print_conf(struct r10conf *conf)
1704 struct raid10_info *tmp;
1706 printk(KERN_DEBUG "RAID10 conf printout:\n");
1708 printk(KERN_DEBUG "(!conf)\n");
1711 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1712 conf->geo.raid_disks);
1714 for (i = 0; i < conf->geo.raid_disks; i++) {
1715 char b[BDEVNAME_SIZE];
1716 tmp = conf->mirrors + i;
1718 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1719 i, !test_bit(In_sync, &tmp->rdev->flags),
1720 !test_bit(Faulty, &tmp->rdev->flags),
1721 bdevname(tmp->rdev->bdev,b));
1725 static void close_sync(struct r10conf *conf)
1728 allow_barrier(conf);
1730 mempool_destroy(conf->r10buf_pool);
1731 conf->r10buf_pool = NULL;
1734 static int raid10_spare_active(struct mddev *mddev)
1737 struct r10conf *conf = mddev->private;
1738 struct raid10_info *tmp;
1740 unsigned long flags;
1743 * Find all non-in_sync disks within the RAID10 configuration
1744 * and mark them in_sync
1746 for (i = 0; i < conf->geo.raid_disks; i++) {
1747 tmp = conf->mirrors + i;
1748 if (tmp->replacement
1749 && tmp->replacement->recovery_offset == MaxSector
1750 && !test_bit(Faulty, &tmp->replacement->flags)
1751 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1752 /* Replacement has just become active */
1754 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1757 /* Replaced device not technically faulty,
1758 * but we need to be sure it gets removed
1759 * and never re-added.
1761 set_bit(Faulty, &tmp->rdev->flags);
1762 sysfs_notify_dirent_safe(
1763 tmp->rdev->sysfs_state);
1765 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1766 } else if (tmp->rdev
1767 && !test_bit(Faulty, &tmp->rdev->flags)
1768 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1770 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1773 spin_lock_irqsave(&conf->device_lock, flags);
1774 mddev->degraded -= count;
1775 spin_unlock_irqrestore(&conf->device_lock, flags);
1782 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1784 struct r10conf *conf = mddev->private;
1788 int last = conf->geo.raid_disks - 1;
1789 struct request_queue *q = bdev_get_queue(rdev->bdev);
1791 if (mddev->recovery_cp < MaxSector)
1792 /* only hot-add to in-sync arrays, as recovery is
1793 * very different from resync
1796 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1799 if (rdev->raid_disk >= 0)
1800 first = last = rdev->raid_disk;
1802 if (q->merge_bvec_fn) {
1803 set_bit(Unmerged, &rdev->flags);
1804 mddev->merge_check_needed = 1;
1807 if (rdev->saved_raid_disk >= first &&
1808 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1809 mirror = rdev->saved_raid_disk;
1812 for ( ; mirror <= last ; mirror++) {
1813 struct raid10_info *p = &conf->mirrors[mirror];
1814 if (p->recovery_disabled == mddev->recovery_disabled)
1817 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1818 p->replacement != NULL)
1820 clear_bit(In_sync, &rdev->flags);
1821 set_bit(Replacement, &rdev->flags);
1822 rdev->raid_disk = mirror;
1824 disk_stack_limits(mddev->gendisk, rdev->bdev,
1825 rdev->data_offset << 9);
1827 rcu_assign_pointer(p->replacement, rdev);
1831 disk_stack_limits(mddev->gendisk, rdev->bdev,
1832 rdev->data_offset << 9);
1834 p->head_position = 0;
1835 p->recovery_disabled = mddev->recovery_disabled - 1;
1836 rdev->raid_disk = mirror;
1838 if (rdev->saved_raid_disk != mirror)
1840 rcu_assign_pointer(p->rdev, rdev);
1843 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1844 /* Some requests might not have seen this new
1845 * merge_bvec_fn. We must wait for them to complete
1846 * before merging the device fully.
1847 * First we make sure any code which has tested
1848 * our function has submitted the request, then
1849 * we wait for all outstanding requests to complete.
1851 synchronize_sched();
1852 raise_barrier(conf, 0);
1853 lower_barrier(conf);
1854 clear_bit(Unmerged, &rdev->flags);
1856 md_integrity_add_rdev(rdev, mddev);
1857 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1858 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1864 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1866 struct r10conf *conf = mddev->private;
1868 int number = rdev->raid_disk;
1869 struct md_rdev **rdevp;
1870 struct raid10_info *p = conf->mirrors + number;
1873 if (rdev == p->rdev)
1875 else if (rdev == p->replacement)
1876 rdevp = &p->replacement;
1880 if (test_bit(In_sync, &rdev->flags) ||
1881 atomic_read(&rdev->nr_pending)) {
1885 /* Only remove faulty devices if recovery
1888 if (!test_bit(Faulty, &rdev->flags) &&
1889 mddev->recovery_disabled != p->recovery_disabled &&
1890 (!p->replacement || p->replacement == rdev) &&
1891 number < conf->geo.raid_disks &&
1898 if (atomic_read(&rdev->nr_pending)) {
1899 /* lost the race, try later */
1903 } else if (p->replacement) {
1904 /* We must have just cleared 'rdev' */
1905 p->rdev = p->replacement;
1906 clear_bit(Replacement, &p->replacement->flags);
1907 smp_mb(); /* Make sure other CPUs may see both as identical
1908 * but will never see neither -- if they are careful.
1910 p->replacement = NULL;
1911 clear_bit(WantReplacement, &rdev->flags);
1913 /* We might have just remove the Replacement as faulty
1914 * Clear the flag just in case
1916 clear_bit(WantReplacement, &rdev->flags);
1918 err = md_integrity_register(mddev);
1927 static void end_sync_read(struct bio *bio, int error)
1929 struct r10bio *r10_bio = bio->bi_private;
1930 struct r10conf *conf = r10_bio->mddev->private;
1933 if (bio == r10_bio->master_bio) {
1934 /* this is a reshape read */
1935 d = r10_bio->read_slot; /* really the read dev */
1937 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1939 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1940 set_bit(R10BIO_Uptodate, &r10_bio->state);
1942 /* The write handler will notice the lack of
1943 * R10BIO_Uptodate and record any errors etc
1945 atomic_add(r10_bio->sectors,
1946 &conf->mirrors[d].rdev->corrected_errors);
1948 /* for reconstruct, we always reschedule after a read.
1949 * for resync, only after all reads
1951 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1952 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1953 atomic_dec_and_test(&r10_bio->remaining)) {
1954 /* we have read all the blocks,
1955 * do the comparison in process context in raid10d
1957 reschedule_retry(r10_bio);
1961 static void end_sync_request(struct r10bio *r10_bio)
1963 struct mddev *mddev = r10_bio->mddev;
1965 while (atomic_dec_and_test(&r10_bio->remaining)) {
1966 if (r10_bio->master_bio == NULL) {
1967 /* the primary of several recovery bios */
1968 sector_t s = r10_bio->sectors;
1969 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1970 test_bit(R10BIO_WriteError, &r10_bio->state))
1971 reschedule_retry(r10_bio);
1974 md_done_sync(mddev, s, 1);
1977 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1978 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1979 test_bit(R10BIO_WriteError, &r10_bio->state))
1980 reschedule_retry(r10_bio);
1988 static void end_sync_write(struct bio *bio, int error)
1990 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1991 struct r10bio *r10_bio = bio->bi_private;
1992 struct mddev *mddev = r10_bio->mddev;
1993 struct r10conf *conf = mddev->private;
1999 struct md_rdev *rdev = NULL;
2001 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2003 rdev = conf->mirrors[d].replacement;
2005 rdev = conf->mirrors[d].rdev;
2009 md_error(mddev, rdev);
2011 set_bit(WriteErrorSeen, &rdev->flags);
2012 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2013 set_bit(MD_RECOVERY_NEEDED,
2014 &rdev->mddev->recovery);
2015 set_bit(R10BIO_WriteError, &r10_bio->state);
2017 } else if (is_badblock(rdev,
2018 r10_bio->devs[slot].addr,
2020 &first_bad, &bad_sectors))
2021 set_bit(R10BIO_MadeGood, &r10_bio->state);
2023 rdev_dec_pending(rdev, mddev);
2025 end_sync_request(r10_bio);
2029 * Note: sync and recover and handled very differently for raid10
2030 * This code is for resync.
2031 * For resync, we read through virtual addresses and read all blocks.
2032 * If there is any error, we schedule a write. The lowest numbered
2033 * drive is authoritative.
2034 * However requests come for physical address, so we need to map.
2035 * For every physical address there are raid_disks/copies virtual addresses,
2036 * which is always are least one, but is not necessarly an integer.
2037 * This means that a physical address can span multiple chunks, so we may
2038 * have to submit multiple io requests for a single sync request.
2041 * We check if all blocks are in-sync and only write to blocks that
2044 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2046 struct r10conf *conf = mddev->private;
2048 struct bio *tbio, *fbio;
2051 atomic_set(&r10_bio->remaining, 1);
2053 /* find the first device with a block */
2054 for (i=0; i<conf->copies; i++)
2055 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2058 if (i == conf->copies)
2062 fbio = r10_bio->devs[i].bio;
2064 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2065 /* now find blocks with errors */
2066 for (i=0 ; i < conf->copies ; i++) {
2069 tbio = r10_bio->devs[i].bio;
2071 if (tbio->bi_end_io != end_sync_read)
2075 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2076 /* We know that the bi_io_vec layout is the same for
2077 * both 'first' and 'i', so we just compare them.
2078 * All vec entries are PAGE_SIZE;
2080 for (j = 0; j < vcnt; j++)
2081 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2082 page_address(tbio->bi_io_vec[j].bv_page),
2083 fbio->bi_io_vec[j].bv_len))
2087 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2088 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2089 /* Don't fix anything. */
2092 /* Ok, we need to write this bio, either to correct an
2093 * inconsistency or to correct an unreadable block.
