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 )
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
50 * near_copies and far_copies must be at least one, and their product is at most
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
59 * Number of guaranteed r10bios in case of extreme VM load:
61 #define NR_RAID10_BIOS 256
63 /* when we get a read error on a read-only array, we redirect to another
64 * device without failing the first device, or trying to over-write to
65 * correct the read error. To keep track of bad blocks on a per-bio
66 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70 * bad-block marking which must be done from process context. So we record
71 * the success by setting devs[n].bio to IO_MADE_GOOD
73 #define IO_MADE_GOOD ((struct bio *)2)
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
77 /* When there are this many requests queued to be written by
78 * the raid10 thread, we become 'congested' to provide back-pressure
81 static int max_queued_requests = 1024;
83 static void allow_barrier(struct r10conf *conf);
84 static void lower_barrier(struct r10conf *conf);
85 static int enough(struct r10conf *conf, int ignore);
86 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
88 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
89 static void end_reshape_write(struct bio *bio, int error);
90 static void end_reshape(struct r10conf *conf);
92 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
94 struct r10conf *conf = data;
95 int size = offsetof(struct r10bio, devs[conf->copies]);
97 /* allocate a r10bio with room for raid_disks entries in the
99 return kzalloc(size, gfp_flags);
102 static void r10bio_pool_free(void *r10_bio, void *data)
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
116 * When performing a resync, we need to read and compare, so
117 * we need as many pages are there are copies.
118 * When performing a recovery, we need 2 bios, one for read,
119 * one for write (we recover only one drive per r10buf)
122 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
124 struct r10conf *conf = data;
126 struct r10bio *r10_bio;
131 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
135 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
136 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
137 nalloc = conf->copies; /* resync */
139 nalloc = 2; /* recovery */
144 for (j = nalloc ; j-- ; ) {
145 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
148 r10_bio->devs[j].bio = bio;
149 if (!conf->have_replacement)
151 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
154 r10_bio->devs[j].repl_bio = bio;
157 * Allocate RESYNC_PAGES data pages and attach them
160 for (j = 0 ; j < nalloc; j++) {
161 struct bio *rbio = r10_bio->devs[j].repl_bio;
162 bio = r10_bio->devs[j].bio;
163 for (i = 0; i < RESYNC_PAGES; i++) {
164 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
165 &conf->mddev->recovery)) {
166 /* we can share bv_page's during recovery
168 struct bio *rbio = r10_bio->devs[0].bio;
169 page = rbio->bi_io_vec[i].bv_page;
172 page = alloc_page(gfp_flags);
176 bio->bi_io_vec[i].bv_page = page;
178 rbio->bi_io_vec[i].bv_page = page;
186 safe_put_page(bio->bi_io_vec[i-1].bv_page);
188 for (i = 0; i < RESYNC_PAGES ; i++)
189 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
192 for ( ; j < nalloc; j++) {
193 if (r10_bio->devs[j].bio)
194 bio_put(r10_bio->devs[j].bio);
195 if (r10_bio->devs[j].repl_bio)
196 bio_put(r10_bio->devs[j].repl_bio);
198 r10bio_pool_free(r10_bio, conf);
202 static void r10buf_pool_free(void *__r10_bio, void *data)
205 struct r10conf *conf = data;
206 struct r10bio *r10bio = __r10_bio;
209 for (j=0; j < conf->copies; j++) {
210 struct bio *bio = r10bio->devs[j].bio;
212 for (i = 0; i < RESYNC_PAGES; i++) {
213 safe_put_page(bio->bi_io_vec[i].bv_page);
214 bio->bi_io_vec[i].bv_page = NULL;
218 bio = r10bio->devs[j].repl_bio;
222 r10bio_pool_free(r10bio, conf);
225 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
229 for (i = 0; i < conf->copies; i++) {
230 struct bio **bio = & r10_bio->devs[i].bio;
231 if (!BIO_SPECIAL(*bio))
234 bio = &r10_bio->devs[i].repl_bio;
235 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
241 static void free_r10bio(struct r10bio *r10_bio)
243 struct r10conf *conf = r10_bio->mddev->private;
245 put_all_bios(conf, r10_bio);
246 mempool_free(r10_bio, conf->r10bio_pool);
249 static void put_buf(struct r10bio *r10_bio)
251 struct r10conf *conf = r10_bio->mddev->private;
253 mempool_free(r10_bio, conf->r10buf_pool);
258 static void reschedule_retry(struct r10bio *r10_bio)
261 struct mddev *mddev = r10_bio->mddev;
262 struct r10conf *conf = mddev->private;
264 spin_lock_irqsave(&conf->device_lock, flags);
265 list_add(&r10_bio->retry_list, &conf->retry_list);
267 spin_unlock_irqrestore(&conf->device_lock, flags);
269 /* wake up frozen array... */
270 wake_up(&conf->wait_barrier);
272 md_wakeup_thread(mddev->thread);
276 * raid_end_bio_io() is called when we have finished servicing a mirrored
277 * operation and are ready to return a success/failure code to the buffer
280 static void raid_end_bio_io(struct r10bio *r10_bio)
282 struct bio *bio = r10_bio->master_bio;
284 struct r10conf *conf = r10_bio->mddev->private;
286 if (bio->bi_phys_segments) {
288 spin_lock_irqsave(&conf->device_lock, flags);
289 bio->bi_phys_segments--;
290 done = (bio->bi_phys_segments == 0);
291 spin_unlock_irqrestore(&conf->device_lock, flags);
294 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
295 clear_bit(BIO_UPTODATE, &bio->bi_flags);
299 * Wake up any possible resync thread that waits for the device
304 free_r10bio(r10_bio);
308 * Update disk head position estimator based on IRQ completion info.
310 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
312 struct r10conf *conf = r10_bio->mddev->private;
314 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
315 r10_bio->devs[slot].addr + (r10_bio->sectors);
319 * Find the disk number which triggered given bio
321 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
322 struct bio *bio, int *slotp, int *replp)
327 for (slot = 0; slot < conf->copies; slot++) {
328 if (r10_bio->devs[slot].bio == bio)
330 if (r10_bio->devs[slot].repl_bio == bio) {
336 BUG_ON(slot == conf->copies);
337 update_head_pos(slot, r10_bio);
343 return r10_bio->devs[slot].devnum;
346 static void raid10_end_read_request(struct bio *bio, int error)
348 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
349 struct r10bio *r10_bio = bio->bi_private;
351 struct md_rdev *rdev;
352 struct r10conf *conf = r10_bio->mddev->private;
355 slot = r10_bio->read_slot;
356 dev = r10_bio->devs[slot].devnum;
357 rdev = r10_bio->devs[slot].rdev;
359 * this branch is our 'one mirror IO has finished' event handler:
361 update_head_pos(slot, r10_bio);
365 * Set R10BIO_Uptodate in our master bio, so that
366 * we will return a good error code to the higher
367 * levels even if IO on some other mirrored buffer fails.
369 * The 'master' represents the composite IO operation to
370 * user-side. So if something waits for IO, then it will
371 * wait for the 'master' bio.
373 set_bit(R10BIO_Uptodate, &r10_bio->state);
375 /* If all other devices that store this block have
376 * failed, we want to return the error upwards rather
377 * than fail the last device. Here we redefine
378 * "uptodate" to mean "Don't want to retry"
381 spin_lock_irqsave(&conf->device_lock, flags);
382 if (!enough(conf, rdev->raid_disk))
384 spin_unlock_irqrestore(&conf->device_lock, flags);
387 raid_end_bio_io(r10_bio);
388 rdev_dec_pending(rdev, conf->mddev);
391 * oops, read error - keep the refcount on the rdev
393 char b[BDEVNAME_SIZE];
394 printk_ratelimited(KERN_ERR
395 "md/raid10:%s: %s: rescheduling sector %llu\n",
397 bdevname(rdev->bdev, b),
398 (unsigned long long)r10_bio->sector);
399 set_bit(R10BIO_ReadError, &r10_bio->state);
400 reschedule_retry(r10_bio);
404 static void close_write(struct r10bio *r10_bio)
406 /* clear the bitmap if all writes complete successfully */
407 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
409 !test_bit(R10BIO_Degraded, &r10_bio->state),
411 md_write_end(r10_bio->mddev);
414 static void one_write_done(struct r10bio *r10_bio)
416 if (atomic_dec_and_test(&r10_bio->remaining)) {
417 if (test_bit(R10BIO_WriteError, &r10_bio->state))
418 reschedule_retry(r10_bio);
420 close_write(r10_bio);
421 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
422 reschedule_retry(r10_bio);
424 raid_end_bio_io(r10_bio);
429 static void raid10_end_write_request(struct bio *bio, int error)
431 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
432 struct r10bio *r10_bio = bio->bi_private;
435 struct r10conf *conf = r10_bio->mddev->private;
437 struct md_rdev *rdev = NULL;
439 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
442 rdev = conf->mirrors[dev].replacement;
446 rdev = conf->mirrors[dev].rdev;
449 * this branch is our 'one mirror IO has finished' event handler:
453 /* Never record new bad blocks to replacement,
456 md_error(rdev->mddev, rdev);
458 set_bit(WriteErrorSeen, &rdev->flags);
459 if (!test_and_set_bit(WantReplacement, &rdev->flags))
460 set_bit(MD_RECOVERY_NEEDED,
461 &rdev->mddev->recovery);
462 set_bit(R10BIO_WriteError, &r10_bio->state);
467 * Set R10BIO_Uptodate in our master bio, so that
468 * we will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer fails.
471 * The 'master' represents the composite IO operation to
472 * user-side. So if something waits for IO, then it will
473 * wait for the 'master' bio.
478 set_bit(R10BIO_Uptodate, &r10_bio->state);
480 /* Maybe we can clear some bad blocks. */
481 if (is_badblock(rdev,
482 r10_bio->devs[slot].addr,
484 &first_bad, &bad_sectors)) {
487 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
489 r10_bio->devs[slot].bio = IO_MADE_GOOD;
491 set_bit(R10BIO_MadeGood, &r10_bio->state);
497 * Let's see if all mirrored write operations have finished
500 one_write_done(r10_bio);
502 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
506 * RAID10 layout manager
507 * As well as the chunksize and raid_disks count, there are two
508 * parameters: near_copies and far_copies.
509 * near_copies * far_copies must be <= raid_disks.
510 * Normally one of these will be 1.
511 * If both are 1, we get raid0.
512 * If near_copies == raid_disks, we get raid1.
514 * Chunks are laid out in raid0 style with near_copies copies of the
515 * first chunk, followed by near_copies copies of the next chunk and
517 * If far_copies > 1, then after 1/far_copies of the array has been assigned
518 * as described above, we start again with a device offset of near_copies.
519 * So we effectively have another copy of the whole array further down all
520 * the drives, but with blocks on different drives.
521 * With this layout, and block is never stored twice on the one device.
523 * raid10_find_phys finds the sector offset of a given virtual sector
524 * on each device that it is on.
526 * raid10_find_virt does the reverse mapping, from a device and a
527 * sector offset to a virtual address
530 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
539 /* now calculate first sector/dev */
540 chunk = r10bio->sector >> geo->chunk_shift;
541 sector = r10bio->sector & geo->chunk_mask;
543 chunk *= geo->near_copies;
545 dev = sector_div(stripe, geo->raid_disks);
547 stripe *= geo->far_copies;
549 sector += stripe << geo->chunk_shift;
551 /* and calculate all the others */
552 for (n = 0; n < geo->near_copies; n++) {
555 r10bio->devs[slot].addr = sector;
556 r10bio->devs[slot].devnum = d;
559 for (f = 1; f < geo->far_copies; f++) {
560 d += geo->near_copies;
561 if (d >= geo->raid_disks)
562 d -= geo->raid_disks;
564 r10bio->devs[slot].devnum = d;
565 r10bio->devs[slot].addr = s;
569 if (dev >= geo->raid_disks) {
571 sector += (geo->chunk_mask + 1);
576 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
578 struct geom *geo = &conf->geo;
580 if (conf->reshape_progress != MaxSector &&
581 ((r10bio->sector >= conf->reshape_progress) !=
582 conf->mddev->reshape_backwards)) {
583 set_bit(R10BIO_Previous, &r10bio->state);
586 clear_bit(R10BIO_Previous, &r10bio->state);
588 __raid10_find_phys(geo, r10bio);
591 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
593 sector_t offset, chunk, vchunk;
594 /* Never use conf->prev as this is only called during resync
595 * or recovery, so reshape isn't happening
597 struct geom *geo = &conf->geo;
599 offset = sector & geo->chunk_mask;
600 if (geo->far_offset) {
602 chunk = sector >> geo->chunk_shift;
603 fc = sector_div(chunk, geo->far_copies);
604 dev -= fc * geo->near_copies;
606 dev += geo->raid_disks;
608 while (sector >= geo->stride) {
609 sector -= geo->stride;
610 if (dev < geo->near_copies)
611 dev += geo->raid_disks - geo->near_copies;
613 dev -= geo->near_copies;
615 chunk = sector >> geo->chunk_shift;
617 vchunk = chunk * geo->raid_disks + dev;
618 sector_div(vchunk, geo->near_copies);
619 return (vchunk << geo->chunk_shift) + offset;
623 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
625 * @bvm: properties of new bio
626 * @biovec: the request that could be merged to it.
628 * Return amount of bytes we can accept at this offset
629 * This requires checking for end-of-chunk if near_copies != raid_disks,
630 * and for subordinate merge_bvec_fns if merge_check_needed.
