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>
33 * RAID10 provides a combination of RAID0 and RAID1 functionality.
34 * The layout of data is defined by
37 * near_copies (stored in low byte of layout)
38 * far_copies (stored in second byte of layout)
39 * far_offset (stored in bit 16 of layout )
41 * The data to be stored is divided into chunks using chunksize.
42 * Each device is divided into far_copies sections.
43 * In each section, chunks are laid out in a style similar to raid0, but
44 * near_copies copies of each chunk is stored (each on a different drive).
45 * The starting device for each section is offset near_copies from the starting
46 * device of the previous section.
47 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * near_copies and far_copies must be at least one, and their product is at most
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of be very far apart
54 * on disk, there are adjacent stripes.
58 * Number of guaranteed r10bios in case of extreme VM load:
60 #define NR_RAID10_BIOS 256
62 /* When there are this many requests queue to be written by
63 * the raid10 thread, we become 'congested' to provide back-pressure
66 static int max_queued_requests = 1024;
68 static void allow_barrier(struct r10conf *conf);
69 static void lower_barrier(struct r10conf *conf);
71 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
73 struct r10conf *conf = data;
74 int size = offsetof(struct r10bio, devs[conf->copies]);
76 /* allocate a r10bio with room for raid_disks entries in the
78 return kzalloc(size, gfp_flags);
81 static void r10bio_pool_free(void *r10_bio, void *data)
86 /* Maximum size of each resync request */
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
89 /* amount of memory to reserve for resync requests */
90 #define RESYNC_WINDOW (1024*1024)
91 /* maximum number of concurrent requests, memory permitting */
92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
95 * When performing a resync, we need to read and compare, so
96 * we need as many pages are there are copies.
97 * When performing a recovery, we need 2 bios, one for read,
98 * one for write (we recover only one drive per r10buf)
101 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
103 struct r10conf *conf = data;
105 struct r10bio *r10_bio;
110 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
114 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
115 nalloc = conf->copies; /* resync */
117 nalloc = 2; /* recovery */
122 for (j = nalloc ; j-- ; ) {
123 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
126 r10_bio->devs[j].bio = bio;
127 if (!conf->have_replacement)
129 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
132 r10_bio->devs[j].repl_bio = bio;
135 * Allocate RESYNC_PAGES data pages and attach them
138 for (j = 0 ; j < nalloc; j++) {
139 struct bio *rbio = r10_bio->devs[j].repl_bio;
140 bio = r10_bio->devs[j].bio;
141 for (i = 0; i < RESYNC_PAGES; i++) {
142 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
143 &conf->mddev->recovery)) {
144 /* we can share bv_page's during recovery */
145 struct bio *rbio = r10_bio->devs[0].bio;
146 page = rbio->bi_io_vec[i].bv_page;
149 page = alloc_page(gfp_flags);
153 bio->bi_io_vec[i].bv_page = page;
155 rbio->bi_io_vec[i].bv_page = page;
163 safe_put_page(bio->bi_io_vec[i-1].bv_page);
165 for (i = 0; i < RESYNC_PAGES ; i++)
166 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
169 while (++j < nalloc) {
170 bio_put(r10_bio->devs[j].bio);
171 if (r10_bio->devs[j].repl_bio)
172 bio_put(r10_bio->devs[j].repl_bio);
174 r10bio_pool_free(r10_bio, conf);
178 static void r10buf_pool_free(void *__r10_bio, void *data)
181 struct r10conf *conf = data;
182 struct r10bio *r10bio = __r10_bio;
185 for (j=0; j < conf->copies; j++) {
186 struct bio *bio = r10bio->devs[j].bio;
188 for (i = 0; i < RESYNC_PAGES; i++) {
189 safe_put_page(bio->bi_io_vec[i].bv_page);
190 bio->bi_io_vec[i].bv_page = NULL;
194 bio = r10bio->devs[j].repl_bio;
198 r10bio_pool_free(r10bio, conf);
201 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
205 for (i = 0; i < conf->copies; i++) {
206 struct bio **bio = & r10_bio->devs[i].bio;
207 if (!BIO_SPECIAL(*bio))
210 bio = &r10_bio->devs[i].repl_bio;
211 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
217 static void free_r10bio(struct r10bio *r10_bio)
219 struct r10conf *conf = r10_bio->mddev->private;
221 put_all_bios(conf, r10_bio);
222 mempool_free(r10_bio, conf->r10bio_pool);
225 static void put_buf(struct r10bio *r10_bio)
