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;
1516 if (conf->mirrors[first].rdev &&
1519 first = (first+1) % geo->raid_disks;
1523 } while (first != 0);
1527 static int enough(struct r10conf *conf, int ignore)
1529 return _enough(conf, &conf->geo, ignore) &&
1530 _enough(conf, &conf->prev, ignore);
1533 static void error(struct mddev *mddev, struct md_rdev *rdev)
1535 char b[BDEVNAME_SIZE];
1536 struct r10conf *conf = mddev->private;
1539 * If it is not operational, then we have already marked it as dead
1540 * else if it is the last working disks, ignore the error, let the
1541 * next level up know.
1542 * else mark the drive as failed
1544 if (test_bit(In_sync, &rdev->flags)
1545 && !enough(conf, rdev->raid_disk))
1547 * Don't fail the drive, just return an IO error.
1550 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1551 unsigned long flags;
1552 spin_lock_irqsave(&conf->device_lock, flags);
1554 spin_unlock_irqrestore(&conf->device_lock, flags);
1556 * if recovery is running, make sure it aborts.
1558 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1560 set_bit(Blocked, &rdev->flags);
1561 set_bit(Faulty, &rdev->flags);
1562 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1564 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1565 "md/raid10:%s: Operation continuing on %d devices.\n",
1566 mdname(mddev), bdevname(rdev->bdev, b),
1567 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1570 static void print_conf(struct r10conf *conf)
1573 struct raid10_info *tmp;
1575 printk(KERN_DEBUG "RAID10 conf printout:\n");
1577 printk(KERN_DEBUG "(!conf)\n");
1580 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1581 conf->geo.raid_disks);
1583 for (i = 0; i < conf->geo.raid_disks; i++) {
1584 char b[BDEVNAME_SIZE];
1585 tmp = conf->mirrors + i;
1587 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1588 i, !test_bit(In_sync, &tmp->rdev->flags),
1589 !test_bit(Faulty, &tmp->rdev->flags),
1590 bdevname(tmp->rdev->bdev,b));
1594 static void close_sync(struct r10conf *conf)
1597 allow_barrier(conf);
1599 mempool_destroy(conf->r10buf_pool);
1600 conf->r10buf_pool = NULL;
1603 static int raid10_spare_active(struct mddev *mddev)
1606 struct r10conf *conf = mddev->private;
1607 struct raid10_info *tmp;
1609 unsigned long flags;
1612 * Find all non-in_sync disks within the RAID10 configuration
1613 * and mark them in_sync
1615 for (i = 0; i < conf->geo.raid_disks; i++) {
1616 tmp = conf->mirrors + i;
1617 if (tmp->replacement
1618 && tmp->replacement->recovery_offset == MaxSector
1619 && !test_bit(Faulty, &tmp->replacement->flags)
1620 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1621 /* Replacement has just become active */
1623 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1626 /* Replaced device not technically faulty,
1627 * but we need to be sure it gets removed
1628 * and never re-added.
1630 set_bit(Faulty, &tmp->rdev->flags);
1631 sysfs_notify_dirent_safe(
1632 tmp->rdev->sysfs_state);
1634 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1635 } else if (tmp->rdev
1636 && !test_bit(Faulty, &tmp->rdev->flags)
1637 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1639 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1642 spin_lock_irqsave(&conf->device_lock, flags);
1643 mddev->degraded -= count;
1644 spin_unlock_irqrestore(&conf->device_lock, flags);
1651 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1653 struct r10conf *conf = mddev->private;
1657 int last = conf->geo.raid_disks - 1;
1658 struct request_queue *q = bdev_get_queue(rdev->bdev);
1660 if (mddev->recovery_cp < MaxSector)
1661 /* only hot-add to in-sync arrays, as recovery is
1662 * very different from resync
1665 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1668 if (rdev->raid_disk >= 0)
1669 first = last = rdev->raid_disk;
1671 if (q->merge_bvec_fn) {
1672 set_bit(Unmerged, &rdev->flags);
1673 mddev->merge_check_needed = 1;
1676 if (rdev->saved_raid_disk >= first &&
1677 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1678 mirror = rdev->saved_raid_disk;
1681 for ( ; mirror <= last ; mirror++) {
1682 struct raid10_info *p = &conf->mirrors[mirror];
1683 if (p->recovery_disabled == mddev->recovery_disabled)
1686 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1687 p->replacement != NULL)
1689 clear_bit(In_sync, &rdev->flags);
1690 set_bit(Replacement, &rdev->flags);
1691 rdev->raid_disk = mirror;
1693 disk_stack_limits(mddev->gendisk, rdev->bdev,
1694 rdev->data_offset << 9);
1696 rcu_assign_pointer(p->replacement, rdev);
1700 disk_stack_limits(mddev->gendisk, rdev->bdev,
1701 rdev->data_offset << 9);
1703 p->head_position = 0;
1704 p->recovery_disabled = mddev->recovery_disabled - 1;
1705 rdev->raid_disk = mirror;
1707 if (rdev->saved_raid_disk != mirror)
1709 rcu_assign_pointer(p->rdev, rdev);
1712 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1713 /* Some requests might not have seen this new
1714 * merge_bvec_fn. We must wait for them to complete
1715 * before merging the device fully.
1716 * First we make sure any code which has tested
1717 * our function has submitted the request, then
1718 * we wait for all outstanding requests to complete.
1720 synchronize_sched();
1721 raise_barrier(conf, 0);
1722 lower_barrier(conf);
1723 clear_bit(Unmerged, &rdev->flags);
1725 md_integrity_add_rdev(rdev, mddev);
1730 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1732 struct r10conf *conf = mddev->private;
1734 int number = rdev->raid_disk;
1735 struct md_rdev **rdevp;
1736 struct raid10_info *p = conf->mirrors + number;
1739 if (rdev == p->rdev)
1741 else if (rdev == p->replacement)
1742 rdevp = &p->replacement;
1746 if (test_bit(In_sync, &rdev->flags) ||
1747 atomic_read(&rdev->nr_pending)) {
1751 /* Only remove faulty devices if recovery
1754 if (!test_bit(Faulty, &rdev->flags) &&
1755 mddev->recovery_disabled != p->recovery_disabled &&
1756 (!p->replacement || p->replacement == rdev) &&
1757 number < conf->geo.raid_disks &&
1764 if (atomic_read(&rdev->nr_pending)) {
1765 /* lost the race, try later */
1769 } else if (p->replacement) {
1770 /* We must have just cleared 'rdev' */
1771 p->rdev = p->replacement;
1772 clear_bit(Replacement, &p->replacement->flags);
1773 smp_mb(); /* Make sure other CPUs may see both as identical
1774 * but will never see neither -- if they are careful.
1776 p->replacement = NULL;
1777 clear_bit(WantReplacement, &rdev->flags);
1779 /* We might have just remove the Replacement as faulty
1780 * Clear the flag just in case
1782 clear_bit(WantReplacement, &rdev->flags);
1784 err = md_integrity_register(mddev);
1793 static void end_sync_read(struct bio *bio, int error)
1795 struct r10bio *r10_bio = bio->bi_private;
1796 struct r10conf *conf = r10_bio->mddev->private;
1799 if (bio == r10_bio->master_bio) {
1800 /* this is a reshape read */
1801 d = r10_bio->read_slot; /* really the read dev */
1803 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1805 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1806 set_bit(R10BIO_Uptodate, &r10_bio->state);
1808 /* The write handler will notice the lack of
1809 * R10BIO_Uptodate and record any errors etc
1811 atomic_add(r10_bio->sectors,
1812 &conf->mirrors[d].rdev->corrected_errors);
1814 /* for reconstruct, we always reschedule after a read.
1815 * for resync, only after all reads
1817 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1818 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1819 atomic_dec_and_test(&r10_bio->remaining)) {
1820 /* we have read all the blocks,
1821 * do the comparison in process context in raid10d
1823 reschedule_retry(r10_bio);
1827 static void end_sync_request(struct r10bio *r10_bio)
1829 struct mddev *mddev = r10_bio->mddev;
1831 while (atomic_dec_and_test(&r10_bio->remaining)) {
1832 if (r10_bio->master_bio == NULL) {
1833 /* the primary of several recovery bios */
1834 sector_t s = r10_bio->sectors;
1835 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1836 test_bit(R10BIO_WriteError, &r10_bio->state))
1837 reschedule_retry(r10_bio);
1840 md_done_sync(mddev, s, 1);
1843 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1844 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1845 test_bit(R10BIO_WriteError, &r10_bio->state))
1846 reschedule_retry(r10_bio);
1854 static void end_sync_write(struct bio *bio, int error)
1856 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1857 struct r10bio *r10_bio = bio->bi_private;
1858 struct mddev *mddev = r10_bio->mddev;
1859 struct r10conf *conf = mddev->private;
1865 struct md_rdev *rdev = NULL;
1867 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1869 rdev = conf->mirrors[d].replacement;
1871 rdev = conf->mirrors[d].rdev;
1875 md_error(mddev, rdev);
1877 set_bit(WriteErrorSeen, &rdev->flags);
1878 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1879 set_bit(MD_RECOVERY_NEEDED,
1880 &rdev->mddev->recovery);
1881 set_bit(R10BIO_WriteError, &r10_bio->state);
1883 } else if (is_badblock(rdev,
1884 r10_bio->devs[slot].addr,
1886 &first_bad, &bad_sectors))
1887 set_bit(R10BIO_MadeGood, &r10_bio->state);
1889 rdev_dec_pending(rdev, mddev);
1891 end_sync_request(r10_bio);
1895 * Note: sync and recover and handled very differently for raid10
1896 * This code is for resync.
1897 * For resync, we read through virtual addresses and read all blocks.
1898 * If there is any error, we schedule a write. The lowest numbered
1899 * drive is authoritative.
1900 * However requests come for physical address, so we need to map.
1901 * For every physical address there are raid_disks/copies virtual addresses,
1902 * which is always are least one, but is not necessarly an integer.
1903 * This means that a physical address can span multiple chunks, so we may
1904 * have to submit multiple io requests for a single sync request.