632 static int raid10_mergeable_bvec(struct request_queue *q,
633 struct bvec_merge_data *bvm,
634 struct bio_vec *biovec)
636 struct mddev *mddev = q->queuedata;
637 struct r10conf *conf = mddev->private;
638 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
640 unsigned int chunk_sectors;
641 unsigned int bio_sectors = bvm->bi_size >> 9;
642 struct geom *geo = &conf->geo;
644 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
645 if (conf->reshape_progress != MaxSector &&
646 ((sector >= conf->reshape_progress) !=
647 conf->mddev->reshape_backwards))
650 if (geo->near_copies < geo->raid_disks) {
651 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
652 + bio_sectors)) << 9;
654 /* bio_add cannot handle a negative return */
656 if (max <= biovec->bv_len && bio_sectors == 0)
657 return biovec->bv_len;
659 max = biovec->bv_len;
661 if (mddev->merge_check_needed) {
663 struct r10bio r10_bio;
664 struct r10dev devs[conf->copies];
666 struct r10bio *r10_bio = &on_stack.r10_bio;
668 if (conf->reshape_progress != MaxSector) {
669 /* Cannot give any guidance during reshape */
670 if (max <= biovec->bv_len && bio_sectors == 0)
671 return biovec->bv_len;
674 r10_bio->sector = sector;
675 raid10_find_phys(conf, r10_bio);
677 for (s = 0; s < conf->copies; s++) {
678 int disk = r10_bio->devs[s].devnum;
679 struct md_rdev *rdev = rcu_dereference(
680 conf->mirrors[disk].rdev);
681 if (rdev && !test_bit(Faulty, &rdev->flags)) {
682 struct request_queue *q =
683 bdev_get_queue(rdev->bdev);
684 if (q->merge_bvec_fn) {
685 bvm->bi_sector = r10_bio->devs[s].addr
687 bvm->bi_bdev = rdev->bdev;
688 max = min(max, q->merge_bvec_fn(
692 rdev = rcu_dereference(conf->mirrors[disk].replacement);
693 if (rdev && !test_bit(Faulty, &rdev->flags)) {
694 struct request_queue *q =
695 bdev_get_queue(rdev->bdev);
696 if (q->merge_bvec_fn) {
697 bvm->bi_sector = r10_bio->devs[s].addr
699 bvm->bi_bdev = rdev->bdev;
700 max = min(max, q->merge_bvec_fn(
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
722 * The rdev for the device selected will have nr_pending incremented.
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
729 static struct md_rdev *read_balance(struct r10conf *conf,
730 struct r10bio *r10_bio,
733 const sector_t this_sector = r10_bio->sector;
735 int sectors = r10_bio->sectors;
736 int best_good_sectors;
737 sector_t new_distance, best_dist;
738 struct md_rdev *best_rdev, *rdev = NULL;
741 struct geom *geo = &conf->geo;
743 raid10_find_phys(conf, r10_bio);
746 sectors = r10_bio->sectors;
749 best_dist = MaxSector;
750 best_good_sectors = 0;
753 * Check if we can balance. We can balance on the whole
754 * device if no resync is going on (recovery is ok), or below
755 * the resync window. We take the first readable disk when
756 * above the resync window.
758 if (conf->mddev->recovery_cp < MaxSector
759 && (this_sector + sectors >= conf->next_resync))
762 for (slot = 0; slot < conf->copies ; slot++) {
767 if (r10_bio->devs[slot].bio == IO_BLOCKED)
769 disk = r10_bio->devs[slot].devnum;
770 rdev = rcu_dereference(conf->mirrors[disk].replacement);
771 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
772 test_bit(Unmerged, &rdev->flags) ||
773 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
774 rdev = rcu_dereference(conf->mirrors[disk].rdev);
776 test_bit(Faulty, &rdev->flags) ||
777 test_bit(Unmerged, &rdev->flags))
779 if (!test_bit(In_sync, &rdev->flags) &&
780 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
783 dev_sector = r10_bio->devs[slot].addr;
784 if (is_badblock(rdev, dev_sector, sectors,
785 &first_bad, &bad_sectors)) {
786 if (best_dist < MaxSector)
787 /* Already have a better slot */
789 if (first_bad <= dev_sector) {
790 /* Cannot read here. If this is the
791 * 'primary' device, then we must not read
792 * beyond 'bad_sectors' from another device.
794 bad_sectors -= (dev_sector - first_bad);
795 if (!do_balance && sectors > bad_sectors)
796 sectors = bad_sectors;
797 if (best_good_sectors > sectors)
798 best_good_sectors = sectors;
800 sector_t good_sectors =
801 first_bad - dev_sector;
802 if (good_sectors > best_good_sectors) {
803 best_good_sectors = good_sectors;
808 /* Must read from here */
813 best_good_sectors = sectors;
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
822 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
825 /* for far > 1 always use the lowest address */
826 if (geo->far_copies > 1)
827 new_distance = r10_bio->devs[slot].addr;
829 new_distance = abs(r10_bio->devs[slot].addr -
830 conf->mirrors[disk].head_position);
831 if (new_distance < best_dist) {
832 best_dist = new_distance;
837 if (slot >= conf->copies) {
843 atomic_inc(&rdev->nr_pending);
844 if (test_bit(Faulty, &rdev->flags)) {
845 /* Cannot risk returning a device that failed
846 * before we inc'ed nr_pending
848 rdev_dec_pending(rdev, conf->mddev);
851 r10_bio->read_slot = slot;
855 *max_sectors = best_good_sectors;
860 int md_raid10_congested(struct mddev *mddev, int bits)
862 struct r10conf *conf = mddev->private;
865 if ((bits & (1 << BDI_async_congested)) &&
866 conf->pending_count >= max_queued_requests)
871 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
874 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
875 if (rdev && !test_bit(Faulty, &rdev->flags)) {
876 struct request_queue *q = bdev_get_queue(rdev->bdev);
878 ret |= bdi_congested(&q->backing_dev_info, bits);
884 EXPORT_SYMBOL_GPL(md_raid10_congested);
886 static int raid10_congested(void *data, int bits)
888 struct mddev *mddev = data;
890 return mddev_congested(mddev, bits) ||
891 md_raid10_congested(mddev, bits);
894 static void flush_pending_writes(struct r10conf *conf)
896 /* Any writes that have been queued but are awaiting
897 * bitmap updates get flushed here.
899 spin_lock_irq(&conf->device_lock);
901 if (conf->pending_bio_list.head) {
903 bio = bio_list_get(&conf->pending_bio_list);
904 conf->pending_count = 0;
905 spin_unlock_irq(&conf->device_lock);
906 /* flush any pending bitmap writes to disk
907 * before proceeding w/ I/O */
908 bitmap_unplug(conf->mddev->bitmap);
909 wake_up(&conf->wait_barrier);
911 while (bio) { /* submit pending writes */
912 struct bio *next = bio->bi_next;
914 generic_make_request(bio);
918 spin_unlock_irq(&conf->device_lock);
922 * Sometimes we need to suspend IO while we do something else,
923 * either some resync/recovery, or reconfigure the array.
924 * To do this we raise a 'barrier'.
925 * The 'barrier' is a counter that can be raised multiple times
926 * to count how many activities are happening which preclude
928 * We can only raise the barrier if there is no pending IO.
929 * i.e. if nr_pending == 0.
930 * We choose only to raise the barrier if no-one is waiting for the
931 * barrier to go down. This means that as soon as an IO request
932 * is ready, no other operations which require a barrier will start
933 * until the IO request has had a chance.
935 * So: regular IO calls 'wait_barrier'. When that returns there
936 * is no backgroup IO happening, It must arrange to call
937 * allow_barrier when it has finished its IO.
938 * backgroup IO calls must call raise_barrier. Once that returns
939 * there is no normal IO happeing. It must arrange to call
940 * lower_barrier when the particular background IO completes.
943 static void raise_barrier(struct r10conf *conf, int force)
945 BUG_ON(force && !conf->barrier);
946 spin_lock_irq(&conf->resync_lock);
948 /* Wait until no block IO is waiting (unless 'force') */
949 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
950 conf->resync_lock, );
952 /* block any new IO from starting */
955 /* Now wait for all pending IO to complete */
956 wait_event_lock_irq(conf->wait_barrier,
957 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
958 conf->resync_lock, );
960 spin_unlock_irq(&conf->resync_lock);
963 static void lower_barrier(struct r10conf *conf)
966 spin_lock_irqsave(&conf->resync_lock, flags);
968 spin_unlock_irqrestore(&conf->resync_lock, flags);
969 wake_up(&conf->wait_barrier);
972 static void wait_barrier(struct r10conf *conf)
974 spin_lock_irq(&conf->resync_lock);
977 /* Wait for the barrier to drop.
978 * However if there are already pending
979 * requests (preventing the barrier from
980 * rising completely), and the
981 * pre-process bio queue isn't empty,
982 * then don't wait, as we need to empty
983 * that queue to get the nr_pending
986 wait_event_lock_irq(conf->wait_barrier,
990 !bio_list_empty(current->bio_list)),
996 spin_unlock_irq(&conf->resync_lock);
999 static void allow_barrier(struct r10conf *conf)
1001 unsigned long flags;
1002 spin_lock_irqsave(&conf->resync_lock, flags);
1004 spin_unlock_irqrestore(&conf->resync_lock, flags);
1005 wake_up(&conf->wait_barrier);
1008 static void freeze_array(struct r10conf *conf)
1010 /* stop syncio and normal IO and wait for everything to
1012 * We increment barrier and nr_waiting, and then
1013 * wait until nr_pending match nr_queued+1
1014 * This is called in the context of one normal IO request
1015 * that has failed. Thus any sync request that might be pending
1016 * will be blocked by nr_pending, and we need to wait for
1017 * pending IO requests to complete or be queued for re-try.
1018 * Thus the number queued (nr_queued) plus this request (1)
1019 * must match the number of pending IOs (nr_pending) before
1022 spin_lock_irq(&conf->resync_lock);
1025 wait_event_lock_irq(conf->wait_barrier,
1026 conf->nr_pending == conf->nr_queued+1,
1028 flush_pending_writes(conf));
1030 spin_unlock_irq(&conf->resync_lock);
1033 static void unfreeze_array(struct r10conf *conf)
1035 /* reverse the effect of the freeze */
1036 spin_lock_irq(&conf->resync_lock);
1039 wake_up(&conf->wait_barrier);
1040 spin_unlock_irq(&conf->resync_lock);
1043 static sector_t choose_data_offset(struct r10bio *r10_bio,
1044 struct md_rdev *rdev)
1046 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1047 test_bit(R10BIO_Previous, &r10_bio->state))
1048 return rdev->data_offset;
1050 return rdev->new_data_offset;
1053 static void make_request(struct mddev *mddev, struct bio * bio)
1055 struct r10conf *conf = mddev->private;
1056 struct r10bio *r10_bio;
1057 struct bio *read_bio;
1059 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1060 int chunk_sects = chunk_mask + 1;
1061 const int rw = bio_data_dir(bio);
1062 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1063 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1064 unsigned long flags;
1065 struct md_rdev *blocked_rdev;
1066 int sectors_handled;
1070 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1071 md_flush_request(mddev, bio);
1075 /* If this request crosses a chunk boundary, we need to
1076 * split it. This will only happen for 1 PAGE (or less) requests.
1078 if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1080 && (conf->geo.near_copies < conf->geo.raid_disks
1081 || conf->prev.near_copies < conf->prev.raid_disks))) {
1082 struct bio_pair *bp;
1083 /* Sanity check -- queue functions should prevent this happening */
1084 if (bio->bi_vcnt != 1 ||
1087 /* This is a one page bio that upper layers
1088 * refuse to split for us, so we need to split it.
1091 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1093 /* Each of these 'make_request' calls will call 'wait_barrier'.
1094 * If the first succeeds but the second blocks due to the resync
1095 * thread raising the barrier, we will deadlock because the
1096 * IO to the underlying device will be queued in generic_make_request
1097 * and will never complete, so will never reduce nr_pending.
1098 * So increment nr_waiting here so no new raise_barriers will
1099 * succeed, and so the second wait_barrier cannot block.
1101 spin_lock_irq(&conf->resync_lock);
1103 spin_unlock_irq(&conf->resync_lock);
1105 make_request(mddev, &bp->bio1);
1106 make_request(mddev, &bp->bio2);
1108 spin_lock_irq(&conf->resync_lock);
1110 wake_up(&conf->wait_barrier);
1111 spin_unlock_irq(&conf->resync_lock);
1113 bio_pair_release(bp);
1116 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1117 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1118 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1124 md_write_start(mddev, bio);
1127 * Register the new request and wait if the reconstruction
1128 * thread has put up a bar for new requests.
1129 * Continue immediately if no resync is active currently.
1133 sectors = bio->bi_size >> 9;
1134 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1135 bio->bi_sector < conf->reshape_progress &&
1136 bio->bi_sector + sectors > conf->reshape_progress) {
1137 /* IO spans the reshape position. Need to wait for
1140 allow_barrier(conf);
1141 wait_event(conf->wait_barrier,
1142 conf->reshape_progress <= bio->bi_sector ||
1143 conf->reshape_progress >= bio->bi_sector + sectors);
1146 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1147 bio_data_dir(bio) == WRITE &&
1148 (mddev->reshape_backwards
1149 ? (bio->bi_sector < conf->reshape_safe &&
1150 bio->bi_sector + sectors > conf->reshape_progress)
1151 : (bio->bi_sector + sectors > conf->reshape_safe &&
1152 bio->bi_sector < conf->reshape_progress))) {
1153 /* Need to update reshape_position in metadata */
1154 mddev->reshape_position = conf->reshape_progress;
1155 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1156 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1157 md_wakeup_thread(mddev->thread);
1158 wait_event(mddev->sb_wait,
1159 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1161 conf->reshape_safe = mddev->reshape_position;
1164 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1166 r10_bio->master_bio = bio;
1167 r10_bio->sectors = sectors;
1169 r10_bio->mddev = mddev;
1170 r10_bio->sector = bio->bi_sector;
1173 /* We might need to issue multiple reads to different
1174 * devices if there are bad blocks around, so we keep
1175 * track of the number of reads in bio->bi_phys_segments.
1176 * If this is 0, there is only one r10_bio and no locking
1177 * will be needed when the request completes. If it is
1178 * non-zero, then it is the number of not-completed requests.
1180 bio->bi_phys_segments = 0;
1181 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1185 * read balancing logic:
1187 struct md_rdev *rdev;
1191 rdev = read_balance(conf, r10_bio, &max_sectors);
1193 raid_end_bio_io(r10_bio);
1196 slot = r10_bio->read_slot;
1198 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1199 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1202 r10_bio->devs[slot].bio = read_bio;
1203 r10_bio->devs[slot].rdev = rdev;
1205 read_bio->bi_sector = r10_bio->devs[slot].addr +
1206 choose_data_offset(r10_bio, rdev);
1207 read_bio->bi_bdev = rdev->bdev;
1208 read_bio->bi_end_io = raid10_end_read_request;
1209 read_bio->bi_rw = READ | do_sync;
1210 read_bio->bi_private = r10_bio;
1212 if (max_sectors < r10_bio->sectors) {
1213 /* Could not read all from this device, so we will
1214 * need another r10_bio.
1216 sectors_handled = (r10_bio->sectors + max_sectors
1218 r10_bio->sectors = max_sectors;
1219 spin_lock_irq(&conf->device_lock);
1220 if (bio->bi_phys_segments == 0)
1221 bio->bi_phys_segments = 2;
1223 bio->bi_phys_segments++;
1224 spin_unlock(&conf->device_lock);
1225 /* Cannot call generic_make_request directly
1226 * as that will be queued in __generic_make_request
1227 * and subsequent mempool_alloc might block
1228 * waiting for it. so hand bio over to raid10d.