2094 * First we need to fixup bv_offset, bv_len and
2095 * bi_vecs, as the read request might have corrupted these
2097 tbio->bi_vcnt = vcnt;
2098 tbio->bi_size = r10_bio->sectors << 9;
2100 tbio->bi_phys_segments = 0;
2101 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
2102 tbio->bi_flags |= 1 << BIO_UPTODATE;
2103 tbio->bi_next = NULL;
2104 tbio->bi_rw = WRITE;
2105 tbio->bi_private = r10_bio;
2106 tbio->bi_sector = r10_bio->devs[i].addr;
2108 for (j=0; j < vcnt ; j++) {
2109 tbio->bi_io_vec[j].bv_offset = 0;
2110 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2112 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2113 page_address(fbio->bi_io_vec[j].bv_page),
2116 tbio->bi_end_io = end_sync_write;
2118 d = r10_bio->devs[i].devnum;
2119 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2120 atomic_inc(&r10_bio->remaining);
2121 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
2123 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2124 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2125 generic_make_request(tbio);
2128 /* Now write out to any replacement devices
2131 for (i = 0; i < conf->copies; i++) {
2134 tbio = r10_bio->devs[i].repl_bio;
2135 if (!tbio || !tbio->bi_end_io)
2137 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2138 && r10_bio->devs[i].bio != fbio)
2139 for (j = 0; j < vcnt; j++)
2140 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2141 page_address(fbio->bi_io_vec[j].bv_page),
2143 d = r10_bio->devs[i].devnum;
2144 atomic_inc(&r10_bio->remaining);
2145 md_sync_acct(conf->mirrors[d].replacement->bdev,
2146 tbio->bi_size >> 9);
2147 generic_make_request(tbio);
2151 if (atomic_dec_and_test(&r10_bio->remaining)) {
2152 md_done_sync(mddev, r10_bio->sectors, 1);
2158 * Now for the recovery code.
2159 * Recovery happens across physical sectors.
2160 * We recover all non-is_sync drives by finding the virtual address of
2161 * each, and then choose a working drive that also has that virt address.
2162 * There is a separate r10_bio for each non-in_sync drive.
2163 * Only the first two slots are in use. The first for reading,
2164 * The second for writing.
2167 static void fix_recovery_read_error(struct r10bio *r10_bio)
2169 /* We got a read error during recovery.
2170 * We repeat the read in smaller page-sized sections.
2171 * If a read succeeds, write it to the new device or record
2172 * a bad block if we cannot.
2173 * If a read fails, record a bad block on both old and
2176 struct mddev *mddev = r10_bio->mddev;
2177 struct r10conf *conf = mddev->private;
2178 struct bio *bio = r10_bio->devs[0].bio;
2180 int sectors = r10_bio->sectors;
2182 int dr = r10_bio->devs[0].devnum;
2183 int dw = r10_bio->devs[1].devnum;
2187 struct md_rdev *rdev;
2191 if (s > (PAGE_SIZE>>9))
2194 rdev = conf->mirrors[dr].rdev;
2195 addr = r10_bio->devs[0].addr + sect,
2196 ok = sync_page_io(rdev,
2199 bio->bi_io_vec[idx].bv_page,
2202 rdev = conf->mirrors[dw].rdev;
2203 addr = r10_bio->devs[1].addr + sect;
2204 ok = sync_page_io(rdev,
2207 bio->bi_io_vec[idx].bv_page,
2210 set_bit(WriteErrorSeen, &rdev->flags);
2211 if (!test_and_set_bit(WantReplacement,
2213 set_bit(MD_RECOVERY_NEEDED,
2214 &rdev->mddev->recovery);
2218 /* We don't worry if we cannot set a bad block -
2219 * it really is bad so there is no loss in not
2222 rdev_set_badblocks(rdev, addr, s, 0);
2224 if (rdev != conf->mirrors[dw].rdev) {
2225 /* need bad block on destination too */
2226 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2227 addr = r10_bio->devs[1].addr + sect;
2228 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2230 /* just abort the recovery */
2232 "md/raid10:%s: recovery aborted"
2233 " due to read error\n",
2236 conf->mirrors[dw].recovery_disabled
2237 = mddev->recovery_disabled;
2238 set_bit(MD_RECOVERY_INTR,
2251 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2253 struct r10conf *conf = mddev->private;
2255 struct bio *wbio, *wbio2;
2257 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2258 fix_recovery_read_error(r10_bio);
2259 end_sync_request(r10_bio);
2264 * share the pages with the first bio
2265 * and submit the write request
2267 d = r10_bio->devs[1].devnum;
2268 wbio = r10_bio->devs[1].bio;
2269 wbio2 = r10_bio->devs[1].repl_bio;
2270 if (wbio->bi_end_io) {
2271 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2272 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2273 generic_make_request(wbio);
2275 if (wbio2 && wbio2->bi_end_io) {
2276 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2277 md_sync_acct(conf->mirrors[d].replacement->bdev,
2278 wbio2->bi_size >> 9);
2279 generic_make_request(wbio2);
2285 * Used by fix_read_error() to decay the per rdev read_errors.
2286 * We halve the read error count for every hour that has elapsed
2287 * since the last recorded read error.
2290 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2292 struct timespec cur_time_mon;
2293 unsigned long hours_since_last;
2294 unsigned int read_errors = atomic_read(&rdev->read_errors);
2296 ktime_get_ts(&cur_time_mon);
2298 if (rdev->last_read_error.tv_sec == 0 &&
2299 rdev->last_read_error.tv_nsec == 0) {
2300 /* first time we've seen a read error */
2301 rdev->last_read_error = cur_time_mon;
2305 hours_since_last = (cur_time_mon.tv_sec -
2306 rdev->last_read_error.tv_sec) / 3600;
2308 rdev->last_read_error = cur_time_mon;
2311 * if hours_since_last is > the number of bits in read_errors
2312 * just set read errors to 0. We do this to avoid
2313 * overflowing the shift of read_errors by hours_since_last.
2315 if (hours_since_last >= 8 * sizeof(read_errors))
2316 atomic_set(&rdev->read_errors, 0);
2318 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2321 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2322 int sectors, struct page *page, int rw)
2327 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2328 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2330 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2334 set_bit(WriteErrorSeen, &rdev->flags);
2335 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2336 set_bit(MD_RECOVERY_NEEDED,
2337 &rdev->mddev->recovery);
2339 /* need to record an error - either for the block or the device */
2340 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2341 md_error(rdev->mddev, rdev);
2346 * This is a kernel thread which:
2348 * 1. Retries failed read operations on working mirrors.
2349 * 2. Updates the raid superblock when problems encounter.
2350 * 3. Performs writes following reads for array synchronising.
2353 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2355 int sect = 0; /* Offset from r10_bio->sector */
2356 int sectors = r10_bio->sectors;
2357 struct md_rdev*rdev;
2358 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2359 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2361 /* still own a reference to this rdev, so it cannot
2362 * have been cleared recently.
2364 rdev = conf->mirrors[d].rdev;
2366 if (test_bit(Faulty, &rdev->flags))
2367 /* drive has already been failed, just ignore any
2368 more fix_read_error() attempts */
2371 check_decay_read_errors(mddev, rdev);
2372 atomic_inc(&rdev->read_errors);
2373 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2374 char b[BDEVNAME_SIZE];
2375 bdevname(rdev->bdev, b);
2378 "md/raid10:%s: %s: Raid device exceeded "
2379 "read_error threshold [cur %d:max %d]\n",
2381 atomic_read(&rdev->read_errors), max_read_errors);
2383 "md/raid10:%s: %s: Failing raid device\n",
2385 md_error(mddev, conf->mirrors[d].rdev);
2386 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2392 int sl = r10_bio->read_slot;
2396 if (s > (PAGE_SIZE>>9))
2404 d = r10_bio->devs[sl].devnum;
2405 rdev = rcu_dereference(conf->mirrors[d].rdev);
2407 !test_bit(Unmerged, &rdev->flags) &&
2408 test_bit(In_sync, &rdev->flags) &&
2409 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2410 &first_bad, &bad_sectors) == 0) {
2411 atomic_inc(&rdev->nr_pending);
2413 success = sync_page_io(rdev,
2414 r10_bio->devs[sl].addr +
2417 conf->tmppage, READ, false);
2418 rdev_dec_pending(rdev, mddev);
2424 if (sl == conf->copies)
2426 } while (!success && sl != r10_bio->read_slot);
2430 /* Cannot read from anywhere, just mark the block
2431 * as bad on the first device to discourage future
2434 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2435 rdev = conf->mirrors[dn].rdev;
2437 if (!rdev_set_badblocks(
2439 r10_bio->devs[r10_bio->read_slot].addr
2442 md_error(mddev, rdev);
2443 r10_bio->devs[r10_bio->read_slot].bio
2450 /* write it back and re-read */
2452 while (sl != r10_bio->read_slot) {
2453 char b[BDEVNAME_SIZE];
2458 d = r10_bio->devs[sl].devnum;
2459 rdev = rcu_dereference(conf->mirrors[d].rdev);
2461 test_bit(Unmerged, &rdev->flags) ||
2462 !test_bit(In_sync, &rdev->flags))
2465 atomic_inc(&rdev->nr_pending);
2467 if (r10_sync_page_io(rdev,
2468 r10_bio->devs[sl].addr +
2470 s, conf->tmppage, WRITE)
2472 /* Well, this device is dead */
2474 "md/raid10:%s: read correction "
2476 " (%d sectors at %llu on %s)\n",
2478 (unsigned long long)(
2480 choose_data_offset(r10_bio,
2482 bdevname(rdev->bdev, b));
2483 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2486 bdevname(rdev->bdev, b));
2488 rdev_dec_pending(rdev, mddev);
2492 while (sl != r10_bio->read_slot) {
2493 char b[BDEVNAME_SIZE];
2498 d = r10_bio->devs[sl].devnum;
2499 rdev = rcu_dereference(conf->mirrors[d].rdev);
2501 !test_bit(In_sync, &rdev->flags))
2504 atomic_inc(&rdev->nr_pending);
2506 switch (r10_sync_page_io(rdev,
2507 r10_bio->devs[sl].addr +
2512 /* Well, this device is dead */
2514 "md/raid10:%s: unable to read back "
2516 " (%d sectors at %llu on %s)\n",
2518 (unsigned long long)(
2520 choose_data_offset(r10_bio, rdev)),
2521 bdevname(rdev->bdev, b));
2522 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2525 bdevname(rdev->bdev, b));
2529 "md/raid10:%s: read error corrected"
2530 " (%d sectors at %llu on %s)\n",
2532 (unsigned long long)(
2534 choose_data_offset(r10_bio, rdev)),
2535 bdevname(rdev->bdev, b));
2536 atomic_add(s, &rdev->corrected_errors);
2539 rdev_dec_pending(rdev, mddev);
2549 static void bi_complete(struct bio *bio, int error)
2551 complete((struct completion *)bio->bi_private);
2554 static int submit_bio_wait(int rw, struct bio *bio)
2556 struct completion event;
2559 init_completion(&event);
2560 bio->bi_private = &event;
2561 bio->bi_end_io = bi_complete;
2562 submit_bio(rw, bio);
2563 wait_for_completion(&event);
2565 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2568 static int narrow_write_error(struct r10bio *r10_bio, int i)
2570 struct bio *bio = r10_bio->master_bio;
2571 struct mddev *mddev = r10_bio->mddev;
2572 struct r10conf *conf = mddev->private;
2573 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2574 /* bio has the data to be written to slot 'i' where
2575 * we just recently had a write error.