227 struct r10conf *conf = r10_bio->mddev->private;
229 mempool_free(r10_bio, conf->r10buf_pool);
234 static void reschedule_retry(struct r10bio *r10_bio)
237 struct mddev *mddev = r10_bio->mddev;
238 struct r10conf *conf = mddev->private;
240 spin_lock_irqsave(&conf->device_lock, flags);
241 list_add(&r10_bio->retry_list, &conf->retry_list);
243 spin_unlock_irqrestore(&conf->device_lock, flags);
245 /* wake up frozen array... */
246 wake_up(&conf->wait_barrier);
248 md_wakeup_thread(mddev->thread);
252 * raid_end_bio_io() is called when we have finished servicing a mirrored
253 * operation and are ready to return a success/failure code to the buffer
256 static void raid_end_bio_io(struct r10bio *r10_bio)
258 struct bio *bio = r10_bio->master_bio;
260 struct r10conf *conf = r10_bio->mddev->private;
262 if (bio->bi_phys_segments) {
264 spin_lock_irqsave(&conf->device_lock, flags);
265 bio->bi_phys_segments--;
266 done = (bio->bi_phys_segments == 0);
267 spin_unlock_irqrestore(&conf->device_lock, flags);
270 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
271 clear_bit(BIO_UPTODATE, &bio->bi_flags);
275 * Wake up any possible resync thread that waits for the device
280 free_r10bio(r10_bio);
284 * Update disk head position estimator based on IRQ completion info.
286 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
288 struct r10conf *conf = r10_bio->mddev->private;
290 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
291 r10_bio->devs[slot].addr + (r10_bio->sectors);
295 * Find the disk number which triggered given bio
297 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
298 struct bio *bio, int *slotp, int *replp)
303 for (slot = 0; slot < conf->copies; slot++) {
304 if (r10_bio->devs[slot].bio == bio)
306 if (r10_bio->devs[slot].repl_bio == bio) {
312 BUG_ON(slot == conf->copies);
313 update_head_pos(slot, r10_bio);
319 return r10_bio->devs[slot].devnum;
322 static void raid10_end_read_request(struct bio *bio, int error)
324 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
325 struct r10bio *r10_bio = bio->bi_private;
327 struct md_rdev *rdev;
328 struct r10conf *conf = r10_bio->mddev->private;
331 slot = r10_bio->read_slot;
332 dev = r10_bio->devs[slot].devnum;
333 rdev = r10_bio->devs[slot].rdev;
335 * this branch is our 'one mirror IO has finished' event handler:
337 update_head_pos(slot, r10_bio);
341 * Set R10BIO_Uptodate in our master bio, so that
342 * we will return a good error code to the higher
343 * levels even if IO on some other mirrored buffer fails.
345 * The 'master' represents the composite IO operation to
346 * user-side. So if something waits for IO, then it will
347 * wait for the 'master' bio.
349 set_bit(R10BIO_Uptodate, &r10_bio->state);
350 raid_end_bio_io(r10_bio);
351 rdev_dec_pending(rdev, conf->mddev);
354 * oops, read error - keep the refcount on the rdev
356 char b[BDEVNAME_SIZE];
357 printk_ratelimited(KERN_ERR
358 "md/raid10:%s: %s: rescheduling sector %llu\n",
360 bdevname(rdev->bdev, b),
361 (unsigned long long)r10_bio->sector);
362 set_bit(R10BIO_ReadError, &r10_bio->state);
363 reschedule_retry(r10_bio);
367 static void close_write(struct r10bio *r10_bio)
369 /* clear the bitmap if all writes complete successfully */
370 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
372 !test_bit(R10BIO_Degraded, &r10_bio->state),
374 md_write_end(r10_bio->mddev);
377 static void one_write_done(struct r10bio *r10_bio)
379 if (atomic_dec_and_test(&r10_bio->remaining)) {
380 if (test_bit(R10BIO_WriteError, &r10_bio->state))
381 reschedule_retry(r10_bio);
383 close_write(r10_bio);
384 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
385 reschedule_retry(r10_bio);
387 raid_end_bio_io(r10_bio);
392 static void raid10_end_write_request(struct bio *bio, int error)
394 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
395 struct r10bio *r10_bio = bio->bi_private;
398 struct r10conf *conf = r10_bio->mddev->private;
400 struct md_rdev *rdev;
402 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
405 rdev = conf->mirrors[dev].replacement;
407 rdev = conf->mirrors[dev].rdev;
409 * this branch is our 'one mirror IO has finished' event handler:
413 /* Never record new bad blocks to replacement,
416 md_error(rdev->mddev, rdev);
418 set_bit(WriteErrorSeen, &rdev->flags);
419 set_bit(R10BIO_WriteError, &r10_bio->state);
424 * Set R10BIO_Uptodate in our master bio, so that
425 * we will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer fails.