1230 reschedule_retry(r10_bio);
1232 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1234 r10_bio->master_bio = bio;
1235 r10_bio->sectors = ((bio->bi_size >> 9)
1238 r10_bio->mddev = mddev;
1239 r10_bio->sector = bio->bi_sector + sectors_handled;
1242 generic_make_request(read_bio);
1249 if (conf->pending_count >= max_queued_requests) {
1250 md_wakeup_thread(mddev->thread);
1251 wait_event(conf->wait_barrier,
1252 conf->pending_count < max_queued_requests);
1254 /* first select target devices under rcu_lock and
1255 * inc refcount on their rdev. Record them by setting
1257 * If there are known/acknowledged bad blocks on any device
1258 * on which we have seen a write error, we want to avoid
1259 * writing to those blocks. This potentially requires several
1260 * writes to write around the bad blocks. Each set of writes
1261 * gets its own r10_bio with a set of bios attached. The number
1262 * of r10_bios is recored in bio->bi_phys_segments just as with
1266 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1267 raid10_find_phys(conf, r10_bio);
1269 blocked_rdev = NULL;
1271 max_sectors = r10_bio->sectors;
1273 for (i = 0; i < conf->copies; i++) {
1274 int d = r10_bio->devs[i].devnum;
1275 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1276 struct md_rdev *rrdev = rcu_dereference(
1277 conf->mirrors[d].replacement);
1280 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1281 atomic_inc(&rdev->nr_pending);
1282 blocked_rdev = rdev;
1285 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1286 atomic_inc(&rrdev->nr_pending);
1287 blocked_rdev = rrdev;
1290 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1291 || test_bit(Unmerged, &rrdev->flags)))
1294 r10_bio->devs[i].bio = NULL;
1295 r10_bio->devs[i].repl_bio = NULL;
1296 if (!rdev || test_bit(Faulty, &rdev->flags) ||
1297 test_bit(Unmerged, &rdev->flags)) {
1298 set_bit(R10BIO_Degraded, &r10_bio->state);
1301 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1303 sector_t dev_sector = r10_bio->devs[i].addr;
1307 is_bad = is_badblock(rdev, dev_sector,
1309 &first_bad, &bad_sectors);
1311 /* Mustn't write here until the bad block
1314 atomic_inc(&rdev->nr_pending);
1315 set_bit(BlockedBadBlocks, &rdev->flags);
1316 blocked_rdev = rdev;
1319 if (is_bad && first_bad <= dev_sector) {
1320 /* Cannot write here at all */
1321 bad_sectors -= (dev_sector - first_bad);
1322 if (bad_sectors < max_sectors)
1323 /* Mustn't write more than bad_sectors
1324 * to other devices yet
1326 max_sectors = bad_sectors;
1327 /* We don't set R10BIO_Degraded as that
1328 * only applies if the disk is missing,
1329 * so it might be re-added, and we want to
1330 * know to recover this chunk.
1331 * In this case the device is here, and the
1332 * fact that this chunk is not in-sync is
1333 * recorded in the bad block log.
1338 int good_sectors = first_bad - dev_sector;
1339 if (good_sectors < max_sectors)
1340 max_sectors = good_sectors;
1343 r10_bio->devs[i].bio = bio;
1344 atomic_inc(&rdev->nr_pending);
1346 r10_bio->devs[i].repl_bio = bio;
1347 atomic_inc(&rrdev->nr_pending);
1352 if (unlikely(blocked_rdev)) {
1353 /* Have to wait for this device to get unblocked, then retry */
1357 for (j = 0; j < i; j++) {
1358 if (r10_bio->devs[j].bio) {
1359 d = r10_bio->devs[j].devnum;
1360 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1362 if (r10_bio->devs[j].repl_bio) {
1363 struct md_rdev *rdev;
1364 d = r10_bio->devs[j].devnum;
1365 rdev = conf->mirrors[d].replacement;
1367 /* Race with remove_disk */
1369 rdev = conf->mirrors[d].rdev;
1371 rdev_dec_pending(rdev, mddev);
1374 allow_barrier(conf);
1375 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1380 if (max_sectors < r10_bio->sectors) {
1381 /* We are splitting this into multiple parts, so
1382 * we need to prepare for allocating another r10_bio.
1384 r10_bio->sectors = max_sectors;
1385 spin_lock_irq(&conf->device_lock);
1386 if (bio->bi_phys_segments == 0)
1387 bio->bi_phys_segments = 2;
1389 bio->bi_phys_segments++;
1390 spin_unlock_irq(&conf->device_lock);
1392 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1394 atomic_set(&r10_bio->remaining, 1);
1395 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1397 for (i = 0; i < conf->copies; i++) {
1399 int d = r10_bio->devs[i].devnum;
1400 if (!r10_bio->devs[i].bio)
1403 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1404 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1406 r10_bio->devs[i].bio = mbio;
1408 mbio->bi_sector = (r10_bio->devs[i].addr+
1409 choose_data_offset(r10_bio,
1410 conf->mirrors[d].rdev));
1411 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1412 mbio->bi_end_io = raid10_end_write_request;
1413 mbio->bi_rw = WRITE | do_sync | do_fua;
1414 mbio->bi_private = r10_bio;
1416 atomic_inc(&r10_bio->remaining);
1417 spin_lock_irqsave(&conf->device_lock, flags);
1418 bio_list_add(&conf->pending_bio_list, mbio);
1419 conf->pending_count++;
1420 spin_unlock_irqrestore(&conf->device_lock, flags);
1421 if (!mddev_check_plugged(mddev))
1422 md_wakeup_thread(mddev->thread);
1424 if (!r10_bio->devs[i].repl_bio)
1427 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1428 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1430 r10_bio->devs[i].repl_bio = mbio;
1432 /* We are actively writing to the original device
1433 * so it cannot disappear, so the replacement cannot
1436 mbio->bi_sector = (r10_bio->devs[i].addr +
1439 conf->mirrors[d].replacement));
1440 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1441 mbio->bi_end_io = raid10_end_write_request;
1442 mbio->bi_rw = WRITE | do_sync | do_fua;
1443 mbio->bi_private = r10_bio;
1445 atomic_inc(&r10_bio->remaining);
1446 spin_lock_irqsave(&conf->device_lock, flags);
1447 bio_list_add(&conf->pending_bio_list, mbio);
1448 conf->pending_count++;
1449 spin_unlock_irqrestore(&conf->device_lock, flags);
1450 if (!mddev_check_plugged(mddev))
1451 md_wakeup_thread(mddev->thread);
1454 /* Don't remove the bias on 'remaining' (one_write_done) until
1455 * after checking if we need to go around again.
1458 if (sectors_handled < (bio->bi_size >> 9)) {
1459 one_write_done(r10_bio);
1460 /* We need another r10_bio. It has already been counted
1461 * in bio->bi_phys_segments.
1463 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1465 r10_bio->master_bio = bio;
1466 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1468 r10_bio->mddev = mddev;
1469 r10_bio->sector = bio->bi_sector + sectors_handled;
1473 one_write_done(r10_bio);
1475 /* In case raid10d snuck in to freeze_array */
1476 wake_up(&conf->wait_barrier);
1479 static void status(struct seq_file *seq, struct mddev *mddev)
1481 struct r10conf *conf = mddev->private;
1484 if (conf->geo.near_copies < conf->geo.raid_disks)
1485 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1486 if (conf->geo.near_copies > 1)
1487 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1488 if (conf->geo.far_copies > 1) {
1489 if (conf->geo.far_offset)
1490 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1492 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1494 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1495 conf->geo.raid_disks - mddev->degraded);
1496 for (i = 0; i < conf->geo.raid_disks; i++)
1497 seq_printf(seq, "%s",
1498 conf->mirrors[i].rdev &&
1499 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1500 seq_printf(seq, "]");
1503 /* check if there are enough drives for
1504 * every block to appear on atleast one.
1505 * Don't consider the device numbered 'ignore'
1506 * as we might be about to remove it.
1508 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1513 int n = conf->copies;
1517 if (conf->mirrors[this].rdev &&
1520 this = (this+1) % geo->raid_disks;
1524 first = (first + geo->near_copies) % geo->raid_disks;
1525 } while (first != 0);
1529 static int enough(struct r10conf *conf, int ignore)
1531 return _enough(conf, &conf->geo, ignore) &&
1532 _enough(conf, &conf->prev, ignore);
1535 static void error(struct mddev *mddev, struct md_rdev *rdev)
1537 char b[BDEVNAME_SIZE];
1538 struct r10conf *conf = mddev->private;
1541 * If it is not operational, then we have already marked it as dead
1542 * else if it is the last working disks, ignore the error, let the
1543 * next level up know.
1544 * else mark the drive as failed
1546 if (test_bit(In_sync, &rdev->flags)
1547 && !enough(conf, rdev->raid_disk))
1549 * Don't fail the drive, just return an IO error.
1552 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1553 unsigned long flags;
1554 spin_lock_irqsave(&conf->device_lock, flags);
1556 spin_unlock_irqrestore(&conf->device_lock, flags);
1558 * if recovery is running, make sure it aborts.
1560 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1562 set_bit(Blocked, &rdev->flags);
1563 set_bit(Faulty, &rdev->flags);
1564 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1566 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1567 "md/raid10:%s: Operation continuing on %d devices.\n",
1568 mdname(mddev), bdevname(rdev->bdev, b),
1569 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1572 static void print_conf(struct r10conf *conf)
1575 struct raid10_info *tmp;
1577 printk(KERN_DEBUG "RAID10 conf printout:\n");
1579 printk(KERN_DEBUG "(!conf)\n");
1582 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1583 conf->geo.raid_disks);
1585 for (i = 0; i < conf->geo.raid_disks; i++) {
1586 char b[BDEVNAME_SIZE];
1587 tmp = conf->mirrors + i;
1589 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1590 i, !test_bit(In_sync, &tmp->rdev->flags),
1591 !test_bit(Faulty, &tmp->rdev->flags),
1592 bdevname(tmp->rdev->bdev,b));
1596 static void close_sync(struct r10conf *conf)
1599 allow_barrier(conf);
1601 mempool_destroy(conf->r10buf_pool);
1602 conf->r10buf_pool = NULL;
1605 static int raid10_spare_active(struct mddev *mddev)
1608 struct r10conf *conf = mddev->private;
1609 struct raid10_info *tmp;
1611 unsigned long flags;
1614 * Find all non-in_sync disks within the RAID10 configuration
1615 * and mark them in_sync
1617 for (i = 0; i < conf->geo.raid_disks; i++) {
1618 tmp = conf->mirrors + i;
1619 if (tmp->replacement
1620 && tmp->replacement->recovery_offset == MaxSector
1621 && !test_bit(Faulty, &tmp->replacement->flags)
1622 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1623 /* Replacement has just become active */
1625 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1628 /* Replaced device not technically faulty,
1629 * but we need to be sure it gets removed
1630 * and never re-added.
1632 set_bit(Faulty, &tmp->rdev->flags);
1633 sysfs_notify_dirent_safe(
1634 tmp->rdev->sysfs_state);
1636 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1637 } else if (tmp->rdev
1638 && !test_bit(Faulty, &tmp->rdev->flags)
1639 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1641 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1644 spin_lock_irqsave(&conf->device_lock, flags);
1645 mddev->degraded -= count;
1646 spin_unlock_irqrestore(&conf->device_lock, flags);
1653 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1655 struct r10conf *conf = mddev->private;
1659 int last = conf->geo.raid_disks - 1;
1660 struct request_queue *q = bdev_get_queue(rdev->bdev);
1662 if (mddev->recovery_cp < MaxSector)
1663 /* only hot-add to in-sync arrays, as recovery is
1664 * very different from resync
1667 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1670 if (rdev->raid_disk >= 0)
1671 first = last = rdev->raid_disk;
1673 if (q->merge_bvec_fn) {
1674 set_bit(Unmerged, &rdev->flags);
1675 mddev->merge_check_needed = 1;
1678 if (rdev->saved_raid_disk >= first &&
1679 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1680 mirror = rdev->saved_raid_disk;
1683 for ( ; mirror <= last ; mirror++) {
1684 struct raid10_info *p = &conf->mirrors[mirror];
1685 if (p->recovery_disabled == mddev->recovery_disabled)
1688 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1689 p->replacement != NULL)
1691 clear_bit(In_sync, &rdev->flags);
1692 set_bit(Replacement, &rdev->flags);
1693 rdev->raid_disk = mirror;
1695 disk_stack_limits(mddev->gendisk, rdev->bdev,
1696 rdev->data_offset << 9);
1698 rcu_assign_pointer(p->replacement, rdev);
1702 disk_stack_limits(mddev->gendisk, rdev->bdev,
1703 rdev->data_offset << 9);
1705 p->head_position = 0;
1706 p->recovery_disabled = mddev->recovery_disabled - 1;
1707 rdev->raid_disk = mirror;
1709 if (rdev->saved_raid_disk != mirror)
1711 rcu_assign_pointer(p->rdev, rdev);
1714 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1715 /* Some requests might not have seen this new
1716 * merge_bvec_fn. We must wait for them to complete
1717 * before merging the device fully.
1718 * First we make sure any code which has tested
1719 * our function has submitted the request, then
1720 * we wait for all outstanding requests to complete.
1722 synchronize_sched();
1723 raise_barrier(conf, 0);
1724 lower_barrier(conf);
1725 clear_bit(Unmerged, &rdev->flags);
1727 md_integrity_add_rdev(rdev, mddev);
1732 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1734 struct r10conf *conf = mddev->private;
1736 int number = rdev->raid_disk;
1737 struct md_rdev **rdevp;
1738 struct raid10_info *p = conf->mirrors + number;
1741 if (rdev == p->rdev)
1743 else if (rdev == p->replacement)
1744 rdevp = &p->replacement;
1748 if (test_bit(In_sync, &rdev->flags) ||
1749 atomic_read(&rdev->nr_pending)) {
1753 /* Only remove faulty devices if recovery
1756 if (!test_bit(Faulty, &rdev->flags) &&
1757 mddev->recovery_disabled != p->recovery_disabled &&
1758 (!p->replacement || p->replacement == rdev) &&
1759 number < conf->geo.raid_disks &&
1766 if (atomic_read(&rdev->nr_pending)) {
1767 /* lost the race, try later */
1771 } else if (p->replacement) {
1772 /* We must have just cleared 'rdev' */
1773 p->rdev = p->replacement;
1774 clear_bit(Replacement, &p->replacement->flags);
1775 smp_mb(); /* Make sure other CPUs may see both as identical
1776 * but will never see neither -- if they are careful.
