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, int extra)
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+extra
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 (extra)
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+extra,
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_sectors(bio)
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_segments(bio) > 1)
1195 /* This is a one page bio that upper layers
1196 * refuse to split for us, so we need to split it.
1199 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1201 /* Each of these 'make_request' calls will call 'wait_barrier'.
1202 * If the first succeeds but the second blocks due to the resync
1203 * thread raising the barrier, we will deadlock because the
1204 * IO to the underlying device will be queued in generic_make_request
1205 * and will never complete, so will never reduce nr_pending.
1206 * So increment nr_waiting here so no new raise_barriers will
1207 * succeed, and so the second wait_barrier cannot block.
1209 spin_lock_irq(&conf->resync_lock);
1211 spin_unlock_irq(&conf->resync_lock);
1213 make_request(mddev, &bp->bio1);
1214 make_request(mddev, &bp->bio2);
1216 spin_lock_irq(&conf->resync_lock);
1218 wake_up(&conf->wait_barrier);
1219 spin_unlock_irq(&conf->resync_lock);
1221 bio_pair_release(bp);
1224 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1225 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1226 (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1232 md_write_start(mddev, bio);
1235 * Register the new request and wait if the reconstruction
1236 * thread has put up a bar for new requests.
1237 * Continue immediately if no resync is active currently.
1241 sectors = bio_sectors(bio);
1242 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1243 bio->bi_sector < conf->reshape_progress &&
1244 bio->bi_sector + sectors > conf->reshape_progress) {
1245 /* IO spans the reshape position. Need to wait for
1248 allow_barrier(conf);
1249 wait_event(conf->wait_barrier,
1250 conf->reshape_progress <= bio->bi_sector ||
1251 conf->reshape_progress >= bio->bi_sector + sectors);
1254 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1255 bio_data_dir(bio) == WRITE &&
1256 (mddev->reshape_backwards
1257 ? (bio->bi_sector < conf->reshape_safe &&
1258 bio->bi_sector + sectors > conf->reshape_progress)
1259 : (bio->bi_sector + sectors > conf->reshape_safe &&
1260 bio->bi_sector < conf->reshape_progress))) {
1261 /* Need to update reshape_position in metadata */
1262 mddev->reshape_position = conf->reshape_progress;
1263 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1264 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1265 md_wakeup_thread(mddev->thread);
1266 wait_event(mddev->sb_wait,
1267 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1269 conf->reshape_safe = mddev->reshape_position;
1272 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1274 r10_bio->master_bio = bio;
1275 r10_bio->sectors = sectors;
1277 r10_bio->mddev = mddev;
1278 r10_bio->sector = bio->bi_sector;
1281 /* We might need to issue multiple reads to different
1282 * devices if there are bad blocks around, so we keep
1283 * track of the number of reads in bio->bi_phys_segments.
1284 * If this is 0, there is only one r10_bio and no locking
1285 * will be needed when the request completes. If it is
1286 * non-zero, then it is the number of not-completed requests.
1288 bio->bi_phys_segments = 0;
1289 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1293 * read balancing logic:
1295 struct md_rdev *rdev;
1299 rdev = read_balance(conf, r10_bio, &max_sectors);
1301 raid_end_bio_io(r10_bio);
1304 slot = r10_bio->read_slot;
1306 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1307 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1310 r10_bio->devs[slot].bio = read_bio;
1311 r10_bio->devs[slot].rdev = rdev;
1313 read_bio->bi_sector = r10_bio->devs[slot].addr +
1314 choose_data_offset(r10_bio, rdev);
1315 read_bio->bi_bdev = rdev->bdev;
1316 read_bio->bi_end_io = raid10_end_read_request;
1317 read_bio->bi_rw = READ | do_sync;
1318 read_bio->bi_private = r10_bio;
1320 if (max_sectors < r10_bio->sectors) {
1321 /* Could not read all from this device, so we will
1322 * need another r10_bio.
1324 sectors_handled = (r10_bio->sectors + max_sectors
1326 r10_bio->sectors = max_sectors;
1327 spin_lock_irq(&conf->device_lock);
1328 if (bio->bi_phys_segments == 0)
1329 bio->bi_phys_segments = 2;
1331 bio->bi_phys_segments++;
1332 spin_unlock(&conf->device_lock);
1333 /* Cannot call generic_make_request directly
1334 * as that will be queued in __generic_make_request
1335 * and subsequent mempool_alloc might block
1336 * waiting for it. so hand bio over to raid10d.
1338 reschedule_retry(r10_bio);
1340 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1342 r10_bio->master_bio = bio;
1343 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1345 r10_bio->mddev = mddev;
1346 r10_bio->sector = bio->bi_sector + sectors_handled;
1349 generic_make_request(read_bio);
1356 if (conf->pending_count >= max_queued_requests) {
1357 md_wakeup_thread(mddev->thread);
1358 wait_event(conf->wait_barrier,
1359 conf->pending_count < max_queued_requests);
1361 /* first select target devices under rcu_lock and
1362 * inc refcount on their rdev. Record them by setting
1364 * If there are known/acknowledged bad blocks on any device
1365 * on which we have seen a write error, we want to avoid
1366 * writing to those blocks. This potentially requires several
1367 * writes to write around the bad blocks. Each set of writes
1368 * gets its own r10_bio with a set of bios attached. The number
1369 * of r10_bios is recored in bio->bi_phys_segments just as with
1373 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1374 raid10_find_phys(conf, r10_bio);
1376 blocked_rdev = NULL;
1378 max_sectors = r10_bio->sectors;
1380 for (i = 0; i < conf->copies; i++) {
1381 int d = r10_bio->devs[i].devnum;
1382 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1383 struct md_rdev *rrdev = rcu_dereference(
1384 conf->mirrors[d].replacement);
1387 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1388 atomic_inc(&rdev->nr_pending);
1389 blocked_rdev = rdev;
1392 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1393 atomic_inc(&rrdev->nr_pending);
1394 blocked_rdev = rrdev;
1397 if (rdev && (test_bit(Faulty, &rdev->flags)
1398 || test_bit(Unmerged, &rdev->flags)))
1400 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1401 || test_bit(Unmerged, &rrdev->flags)))
1404 r10_bio->devs[i].bio = NULL;
1405 r10_bio->devs[i].repl_bio = NULL;
1407 if (!rdev && !rrdev) {
1408 set_bit(R10BIO_Degraded, &r10_bio->state);
1411 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1413 sector_t dev_sector = r10_bio->devs[i].addr;
1417 is_bad = is_badblock(rdev, dev_sector,
1419 &first_bad, &bad_sectors);
1421 /* Mustn't write here until the bad block
1424 atomic_inc(&rdev->nr_pending);
1425 set_bit(BlockedBadBlocks, &rdev->flags);
1426 blocked_rdev = rdev;
1429 if (is_bad && first_bad <= dev_sector) {
1430 /* Cannot write here at all */
1431 bad_sectors -= (dev_sector - first_bad);
1432 if (bad_sectors < max_sectors)
1433 /* Mustn't write more than bad_sectors
1434 * to other devices yet
1436 max_sectors = bad_sectors;
1437 /* We don't set R10BIO_Degraded as that
1438 * only applies if the disk is missing,
1439 * so it might be re-added, and we want to
1440 * know to recover this chunk.
1441 * In this case the device is here, and the
1442 * fact that this chunk is not in-sync is
1443 * recorded in the bad block log.
1448 int good_sectors = first_bad - dev_sector;
1449 if (good_sectors < max_sectors)
1450 max_sectors = good_sectors;
1454 r10_bio->devs[i].bio = bio;
1455 atomic_inc(&rdev->nr_pending);
1458 r10_bio->devs[i].repl_bio = bio;
1459 atomic_inc(&rrdev->nr_pending);
1464 if (unlikely(blocked_rdev)) {
1465 /* Have to wait for this device to get unblocked, then retry */
1469 for (j = 0; j < i; j++) {
1470 if (r10_bio->devs[j].bio) {
1471 d = r10_bio->devs[j].devnum;
1472 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1474 if (r10_bio->devs[j].repl_bio) {
1475 struct md_rdev *rdev;
1476 d = r10_bio->devs[j].devnum;
1477 rdev = conf->mirrors[d].replacement;
1479 /* Race with remove_disk */
1481 rdev = conf->mirrors[d].rdev;
1483 rdev_dec_pending(rdev, mddev);
1486 allow_barrier(conf);
1487 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1492 if (max_sectors < r10_bio->sectors) {
1493 /* We are splitting this into multiple parts, so
1494 * we need to prepare for allocating another r10_bio.
1496 r10_bio->sectors = max_sectors;
1497 spin_lock_irq(&conf->device_lock);
1498 if (bio->bi_phys_segments == 0)
1499 bio->bi_phys_segments = 2;
1501 bio->bi_phys_segments++;
1502 spin_unlock_irq(&conf->device_lock);
1504 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1506 atomic_set(&r10_bio->remaining, 1);
1507 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1509 for (i = 0; i < conf->copies; i++) {
1511 int d = r10_bio->devs[i].devnum;
1512 if (r10_bio->devs[i].bio) {
1513 struct md_rdev *rdev = conf->mirrors[d].rdev;
1514 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1515 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1517 r10_bio->devs[i].bio = mbio;
1519 mbio->bi_sector = (r10_bio->devs[i].addr+
1520 choose_data_offset(r10_bio,
1522 mbio->bi_bdev = rdev->bdev;
1523 mbio->bi_end_io = raid10_end_write_request;
1525 WRITE | do_sync | do_fua | do_discard | do_same;
1526 mbio->bi_private = r10_bio;
1528 atomic_inc(&r10_bio->remaining);
1530 cb = blk_check_plugged(raid10_unplug, mddev,
1533 plug = container_of(cb, struct raid10_plug_cb,
1537 spin_lock_irqsave(&conf->device_lock, flags);
1539 bio_list_add(&plug->pending, mbio);
1540 plug->pending_cnt++;
1542 bio_list_add(&conf->pending_bio_list, mbio);
1543 conf->pending_count++;
1545 spin_unlock_irqrestore(&conf->device_lock, flags);
1547 md_wakeup_thread(mddev->thread);
1550 if (r10_bio->devs[i].repl_bio) {
1551 struct md_rdev *rdev = conf->mirrors[d].replacement;
1553 /* Replacement just got moved to main 'rdev' */
1555 rdev = conf->mirrors[d].rdev;
1557 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1558 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1560 r10_bio->devs[i].repl_bio = mbio;
1562 mbio->bi_sector = (r10_bio->devs[i].addr +
1565 mbio->bi_bdev = rdev->bdev;
1566 mbio->bi_end_io = raid10_end_write_request;
1568 WRITE | do_sync | do_fua | do_discard | do_same;
1569 mbio->bi_private = r10_bio;
1571 atomic_inc(&r10_bio->remaining);
1572 spin_lock_irqsave(&conf->device_lock, flags);
1573 bio_list_add(&conf->pending_bio_list, mbio);
1574 conf->pending_count++;
1575 spin_unlock_irqrestore(&conf->device_lock, flags);
1576 if (!mddev_check_plugged(mddev))
1577 md_wakeup_thread(mddev->thread);
1581 /* Don't remove the bias on 'remaining' (one_write_done) until
1582 * after checking if we need to go around again.
1585 if (sectors_handled < bio_sectors(bio)) {
1586 one_write_done(r10_bio);
1587 /* We need another r10_bio. It has already been counted
1588 * in bio->bi_phys_segments.