428 * The 'master' represents the composite IO operation to
429 * user-side. So if something waits for IO, then it will
430 * wait for the 'master' bio.
435 set_bit(R10BIO_Uptodate, &r10_bio->state);
437 /* Maybe we can clear some bad blocks. */
438 if (is_badblock(rdev,
439 r10_bio->devs[slot].addr,
441 &first_bad, &bad_sectors)) {
444 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
446 r10_bio->devs[slot].bio = IO_MADE_GOOD;
448 set_bit(R10BIO_MadeGood, &r10_bio->state);
454 * Let's see if all mirrored write operations have finished
457 one_write_done(r10_bio);
459 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
463 * RAID10 layout manager
464 * As well as the chunksize and raid_disks count, there are two
465 * parameters: near_copies and far_copies.
466 * near_copies * far_copies must be <= raid_disks.
467 * Normally one of these will be 1.
468 * If both are 1, we get raid0.
469 * If near_copies == raid_disks, we get raid1.
471 * Chunks are laid out in raid0 style with near_copies copies of the
472 * first chunk, followed by near_copies copies of the next chunk and
474 * If far_copies > 1, then after 1/far_copies of the array has been assigned
475 * as described above, we start again with a device offset of near_copies.
476 * So we effectively have another copy of the whole array further down all
477 * the drives, but with blocks on different drives.
478 * With this layout, and block is never stored twice on the one device.
480 * raid10_find_phys finds the sector offset of a given virtual sector
481 * on each device that it is on.
483 * raid10_find_virt does the reverse mapping, from a device and a
484 * sector offset to a virtual address
487 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
497 /* now calculate first sector/dev */
498 chunk = r10bio->sector >> conf->chunk_shift;
499 sector = r10bio->sector & conf->chunk_mask;
501 chunk *= conf->near_copies;
503 dev = sector_div(stripe, conf->raid_disks);
504 if (conf->far_offset)
505 stripe *= conf->far_copies;
507 sector += stripe << conf->chunk_shift;
509 /* and calculate all the others */
510 for (n=0; n < conf->near_copies; n++) {
513 r10bio->devs[slot].addr = sector;
514 r10bio->devs[slot].devnum = d;
517 for (f = 1; f < conf->far_copies; f++) {
518 d += conf->near_copies;
519 if (d >= conf->raid_disks)
520 d -= conf->raid_disks;
522 r10bio->devs[slot].devnum = d;
523 r10bio->devs[slot].addr = s;
527 if (dev >= conf->raid_disks) {
529 sector += (conf->chunk_mask + 1);
532 BUG_ON(slot != conf->copies);
535 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
537 sector_t offset, chunk, vchunk;
539 offset = sector & conf->chunk_mask;
540 if (conf->far_offset) {
542 chunk = sector >> conf->chunk_shift;
543 fc = sector_div(chunk, conf->far_copies);
544 dev -= fc * conf->near_copies;
546 dev += conf->raid_disks;
548 while (sector >= conf->stride) {
549 sector -= conf->stride;
550 if (dev < conf->near_copies)
551 dev += conf->raid_disks - conf->near_copies;
553 dev -= conf->near_copies;
555 chunk = sector >> conf->chunk_shift;
557 vchunk = chunk * conf->raid_disks + dev;
558 sector_div(vchunk, conf->near_copies);
559 return (vchunk << conf->chunk_shift) + offset;
563 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
565 * @bvm: properties of new bio
566 * @biovec: the request that could be merged to it.