1778 p->replacement = NULL;
1779 clear_bit(WantReplacement, &rdev->flags);
1781 /* We might have just remove the Replacement as faulty
1782 * Clear the flag just in case
1784 clear_bit(WantReplacement, &rdev->flags);
1786 err = md_integrity_register(mddev);
1795 static void end_sync_read(struct bio *bio, int error)
1797 struct r10bio *r10_bio = bio->bi_private;
1798 struct r10conf *conf = r10_bio->mddev->private;
1801 if (bio == r10_bio->master_bio) {
1802 /* this is a reshape read */
1803 d = r10_bio->read_slot; /* really the read dev */
1805 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1807 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1808 set_bit(R10BIO_Uptodate, &r10_bio->state);
1810 /* The write handler will notice the lack of
1811 * R10BIO_Uptodate and record any errors etc
1813 atomic_add(r10_bio->sectors,
1814 &conf->mirrors[d].rdev->corrected_errors);
1816 /* for reconstruct, we always reschedule after a read.
1817 * for resync, only after all reads
1819 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1820 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1821 atomic_dec_and_test(&r10_bio->remaining)) {
1822 /* we have read all the blocks,
1823 * do the comparison in process context in raid10d
1825 reschedule_retry(r10_bio);
1829 static void end_sync_request(struct r10bio *r10_bio)
1831 struct mddev *mddev = r10_bio->mddev;
1833 while (atomic_dec_and_test(&r10_bio->remaining)) {
1834 if (r10_bio->master_bio == NULL) {
1835 /* the primary of several recovery bios */
1836 sector_t s = r10_bio->sectors;
1837 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1838 test_bit(R10BIO_WriteError, &r10_bio->state))
1839 reschedule_retry(r10_bio);
1842 md_done_sync(mddev, s, 1);
1845 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1846 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1847 test_bit(R10BIO_WriteError, &r10_bio->state))
1848 reschedule_retry(r10_bio);
1856 static void end_sync_write(struct bio *bio, int error)
1858 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1859 struct r10bio *r10_bio = bio->bi_private;
1860 struct mddev *mddev = r10_bio->mddev;
1861 struct r10conf *conf = mddev->private;
1867 struct md_rdev *rdev = NULL;
1869 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1871 rdev = conf->mirrors[d].replacement;
1873 rdev = conf->mirrors[d].rdev;
1877 md_error(mddev, rdev);
1879 set_bit(WriteErrorSeen, &rdev->flags);
1880 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1881 set_bit(MD_RECOVERY_NEEDED,
1882 &rdev->mddev->recovery);
1883 set_bit(R10BIO_WriteError, &r10_bio->state);
1885 } else if (is_badblock(rdev,
1886 r10_bio->devs[slot].addr,
1888 &first_bad, &bad_sectors))
1889 set_bit(R10BIO_MadeGood, &r10_bio->state);
1891 rdev_dec_pending(rdev, mddev);
1893 end_sync_request(r10_bio);
1897 * Note: sync and recover and handled very differently for raid10
1898 * This code is for resync.
1899 * For resync, we read through virtual addresses and read all blocks.
1900 * If there is any error, we schedule a write. The lowest numbered
1901 * drive is authoritative.
1902 * However requests come for physical address, so we need to map.
1903 * For every physical address there are raid_disks/copies virtual addresses,
1904 * which is always are least one, but is not necessarly an integer.
1905 * This means that a physical address can span multiple chunks, so we may
1906 * have to submit multiple io requests for a single sync request.
1909 * We check if all blocks are in-sync and only write to blocks that
1912 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1914 struct r10conf *conf = mddev->private;
1916 struct bio *tbio, *fbio;
1919 atomic_set(&r10_bio->remaining, 1);
1921 /* find the first device with a block */
1922 for (i=0; i<conf->copies; i++)
1923 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1926 if (i == conf->copies)
1930 fbio = r10_bio->devs[i].bio;
1932 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1933 /* now find blocks with errors */
1934 for (i=0 ; i < conf->copies ; i++) {
1937 tbio = r10_bio->devs[i].bio;
1939 if (tbio->bi_end_io != end_sync_read)
1943 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1944 /* We know that the bi_io_vec layout is the same for
1945 * both 'first' and 'i', so we just compare them.
1946 * All vec entries are PAGE_SIZE;
1948 for (j = 0; j < vcnt; j++)
1949 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1950 page_address(tbio->bi_io_vec[j].bv_page),
1951 fbio->bi_io_vec[j].bv_len))
1955 mddev->resync_mismatches += r10_bio->sectors;
1956 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1957 /* Don't fix anything. */
1960 /* Ok, we need to write this bio, either to correct an
1961 * inconsistency or to correct an unreadable block.
1962 * First we need to fixup bv_offset, bv_len and
1963 * bi_vecs, as the read request might have corrupted these
1965 tbio->bi_vcnt = vcnt;
1966 tbio->bi_size = r10_bio->sectors << 9;
1968 tbio->bi_phys_segments = 0;
1969 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1970 tbio->bi_flags |= 1 << BIO_UPTODATE;
1971 tbio->bi_next = NULL;
1972 tbio->bi_rw = WRITE;
1973 tbio->bi_private = r10_bio;
1974 tbio->bi_sector = r10_bio->devs[i].addr;
1976 for (j=0; j < vcnt ; j++) {
1977 tbio->bi_io_vec[j].bv_offset = 0;
1978 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1980 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1981 page_address(fbio->bi_io_vec[j].bv_page),
1984 tbio->bi_end_io = end_sync_write;
1986 d = r10_bio->devs[i].devnum;
1987 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1988 atomic_inc(&r10_bio->remaining);
1989 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1991 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1992 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1993 generic_make_request(tbio);
1996 /* Now write out to any replacement devices
1999 for (i = 0; i < conf->copies; i++) {
2002 tbio = r10_bio->devs[i].repl_bio;
2003 if (!tbio || !tbio->bi_end_io)
2005 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2006 && r10_bio->devs[i].bio != fbio)
2007 for (j = 0; j < vcnt; j++)
2008 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2009 page_address(fbio->bi_io_vec[j].bv_page),
2011 d = r10_bio->devs[i].devnum;
2012 atomic_inc(&r10_bio->remaining);
2013 md_sync_acct(conf->mirrors[d].replacement->bdev,
2014 tbio->bi_size >> 9);
2015 generic_make_request(tbio);
2019 if (atomic_dec_and_test(&r10_bio->remaining)) {
2020 md_done_sync(mddev, r10_bio->sectors, 1);
2026 * Now for the recovery code.
2027 * Recovery happens across physical sectors.
2028 * We recover all non-is_sync drives by finding the virtual address of
2029 * each, and then choose a working drive that also has that virt address.
2030 * There is a separate r10_bio for each non-in_sync drive.
2031 * Only the first two slots are in use. The first for reading,
2032 * The second for writing.
2035 static void fix_recovery_read_error(struct r10bio *r10_bio)
2037 /* We got a read error during recovery.
2038 * We repeat the read in smaller page-sized sections.
2039 * If a read succeeds, write it to the new device or record
2040 * a bad block if we cannot.
2041 * If a read fails, record a bad block on both old and
2044 struct mddev *mddev = r10_bio->mddev;
2045 struct r10conf *conf = mddev->private;
2046 struct bio *bio = r10_bio->devs[0].bio;
2048 int sectors = r10_bio->sectors;
2050 int dr = r10_bio->devs[0].devnum;
2051 int dw = r10_bio->devs[1].devnum;
2055 struct md_rdev *rdev;
2059 if (s > (PAGE_SIZE>>9))
2062 rdev = conf->mirrors[dr].rdev;
2063 addr = r10_bio->devs[0].addr + sect,
2064 ok = sync_page_io(rdev,
2067 bio->bi_io_vec[idx].bv_page,
2070 rdev = conf->mirrors[dw].rdev;
2071 addr = r10_bio->devs[1].addr + sect;
2072 ok = sync_page_io(rdev,
2075 bio->bi_io_vec[idx].bv_page,
2078 set_bit(WriteErrorSeen, &rdev->flags);
2079 if (!test_and_set_bit(WantReplacement,
2081 set_bit(MD_RECOVERY_NEEDED,
2082 &rdev->mddev->recovery);
2086 /* We don't worry if we cannot set a bad block -
2087 * it really is bad so there is no loss in not
2090 rdev_set_badblocks(rdev, addr, s, 0);
2092 if (rdev != conf->mirrors[dw].rdev) {
2093 /* need bad block on destination too */
2094 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2095 addr = r10_bio->devs[1].addr + sect;
2096 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2098 /* just abort the recovery */
2100 "md/raid10:%s: recovery aborted"
2101 " due to read error\n",
2104 conf->mirrors[dw].recovery_disabled
2105 = mddev->recovery_disabled;
2106 set_bit(MD_RECOVERY_INTR,
2119 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2121 struct r10conf *conf = mddev->private;
2123 struct bio *wbio, *wbio2;
2125 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2126 fix_recovery_read_error(r10_bio);
2127 end_sync_request(r10_bio);
2132 * share the pages with the first bio
2133 * and submit the write request
2135 d = r10_bio->devs[1].devnum;
2136 wbio = r10_bio->devs[1].bio;
2137 wbio2 = r10_bio->devs[1].repl_bio;
2138 if (wbio->bi_end_io) {
2139 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2140 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2141 generic_make_request(wbio);
2143 if (wbio2 && wbio2->bi_end_io) {
2144 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2145 md_sync_acct(conf->mirrors[d].replacement->bdev,
2146 wbio2->bi_size >> 9);
2147 generic_make_request(wbio2);
2153 * Used by fix_read_error() to decay the per rdev read_errors.
2154 * We halve the read error count for every hour that has elapsed
2155 * since the last recorded read error.
2158 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2160 struct timespec cur_time_mon;
2161 unsigned long hours_since_last;
2162 unsigned int read_errors = atomic_read(&rdev->read_errors);
2164 ktime_get_ts(&cur_time_mon);
2166 if (rdev->last_read_error.tv_sec == 0 &&
2167 rdev->last_read_error.tv_nsec == 0) {
2168 /* first time we've seen a read error */
2169 rdev->last_read_error = cur_time_mon;
2173 hours_since_last = (cur_time_mon.tv_sec -
2174 rdev->last_read_error.tv_sec) / 3600;
2176 rdev->last_read_error = cur_time_mon;
2179 * if hours_since_last is > the number of bits in read_errors
2180 * just set read errors to 0. We do this to avoid
2181 * overflowing the shift of read_errors by hours_since_last.
2183 if (hours_since_last >= 8 * sizeof(read_errors))
2184 atomic_set(&rdev->read_errors, 0);
2186 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2189 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2190 int sectors, struct page *page, int rw)
2195 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2196 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2198 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2202 set_bit(WriteErrorSeen, &rdev->flags);
2203 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2204 set_bit(MD_RECOVERY_NEEDED,
2205 &rdev->mddev->recovery);
2207 /* need to record an error - either for the block or the device */
2208 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2209 md_error(rdev->mddev, rdev);
2214 * This is a kernel thread which:
2216 * 1. Retries failed read operations on working mirrors.
2217 * 2. Updates the raid superblock when problems encounter.
2218 * 3. Performs writes following reads for array synchronising.
2221 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2223 int sect = 0; /* Offset from r10_bio->sector */
2224 int sectors = r10_bio->sectors;
2225 struct md_rdev*rdev;
2226 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2227 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2229 /* still own a reference to this rdev, so it cannot
2230 * have been cleared recently.
2232 rdev = conf->mirrors[d].rdev;
2234 if (test_bit(Faulty, &rdev->flags))
2235 /* drive has already been failed, just ignore any
2236 more fix_read_error() attempts */
2239 check_decay_read_errors(mddev, rdev);
2240 atomic_inc(&rdev->read_errors);
2241 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2242 char b[BDEVNAME_SIZE];
2243 bdevname(rdev->bdev, b);
2246 "md/raid10:%s: %s: Raid device exceeded "
2247 "read_error threshold [cur %d:max %d]\n",
2249 atomic_read(&rdev->read_errors), max_read_errors);
2251 "md/raid10:%s: %s: Failing raid device\n",
2253 md_error(mddev, conf->mirrors[d].rdev);
2254 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2260 int sl = r10_bio->read_slot;
2264 if (s > (PAGE_SIZE>>9))
2272 d = r10_bio->devs[sl].devnum;
2273 rdev = rcu_dereference(conf->mirrors[d].rdev);
2275 !test_bit(Unmerged, &rdev->flags) &&
2276 test_bit(In_sync, &rdev->flags) &&
2277 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2278 &first_bad, &bad_sectors) == 0) {
2279 atomic_inc(&rdev->nr_pending);
2281 success = sync_page_io(rdev,
2282 r10_bio->devs[sl].addr +
2285 conf->tmppage, READ, false);
2286 rdev_dec_pending(rdev, mddev);
2292 if (sl == conf->copies)
2294 } while (!success && sl != r10_bio->read_slot);
2298 /* Cannot read from anywhere, just mark the block
2299 * as bad on the first device to discourage future
2302 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2303 rdev = conf->mirrors[dn].rdev;
2305 if (!rdev_set_badblocks(
2307 r10_bio->devs[r10_bio->read_slot].addr
2310 md_error(mddev, rdev);
2311 r10_bio->devs[r10_bio->read_slot].bio
2318 /* write it back and re-read */
2320 while (sl != r10_bio->read_slot) {
2321 char b[BDEVNAME_SIZE];
2326 d = r10_bio->devs[sl].devnum;
2327 rdev = rcu_dereference(conf->mirrors[d].rdev);
2329 test_bit(Unmerged, &rdev->flags) ||
2330 !test_bit(In_sync, &rdev->flags))
2333 atomic_inc(&rdev->nr_pending);
2335 if (r10_sync_page_io(rdev,
2336 r10_bio->devs[sl].addr +
2338 s, conf->tmppage, WRITE)
2340 /* Well, this device is dead */
2342 "md/raid10:%s: read correction "
2344 " (%d sectors at %llu on %s)\n",
2346 (unsigned long long)(
2348 choose_data_offset(r10_bio,
2350 bdevname(rdev->bdev, b));
2351 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2354 bdevname(rdev->bdev, b));
2356 rdev_dec_pending(rdev, mddev);
2360 while (sl != r10_bio->read_slot) {
2361 char b[BDEVNAME_SIZE];
2366 d = r10_bio->devs[sl].devnum;
2367 rdev = rcu_dereference(conf->mirrors[d].rdev);
2369 !test_bit(In_sync, &rdev->flags))
2372 atomic_inc(&rdev->nr_pending);
2374 switch (r10_sync_page_io(rdev,
2375 r10_bio->devs[sl].addr +
2380 /* Well, this device is dead */
2382 "md/raid10:%s: unable to read back "
2384 " (%d sectors at %llu on %s)\n",
2386 (unsigned long long)(
2388 choose_data_offset(r10_bio, rdev)),
2389 bdevname(rdev->bdev, b));
2390 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2393 bdevname(rdev->bdev, b));
2397 "md/raid10:%s: read error corrected"
2398 " (%d sectors at %llu on %s)\n",
2400 (unsigned long long)(
2402 choose_data_offset(r10_bio, rdev)),
2403 bdevname(rdev->bdev, b));
2404 atomic_add(s, &rdev->corrected_errors);
2407 rdev_dec_pending(rdev, mddev);
2417 static void bi_complete(struct bio *bio, int error)
2419 complete((struct completion *)bio->bi_private);
2422 static int submit_bio_wait(int rw, struct bio *bio)
2424 struct completion event;
2427 init_completion(&event);
2428 bio->bi_private = &event;
2429 bio->bi_end_io = bi_complete;
2430 submit_bio(rw, bio);
2431 wait_for_completion(&event);
2433 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2436 static int narrow_write_error(struct r10bio *r10_bio, int i)
2438 struct bio *bio = r10_bio->master_bio;
2439 struct mddev *mddev = r10_bio->mddev;
2440 struct r10conf *conf = mddev->private;
2441 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2442 /* bio has the data to be written to slot 'i' where
2443 * we just recently had a write error.