2576 * We repeatedly clone the bio and trim down to one block,
2577 * then try the write. Where the write fails we record
2579 * It is conceivable that the bio doesn't exactly align with
2580 * blocks. We must handle this.
2582 * We currently own a reference to the rdev.
2588 int sect_to_write = r10_bio->sectors;
2591 if (rdev->badblocks.shift < 0)
2594 block_sectors = 1 << rdev->badblocks.shift;
2595 sector = r10_bio->sector;
2596 sectors = ((r10_bio->sector + block_sectors)
2597 & ~(sector_t)(block_sectors - 1))
2600 while (sect_to_write) {
2602 if (sectors > sect_to_write)
2603 sectors = sect_to_write;
2604 /* Write at 'sector' for 'sectors' */
2605 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2606 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2607 wbio->bi_sector = (r10_bio->devs[i].addr+
2608 choose_data_offset(r10_bio, rdev) +
2609 (sector - r10_bio->sector));
2610 wbio->bi_bdev = rdev->bdev;
2611 if (submit_bio_wait(WRITE, wbio) == 0)
2613 ok = rdev_set_badblocks(rdev, sector,
2618 sect_to_write -= sectors;
2620 sectors = block_sectors;
2625 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2627 int slot = r10_bio->read_slot;
2629 struct r10conf *conf = mddev->private;
2630 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2631 char b[BDEVNAME_SIZE];
2632 unsigned long do_sync;
2635 /* we got a read error. Maybe the drive is bad. Maybe just
2636 * the block and we can fix it.
2637 * We freeze all other IO, and try reading the block from
2638 * other devices. When we find one, we re-write
2639 * and check it that fixes the read error.
2640 * This is all done synchronously while the array is
2643 bio = r10_bio->devs[slot].bio;
2644 bdevname(bio->bi_bdev, b);
2646 r10_bio->devs[slot].bio = NULL;
2648 if (mddev->ro == 0) {
2650 fix_read_error(conf, mddev, r10_bio);
2651 unfreeze_array(conf);
2653 r10_bio->devs[slot].bio = IO_BLOCKED;
2655 rdev_dec_pending(rdev, mddev);
2658 rdev = read_balance(conf, r10_bio, &max_sectors);
2660 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2661 " read error for block %llu\n",
2663 (unsigned long long)r10_bio->sector);
2664 raid_end_bio_io(r10_bio);
2668 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2669 slot = r10_bio->read_slot;
2672 "md/raid10:%s: %s: redirecting "
2673 "sector %llu to another mirror\n",
2675 bdevname(rdev->bdev, b),
2676 (unsigned long long)r10_bio->sector);
2677 bio = bio_clone_mddev(r10_bio->master_bio,
2680 r10_bio->sector - bio->bi_sector,
2682 r10_bio->devs[slot].bio = bio;
2683 r10_bio->devs[slot].rdev = rdev;
2684 bio->bi_sector = r10_bio->devs[slot].addr
2685 + choose_data_offset(r10_bio, rdev);
2686 bio->bi_bdev = rdev->bdev;
2687 bio->bi_rw = READ | do_sync;
2688 bio->bi_private = r10_bio;
2689 bio->bi_end_io = raid10_end_read_request;
2690 if (max_sectors < r10_bio->sectors) {
2691 /* Drat - have to split this up more */
2692 struct bio *mbio = r10_bio->master_bio;
2693 int sectors_handled =
2694 r10_bio->sector + max_sectors
2696 r10_bio->sectors = max_sectors;
2697 spin_lock_irq(&conf->device_lock);
2698 if (mbio->bi_phys_segments == 0)
2699 mbio->bi_phys_segments = 2;
2701 mbio->bi_phys_segments++;
2702 spin_unlock_irq(&conf->device_lock);
2703 generic_make_request(bio);
2705 r10_bio = mempool_alloc(conf->r10bio_pool,
2707 r10_bio->master_bio = mbio;
2708 r10_bio->sectors = (mbio->bi_size >> 9)
2711 set_bit(R10BIO_ReadError,
2713 r10_bio->mddev = mddev;
2714 r10_bio->sector = mbio->bi_sector
2719 generic_make_request(bio);
2722 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2724 /* Some sort of write request has finished and it
2725 * succeeded in writing where we thought there was a
2726 * bad block. So forget the bad block.
2727 * Or possibly if failed and we need to record
2731 struct md_rdev *rdev;
2733 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2734 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2735 for (m = 0; m < conf->copies; m++) {
2736 int dev = r10_bio->devs[m].devnum;
2737 rdev = conf->mirrors[dev].rdev;
2738 if (r10_bio->devs[m].bio == NULL)
2740 if (test_bit(BIO_UPTODATE,
2741 &r10_bio->devs[m].bio->bi_flags)) {
2742 rdev_clear_badblocks(
2744 r10_bio->devs[m].addr,
2745 r10_bio->sectors, 0);
2747 if (!rdev_set_badblocks(
2749 r10_bio->devs[m].addr,
2750 r10_bio->sectors, 0))
2751 md_error(conf->mddev, rdev);
2753 rdev = conf->mirrors[dev].replacement;
2754 if (r10_bio->devs[m].repl_bio == NULL)
2756 if (test_bit(BIO_UPTODATE,
2757 &r10_bio->devs[m].repl_bio->bi_flags)) {
2758 rdev_clear_badblocks(
2760 r10_bio->devs[m].addr,
2761 r10_bio->sectors, 0);
2763 if (!rdev_set_badblocks(
2765 r10_bio->devs[m].addr,
2766 r10_bio->sectors, 0))
2767 md_error(conf->mddev, rdev);
2772 for (m = 0; m < conf->copies; m++) {
2773 int dev = r10_bio->devs[m].devnum;
2774 struct bio *bio = r10_bio->devs[m].bio;
2775 rdev = conf->mirrors[dev].rdev;
2776 if (bio == IO_MADE_GOOD) {
2777 rdev_clear_badblocks(
2779 r10_bio->devs[m].addr,
2780 r10_bio->sectors, 0);
2781 rdev_dec_pending(rdev, conf->mddev);
2782 } else if (bio != NULL &&
2783 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2784 if (!narrow_write_error(r10_bio, m)) {
2785 md_error(conf->mddev, rdev);
2786 set_bit(R10BIO_Degraded,
2789 rdev_dec_pending(rdev, conf->mddev);
2791 bio = r10_bio->devs[m].repl_bio;
2792 rdev = conf->mirrors[dev].replacement;
2793 if (rdev && bio == IO_MADE_GOOD) {
2794 rdev_clear_badblocks(
2796 r10_bio->devs[m].addr,
2797 r10_bio->sectors, 0);
2798 rdev_dec_pending(rdev, conf->mddev);
2801 if (test_bit(R10BIO_WriteError,
2803 close_write(r10_bio);
2804 raid_end_bio_io(r10_bio);
2808 static void raid10d(struct md_thread *thread)
2810 struct mddev *mddev = thread->mddev;
2811 struct r10bio *r10_bio;
2812 unsigned long flags;
2813 struct r10conf *conf = mddev->private;
2814 struct list_head *head = &conf->retry_list;
2815 struct blk_plug plug;
2817 md_check_recovery(mddev);
2819 blk_start_plug(&plug);
2822 flush_pending_writes(conf);
2824 spin_lock_irqsave(&conf->device_lock, flags);
2825 if (list_empty(head)) {
2826 spin_unlock_irqrestore(&conf->device_lock, flags);
2829 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2830 list_del(head->prev);
2832 spin_unlock_irqrestore(&conf->device_lock, flags);
2834 mddev = r10_bio->mddev;
2835 conf = mddev->private;
2836 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2837 test_bit(R10BIO_WriteError, &r10_bio->state))
2838 handle_write_completed(conf, r10_bio);
2839 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2840 reshape_request_write(mddev, r10_bio);
2841 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2842 sync_request_write(mddev, r10_bio);
2843 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2844 recovery_request_write(mddev, r10_bio);
2845 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2846 handle_read_error(mddev, r10_bio);
2848 /* just a partial read to be scheduled from a
2851 int slot = r10_bio->read_slot;
2852 generic_make_request(r10_bio->devs[slot].bio);
2856 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2857 md_check_recovery(mddev);
2859 blk_finish_plug(&plug);
2863 static int init_resync(struct r10conf *conf)
2868 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2869 BUG_ON(conf->r10buf_pool);
2870 conf->have_replacement = 0;
2871 for (i = 0; i < conf->geo.raid_disks; i++)
2872 if (conf->mirrors[i].replacement)
2873 conf->have_replacement = 1;
2874 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2875 if (!conf->r10buf_pool)
2877 conf->next_resync = 0;
2882 * perform a "sync" on one "block"
2884 * We need to make sure that no normal I/O request - particularly write
2885 * requests - conflict with active sync requests.