1907 * We check if all blocks are in-sync and only write to blocks that
1910 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1912 struct r10conf *conf = mddev->private;
1914 struct bio *tbio, *fbio;
1917 atomic_set(&r10_bio->remaining, 1);
1919 /* find the first device with a block */
1920 for (i=0; i<conf->copies; i++)
1921 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1924 if (i == conf->copies)
1928 fbio = r10_bio->devs[i].bio;
1930 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1931 /* now find blocks with errors */
1932 for (i=0 ; i < conf->copies ; i++) {
1935 tbio = r10_bio->devs[i].bio;
1937 if (tbio->bi_end_io != end_sync_read)
1941 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1942 /* We know that the bi_io_vec layout is the same for
1943 * both 'first' and 'i', so we just compare them.
1944 * All vec entries are PAGE_SIZE;
1946 for (j = 0; j < vcnt; j++)
1947 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1948 page_address(tbio->bi_io_vec[j].bv_page),
1949 fbio->bi_io_vec[j].bv_len))
1953 mddev->resync_mismatches += r10_bio->sectors;
1954 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1955 /* Don't fix anything. */
1958 /* Ok, we need to write this bio, either to correct an
1959 * inconsistency or to correct an unreadable block.
1960 * First we need to fixup bv_offset, bv_len and
1961 * bi_vecs, as the read request might have corrupted these
1963 tbio->bi_vcnt = vcnt;
1964 tbio->bi_size = r10_bio->sectors << 9;
1966 tbio->bi_phys_segments = 0;
1967 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1968 tbio->bi_flags |= 1 << BIO_UPTODATE;
1969 tbio->bi_next = NULL;
1970 tbio->bi_rw = WRITE;
1971 tbio->bi_private = r10_bio;
1972 tbio->bi_sector = r10_bio->devs[i].addr;
1974 for (j=0; j < vcnt ; j++) {
1975 tbio->bi_io_vec[j].bv_offset = 0;
1976 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1978 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1979 page_address(fbio->bi_io_vec[j].bv_page),
1982 tbio->bi_end_io = end_sync_write;
1984 d = r10_bio->devs[i].devnum;
1985 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1986 atomic_inc(&r10_bio->remaining);
1987 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1989 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1990 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1991 generic_make_request(tbio);
1994 /* Now write out to any replacement devices
1997 for (i = 0; i < conf->copies; i++) {
2000 tbio = r10_bio->devs[i].repl_bio;
2001 if (!tbio || !tbio->bi_end_io)
2003 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2004 && r10_bio->devs[i].bio != fbio)
2005 for (j = 0; j < vcnt; j++)
2006 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2007 page_address(fbio->bi_io_vec[j].bv_page),
2009 d = r10_bio->devs[i].devnum;
2010 atomic_inc(&r10_bio->remaining);
2011 md_sync_acct(conf->mirrors[d].replacement->bdev,
2012 tbio->bi_size >> 9);
2013 generic_make_request(tbio);
2017 if (atomic_dec_and_test(&r10_bio->remaining)) {
2018 md_done_sync(mddev, r10_bio->sectors, 1);
2024 * Now for the recovery code.
2025 * Recovery happens across physical sectors.
2026 * We recover all non-is_sync drives by finding the virtual address of
2027 * each, and then choose a working drive that also has that virt address.
2028 * There is a separate r10_bio for each non-in_sync drive.
2029 * Only the first two slots are in use. The first for reading,
2030 * The second for writing.
2033 static void fix_recovery_read_error(struct r10bio *r10_bio)
2035 /* We got a read error during recovery.
2036 * We repeat the read in smaller page-sized sections.
2037 * If a read succeeds, write it to the new device or record
2038 * a bad block if we cannot.
2039 * If a read fails, record a bad block on both old and
2042 struct mddev *mddev = r10_bio->mddev;
2043 struct r10conf *conf = mddev->private;
2044 struct bio *bio = r10_bio->devs[0].bio;
2046 int sectors = r10_bio->sectors;
2048 int dr = r10_bio->devs[0].devnum;
2049 int dw = r10_bio->devs[1].devnum;
2053 struct md_rdev *rdev;
2057 if (s > (PAGE_SIZE>>9))
2060 rdev = conf->mirrors[dr].rdev;
2061 addr = r10_bio->devs[0].addr + sect,
2062 ok = sync_page_io(rdev,
2065 bio->bi_io_vec[idx].bv_page,
2068 rdev = conf->mirrors[dw].rdev;
2069 addr = r10_bio->devs[1].addr + sect;
2070 ok = sync_page_io(rdev,
2073 bio->bi_io_vec[idx].bv_page,
2076 set_bit(WriteErrorSeen, &rdev->flags);
2077 if (!test_and_set_bit(WantReplacement,
2079 set_bit(MD_RECOVERY_NEEDED,
2080 &rdev->mddev->recovery);
2084 /* We don't worry if we cannot set a bad block -
2085 * it really is bad so there is no loss in not
2088 rdev_set_badblocks(rdev, addr, s, 0);
2090 if (rdev != conf->mirrors[dw].rdev) {
2091 /* need bad block on destination too */
2092 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2093 addr = r10_bio->devs[1].addr + sect;
2094 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2096 /* just abort the recovery */
2098 "md/raid10:%s: recovery aborted"
2099 " due to read error\n",
2102 conf->mirrors[dw].recovery_disabled
2103 = mddev->recovery_disabled;
2104 set_bit(MD_RECOVERY_INTR,
2117 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2119 struct r10conf *conf = mddev->private;
2121 struct bio *wbio, *wbio2;
2123 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2124 fix_recovery_read_error(r10_bio);
2125 end_sync_request(r10_bio);
2130 * share the pages with the first bio
2131 * and submit the write request
2133 d = r10_bio->devs[1].devnum;
2134 wbio = r10_bio->devs[1].bio;
2135 wbio2 = r10_bio->devs[1].repl_bio;
2136 if (wbio->bi_end_io) {
2137 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2138 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2139 generic_make_request(wbio);
2141 if (wbio2 && wbio2->bi_end_io) {
2142 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2143 md_sync_acct(conf->mirrors[d].replacement->bdev,
2144 wbio2->bi_size >> 9);
2145 generic_make_request(wbio2);
2151 * Used by fix_read_error() to decay the per rdev read_errors.
2152 * We halve the read error count for every hour that has elapsed
2153 * since the last recorded read error.
2156 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2158 struct timespec cur_time_mon;
2159 unsigned long hours_since_last;
2160 unsigned int read_errors = atomic_read(&rdev->read_errors);
2162 ktime_get_ts(&cur_time_mon);
2164 if (rdev->last_read_error.tv_sec == 0 &&
2165 rdev->last_read_error.tv_nsec == 0) {
2166 /* first time we've seen a read error */
2167 rdev->last_read_error = cur_time_mon;
2171 hours_since_last = (cur_time_mon.tv_sec -
2172 rdev->last_read_error.tv_sec) / 3600;
2174 rdev->last_read_error = cur_time_mon;
2177 * if hours_since_last is > the number of bits in read_errors
2178 * just set read errors to 0. We do this to avoid
2179 * overflowing the shift of read_errors by hours_since_last.
2181 if (hours_since_last >= 8 * sizeof(read_errors))
2182 atomic_set(&rdev->read_errors, 0);
2184 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2187 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2188 int sectors, struct page *page, int rw)
2193 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2194 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2196 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2200 set_bit(WriteErrorSeen, &rdev->flags);
2201 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2202 set_bit(MD_RECOVERY_NEEDED,
2203 &rdev->mddev->recovery);
2205 /* need to record an error - either for the block or the device */
2206 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2207 md_error(rdev->mddev, rdev);
2212 * This is a kernel thread which:
2214 * 1. Retries failed read operations on working mirrors.
2215 * 2. Updates the raid superblock when problems encounter.
2216 * 3. Performs writes following reads for array synchronising.
2219 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2221 int sect = 0; /* Offset from r10_bio->sector */
2222 int sectors = r10_bio->sectors;
2223 struct md_rdev*rdev;
2224 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2225 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2227 /* still own a reference to this rdev, so it cannot
2228 * have been cleared recently.