2444 * We repeatedly clone the bio and trim down to one block,
2445 * then try the write. Where the write fails we record
2447 * It is conceivable that the bio doesn't exactly align with
2448 * blocks. We must handle this.
2450 * We currently own a reference to the rdev.
2456 int sect_to_write = r10_bio->sectors;
2459 if (rdev->badblocks.shift < 0)
2462 block_sectors = 1 << rdev->badblocks.shift;
2463 sector = r10_bio->sector;
2464 sectors = ((r10_bio->sector + block_sectors)
2465 & ~(sector_t)(block_sectors - 1))
2468 while (sect_to_write) {
2470 if (sectors > sect_to_write)
2471 sectors = sect_to_write;
2472 /* Write at 'sector' for 'sectors' */
2473 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2474 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2475 wbio->bi_sector = (r10_bio->devs[i].addr+
2476 choose_data_offset(r10_bio, rdev) +
2477 (sector - r10_bio->sector));
2478 wbio->bi_bdev = rdev->bdev;
2479 if (submit_bio_wait(WRITE, wbio) == 0)
2481 ok = rdev_set_badblocks(rdev, sector,
2486 sect_to_write -= sectors;
2488 sectors = block_sectors;
2493 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2495 int slot = r10_bio->read_slot;
2497 struct r10conf *conf = mddev->private;
2498 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2499 char b[BDEVNAME_SIZE];
2500 unsigned long do_sync;
2503 /* we got a read error. Maybe the drive is bad. Maybe just
2504 * the block and we can fix it.
2505 * We freeze all other IO, and try reading the block from
2506 * other devices. When we find one, we re-write
2507 * and check it that fixes the read error.
2508 * This is all done synchronously while the array is
2511 bio = r10_bio->devs[slot].bio;
2512 bdevname(bio->bi_bdev, b);
2514 r10_bio->devs[slot].bio = NULL;
2516 if (mddev->ro == 0) {
2518 fix_read_error(conf, mddev, r10_bio);
2519 unfreeze_array(conf);
2521 r10_bio->devs[slot].bio = IO_BLOCKED;
2523 rdev_dec_pending(rdev, mddev);
2526 rdev = read_balance(conf, r10_bio, &max_sectors);
2528 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2529 " read error for block %llu\n",
2531 (unsigned long long)r10_bio->sector);
2532 raid_end_bio_io(r10_bio);
2536 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2537 slot = r10_bio->read_slot;
2540 "md/raid10:%s: %s: redirecting "
2541 "sector %llu to another mirror\n",
2543 bdevname(rdev->bdev, b),
2544 (unsigned long long)r10_bio->sector);
2545 bio = bio_clone_mddev(r10_bio->master_bio,
2548 r10_bio->sector - bio->bi_sector,
2550 r10_bio->devs[slot].bio = bio;
2551 r10_bio->devs[slot].rdev = rdev;
2552 bio->bi_sector = r10_bio->devs[slot].addr
2553 + choose_data_offset(r10_bio, rdev);
2554 bio->bi_bdev = rdev->bdev;
2555 bio->bi_rw = READ | do_sync;
2556 bio->bi_private = r10_bio;
2557 bio->bi_end_io = raid10_end_read_request;
2558 if (max_sectors < r10_bio->sectors) {
2559 /* Drat - have to split this up more */
2560 struct bio *mbio = r10_bio->master_bio;
2561 int sectors_handled =
2562 r10_bio->sector + max_sectors
2564 r10_bio->sectors = max_sectors;
2565 spin_lock_irq(&conf->device_lock);
2566 if (mbio->bi_phys_segments == 0)
2567 mbio->bi_phys_segments = 2;
2569 mbio->bi_phys_segments++;
2570 spin_unlock_irq(&conf->device_lock);
2571 generic_make_request(bio);
2573 r10_bio = mempool_alloc(conf->r10bio_pool,
2575 r10_bio->master_bio = mbio;
2576 r10_bio->sectors = (mbio->bi_size >> 9)
2579 set_bit(R10BIO_ReadError,
2581 r10_bio->mddev = mddev;
2582 r10_bio->sector = mbio->bi_sector
2587 generic_make_request(bio);
2590 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2592 /* Some sort of write request has finished and it
2593 * succeeded in writing where we thought there was a
2594 * bad block. So forget the bad block.
2595 * Or possibly if failed and we need to record
2599 struct md_rdev *rdev;
2601 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2602 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2603 for (m = 0; m < conf->copies; m++) {
2604 int dev = r10_bio->devs[m].devnum;
2605 rdev = conf->mirrors[dev].rdev;
2606 if (r10_bio->devs[m].bio == NULL)
2608 if (test_bit(BIO_UPTODATE,
2609 &r10_bio->devs[m].bio->bi_flags)) {
2610 rdev_clear_badblocks(
2612 r10_bio->devs[m].addr,
2613 r10_bio->sectors, 0);
2615 if (!rdev_set_badblocks(
2617 r10_bio->devs[m].addr,
2618 r10_bio->sectors, 0))
2619 md_error(conf->mddev, rdev);
2621 rdev = conf->mirrors[dev].replacement;
2622 if (r10_bio->devs[m].repl_bio == NULL)
2624 if (test_bit(BIO_UPTODATE,
2625 &r10_bio->devs[m].repl_bio->bi_flags)) {
2626 rdev_clear_badblocks(
2628 r10_bio->devs[m].addr,
2629 r10_bio->sectors, 0);
2631 if (!rdev_set_badblocks(
2633 r10_bio->devs[m].addr,
2634 r10_bio->sectors, 0))
2635 md_error(conf->mddev, rdev);
2640 for (m = 0; m < conf->copies; m++) {
2641 int dev = r10_bio->devs[m].devnum;
2642 struct bio *bio = r10_bio->devs[m].bio;
2643 rdev = conf->mirrors[dev].rdev;
2644 if (bio == IO_MADE_GOOD) {
2645 rdev_clear_badblocks(
2647 r10_bio->devs[m].addr,
2648 r10_bio->sectors, 0);
2649 rdev_dec_pending(rdev, conf->mddev);
2650 } else if (bio != NULL &&
2651 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2652 if (!narrow_write_error(r10_bio, m)) {
2653 md_error(conf->mddev, rdev);
2654 set_bit(R10BIO_Degraded,
2657 rdev_dec_pending(rdev, conf->mddev);
2659 bio = r10_bio->devs[m].repl_bio;
2660 rdev = conf->mirrors[dev].replacement;
2661 if (rdev && bio == IO_MADE_GOOD) {
2662 rdev_clear_badblocks(
2664 r10_bio->devs[m].addr,
2665 r10_bio->sectors, 0);
2666 rdev_dec_pending(rdev, conf->mddev);
2669 if (test_bit(R10BIO_WriteError,
2671 close_write(r10_bio);
2672 raid_end_bio_io(r10_bio);
2676 static void raid10d(struct mddev *mddev)
2678 struct r10bio *r10_bio;
2679 unsigned long flags;
2680 struct r10conf *conf = mddev->private;
2681 struct list_head *head = &conf->retry_list;
2682 struct blk_plug plug;
2684 md_check_recovery(mddev);
2686 blk_start_plug(&plug);
2689 flush_pending_writes(conf);
2691 spin_lock_irqsave(&conf->device_lock, flags);
2692 if (list_empty(head)) {
2693 spin_unlock_irqrestore(&conf->device_lock, flags);
2696 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2697 list_del(head->prev);
2699 spin_unlock_irqrestore(&conf->device_lock, flags);
2701 mddev = r10_bio->mddev;
2702 conf = mddev->private;
2703 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2704 test_bit(R10BIO_WriteError, &r10_bio->state))
2705 handle_write_completed(conf, r10_bio);
2706 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2707 reshape_request_write(mddev, r10_bio);
2708 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2709 sync_request_write(mddev, r10_bio);
2710 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2711 recovery_request_write(mddev, r10_bio);
2712 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2713 handle_read_error(mddev, r10_bio);
2715 /* just a partial read to be scheduled from a
2718 int slot = r10_bio->read_slot;
2719 generic_make_request(r10_bio->devs[slot].bio);
2723 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2724 md_check_recovery(mddev);
2726 blk_finish_plug(&plug);
2730 static int init_resync(struct r10conf *conf)
2735 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2736 BUG_ON(conf->r10buf_pool);
2737 conf->have_replacement = 0;
2738 for (i = 0; i < conf->geo.raid_disks; i++)
2739 if (conf->mirrors[i].replacement)
2740 conf->have_replacement = 1;
2741 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2742 if (!conf->r10buf_pool)
2744 conf->next_resync = 0;
2749 * perform a "sync" on one "block"
2751 * We need to make sure that no normal I/O request - particularly write
2752 * requests - conflict with active sync requests.
2754 * This is achieved by tracking pending requests and a 'barrier' concept
2755 * that can be installed to exclude normal IO requests.
2757 * Resync and recovery are handled very differently.
2758 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2760 * For resync, we iterate over virtual addresses, read all copies,
2761 * and update if there are differences. If only one copy is live,
2763 * For recovery, we iterate over physical addresses, read a good
2764 * value for each non-in_sync drive, and over-write.
2766 * So, for recovery we may have several outstanding complex requests for a
2767 * given address, one for each out-of-sync device. We model this by allocating
2768 * a number of r10_bio structures, one for each out-of-sync device.
2769 * As we setup these structures, we collect all bio's together into a list
2770 * which we then process collectively to add pages, and then process again
2771 * to pass to generic_make_request.
2773 * The r10_bio structures are linked using a borrowed master_bio pointer.
2774 * This link is counted in ->remaining. When the r10_bio that points to NULL
2775 * has its remaining count decremented to 0, the whole complex operation
2780 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2781 int *skipped, int go_faster)
2783 struct r10conf *conf = mddev->private;
2784 struct r10bio *r10_bio;
2785 struct bio *biolist = NULL, *bio;
2786 sector_t max_sector, nr_sectors;
2789 sector_t sync_blocks;
2790 sector_t sectors_skipped = 0;
2791 int chunks_skipped = 0;
2792 sector_t chunk_mask = conf->geo.chunk_mask;
2794 if (!conf->r10buf_pool)
2795 if (init_resync(conf))
2799 max_sector = mddev->dev_sectors;
2800 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2801 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2802 max_sector = mddev->resync_max_sectors;
2803 if (sector_nr >= max_sector) {
2804 /* If we aborted, we need to abort the
2805 * sync on the 'current' bitmap chucks (there can
2806 * be several when recovering multiple devices).
2807 * as we may have started syncing it but not finished.
2808 * We can find the current address in
2809 * mddev->curr_resync, but for recovery,
2810 * we need to convert that to several
2811 * virtual addresses.
2813 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2818 if (mddev->curr_resync < max_sector) { /* aborted */
2819 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2820 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2822 else for (i = 0; i < conf->geo.raid_disks; i++) {
2824 raid10_find_virt(conf, mddev->curr_resync, i);
2825 bitmap_end_sync(mddev->bitmap, sect,
2829 /* completed sync */
2830 if ((!mddev->bitmap || conf->fullsync)
2831 && conf->have_replacement
2832 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2833 /* Completed a full sync so the replacements
2834 * are now fully recovered.
2836 for (i = 0; i < conf->geo.raid_disks; i++)
2837 if (conf->mirrors[i].replacement)
2838 conf->mirrors[i].replacement
2844 bitmap_close_sync(mddev->bitmap);
2847 return sectors_skipped;
2850 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2851 return reshape_request(mddev, sector_nr, skipped);
2853 if (chunks_skipped >= conf->geo.raid_disks) {
2854 /* if there has been nothing to do on any drive,
2855 * then there is nothing to do at all..
2858 return (max_sector - sector_nr) + sectors_skipped;
2861 if (max_sector > mddev->resync_max)
2862 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2864 /* make sure whole request will fit in a chunk - if chunks
2867 if (conf->geo.near_copies < conf->geo.raid_disks &&
2868 max_sector > (sector_nr | chunk_mask))
2869 max_sector = (sector_nr | chunk_mask) + 1;
2871 * If there is non-resync activity waiting for us then
2872 * put in a delay to throttle resync.
2874 if (!go_faster && conf->nr_waiting)
2875 msleep_interruptible(1000);
2877 /* Again, very different code for resync and recovery.
2878 * Both must result in an r10bio with a list of bios that
2879 * have bi_end_io, bi_sector, bi_bdev set,
2880 * and bi_private set to the r10bio.
2881 * For recovery, we may actually create several r10bios
2882 * with 2 bios in each, that correspond to the bios in the main one.
2883 * In this case, the subordinate r10bios link back through a
2884 * borrowed master_bio pointer, and the counter in the master
2885 * includes a ref from each subordinate.
2887 /* First, we decide what to do and set ->bi_end_io
2888 * To end_sync_read if we want to read, and
2889 * end_sync_write if we will want to write.