1590 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1592 r10_bio->master_bio = bio;
1593 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1595 r10_bio->mddev = mddev;
1596 r10_bio->sector = bio->bi_sector + sectors_handled;
1600 one_write_done(r10_bio);
1602 /* In case raid10d snuck in to freeze_array */
1603 wake_up(&conf->wait_barrier);
1606 static void status(struct seq_file *seq, struct mddev *mddev)
1608 struct r10conf *conf = mddev->private;
1611 if (conf->geo.near_copies < conf->geo.raid_disks)
1612 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1613 if (conf->geo.near_copies > 1)
1614 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1615 if (conf->geo.far_copies > 1) {
1616 if (conf->geo.far_offset)
1617 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1619 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1621 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1622 conf->geo.raid_disks - mddev->degraded);
1623 for (i = 0; i < conf->geo.raid_disks; i++)
1624 seq_printf(seq, "%s",
1625 conf->mirrors[i].rdev &&
1626 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1627 seq_printf(seq, "]");
1630 /* check if there are enough drives for
1631 * every block to appear on atleast one.
1632 * Don't consider the device numbered 'ignore'
1633 * as we might be about to remove it.
1635 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1640 int n = conf->copies;
1644 if (conf->mirrors[this].rdev &&
1647 this = (this+1) % geo->raid_disks;
1651 first = (first + geo->near_copies) % geo->raid_disks;
1652 } while (first != 0);
1656 static int enough(struct r10conf *conf, int ignore)
1658 return _enough(conf, &conf->geo, ignore) &&
1659 _enough(conf, &conf->prev, ignore);
1662 static void error(struct mddev *mddev, struct md_rdev *rdev)
1664 char b[BDEVNAME_SIZE];
1665 struct r10conf *conf = mddev->private;
1668 * If it is not operational, then we have already marked it as dead
1669 * else if it is the last working disks, ignore the error, let the
1670 * next level up know.
1671 * else mark the drive as failed
1673 if (test_bit(In_sync, &rdev->flags)
1674 && !enough(conf, rdev->raid_disk))
1676 * Don't fail the drive, just return an IO error.
1679 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1680 unsigned long flags;
1681 spin_lock_irqsave(&conf->device_lock, flags);
1683 spin_unlock_irqrestore(&conf->device_lock, flags);
1685 * if recovery is running, make sure it aborts.
1687 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1689 set_bit(Blocked, &rdev->flags);
1690 set_bit(Faulty, &rdev->flags);
1691 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1693 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1694 "md/raid10:%s: Operation continuing on %d devices.\n",
1695 mdname(mddev), bdevname(rdev->bdev, b),
1696 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1699 static void print_conf(struct r10conf *conf)
1702 struct raid10_info *tmp;
1704 printk(KERN_DEBUG "RAID10 conf printout:\n");
1706 printk(KERN_DEBUG "(!conf)\n");
1709 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1710 conf->geo.raid_disks);
1712 for (i = 0; i < conf->geo.raid_disks; i++) {
1713 char b[BDEVNAME_SIZE];
1714 tmp = conf->mirrors + i;
1716 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1717 i, !test_bit(In_sync, &tmp->rdev->flags),
1718 !test_bit(Faulty, &tmp->rdev->flags),
1719 bdevname(tmp->rdev->bdev,b));
1723 static void close_sync(struct r10conf *conf)
1726 allow_barrier(conf);
1728 mempool_destroy(conf->r10buf_pool);
1729 conf->r10buf_pool = NULL;
1732 static int raid10_spare_active(struct mddev *mddev)
1735 struct r10conf *conf = mddev->private;
1736 struct raid10_info *tmp;
1738 unsigned long flags;
1741 * Find all non-in_sync disks within the RAID10 configuration
1742 * and mark them in_sync
1744 for (i = 0; i < conf->geo.raid_disks; i++) {
1745 tmp = conf->mirrors + i;
1746 if (tmp->replacement
1747 && tmp->replacement->recovery_offset == MaxSector
1748 && !test_bit(Faulty, &tmp->replacement->flags)
1749 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1750 /* Replacement has just become active */
1752 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1755 /* Replaced device not technically faulty,
1756 * but we need to be sure it gets removed
1757 * and never re-added.
1759 set_bit(Faulty, &tmp->rdev->flags);
1760 sysfs_notify_dirent_safe(
1761 tmp->rdev->sysfs_state);
1763 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1764 } else if (tmp->rdev
1765 && !test_bit(Faulty, &tmp->rdev->flags)
1766 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1768 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1771 spin_lock_irqsave(&conf->device_lock, flags);
1772 mddev->degraded -= count;
1773 spin_unlock_irqrestore(&conf->device_lock, flags);
1780 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1782 struct r10conf *conf = mddev->private;
1786 int last = conf->geo.raid_disks - 1;
1787 struct request_queue *q = bdev_get_queue(rdev->bdev);
1789 if (mddev->recovery_cp < MaxSector)
1790 /* only hot-add to in-sync arrays, as recovery is
1791 * very different from resync
1794 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1797 if (rdev->raid_disk >= 0)
1798 first = last = rdev->raid_disk;
1800 if (q->merge_bvec_fn) {
1801 set_bit(Unmerged, &rdev->flags);
1802 mddev->merge_check_needed = 1;
1805 if (rdev->saved_raid_disk >= first &&
1806 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1807 mirror = rdev->saved_raid_disk;
1810 for ( ; mirror <= last ; mirror++) {
1811 struct raid10_info *p = &conf->mirrors[mirror];
1812 if (p->recovery_disabled == mddev->recovery_disabled)
1815 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1816 p->replacement != NULL)
1818 clear_bit(In_sync, &rdev->flags);
1819 set_bit(Replacement, &rdev->flags);
1820 rdev->raid_disk = mirror;
1822 disk_stack_limits(mddev->gendisk, rdev->bdev,
1823 rdev->data_offset << 9);
1825 rcu_assign_pointer(p->replacement, rdev);
1829 disk_stack_limits(mddev->gendisk, rdev->bdev,
1830 rdev->data_offset << 9);
1832 p->head_position = 0;
1833 p->recovery_disabled = mddev->recovery_disabled - 1;
1834 rdev->raid_disk = mirror;
1836 if (rdev->saved_raid_disk != mirror)
1838 rcu_assign_pointer(p->rdev, rdev);
1841 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1842 /* Some requests might not have seen this new
1843 * merge_bvec_fn. We must wait for them to complete
1844 * before merging the device fully.
1845 * First we make sure any code which has tested
1846 * our function has submitted the request, then
1847 * we wait for all outstanding requests to complete.
1849 synchronize_sched();
1850 freeze_array(conf, 0);
1851 unfreeze_array(conf);
1852 clear_bit(Unmerged, &rdev->flags);
1854 md_integrity_add_rdev(rdev, mddev);
1855 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1856 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1862 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1864 struct r10conf *conf = mddev->private;
1866 int number = rdev->raid_disk;
1867 struct md_rdev **rdevp;
1868 struct raid10_info *p = conf->mirrors + number;
1871 if (rdev == p->rdev)
1873 else if (rdev == p->replacement)
1874 rdevp = &p->replacement;
1878 if (test_bit(In_sync, &rdev->flags) ||
1879 atomic_read(&rdev->nr_pending)) {
1883 /* Only remove faulty devices if recovery
1886 if (!test_bit(Faulty, &rdev->flags) &&
1887 mddev->recovery_disabled != p->recovery_disabled &&
1888 (!p->replacement || p->replacement == rdev) &&
1889 number < conf->geo.raid_disks &&
1896 if (atomic_read(&rdev->nr_pending)) {
1897 /* lost the race, try later */
1901 } else if (p->replacement) {
1902 /* We must have just cleared 'rdev' */
1903 p->rdev = p->replacement;
1904 clear_bit(Replacement, &p->replacement->flags);
1905 smp_mb(); /* Make sure other CPUs may see both as identical
1906 * but will never see neither -- if they are careful.
1908 p->replacement = NULL;
1909 clear_bit(WantReplacement, &rdev->flags);
1911 /* We might have just remove the Replacement as faulty
1912 * Clear the flag just in case
1914 clear_bit(WantReplacement, &rdev->flags);
1916 err = md_integrity_register(mddev);
1925 static void end_sync_read(struct bio *bio, int error)
1927 struct r10bio *r10_bio = bio->bi_private;
1928 struct r10conf *conf = r10_bio->mddev->private;
1931 if (bio == r10_bio->master_bio) {
1932 /* this is a reshape read */
1933 d = r10_bio->read_slot; /* really the read dev */
1935 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1937 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1938 set_bit(R10BIO_Uptodate, &r10_bio->state);
1940 /* The write handler will notice the lack of
1941 * R10BIO_Uptodate and record any errors etc
1943 atomic_add(r10_bio->sectors,
1944 &conf->mirrors[d].rdev->corrected_errors);
1946 /* for reconstruct, we always reschedule after a read.
1947 * for resync, only after all reads
1949 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1950 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1951 atomic_dec_and_test(&r10_bio->remaining)) {
1952 /* we have read all the blocks,
1953 * do the comparison in process context in raid10d
1955 reschedule_retry(r10_bio);
1959 static void end_sync_request(struct r10bio *r10_bio)
1961 struct mddev *mddev = r10_bio->mddev;
1963 while (atomic_dec_and_test(&r10_bio->remaining)) {
1964 if (r10_bio->master_bio == NULL) {
1965 /* the primary of several recovery bios */
1966 sector_t s = r10_bio->sectors;
1967 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1968 test_bit(R10BIO_WriteError, &r10_bio->state))
1969 reschedule_retry(r10_bio);
1972 md_done_sync(mddev, s, 1);
1975 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1976 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1977 test_bit(R10BIO_WriteError, &r10_bio->state))
1978 reschedule_retry(r10_bio);
1986 static void end_sync_write(struct bio *bio, int error)
1988 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1989 struct r10bio *r10_bio = bio->bi_private;
1990 struct mddev *mddev = r10_bio->mddev;
1991 struct r10conf *conf = mddev->private;
1997 struct md_rdev *rdev = NULL;
1999 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2001 rdev = conf->mirrors[d].replacement;
2003 rdev = conf->mirrors[d].rdev;
2007 md_error(mddev, rdev);
2009 set_bit(WriteErrorSeen, &rdev->flags);
2010 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2011 set_bit(MD_RECOVERY_NEEDED,
2012 &rdev->mddev->recovery);
2013 set_bit(R10BIO_WriteError, &r10_bio->state);
2015 } else if (is_badblock(rdev,
2016 r10_bio->devs[slot].addr,
2018 &first_bad, &bad_sectors))
2019 set_bit(R10BIO_MadeGood, &r10_bio->state);
2021 rdev_dec_pending(rdev, mddev);
2023 end_sync_request(r10_bio);
2027 * Note: sync and recover and handled very differently for raid10
2028 * This code is for resync.
2029 * For resync, we read through virtual addresses and read all blocks.
2030 * If there is any error, we schedule a write. The lowest numbered
2031 * drive is authoritative.
2032 * However requests come for physical address, so we need to map.
2033 * For every physical address there are raid_disks/copies virtual addresses,
2034 * which is always are least one, but is not necessarly an integer.
2035 * This means that a physical address can span multiple chunks, so we may
2036 * have to submit multiple io requests for a single sync request.
2039 * We check if all blocks are in-sync and only write to blocks that
2042 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2044 struct r10conf *conf = mddev->private;
2046 struct bio *tbio, *fbio;
2049 atomic_set(&r10_bio->remaining, 1);
2051 /* find the first device with a block */
2052 for (i=0; i<conf->copies; i++)
2053 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2056 if (i == conf->copies)
2060 fbio = r10_bio->devs[i].bio;
2062 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2063 /* now find blocks with errors */
2064 for (i=0 ; i < conf->copies ; i++) {
2067 tbio = r10_bio->devs[i].bio;
2069 if (tbio->bi_end_io != end_sync_read)
2073 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2074 /* We know that the bi_io_vec layout is the same for
2075 * both 'first' and 'i', so we just compare them.