2887 * This is achieved by tracking pending requests and a 'barrier' concept
2888 * that can be installed to exclude normal IO requests.
2890 * Resync and recovery are handled very differently.
2891 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2893 * For resync, we iterate over virtual addresses, read all copies,
2894 * and update if there are differences. If only one copy is live,
2896 * For recovery, we iterate over physical addresses, read a good
2897 * value for each non-in_sync drive, and over-write.
2899 * So, for recovery we may have several outstanding complex requests for a
2900 * given address, one for each out-of-sync device. We model this by allocating
2901 * a number of r10_bio structures, one for each out-of-sync device.
2902 * As we setup these structures, we collect all bio's together into a list
2903 * which we then process collectively to add pages, and then process again
2904 * to pass to generic_make_request.
2906 * The r10_bio structures are linked using a borrowed master_bio pointer.
2907 * This link is counted in ->remaining. When the r10_bio that points to NULL
2908 * has its remaining count decremented to 0, the whole complex operation
2913 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2914 int *skipped, int go_faster)
2916 struct r10conf *conf = mddev->private;
2917 struct r10bio *r10_bio;
2918 struct bio *biolist = NULL, *bio;
2919 sector_t max_sector, nr_sectors;
2922 sector_t sync_blocks;
2923 sector_t sectors_skipped = 0;
2924 int chunks_skipped = 0;
2925 sector_t chunk_mask = conf->geo.chunk_mask;
2927 if (!conf->r10buf_pool)
2928 if (init_resync(conf))
2932 * Allow skipping a full rebuild for incremental assembly
2933 * of a clean array, like RAID1 does.
2935 if (mddev->bitmap == NULL &&
2936 mddev->recovery_cp == MaxSector &&
2937 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2938 conf->fullsync == 0) {
2940 max_sector = mddev->dev_sectors;
2941 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2942 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2943 max_sector = mddev->resync_max_sectors;
2944 return max_sector - sector_nr;
2948 max_sector = mddev->dev_sectors;
2949 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2950 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2951 max_sector = mddev->resync_max_sectors;
2952 if (sector_nr >= max_sector) {
2953 /* If we aborted, we need to abort the
2954 * sync on the 'current' bitmap chucks (there can
2955 * be several when recovering multiple devices).
2956 * as we may have started syncing it but not finished.
2957 * We can find the current address in
2958 * mddev->curr_resync, but for recovery,
2959 * we need to convert that to several
2960 * virtual addresses.
2962 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2967 if (mddev->curr_resync < max_sector) { /* aborted */
2968 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2969 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2971 else for (i = 0; i < conf->geo.raid_disks; i++) {
2973 raid10_find_virt(conf, mddev->curr_resync, i);
2974 bitmap_end_sync(mddev->bitmap, sect,
2978 /* completed sync */
2979 if ((!mddev->bitmap || conf->fullsync)
2980 && conf->have_replacement
2981 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2982 /* Completed a full sync so the replacements
2983 * are now fully recovered.
2985 for (i = 0; i < conf->geo.raid_disks; i++)
2986 if (conf->mirrors[i].replacement)
2987 conf->mirrors[i].replacement
2993 bitmap_close_sync(mddev->bitmap);
2996 return sectors_skipped;
2999 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3000 return reshape_request(mddev, sector_nr, skipped);
3002 if (chunks_skipped >= conf->geo.raid_disks) {
3003 /* if there has been nothing to do on any drive,
3004 * then there is nothing to do at all..
3007 return (max_sector - sector_nr) + sectors_skipped;
3010 if (max_sector > mddev->resync_max)
3011 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3013 /* make sure whole request will fit in a chunk - if chunks
3016 if (conf->geo.near_copies < conf->geo.raid_disks &&
3017 max_sector > (sector_nr | chunk_mask))
3018 max_sector = (sector_nr | chunk_mask) + 1;
3020 * If there is non-resync activity waiting for us then
3021 * put in a delay to throttle resync.
3023 if (!go_faster && conf->nr_waiting)
3024 msleep_interruptible(1000);
3026 /* Again, very different code for resync and recovery.
3027 * Both must result in an r10bio with a list of bios that
3028 * have bi_end_io, bi_sector, bi_bdev set,
3029 * and bi_private set to the r10bio.
3030 * For recovery, we may actually create several r10bios
3031 * with 2 bios in each, that correspond to the bios in the main one.
3032 * In this case, the subordinate r10bios link back through a
3033 * borrowed master_bio pointer, and the counter in the master
3034 * includes a ref from each subordinate.
3036 /* First, we decide what to do and set ->bi_end_io
3037 * To end_sync_read if we want to read, and
3038 * end_sync_write if we will want to write.
3041 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3042 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3043 /* recovery... the complicated one */
3047 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3053 struct raid10_info *mirror = &conf->mirrors[i];
3055 if ((mirror->rdev == NULL ||
3056 test_bit(In_sync, &mirror->rdev->flags))
3058 (mirror->replacement == NULL ||
3060 &mirror->replacement->flags)))
3064 /* want to reconstruct this device */
3066 sect = raid10_find_virt(conf, sector_nr, i);
3067 if (sect >= mddev->resync_max_sectors) {
3068 /* last stripe is not complete - don't
3069 * try to recover this sector.
3073 /* Unless we are doing a full sync, or a replacement
3074 * we only need to recover the block if it is set in
3077 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3079 if (sync_blocks < max_sync)
3080 max_sync = sync_blocks;
3082 mirror->replacement == NULL &&
3084 /* yep, skip the sync_blocks here, but don't assume
3085 * that there will never be anything to do here
3087 chunks_skipped = -1;
3091 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3092 raise_barrier(conf, rb2 != NULL);
3093 atomic_set(&r10_bio->remaining, 0);
3095 r10_bio->master_bio = (struct bio*)rb2;
3097 atomic_inc(&rb2->remaining);
3098 r10_bio->mddev = mddev;
3099 set_bit(R10BIO_IsRecover, &r10_bio->state);
3100 r10_bio->sector = sect;
3102 raid10_find_phys(conf, r10_bio);
3104 /* Need to check if the array will still be
3107 for (j = 0; j < conf->geo.raid_disks; j++)
3108 if (conf->mirrors[j].rdev == NULL ||
3109 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3114 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3115 &sync_blocks, still_degraded);
3118 for (j=0; j<conf->copies;j++) {
3120 int d = r10_bio->devs[j].devnum;
3121 sector_t from_addr, to_addr;
3122 struct md_rdev *rdev;
3123 sector_t sector, first_bad;
3125 if (!conf->mirrors[d].rdev ||
3126 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3128 /* This is where we read from */
3130 rdev = conf->mirrors[d].rdev;
3131 sector = r10_bio->devs[j].addr;
3133 if (is_badblock(rdev, sector, max_sync,
3134 &first_bad, &bad_sectors)) {
3135 if (first_bad > sector)
3136 max_sync = first_bad - sector;
3138 bad_sectors -= (sector
3140 if (max_sync > bad_sectors)
3141 max_sync = bad_sectors;
3145 bio = r10_bio->devs[0].bio;
3146 bio->bi_next = biolist;
3148 bio->bi_private = r10_bio;
3149 bio->bi_end_io = end_sync_read;
3151 from_addr = r10_bio->devs[j].addr;
3152 bio->bi_sector = from_addr + rdev->data_offset;
3153 bio->bi_bdev = rdev->bdev;
3154 atomic_inc(&rdev->nr_pending);
3155 /* and we write to 'i' (if not in_sync) */
3157 for (k=0; k<conf->copies; k++)
3158 if (r10_bio->devs[k].devnum == i)
3160 BUG_ON(k == conf->copies);
3161 to_addr = r10_bio->devs[k].addr;
3162 r10_bio->devs[0].devnum = d;
3163 r10_bio->devs[0].addr = from_addr;
3164 r10_bio->devs[1].devnum = i;
3165 r10_bio->devs[1].addr = to_addr;
3167 rdev = mirror->rdev;
3168 if (!test_bit(In_sync, &rdev->flags)) {
3169 bio = r10_bio->devs[1].bio;
3170 bio->bi_next = biolist;
3172 bio->bi_private = r10_bio;
3173 bio->bi_end_io = end_sync_write;
3175 bio->bi_sector = to_addr
3176 + rdev->data_offset;
3177 bio->bi_bdev = rdev->bdev;
3178 atomic_inc(&r10_bio->remaining);
3180 r10_bio->devs[1].bio->bi_end_io = NULL;
3182 /* and maybe write to replacement */
3183 bio = r10_bio->devs[1].repl_bio;
3185 bio->bi_end_io = NULL;
3186 rdev = mirror->replacement;
3187 /* Note: if rdev != NULL, then bio
3188 * cannot be NULL as r10buf_pool_alloc will
3189 * have allocated it.
3190 * So the second test here is pointless.