2230 rdev = conf->mirrors[d].rdev;
2232 if (test_bit(Faulty, &rdev->flags))
2233 /* drive has already been failed, just ignore any
2234 more fix_read_error() attempts */
2237 check_decay_read_errors(mddev, rdev);
2238 atomic_inc(&rdev->read_errors);
2239 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2240 char b[BDEVNAME_SIZE];
2241 bdevname(rdev->bdev, b);
2244 "md/raid10:%s: %s: Raid device exceeded "
2245 "read_error threshold [cur %d:max %d]\n",
2247 atomic_read(&rdev->read_errors), max_read_errors);
2249 "md/raid10:%s: %s: Failing raid device\n",
2251 md_error(mddev, conf->mirrors[d].rdev);
2252 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2258 int sl = r10_bio->read_slot;
2262 if (s > (PAGE_SIZE>>9))
2270 d = r10_bio->devs[sl].devnum;
2271 rdev = rcu_dereference(conf->mirrors[d].rdev);
2273 !test_bit(Unmerged, &rdev->flags) &&
2274 test_bit(In_sync, &rdev->flags) &&
2275 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2276 &first_bad, &bad_sectors) == 0) {
2277 atomic_inc(&rdev->nr_pending);
2279 success = sync_page_io(rdev,
2280 r10_bio->devs[sl].addr +
2283 conf->tmppage, READ, false);
2284 rdev_dec_pending(rdev, mddev);
2290 if (sl == conf->copies)
2292 } while (!success && sl != r10_bio->read_slot);
2296 /* Cannot read from anywhere, just mark the block
2297 * as bad on the first device to discourage future
2300 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2301 rdev = conf->mirrors[dn].rdev;
2303 if (!rdev_set_badblocks(
2305 r10_bio->devs[r10_bio->read_slot].addr
2308 md_error(mddev, rdev);
2309 r10_bio->devs[r10_bio->read_slot].bio
2316 /* write it back and re-read */
2318 while (sl != r10_bio->read_slot) {
2319 char b[BDEVNAME_SIZE];
2324 d = r10_bio->devs[sl].devnum;
2325 rdev = rcu_dereference(conf->mirrors[d].rdev);
2327 test_bit(Unmerged, &rdev->flags) ||
2328 !test_bit(In_sync, &rdev->flags))
2331 atomic_inc(&rdev->nr_pending);
2333 if (r10_sync_page_io(rdev,
2334 r10_bio->devs[sl].addr +
2336 s, conf->tmppage, WRITE)
2338 /* Well, this device is dead */
2340 "md/raid10:%s: read correction "
2342 " (%d sectors at %llu on %s)\n",
2344 (unsigned long long)(
2346 choose_data_offset(r10_bio,
2348 bdevname(rdev->bdev, b));
2349 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2352 bdevname(rdev->bdev, b));
2354 rdev_dec_pending(rdev, mddev);
2358 while (sl != r10_bio->read_slot) {
2359 char b[BDEVNAME_SIZE];
2364 d = r10_bio->devs[sl].devnum;
2365 rdev = rcu_dereference(conf->mirrors[d].rdev);
2367 !test_bit(In_sync, &rdev->flags))
2370 atomic_inc(&rdev->nr_pending);
2372 switch (r10_sync_page_io(rdev,
2373 r10_bio->devs[sl].addr +
2378 /* Well, this device is dead */
2380 "md/raid10:%s: unable to read back "
2382 " (%d sectors at %llu on %s)\n",
2384 (unsigned long long)(
2386 choose_data_offset(r10_bio, rdev)),
2387 bdevname(rdev->bdev, b));
2388 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2391 bdevname(rdev->bdev, b));
2395 "md/raid10:%s: read error corrected"
2396 " (%d sectors at %llu on %s)\n",
2398 (unsigned long long)(
2400 choose_data_offset(r10_bio, rdev)),
2401 bdevname(rdev->bdev, b));
2402 atomic_add(s, &rdev->corrected_errors);
2405 rdev_dec_pending(rdev, mddev);
2415 static void bi_complete(struct bio *bio, int error)
2417 complete((struct completion *)bio->bi_private);
2420 static int submit_bio_wait(int rw, struct bio *bio)
2422 struct completion event;
2425 init_completion(&event);
2426 bio->bi_private = &event;
2427 bio->bi_end_io = bi_complete;
2428 submit_bio(rw, bio);
2429 wait_for_completion(&event);
2431 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2434 static int narrow_write_error(struct r10bio *r10_bio, int i)
2436 struct bio *bio = r10_bio->master_bio;
2437 struct mddev *mddev = r10_bio->mddev;
2438 struct r10conf *conf = mddev->private;
2439 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2440 /* bio has the data to be written to slot 'i' where
2441 * we just recently had a write error.
2442 * We repeatedly clone the bio and trim down to one block,
2443 * then try the write. Where the write fails we record
2445 * It is conceivable that the bio doesn't exactly align with
2446 * blocks. We must handle this.
2448 * We currently own a reference to the rdev.
2454 int sect_to_write = r10_bio->sectors;
2457 if (rdev->badblocks.shift < 0)
2460 block_sectors = 1 << rdev->badblocks.shift;
2461 sector = r10_bio->sector;
2462 sectors = ((r10_bio->sector + block_sectors)
2463 & ~(sector_t)(block_sectors - 1))
2466 while (sect_to_write) {
2468 if (sectors > sect_to_write)
2469 sectors = sect_to_write;
2470 /* Write at 'sector' for 'sectors' */
2471 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2472 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2473 wbio->bi_sector = (r10_bio->devs[i].addr+
2474 choose_data_offset(r10_bio, rdev) +
2475 (sector - r10_bio->sector));
2476 wbio->bi_bdev = rdev->bdev;
2477 if (submit_bio_wait(WRITE, wbio) == 0)
2479 ok = rdev_set_badblocks(rdev, sector,
2484 sect_to_write -= sectors;
2486 sectors = block_sectors;
2491 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2493 int slot = r10_bio->read_slot;
2495 struct r10conf *conf = mddev->private;
2496 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2497 char b[BDEVNAME_SIZE];
2498 unsigned long do_sync;
2501 /* we got a read error. Maybe the drive is bad. Maybe just
2502 * the block and we can fix it.
2503 * We freeze all other IO, and try reading the block from
2504 * other devices. When we find one, we re-write
2505 * and check it that fixes the read error.
2506 * This is all done synchronously while the array is
2509 bio = r10_bio->devs[slot].bio;
2510 bdevname(bio->bi_bdev, b);
2512 r10_bio->devs[slot].bio = NULL;
2514 if (mddev->ro == 0) {
2516 fix_read_error(conf, mddev, r10_bio);
2517 unfreeze_array(conf);
2519 r10_bio->devs[slot].bio = IO_BLOCKED;
2521 rdev_dec_pending(rdev, mddev);
2524 rdev = read_balance(conf, r10_bio, &max_sectors);
2526 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2527 " read error for block %llu\n",
2529 (unsigned long long)r10_bio->sector);
2530 raid_end_bio_io(r10_bio);
2534 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2535 slot = r10_bio->read_slot;
2538 "md/raid10:%s: %s: redirecting "
2539 "sector %llu to another mirror\n",
2541 bdevname(rdev->bdev, b),
2542 (unsigned long long)r10_bio->sector);
2543 bio = bio_clone_mddev(r10_bio->master_bio,
2546 r10_bio->sector - bio->bi_sector,
2548 r10_bio->devs[slot].bio = bio;
2549 r10_bio->devs[slot].rdev = rdev;
2550 bio->bi_sector = r10_bio->devs[slot].addr
2551 + choose_data_offset(r10_bio, rdev);
2552 bio->bi_bdev = rdev->bdev;
2553 bio->bi_rw = READ | do_sync;
2554 bio->bi_private = r10_bio;
2555 bio->bi_end_io = raid10_end_read_request;
2556 if (max_sectors < r10_bio->sectors) {
2557 /* Drat - have to split this up more */
2558 struct bio *mbio = r10_bio->master_bio;
2559 int sectors_handled =
2560 r10_bio->sector + max_sectors
2562 r10_bio->sectors = max_sectors;
2563 spin_lock_irq(&conf->device_lock);
2564 if (mbio->bi_phys_segments == 0)
2565 mbio->bi_phys_segments = 2;
2567 mbio->bi_phys_segments++;
2568 spin_unlock_irq(&conf->device_lock);
2569 generic_make_request(bio);
2571 r10_bio = mempool_alloc(conf->r10bio_pool,
2573 r10_bio->master_bio = mbio;
2574 r10_bio->sectors = (mbio->bi_size >> 9)
2577 set_bit(R10BIO_ReadError,
2579 r10_bio->mddev = mddev;
2580 r10_bio->sector = mbio->bi_sector
2585 generic_make_request(bio);
2588 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2590 /* Some sort of write request has finished and it
2591 * succeeded in writing where we thought there was a
2592 * bad block. So forget the bad block.
2593 * Or possibly if failed and we need to record
2597 struct md_rdev *rdev;
2599 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2600 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2601 for (m = 0; m < conf->copies; m++) {
2602 int dev = r10_bio->devs[m].devnum;
2603 rdev = conf->mirrors[dev].rdev;
2604 if (r10_bio->devs[m].bio == NULL)
2606 if (test_bit(BIO_UPTODATE,
2607 &r10_bio->devs[m].bio->bi_flags)) {
2608 rdev_clear_badblocks(
2610 r10_bio->devs[m].addr,
2611 r10_bio->sectors, 0);
2613 if (!rdev_set_badblocks(
2615 r10_bio->devs[m].addr,
2616 r10_bio->sectors, 0))
2617 md_error(conf->mddev, rdev);
2619 rdev = conf->mirrors[dev].replacement;
2620 if (r10_bio->devs[m].repl_bio == NULL)
2622 if (test_bit(BIO_UPTODATE,
2623 &r10_bio->devs[m].repl_bio->bi_flags)) {
2624 rdev_clear_badblocks(
2626 r10_bio->devs[m].addr,
2627 r10_bio->sectors, 0);
2629 if (!rdev_set_badblocks(
2631 r10_bio->devs[m].addr,
2632 r10_bio->sectors, 0))
2633 md_error(conf->mddev, rdev);
2638 for (m = 0; m < conf->copies; m++) {
2639 int dev = r10_bio->devs[m].devnum;
2640 struct bio *bio = r10_bio->devs[m].bio;
2641 rdev = conf->mirrors[dev].rdev;
2642 if (bio == IO_MADE_GOOD) {
2643 rdev_clear_badblocks(
2645 r10_bio->devs[m].addr,
2646 r10_bio->sectors, 0);
2647 rdev_dec_pending(rdev, conf->mddev);
2648 } else if (bio != NULL &&
2649 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2650 if (!narrow_write_error(r10_bio, m)) {
2651 md_error(conf->mddev, rdev);
2652 set_bit(R10BIO_Degraded,
2655 rdev_dec_pending(rdev, conf->mddev);
2657 bio = r10_bio->devs[m].repl_bio;
2658 rdev = conf->mirrors[dev].replacement;
2659 if (rdev && bio == IO_MADE_GOOD) {
2660 rdev_clear_badblocks(
2662 r10_bio->devs[m].addr,
2663 r10_bio->sectors, 0);
2664 rdev_dec_pending(rdev, conf->mddev);
2667 if (test_bit(R10BIO_WriteError,
2669 close_write(r10_bio);
2670 raid_end_bio_io(r10_bio);
2674 static void raid10d(struct mddev *mddev)
2676 struct r10bio *r10_bio;
2677 unsigned long flags;
2678 struct r10conf *conf = mddev->private;
2679 struct list_head *head = &conf->retry_list;
2680 struct blk_plug plug;
2682 md_check_recovery(mddev);
2684 blk_start_plug(&plug);
2687 flush_pending_writes(conf);
2689 spin_lock_irqsave(&conf->device_lock, flags);
2690 if (list_empty(head)) {
2691 spin_unlock_irqrestore(&conf->device_lock, flags);
2694 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2695 list_del(head->prev);
2697 spin_unlock_irqrestore(&conf->device_lock, flags);
2699 mddev = r10_bio->mddev;
2700 conf = mddev->private;
2701 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2702 test_bit(R10BIO_WriteError, &r10_bio->state))
2703 handle_write_completed(conf, r10_bio);
2704 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2705 reshape_request_write(mddev, r10_bio);
2706 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2707 sync_request_write(mddev, r10_bio);
2708 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2709 recovery_request_write(mddev, r10_bio);
2710 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2711 handle_read_error(mddev, r10_bio);
2713 /* just a partial read to be scheduled from a
2716 int slot = r10_bio->read_slot;
2717 generic_make_request(r10_bio->devs[slot].bio);
2721 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2722 md_check_recovery(mddev);
2724 blk_finish_plug(&plug);
2728 static int init_resync(struct r10conf *conf)
2733 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2734 BUG_ON(conf->r10buf_pool);
2735 conf->have_replacement = 0;
2736 for (i = 0; i < conf->geo.raid_disks; i++)
2737 if (conf->mirrors[i].replacement)
2738 conf->have_replacement = 1;
2739 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2740 if (!conf->r10buf_pool)
2742 conf->next_resync = 0;
2747 * perform a "sync" on one "block"
2749 * We need to make sure that no normal I/O request - particularly write
2750 * requests - conflict with active sync requests.