2892 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2893 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2894 /* recovery... the complicated one */
2898 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2904 struct raid10_info *mirror = &conf->mirrors[i];
2906 if ((mirror->rdev == NULL ||
2907 test_bit(In_sync, &mirror->rdev->flags))
2909 (mirror->replacement == NULL ||
2911 &mirror->replacement->flags)))
2915 /* want to reconstruct this device */
2917 sect = raid10_find_virt(conf, sector_nr, i);
2918 if (sect >= mddev->resync_max_sectors) {
2919 /* last stripe is not complete - don't
2920 * try to recover this sector.
2924 /* Unless we are doing a full sync, or a replacement
2925 * we only need to recover the block if it is set in
2928 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2930 if (sync_blocks < max_sync)
2931 max_sync = sync_blocks;
2933 mirror->replacement == NULL &&
2935 /* yep, skip the sync_blocks here, but don't assume
2936 * that there will never be anything to do here
2938 chunks_skipped = -1;
2942 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2943 raise_barrier(conf, rb2 != NULL);
2944 atomic_set(&r10_bio->remaining, 0);
2946 r10_bio->master_bio = (struct bio*)rb2;
2948 atomic_inc(&rb2->remaining);
2949 r10_bio->mddev = mddev;
2950 set_bit(R10BIO_IsRecover, &r10_bio->state);
2951 r10_bio->sector = sect;
2953 raid10_find_phys(conf, r10_bio);
2955 /* Need to check if the array will still be
2958 for (j = 0; j < conf->geo.raid_disks; j++)
2959 if (conf->mirrors[j].rdev == NULL ||
2960 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2965 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2966 &sync_blocks, still_degraded);
2969 for (j=0; j<conf->copies;j++) {
2971 int d = r10_bio->devs[j].devnum;
2972 sector_t from_addr, to_addr;
2973 struct md_rdev *rdev;
2974 sector_t sector, first_bad;
2976 if (!conf->mirrors[d].rdev ||
2977 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2979 /* This is where we read from */
2981 rdev = conf->mirrors[d].rdev;
2982 sector = r10_bio->devs[j].addr;
2984 if (is_badblock(rdev, sector, max_sync,
2985 &first_bad, &bad_sectors)) {
2986 if (first_bad > sector)
2987 max_sync = first_bad - sector;
2989 bad_sectors -= (sector
2991 if (max_sync > bad_sectors)
2992 max_sync = bad_sectors;
2996 bio = r10_bio->devs[0].bio;
2997 bio->bi_next = biolist;
2999 bio->bi_private = r10_bio;
3000 bio->bi_end_io = end_sync_read;
3002 from_addr = r10_bio->devs[j].addr;
3003 bio->bi_sector = from_addr + rdev->data_offset;
3004 bio->bi_bdev = rdev->bdev;
3005 atomic_inc(&rdev->nr_pending);
3006 /* and we write to 'i' (if not in_sync) */
3008 for (k=0; k<conf->copies; k++)
3009 if (r10_bio->devs[k].devnum == i)
3011 BUG_ON(k == conf->copies);
3012 to_addr = r10_bio->devs[k].addr;
3013 r10_bio->devs[0].devnum = d;
3014 r10_bio->devs[0].addr = from_addr;
3015 r10_bio->devs[1].devnum = i;
3016 r10_bio->devs[1].addr = to_addr;
3018 rdev = mirror->rdev;
3019 if (!test_bit(In_sync, &rdev->flags)) {
3020 bio = r10_bio->devs[1].bio;
3021 bio->bi_next = biolist;
3023 bio->bi_private = r10_bio;
3024 bio->bi_end_io = end_sync_write;
3026 bio->bi_sector = to_addr
3027 + rdev->data_offset;
3028 bio->bi_bdev = rdev->bdev;
3029 atomic_inc(&r10_bio->remaining);
3031 r10_bio->devs[1].bio->bi_end_io = NULL;
3033 /* and maybe write to replacement */
3034 bio = r10_bio->devs[1].repl_bio;
3036 bio->bi_end_io = NULL;
3037 rdev = mirror->replacement;
3038 /* Note: if rdev != NULL, then bio
3039 * cannot be NULL as r10buf_pool_alloc will
3040 * have allocated it.
3041 * So the second test here is pointless.
3042 * But it keeps semantic-checkers happy, and
3043 * this comment keeps human reviewers
3046 if (rdev == NULL || bio == NULL ||
3047 test_bit(Faulty, &rdev->flags))
3049 bio->bi_next = biolist;
3051 bio->bi_private = r10_bio;
3052 bio->bi_end_io = end_sync_write;
3054 bio->bi_sector = to_addr + rdev->data_offset;
3055 bio->bi_bdev = rdev->bdev;
3056 atomic_inc(&r10_bio->remaining);
3059 if (j == conf->copies) {
3060 /* Cannot recover, so abort the recovery or
3061 * record a bad block */
3064 atomic_dec(&rb2->remaining);
3067 /* problem is that there are bad blocks
3068 * on other device(s)
3071 for (k = 0; k < conf->copies; k++)
3072 if (r10_bio->devs[k].devnum == i)
3074 if (!test_bit(In_sync,
3075 &mirror->rdev->flags)
3076 && !rdev_set_badblocks(
3078 r10_bio->devs[k].addr,
3081 if (mirror->replacement &&
3082 !rdev_set_badblocks(
3083 mirror->replacement,
3084 r10_bio->devs[k].addr,
3089 if (!test_and_set_bit(MD_RECOVERY_INTR,
3091 printk(KERN_INFO "md/raid10:%s: insufficient "
3092 "working devices for recovery.\n",
3094 mirror->recovery_disabled
3095 = mddev->recovery_disabled;
3100 if (biolist == NULL) {
3102 struct r10bio *rb2 = r10_bio;
3103 r10_bio = (struct r10bio*) rb2->master_bio;
3104 rb2->master_bio = NULL;
3110 /* resync. Schedule a read for every block at this virt offset */
3113 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3115 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3116 &sync_blocks, mddev->degraded) &&
3117 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3118 &mddev->recovery)) {
3119 /* We can skip this block */
3121 return sync_blocks + sectors_skipped;
3123 if (sync_blocks < max_sync)
3124 max_sync = sync_blocks;
3125 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3127 r10_bio->mddev = mddev;
3128 atomic_set(&r10_bio->remaining, 0);
3129 raise_barrier(conf, 0);
3130 conf->next_resync = sector_nr;
3132 r10_bio->master_bio = NULL;
3133 r10_bio->sector = sector_nr;
3134 set_bit(R10BIO_IsSync, &r10_bio->state);
3135 raid10_find_phys(conf, r10_bio);
3136 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3138 for (i = 0; i < conf->copies; i++) {
3139 int d = r10_bio->devs[i].devnum;
3140 sector_t first_bad, sector;
3143 if (r10_bio->devs[i].repl_bio)
3144 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3146 bio = r10_bio->devs[i].bio;
3147 bio->bi_end_io = NULL;
3148 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3149 if (conf->mirrors[d].rdev == NULL ||
3150 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3152 sector = r10_bio->devs[i].addr;
3153 if (is_badblock(conf->mirrors[d].rdev,
3155 &first_bad, &bad_sectors)) {
3156 if (first_bad > sector)
3157 max_sync = first_bad - sector;
3159 bad_sectors -= (sector - first_bad);
3160 if (max_sync > bad_sectors)
3161 max_sync = max_sync;
3165 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3166 atomic_inc(&r10_bio->remaining);
3167 bio->bi_next = biolist;
3169 bio->bi_private = r10_bio;
3170 bio->bi_end_io = end_sync_read;
3172 bio->bi_sector = sector +
3173 conf->mirrors[d].rdev->data_offset;
3174 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3177 if (conf->mirrors[d].replacement == NULL ||
3179 &conf->mirrors[d].replacement->flags))
3182 /* Need to set up for writing to the replacement */
3183 bio = r10_bio->devs[i].repl_bio;
3184 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3186 sector = r10_bio->devs[i].addr;
3187 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3188 bio->bi_next = biolist;
3190 bio->bi_private = r10_bio;
3191 bio->bi_end_io = end_sync_write;
3193 bio->bi_sector = sector +
3194 conf->mirrors[d].replacement->data_offset;
3195 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3200 for (i=0; i<conf->copies; i++) {
3201 int d = r10_bio->devs[i].devnum;
3202 if (r10_bio->devs[i].bio->bi_end_io)
3203 rdev_dec_pending(conf->mirrors[d].rdev,
3205 if (r10_bio->devs[i].repl_bio &&
3206 r10_bio->devs[i].repl_bio->bi_end_io)
3208 conf->mirrors[d].replacement,
3217 for (bio = biolist; bio ; bio=bio->bi_next) {
3219 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3221 bio->bi_flags |= 1 << BIO_UPTODATE;
3224 bio->bi_phys_segments = 0;
3229 if (sector_nr + max_sync < max_sector)
3230 max_sector = sector_nr + max_sync;
3233 int len = PAGE_SIZE;
3234 if (sector_nr + (len>>9) > max_sector)
3235 len = (max_sector - sector_nr) << 9;
3238 for (bio= biolist ; bio ; bio=bio->bi_next) {
3240 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3241 if (bio_add_page(bio, page, len, 0))
3245 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3246 for (bio2 = biolist;
3247 bio2 && bio2 != bio;
3248 bio2 = bio2->bi_next) {
3249 /* remove last page from this bio */
3251 bio2->bi_size -= len;
3252 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3256 nr_sectors += len>>9;
3257 sector_nr += len>>9;
3258 } while (biolist->bi_vcnt < RESYNC_PAGES);
3260 r10_bio->sectors = nr_sectors;
3264 biolist = biolist->bi_next;
3266 bio->bi_next = NULL;
3267 r10_bio = bio->bi_private;
3268 r10_bio->sectors = nr_sectors;
3270 if (bio->bi_end_io == end_sync_read) {
3271 md_sync_acct(bio->bi_bdev, nr_sectors);
3272 generic_make_request(bio);
3276 if (sectors_skipped)
3277 /* pretend they weren't skipped, it makes
3278 * no important difference in this case
3280 md_done_sync(mddev, sectors_skipped, 1);
3282 return sectors_skipped + nr_sectors;
3284 /* There is nowhere to write, so all non-sync
3285 * drives must be failed or in resync, all drives
3286 * have a bad block, so try the next chunk...
3288 if (sector_nr + max_sync < max_sector)
3289 max_sector = sector_nr + max_sync;
3291 sectors_skipped += (max_sector - sector_nr);
3293 sector_nr = max_sector;
3298 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3301 struct r10conf *conf = mddev->private;
3304 raid_disks = min(conf->geo.raid_disks,
3305 conf->prev.raid_disks);
3307 sectors = conf->dev_sectors;
3309 size = sectors >> conf->geo.chunk_shift;
3310 sector_div(size, conf->geo.far_copies);
3311 size = size * raid_disks;
3312 sector_div(size, conf->geo.near_copies);
3314 return size << conf->geo.chunk_shift;
3317 static void calc_sectors(struct r10conf *conf, sector_t size)
3319 /* Calculate the number of sectors-per-device that will
3320 * actually be used, and set conf->dev_sectors and
3324 size = size >> conf->geo.chunk_shift;
3325 sector_div(size, conf->geo.far_copies);
3326 size = size * conf->geo.raid_disks;
3327 sector_div(size, conf->geo.near_copies);
3328 /* 'size' is now the number of chunks in the array */
3329 /* calculate "used chunks per device" */
3330 size = size * conf->copies;
3332 /* We need to round up when dividing by raid_disks to
3333 * get the stride size.
3335 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3337 conf->dev_sectors = size << conf->geo.chunk_shift;
3339 if (conf->geo.far_offset)
3340 conf->geo.stride = 1 << conf->geo.chunk_shift;
3342 sector_div(size, conf->geo.far_copies);
3343 conf->geo.stride = size << conf->geo.chunk_shift;
3347 enum geo_type {geo_new, geo_old, geo_start};
3348 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3351 int layout, chunk, disks;
3354 layout = mddev->layout;
3355 chunk = mddev->chunk_sectors;
3356 disks = mddev->raid_disks - mddev->delta_disks;
3359 layout = mddev->new_layout;
3360 chunk = mddev->new_chunk_sectors;
3361 disks = mddev->raid_disks;
3363 default: /* avoid 'may be unused' warnings */
3364 case geo_start: /* new when starting reshape - raid_disks not
3366 layout = mddev->new_layout;
3367 chunk = mddev->new_chunk_sectors;
3368 disks = mddev->raid_disks + mddev->delta_disks;
3373 if (chunk < (PAGE_SIZE >> 9) ||
3374 !is_power_of_2(chunk))
3377 fc = (layout >> 8) & 255;
3378 fo = layout & (1<<16);
3379 geo->raid_disks = disks;
3380 geo->near_copies = nc;
3381 geo->far_copies = fc;
3382 geo->far_offset = fo;
3383 geo->chunk_mask = chunk - 1;
3384 geo->chunk_shift = ffz(~chunk);
3388 static struct r10conf *setup_conf(struct mddev *mddev)
3390 struct r10conf *conf = NULL;
3395 copies = setup_geo(&geo, mddev, geo_new);
3398 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3399 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3400 mdname(mddev), PAGE_SIZE);
3404 if (copies < 2 || copies > mddev->raid_disks) {
3405 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3406 mdname(mddev), mddev->new_layout);
3411 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3415 /* FIXME calc properly */
3416 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3417 max(0,mddev->delta_disks)),
3422 conf->tmppage = alloc_page(GFP_KERNEL);
3427 conf->copies = copies;
3428 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3429 r10bio_pool_free, conf);
3430 if (!conf->r10bio_pool)
3433 calc_sectors(conf, mddev->dev_sectors);
3434 if (mddev->reshape_position == MaxSector) {
3435 conf->prev = conf->geo;
3436 conf->reshape_progress = MaxSector;
3438 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3442 conf->reshape_progress = mddev->reshape_position;
3443 if (conf->prev.far_offset)
3444 conf->prev.stride = 1 << conf->prev.chunk_shift;
3446 /* far_copies must be 1 */
3447 conf->prev.stride = conf->dev_sectors;
3449 spin_lock_init(&conf->device_lock);
3450 INIT_LIST_HEAD(&conf->retry_list);
3452 spin_lock_init(&conf->resync_lock);
3453 init_waitqueue_head(&conf->wait_barrier);
3455 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3459 conf->mddev = mddev;
3464 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3467 if (conf->r10bio_pool)
3468 mempool_destroy(conf->r10bio_pool);
3469 kfree(conf->mirrors);
3470 safe_put_page(conf->tmppage);
3473 return ERR_PTR(err);
3476 static int run(struct mddev *mddev)
3478 struct r10conf *conf;
3479 int i, disk_idx, chunk_size;
3480 struct raid10_info *disk;
3481 struct md_rdev *rdev;
3483 sector_t min_offset_diff = 0;
3486 if (mddev->private == NULL) {
3487 conf = setup_conf(mddev);
3489 return PTR_ERR(conf);
3490 mddev->private = conf;
3492 conf = mddev->private;
3496 mddev->thread = conf->thread;
3497 conf->thread = NULL;
3499 chunk_size = mddev->chunk_sectors << 9;
3501 blk_queue_io_min(mddev->queue, chunk_size);
3502 if (conf->geo.raid_disks % conf->geo.near_copies)
3503 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3505 blk_queue_io_opt(mddev->queue, chunk_size *
3506 (conf->geo.raid_disks / conf->geo.near_copies));
3509 rdev_for_each(rdev, mddev) {
3511 struct request_queue *q;
3513 disk_idx = rdev->raid_disk;
3516 if (disk_idx >= conf->geo.raid_disks &&
3517 disk_idx >= conf->prev.raid_disks)
3519 disk = conf->mirrors + disk_idx;
3521 if (test_bit(Replacement, &rdev->flags)) {
3522 if (disk->replacement)
3524 disk->replacement = rdev;
3530 q = bdev_get_queue(rdev->bdev);
3531 if (q->merge_bvec_fn)
3532 mddev->merge_check_needed = 1;
3533 diff = (rdev->new_data_offset - rdev->data_offset);
3534 if (!mddev->reshape_backwards)
3538 if (first || diff < min_offset_diff)
3539 min_offset_diff = diff;
3542 disk_stack_limits(mddev->gendisk, rdev->bdev,
3543 rdev->data_offset << 9);
3545 disk->head_position = 0;
3548 /* need to check that every block has at least one working mirror */
3549 if (!enough(conf, -1)) {
3550 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3555 if (conf->reshape_progress != MaxSector) {
3556 /* must ensure that shape change is supported */
3557 if (conf->geo.far_copies != 1 &&
3558 conf->geo.far_offset == 0)
3560 if (conf->prev.far_copies != 1 &&
3561 conf->geo.far_offset == 0)
3565 mddev->degraded = 0;
3567 i < conf->geo.raid_disks
3568 || i < conf->prev.raid_disks;
3571 disk = conf->mirrors + i;
3573 if (!disk->rdev && disk->replacement) {
3574 /* The replacement is all we have - use it */
3575 disk->rdev = disk->replacement;
3576 disk->replacement = NULL;
3577 clear_bit(Replacement, &disk->rdev->flags);
3581 !test_bit(In_sync, &disk->rdev->flags)) {
3582 disk->head_position = 0;
3587 disk->recovery_disabled = mddev->recovery_disabled - 1;
3590 if (mddev->recovery_cp != MaxSector)
3591 printk(KERN_NOTICE "md/raid10:%s: not clean"
3592 " -- starting background reconstruction\n",
3595 "md/raid10:%s: active with %d out of %d devices\n",
3596 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3597 conf->geo.raid_disks);
3599 * Ok, everything is just fine now
3601 mddev->dev_sectors = conf->dev_sectors;
3602 size = raid10_size(mddev, 0, 0);
3603 md_set_array_sectors(mddev, size);
3604 mddev->resync_max_sectors = size;
3607 int stripe = conf->geo.raid_disks *
3608 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3609 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3610 mddev->queue->backing_dev_info.congested_data = mddev;
3612 /* Calculate max read-ahead size.