3191 * But it keeps semantic-checkers happy, and
3192 * this comment keeps human reviewers
3195 if (rdev == NULL || bio == NULL ||
3196 test_bit(Faulty, &rdev->flags))
3198 bio->bi_next = biolist;
3200 bio->bi_private = r10_bio;
3201 bio->bi_end_io = end_sync_write;
3203 bio->bi_sector = to_addr + rdev->data_offset;
3204 bio->bi_bdev = rdev->bdev;
3205 atomic_inc(&r10_bio->remaining);
3208 if (j == conf->copies) {
3209 /* Cannot recover, so abort the recovery or
3210 * record a bad block */
3213 atomic_dec(&rb2->remaining);
3216 /* problem is that there are bad blocks
3217 * on other device(s)
3220 for (k = 0; k < conf->copies; k++)
3221 if (r10_bio->devs[k].devnum == i)
3223 if (!test_bit(In_sync,
3224 &mirror->rdev->flags)
3225 && !rdev_set_badblocks(
3227 r10_bio->devs[k].addr,
3230 if (mirror->replacement &&
3231 !rdev_set_badblocks(
3232 mirror->replacement,
3233 r10_bio->devs[k].addr,
3238 if (!test_and_set_bit(MD_RECOVERY_INTR,
3240 printk(KERN_INFO "md/raid10:%s: insufficient "
3241 "working devices for recovery.\n",
3243 mirror->recovery_disabled
3244 = mddev->recovery_disabled;
3249 if (biolist == NULL) {
3251 struct r10bio *rb2 = r10_bio;
3252 r10_bio = (struct r10bio*) rb2->master_bio;
3253 rb2->master_bio = NULL;
3259 /* resync. Schedule a read for every block at this virt offset */
3262 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3264 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3265 &sync_blocks, mddev->degraded) &&
3266 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3267 &mddev->recovery)) {
3268 /* We can skip this block */
3270 return sync_blocks + sectors_skipped;
3272 if (sync_blocks < max_sync)
3273 max_sync = sync_blocks;
3274 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3276 r10_bio->mddev = mddev;
3277 atomic_set(&r10_bio->remaining, 0);
3278 raise_barrier(conf, 0);
3279 conf->next_resync = sector_nr;
3281 r10_bio->master_bio = NULL;
3282 r10_bio->sector = sector_nr;
3283 set_bit(R10BIO_IsSync, &r10_bio->state);
3284 raid10_find_phys(conf, r10_bio);
3285 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3287 for (i = 0; i < conf->copies; i++) {
3288 int d = r10_bio->devs[i].devnum;
3289 sector_t first_bad, sector;
3292 if (r10_bio->devs[i].repl_bio)
3293 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3295 bio = r10_bio->devs[i].bio;
3296 bio->bi_end_io = NULL;
3297 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3298 if (conf->mirrors[d].rdev == NULL ||
3299 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3301 sector = r10_bio->devs[i].addr;
3302 if (is_badblock(conf->mirrors[d].rdev,
3304 &first_bad, &bad_sectors)) {
3305 if (first_bad > sector)
3306 max_sync = first_bad - sector;
3308 bad_sectors -= (sector - first_bad);
3309 if (max_sync > bad_sectors)
3310 max_sync = bad_sectors;
3314 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3315 atomic_inc(&r10_bio->remaining);
3316 bio->bi_next = biolist;
3318 bio->bi_private = r10_bio;
3319 bio->bi_end_io = end_sync_read;
3321 bio->bi_sector = sector +
3322 conf->mirrors[d].rdev->data_offset;
3323 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3326 if (conf->mirrors[d].replacement == NULL ||
3328 &conf->mirrors[d].replacement->flags))
3331 /* Need to set up for writing to the replacement */
3332 bio = r10_bio->devs[i].repl_bio;
3333 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3335 sector = r10_bio->devs[i].addr;
3336 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3337 bio->bi_next = biolist;
3339 bio->bi_private = r10_bio;
3340 bio->bi_end_io = end_sync_write;
3342 bio->bi_sector = sector +
3343 conf->mirrors[d].replacement->data_offset;
3344 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3349 for (i=0; i<conf->copies; i++) {
3350 int d = r10_bio->devs[i].devnum;
3351 if (r10_bio->devs[i].bio->bi_end_io)
3352 rdev_dec_pending(conf->mirrors[d].rdev,
3354 if (r10_bio->devs[i].repl_bio &&
3355 r10_bio->devs[i].repl_bio->bi_end_io)
3357 conf->mirrors[d].replacement,
3366 for (bio = biolist; bio ; bio=bio->bi_next) {
3368 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3370 bio->bi_flags |= 1 << BIO_UPTODATE;
3373 bio->bi_phys_segments = 0;
3378 if (sector_nr + max_sync < max_sector)
3379 max_sector = sector_nr + max_sync;
3382 int len = PAGE_SIZE;
3383 if (sector_nr + (len>>9) > max_sector)
3384 len = (max_sector - sector_nr) << 9;
3387 for (bio= biolist ; bio ; bio=bio->bi_next) {
3389 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3390 if (bio_add_page(bio, page, len, 0))
3394 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3395 for (bio2 = biolist;
3396 bio2 && bio2 != bio;
3397 bio2 = bio2->bi_next) {
3398 /* remove last page from this bio */
3400 bio2->bi_size -= len;
3401 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3405 nr_sectors += len>>9;
3406 sector_nr += len>>9;
3407 } while (biolist->bi_vcnt < RESYNC_PAGES);
3409 r10_bio->sectors = nr_sectors;
3413 biolist = biolist->bi_next;
3415 bio->bi_next = NULL;
3416 r10_bio = bio->bi_private;
3417 r10_bio->sectors = nr_sectors;
3419 if (bio->bi_end_io == end_sync_read) {
3420 md_sync_acct(bio->bi_bdev, nr_sectors);
3421 generic_make_request(bio);
3425 if (sectors_skipped)
3426 /* pretend they weren't skipped, it makes
3427 * no important difference in this case
3429 md_done_sync(mddev, sectors_skipped, 1);
3431 return sectors_skipped + nr_sectors;
3433 /* There is nowhere to write, so all non-sync
3434 * drives must be failed or in resync, all drives
3435 * have a bad block, so try the next chunk...
3437 if (sector_nr + max_sync < max_sector)
3438 max_sector = sector_nr + max_sync;
3440 sectors_skipped += (max_sector - sector_nr);
3442 sector_nr = max_sector;
3447 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3450 struct r10conf *conf = mddev->private;
3453 raid_disks = min(conf->geo.raid_disks,
3454 conf->prev.raid_disks);
3456 sectors = conf->dev_sectors;
3458 size = sectors >> conf->geo.chunk_shift;
3459 sector_div(size, conf->geo.far_copies);
3460 size = size * raid_disks;
3461 sector_div(size, conf->geo.near_copies);
3463 return size << conf->geo.chunk_shift;
3466 static void calc_sectors(struct r10conf *conf, sector_t size)
3468 /* Calculate the number of sectors-per-device that will
3469 * actually be used, and set conf->dev_sectors and
3473 size = size >> conf->geo.chunk_shift;
3474 sector_div(size, conf->geo.far_copies);
3475 size = size * conf->geo.raid_disks;
3476 sector_div(size, conf->geo.near_copies);
3477 /* 'size' is now the number of chunks in the array */
3478 /* calculate "used chunks per device" */
3479 size = size * conf->copies;
3481 /* We need to round up when dividing by raid_disks to
3482 * get the stride size.
3484 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3486 conf->dev_sectors = size << conf->geo.chunk_shift;
3488 if (conf->geo.far_offset)
3489 conf->geo.stride = 1 << conf->geo.chunk_shift;
3491 sector_div(size, conf->geo.far_copies);
3492 conf->geo.stride = size << conf->geo.chunk_shift;
3496 enum geo_type {geo_new, geo_old, geo_start};
3497 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3500 int layout, chunk, disks;
3503 layout = mddev->layout;
3504 chunk = mddev->chunk_sectors;
3505 disks = mddev->raid_disks - mddev->delta_disks;
3508 layout = mddev->new_layout;
3509 chunk = mddev->new_chunk_sectors;
3510 disks = mddev->raid_disks;
3512 default: /* avoid 'may be unused' warnings */
3513 case geo_start: /* new when starting reshape - raid_disks not
3515 layout = mddev->new_layout;
3516 chunk = mddev->new_chunk_sectors;
3517 disks = mddev->raid_disks + mddev->delta_disks;
3522 if (chunk < (PAGE_SIZE >> 9) ||
3523 !is_power_of_2(chunk))
3526 fc = (layout >> 8) & 255;
3527 fo = layout & (1<<16);
3528 geo->raid_disks = disks;
3529 geo->near_copies = nc;
3530 geo->far_copies = fc;
3531 geo->far_offset = fo;
3532 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3533 geo->chunk_mask = chunk - 1;
3534 geo->chunk_shift = ffz(~chunk);
3538 static struct r10conf *setup_conf(struct mddev *mddev)
3540 struct r10conf *conf = NULL;
3545 copies = setup_geo(&geo, mddev, geo_new);
3548 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3549 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3550 mdname(mddev), PAGE_SIZE);
3554 if (copies < 2 || copies > mddev->raid_disks) {
3555 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3556 mdname(mddev), mddev->new_layout);
3561 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3565 /* FIXME calc properly */
3566 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3567 max(0,mddev->delta_disks)),
3572 conf->tmppage = alloc_page(GFP_KERNEL);
3577 conf->copies = copies;
3578 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3579 r10bio_pool_free, conf);
3580 if (!conf->r10bio_pool)
3583 calc_sectors(conf, mddev->dev_sectors);
3584 if (mddev->reshape_position == MaxSector) {
3585 conf->prev = conf->geo;
3586 conf->reshape_progress = MaxSector;
3588 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3592 conf->reshape_progress = mddev->reshape_position;
3593 if (conf->prev.far_offset)
3594 conf->prev.stride = 1 << conf->prev.chunk_shift;
3596 /* far_copies must be 1 */
3597 conf->prev.stride = conf->dev_sectors;
3599 spin_lock_init(&conf->device_lock);
3600 INIT_LIST_HEAD(&conf->retry_list);
3602 spin_lock_init(&conf->resync_lock);
3603 init_waitqueue_head(&conf->wait_barrier);
3605 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3609 conf->mddev = mddev;
3614 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3617 if (conf->r10bio_pool)
3618 mempool_destroy(conf->r10bio_pool);
3619 kfree(conf->mirrors);