2752 * This is achieved by tracking pending requests and a 'barrier' concept
2753 * that can be installed to exclude normal IO requests.
2755 * Resync and recovery are handled very differently.
2756 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2758 * For resync, we iterate over virtual addresses, read all copies,
2759 * and update if there are differences. If only one copy is live,
2761 * For recovery, we iterate over physical addresses, read a good
2762 * value for each non-in_sync drive, and over-write.
2764 * So, for recovery we may have several outstanding complex requests for a
2765 * given address, one for each out-of-sync device. We model this by allocating
2766 * a number of r10_bio structures, one for each out-of-sync device.
2767 * As we setup these structures, we collect all bio's together into a list
2768 * which we then process collectively to add pages, and then process again
2769 * to pass to generic_make_request.
2771 * The r10_bio structures are linked using a borrowed master_bio pointer.
2772 * This link is counted in ->remaining. When the r10_bio that points to NULL
2773 * has its remaining count decremented to 0, the whole complex operation
2778 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2779 int *skipped, int go_faster)
2781 struct r10conf *conf = mddev->private;
2782 struct r10bio *r10_bio;
2783 struct bio *biolist = NULL, *bio;
2784 sector_t max_sector, nr_sectors;
2787 sector_t sync_blocks;
2788 sector_t sectors_skipped = 0;
2789 int chunks_skipped = 0;
2790 sector_t chunk_mask = conf->geo.chunk_mask;
2792 if (!conf->r10buf_pool)
2793 if (init_resync(conf))
2797 max_sector = mddev->dev_sectors;
2798 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2799 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2800 max_sector = mddev->resync_max_sectors;
2801 if (sector_nr >= max_sector) {
2802 /* If we aborted, we need to abort the
2803 * sync on the 'current' bitmap chucks (there can
2804 * be several when recovering multiple devices).
2805 * as we may have started syncing it but not finished.
2806 * We can find the current address in
2807 * mddev->curr_resync, but for recovery,
2808 * we need to convert that to several
2809 * virtual addresses.
2811 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2816 if (mddev->curr_resync < max_sector) { /* aborted */
2817 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2818 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2820 else for (i = 0; i < conf->geo.raid_disks; i++) {
2822 raid10_find_virt(conf, mddev->curr_resync, i);
2823 bitmap_end_sync(mddev->bitmap, sect,
2827 /* completed sync */
2828 if ((!mddev->bitmap || conf->fullsync)
2829 && conf->have_replacement
2830 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2831 /* Completed a full sync so the replacements
2832 * are now fully recovered.
2834 for (i = 0; i < conf->geo.raid_disks; i++)
2835 if (conf->mirrors[i].replacement)
2836 conf->mirrors[i].replacement
2842 bitmap_close_sync(mddev->bitmap);
2845 return sectors_skipped;
2848 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2849 return reshape_request(mddev, sector_nr, skipped);
2851 if (chunks_skipped >= conf->geo.raid_disks) {
2852 /* if there has been nothing to do on any drive,
2853 * then there is nothing to do at all..
2856 return (max_sector - sector_nr) + sectors_skipped;
2859 if (max_sector > mddev->resync_max)
2860 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2862 /* make sure whole request will fit in a chunk - if chunks
2865 if (conf->geo.near_copies < conf->geo.raid_disks &&
2866 max_sector > (sector_nr | chunk_mask))
2867 max_sector = (sector_nr | chunk_mask) + 1;
2869 * If there is non-resync activity waiting for us then
2870 * put in a delay to throttle resync.
2872 if (!go_faster && conf->nr_waiting)
2873 msleep_interruptible(1000);
2875 /* Again, very different code for resync and recovery.
2876 * Both must result in an r10bio with a list of bios that
2877 * have bi_end_io, bi_sector, bi_bdev set,
2878 * and bi_private set to the r10bio.
2879 * For recovery, we may actually create several r10bios
2880 * with 2 bios in each, that correspond to the bios in the main one.
2881 * In this case, the subordinate r10bios link back through a
2882 * borrowed master_bio pointer, and the counter in the master
2883 * includes a ref from each subordinate.
2885 /* First, we decide what to do and set ->bi_end_io
2886 * To end_sync_read if we want to read, and
2887 * end_sync_write if we will want to write.
2890 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2891 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2892 /* recovery... the complicated one */
2896 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2902 struct raid10_info *mirror = &conf->mirrors[i];
2904 if ((mirror->rdev == NULL ||
2905 test_bit(In_sync, &mirror->rdev->flags))
2907 (mirror->replacement == NULL ||
2909 &mirror->replacement->flags)))
2913 /* want to reconstruct this device */
2915 sect = raid10_find_virt(conf, sector_nr, i);
2916 if (sect >= mddev->resync_max_sectors) {
2917 /* last stripe is not complete - don't
2918 * try to recover this sector.
2922 /* Unless we are doing a full sync, or a replacement
2923 * we only need to recover the block if it is set in
2926 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2928 if (sync_blocks < max_sync)
2929 max_sync = sync_blocks;
2931 mirror->replacement == NULL &&
2933 /* yep, skip the sync_blocks here, but don't assume
2934 * that there will never be anything to do here
2936 chunks_skipped = -1;
2940 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2941 raise_barrier(conf, rb2 != NULL);
2942 atomic_set(&r10_bio->remaining, 0);
2944 r10_bio->master_bio = (struct bio*)rb2;
2946 atomic_inc(&rb2->remaining);
2947 r10_bio->mddev = mddev;
2948 set_bit(R10BIO_IsRecover, &r10_bio->state);
2949 r10_bio->sector = sect;
2951 raid10_find_phys(conf, r10_bio);
2953 /* Need to check if the array will still be
2956 for (j = 0; j < conf->geo.raid_disks; j++)
2957 if (conf->mirrors[j].rdev == NULL ||
2958 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2963 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2964 &sync_blocks, still_degraded);
2967 for (j=0; j<conf->copies;j++) {
2969 int d = r10_bio->devs[j].devnum;
2970 sector_t from_addr, to_addr;
2971 struct md_rdev *rdev;
2972 sector_t sector, first_bad;
2974 if (!conf->mirrors[d].rdev ||
2975 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2977 /* This is where we read from */
2979 rdev = conf->mirrors[d].rdev;
2980 sector = r10_bio->devs[j].addr;
2982 if (is_badblock(rdev, sector, max_sync,
2983 &first_bad, &bad_sectors)) {
2984 if (first_bad > sector)
2985 max_sync = first_bad - sector;
2987 bad_sectors -= (sector
2989 if (max_sync > bad_sectors)
2990 max_sync = bad_sectors;
2994 bio = r10_bio->devs[0].bio;
2995 bio->bi_next = biolist;
2997 bio->bi_private = r10_bio;
2998 bio->bi_end_io = end_sync_read;
3000 from_addr = r10_bio->devs[j].addr;
3001 bio->bi_sector = from_addr + rdev->data_offset;
3002 bio->bi_bdev = rdev->bdev;
3003 atomic_inc(&rdev->nr_pending);
3004 /* and we write to 'i' (if not in_sync) */
3006 for (k=0; k<conf->copies; k++)
3007 if (r10_bio->devs[k].devnum == i)
3009 BUG_ON(k == conf->copies);
3010 to_addr = r10_bio->devs[k].addr;
3011 r10_bio->devs[0].devnum = d;
3012 r10_bio->devs[0].addr = from_addr;
3013 r10_bio->devs[1].devnum = i;
3014 r10_bio->devs[1].addr = to_addr;
3016 rdev = mirror->rdev;
3017 if (!test_bit(In_sync, &rdev->flags)) {
3018 bio = r10_bio->devs[1].bio;
3019 bio->bi_next = biolist;
3021 bio->bi_private = r10_bio;
3022 bio->bi_end_io = end_sync_write;
3024 bio->bi_sector = to_addr
3025 + rdev->data_offset;
3026 bio->bi_bdev = rdev->bdev;
3027 atomic_inc(&r10_bio->remaining);
3029 r10_bio->devs[1].bio->bi_end_io = NULL;
3031 /* and maybe write to replacement */
3032 bio = r10_bio->devs[1].repl_bio;
3034 bio->bi_end_io = NULL;
3035 rdev = mirror->replacement;
3036 /* Note: if rdev != NULL, then bio
3037 * cannot be NULL as r10buf_pool_alloc will
3038 * have allocated it.
3039 * So the second test here is pointless.