2076 * All vec entries are PAGE_SIZE;
2078 int sectors = r10_bio->sectors;
2079 for (j = 0; j < vcnt; j++) {
2080 int len = PAGE_SIZE;
2081 if (sectors < (len / 512))
2082 len = sectors * 512;
2083 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2084 page_address(tbio->bi_io_vec[j].bv_page),
2091 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2092 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2093 /* Don't fix anything. */
2096 /* Ok, we need to write this bio, either to correct an
2097 * inconsistency or to correct an unreadable block.
2098 * First we need to fixup bv_offset, bv_len and
2099 * bi_vecs, as the read request might have corrupted these
2103 tbio->bi_vcnt = vcnt;
2104 tbio->bi_size = r10_bio->sectors << 9;
2105 tbio->bi_rw = WRITE;
2106 tbio->bi_private = r10_bio;
2107 tbio->bi_sector = r10_bio->devs[i].addr;
2109 for (j=0; j < vcnt ; j++) {
2110 tbio->bi_io_vec[j].bv_offset = 0;
2111 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2113 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2114 page_address(fbio->bi_io_vec[j].bv_page),
2117 tbio->bi_end_io = end_sync_write;
2119 d = r10_bio->devs[i].devnum;
2120 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2121 atomic_inc(&r10_bio->remaining);
2122 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2124 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2125 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2126 generic_make_request(tbio);
2129 /* Now write out to any replacement devices
2132 for (i = 0; i < conf->copies; i++) {
2135 tbio = r10_bio->devs[i].repl_bio;
2136 if (!tbio || !tbio->bi_end_io)
2138 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2139 && r10_bio->devs[i].bio != fbio)
2140 for (j = 0; j < vcnt; j++)
2141 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2142 page_address(fbio->bi_io_vec[j].bv_page),
2144 d = r10_bio->devs[i].devnum;
2145 atomic_inc(&r10_bio->remaining);
2146 md_sync_acct(conf->mirrors[d].replacement->bdev,
2148 generic_make_request(tbio);
2152 if (atomic_dec_and_test(&r10_bio->remaining)) {
2153 md_done_sync(mddev, r10_bio->sectors, 1);
2159 * Now for the recovery code.
2160 * Recovery happens across physical sectors.
2161 * We recover all non-is_sync drives by finding the virtual address of
2162 * each, and then choose a working drive that also has that virt address.
2163 * There is a separate r10_bio for each non-in_sync drive.
2164 * Only the first two slots are in use. The first for reading,
2165 * The second for writing.
2168 static void fix_recovery_read_error(struct r10bio *r10_bio)
2170 /* We got a read error during recovery.
2171 * We repeat the read in smaller page-sized sections.
2172 * If a read succeeds, write it to the new device or record
2173 * a bad block if we cannot.
2174 * If a read fails, record a bad block on both old and
2177 struct mddev *mddev = r10_bio->mddev;
2178 struct r10conf *conf = mddev->private;
2179 struct bio *bio = r10_bio->devs[0].bio;
2181 int sectors = r10_bio->sectors;
2183 int dr = r10_bio->devs[0].devnum;
2184 int dw = r10_bio->devs[1].devnum;
2188 struct md_rdev *rdev;
2192 if (s > (PAGE_SIZE>>9))
2195 rdev = conf->mirrors[dr].rdev;
2196 addr = r10_bio->devs[0].addr + sect,
2197 ok = sync_page_io(rdev,
2200 bio->bi_io_vec[idx].bv_page,
2203 rdev = conf->mirrors[dw].rdev;
2204 addr = r10_bio->devs[1].addr + sect;
2205 ok = sync_page_io(rdev,
2208 bio->bi_io_vec[idx].bv_page,
2211 set_bit(WriteErrorSeen, &rdev->flags);
2212 if (!test_and_set_bit(WantReplacement,
2214 set_bit(MD_RECOVERY_NEEDED,
2215 &rdev->mddev->recovery);
2219 /* We don't worry if we cannot set a bad block -
2220 * it really is bad so there is no loss in not
2223 rdev_set_badblocks(rdev, addr, s, 0);
2225 if (rdev != conf->mirrors[dw].rdev) {
2226 /* need bad block on destination too */
2227 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2228 addr = r10_bio->devs[1].addr + sect;
2229 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2231 /* just abort the recovery */
2233 "md/raid10:%s: recovery aborted"
2234 " due to read error\n",
2237 conf->mirrors[dw].recovery_disabled
2238 = mddev->recovery_disabled;
2239 set_bit(MD_RECOVERY_INTR,
2252 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2254 struct r10conf *conf = mddev->private;
2256 struct bio *wbio, *wbio2;
2258 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2259 fix_recovery_read_error(r10_bio);
2260 end_sync_request(r10_bio);
2265 * share the pages with the first bio
2266 * and submit the write request
2268 d = r10_bio->devs[1].devnum;
2269 wbio = r10_bio->devs[1].bio;
2270 wbio2 = r10_bio->devs[1].repl_bio;
2271 /* Need to test wbio2->bi_end_io before we call
2272 * generic_make_request as if the former is NULL,
2273 * the latter is free to free wbio2.
2275 if (wbio2 && !wbio2->bi_end_io)
2277 if (wbio->bi_end_io) {
2278 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2279 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2280 generic_make_request(wbio);
2283 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2284 md_sync_acct(conf->mirrors[d].replacement->bdev,
2285 bio_sectors(wbio2));
2286 generic_make_request(wbio2);
2292 * Used by fix_read_error() to decay the per rdev read_errors.
2293 * We halve the read error count for every hour that has elapsed
2294 * since the last recorded read error.
2297 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2299 struct timespec cur_time_mon;
2300 unsigned long hours_since_last;
2301 unsigned int read_errors = atomic_read(&rdev->read_errors);
2303 ktime_get_ts(&cur_time_mon);
2305 if (rdev->last_read_error.tv_sec == 0 &&
2306 rdev->last_read_error.tv_nsec == 0) {
2307 /* first time we've seen a read error */
2308 rdev->last_read_error = cur_time_mon;
2312 hours_since_last = (cur_time_mon.tv_sec -
2313 rdev->last_read_error.tv_sec) / 3600;
2315 rdev->last_read_error = cur_time_mon;
2318 * if hours_since_last is > the number of bits in read_errors
2319 * just set read errors to 0. We do this to avoid
2320 * overflowing the shift of read_errors by hours_since_last.
2322 if (hours_since_last >= 8 * sizeof(read_errors))
2323 atomic_set(&rdev->read_errors, 0);
2325 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2328 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2329 int sectors, struct page *page, int rw)
2334 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2335 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2337 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2341 set_bit(WriteErrorSeen, &rdev->flags);
2342 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2343 set_bit(MD_RECOVERY_NEEDED,
2344 &rdev->mddev->recovery);
2346 /* need to record an error - either for the block or the device */
2347 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2348 md_error(rdev->mddev, rdev);
2353 * This is a kernel thread which:
2355 * 1. Retries failed read operations on working mirrors.
2356 * 2. Updates the raid superblock when problems encounter.
2357 * 3. Performs writes following reads for array synchronising.
2360 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2362 int sect = 0; /* Offset from r10_bio->sector */
2363 int sectors = r10_bio->sectors;
2364 struct md_rdev*rdev;
2365 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2366 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2368 /* still own a reference to this rdev, so it cannot
2369 * have been cleared recently.
2371 rdev = conf->mirrors[d].rdev;
2373 if (test_bit(Faulty, &rdev->flags))
2374 /* drive has already been failed, just ignore any
2375 more fix_read_error() attempts */
2378 check_decay_read_errors(mddev, rdev);
2379 atomic_inc(&rdev->read_errors);
2380 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2381 char b[BDEVNAME_SIZE];
2382 bdevname(rdev->bdev, b);
2385 "md/raid10:%s: %s: Raid device exceeded "
2386 "read_error threshold [cur %d:max %d]\n",
2388 atomic_read(&rdev->read_errors), max_read_errors);
2390 "md/raid10:%s: %s: Failing raid device\n",
2392 md_error(mddev, conf->mirrors[d].rdev);
2393 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2399 int sl = r10_bio->read_slot;
2403 if (s > (PAGE_SIZE>>9))
2411 d = r10_bio->devs[sl].devnum;
2412 rdev = rcu_dereference(conf->mirrors[d].rdev);
2414 !test_bit(Unmerged, &rdev->flags) &&
2415 test_bit(In_sync, &rdev->flags) &&
2416 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2417 &first_bad, &bad_sectors) == 0) {
2418 atomic_inc(&rdev->nr_pending);
2420 success = sync_page_io(rdev,
2421 r10_bio->devs[sl].addr +
2424 conf->tmppage, READ, false);
2425 rdev_dec_pending(rdev, mddev);
2431 if (sl == conf->copies)
2433 } while (!success && sl != r10_bio->read_slot);
2437 /* Cannot read from anywhere, just mark the block
2438 * as bad on the first device to discourage future
2441 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2442 rdev = conf->mirrors[dn].rdev;
2444 if (!rdev_set_badblocks(
2446 r10_bio->devs[r10_bio->read_slot].addr
2449 md_error(mddev, rdev);
2450 r10_bio->devs[r10_bio->read_slot].bio
2457 /* write it back and re-read */
2459 while (sl != r10_bio->read_slot) {
2460 char b[BDEVNAME_SIZE];
2465 d = r10_bio->devs[sl].devnum;
2466 rdev = rcu_dereference(conf->mirrors[d].rdev);
2468 test_bit(Unmerged, &rdev->flags) ||
2469 !test_bit(In_sync, &rdev->flags))
2472 atomic_inc(&rdev->nr_pending);
2474 if (r10_sync_page_io(rdev,
2475 r10_bio->devs[sl].addr +
2477 s, conf->tmppage, WRITE)
2479 /* Well, this device is dead */
2481 "md/raid10:%s: read correction "
2483 " (%d sectors at %llu on %s)\n",
2485 (unsigned long long)(
2487 choose_data_offset(r10_bio,
2489 bdevname(rdev->bdev, b));
2490 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2493 bdevname(rdev->bdev, b));
2495 rdev_dec_pending(rdev, mddev);
2499 while (sl != r10_bio->read_slot) {
2500 char b[BDEVNAME_SIZE];
2505 d = r10_bio->devs[sl].devnum;
2506 rdev = rcu_dereference(conf->mirrors[d].rdev);
2508 !test_bit(In_sync, &rdev->flags))
2511 atomic_inc(&rdev->nr_pending);
2513 switch (r10_sync_page_io(rdev,
2514 r10_bio->devs[sl].addr +
2519 /* Well, this device is dead */
2521 "md/raid10:%s: unable to read back "
2523 " (%d sectors at %llu on %s)\n",
2525 (unsigned long long)(
2527 choose_data_offset(r10_bio, rdev)),
2528 bdevname(rdev->bdev, b));
2529 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2532 bdevname(rdev->bdev, b));
2536 "md/raid10:%s: read error corrected"
2537 " (%d sectors at %llu on %s)\n",
2539 (unsigned long long)(
2541 choose_data_offset(r10_bio, rdev)),
2542 bdevname(rdev->bdev, b));
2543 atomic_add(s, &rdev->corrected_errors);
2546 rdev_dec_pending(rdev, mddev);
2556 static int narrow_write_error(struct r10bio *r10_bio, int i)
2558 struct bio *bio = r10_bio->master_bio;
2559 struct mddev *mddev = r10_bio->mddev;
2560 struct r10conf *conf = mddev->private;
2561 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2562 /* bio has the data to be written to slot 'i' where
2563 * we just recently had a write error.