3620 safe_put_page(conf->tmppage);
3623 return ERR_PTR(err);
3626 static int run(struct mddev *mddev)
3628 struct r10conf *conf;
3629 int i, disk_idx, chunk_size;
3630 struct raid10_info *disk;
3631 struct md_rdev *rdev;
3633 sector_t min_offset_diff = 0;
3635 bool discard_supported = false;
3637 if (mddev->private == NULL) {
3638 conf = setup_conf(mddev);
3640 return PTR_ERR(conf);
3641 mddev->private = conf;
3643 conf = mddev->private;
3647 mddev->thread = conf->thread;
3648 conf->thread = NULL;
3650 chunk_size = mddev->chunk_sectors << 9;
3652 blk_queue_max_discard_sectors(mddev->queue,
3653 mddev->chunk_sectors);
3654 blk_queue_max_write_same_sectors(mddev->queue,
3655 mddev->chunk_sectors);
3656 blk_queue_io_min(mddev->queue, chunk_size);
3657 if (conf->geo.raid_disks % conf->geo.near_copies)
3658 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3660 blk_queue_io_opt(mddev->queue, chunk_size *
3661 (conf->geo.raid_disks / conf->geo.near_copies));
3664 rdev_for_each(rdev, mddev) {
3666 struct request_queue *q;
3668 disk_idx = rdev->raid_disk;
3671 if (disk_idx >= conf->geo.raid_disks &&
3672 disk_idx >= conf->prev.raid_disks)
3674 disk = conf->mirrors + disk_idx;
3676 if (test_bit(Replacement, &rdev->flags)) {
3677 if (disk->replacement)
3679 disk->replacement = rdev;
3685 q = bdev_get_queue(rdev->bdev);
3686 if (q->merge_bvec_fn)
3687 mddev->merge_check_needed = 1;
3688 diff = (rdev->new_data_offset - rdev->data_offset);
3689 if (!mddev->reshape_backwards)
3693 if (first || diff < min_offset_diff)
3694 min_offset_diff = diff;
3697 disk_stack_limits(mddev->gendisk, rdev->bdev,
3698 rdev->data_offset << 9);
3700 disk->head_position = 0;
3702 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3703 discard_supported = true;
3707 if (discard_supported)
3708 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3711 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3714 /* need to check that every block has at least one working mirror */
3715 if (!enough(conf, -1)) {
3716 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3721 if (conf->reshape_progress != MaxSector) {
3722 /* must ensure that shape change is supported */
3723 if (conf->geo.far_copies != 1 &&
3724 conf->geo.far_offset == 0)
3726 if (conf->prev.far_copies != 1 &&
3727 conf->geo.far_offset == 0)
3731 mddev->degraded = 0;
3733 i < conf->geo.raid_disks
3734 || i < conf->prev.raid_disks;
3737 disk = conf->mirrors + i;
3739 if (!disk->rdev && disk->replacement) {
3740 /* The replacement is all we have - use it */
3741 disk->rdev = disk->replacement;
3742 disk->replacement = NULL;
3743 clear_bit(Replacement, &disk->rdev->flags);
3747 !test_bit(In_sync, &disk->rdev->flags)) {
3748 disk->head_position = 0;
3753 disk->recovery_disabled = mddev->recovery_disabled - 1;
3756 if (mddev->recovery_cp != MaxSector)
3757 printk(KERN_NOTICE "md/raid10:%s: not clean"
3758 " -- starting background reconstruction\n",
3761 "md/raid10:%s: active with %d out of %d devices\n",
3762 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3763 conf->geo.raid_disks);
3765 * Ok, everything is just fine now
3767 mddev->dev_sectors = conf->dev_sectors;
3768 size = raid10_size(mddev, 0, 0);
3769 md_set_array_sectors(mddev, size);
3770 mddev->resync_max_sectors = size;
3773 int stripe = conf->geo.raid_disks *
3774 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3775 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3776 mddev->queue->backing_dev_info.congested_data = mddev;
3778 /* Calculate max read-ahead size.
3779 * We need to readahead at least twice a whole stripe....
3782 stripe /= conf->geo.near_copies;
3783 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3784 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3785 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3789 if (md_integrity_register(mddev))
3792 if (conf->reshape_progress != MaxSector) {
3793 unsigned long before_length, after_length;
3795 before_length = ((1 << conf->prev.chunk_shift) *
3796 conf->prev.far_copies);
3797 after_length = ((1 << conf->geo.chunk_shift) *
3798 conf->geo.far_copies);
3800 if (max(before_length, after_length) > min_offset_diff) {
3801 /* This cannot work */
3802 printk("md/raid10: offset difference not enough to continue reshape\n");
3805 conf->offset_diff = min_offset_diff;
3807 conf->reshape_safe = conf->reshape_progress;
3808 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3809 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3810 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3811 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3812 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3819 md_unregister_thread(&mddev->thread);
3820 if (conf->r10bio_pool)
3821 mempool_destroy(conf->r10bio_pool);
3822 safe_put_page(conf->tmppage);
3823 kfree(conf->mirrors);
3825 mddev->private = NULL;
3830 static int stop(struct mddev *mddev)
3832 struct r10conf *conf = mddev->private;
3834 raise_barrier(conf, 0);
3835 lower_barrier(conf);
3837 md_unregister_thread(&mddev->thread);
3839 /* the unplug fn references 'conf'*/
3840 blk_sync_queue(mddev->queue);
3842 if (conf->r10bio_pool)
3843 mempool_destroy(conf->r10bio_pool);
3844 safe_put_page(conf->tmppage);
3845 kfree(conf->mirrors);
3847 mddev->private = NULL;
3851 static void raid10_quiesce(struct mddev *mddev, int state)
3853 struct r10conf *conf = mddev->private;
3857 raise_barrier(conf, 0);
3860 lower_barrier(conf);
3865 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3867 /* Resize of 'far' arrays is not supported.
3868 * For 'near' and 'offset' arrays we can set the
3869 * number of sectors used to be an appropriate multiple
3870 * of the chunk size.
3871 * For 'offset', this is far_copies*chunksize.
3872 * For 'near' the multiplier is the LCM of
3873 * near_copies and raid_disks.
3874 * So if far_copies > 1 && !far_offset, fail.
3875 * Else find LCM(raid_disks, near_copy)*far_copies and
3876 * multiply by chunk_size. Then round to this number.
3877 * This is mostly done by raid10_size()
3879 struct r10conf *conf = mddev->private;
3880 sector_t oldsize, size;
3882 if (mddev->reshape_position != MaxSector)
3885 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3888 oldsize = raid10_size(mddev, 0, 0);
3889 size = raid10_size(mddev, sectors, 0);
3890 if (mddev->external_size &&
3891 mddev->array_sectors > size)
3893 if (mddev->bitmap) {
3894 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3898 md_set_array_sectors(mddev, size);
3899 set_capacity(mddev->gendisk, mddev->array_sectors);
3900 revalidate_disk(mddev->gendisk);
3901 if (sectors > mddev->dev_sectors &&
3902 mddev->recovery_cp > oldsize) {
3903 mddev->recovery_cp = oldsize;
3904 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3906 calc_sectors(conf, sectors);
3907 mddev->dev_sectors = conf->dev_sectors;
3908 mddev->resync_max_sectors = size;
3912 static void *raid10_takeover_raid0(struct mddev *mddev)
3914 struct md_rdev *rdev;
3915 struct r10conf *conf;
3917 if (mddev->degraded > 0) {
3918 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3920 return ERR_PTR(-EINVAL);
3923 /* Set new parameters */
3924 mddev->new_level = 10;
3925 /* new layout: far_copies = 1, near_copies = 2 */
3926 mddev->new_layout = (1<<8) + 2;
3927 mddev->new_chunk_sectors = mddev->chunk_sectors;
3928 mddev->delta_disks = mddev->raid_disks;
3929 mddev->raid_disks *= 2;
3930 /* make sure it will be not marked as dirty */
3931 mddev->recovery_cp = MaxSector;
3933 conf = setup_conf(mddev);
3934 if (!IS_ERR(conf)) {
3935 rdev_for_each(rdev, mddev)
3936 if (rdev->raid_disk >= 0)
3937 rdev->new_raid_disk = rdev->raid_disk * 2;
3944 static void *raid10_takeover(struct mddev *mddev)
3946 struct r0conf *raid0_conf;
3948 /* raid10 can take over:
3949 * raid0 - providing it has only two drives
3951 if (mddev->level == 0) {
3952 /* for raid0 takeover only one zone is supported */
3953 raid0_conf = mddev->private;
3954 if (raid0_conf->nr_strip_zones > 1) {
3955 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3956 " with more than one zone.\n",
3958 return ERR_PTR(-EINVAL);
3960 return raid10_takeover_raid0(mddev);
3962 return ERR_PTR(-EINVAL);
3965 static int raid10_check_reshape(struct mddev *mddev)
3967 /* Called when there is a request to change
3968 * - layout (to ->new_layout)
3969 * - chunk size (to ->new_chunk_sectors)
3970 * - raid_disks (by delta_disks)
3971 * or when trying to restart a reshape that was ongoing.
3973 * We need to validate the request and possibly allocate
3974 * space if that might be an issue later.
3976 * Currently we reject any reshape of a 'far' mode array,
3977 * allow chunk size to change if new is generally acceptable,
3978 * allow raid_disks to increase, and allow
3979 * a switch between 'near' mode and 'offset' mode.
3981 struct r10conf *conf = mddev->private;
3984 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3987 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3988 /* mustn't change number of copies */
3990 if (geo.far_copies > 1 && !geo.far_offset)
3991 /* Cannot switch to 'far' mode */
3994 if (mddev->array_sectors & geo.chunk_mask)
3995 /* not factor of array size */
3998 if (!enough(conf, -1))
4001 kfree(conf->mirrors_new);
4002 conf->mirrors_new = NULL;
4003 if (mddev->delta_disks > 0) {
4004 /* allocate new 'mirrors' list */
4005 conf->mirrors_new = kzalloc(
4006 sizeof(struct raid10_info)
4007 *(mddev->raid_disks +
4008 mddev->delta_disks),
4010 if (!conf->mirrors_new)
4017 * Need to check if array has failed when deciding whether to:
4019 * - remove non-faulty devices
4022 * This determination is simple when no reshape is happening.