568 * Return amount of bytes we can accept at this offset
569 * If near_copies == raid_disk, there are no striping issues,
570 * but in that case, the function isn't called at all.
572 static int raid10_mergeable_bvec(struct request_queue *q,
573 struct bvec_merge_data *bvm,
574 struct bio_vec *biovec)
576 struct mddev *mddev = q->queuedata;
577 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
579 unsigned int chunk_sectors = mddev->chunk_sectors;
580 unsigned int bio_sectors = bvm->bi_size >> 9;
582 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
583 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
584 if (max <= biovec->bv_len && bio_sectors == 0)
585 return biovec->bv_len;
591 * This routine returns the disk from which the requested read should
592 * be done. There is a per-array 'next expected sequential IO' sector
593 * number - if this matches on the next IO then we use the last disk.
594 * There is also a per-disk 'last know head position' sector that is
595 * maintained from IRQ contexts, both the normal and the resync IO
596 * completion handlers update this position correctly. If there is no
597 * perfect sequential match then we pick the disk whose head is closest.
599 * If there are 2 mirrors in the same 2 devices, performance degrades
600 * because position is mirror, not device based.
602 * The rdev for the device selected will have nr_pending incremented.
606 * FIXME: possibly should rethink readbalancing and do it differently
607 * depending on near_copies / far_copies geometry.
609 static struct md_rdev *read_balance(struct r10conf *conf,
610 struct r10bio *r10_bio,
613 const sector_t this_sector = r10_bio->sector;
615 int sectors = r10_bio->sectors;
616 int best_good_sectors;
617 sector_t new_distance, best_dist;
618 struct md_rdev *rdev, *best_rdev;
622 raid10_find_phys(conf, r10_bio);
625 sectors = r10_bio->sectors;
628 best_dist = MaxSector;
629 best_good_sectors = 0;
632 * Check if we can balance. We can balance on the whole
633 * device if no resync is going on (recovery is ok), or below
634 * the resync window. We take the first readable disk when
635 * above the resync window.
637 if (conf->mddev->recovery_cp < MaxSector
638 && (this_sector + sectors >= conf->next_resync))
641 for (slot = 0; slot < conf->copies ; slot++) {
646 if (r10_bio->devs[slot].bio == IO_BLOCKED)
648 disk = r10_bio->devs[slot].devnum;
649 rdev = rcu_dereference(conf->mirrors[disk].replacement);
650 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
651 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
652 rdev = rcu_dereference(conf->mirrors[disk].rdev);
655 if (test_bit(Faulty, &rdev->flags))
657 if (!test_bit(In_sync, &rdev->flags) &&
658 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
661 dev_sector = r10_bio->devs[slot].addr;
662 if (is_badblock(rdev, dev_sector, sectors,
663 &first_bad, &bad_sectors)) {
664 if (best_dist < MaxSector)
665 /* Already have a better slot */
667 if (first_bad <= dev_sector) {
668 /* Cannot read here. If this is the
669 * 'primary' device, then we must not read
670 * beyond 'bad_sectors' from another device.
672 bad_sectors -= (dev_sector - first_bad);
673 if (!do_balance && sectors > bad_sectors)
674 sectors = bad_sectors;
675 if (best_good_sectors > sectors)
676 best_good_sectors = sectors;
678 sector_t good_sectors =
679 first_bad - dev_sector;
680 if (good_sectors > best_good_sectors) {
681 best_good_sectors = good_sectors;
686 /* Must read from here */
691 best_good_sectors = sectors;
696 /* This optimisation is debatable, and completely destroys
697 * sequential read speed for 'far copies' arrays. So only
698 * keep it for 'near' arrays, and review those later.