2564 * We repeatedly clone the bio and trim down to one block,
2565 * then try the write. Where the write fails we record
2567 * It is conceivable that the bio doesn't exactly align with
2568 * blocks. We must handle this.
2570 * We currently own a reference to the rdev.
2576 int sect_to_write = r10_bio->sectors;
2579 if (rdev->badblocks.shift < 0)
2582 block_sectors = 1 << rdev->badblocks.shift;
2583 sector = r10_bio->sector;
2584 sectors = ((r10_bio->sector + block_sectors)
2585 & ~(sector_t)(block_sectors - 1))
2588 while (sect_to_write) {
2590 if (sectors > sect_to_write)
2591 sectors = sect_to_write;
2592 /* Write at 'sector' for 'sectors' */
2593 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2594 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2595 wbio->bi_sector = (r10_bio->devs[i].addr+
2596 choose_data_offset(r10_bio, rdev) +
2597 (sector - r10_bio->sector));
2598 wbio->bi_bdev = rdev->bdev;
2599 if (submit_bio_wait(WRITE, wbio) == 0)
2601 ok = rdev_set_badblocks(rdev, sector,
2606 sect_to_write -= sectors;
2608 sectors = block_sectors;
2613 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2615 int slot = r10_bio->read_slot;
2617 struct r10conf *conf = mddev->private;
2618 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2619 char b[BDEVNAME_SIZE];
2620 unsigned long do_sync;
2623 /* we got a read error. Maybe the drive is bad. Maybe just
2624 * the block and we can fix it.
2625 * We freeze all other IO, and try reading the block from
2626 * other devices. When we find one, we re-write
2627 * and check it that fixes the read error.
2628 * This is all done synchronously while the array is
2631 bio = r10_bio->devs[slot].bio;
2632 bdevname(bio->bi_bdev, b);
2634 r10_bio->devs[slot].bio = NULL;
2636 if (mddev->ro == 0) {
2637 freeze_array(conf, 1);
2638 fix_read_error(conf, mddev, r10_bio);
2639 unfreeze_array(conf);
2641 r10_bio->devs[slot].bio = IO_BLOCKED;
2643 rdev_dec_pending(rdev, mddev);
2646 rdev = read_balance(conf, r10_bio, &max_sectors);
2648 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2649 " read error for block %llu\n",
2651 (unsigned long long)r10_bio->sector);
2652 raid_end_bio_io(r10_bio);
2656 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2657 slot = r10_bio->read_slot;
2660 "md/raid10:%s: %s: redirecting "
2661 "sector %llu to another mirror\n",
2663 bdevname(rdev->bdev, b),
2664 (unsigned long long)r10_bio->sector);
2665 bio = bio_clone_mddev(r10_bio->master_bio,
2668 r10_bio->sector - bio->bi_sector,
2670 r10_bio->devs[slot].bio = bio;
2671 r10_bio->devs[slot].rdev = rdev;
2672 bio->bi_sector = r10_bio->devs[slot].addr
2673 + choose_data_offset(r10_bio, rdev);
2674 bio->bi_bdev = rdev->bdev;
2675 bio->bi_rw = READ | do_sync;
2676 bio->bi_private = r10_bio;
2677 bio->bi_end_io = raid10_end_read_request;
2678 if (max_sectors < r10_bio->sectors) {
2679 /* Drat - have to split this up more */
2680 struct bio *mbio = r10_bio->master_bio;
2681 int sectors_handled =
2682 r10_bio->sector + max_sectors
2684 r10_bio->sectors = max_sectors;
2685 spin_lock_irq(&conf->device_lock);
2686 if (mbio->bi_phys_segments == 0)
2687 mbio->bi_phys_segments = 2;
2689 mbio->bi_phys_segments++;
2690 spin_unlock_irq(&conf->device_lock);
2691 generic_make_request(bio);
2693 r10_bio = mempool_alloc(conf->r10bio_pool,
2695 r10_bio->master_bio = mbio;
2696 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2698 set_bit(R10BIO_ReadError,
2700 r10_bio->mddev = mddev;
2701 r10_bio->sector = mbio->bi_sector
2706 generic_make_request(bio);
2709 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2711 /* Some sort of write request has finished and it
2712 * succeeded in writing where we thought there was a
2713 * bad block. So forget the bad block.
2714 * Or possibly if failed and we need to record
2718 struct md_rdev *rdev;
2720 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2721 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2722 for (m = 0; m < conf->copies; m++) {
2723 int dev = r10_bio->devs[m].devnum;
2724 rdev = conf->mirrors[dev].rdev;
2725 if (r10_bio->devs[m].bio == NULL)
2727 if (test_bit(BIO_UPTODATE,
2728 &r10_bio->devs[m].bio->bi_flags)) {
2729 rdev_clear_badblocks(
2731 r10_bio->devs[m].addr,
2732 r10_bio->sectors, 0);
2734 if (!rdev_set_badblocks(
2736 r10_bio->devs[m].addr,
2737 r10_bio->sectors, 0))
2738 md_error(conf->mddev, rdev);
2740 rdev = conf->mirrors[dev].replacement;
2741 if (r10_bio->devs[m].repl_bio == NULL)
2743 if (test_bit(BIO_UPTODATE,
2744 &r10_bio->devs[m].repl_bio->bi_flags)) {
2745 rdev_clear_badblocks(
2747 r10_bio->devs[m].addr,
2748 r10_bio->sectors, 0);
2750 if (!rdev_set_badblocks(
2752 r10_bio->devs[m].addr,
2753 r10_bio->sectors, 0))
2754 md_error(conf->mddev, rdev);
2759 for (m = 0; m < conf->copies; m++) {
2760 int dev = r10_bio->devs[m].devnum;
2761 struct bio *bio = r10_bio->devs[m].bio;
2762 rdev = conf->mirrors[dev].rdev;
2763 if (bio == IO_MADE_GOOD) {
2764 rdev_clear_badblocks(
2766 r10_bio->devs[m].addr,
2767 r10_bio->sectors, 0);
2768 rdev_dec_pending(rdev, conf->mddev);
2769 } else if (bio != NULL &&
2770 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2771 if (!narrow_write_error(r10_bio, m)) {
2772 md_error(conf->mddev, rdev);
2773 set_bit(R10BIO_Degraded,
2776 rdev_dec_pending(rdev, conf->mddev);
2778 bio = r10_bio->devs[m].repl_bio;
2779 rdev = conf->mirrors[dev].replacement;
2780 if (rdev && bio == IO_MADE_GOOD) {
2781 rdev_clear_badblocks(
2783 r10_bio->devs[m].addr,
2784 r10_bio->sectors, 0);
2785 rdev_dec_pending(rdev, conf->mddev);
2788 if (test_bit(R10BIO_WriteError,
2790 close_write(r10_bio);
2791 raid_end_bio_io(r10_bio);
2795 static void raid10d(struct md_thread *thread)
2797 struct mddev *mddev = thread->mddev;
2798 struct r10bio *r10_bio;
2799 unsigned long flags;
2800 struct r10conf *conf = mddev->private;
2801 struct list_head *head = &conf->retry_list;
2802 struct blk_plug plug;
2804 md_check_recovery(mddev);
2806 blk_start_plug(&plug);
2809 flush_pending_writes(conf);
2811 spin_lock_irqsave(&conf->device_lock, flags);
2812 if (list_empty(head)) {
2813 spin_unlock_irqrestore(&conf->device_lock, flags);
2816 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2817 list_del(head->prev);
2819 spin_unlock_irqrestore(&conf->device_lock, flags);
2821 mddev = r10_bio->mddev;
2822 conf = mddev->private;
2823 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2824 test_bit(R10BIO_WriteError, &r10_bio->state))
2825 handle_write_completed(conf, r10_bio);
2826 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2827 reshape_request_write(mddev, r10_bio);
2828 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2829 sync_request_write(mddev, r10_bio);
2830 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2831 recovery_request_write(mddev, r10_bio);
2832 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2833 handle_read_error(mddev, r10_bio);
2835 /* just a partial read to be scheduled from a
2838 int slot = r10_bio->read_slot;
2839 generic_make_request(r10_bio->devs[slot].bio);
2843 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2844 md_check_recovery(mddev);
2846 blk_finish_plug(&plug);
2850 static int init_resync(struct r10conf *conf)
2855 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2856 BUG_ON(conf->r10buf_pool);
2857 conf->have_replacement = 0;
2858 for (i = 0; i < conf->geo.raid_disks; i++)
2859 if (conf->mirrors[i].replacement)
2860 conf->have_replacement = 1;
2861 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2862 if (!conf->r10buf_pool)
2864 conf->next_resync = 0;
2869 * perform a "sync" on one "block"
2871 * We need to make sure that no normal I/O request - particularly write
2872 * requests - conflict with active sync requests.
2874 * This is achieved by tracking pending requests and a 'barrier' concept
2875 * that can be installed to exclude normal IO requests.
2877 * Resync and recovery are handled very differently.
2878 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2880 * For resync, we iterate over virtual addresses, read all copies,
2881 * and update if there are differences. If only one copy is live,
2883 * For recovery, we iterate over physical addresses, read a good
2884 * value for each non-in_sync drive, and over-write.
2886 * So, for recovery we may have several outstanding complex requests for a
2887 * given address, one for each out-of-sync device. We model this by allocating
2888 * a number of r10_bio structures, one for each out-of-sync device.
2889 * As we setup these structures, we collect all bio's together into a list
2890 * which we then process collectively to add pages, and then process again
2891 * to pass to generic_make_request.
2893 * The r10_bio structures are linked using a borrowed master_bio pointer.
2894 * This link is counted in ->remaining. When the r10_bio that points to NULL
2895 * has its remaining count decremented to 0, the whole complex operation
2900 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2901 int *skipped, int go_faster)
2903 struct r10conf *conf = mddev->private;
2904 struct r10bio *r10_bio;
2905 struct bio *biolist = NULL, *bio;
2906 sector_t max_sector, nr_sectors;
2909 sector_t sync_blocks;
2910 sector_t sectors_skipped = 0;
2911 int chunks_skipped = 0;
2912 sector_t chunk_mask = conf->geo.chunk_mask;
2914 if (!conf->r10buf_pool)
2915 if (init_resync(conf))
2919 * Allow skipping a full rebuild for incremental assembly
2920 * of a clean array, like RAID1 does.
2922 if (mddev->bitmap == NULL &&
2923 mddev->recovery_cp == MaxSector &&
2924 mddev->reshape_position == MaxSector &&
2925 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2926 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2927 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2928 conf->fullsync == 0) {
2930 return mddev->dev_sectors - sector_nr;
2934 max_sector = mddev->dev_sectors;
2935 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2936 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2937 max_sector = mddev->resync_max_sectors;
2938 if (sector_nr >= max_sector) {
2939 /* If we aborted, we need to abort the
2940 * sync on the 'current' bitmap chucks (there can
2941 * be several when recovering multiple devices).
2942 * as we may have started syncing it but not finished.
2943 * We can find the current address in
2944 * mddev->curr_resync, but for recovery,
2945 * we need to convert that to several
2946 * virtual addresses.