4023 * However if there is a reshape, we need to carefully check
4024 * both the before and after sections.
4025 * This is because some failed devices may only affect one
4026 * of the two sections, and some non-in_sync devices may
4027 * be insync in the section most affected by failed devices.
4029 static int calc_degraded(struct r10conf *conf)
4031 int degraded, degraded2;
4036 /* 'prev' section first */
4037 for (i = 0; i < conf->prev.raid_disks; i++) {
4038 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4039 if (!rdev || test_bit(Faulty, &rdev->flags))
4041 else if (!test_bit(In_sync, &rdev->flags))
4042 /* When we can reduce the number of devices in
4043 * an array, this might not contribute to
4044 * 'degraded'. It does now.
4049 if (conf->geo.raid_disks == conf->prev.raid_disks)
4053 for (i = 0; i < conf->geo.raid_disks; i++) {
4054 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4055 if (!rdev || test_bit(Faulty, &rdev->flags))
4057 else if (!test_bit(In_sync, &rdev->flags)) {
4058 /* If reshape is increasing the number of devices,
4059 * this section has already been recovered, so
4060 * it doesn't contribute to degraded.
4063 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4068 if (degraded2 > degraded)
4073 static int raid10_start_reshape(struct mddev *mddev)
4075 /* A 'reshape' has been requested. This commits
4076 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4077 * This also checks if there are enough spares and adds them
4079 * We currently require enough spares to make the final
4080 * array non-degraded. We also require that the difference
4081 * between old and new data_offset - on each device - is
4082 * enough that we never risk over-writing.
4085 unsigned long before_length, after_length;
4086 sector_t min_offset_diff = 0;
4089 struct r10conf *conf = mddev->private;
4090 struct md_rdev *rdev;
4094 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4097 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4100 before_length = ((1 << conf->prev.chunk_shift) *
4101 conf->prev.far_copies);
4102 after_length = ((1 << conf->geo.chunk_shift) *
4103 conf->geo.far_copies);
4105 rdev_for_each(rdev, mddev) {
4106 if (!test_bit(In_sync, &rdev->flags)
4107 && !test_bit(Faulty, &rdev->flags))
4109 if (rdev->raid_disk >= 0) {
4110 long long diff = (rdev->new_data_offset
4111 - rdev->data_offset);
4112 if (!mddev->reshape_backwards)
4116 if (first || diff < min_offset_diff)
4117 min_offset_diff = diff;
4121 if (max(before_length, after_length) > min_offset_diff)
4124 if (spares < mddev->delta_disks)
4127 conf->offset_diff = min_offset_diff;
4128 spin_lock_irq(&conf->device_lock);
4129 if (conf->mirrors_new) {
4130 memcpy(conf->mirrors_new, conf->mirrors,
4131 sizeof(struct raid10_info)*conf->prev.raid_disks);
4133 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4134 conf->mirrors_old = conf->mirrors;
4135 conf->mirrors = conf->mirrors_new;
4136 conf->mirrors_new = NULL;
4138 setup_geo(&conf->geo, mddev, geo_start);
4140 if (mddev->reshape_backwards) {
4141 sector_t size = raid10_size(mddev, 0, 0);
4142 if (size < mddev->array_sectors) {
4143 spin_unlock_irq(&conf->device_lock);
4144 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4148 mddev->resync_max_sectors = size;
4149 conf->reshape_progress = size;
4151 conf->reshape_progress = 0;
4152 spin_unlock_irq(&conf->device_lock);
4154 if (mddev->delta_disks && mddev->bitmap) {
4155 ret = bitmap_resize(mddev->bitmap,
4156 raid10_size(mddev, 0,
4157 conf->geo.raid_disks),
4162 if (mddev->delta_disks > 0) {
4163 rdev_for_each(rdev, mddev)
4164 if (rdev->raid_disk < 0 &&
4165 !test_bit(Faulty, &rdev->flags)) {
4166 if (raid10_add_disk(mddev, rdev) == 0) {
4167 if (rdev->raid_disk >=
4168 conf->prev.raid_disks)
4169 set_bit(In_sync, &rdev->flags);
4171 rdev->recovery_offset = 0;
4173 if (sysfs_link_rdev(mddev, rdev))
4174 /* Failure here is OK */;
4176 } else if (rdev->raid_disk >= conf->prev.raid_disks
4177 && !test_bit(Faulty, &rdev->flags)) {
4178 /* This is a spare that was manually added */
4179 set_bit(In_sync, &rdev->flags);
4182 /* When a reshape changes the number of devices,
4183 * ->degraded is measured against the larger of the
4184 * pre and post numbers.
4186 spin_lock_irq(&conf->device_lock);
4187 mddev->degraded = calc_degraded(conf);
4188 spin_unlock_irq(&conf->device_lock);
4189 mddev->raid_disks = conf->geo.raid_disks;
4190 mddev->reshape_position = conf->reshape_progress;
4191 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4193 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4194 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4195 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4196 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4198 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4200 if (!mddev->sync_thread) {
4204 conf->reshape_checkpoint = jiffies;
4205 md_wakeup_thread(mddev->sync_thread);
4206 md_new_event(mddev);
4210 mddev->recovery = 0;
4211 spin_lock_irq(&conf->device_lock);
4212 conf->geo = conf->prev;
4213 mddev->raid_disks = conf->geo.raid_disks;
4214 rdev_for_each(rdev, mddev)
4215 rdev->new_data_offset = rdev->data_offset;
4217 conf->reshape_progress = MaxSector;
4218 mddev->reshape_position = MaxSector;
4219 spin_unlock_irq(&conf->device_lock);
4223 /* Calculate the last device-address that could contain
4224 * any block from the chunk that includes the array-address 's'
4225 * and report the next address.
4226 * i.e. the address returned will be chunk-aligned and after
4227 * any data that is in the chunk containing 's'.
4229 static sector_t last_dev_address(sector_t s, struct geom *geo)
4231 s = (s | geo->chunk_mask) + 1;
4232 s >>= geo->chunk_shift;
4233 s *= geo->near_copies;
4234 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4235 s *= geo->far_copies;
4236 s <<= geo->chunk_shift;
4240 /* Calculate the first device-address that could contain
4241 * any block from the chunk that includes the array-address 's'.
4242 * This too will be the start of a chunk
4244 static sector_t first_dev_address(sector_t s, struct geom *geo)
4246 s >>= geo->chunk_shift;
4247 s *= geo->near_copies;
4248 sector_div(s, geo->raid_disks);
4249 s *= geo->far_copies;
4250 s <<= geo->chunk_shift;
4254 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4257 /* We simply copy at most one chunk (smallest of old and new)
4258 * at a time, possibly less if that exceeds RESYNC_PAGES,
4259 * or we hit a bad block or something.
4260 * This might mean we pause for normal IO in the middle of
4261 * a chunk, but that is not a problem was mddev->reshape_position
4262 * can record any location.
4264 * If we will want to write to a location that isn't
4265 * yet recorded as 'safe' (i.e. in metadata on disk) then
4266 * we need to flush all reshape requests and update the metadata.
4268 * When reshaping forwards (e.g. to more devices), we interpret
4269 * 'safe' as the earliest block which might not have been copied
4270 * down yet. We divide this by previous stripe size and multiply
4271 * by previous stripe length to get lowest device offset that we
4272 * cannot write to yet.
4273 * We interpret 'sector_nr' as an address that we want to write to.
4274 * From this we use last_device_address() to find where we might
4275 * write to, and first_device_address on the 'safe' position.
4276 * If this 'next' write position is after the 'safe' position,
4277 * we must update the metadata to increase the 'safe' position.
4279 * When reshaping backwards, we round in the opposite direction
4280 * and perform the reverse test: next write position must not be
4281 * less than current safe position.
4283 * In all this the minimum difference in data offsets
4284 * (conf->offset_diff - always positive) allows a bit of slack,
4285 * so next can be after 'safe', but not by more than offset_disk
4287 * We need to prepare all the bios here before we start any IO
4288 * to ensure the size we choose is acceptable to all devices.
4289 * The means one for each copy for write-out and an extra one for
4291 * We store the read-in bio in ->master_bio and the others in
4292 * ->devs[x].bio and ->devs[x].repl_bio.
4294 struct r10conf *conf = mddev->private;
4295 struct r10bio *r10_bio;
4296 sector_t next, safe, last;
4300 struct md_rdev *rdev;
4303 struct bio *bio, *read_bio;
4304 int sectors_done = 0;
4306 if (sector_nr == 0) {
4307 /* If restarting in the middle, skip the initial sectors */
4308 if (mddev->reshape_backwards &&
4309 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4310 sector_nr = (raid10_size(mddev, 0, 0)
4311 - conf->reshape_progress);
4312 } else if (!mddev->reshape_backwards &&
4313 conf->reshape_progress > 0)
4314 sector_nr = conf->reshape_progress;
4316 mddev->curr_resync_completed = sector_nr;
4317 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4323 /* We don't use sector_nr to track where we are up to
4324 * as that doesn't work well for ->reshape_backwards.
4325 * So just use ->reshape_progress.