700 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
703 /* for far > 1 always use the lowest address */
704 if (conf->far_copies > 1)
705 new_distance = r10_bio->devs[slot].addr;
707 new_distance = abs(r10_bio->devs[slot].addr -
708 conf->mirrors[disk].head_position);
709 if (new_distance < best_dist) {
710 best_dist = new_distance;
715 if (slot >= conf->copies) {
721 atomic_inc(&rdev->nr_pending);
722 if (test_bit(Faulty, &rdev->flags)) {
723 /* Cannot risk returning a device that failed
724 * before we inc'ed nr_pending
726 rdev_dec_pending(rdev, conf->mddev);
729 r10_bio->read_slot = slot;
733 *max_sectors = best_good_sectors;
738 static int raid10_congested(void *data, int bits)
740 struct mddev *mddev = data;
741 struct r10conf *conf = mddev->private;
744 if ((bits & (1 << BDI_async_congested)) &&
745 conf->pending_count >= max_queued_requests)
748 if (mddev_congested(mddev, bits))
751 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
752 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
753 if (rdev && !test_bit(Faulty, &rdev->flags)) {
754 struct request_queue *q = bdev_get_queue(rdev->bdev);
756 ret |= bdi_congested(&q->backing_dev_info, bits);
763 static void flush_pending_writes(struct r10conf *conf)
765 /* Any writes that have been queued but are awaiting
766 * bitmap updates get flushed here.
768 spin_lock_irq(&conf->device_lock);
770 if (conf->pending_bio_list.head) {
772 bio = bio_list_get(&conf->pending_bio_list);
773 conf->pending_count = 0;
774 spin_unlock_irq(&conf->device_lock);
775 /* flush any pending bitmap writes to disk
776 * before proceeding w/ I/O */
777 bitmap_unplug(conf->mddev->bitmap);
778 wake_up(&conf->wait_barrier);
780 while (bio) { /* submit pending writes */
781 struct bio *next = bio->bi_next;
783 generic_make_request(bio);
787 spin_unlock_irq(&conf->device_lock);
791 * Sometimes we need to suspend IO while we do something else,
792 * either some resync/recovery, or reconfigure the array.
793 * To do this we raise a 'barrier'.
794 * The 'barrier' is a counter that can be raised multiple times
795 * to count how many activities are happening which preclude
797 * We can only raise the barrier if there is no pending IO.
798 * i.e. if nr_pending == 0.
799 * We choose only to raise the barrier if no-one is waiting for the
800 * barrier to go down. This means that as soon as an IO request
801 * is ready, no other operations which require a barrier will start
802 * until the IO request has had a chance.
804 * So: regular IO calls 'wait_barrier'. When that returns there
805 * is no backgroup IO happening, It must arrange to call
806 * allow_barrier when it has finished its IO.
807 * backgroup IO calls must call raise_barrier. Once that returns
808 * there is no normal IO happeing. It must arrange to call
809 * lower_barrier when the particular background IO completes.
812 static void raise_barrier(struct r10conf *conf, int force)
814 BUG_ON(force && !conf->barrier);
815 spin_lock_irq(&conf->resync_lock);
817 /* Wait until no block IO is waiting (unless 'force') */
818 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
819 conf->resync_lock, );
821 /* block any new IO from starting */
824 /* Now wait for all pending IO to complete */
825 wait_event_lock_irq(conf->wait_barrier,
826 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
827 conf->resync_lock, );
829 spin_unlock_irq(&conf->resync_lock);
832 static void lower_barrier(struct r10conf *conf)
835 spin_lock_irqsave(&conf->resync_lock, flags);
837 spin_unlock_irqrestore(&conf->resync_lock, flags);
838 wake_up(&conf->wait_barrier);
841 static void wait_barrier(struct r10conf *conf)
843 spin_lock_irq(&conf->resync_lock);
846 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
852 spin_unlock_irq(&conf->resync_lock);
855 static void allow_barrier(struct r10conf *conf)
858 spin_lock_irqsave(&conf->resync_lock, flags);
860 spin_unlock_irqrestore(&conf->resync_lock, flags);
861 wake_up(&conf->wait_barrier);
864 static void freeze_array(struct r10conf *conf)
866 /* stop syncio and normal IO and wait for everything to
868 * We increment barrier and nr_waiting, and then
869 * wait until nr_pending match nr_queued+1
870 * This is called in the context of one normal IO request
871 * that has failed. Thus any sync request that might be pending
872 * will be blocked by nr_pending, and we need to wait for
873 * pending IO requests to complete or be queued for re-try.