2948 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2953 if (mddev->curr_resync < max_sector) { /* aborted */
2954 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2955 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2957 else for (i = 0; i < conf->geo.raid_disks; i++) {
2959 raid10_find_virt(conf, mddev->curr_resync, i);
2960 bitmap_end_sync(mddev->bitmap, sect,
2964 /* completed sync */
2965 if ((!mddev->bitmap || conf->fullsync)
2966 && conf->have_replacement
2967 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2968 /* Completed a full sync so the replacements
2969 * are now fully recovered.
2971 for (i = 0; i < conf->geo.raid_disks; i++)
2972 if (conf->mirrors[i].replacement)
2973 conf->mirrors[i].replacement
2979 bitmap_close_sync(mddev->bitmap);
2982 return sectors_skipped;
2985 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2986 return reshape_request(mddev, sector_nr, skipped);
2988 if (chunks_skipped >= conf->geo.raid_disks) {
2989 /* if there has been nothing to do on any drive,
2990 * then there is nothing to do at all..
2993 return (max_sector - sector_nr) + sectors_skipped;
2996 if (max_sector > mddev->resync_max)
2997 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2999 /* make sure whole request will fit in a chunk - if chunks
3002 if (conf->geo.near_copies < conf->geo.raid_disks &&
3003 max_sector > (sector_nr | chunk_mask))
3004 max_sector = (sector_nr | chunk_mask) + 1;
3006 * If there is non-resync activity waiting for us then
3007 * put in a delay to throttle resync.
3009 if (!go_faster && conf->nr_waiting)
3010 msleep_interruptible(1000);
3012 /* Again, very different code for resync and recovery.
3013 * Both must result in an r10bio with a list of bios that
3014 * have bi_end_io, bi_sector, bi_bdev set,
3015 * and bi_private set to the r10bio.
3016 * For recovery, we may actually create several r10bios
3017 * with 2 bios in each, that correspond to the bios in the main one.
3018 * In this case, the subordinate r10bios link back through a
3019 * borrowed master_bio pointer, and the counter in the master
3020 * includes a ref from each subordinate.
3022 /* First, we decide what to do and set ->bi_end_io
3023 * To end_sync_read if we want to read, and
3024 * end_sync_write if we will want to write.
3027 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3028 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3029 /* recovery... the complicated one */
3033 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3039 struct raid10_info *mirror = &conf->mirrors[i];
3041 if ((mirror->rdev == NULL ||
3042 test_bit(In_sync, &mirror->rdev->flags))
3044 (mirror->replacement == NULL ||
3046 &mirror->replacement->flags)))
3050 /* want to reconstruct this device */
3052 sect = raid10_find_virt(conf, sector_nr, i);
3053 if (sect >= mddev->resync_max_sectors) {
3054 /* last stripe is not complete - don't
3055 * try to recover this sector.
3059 /* Unless we are doing a full sync, or a replacement
3060 * we only need to recover the block if it is set in
3063 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3065 if (sync_blocks < max_sync)
3066 max_sync = sync_blocks;
3068 mirror->replacement == NULL &&
3070 /* yep, skip the sync_blocks here, but don't assume
3071 * that there will never be anything to do here
3073 chunks_skipped = -1;
3077 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3078 raise_barrier(conf, rb2 != NULL);
3079 atomic_set(&r10_bio->remaining, 0);
3081 r10_bio->master_bio = (struct bio*)rb2;
3083 atomic_inc(&rb2->remaining);
3084 r10_bio->mddev = mddev;
3085 set_bit(R10BIO_IsRecover, &r10_bio->state);
3086 r10_bio->sector = sect;
3088 raid10_find_phys(conf, r10_bio);
3090 /* Need to check if the array will still be
3093 for (j = 0; j < conf->geo.raid_disks; j++)
3094 if (conf->mirrors[j].rdev == NULL ||
3095 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3100 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3101 &sync_blocks, still_degraded);
3104 for (j=0; j<conf->copies;j++) {
3106 int d = r10_bio->devs[j].devnum;
3107 sector_t from_addr, to_addr;
3108 struct md_rdev *rdev;
3109 sector_t sector, first_bad;
3111 if (!conf->mirrors[d].rdev ||
3112 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3114 /* This is where we read from */
3116 rdev = conf->mirrors[d].rdev;
3117 sector = r10_bio->devs[j].addr;
3119 if (is_badblock(rdev, sector, max_sync,
3120 &first_bad, &bad_sectors)) {
3121 if (first_bad > sector)
3122 max_sync = first_bad - sector;
3124 bad_sectors -= (sector
3126 if (max_sync > bad_sectors)
3127 max_sync = bad_sectors;
3131 bio = r10_bio->devs[0].bio;
3133 bio->bi_next = biolist;
3135 bio->bi_private = r10_bio;
3136 bio->bi_end_io = end_sync_read;
3138 from_addr = r10_bio->devs[j].addr;
3139 bio->bi_sector = from_addr + rdev->data_offset;
3140 bio->bi_bdev = rdev->bdev;
3141 atomic_inc(&rdev->nr_pending);
3142 /* and we write to 'i' (if not in_sync) */
3144 for (k=0; k<conf->copies; k++)
3145 if (r10_bio->devs[k].devnum == i)
3147 BUG_ON(k == conf->copies);
3148 to_addr = r10_bio->devs[k].addr;
3149 r10_bio->devs[0].devnum = d;
3150 r10_bio->devs[0].addr = from_addr;
3151 r10_bio->devs[1].devnum = i;
3152 r10_bio->devs[1].addr = to_addr;
3154 rdev = mirror->rdev;
3155 if (!test_bit(In_sync, &rdev->flags)) {
3156 bio = r10_bio->devs[1].bio;
3158 bio->bi_next = biolist;
3160 bio->bi_private = r10_bio;
3161 bio->bi_end_io = end_sync_write;
3163 bio->bi_sector = to_addr
3164 + rdev->data_offset;
3165 bio->bi_bdev = rdev->bdev;
3166 atomic_inc(&r10_bio->remaining);
3168 r10_bio->devs[1].bio->bi_end_io = NULL;
3170 /* and maybe write to replacement */
3171 bio = r10_bio->devs[1].repl_bio;
3173 bio->bi_end_io = NULL;
3174 rdev = mirror->replacement;
3175 /* Note: if rdev != NULL, then bio
3176 * cannot be NULL as r10buf_pool_alloc will
3177 * have allocated it.
3178 * So the second test here is pointless.
3179 * But it keeps semantic-checkers happy, and
3180 * this comment keeps human reviewers
3183 if (rdev == NULL || bio == NULL ||
3184 test_bit(Faulty, &rdev->flags))
3187 bio->bi_next = biolist;
3189 bio->bi_private = r10_bio;
3190 bio->bi_end_io = end_sync_write;
3192 bio->bi_sector = to_addr + rdev->data_offset;
3193 bio->bi_bdev = rdev->bdev;
3194 atomic_inc(&r10_bio->remaining);
3197 if (j == conf->copies) {
3198 /* Cannot recover, so abort the recovery or
3199 * record a bad block */
3202 atomic_dec(&rb2->remaining);
3205 /* problem is that there are bad blocks
3206 * on other device(s)
3209 for (k = 0; k < conf->copies; k++)
3210 if (r10_bio->devs[k].devnum == i)
3212 if (!test_bit(In_sync,
3213 &mirror->rdev->flags)
3214 && !rdev_set_badblocks(
3216 r10_bio->devs[k].addr,
3219 if (mirror->replacement &&
3220 !rdev_set_badblocks(
3221 mirror->replacement,
3222 r10_bio->devs[k].addr,
3227 if (!test_and_set_bit(MD_RECOVERY_INTR,
3229 printk(KERN_INFO "md/raid10:%s: insufficient "
3230 "working devices for recovery.\n",
3232 mirror->recovery_disabled
3233 = mddev->recovery_disabled;
3238 if (biolist == NULL) {
3240 struct r10bio *rb2 = r10_bio;
3241 r10_bio = (struct r10bio*) rb2->master_bio;
3242 rb2->master_bio = NULL;
3248 /* resync. Schedule a read for every block at this virt offset */
3251 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3253 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3254 &sync_blocks, mddev->degraded) &&
3255 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3256 &mddev->recovery)) {
3257 /* We can skip this block */
3259 return sync_blocks + sectors_skipped;
3261 if (sync_blocks < max_sync)
3262 max_sync = sync_blocks;
3263 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3265 r10_bio->mddev = mddev;
3266 atomic_set(&r10_bio->remaining, 0);
3267 raise_barrier(conf, 0);
3268 conf->next_resync = sector_nr;
3270 r10_bio->master_bio = NULL;
3271 r10_bio->sector = sector_nr;
3272 set_bit(R10BIO_IsSync, &r10_bio->state);
3273 raid10_find_phys(conf, r10_bio);
3274 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3276 for (i = 0; i < conf->copies; i++) {
3277 int d = r10_bio->devs[i].devnum;
3278 sector_t first_bad, sector;
3281 if (r10_bio->devs[i].repl_bio)
3282 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3284 bio = r10_bio->devs[i].bio;
3286 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3287 if (conf->mirrors[d].rdev == NULL ||
3288 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3290 sector = r10_bio->devs[i].addr;
3291 if (is_badblock(conf->mirrors[d].rdev,
3293 &first_bad, &bad_sectors)) {
3294 if (first_bad > sector)
3295 max_sync = first_bad - sector;
3297 bad_sectors -= (sector - first_bad);
3298 if (max_sync > bad_sectors)
3299 max_sync = bad_sectors;
3303 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3304 atomic_inc(&r10_bio->remaining);
3305 bio->bi_next = biolist;
3307 bio->bi_private = r10_bio;
3308 bio->bi_end_io = end_sync_read;
3310 bio->bi_sector = sector +
3311 conf->mirrors[d].rdev->data_offset;
3312 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3315 if (conf->mirrors[d].replacement == NULL ||
3317 &conf->mirrors[d].replacement->flags))
3320 /* Need to set up for writing to the replacement */
3321 bio = r10_bio->devs[i].repl_bio;
3323 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3325 sector = r10_bio->devs[i].addr;
3326 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3327 bio->bi_next = biolist;
3329 bio->bi_private = r10_bio;
3330 bio->bi_end_io = end_sync_write;
3332 bio->bi_sector = sector +
3333 conf->mirrors[d].replacement->data_offset;
3334 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3339 for (i=0; i<conf->copies; i++) {
3340 int d = r10_bio->devs[i].devnum;
3341 if (r10_bio->devs[i].bio->bi_end_io)
3342 rdev_dec_pending(conf->mirrors[d].rdev,
3344 if (r10_bio->devs[i].repl_bio &&
3345 r10_bio->devs[i].repl_bio->bi_end_io)
3347 conf->mirrors[d].replacement,
3357 if (sector_nr + max_sync < max_sector)
3358 max_sector = sector_nr + max_sync;
3361 int len = PAGE_SIZE;
3362 if (sector_nr + (len>>9) > max_sector)
3363 len = (max_sector - sector_nr) << 9;
3366 for (bio= biolist ; bio ; bio=bio->bi_next) {
3368 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3369 if (bio_add_page(bio, page, len, 0))
3373 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3374 for (bio2 = biolist;
3375 bio2 && bio2 != bio;
3376 bio2 = bio2->bi_next) {
3377 /* remove last page from this bio */
3379 bio2->bi_size -= len;
3380 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3384 nr_sectors += len>>9;
3385 sector_nr += len>>9;
3386 } while (biolist->bi_vcnt < RESYNC_PAGES);
3388 r10_bio->sectors = nr_sectors;
3392 biolist = biolist->bi_next;
3394 bio->bi_next = NULL;
3395 r10_bio = bio->bi_private;
3396 r10_bio->sectors = nr_sectors;
3398 if (bio->bi_end_io == end_sync_read) {
3399 md_sync_acct(bio->bi_bdev, nr_sectors);
3400 set_bit(BIO_UPTODATE, &bio->bi_flags);
3401 generic_make_request(bio);
3405 if (sectors_skipped)
3406 /* pretend they weren't skipped, it makes
3407 * no important difference in this case
3409 md_done_sync(mddev, sectors_skipped, 1);
3411 return sectors_skipped + nr_sectors;
3413 /* There is nowhere to write, so all non-sync
3414 * drives must be failed or in resync, all drives
3415 * have a bad block, so try the next chunk...