3040 * But it keeps semantic-checkers happy, and
3041 * this comment keeps human reviewers
3044 if (rdev == NULL || bio == NULL ||
3045 test_bit(Faulty, &rdev->flags))
3047 bio->bi_next = biolist;
3049 bio->bi_private = r10_bio;
3050 bio->bi_end_io = end_sync_write;
3052 bio->bi_sector = to_addr + rdev->data_offset;
3053 bio->bi_bdev = rdev->bdev;
3054 atomic_inc(&r10_bio->remaining);
3057 if (j == conf->copies) {
3058 /* Cannot recover, so abort the recovery or
3059 * record a bad block */
3062 atomic_dec(&rb2->remaining);
3065 /* problem is that there are bad blocks
3066 * on other device(s)
3069 for (k = 0; k < conf->copies; k++)
3070 if (r10_bio->devs[k].devnum == i)
3072 if (!test_bit(In_sync,
3073 &mirror->rdev->flags)
3074 && !rdev_set_badblocks(
3076 r10_bio->devs[k].addr,
3079 if (mirror->replacement &&
3080 !rdev_set_badblocks(
3081 mirror->replacement,
3082 r10_bio->devs[k].addr,
3087 if (!test_and_set_bit(MD_RECOVERY_INTR,
3089 printk(KERN_INFO "md/raid10:%s: insufficient "
3090 "working devices for recovery.\n",
3092 mirror->recovery_disabled
3093 = mddev->recovery_disabled;
3098 if (biolist == NULL) {
3100 struct r10bio *rb2 = r10_bio;
3101 r10_bio = (struct r10bio*) rb2->master_bio;
3102 rb2->master_bio = NULL;
3108 /* resync. Schedule a read for every block at this virt offset */
3111 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3113 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3114 &sync_blocks, mddev->degraded) &&
3115 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3116 &mddev->recovery)) {
3117 /* We can skip this block */
3119 return sync_blocks + sectors_skipped;
3121 if (sync_blocks < max_sync)
3122 max_sync = sync_blocks;
3123 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3125 r10_bio->mddev = mddev;
3126 atomic_set(&r10_bio->remaining, 0);
3127 raise_barrier(conf, 0);
3128 conf->next_resync = sector_nr;
3130 r10_bio->master_bio = NULL;
3131 r10_bio->sector = sector_nr;
3132 set_bit(R10BIO_IsSync, &r10_bio->state);
3133 raid10_find_phys(conf, r10_bio);
3134 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3136 for (i = 0; i < conf->copies; i++) {
3137 int d = r10_bio->devs[i].devnum;
3138 sector_t first_bad, sector;
3141 if (r10_bio->devs[i].repl_bio)
3142 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3144 bio = r10_bio->devs[i].bio;
3145 bio->bi_end_io = NULL;
3146 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3147 if (conf->mirrors[d].rdev == NULL ||
3148 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3150 sector = r10_bio->devs[i].addr;
3151 if (is_badblock(conf->mirrors[d].rdev,
3153 &first_bad, &bad_sectors)) {
3154 if (first_bad > sector)
3155 max_sync = first_bad - sector;
3157 bad_sectors -= (sector - first_bad);
3158 if (max_sync > bad_sectors)
3159 max_sync = max_sync;
3163 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3164 atomic_inc(&r10_bio->remaining);
3165 bio->bi_next = biolist;
3167 bio->bi_private = r10_bio;
3168 bio->bi_end_io = end_sync_read;
3170 bio->bi_sector = sector +
3171 conf->mirrors[d].rdev->data_offset;
3172 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3175 if (conf->mirrors[d].replacement == NULL ||
3177 &conf->mirrors[d].replacement->flags))
3180 /* Need to set up for writing to the replacement */
3181 bio = r10_bio->devs[i].repl_bio;
3182 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3184 sector = r10_bio->devs[i].addr;
3185 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3186 bio->bi_next = biolist;
3188 bio->bi_private = r10_bio;
3189 bio->bi_end_io = end_sync_write;
3191 bio->bi_sector = sector +
3192 conf->mirrors[d].replacement->data_offset;
3193 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3198 for (i=0; i<conf->copies; i++) {
3199 int d = r10_bio->devs[i].devnum;
3200 if (r10_bio->devs[i].bio->bi_end_io)
3201 rdev_dec_pending(conf->mirrors[d].rdev,
3203 if (r10_bio->devs[i].repl_bio &&
3204 r10_bio->devs[i].repl_bio->bi_end_io)
3206 conf->mirrors[d].replacement,
3215 for (bio = biolist; bio ; bio=bio->bi_next) {
3217 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3219 bio->bi_flags |= 1 << BIO_UPTODATE;
3222 bio->bi_phys_segments = 0;
3227 if (sector_nr + max_sync < max_sector)
3228 max_sector = sector_nr + max_sync;
3231 int len = PAGE_SIZE;
3232 if (sector_nr + (len>>9) > max_sector)
3233 len = (max_sector - sector_nr) << 9;
3236 for (bio= biolist ; bio ; bio=bio->bi_next) {
3238 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3239 if (bio_add_page(bio, page, len, 0))
3243 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3244 for (bio2 = biolist;
3245 bio2 && bio2 != bio;
3246 bio2 = bio2->bi_next) {
3247 /* remove last page from this bio */
3249 bio2->bi_size -= len;
3250 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3254 nr_sectors += len>>9;
3255 sector_nr += len>>9;
3256 } while (biolist->bi_vcnt < RESYNC_PAGES);
3258 r10_bio->sectors = nr_sectors;
3262 biolist = biolist->bi_next;
3264 bio->bi_next = NULL;
3265 r10_bio = bio->bi_private;
3266 r10_bio->sectors = nr_sectors;
3268 if (bio->bi_end_io == end_sync_read) {
3269 md_sync_acct(bio->bi_bdev, nr_sectors);
3270 generic_make_request(bio);
3274 if (sectors_skipped)
3275 /* pretend they weren't skipped, it makes
3276 * no important difference in this case
3278 md_done_sync(mddev, sectors_skipped, 1);
3280 return sectors_skipped + nr_sectors;
3282 /* There is nowhere to write, so all non-sync
3283 * drives must be failed or in resync, all drives
3284 * have a bad block, so try the next chunk...
3286 if (sector_nr + max_sync < max_sector)
3287 max_sector = sector_nr + max_sync;
3289 sectors_skipped += (max_sector - sector_nr);
3291 sector_nr = max_sector;
3296 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3299 struct r10conf *conf = mddev->private;
3302 raid_disks = min(conf->geo.raid_disks,
3303 conf->prev.raid_disks);
3305 sectors = conf->dev_sectors;
3307 size = sectors >> conf->geo.chunk_shift;
3308 sector_div(size, conf->geo.far_copies);
3309 size = size * raid_disks;
3310 sector_div(size, conf->geo.near_copies);
3312 return size << conf->geo.chunk_shift;
3315 static void calc_sectors(struct r10conf *conf, sector_t size)
3317 /* Calculate the number of sectors-per-device that will
3318 * actually be used, and set conf->dev_sectors and
3322 size = size >> conf->geo.chunk_shift;
3323 sector_div(size, conf->geo.far_copies);
3324 size = size * conf->geo.raid_disks;
3325 sector_div(size, conf->geo.near_copies);
3326 /* 'size' is now the number of chunks in the array */
3327 /* calculate "used chunks per device" */
3328 size = size * conf->copies;
3330 /* We need to round up when dividing by raid_disks to
3331 * get the stride size.