874 * Thus the number queued (nr_queued) plus this request (1)
875 * must match the number of pending IOs (nr_pending) before
878 spin_lock_irq(&conf->resync_lock);
881 wait_event_lock_irq(conf->wait_barrier,
882 conf->nr_pending == conf->nr_queued+1,
884 flush_pending_writes(conf));
886 spin_unlock_irq(&conf->resync_lock);
889 static void unfreeze_array(struct r10conf *conf)
891 /* reverse the effect of the freeze */
892 spin_lock_irq(&conf->resync_lock);
895 wake_up(&conf->wait_barrier);
896 spin_unlock_irq(&conf->resync_lock);
899 static void make_request(struct mddev *mddev, struct bio * bio)
901 struct r10conf *conf = mddev->private;
902 struct r10bio *r10_bio;
903 struct bio *read_bio;
905 int chunk_sects = conf->chunk_mask + 1;
906 const int rw = bio_data_dir(bio);
907 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
908 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
910 struct md_rdev *blocked_rdev;
915 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
916 md_flush_request(mddev, bio);
920 /* If this request crosses a chunk boundary, we need to
921 * split it. This will only happen for 1 PAGE (or less) requests.
923 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
925 conf->near_copies < conf->raid_disks)) {
927 /* Sanity check -- queue functions should prevent this happening */
928 if (bio->bi_vcnt != 1 ||
931 /* This is a one page bio that upper layers
932 * refuse to split for us, so we need to split it.
935 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
937 /* Each of these 'make_request' calls will call 'wait_barrier'.
938 * If the first succeeds but the second blocks due to the resync
939 * thread raising the barrier, we will deadlock because the
940 * IO to the underlying device will be queued in generic_make_request
941 * and will never complete, so will never reduce nr_pending.
942 * So increment nr_waiting here so no new raise_barriers will
943 * succeed, and so the second wait_barrier cannot block.
945 spin_lock_irq(&conf->resync_lock);
947 spin_unlock_irq(&conf->resync_lock);
949 make_request(mddev, &bp->bio1);
950 make_request(mddev, &bp->bio2);
952 spin_lock_irq(&conf->resync_lock);
954 wake_up(&conf->wait_barrier);
955 spin_unlock_irq(&conf->resync_lock);
957 bio_pair_release(bp);
960 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
961 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
962 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
968 md_write_start(mddev, bio);
971 * Register the new request and wait if the reconstruction
972 * thread has put up a bar for new requests.
973 * Continue immediately if no resync is active currently.
977 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
979 r10_bio->master_bio = bio;
980 r10_bio->sectors = bio->bi_size >> 9;
982 r10_bio->mddev = mddev;
983 r10_bio->sector = bio->bi_sector;
986 /* We might need to issue multiple reads to different
987 * devices if there are bad blocks around, so we keep
988 * track of the number of reads in bio->bi_phys_segments.
989 * If this is 0, there is only one r10_bio and no locking
990 * will be needed when the request completes. If it is
991 * non-zero, then it is the number of not-completed requests.
993 bio->bi_phys_segments = 0;
994 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
998 * read balancing logic:
1000 struct md_rdev *rdev;
1004 rdev = read_balance(conf, r10_bio, &max_sectors);
1006 raid_end_bio_io(r10_bio);
1009 slot = r10_bio->read_slot;
1011 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1012 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1015 r10_bio->devs[slot].bio = read_bio;
1016 r10_bio->devs[slot].rdev = rdev;
1018 read_bio->bi_sector = r10_bio->devs[slot].addr +
1020 read_bio->bi_bdev = rdev->bdev;
1021 read_bio->bi_end_io = raid10_end_read_request;
1022 read_bio->bi_rw = READ | do_sync;
1023 read_bio->bi_private = r10_bio;
1025 if (max_sectors < r10_bio->sectors) {
1026 /* Could not read all from this device, so we will
1027 * need another r10_bio.
1029 sectors_handled = (r10_bio->sectors + max_sectors
1031 r10_bio->sectors = max_sectors;
1032 spin_lock_irq(&conf->device_lock);
1033 if (bio->bi_phys_segments == 0)
1034 bio->bi_phys_segments = 2;
1036 bio->bi_phys_segments++;
1037 spin_unlock(&conf->device_lock);
1038 /* Cannot call generic_make_request directly
1039 * as that will be queued in __generic_make_request
1040 * and subsequent mempool_alloc might block
1041 * waiting for it. so hand bio over to raid10d.