4327 if (mddev->reshape_backwards) {
4328 /* 'next' is the earliest device address that we might
4329 * write to for this chunk in the new layout
4331 next = first_dev_address(conf->reshape_progress - 1,
4334 /* 'safe' is the last device address that we might read from
4335 * in the old layout after a restart
4337 safe = last_dev_address(conf->reshape_safe - 1,
4340 if (next + conf->offset_diff < safe)
4343 last = conf->reshape_progress - 1;
4344 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4345 & conf->prev.chunk_mask);
4346 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4347 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4349 /* 'next' is after the last device address that we
4350 * might write to for this chunk in the new layout
4352 next = last_dev_address(conf->reshape_progress, &conf->geo);
4354 /* 'safe' is the earliest device address that we might
4355 * read from in the old layout after a restart
4357 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4359 /* Need to update metadata if 'next' might be beyond 'safe'
4360 * as that would possibly corrupt data
4362 if (next > safe + conf->offset_diff)
4365 sector_nr = conf->reshape_progress;
4366 last = sector_nr | (conf->geo.chunk_mask
4367 & conf->prev.chunk_mask);
4369 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4370 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4374 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4375 /* Need to update reshape_position in metadata */
4377 mddev->reshape_position = conf->reshape_progress;
4378 if (mddev->reshape_backwards)
4379 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4380 - conf->reshape_progress;
4382 mddev->curr_resync_completed = conf->reshape_progress;
4383 conf->reshape_checkpoint = jiffies;
4384 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4385 md_wakeup_thread(mddev->thread);
4386 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4387 kthread_should_stop());
4388 conf->reshape_safe = mddev->reshape_position;
4389 allow_barrier(conf);
4393 /* Now schedule reads for blocks from sector_nr to last */
4394 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4395 raise_barrier(conf, sectors_done != 0);
4396 atomic_set(&r10_bio->remaining, 0);
4397 r10_bio->mddev = mddev;
4398 r10_bio->sector = sector_nr;
4399 set_bit(R10BIO_IsReshape, &r10_bio->state);
4400 r10_bio->sectors = last - sector_nr + 1;
4401 rdev = read_balance(conf, r10_bio, &max_sectors);
4402 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4405 /* Cannot read from here, so need to record bad blocks
4406 * on all the target devices.
4409 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4410 return sectors_done;
4413 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4415 read_bio->bi_bdev = rdev->bdev;
4416 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4417 + rdev->data_offset);
4418 read_bio->bi_private = r10_bio;
4419 read_bio->bi_end_io = end_sync_read;
4420 read_bio->bi_rw = READ;
4421 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4422 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4423 read_bio->bi_vcnt = 0;
4424 read_bio->bi_idx = 0;
4425 read_bio->bi_size = 0;
4426 r10_bio->master_bio = read_bio;
4427 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4429 /* Now find the locations in the new layout */
4430 __raid10_find_phys(&conf->geo, r10_bio);
4433 read_bio->bi_next = NULL;
4435 for (s = 0; s < conf->copies*2; s++) {
4437 int d = r10_bio->devs[s/2].devnum;
4438 struct md_rdev *rdev2;
4440 rdev2 = conf->mirrors[d].replacement;
4441 b = r10_bio->devs[s/2].repl_bio;
4443 rdev2 = conf->mirrors[d].rdev;
4444 b = r10_bio->devs[s/2].bio;
4446 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4448 b->bi_bdev = rdev2->bdev;
4449 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4450 b->bi_private = r10_bio;
4451 b->bi_end_io = end_reshape_write;
4453 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4454 b->bi_flags |= 1 << BIO_UPTODATE;
4462 /* Now add as many pages as possible to all of these bios. */
4465 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4466 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4467 int len = (max_sectors - s) << 9;
4468 if (len > PAGE_SIZE)
4470 for (bio = blist; bio ; bio = bio->bi_next) {
4472 if (bio_add_page(bio, page, len, 0))
4475 /* Didn't fit, must stop */
4477 bio2 && bio2 != bio;
4478 bio2 = bio2->bi_next) {
4479 /* Remove last page from this bio */
4481 bio2->bi_size -= len;
4482 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4486 sector_nr += len >> 9;
4487 nr_sectors += len >> 9;
4490 r10_bio->sectors = nr_sectors;
4492 /* Now submit the read */
4493 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4494 atomic_inc(&r10_bio->remaining);
4495 read_bio->bi_next = NULL;
4496 generic_make_request(read_bio);
4497 sector_nr += nr_sectors;
4498 sectors_done += nr_sectors;
4499 if (sector_nr <= last)
4502 /* Now that we have done the whole section we can
4503 * update reshape_progress
4505 if (mddev->reshape_backwards)
4506 conf->reshape_progress -= sectors_done;
4508 conf->reshape_progress += sectors_done;
4510 return sectors_done;
4513 static void end_reshape_request(struct r10bio *r10_bio);
4514 static int handle_reshape_read_error(struct mddev *mddev,
4515 struct r10bio *r10_bio);
4516 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4518 /* Reshape read completed. Hopefully we have a block
4520 * If we got a read error then we do sync 1-page reads from
4521 * elsewhere until we find the data - or give up.
4523 struct r10conf *conf = mddev->private;
4526 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4527 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4528 /* Reshape has been aborted */
4529 md_done_sync(mddev, r10_bio->sectors, 0);
4533 /* We definitely have the data in the pages, schedule the
4536 atomic_set(&r10_bio->remaining, 1);
4537 for (s = 0; s < conf->copies*2; s++) {
4539 int d = r10_bio->devs[s/2].devnum;
4540 struct md_rdev *rdev;
4542 rdev = conf->mirrors[d].replacement;
4543 b = r10_bio->devs[s/2].repl_bio;
4545 rdev = conf->mirrors[d].rdev;
4546 b = r10_bio->devs[s/2].bio;
4548 if (!rdev || test_bit(Faulty, &rdev->flags))
4550 atomic_inc(&rdev->nr_pending);
4551 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4552 atomic_inc(&r10_bio->remaining);
4554 generic_make_request(b);
4556 end_reshape_request(r10_bio);
4559 static void end_reshape(struct r10conf *conf)
4561 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4564 spin_lock_irq(&conf->device_lock);
4565 conf->prev = conf->geo;
4566 md_finish_reshape(conf->mddev);
4568 conf->reshape_progress = MaxSector;
4569 spin_unlock_irq(&conf->device_lock);
4571 /* read-ahead size must cover two whole stripes, which is
4572 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4574 if (conf->mddev->queue) {
4575 int stripe = conf->geo.raid_disks *
4576 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4577 stripe /= conf->geo.near_copies;
4578 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4579 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4585 static int handle_reshape_read_error(struct mddev *mddev,
4586 struct r10bio *r10_bio)
4588 /* Use sync reads to get the blocks from somewhere else */
4589 int sectors = r10_bio->sectors;
4590 struct r10conf *conf = mddev->private;
4592 struct r10bio r10_bio;
4593 struct r10dev devs[conf->copies];
4595 struct r10bio *r10b = &on_stack.r10_bio;
4598 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4600 r10b->sector = r10_bio->sector;
4601 __raid10_find_phys(&conf->prev, r10b);
4606 int first_slot = slot;
4608 if (s > (PAGE_SIZE >> 9))
4612 int d = r10b->devs[slot].devnum;
4613 struct md_rdev *rdev = conf->mirrors[d].rdev;
4616 test_bit(Faulty, &rdev->flags) ||
4617 !test_bit(In_sync, &rdev->flags))
4620 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4621 success = sync_page_io(rdev,
4630 if (slot >= conf->copies)
4632 if (slot == first_slot)
4636 /* couldn't read this block, must give up */
4637 set_bit(MD_RECOVERY_INTR,
4647 static void end_reshape_write(struct bio *bio, int error)
4649 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4650 struct r10bio *r10_bio = bio->bi_private;
4651 struct mddev *mddev = r10_bio->mddev;
4652 struct r10conf *conf = mddev->private;
4656 struct md_rdev *rdev = NULL;
4658 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4660 rdev = conf->mirrors[d].replacement;
4663 rdev = conf->mirrors[d].rdev;
4667 /* FIXME should record badblock */
4668 md_error(mddev, rdev);
4671 rdev_dec_pending(rdev, mddev);
4672 end_reshape_request(r10_bio);
4675 static void end_reshape_request(struct r10bio *r10_bio)
4677 if (!atomic_dec_and_test(&r10_bio->remaining))
4679 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4680 bio_put(r10_bio->master_bio);
4684 static void raid10_finish_reshape(struct mddev *mddev)
4686 struct r10conf *conf = mddev->private;
4688 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4691 if (mddev->delta_disks > 0) {
4692 sector_t size = raid10_size(mddev, 0, 0);
4693 md_set_array_sectors(mddev, size);
4694 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4695 mddev->recovery_cp = mddev->resync_max_sectors;
4696 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4698 mddev->resync_max_sectors = size;
4699 set_capacity(mddev->gendisk, mddev->array_sectors);
4700 revalidate_disk(mddev->gendisk);
4703 for (d = conf->geo.raid_disks ;
4704 d < conf->geo.raid_disks - mddev->delta_disks;
4706 struct md_rdev *rdev = conf->mirrors[d].rdev;
4708 clear_bit(In_sync, &rdev->flags);
4709 rdev = conf->mirrors[d].replacement;
4711 clear_bit(In_sync, &rdev->flags);
4714 mddev->layout = mddev->new_layout;
4715 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4716 mddev->reshape_position = MaxSector;
4717 mddev->delta_disks = 0;
4718 mddev->reshape_backwards = 0;
4721 static struct md_personality raid10_personality =
4725 .owner = THIS_MODULE,
4726 .make_request = make_request,
4730 .error_handler = error,
4731 .hot_add_disk = raid10_add_disk,
4732 .hot_remove_disk= raid10_remove_disk,
4733 .spare_active = raid10_spare_active,
4734 .sync_request = sync_request,
4735 .quiesce = raid10_quiesce,
4736 .size = raid10_size,
4737 .resize = raid10_resize,
4738 .takeover = raid10_takeover,
4739 .check_reshape = raid10_check_reshape,
4740 .start_reshape = raid10_start_reshape,
4741 .finish_reshape = raid10_finish_reshape,
4744 static int __init raid_init(void)
4746 return register_md_personality(&raid10_personality);
4749 static void raid_exit(void)
4751 unregister_md_personality(&raid10_personality);
4754 module_init(raid_init);
4755 module_exit(raid_exit);
4756 MODULE_LICENSE("GPL");
4757 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4758 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4759 MODULE_ALIAS("md-raid10");
4760 MODULE_ALIAS("md-level-10");
4762 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);