3333 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3335 conf->dev_sectors = size << conf->geo.chunk_shift;
3337 if (conf->geo.far_offset)
3338 conf->geo.stride = 1 << conf->geo.chunk_shift;
3340 sector_div(size, conf->geo.far_copies);
3341 conf->geo.stride = size << conf->geo.chunk_shift;
3345 enum geo_type {geo_new, geo_old, geo_start};
3346 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3349 int layout, chunk, disks;
3352 layout = mddev->layout;
3353 chunk = mddev->chunk_sectors;
3354 disks = mddev->raid_disks - mddev->delta_disks;
3357 layout = mddev->new_layout;
3358 chunk = mddev->new_chunk_sectors;
3359 disks = mddev->raid_disks;
3361 default: /* avoid 'may be unused' warnings */
3362 case geo_start: /* new when starting reshape - raid_disks not
3364 layout = mddev->new_layout;
3365 chunk = mddev->new_chunk_sectors;
3366 disks = mddev->raid_disks + mddev->delta_disks;
3371 if (chunk < (PAGE_SIZE >> 9) ||
3372 !is_power_of_2(chunk))
3375 fc = (layout >> 8) & 255;
3376 fo = layout & (1<<16);
3377 geo->raid_disks = disks;
3378 geo->near_copies = nc;
3379 geo->far_copies = fc;
3380 geo->far_offset = fo;
3381 geo->chunk_mask = chunk - 1;
3382 geo->chunk_shift = ffz(~chunk);
3386 static struct r10conf *setup_conf(struct mddev *mddev)
3388 struct r10conf *conf = NULL;
3393 copies = setup_geo(&geo, mddev, geo_new);
3396 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3397 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3398 mdname(mddev), PAGE_SIZE);
3402 if (copies < 2 || copies > mddev->raid_disks) {
3403 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3404 mdname(mddev), mddev->new_layout);
3409 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3413 /* FIXME calc properly */
3414 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3415 max(0,mddev->delta_disks)),
3420 conf->tmppage = alloc_page(GFP_KERNEL);
3425 conf->copies = copies;
3426 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3427 r10bio_pool_free, conf);
3428 if (!conf->r10bio_pool)
3431 calc_sectors(conf, mddev->dev_sectors);
3432 if (mddev->reshape_position == MaxSector) {
3433 conf->prev = conf->geo;
3434 conf->reshape_progress = MaxSector;
3436 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3440 conf->reshape_progress = mddev->reshape_position;
3441 if (conf->prev.far_offset)
3442 conf->prev.stride = 1 << conf->prev.chunk_shift;
3444 /* far_copies must be 1 */
3445 conf->prev.stride = conf->dev_sectors;
3447 spin_lock_init(&conf->device_lock);
3448 INIT_LIST_HEAD(&conf->retry_list);
3450 spin_lock_init(&conf->resync_lock);
3451 init_waitqueue_head(&conf->wait_barrier);
3453 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3457 conf->mddev = mddev;
3462 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3465 if (conf->r10bio_pool)
3466 mempool_destroy(conf->r10bio_pool);
3467 kfree(conf->mirrors);
3468 safe_put_page(conf->tmppage);
3471 return ERR_PTR(err);
3474 static int run(struct mddev *mddev)
3476 struct r10conf *conf;
3477 int i, disk_idx, chunk_size;
3478 struct raid10_info *disk;
3479 struct md_rdev *rdev;
3481 sector_t min_offset_diff = 0;
3484 if (mddev->private == NULL) {
3485 conf = setup_conf(mddev);
3487 return PTR_ERR(conf);
3488 mddev->private = conf;
3490 conf = mddev->private;
3494 mddev->thread = conf->thread;
3495 conf->thread = NULL;
3497 chunk_size = mddev->chunk_sectors << 9;
3499 blk_queue_io_min(mddev->queue, chunk_size);
3500 if (conf->geo.raid_disks % conf->geo.near_copies)
3501 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3503 blk_queue_io_opt(mddev->queue, chunk_size *
3504 (conf->geo.raid_disks / conf->geo.near_copies));
3507 rdev_for_each(rdev, mddev) {
3509 struct request_queue *q;
3511 disk_idx = rdev->raid_disk;
3514 if (disk_idx >= conf->geo.raid_disks &&
3515 disk_idx >= conf->prev.raid_disks)
3517 disk = conf->mirrors + disk_idx;
3519 if (test_bit(Replacement, &rdev->flags)) {
3520 if (disk->replacement)
3522 disk->replacement = rdev;
3528 q = bdev_get_queue(rdev->bdev);
3529 if (q->merge_bvec_fn)
3530 mddev->merge_check_needed = 1;
3531 diff = (rdev->new_data_offset - rdev->data_offset);
3532 if (!mddev->reshape_backwards)
3536 if (first || diff < min_offset_diff)
3537 min_offset_diff = diff;
3540 disk_stack_limits(mddev->gendisk, rdev->bdev,
3541 rdev->data_offset << 9);
3543 disk->head_position = 0;
3546 /* need to check that every block has at least one working mirror */
3547 if (!enough(conf, -1)) {
3548 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3553 if (conf->reshape_progress != MaxSector) {
3554 /* must ensure that shape change is supported */
3555 if (conf->geo.far_copies != 1 &&
3556 conf->geo.far_offset == 0)
3558 if (conf->prev.far_copies != 1 &&
3559 conf->geo.far_offset == 0)
3563 mddev->degraded = 0;
3565 i < conf->geo.raid_disks
3566 || i < conf->prev.raid_disks;
3569 disk = conf->mirrors + i;
3571 if (!disk->rdev && disk->replacement) {
3572 /* The replacement is all we have - use it */
3573 disk->rdev = disk->replacement;
3574 disk->replacement = NULL;
3575 clear_bit(Replacement, &disk->rdev->flags);
3579 !test_bit(In_sync, &disk->rdev->flags)) {
3580 disk->head_position = 0;
3585 disk->recovery_disabled = mddev->recovery_disabled - 1;
3588 if (mddev->recovery_cp != MaxSector)
3589 printk(KERN_NOTICE "md/raid10:%s: not clean"
3590 " -- starting background reconstruction\n",
3593 "md/raid10:%s: active with %d out of %d devices\n",
3594 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3595 conf->geo.raid_disks);
3597 * Ok, everything is just fine now
3599 mddev->dev_sectors = conf->dev_sectors;
3600 size = raid10_size(mddev, 0, 0);
3601 md_set_array_sectors(mddev, size);
3602 mddev->resync_max_sectors = size;
3605 int stripe = conf->geo.raid_disks *
3606 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3607 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3608 mddev->queue->backing_dev_info.congested_data = mddev;
3610 /* Calculate max read-ahead size.
3611 * We need to readahead at least twice a whole stripe....
3614 stripe /= conf->geo.near_copies;
3615 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3616 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3617 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3621 if (md_integrity_register(mddev))
3624 if (conf->reshape_progress != MaxSector) {
3625 unsigned long before_length, after_length;
3627 before_length = ((1 << conf->prev.chunk_shift) *
3628 conf->prev.far_copies);
3629 after_length = ((1 << conf->geo.chunk_shift) *
3630 conf->geo.far_copies);
3632 if (max(before_length, after_length) > min_offset_diff) {
3633 /* This cannot work */
3634 printk("md/raid10: offset difference not enough to continue reshape\n");
3637 conf->offset_diff = min_offset_diff;
3639 conf->reshape_safe = conf->reshape_progress;
3640 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3641 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3642 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3643 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3644 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3651 md_unregister_thread(&mddev->thread);
3652 if (conf->r10bio_pool)
3653 mempool_destroy(conf->r10bio_pool);
3654 safe_put_page(conf->tmppage);
3655 kfree(conf->mirrors);
3657 mddev->private = NULL;
3662 static int stop(struct mddev *mddev)
3664 struct r10conf *conf = mddev->private;
3666 raise_barrier(conf, 0);
3667 lower_barrier(conf);
3669 md_unregister_thread(&mddev->thread);
3671 /* the unplug fn references 'conf'*/
3672 blk_sync_queue(mddev->queue);
3674 if (conf->r10bio_pool)
3675 mempool_destroy(conf->r10bio_pool);
3676 kfree(conf->mirrors);
3678 mddev->private = NULL;
3682 static void raid10_quiesce(struct mddev *mddev, int state)
3684 struct r10conf *conf = mddev->private;
3688 raise_barrier(conf, 0);
3691 lower_barrier(conf);
3696 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3698 /* Resize of 'far' arrays is not supported.
3699 * For 'near' and 'offset' arrays we can set the
3700 * number of sectors used to be an appropriate multiple
3701 * of the chunk size.
3702 * For 'offset', this is far_copies*chunksize.
3703 * For 'near' the multiplier is the LCM of
3704 * near_copies and raid_disks.
3705 * So if far_copies > 1 && !far_offset, fail.
3706 * Else find LCM(raid_disks, near_copy)*far_copies and
3707 * multiply by chunk_size. Then round to this number.
3708 * This is mostly done by raid10_size()
3710 struct r10conf *conf = mddev->private;
3711 sector_t oldsize, size;
3713 if (mddev->reshape_position != MaxSector)
3716 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3719 oldsize = raid10_size(mddev, 0, 0);
3720 size = raid10_size(mddev, sectors, 0);
3721 if (mddev->external_size &&
3722 mddev->array_sectors > size)
3724 if (mddev->bitmap) {
3725 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3729 md_set_array_sectors(mddev, size);
3730 set_capacity(mddev->gendisk, mddev->array_sectors);
3731 revalidate_disk(mddev->gendisk);
3732 if (sectors > mddev->dev_sectors &&
3733 mddev->recovery_cp > oldsize) {
3734 mddev->recovery_cp = oldsize;
3735 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3737 calc_sectors(conf, sectors);
3738 mddev->dev_sectors = conf->dev_sectors;
3739 mddev->resync_max_sectors = size;
3743 static void *raid10_takeover_raid0(struct mddev *mddev)
3745 struct md_rdev *rdev;
3746 struct r10conf *conf;
3748 if (mddev->degraded > 0) {
3749 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3751 return ERR_PTR(-EINVAL);
3754 /* Set new parameters */
3755 mddev->new_level = 10;
3756 /* new layout: far_copies = 1, near_copies = 2 */
3757 mddev->new_layout = (1<<8) + 2;
3758 mddev->new_chunk_sectors = mddev->chunk_sectors;
3759 mddev->delta_disks = mddev->raid_disks;
3760 mddev->raid_disks *= 2;
3761 /* make sure it will be not marked as dirty */
3762 mddev->recovery_cp = MaxSector;
3764 conf = setup_conf(mddev);
3765 if (!IS_ERR(conf)) {
3766 rdev_for_each(rdev, mddev)
3767 if (rdev->raid_disk >= 0)
3768 rdev->new_raid_disk = rdev->raid_disk * 2;
3775 static void *raid10_takeover(struct mddev *mddev)
3777 struct r0conf *raid0_conf;
3779 /* raid10 can take over:
3780 * raid0 - providing it has only two drives
3782 if (mddev->level == 0) {
3783 /* for raid0 takeover only one zone is supported */
3784 raid0_conf = mddev->private;
3785 if (raid0_conf->nr_strip_zones > 1) {
3786 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3787 " with more than one zone.\n",
3789 return ERR_PTR(-EINVAL);
3791 return raid10_takeover_raid0(mddev);
3793 return ERR_PTR(-EINVAL);
3796 static int raid10_check_reshape(struct mddev *mddev)
3798 /* Called when there is a request to change
3799 * - layout (to ->new_layout)
3800 * - chunk size (to ->new_chunk_sectors)
3801 * - raid_disks (by delta_disks)
3802 * or when trying to restart a reshape that was ongoing.
3804 * We need to validate the request and possibly allocate
3805 * space if that might be an issue later.
3807 * Currently we reject any reshape of a 'far' mode array,
3808 * allow chunk size to change if new is generally acceptable,
3809 * allow raid_disks to increase, and allow
3810 * a switch between 'near' mode and 'offset' mode.
3812 struct r10conf *conf = mddev->private;
3815 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3818 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3819 /* mustn't change number of copies */
3821 if (geo.far_copies > 1 && !geo.far_offset)
3822 /* Cannot switch to 'far' mode */
3825 if (mddev->array_sectors & geo.chunk_mask)
3826 /* not factor of array size */
3829 if (!enough(conf, -1))
3832 kfree(conf->mirrors_new);
3833 conf->mirrors_new = NULL;
3834 if (mddev->delta_disks > 0) {
3835 /* allocate new 'mirrors' list */
3836 conf->mirrors_new = kzalloc(
3837 sizeof(struct raid10_info)
3838 *(mddev->raid_disks +
3839 mddev->delta_disks),
3841 if (!conf->mirrors_new)
3848 * Need to check if array has failed when deciding whether to:
3850 * - remove non-faulty devices
3853 * This determination is simple when no reshape is happening.