3613 * We need to readahead at least twice a whole stripe....
3616 stripe /= conf->geo.near_copies;
3617 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3618 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3619 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3623 if (md_integrity_register(mddev))
3626 if (conf->reshape_progress != MaxSector) {
3627 unsigned long before_length, after_length;
3629 before_length = ((1 << conf->prev.chunk_shift) *
3630 conf->prev.far_copies);
3631 after_length = ((1 << conf->geo.chunk_shift) *
3632 conf->geo.far_copies);
3634 if (max(before_length, after_length) > min_offset_diff) {
3635 /* This cannot work */
3636 printk("md/raid10: offset difference not enough to continue reshape\n");
3639 conf->offset_diff = min_offset_diff;
3641 conf->reshape_safe = conf->reshape_progress;
3642 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3643 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3644 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3645 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3646 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3653 md_unregister_thread(&mddev->thread);
3654 if (conf->r10bio_pool)
3655 mempool_destroy(conf->r10bio_pool);
3656 safe_put_page(conf->tmppage);
3657 kfree(conf->mirrors);
3659 mddev->private = NULL;
3664 static int stop(struct mddev *mddev)
3666 struct r10conf *conf = mddev->private;
3668 raise_barrier(conf, 0);
3669 lower_barrier(conf);
3671 md_unregister_thread(&mddev->thread);
3673 /* the unplug fn references 'conf'*/
3674 blk_sync_queue(mddev->queue);
3676 if (conf->r10bio_pool)
3677 mempool_destroy(conf->r10bio_pool);
3678 kfree(conf->mirrors);
3680 mddev->private = NULL;
3684 static void raid10_quiesce(struct mddev *mddev, int state)
3686 struct r10conf *conf = mddev->private;
3690 raise_barrier(conf, 0);
3693 lower_barrier(conf);
3698 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3700 /* Resize of 'far' arrays is not supported.
3701 * For 'near' and 'offset' arrays we can set the
3702 * number of sectors used to be an appropriate multiple
3703 * of the chunk size.
3704 * For 'offset', this is far_copies*chunksize.
3705 * For 'near' the multiplier is the LCM of
3706 * near_copies and raid_disks.
3707 * So if far_copies > 1 && !far_offset, fail.
3708 * Else find LCM(raid_disks, near_copy)*far_copies and
3709 * multiply by chunk_size. Then round to this number.
3710 * This is mostly done by raid10_size()
3712 struct r10conf *conf = mddev->private;
3713 sector_t oldsize, size;
3715 if (mddev->reshape_position != MaxSector)
3718 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3721 oldsize = raid10_size(mddev, 0, 0);
3722 size = raid10_size(mddev, sectors, 0);
3723 if (mddev->external_size &&
3724 mddev->array_sectors > size)
3726 if (mddev->bitmap) {
3727 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3731 md_set_array_sectors(mddev, size);
3732 set_capacity(mddev->gendisk, mddev->array_sectors);
3733 revalidate_disk(mddev->gendisk);
3734 if (sectors > mddev->dev_sectors &&
3735 mddev->recovery_cp > oldsize) {
3736 mddev->recovery_cp = oldsize;
3737 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3739 calc_sectors(conf, sectors);
3740 mddev->dev_sectors = conf->dev_sectors;
3741 mddev->resync_max_sectors = size;
3745 static void *raid10_takeover_raid0(struct mddev *mddev)
3747 struct md_rdev *rdev;
3748 struct r10conf *conf;
3750 if (mddev->degraded > 0) {
3751 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3753 return ERR_PTR(-EINVAL);
3756 /* Set new parameters */
3757 mddev->new_level = 10;
3758 /* new layout: far_copies = 1, near_copies = 2 */
3759 mddev->new_layout = (1<<8) + 2;
3760 mddev->new_chunk_sectors = mddev->chunk_sectors;
3761 mddev->delta_disks = mddev->raid_disks;
3762 mddev->raid_disks *= 2;
3763 /* make sure it will be not marked as dirty */
3764 mddev->recovery_cp = MaxSector;
3766 conf = setup_conf(mddev);
3767 if (!IS_ERR(conf)) {
3768 rdev_for_each(rdev, mddev)
3769 if (rdev->raid_disk >= 0)
3770 rdev->new_raid_disk = rdev->raid_disk * 2;
3777 static void *raid10_takeover(struct mddev *mddev)
3779 struct r0conf *raid0_conf;
3781 /* raid10 can take over:
3782 * raid0 - providing it has only two drives
3784 if (mddev->level == 0) {
3785 /* for raid0 takeover only one zone is supported */
3786 raid0_conf = mddev->private;
3787 if (raid0_conf->nr_strip_zones > 1) {
3788 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3789 " with more than one zone.\n",
3791 return ERR_PTR(-EINVAL);
3793 return raid10_takeover_raid0(mddev);
3795 return ERR_PTR(-EINVAL);
3798 static int raid10_check_reshape(struct mddev *mddev)
3800 /* Called when there is a request to change
3801 * - layout (to ->new_layout)
3802 * - chunk size (to ->new_chunk_sectors)
3803 * - raid_disks (by delta_disks)
3804 * or when trying to restart a reshape that was ongoing.
3806 * We need to validate the request and possibly allocate
3807 * space if that might be an issue later.
3809 * Currently we reject any reshape of a 'far' mode array,
3810 * allow chunk size to change if new is generally acceptable,
3811 * allow raid_disks to increase, and allow
3812 * a switch between 'near' mode and 'offset' mode.
3814 struct r10conf *conf = mddev->private;
3817 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3820 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3821 /* mustn't change number of copies */
3823 if (geo.far_copies > 1 && !geo.far_offset)
3824 /* Cannot switch to 'far' mode */
3827 if (mddev->array_sectors & geo.chunk_mask)
3828 /* not factor of array size */
3831 if (!enough(conf, -1))
3834 kfree(conf->mirrors_new);
3835 conf->mirrors_new = NULL;
3836 if (mddev->delta_disks > 0) {
3837 /* allocate new 'mirrors' list */
3838 conf->mirrors_new = kzalloc(
3839 sizeof(struct raid10_info)
3840 *(mddev->raid_disks +
3841 mddev->delta_disks),
3843 if (!conf->mirrors_new)
3850 * Need to check if array has failed when deciding whether to:
3852 * - remove non-faulty devices
3855 * This determination is simple when no reshape is happening.
3856 * However if there is a reshape, we need to carefully check
3857 * both the before and after sections.
3858 * This is because some failed devices may only affect one
3859 * of the two sections, and some non-in_sync devices may
3860 * be insync in the section most affected by failed devices.
3862 static int calc_degraded(struct r10conf *conf)
3864 int degraded, degraded2;
3869 /* 'prev' section first */
3870 for (i = 0; i < conf->prev.raid_disks; i++) {
3871 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3872 if (!rdev || test_bit(Faulty, &rdev->flags))
3874 else if (!test_bit(In_sync, &rdev->flags))
3875 /* When we can reduce the number of devices in
3876 * an array, this might not contribute to
3877 * 'degraded'. It does now.
3882 if (conf->geo.raid_disks == conf->prev.raid_disks)
3886 for (i = 0; i < conf->geo.raid_disks; i++) {
3887 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3888 if (!rdev || test_bit(Faulty, &rdev->flags))
3890 else if (!test_bit(In_sync, &rdev->flags)) {
3891 /* If reshape is increasing the number of devices,
3892 * this section has already been recovered, so
3893 * it doesn't contribute to degraded.
3896 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3901 if (degraded2 > degraded)
3906 static int raid10_start_reshape(struct mddev *mddev)
3908 /* A 'reshape' has been requested. This commits
3909 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3910 * This also checks if there are enough spares and adds them
3912 * We currently require enough spares to make the final
3913 * array non-degraded. We also require that the difference
3914 * between old and new data_offset - on each device - is
3915 * enough that we never risk over-writing.
3918 unsigned long before_length, after_length;
3919 sector_t min_offset_diff = 0;
3922 struct r10conf *conf = mddev->private;
3923 struct md_rdev *rdev;
3927 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3930 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3933 before_length = ((1 << conf->prev.chunk_shift) *
3934 conf->prev.far_copies);
3935 after_length = ((1 << conf->geo.chunk_shift) *
3936 conf->geo.far_copies);
3938 rdev_for_each(rdev, mddev) {
3939 if (!test_bit(In_sync, &rdev->flags)
3940 && !test_bit(Faulty, &rdev->flags))
3942 if (rdev->raid_disk >= 0) {
3943 long long diff = (rdev->new_data_offset
3944 - rdev->data_offset);
3945 if (!mddev->reshape_backwards)
3949 if (first || diff < min_offset_diff)
3950 min_offset_diff = diff;
3954 if (max(before_length, after_length) > min_offset_diff)
3957 if (spares < mddev->delta_disks)
3960 conf->offset_diff = min_offset_diff;
3961 spin_lock_irq(&conf->device_lock);
3962 if (conf->mirrors_new) {
3963 memcpy(conf->mirrors_new, conf->mirrors,
3964 sizeof(struct raid10_info)*conf->prev.raid_disks);
3966 kfree(conf->mirrors_old); /* FIXME and elsewhere */
3967 conf->mirrors_old = conf->mirrors;
3968 conf->mirrors = conf->mirrors_new;
3969 conf->mirrors_new = NULL;
3971 setup_geo(&conf->geo, mddev, geo_start);
3973 if (mddev->reshape_backwards) {
3974 sector_t size = raid10_size(mddev, 0, 0);
3975 if (size < mddev->array_sectors) {
3976 spin_unlock_irq(&conf->device_lock);
3977 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3981 mddev->resync_max_sectors = size;
3982 conf->reshape_progress = size;
3984 conf->reshape_progress = 0;
3985 spin_unlock_irq(&conf->device_lock);
3987 if (mddev->delta_disks && mddev->bitmap) {
3988 ret = bitmap_resize(mddev->bitmap,
3989 raid10_size(mddev, 0,
3990 conf->geo.raid_disks),
3995 if (mddev->delta_disks > 0) {
3996 rdev_for_each(rdev, mddev)
3997 if (rdev->raid_disk < 0 &&
3998 !test_bit(Faulty, &rdev->flags)) {
3999 if (raid10_add_disk(mddev, rdev) == 0) {
4000 if (rdev->raid_disk >=
4001 conf->prev.raid_disks)
4002 set_bit(In_sync, &rdev->flags);
4004 rdev->recovery_offset = 0;
4006 if (sysfs_link_rdev(mddev, rdev))
4007 /* Failure here is OK */;
4009 } else if (rdev->raid_disk >= conf->prev.raid_disks
4010 && !test_bit(Faulty, &rdev->flags)) {
4011 /* This is a spare that was manually added */
4012 set_bit(In_sync, &rdev->flags);
4015 /* When a reshape changes the number of devices,
4016 * ->degraded is measured against the larger of the
4017 * pre and post numbers.