3417 if (sector_nr + max_sync < max_sector)
3418 max_sector = sector_nr + max_sync;
3420 sectors_skipped += (max_sector - sector_nr);
3422 sector_nr = max_sector;
3427 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3430 struct r10conf *conf = mddev->private;
3433 raid_disks = min(conf->geo.raid_disks,
3434 conf->prev.raid_disks);
3436 sectors = conf->dev_sectors;
3438 size = sectors >> conf->geo.chunk_shift;
3439 sector_div(size, conf->geo.far_copies);
3440 size = size * raid_disks;
3441 sector_div(size, conf->geo.near_copies);
3443 return size << conf->geo.chunk_shift;
3446 static void calc_sectors(struct r10conf *conf, sector_t size)
3448 /* Calculate the number of sectors-per-device that will
3449 * actually be used, and set conf->dev_sectors and
3453 size = size >> conf->geo.chunk_shift;
3454 sector_div(size, conf->geo.far_copies);
3455 size = size * conf->geo.raid_disks;
3456 sector_div(size, conf->geo.near_copies);
3457 /* 'size' is now the number of chunks in the array */
3458 /* calculate "used chunks per device" */
3459 size = size * conf->copies;
3461 /* We need to round up when dividing by raid_disks to
3462 * get the stride size.
3464 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3466 conf->dev_sectors = size << conf->geo.chunk_shift;
3468 if (conf->geo.far_offset)
3469 conf->geo.stride = 1 << conf->geo.chunk_shift;
3471 sector_div(size, conf->geo.far_copies);
3472 conf->geo.stride = size << conf->geo.chunk_shift;
3476 enum geo_type {geo_new, geo_old, geo_start};
3477 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3480 int layout, chunk, disks;
3483 layout = mddev->layout;
3484 chunk = mddev->chunk_sectors;
3485 disks = mddev->raid_disks - mddev->delta_disks;
3488 layout = mddev->new_layout;
3489 chunk = mddev->new_chunk_sectors;
3490 disks = mddev->raid_disks;
3492 default: /* avoid 'may be unused' warnings */
3493 case geo_start: /* new when starting reshape - raid_disks not
3495 layout = mddev->new_layout;
3496 chunk = mddev->new_chunk_sectors;
3497 disks = mddev->raid_disks + mddev->delta_disks;
3502 if (chunk < (PAGE_SIZE >> 9) ||
3503 !is_power_of_2(chunk))
3506 fc = (layout >> 8) & 255;
3507 fo = layout & (1<<16);
3508 geo->raid_disks = disks;
3509 geo->near_copies = nc;
3510 geo->far_copies = fc;
3511 geo->far_offset = fo;
3512 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3513 geo->chunk_mask = chunk - 1;
3514 geo->chunk_shift = ffz(~chunk);
3518 static struct r10conf *setup_conf(struct mddev *mddev)
3520 struct r10conf *conf = NULL;
3525 copies = setup_geo(&geo, mddev, geo_new);
3528 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3529 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3530 mdname(mddev), PAGE_SIZE);
3534 if (copies < 2 || copies > mddev->raid_disks) {
3535 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3536 mdname(mddev), mddev->new_layout);
3541 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3545 /* FIXME calc properly */
3546 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3547 max(0,-mddev->delta_disks)),
3552 conf->tmppage = alloc_page(GFP_KERNEL);
3557 conf->copies = copies;
3558 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3559 r10bio_pool_free, conf);
3560 if (!conf->r10bio_pool)
3563 calc_sectors(conf, mddev->dev_sectors);
3564 if (mddev->reshape_position == MaxSector) {
3565 conf->prev = conf->geo;
3566 conf->reshape_progress = MaxSector;
3568 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3572 conf->reshape_progress = mddev->reshape_position;
3573 if (conf->prev.far_offset)
3574 conf->prev.stride = 1 << conf->prev.chunk_shift;
3576 /* far_copies must be 1 */
3577 conf->prev.stride = conf->dev_sectors;
3579 spin_lock_init(&conf->device_lock);
3580 INIT_LIST_HEAD(&conf->retry_list);
3582 spin_lock_init(&conf->resync_lock);
3583 init_waitqueue_head(&conf->wait_barrier);
3585 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3589 conf->mddev = mddev;
3594 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3597 if (conf->r10bio_pool)
3598 mempool_destroy(conf->r10bio_pool);
3599 kfree(conf->mirrors);
3600 safe_put_page(conf->tmppage);
3603 return ERR_PTR(err);
3606 static int run(struct mddev *mddev)
3608 struct r10conf *conf;
3609 int i, disk_idx, chunk_size;
3610 struct raid10_info *disk;
3611 struct md_rdev *rdev;
3613 sector_t min_offset_diff = 0;
3615 bool discard_supported = false;
3617 if (mddev->private == NULL) {
3618 conf = setup_conf(mddev);
3620 return PTR_ERR(conf);
3621 mddev->private = conf;
3623 conf = mddev->private;
3627 mddev->thread = conf->thread;
3628 conf->thread = NULL;
3630 chunk_size = mddev->chunk_sectors << 9;
3632 blk_queue_max_discard_sectors(mddev->queue,
3633 mddev->chunk_sectors);
3634 blk_queue_max_write_same_sectors(mddev->queue, 0);
3635 blk_queue_io_min(mddev->queue, chunk_size);
3636 if (conf->geo.raid_disks % conf->geo.near_copies)
3637 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3639 blk_queue_io_opt(mddev->queue, chunk_size *
3640 (conf->geo.raid_disks / conf->geo.near_copies));
3643 rdev_for_each(rdev, mddev) {
3645 struct request_queue *q;
3647 disk_idx = rdev->raid_disk;
3650 if (disk_idx >= conf->geo.raid_disks &&
3651 disk_idx >= conf->prev.raid_disks)
3653 disk = conf->mirrors + disk_idx;
3655 if (test_bit(Replacement, &rdev->flags)) {
3656 if (disk->replacement)
3658 disk->replacement = rdev;
3664 q = bdev_get_queue(rdev->bdev);
3665 if (q->merge_bvec_fn)
3666 mddev->merge_check_needed = 1;
3667 diff = (rdev->new_data_offset - rdev->data_offset);
3668 if (!mddev->reshape_backwards)
3672 if (first || diff < min_offset_diff)
3673 min_offset_diff = diff;
3676 disk_stack_limits(mddev->gendisk, rdev->bdev,
3677 rdev->data_offset << 9);
3679 disk->head_position = 0;
3681 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3682 discard_supported = true;
3686 if (discard_supported)
3687 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3690 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3693 /* need to check that every block has at least one working mirror */
3694 if (!enough(conf, -1)) {
3695 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3700 if (conf->reshape_progress != MaxSector) {
3701 /* must ensure that shape change is supported */
3702 if (conf->geo.far_copies != 1 &&
3703 conf->geo.far_offset == 0)
3705 if (conf->prev.far_copies != 1 &&
3706 conf->prev.far_offset == 0)
3710 mddev->degraded = 0;
3712 i < conf->geo.raid_disks
3713 || i < conf->prev.raid_disks;
3716 disk = conf->mirrors + i;
3718 if (!disk->rdev && disk->replacement) {
3719 /* The replacement is all we have - use it */
3720 disk->rdev = disk->replacement;
3721 disk->replacement = NULL;
3722 clear_bit(Replacement, &disk->rdev->flags);
3726 !test_bit(In_sync, &disk->rdev->flags)) {
3727 disk->head_position = 0;
3732 disk->recovery_disabled = mddev->recovery_disabled - 1;
3735 if (mddev->recovery_cp != MaxSector)
3736 printk(KERN_NOTICE "md/raid10:%s: not clean"
3737 " -- starting background reconstruction\n",
3740 "md/raid10:%s: active with %d out of %d devices\n",
3741 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3742 conf->geo.raid_disks);
3744 * Ok, everything is just fine now
3746 mddev->dev_sectors = conf->dev_sectors;
3747 size = raid10_size(mddev, 0, 0);
3748 md_set_array_sectors(mddev, size);
3749 mddev->resync_max_sectors = size;
3752 int stripe = conf->geo.raid_disks *
3753 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3754 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3755 mddev->queue->backing_dev_info.congested_data = mddev;
3757 /* Calculate max read-ahead size.
3758 * We need to readahead at least twice a whole stripe....
3761 stripe /= conf->geo.near_copies;
3762 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3763 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3764 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3768 if (md_integrity_register(mddev))
3771 if (conf->reshape_progress != MaxSector) {
3772 unsigned long before_length, after_length;
3774 before_length = ((1 << conf->prev.chunk_shift) *
3775 conf->prev.far_copies);
3776 after_length = ((1 << conf->geo.chunk_shift) *
3777 conf->geo.far_copies);
3779 if (max(before_length, after_length) > min_offset_diff) {
3780 /* This cannot work */
3781 printk("md/raid10: offset difference not enough to continue reshape\n");
3784 conf->offset_diff = min_offset_diff;
3786 conf->reshape_safe = conf->reshape_progress;
3787 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3788 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3789 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3790 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3791 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3798 md_unregister_thread(&mddev->thread);
3799 if (conf->r10bio_pool)
3800 mempool_destroy(conf->r10bio_pool);
3801 safe_put_page(conf->tmppage);
3802 kfree(conf->mirrors);
3804 mddev->private = NULL;
3809 static int stop(struct mddev *mddev)
3811 struct r10conf *conf = mddev->private;
3813 raise_barrier(conf, 0);
3814 lower_barrier(conf);
3816 md_unregister_thread(&mddev->thread);
3818 /* the unplug fn references 'conf'*/
3819 blk_sync_queue(mddev->queue);
3821 if (conf->r10bio_pool)
3822 mempool_destroy(conf->r10bio_pool);
3823 safe_put_page(conf->tmppage);
3824 kfree(conf->mirrors);
3826 mddev->private = NULL;
3830 static void raid10_quiesce(struct mddev *mddev, int state)
3832 struct r10conf *conf = mddev->private;
3836 raise_barrier(conf, 0);
3839 lower_barrier(conf);
3844 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3846 /* Resize of 'far' arrays is not supported.
3847 * For 'near' and 'offset' arrays we can set the
3848 * number of sectors used to be an appropriate multiple
3849 * of the chunk size.
3850 * For 'offset', this is far_copies*chunksize.
3851 * For 'near' the multiplier is the LCM of
3852 * near_copies and raid_disks.
3853 * So if far_copies > 1 && !far_offset, fail.
3854 * Else find LCM(raid_disks, near_copy)*far_copies and
3855 * multiply by chunk_size. Then round to this number.