3854 * However if there is a reshape, we need to carefully check
3855 * both the before and after sections.
3856 * This is because some failed devices may only affect one
3857 * of the two sections, and some non-in_sync devices may
3858 * be insync in the section most affected by failed devices.
3860 static int calc_degraded(struct r10conf *conf)
3862 int degraded, degraded2;
3867 /* 'prev' section first */
3868 for (i = 0; i < conf->prev.raid_disks; i++) {
3869 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3870 if (!rdev || test_bit(Faulty, &rdev->flags))
3872 else if (!test_bit(In_sync, &rdev->flags))
3873 /* When we can reduce the number of devices in
3874 * an array, this might not contribute to
3875 * 'degraded'. It does now.
3880 if (conf->geo.raid_disks == conf->prev.raid_disks)
3884 for (i = 0; i < conf->geo.raid_disks; i++) {
3885 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3886 if (!rdev || test_bit(Faulty, &rdev->flags))
3888 else if (!test_bit(In_sync, &rdev->flags)) {
3889 /* If reshape is increasing the number of devices,
3890 * this section has already been recovered, so
3891 * it doesn't contribute to degraded.
3894 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3899 if (degraded2 > degraded)
3904 static int raid10_start_reshape(struct mddev *mddev)
3906 /* A 'reshape' has been requested. This commits
3907 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3908 * This also checks if there are enough spares and adds them
3910 * We currently require enough spares to make the final
3911 * array non-degraded. We also require that the difference
3912 * between old and new data_offset - on each device - is
3913 * enough that we never risk over-writing.
3916 unsigned long before_length, after_length;
3917 sector_t min_offset_diff = 0;
3920 struct r10conf *conf = mddev->private;
3921 struct md_rdev *rdev;
3925 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3928 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3931 before_length = ((1 << conf->prev.chunk_shift) *
3932 conf->prev.far_copies);
3933 after_length = ((1 << conf->geo.chunk_shift) *
3934 conf->geo.far_copies);
3936 rdev_for_each(rdev, mddev) {
3937 if (!test_bit(In_sync, &rdev->flags)
3938 && !test_bit(Faulty, &rdev->flags))
3940 if (rdev->raid_disk >= 0) {
3941 long long diff = (rdev->new_data_offset
3942 - rdev->data_offset);
3943 if (!mddev->reshape_backwards)
3947 if (first || diff < min_offset_diff)
3948 min_offset_diff = diff;
3952 if (max(before_length, after_length) > min_offset_diff)
3955 if (spares < mddev->delta_disks)
3958 conf->offset_diff = min_offset_diff;
3959 spin_lock_irq(&conf->device_lock);
3960 if (conf->mirrors_new) {
3961 memcpy(conf->mirrors_new, conf->mirrors,
3962 sizeof(struct raid10_info)*conf->prev.raid_disks);
3964 kfree(conf->mirrors_old); /* FIXME and elsewhere */
3965 conf->mirrors_old = conf->mirrors;
3966 conf->mirrors = conf->mirrors_new;
3967 conf->mirrors_new = NULL;
3969 setup_geo(&conf->geo, mddev, geo_start);
3971 if (mddev->reshape_backwards) {
3972 sector_t size = raid10_size(mddev, 0, 0);
3973 if (size < mddev->array_sectors) {
3974 spin_unlock_irq(&conf->device_lock);
3975 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3979 mddev->resync_max_sectors = size;
3980 conf->reshape_progress = size;
3982 conf->reshape_progress = 0;
3983 spin_unlock_irq(&conf->device_lock);
3985 if (mddev->delta_disks && mddev->bitmap) {
3986 ret = bitmap_resize(mddev->bitmap,
3987 raid10_size(mddev, 0,
3988 conf->geo.raid_disks),
3993 if (mddev->delta_disks > 0) {
3994 rdev_for_each(rdev, mddev)
3995 if (rdev->raid_disk < 0 &&
3996 !test_bit(Faulty, &rdev->flags)) {
3997 if (raid10_add_disk(mddev, rdev) == 0) {
3998 if (rdev->raid_disk >=
3999 conf->prev.raid_disks)
4000 set_bit(In_sync, &rdev->flags);
4002 rdev->recovery_offset = 0;
4004 if (sysfs_link_rdev(mddev, rdev))
4005 /* Failure here is OK */;
4007 } else if (rdev->raid_disk >= conf->prev.raid_disks
4008 && !test_bit(Faulty, &rdev->flags)) {
4009 /* This is a spare that was manually added */
4010 set_bit(In_sync, &rdev->flags);
4013 /* When a reshape changes the number of devices,
4014 * ->degraded is measured against the larger of the
4015 * pre and post numbers.
4017 spin_lock_irq(&conf->device_lock);
4018 mddev->degraded = calc_degraded(conf);
4019 spin_unlock_irq(&conf->device_lock);
4020 mddev->raid_disks = conf->geo.raid_disks;
4021 mddev->reshape_position = conf->reshape_progress;
4022 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4024 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4025 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4026 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4027 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4029 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4031 if (!mddev->sync_thread) {
4035 conf->reshape_checkpoint = jiffies;
4036 md_wakeup_thread(mddev->sync_thread);
4037 md_new_event(mddev);
4041 mddev->recovery = 0;
4042 spin_lock_irq(&conf->device_lock);
4043 conf->geo = conf->prev;
4044 mddev->raid_disks = conf->geo.raid_disks;
4045 rdev_for_each(rdev, mddev)
4046 rdev->new_data_offset = rdev->data_offset;
4048 conf->reshape_progress = MaxSector;
4049 mddev->reshape_position = MaxSector;
4050 spin_unlock_irq(&conf->device_lock);
4054 /* Calculate the last device-address that could contain
4055 * any block from the chunk that includes the array-address 's'
4056 * and report the next address.
4057 * i.e. the address returned will be chunk-aligned and after
4058 * any data that is in the chunk containing 's'.
4060 static sector_t last_dev_address(sector_t s, struct geom *geo)
4062 s = (s | geo->chunk_mask) + 1;
4063 s >>= geo->chunk_shift;
4064 s *= geo->near_copies;
4065 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4066 s *= geo->far_copies;
4067 s <<= geo->chunk_shift;
4071 /* Calculate the first device-address that could contain
4072 * any block from the chunk that includes the array-address 's'.
4073 * This too will be the start of a chunk
4075 static sector_t first_dev_address(sector_t s, struct geom *geo)
4077 s >>= geo->chunk_shift;
4078 s *= geo->near_copies;
4079 sector_div(s, geo->raid_disks);
4080 s *= geo->far_copies;
4081 s <<= geo->chunk_shift;
4085 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4088 /* We simply copy at most one chunk (smallest of old and new)
4089 * at a time, possibly less if that exceeds RESYNC_PAGES,
4090 * or we hit a bad block or something.
4091 * This might mean we pause for normal IO in the middle of
4092 * a chunk, but that is not a problem was mddev->reshape_position
4093 * can record any location.
4095 * If we will want to write to a location that isn't
4096 * yet recorded as 'safe' (i.e. in metadata on disk) then
4097 * we need to flush all reshape requests and update the metadata.
4099 * When reshaping forwards (e.g. to more devices), we interpret
4100 * 'safe' as the earliest block which might not have been copied
4101 * down yet. We divide this by previous stripe size and multiply
4102 * by previous stripe length to get lowest device offset that we
4103 * cannot write to yet.
4104 * We interpret 'sector_nr' as an address that we want to write to.
4105 * From this we use last_device_address() to find where we might
4106 * write to, and first_device_address on the 'safe' position.
4107 * If this 'next' write position is after the 'safe' position,
4108 * we must update the metadata to increase the 'safe' position.
4110 * When reshaping backwards, we round in the opposite direction
4111 * and perform the reverse test: next write position must not be
4112 * less than current safe position.
4114 * In all this the minimum difference in data offsets
4115 * (conf->offset_diff - always positive) allows a bit of slack,
4116 * so next can be after 'safe', but not by more than offset_disk
4118 * We need to prepare all the bios here before we start any IO
4119 * to ensure the size we choose is acceptable to all devices.
4120 * The means one for each copy for write-out and an extra one for
4122 * We store the read-in bio in ->master_bio and the others in
4123 * ->devs[x].bio and ->devs[x].repl_bio.
4125 struct r10conf *conf = mddev->private;
4126 struct r10bio *r10_bio;
4127 sector_t next, safe, last;
4131 struct md_rdev *rdev;
4134 struct bio *bio, *read_bio;
4135 int sectors_done = 0;
4137 if (sector_nr == 0) {
4138 /* If restarting in the middle, skip the initial sectors */
4139 if (mddev->reshape_backwards &&
4140 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4141 sector_nr = (raid10_size(mddev, 0, 0)
4142 - conf->reshape_progress);
4143 } else if (!mddev->reshape_backwards &&
4144 conf->reshape_progress > 0)
4145 sector_nr = conf->reshape_progress;
4147 mddev->curr_resync_completed = sector_nr;
4148 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4154 /* We don't use sector_nr to track where we are up to
4155 * as that doesn't work well for ->reshape_backwards.
4156 * So just use ->reshape_progress.