4019 spin_lock_irq(&conf->device_lock);
4020 mddev->degraded = calc_degraded(conf);
4021 spin_unlock_irq(&conf->device_lock);
4022 mddev->raid_disks = conf->geo.raid_disks;
4023 mddev->reshape_position = conf->reshape_progress;
4024 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4026 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4027 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4028 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4029 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4031 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4033 if (!mddev->sync_thread) {
4037 conf->reshape_checkpoint = jiffies;
4038 md_wakeup_thread(mddev->sync_thread);
4039 md_new_event(mddev);
4043 mddev->recovery = 0;
4044 spin_lock_irq(&conf->device_lock);
4045 conf->geo = conf->prev;
4046 mddev->raid_disks = conf->geo.raid_disks;
4047 rdev_for_each(rdev, mddev)
4048 rdev->new_data_offset = rdev->data_offset;
4050 conf->reshape_progress = MaxSector;
4051 mddev->reshape_position = MaxSector;
4052 spin_unlock_irq(&conf->device_lock);
4056 /* Calculate the last device-address that could contain
4057 * any block from the chunk that includes the array-address 's'
4058 * and report the next address.
4059 * i.e. the address returned will be chunk-aligned and after
4060 * any data that is in the chunk containing 's'.
4062 static sector_t last_dev_address(sector_t s, struct geom *geo)
4064 s = (s | geo->chunk_mask) + 1;
4065 s >>= geo->chunk_shift;
4066 s *= geo->near_copies;
4067 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4068 s *= geo->far_copies;
4069 s <<= geo->chunk_shift;
4073 /* Calculate the first device-address that could contain
4074 * any block from the chunk that includes the array-address 's'.
4075 * This too will be the start of a chunk
4077 static sector_t first_dev_address(sector_t s, struct geom *geo)
4079 s >>= geo->chunk_shift;
4080 s *= geo->near_copies;
4081 sector_div(s, geo->raid_disks);
4082 s *= geo->far_copies;
4083 s <<= geo->chunk_shift;
4087 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4090 /* We simply copy at most one chunk (smallest of old and new)
4091 * at a time, possibly less if that exceeds RESYNC_PAGES,
4092 * or we hit a bad block or something.
4093 * This might mean we pause for normal IO in the middle of
4094 * a chunk, but that is not a problem was mddev->reshape_position
4095 * can record any location.
4097 * If we will want to write to a location that isn't
4098 * yet recorded as 'safe' (i.e. in metadata on disk) then
4099 * we need to flush all reshape requests and update the metadata.
4101 * When reshaping forwards (e.g. to more devices), we interpret
4102 * 'safe' as the earliest block which might not have been copied
4103 * down yet. We divide this by previous stripe size and multiply
4104 * by previous stripe length to get lowest device offset that we
4105 * cannot write to yet.
4106 * We interpret 'sector_nr' as an address that we want to write to.
4107 * From this we use last_device_address() to find where we might
4108 * write to, and first_device_address on the 'safe' position.
4109 * If this 'next' write position is after the 'safe' position,
4110 * we must update the metadata to increase the 'safe' position.
4112 * When reshaping backwards, we round in the opposite direction
4113 * and perform the reverse test: next write position must not be
4114 * less than current safe position.
4116 * In all this the minimum difference in data offsets
4117 * (conf->offset_diff - always positive) allows a bit of slack,
4118 * so next can be after 'safe', but not by more than offset_disk
4120 * We need to prepare all the bios here before we start any IO
4121 * to ensure the size we choose is acceptable to all devices.
4122 * The means one for each copy for write-out and an extra one for
4124 * We store the read-in bio in ->master_bio and the others in
4125 * ->devs[x].bio and ->devs[x].repl_bio.
4127 struct r10conf *conf = mddev->private;
4128 struct r10bio *r10_bio;
4129 sector_t next, safe, last;
4133 struct md_rdev *rdev;
4136 struct bio *bio, *read_bio;
4137 int sectors_done = 0;
4139 if (sector_nr == 0) {
4140 /* If restarting in the middle, skip the initial sectors */
4141 if (mddev->reshape_backwards &&
4142 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4143 sector_nr = (raid10_size(mddev, 0, 0)
4144 - conf->reshape_progress);
4145 } else if (!mddev->reshape_backwards &&
4146 conf->reshape_progress > 0)
4147 sector_nr = conf->reshape_progress;
4149 mddev->curr_resync_completed = sector_nr;
4150 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4156 /* We don't use sector_nr to track where we are up to
4157 * as that doesn't work well for ->reshape_backwards.
4158 * So just use ->reshape_progress.
4160 if (mddev->reshape_backwards) {
4161 /* 'next' is the earliest device address that we might
4162 * write to for this chunk in the new layout
4164 next = first_dev_address(conf->reshape_progress - 1,
4167 /* 'safe' is the last device address that we might read from
4168 * in the old layout after a restart
4170 safe = last_dev_address(conf->reshape_safe - 1,
4173 if (next + conf->offset_diff < safe)
4176 last = conf->reshape_progress - 1;
4177 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4178 & conf->prev.chunk_mask);
4179 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4180 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4182 /* 'next' is after the last device address that we
4183 * might write to for this chunk in the new layout
4185 next = last_dev_address(conf->reshape_progress, &conf->geo);
4187 /* 'safe' is the earliest device address that we might
4188 * read from in the old layout after a restart
4190 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4192 /* Need to update metadata if 'next' might be beyond 'safe'
4193 * as that would possibly corrupt data
4195 if (next > safe + conf->offset_diff)
4198 sector_nr = conf->reshape_progress;
4199 last = sector_nr | (conf->geo.chunk_mask
4200 & conf->prev.chunk_mask);
4202 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4203 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4207 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4208 /* Need to update reshape_position in metadata */
4210 mddev->reshape_position = conf->reshape_progress;
4211 if (mddev->reshape_backwards)
4212 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4213 - conf->reshape_progress;
4215 mddev->curr_resync_completed = conf->reshape_progress;
4216 conf->reshape_checkpoint = jiffies;
4217 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4218 md_wakeup_thread(mddev->thread);
4219 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4220 kthread_should_stop());
4221 conf->reshape_safe = mddev->reshape_position;
4222 allow_barrier(conf);
4226 /* Now schedule reads for blocks from sector_nr to last */
4227 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4228 raise_barrier(conf, sectors_done != 0);
4229 atomic_set(&r10_bio->remaining, 0);
4230 r10_bio->mddev = mddev;
4231 r10_bio->sector = sector_nr;
4232 set_bit(R10BIO_IsReshape, &r10_bio->state);
4233 r10_bio->sectors = last - sector_nr + 1;
4234 rdev = read_balance(conf, r10_bio, &max_sectors);
4235 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4238 /* Cannot read from here, so need to record bad blocks
4239 * on all the target devices.
4242 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4243 return sectors_done;
4246 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4248 read_bio->bi_bdev = rdev->bdev;
4249 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4250 + rdev->data_offset);
4251 read_bio->bi_private = r10_bio;
4252 read_bio->bi_end_io = end_sync_read;
4253 read_bio->bi_rw = READ;
4254 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4255 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4256 read_bio->bi_vcnt = 0;
4257 read_bio->bi_idx = 0;
4258 read_bio->bi_size = 0;
4259 r10_bio->master_bio = read_bio;
4260 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4262 /* Now find the locations in the new layout */
4263 __raid10_find_phys(&conf->geo, r10_bio);
4266 read_bio->bi_next = NULL;
4268 for (s = 0; s < conf->copies*2; s++) {
4270 int d = r10_bio->devs[s/2].devnum;
4271 struct md_rdev *rdev2;
4273 rdev2 = conf->mirrors[d].replacement;
4274 b = r10_bio->devs[s/2].repl_bio;
4276 rdev2 = conf->mirrors[d].rdev;
4277 b = r10_bio->devs[s/2].bio;
4279 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4281 b->bi_bdev = rdev2->bdev;
4282 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4283 b->bi_private = r10_bio;
4284 b->bi_end_io = end_reshape_write;
4286 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4287 b->bi_flags |= 1 << BIO_UPTODATE;
4295 /* Now add as many pages as possible to all of these bios. */
4298 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4299 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4300 int len = (max_sectors - s) << 9;
4301 if (len > PAGE_SIZE)
4303 for (bio = blist; bio ; bio = bio->bi_next) {
4305 if (bio_add_page(bio, page, len, 0))
4308 /* Didn't fit, must stop */
4310 bio2 && bio2 != bio;
4311 bio2 = bio2->bi_next) {
4312 /* Remove last page from this bio */
4314 bio2->bi_size -= len;
4315 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4319 sector_nr += len >> 9;
4320 nr_sectors += len >> 9;
4323 r10_bio->sectors = nr_sectors;
4325 /* Now submit the read */
4326 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4327 atomic_inc(&r10_bio->remaining);
4328 read_bio->bi_next = NULL;
4329 generic_make_request(read_bio);
4330 sector_nr += nr_sectors;
4331 sectors_done += nr_sectors;
4332 if (sector_nr <= last)
4335 /* Now that we have done the whole section we can
4336 * update reshape_progress
4338 if (mddev->reshape_backwards)
4339 conf->reshape_progress -= sectors_done;
4341 conf->reshape_progress += sectors_done;
4343 return sectors_done;
4346 static void end_reshape_request(struct r10bio *r10_bio);
4347 static int handle_reshape_read_error(struct mddev *mddev,
4348 struct r10bio *r10_bio);
4349 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4351 /* Reshape read completed. Hopefully we have a block
4353 * If we got a read error then we do sync 1-page reads from
4354 * elsewhere until we find the data - or give up.
4356 struct r10conf *conf = mddev->private;
4359 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4360 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4361 /* Reshape has been aborted */
4362 md_done_sync(mddev, r10_bio->sectors, 0);
4366 /* We definitely have the data in the pages, schedule the
4369 atomic_set(&r10_bio->remaining, 1);
4370 for (s = 0; s < conf->copies*2; s++) {
4372 int d = r10_bio->devs[s/2].devnum;
4373 struct md_rdev *rdev;
4375 rdev = conf->mirrors[d].replacement;
4376 b = r10_bio->devs[s/2].repl_bio;
4378 rdev = conf->mirrors[d].rdev;
4379 b = r10_bio->devs[s/2].bio;
4381 if (!rdev || test_bit(Faulty, &rdev->flags))
4383 atomic_inc(&rdev->nr_pending);
4384 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4385 atomic_inc(&r10_bio->remaining);
4387 generic_make_request(b);
4389 end_reshape_request(r10_bio);
4392 static void end_reshape(struct r10conf *conf)
4394 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4397 spin_lock_irq(&conf->device_lock);
4398 conf->prev = conf->geo;
4399 md_finish_reshape(conf->mddev);
4401 conf->reshape_progress = MaxSector;
4402 spin_unlock_irq(&conf->device_lock);
4404 /* read-ahead size must cover two whole stripes, which is
4405 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4407 if (conf->mddev->queue) {
4408 int stripe = conf->geo.raid_disks *
4409 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4410 stripe /= conf->geo.near_copies;
4411 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4412 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4418 static int handle_reshape_read_error(struct mddev *mddev,
4419 struct r10bio *r10_bio)
4421 /* Use sync reads to get the blocks from somewhere else */
4422 int sectors = r10_bio->sectors;
4423 struct r10conf *conf = mddev->private;
4425 struct r10bio r10_bio;
4426 struct r10dev devs[conf->copies];
4428 struct r10bio *r10b = &on_stack.r10_bio;
4431 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4433 r10b->sector = r10_bio->sector;
4434 __raid10_find_phys(&conf->prev, r10b);
4439 int first_slot = slot;
4441 if (s > (PAGE_SIZE >> 9))
4445 int d = r10b->devs[slot].devnum;
4446 struct md_rdev *rdev = conf->mirrors[d].rdev;
4449 test_bit(Faulty, &rdev->flags) ||
4450 !test_bit(In_sync, &rdev->flags))
4453 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4454 success = sync_page_io(rdev,
4463 if (slot >= conf->copies)
4465 if (slot == first_slot)
4469 /* couldn't read this block, must give up */
4470 set_bit(MD_RECOVERY_INTR,
4480 static void end_reshape_write(struct bio *bio, int error)
4482 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4483 struct r10bio *r10_bio = bio->bi_private;
4484 struct mddev *mddev = r10_bio->mddev;
4485 struct r10conf *conf = mddev->private;
4489 struct md_rdev *rdev = NULL;
4491 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4493 rdev = conf->mirrors[d].replacement;
4496 rdev = conf->mirrors[d].rdev;
4500 /* FIXME should record badblock */
4501 md_error(mddev, rdev);
4504 rdev_dec_pending(rdev, mddev);
4505 end_reshape_request(r10_bio);
4508 static void end_reshape_request(struct r10bio *r10_bio)
4510 if (!atomic_dec_and_test(&r10_bio->remaining))
4512 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4513 bio_put(r10_bio->master_bio);
4517 static void raid10_finish_reshape(struct mddev *mddev)
4519 struct r10conf *conf = mddev->private;
4521 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4524 if (mddev->delta_disks > 0) {
4525 sector_t size = raid10_size(mddev, 0, 0);
4526 md_set_array_sectors(mddev, size);
4527 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4528 mddev->recovery_cp = mddev->resync_max_sectors;
4529 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4531 mddev->resync_max_sectors = size;
4532 set_capacity(mddev->gendisk, mddev->array_sectors);
4533 revalidate_disk(mddev->gendisk);
4536 for (d = conf->geo.raid_disks ;
4537 d < conf->geo.raid_disks - mddev->delta_disks;
4539 struct md_rdev *rdev = conf->mirrors[d].rdev;
4541 clear_bit(In_sync, &rdev->flags);
4542 rdev = conf->mirrors[d].replacement;
4544 clear_bit(In_sync, &rdev->flags);
4547 mddev->layout = mddev->new_layout;
4548 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4549 mddev->reshape_position = MaxSector;
4550 mddev->delta_disks = 0;
4551 mddev->reshape_backwards = 0;
4554 static struct md_personality raid10_personality =
4558 .owner = THIS_MODULE,
4559 .make_request = make_request,
4563 .error_handler = error,
4564 .hot_add_disk = raid10_add_disk,
4565 .hot_remove_disk= raid10_remove_disk,
4566 .spare_active = raid10_spare_active,
4567 .sync_request = sync_request,
4568 .quiesce = raid10_quiesce,
4569 .size = raid10_size,
4570 .resize = raid10_resize,
4571 .takeover = raid10_takeover,
4572 .check_reshape = raid10_check_reshape,
4573 .start_reshape = raid10_start_reshape,
4574 .finish_reshape = raid10_finish_reshape,
4577 static int __init raid_init(void)
4579 return register_md_personality(&raid10_personality);
4582 static void raid_exit(void)
4584 unregister_md_personality(&raid10_personality);
4587 module_init(raid_init);
4588 module_exit(raid_exit);
4589 MODULE_LICENSE("GPL");
4590 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4591 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4592 MODULE_ALIAS("md-raid10");
4593 MODULE_ALIAS("md-level-10");
4595 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);