3856 * This is mostly done by raid10_size()
3858 struct r10conf *conf = mddev->private;
3859 sector_t oldsize, size;
3861 if (mddev->reshape_position != MaxSector)
3864 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3867 oldsize = raid10_size(mddev, 0, 0);
3868 size = raid10_size(mddev, sectors, 0);
3869 if (mddev->external_size &&
3870 mddev->array_sectors > size)
3872 if (mddev->bitmap) {
3873 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3877 md_set_array_sectors(mddev, size);
3878 set_capacity(mddev->gendisk, mddev->array_sectors);
3879 revalidate_disk(mddev->gendisk);
3880 if (sectors > mddev->dev_sectors &&
3881 mddev->recovery_cp > oldsize) {
3882 mddev->recovery_cp = oldsize;
3883 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3885 calc_sectors(conf, sectors);
3886 mddev->dev_sectors = conf->dev_sectors;
3887 mddev->resync_max_sectors = size;
3891 static void *raid10_takeover_raid0(struct mddev *mddev)
3893 struct md_rdev *rdev;
3894 struct r10conf *conf;
3896 if (mddev->degraded > 0) {
3897 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3899 return ERR_PTR(-EINVAL);
3902 /* Set new parameters */
3903 mddev->new_level = 10;
3904 /* new layout: far_copies = 1, near_copies = 2 */
3905 mddev->new_layout = (1<<8) + 2;
3906 mddev->new_chunk_sectors = mddev->chunk_sectors;
3907 mddev->delta_disks = mddev->raid_disks;
3908 mddev->raid_disks *= 2;
3909 /* make sure it will be not marked as dirty */
3910 mddev->recovery_cp = MaxSector;
3912 conf = setup_conf(mddev);
3913 if (!IS_ERR(conf)) {
3914 rdev_for_each(rdev, mddev)
3915 if (rdev->raid_disk >= 0)
3916 rdev->new_raid_disk = rdev->raid_disk * 2;
3923 static void *raid10_takeover(struct mddev *mddev)
3925 struct r0conf *raid0_conf;
3927 /* raid10 can take over:
3928 * raid0 - providing it has only two drives
3930 if (mddev->level == 0) {
3931 /* for raid0 takeover only one zone is supported */
3932 raid0_conf = mddev->private;
3933 if (raid0_conf->nr_strip_zones > 1) {
3934 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3935 " with more than one zone.\n",
3937 return ERR_PTR(-EINVAL);
3939 return raid10_takeover_raid0(mddev);
3941 return ERR_PTR(-EINVAL);
3944 static int raid10_check_reshape(struct mddev *mddev)
3946 /* Called when there is a request to change
3947 * - layout (to ->new_layout)
3948 * - chunk size (to ->new_chunk_sectors)
3949 * - raid_disks (by delta_disks)
3950 * or when trying to restart a reshape that was ongoing.
3952 * We need to validate the request and possibly allocate
3953 * space if that might be an issue later.
3955 * Currently we reject any reshape of a 'far' mode array,
3956 * allow chunk size to change if new is generally acceptable,
3957 * allow raid_disks to increase, and allow
3958 * a switch between 'near' mode and 'offset' mode.
3960 struct r10conf *conf = mddev->private;
3963 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3966 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3967 /* mustn't change number of copies */
3969 if (geo.far_copies > 1 && !geo.far_offset)
3970 /* Cannot switch to 'far' mode */
3973 if (mddev->array_sectors & geo.chunk_mask)
3974 /* not factor of array size */
3977 if (!enough(conf, -1))
3980 kfree(conf->mirrors_new);
3981 conf->mirrors_new = NULL;
3982 if (mddev->delta_disks > 0) {
3983 /* allocate new 'mirrors' list */
3984 conf->mirrors_new = kzalloc(
3985 sizeof(struct raid10_info)
3986 *(mddev->raid_disks +
3987 mddev->delta_disks),
3989 if (!conf->mirrors_new)
3996 * Need to check if array has failed when deciding whether to:
3998 * - remove non-faulty devices
4001 * This determination is simple when no reshape is happening.
4002 * However if there is a reshape, we need to carefully check
4003 * both the before and after sections.
4004 * This is because some failed devices may only affect one
4005 * of the two sections, and some non-in_sync devices may
4006 * be insync in the section most affected by failed devices.
4008 static int calc_degraded(struct r10conf *conf)
4010 int degraded, degraded2;
4015 /* 'prev' section first */
4016 for (i = 0; i < conf->prev.raid_disks; i++) {
4017 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4018 if (!rdev || test_bit(Faulty, &rdev->flags))
4020 else if (!test_bit(In_sync, &rdev->flags))
4021 /* When we can reduce the number of devices in
4022 * an array, this might not contribute to
4023 * 'degraded'. It does now.
4028 if (conf->geo.raid_disks == conf->prev.raid_disks)
4032 for (i = 0; i < conf->geo.raid_disks; i++) {
4033 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4034 if (!rdev || test_bit(Faulty, &rdev->flags))
4036 else if (!test_bit(In_sync, &rdev->flags)) {
4037 /* If reshape is increasing the number of devices,
4038 * this section has already been recovered, so
4039 * it doesn't contribute to degraded.
4042 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4047 if (degraded2 > degraded)
4052 static int raid10_start_reshape(struct mddev *mddev)
4054 /* A 'reshape' has been requested. This commits
4055 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4056 * This also checks if there are enough spares and adds them
4058 * We currently require enough spares to make the final
4059 * array non-degraded. We also require that the difference
4060 * between old and new data_offset - on each device - is
4061 * enough that we never risk over-writing.
4064 unsigned long before_length, after_length;
4065 sector_t min_offset_diff = 0;
4068 struct r10conf *conf = mddev->private;
4069 struct md_rdev *rdev;
4073 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4076 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4079 before_length = ((1 << conf->prev.chunk_shift) *
4080 conf->prev.far_copies);
4081 after_length = ((1 << conf->geo.chunk_shift) *
4082 conf->geo.far_copies);
4084 rdev_for_each(rdev, mddev) {
4085 if (!test_bit(In_sync, &rdev->flags)
4086 && !test_bit(Faulty, &rdev->flags))
4088 if (rdev->raid_disk >= 0) {
4089 long long diff = (rdev->new_data_offset
4090 - rdev->data_offset);
4091 if (!mddev->reshape_backwards)
4095 if (first || diff < min_offset_diff)
4096 min_offset_diff = diff;
4100 if (max(before_length, after_length) > min_offset_diff)
4103 if (spares < mddev->delta_disks)
4106 conf->offset_diff = min_offset_diff;
4107 spin_lock_irq(&conf->device_lock);
4108 if (conf->mirrors_new) {
4109 memcpy(conf->mirrors_new, conf->mirrors,
4110 sizeof(struct raid10_info)*conf->prev.raid_disks);
4112 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4113 conf->mirrors_old = conf->mirrors;
4114 conf->mirrors = conf->mirrors_new;
4115 conf->mirrors_new = NULL;
4117 setup_geo(&conf->geo, mddev, geo_start);
4119 if (mddev->reshape_backwards) {
4120 sector_t size = raid10_size(mddev, 0, 0);
4121 if (size < mddev->array_sectors) {
4122 spin_unlock_irq(&conf->device_lock);
4123 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4127 mddev->resync_max_sectors = size;
4128 conf->reshape_progress = size;
4130 conf->reshape_progress = 0;
4131 spin_unlock_irq(&conf->device_lock);
4133 if (mddev->delta_disks && mddev->bitmap) {
4134 ret = bitmap_resize(mddev->bitmap,
4135 raid10_size(mddev, 0,
4136 conf->geo.raid_disks),
4141 if (mddev->delta_disks > 0) {
4142 rdev_for_each(rdev, mddev)
4143 if (rdev->raid_disk < 0 &&
4144 !test_bit(Faulty, &rdev->flags)) {
4145 if (raid10_add_disk(mddev, rdev) == 0) {
4146 if (rdev->raid_disk >=
4147 conf->prev.raid_disks)
4148 set_bit(In_sync, &rdev->flags);
4150 rdev->recovery_offset = 0;
4152 if (sysfs_link_rdev(mddev, rdev))
4153 /* Failure here is OK */;
4155 } else if (rdev->raid_disk >= conf->prev.raid_disks
4156 && !test_bit(Faulty, &rdev->flags)) {
4157 /* This is a spare that was manually added */
4158 set_bit(In_sync, &rdev->flags);
4161 /* When a reshape changes the number of devices,
4162 * ->degraded is measured against the larger of the
4163 * pre and post numbers.
4165 spin_lock_irq(&conf->device_lock);
4166 mddev->degraded = calc_degraded(conf);
4167 spin_unlock_irq(&conf->device_lock);
4168 mddev->raid_disks = conf->geo.raid_disks;
4169 mddev->reshape_position = conf->reshape_progress;
4170 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4172 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4173 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4174 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4175 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4177 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4179 if (!mddev->sync_thread) {
4183 conf->reshape_checkpoint = jiffies;
4184 md_wakeup_thread(mddev->sync_thread);
4185 md_new_event(mddev);
4189 mddev->recovery = 0;
4190 spin_lock_irq(&conf->device_lock);
4191 conf->geo = conf->prev;
4192 mddev->raid_disks = conf->geo.raid_disks;
4193 rdev_for_each(rdev, mddev)
4194 rdev->new_data_offset = rdev->data_offset;
4196 conf->reshape_progress = MaxSector;
4197 mddev->reshape_position = MaxSector;
4198 spin_unlock_irq(&conf->device_lock);
4202 /* Calculate the last device-address that could contain
4203 * any block from the chunk that includes the array-address 's'
4204 * and report the next address.
4205 * i.e. the address returned will be chunk-aligned and after
4206 * any data that is in the chunk containing 's'.
4208 static sector_t last_dev_address(sector_t s, struct geom *geo)
4210 s = (s | geo->chunk_mask) + 1;
4211 s >>= geo->chunk_shift;
4212 s *= geo->near_copies;
4213 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4214 s *= geo->far_copies;
4215 s <<= geo->chunk_shift;
4219 /* Calculate the first device-address that could contain
4220 * any block from the chunk that includes the array-address 's'.
4221 * This too will be the start of a chunk
4223 static sector_t first_dev_address(sector_t s, struct geom *geo)
4225 s >>= geo->chunk_shift;
4226 s *= geo->near_copies;
4227 sector_div(s, geo->raid_disks);
4228 s *= geo->far_copies;
4229 s <<= geo->chunk_shift;
4233 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4236 /* We simply copy at most one chunk (smallest of old and new)
4237 * at a time, possibly less if that exceeds RESYNC_PAGES,
4238 * or we hit a bad block or something.
4239 * This might mean we pause for normal IO in the middle of
4240 * a chunk, but that is not a problem was mddev->reshape_position
4241 * can record any location.
4243 * If we will want to write to a location that isn't
4244 * yet recorded as 'safe' (i.e. in metadata on disk) then
4245 * we need to flush all reshape requests and update the metadata.
4247 * When reshaping forwards (e.g. to more devices), we interpret
4248 * 'safe' as the earliest block which might not have been copied
4249 * down yet. We divide this by previous stripe size and multiply
4250 * by previous stripe length to get lowest device offset that we
4251 * cannot write to yet.
4252 * We interpret 'sector_nr' as an address that we want to write to.
4253 * From this we use last_device_address() to find where we might
4254 * write to, and first_device_address on the 'safe' position.
4255 * If this 'next' write position is after the 'safe' position,
4256 * we must update the metadata to increase the 'safe' position.
4258 * When reshaping backwards, we round in the opposite direction
4259 * and perform the reverse test: next write position must not be
4260 * less than current safe position.
4262 * In all this the minimum difference in data offsets
4263 * (conf->offset_diff - always positive) allows a bit of slack,
4264 * so next can be after 'safe', but not by more than offset_disk
4266 * We need to prepare all the bios here before we start any IO
4267 * to ensure the size we choose is acceptable to all devices.