4158 if (mddev->reshape_backwards) {
4159 /* 'next' is the earliest device address that we might
4160 * write to for this chunk in the new layout
4162 next = first_dev_address(conf->reshape_progress - 1,
4165 /* 'safe' is the last device address that we might read from
4166 * in the old layout after a restart
4168 safe = last_dev_address(conf->reshape_safe - 1,
4171 if (next + conf->offset_diff < safe)
4174 last = conf->reshape_progress - 1;
4175 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4176 & conf->prev.chunk_mask);
4177 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4178 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4180 /* 'next' is after the last device address that we
4181 * might write to for this chunk in the new layout
4183 next = last_dev_address(conf->reshape_progress, &conf->geo);
4185 /* 'safe' is the earliest device address that we might
4186 * read from in the old layout after a restart
4188 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4190 /* Need to update metadata if 'next' might be beyond 'safe'
4191 * as that would possibly corrupt data
4193 if (next > safe + conf->offset_diff)
4196 sector_nr = conf->reshape_progress;
4197 last = sector_nr | (conf->geo.chunk_mask
4198 & conf->prev.chunk_mask);
4200 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4201 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4205 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4206 /* Need to update reshape_position in metadata */
4208 mddev->reshape_position = conf->reshape_progress;
4209 if (mddev->reshape_backwards)
4210 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4211 - conf->reshape_progress;
4213 mddev->curr_resync_completed = conf->reshape_progress;
4214 conf->reshape_checkpoint = jiffies;
4215 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4216 md_wakeup_thread(mddev->thread);
4217 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4218 kthread_should_stop());
4219 conf->reshape_safe = mddev->reshape_position;
4220 allow_barrier(conf);
4224 /* Now schedule reads for blocks from sector_nr to last */
4225 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4226 raise_barrier(conf, sectors_done != 0);
4227 atomic_set(&r10_bio->remaining, 0);
4228 r10_bio->mddev = mddev;
4229 r10_bio->sector = sector_nr;
4230 set_bit(R10BIO_IsReshape, &r10_bio->state);
4231 r10_bio->sectors = last - sector_nr + 1;
4232 rdev = read_balance(conf, r10_bio, &max_sectors);
4233 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4236 /* Cannot read from here, so need to record bad blocks
4237 * on all the target devices.
4240 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4241 return sectors_done;
4244 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4246 read_bio->bi_bdev = rdev->bdev;
4247 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4248 + rdev->data_offset);
4249 read_bio->bi_private = r10_bio;
4250 read_bio->bi_end_io = end_sync_read;
4251 read_bio->bi_rw = READ;
4252 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4253 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4254 read_bio->bi_vcnt = 0;
4255 read_bio->bi_idx = 0;
4256 read_bio->bi_size = 0;
4257 r10_bio->master_bio = read_bio;
4258 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4260 /* Now find the locations in the new layout */
4261 __raid10_find_phys(&conf->geo, r10_bio);
4264 read_bio->bi_next = NULL;
4266 for (s = 0; s < conf->copies*2; s++) {
4268 int d = r10_bio->devs[s/2].devnum;
4269 struct md_rdev *rdev2;
4271 rdev2 = conf->mirrors[d].replacement;
4272 b = r10_bio->devs[s/2].repl_bio;
4274 rdev2 = conf->mirrors[d].rdev;
4275 b = r10_bio->devs[s/2].bio;
4277 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4279 b->bi_bdev = rdev2->bdev;
4280 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4281 b->bi_private = r10_bio;
4282 b->bi_end_io = end_reshape_write;
4284 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4285 b->bi_flags |= 1 << BIO_UPTODATE;
4293 /* Now add as many pages as possible to all of these bios. */
4296 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4297 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4298 int len = (max_sectors - s) << 9;
4299 if (len > PAGE_SIZE)
4301 for (bio = blist; bio ; bio = bio->bi_next) {
4303 if (bio_add_page(bio, page, len, 0))
4306 /* Didn't fit, must stop */
4308 bio2 && bio2 != bio;
4309 bio2 = bio2->bi_next) {
4310 /* Remove last page from this bio */
4312 bio2->bi_size -= len;
4313 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4317 sector_nr += len >> 9;
4318 nr_sectors += len >> 9;
4321 r10_bio->sectors = nr_sectors;
4323 /* Now submit the read */
4324 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4325 atomic_inc(&r10_bio->remaining);
4326 read_bio->bi_next = NULL;
4327 generic_make_request(read_bio);
4328 sector_nr += nr_sectors;
4329 sectors_done += nr_sectors;
4330 if (sector_nr <= last)
4333 /* Now that we have done the whole section we can
4334 * update reshape_progress
4336 if (mddev->reshape_backwards)
4337 conf->reshape_progress -= sectors_done;
4339 conf->reshape_progress += sectors_done;
4341 return sectors_done;
4344 static void end_reshape_request(struct r10bio *r10_bio);
4345 static int handle_reshape_read_error(struct mddev *mddev,
4346 struct r10bio *r10_bio);
4347 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4349 /* Reshape read completed. Hopefully we have a block
4351 * If we got a read error then we do sync 1-page reads from
4352 * elsewhere until we find the data - or give up.
4354 struct r10conf *conf = mddev->private;
4357 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4358 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4359 /* Reshape has been aborted */
4360 md_done_sync(mddev, r10_bio->sectors, 0);
4364 /* We definitely have the data in the pages, schedule the
4367 atomic_set(&r10_bio->remaining, 1);
4368 for (s = 0; s < conf->copies*2; s++) {
4370 int d = r10_bio->devs[s/2].devnum;
4371 struct md_rdev *rdev;
4373 rdev = conf->mirrors[d].replacement;
4374 b = r10_bio->devs[s/2].repl_bio;
4376 rdev = conf->mirrors[d].rdev;
4377 b = r10_bio->devs[s/2].bio;
4379 if (!rdev || test_bit(Faulty, &rdev->flags))
4381 atomic_inc(&rdev->nr_pending);
4382 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4383 atomic_inc(&r10_bio->remaining);
4385 generic_make_request(b);
4387 end_reshape_request(r10_bio);
4390 static void end_reshape(struct r10conf *conf)
4392 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4395 spin_lock_irq(&conf->device_lock);
4396 conf->prev = conf->geo;
4397 md_finish_reshape(conf->mddev);
4399 conf->reshape_progress = MaxSector;
4400 spin_unlock_irq(&conf->device_lock);
4402 /* read-ahead size must cover two whole stripes, which is
4403 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4405 if (conf->mddev->queue) {
4406 int stripe = conf->geo.raid_disks *
4407 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4408 stripe /= conf->geo.near_copies;
4409 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4410 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4416 static int handle_reshape_read_error(struct mddev *mddev,
4417 struct r10bio *r10_bio)
4419 /* Use sync reads to get the blocks from somewhere else */
4420 int sectors = r10_bio->sectors;
4421 struct r10conf *conf = mddev->private;
4423 struct r10bio r10_bio;
4424 struct r10dev devs[conf->copies];
4426 struct r10bio *r10b = &on_stack.r10_bio;
4429 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4431 r10b->sector = r10_bio->sector;
4432 __raid10_find_phys(&conf->prev, r10b);
4437 int first_slot = slot;
4439 if (s > (PAGE_SIZE >> 9))
4443 int d = r10b->devs[slot].devnum;
4444 struct md_rdev *rdev = conf->mirrors[d].rdev;
4447 test_bit(Faulty, &rdev->flags) ||
4448 !test_bit(In_sync, &rdev->flags))
4451 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4452 success = sync_page_io(rdev,
4461 if (slot >= conf->copies)
4463 if (slot == first_slot)
4467 /* couldn't read this block, must give up */
4468 set_bit(MD_RECOVERY_INTR,
4478 static void end_reshape_write(struct bio *bio, int error)
4480 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4481 struct r10bio *r10_bio = bio->bi_private;
4482 struct mddev *mddev = r10_bio->mddev;
4483 struct r10conf *conf = mddev->private;
4487 struct md_rdev *rdev = NULL;
4489 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4491 rdev = conf->mirrors[d].replacement;
4494 rdev = conf->mirrors[d].rdev;
4498 /* FIXME should record badblock */
4499 md_error(mddev, rdev);
4502 rdev_dec_pending(rdev, mddev);
4503 end_reshape_request(r10_bio);
4506 static void end_reshape_request(struct r10bio *r10_bio)
4508 if (!atomic_dec_and_test(&r10_bio->remaining))
4510 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4511 bio_put(r10_bio->master_bio);
4515 static void raid10_finish_reshape(struct mddev *mddev)
4517 struct r10conf *conf = mddev->private;
4519 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4522 if (mddev->delta_disks > 0) {
4523 sector_t size = raid10_size(mddev, 0, 0);
4524 md_set_array_sectors(mddev, size);
4525 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4526 mddev->recovery_cp = mddev->resync_max_sectors;
4527 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4529 mddev->resync_max_sectors = size;
4530 set_capacity(mddev->gendisk, mddev->array_sectors);
4531 revalidate_disk(mddev->gendisk);
4534 for (d = conf->geo.raid_disks ;
4535 d < conf->geo.raid_disks - mddev->delta_disks;
4537 struct md_rdev *rdev = conf->mirrors[d].rdev;
4539 clear_bit(In_sync, &rdev->flags);
4540 rdev = conf->mirrors[d].replacement;
4542 clear_bit(In_sync, &rdev->flags);
4545 mddev->layout = mddev->new_layout;
4546 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4547 mddev->reshape_position = MaxSector;
4548 mddev->delta_disks = 0;
4549 mddev->reshape_backwards = 0;
4552 static struct md_personality raid10_personality =
4556 .owner = THIS_MODULE,
4557 .make_request = make_request,
4561 .error_handler = error,
4562 .hot_add_disk = raid10_add_disk,
4563 .hot_remove_disk= raid10_remove_disk,
4564 .spare_active = raid10_spare_active,
4565 .sync_request = sync_request,
4566 .quiesce = raid10_quiesce,
4567 .size = raid10_size,
4568 .resize = raid10_resize,
4569 .takeover = raid10_takeover,
4570 .check_reshape = raid10_check_reshape,
4571 .start_reshape = raid10_start_reshape,
4572 .finish_reshape = raid10_finish_reshape,
4575 static int __init raid_init(void)
4577 return register_md_personality(&raid10_personality);
4580 static void raid_exit(void)
4582 unregister_md_personality(&raid10_personality);
4585 module_init(raid_init);
4586 module_exit(raid_exit);
4587 MODULE_LICENSE("GPL");
4588 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4589 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4590 MODULE_ALIAS("md-raid10");
4591 MODULE_ALIAS("md-level-10");
4593 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);