4268 * The means one for each copy for write-out and an extra one for
4270 * We store the read-in bio in ->master_bio and the others in
4271 * ->devs[x].bio and ->devs[x].repl_bio.
4273 struct r10conf *conf = mddev->private;
4274 struct r10bio *r10_bio;
4275 sector_t next, safe, last;
4279 struct md_rdev *rdev;
4282 struct bio *bio, *read_bio;
4283 int sectors_done = 0;
4285 if (sector_nr == 0) {
4286 /* If restarting in the middle, skip the initial sectors */
4287 if (mddev->reshape_backwards &&
4288 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4289 sector_nr = (raid10_size(mddev, 0, 0)
4290 - conf->reshape_progress);
4291 } else if (!mddev->reshape_backwards &&
4292 conf->reshape_progress > 0)
4293 sector_nr = conf->reshape_progress;
4295 mddev->curr_resync_completed = sector_nr;
4296 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4302 /* We don't use sector_nr to track where we are up to
4303 * as that doesn't work well for ->reshape_backwards.
4304 * So just use ->reshape_progress.
4306 if (mddev->reshape_backwards) {
4307 /* 'next' is the earliest device address that we might
4308 * write to for this chunk in the new layout
4310 next = first_dev_address(conf->reshape_progress - 1,
4313 /* 'safe' is the last device address that we might read from
4314 * in the old layout after a restart
4316 safe = last_dev_address(conf->reshape_safe - 1,
4319 if (next + conf->offset_diff < safe)
4322 last = conf->reshape_progress - 1;
4323 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4324 & conf->prev.chunk_mask);
4325 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4326 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4328 /* 'next' is after the last device address that we
4329 * might write to for this chunk in the new layout
4331 next = last_dev_address(conf->reshape_progress, &conf->geo);
4333 /* 'safe' is the earliest device address that we might
4334 * read from in the old layout after a restart
4336 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4338 /* Need to update metadata if 'next' might be beyond 'safe'
4339 * as that would possibly corrupt data
4341 if (next > safe + conf->offset_diff)
4344 sector_nr = conf->reshape_progress;
4345 last = sector_nr | (conf->geo.chunk_mask
4346 & conf->prev.chunk_mask);
4348 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4349 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4353 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4354 /* Need to update reshape_position in metadata */
4356 mddev->reshape_position = conf->reshape_progress;
4357 if (mddev->reshape_backwards)
4358 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4359 - conf->reshape_progress;
4361 mddev->curr_resync_completed = conf->reshape_progress;
4362 conf->reshape_checkpoint = jiffies;
4363 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4364 md_wakeup_thread(mddev->thread);
4365 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4366 kthread_should_stop());
4367 conf->reshape_safe = mddev->reshape_position;
4368 allow_barrier(conf);
4372 /* Now schedule reads for blocks from sector_nr to last */
4373 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4374 raise_barrier(conf, sectors_done != 0);
4375 atomic_set(&r10_bio->remaining, 0);
4376 r10_bio->mddev = mddev;
4377 r10_bio->sector = sector_nr;
4378 set_bit(R10BIO_IsReshape, &r10_bio->state);
4379 r10_bio->sectors = last - sector_nr + 1;
4380 rdev = read_balance(conf, r10_bio, &max_sectors);
4381 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4384 /* Cannot read from here, so need to record bad blocks
4385 * on all the target devices.
4388 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4389 return sectors_done;
4392 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4394 read_bio->bi_bdev = rdev->bdev;
4395 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4396 + rdev->data_offset);
4397 read_bio->bi_private = r10_bio;
4398 read_bio->bi_end_io = end_sync_read;
4399 read_bio->bi_rw = READ;
4400 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4401 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4402 read_bio->bi_vcnt = 0;
4403 read_bio->bi_size = 0;
4404 r10_bio->master_bio = read_bio;
4405 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4407 /* Now find the locations in the new layout */
4408 __raid10_find_phys(&conf->geo, r10_bio);
4411 read_bio->bi_next = NULL;
4413 for (s = 0; s < conf->copies*2; s++) {
4415 int d = r10_bio->devs[s/2].devnum;
4416 struct md_rdev *rdev2;
4418 rdev2 = conf->mirrors[d].replacement;
4419 b = r10_bio->devs[s/2].repl_bio;
4421 rdev2 = conf->mirrors[d].rdev;
4422 b = r10_bio->devs[s/2].bio;
4424 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4428 b->bi_bdev = rdev2->bdev;
4429 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4430 b->bi_private = r10_bio;
4431 b->bi_end_io = end_reshape_write;
4437 /* Now add as many pages as possible to all of these bios. */
4440 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4441 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4442 int len = (max_sectors - s) << 9;
4443 if (len > PAGE_SIZE)
4445 for (bio = blist; bio ; bio = bio->bi_next) {
4447 if (bio_add_page(bio, page, len, 0))
4450 /* Didn't fit, must stop */
4452 bio2 && bio2 != bio;
4453 bio2 = bio2->bi_next) {
4454 /* Remove last page from this bio */
4456 bio2->bi_size -= len;
4457 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4461 sector_nr += len >> 9;
4462 nr_sectors += len >> 9;
4465 r10_bio->sectors = nr_sectors;
4467 /* Now submit the read */
4468 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4469 atomic_inc(&r10_bio->remaining);
4470 read_bio->bi_next = NULL;
4471 generic_make_request(read_bio);
4472 sector_nr += nr_sectors;
4473 sectors_done += nr_sectors;
4474 if (sector_nr <= last)
4477 /* Now that we have done the whole section we can
4478 * update reshape_progress
4480 if (mddev->reshape_backwards)
4481 conf->reshape_progress -= sectors_done;
4483 conf->reshape_progress += sectors_done;
4485 return sectors_done;
4488 static void end_reshape_request(struct r10bio *r10_bio);
4489 static int handle_reshape_read_error(struct mddev *mddev,
4490 struct r10bio *r10_bio);
4491 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4493 /* Reshape read completed. Hopefully we have a block
4495 * If we got a read error then we do sync 1-page reads from
4496 * elsewhere until we find the data - or give up.
4498 struct r10conf *conf = mddev->private;
4501 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4502 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4503 /* Reshape has been aborted */
4504 md_done_sync(mddev, r10_bio->sectors, 0);
4508 /* We definitely have the data in the pages, schedule the
4511 atomic_set(&r10_bio->remaining, 1);
4512 for (s = 0; s < conf->copies*2; s++) {
4514 int d = r10_bio->devs[s/2].devnum;
4515 struct md_rdev *rdev;
4517 rdev = conf->mirrors[d].replacement;
4518 b = r10_bio->devs[s/2].repl_bio;
4520 rdev = conf->mirrors[d].rdev;
4521 b = r10_bio->devs[s/2].bio;
4523 if (!rdev || test_bit(Faulty, &rdev->flags))
4525 atomic_inc(&rdev->nr_pending);
4526 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4527 atomic_inc(&r10_bio->remaining);
4529 generic_make_request(b);
4531 end_reshape_request(r10_bio);
4534 static void end_reshape(struct r10conf *conf)
4536 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4539 spin_lock_irq(&conf->device_lock);
4540 conf->prev = conf->geo;
4541 md_finish_reshape(conf->mddev);
4543 conf->reshape_progress = MaxSector;
4544 spin_unlock_irq(&conf->device_lock);
4546 /* read-ahead size must cover two whole stripes, which is
4547 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4549 if (conf->mddev->queue) {
4550 int stripe = conf->geo.raid_disks *
4551 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4552 stripe /= conf->geo.near_copies;
4553 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4554 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4560 static int handle_reshape_read_error(struct mddev *mddev,
4561 struct r10bio *r10_bio)
4563 /* Use sync reads to get the blocks from somewhere else */
4564 int sectors = r10_bio->sectors;
4565 struct r10conf *conf = mddev->private;
4567 struct r10bio r10_bio;
4568 struct r10dev devs[conf->copies];
4570 struct r10bio *r10b = &on_stack.r10_bio;
4573 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4575 r10b->sector = r10_bio->sector;
4576 __raid10_find_phys(&conf->prev, r10b);
4581 int first_slot = slot;
4583 if (s > (PAGE_SIZE >> 9))
4587 int d = r10b->devs[slot].devnum;
4588 struct md_rdev *rdev = conf->mirrors[d].rdev;
4591 test_bit(Faulty, &rdev->flags) ||
4592 !test_bit(In_sync, &rdev->flags))
4595 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4596 success = sync_page_io(rdev,
4605 if (slot >= conf->copies)
4607 if (slot == first_slot)
4611 /* couldn't read this block, must give up */
4612 set_bit(MD_RECOVERY_INTR,
4622 static void end_reshape_write(struct bio *bio, int error)
4624 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4625 struct r10bio *r10_bio = bio->bi_private;
4626 struct mddev *mddev = r10_bio->mddev;
4627 struct r10conf *conf = mddev->private;
4631 struct md_rdev *rdev = NULL;
4633 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4635 rdev = conf->mirrors[d].replacement;
4638 rdev = conf->mirrors[d].rdev;
4642 /* FIXME should record badblock */
4643 md_error(mddev, rdev);
4646 rdev_dec_pending(rdev, mddev);
4647 end_reshape_request(r10_bio);
4650 static void end_reshape_request(struct r10bio *r10_bio)
4652 if (!atomic_dec_and_test(&r10_bio->remaining))
4654 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4655 bio_put(r10_bio->master_bio);
4659 static void raid10_finish_reshape(struct mddev *mddev)
4661 struct r10conf *conf = mddev->private;
4663 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4666 if (mddev->delta_disks > 0) {
4667 sector_t size = raid10_size(mddev, 0, 0);
4668 md_set_array_sectors(mddev, size);
4669 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4670 mddev->recovery_cp = mddev->resync_max_sectors;
4671 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4673 mddev->resync_max_sectors = size;
4674 set_capacity(mddev->gendisk, mddev->array_sectors);
4675 revalidate_disk(mddev->gendisk);
4678 for (d = conf->geo.raid_disks ;
4679 d < conf->geo.raid_disks - mddev->delta_disks;
4681 struct md_rdev *rdev = conf->mirrors[d].rdev;
4683 clear_bit(In_sync, &rdev->flags);
4684 rdev = conf->mirrors[d].replacement;
4686 clear_bit(In_sync, &rdev->flags);
4689 mddev->layout = mddev->new_layout;
4690 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4691 mddev->reshape_position = MaxSector;
4692 mddev->delta_disks = 0;
4693 mddev->reshape_backwards = 0;
4696 static struct md_personality raid10_personality =
4700 .owner = THIS_MODULE,
4701 .make_request = make_request,
4705 .error_handler = error,
4706 .hot_add_disk = raid10_add_disk,
4707 .hot_remove_disk= raid10_remove_disk,
4708 .spare_active = raid10_spare_active,
4709 .sync_request = sync_request,
4710 .quiesce = raid10_quiesce,
4711 .size = raid10_size,
4712 .resize = raid10_resize,
4713 .takeover = raid10_takeover,
4714 .check_reshape = raid10_check_reshape,
4715 .start_reshape = raid10_start_reshape,
4716 .finish_reshape = raid10_finish_reshape,
4719 static int __init raid_init(void)
4721 return register_md_personality(&raid10_personality);
4724 static void raid_exit(void)
4726 unregister_md_personality(&raid10_personality);
4729 module_init(raid_init);
4730 module_exit(raid_exit);
4731 MODULE_LICENSE("GPL");
4732 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4733 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4734 MODULE_ALIAS("md-raid10");
4735 MODULE_ALIAS("md-level-10");
4737 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);