1043 reschedule_retry(r10_bio);
1045 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1047 r10_bio->master_bio = bio;
1048 r10_bio->sectors = ((bio->bi_size >> 9)
1051 r10_bio->mddev = mddev;
1052 r10_bio->sector = bio->bi_sector + sectors_handled;
1055 generic_make_request(read_bio);
1062 if (conf->pending_count >= max_queued_requests) {
1063 md_wakeup_thread(mddev->thread);
1064 wait_event(conf->wait_barrier,
1065 conf->pending_count < max_queued_requests);
1067 /* first select target devices under rcu_lock and
1068 * inc refcount on their rdev. Record them by setting
1070 * If there are known/acknowledged bad blocks on any device
1071 * on which we have seen a write error, we want to avoid
1072 * writing to those blocks. This potentially requires several
1073 * writes to write around the bad blocks. Each set of writes
1074 * gets its own r10_bio with a set of bios attached. The number
1075 * of r10_bios is recored in bio->bi_phys_segments just as with
1078 plugged = mddev_check_plugged(mddev);
1080 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1081 raid10_find_phys(conf, r10_bio);
1083 blocked_rdev = NULL;
1085 max_sectors = r10_bio->sectors;
1087 for (i = 0; i < conf->copies; i++) {
1088 int d = r10_bio->devs[i].devnum;
1089 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1090 struct md_rdev *rrdev = rcu_dereference(
1091 conf->mirrors[d].replacement);
1092 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1093 atomic_inc(&rdev->nr_pending);
1094 blocked_rdev = rdev;
1097 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1098 atomic_inc(&rrdev->nr_pending);
1099 blocked_rdev = rrdev;
1102 if (rrdev && test_bit(Faulty, &rrdev->flags))
1105 r10_bio->devs[i].bio = NULL;
1106 r10_bio->devs[i].repl_bio = NULL;
1107 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1108 set_bit(R10BIO_Degraded, &r10_bio->state);
1111 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1113 sector_t dev_sector = r10_bio->devs[i].addr;
1117 is_bad = is_badblock(rdev, dev_sector,
1119 &first_bad, &bad_sectors);
1121 /* Mustn't write here until the bad block
1124 atomic_inc(&rdev->nr_pending);
1125 set_bit(BlockedBadBlocks, &rdev->flags);
1126 blocked_rdev = rdev;
1129 if (is_bad && first_bad <= dev_sector) {
1130 /* Cannot write here at all */
1131 bad_sectors -= (dev_sector - first_bad);
1132 if (bad_sectors < max_sectors)
1133 /* Mustn't write more than bad_sectors
1134 * to other devices yet
1136 max_sectors = bad_sectors;
1137 /* We don't set R10BIO_Degraded as that
1138 * only applies if the disk is missing,
1139 * so it might be re-added, and we want to
1140 * know to recover this chunk.
1141 * In this case the device is here, and the
1142 * fact that this chunk is not in-sync is
1143 * recorded in the bad block log.
1148 int good_sectors = first_bad - dev_sector;
1149 if (good_sectors < max_sectors)
1150 max_sectors = good_sectors;
1153 r10_bio->devs[i].bio = bio;
1154 atomic_inc(&rdev->nr_pending);
1156 r10_bio->devs[i].repl_bio = bio;
1157 atomic_inc(&rrdev->nr_pending);
1162 if (unlikely(blocked_rdev)) {
1163 /* Have to wait for this device to get unblocked, then retry */
1167 for (j = 0; j < i; j++) {
1168 if (r10_bio->devs[j].bio) {
1169 d = r10_bio->devs[j].devnum;
1170 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1172 if (r10_bio->devs[j].repl_bio) {
1173 d = r10_bio->devs[j].devnum;
1175 conf->mirrors[d].replacement, mddev);
1178 allow_barrier(conf);
1179 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1184 if (max_sectors < r10_bio->sectors) {
1185 /* We are splitting this into multiple parts, so
1186 * we need to prepare for allocating another r10_bio.
1188 r10_bio->sectors = max_sectors;
1189 spin_lock_irq(&conf->device_lock);
1190 if (bio->bi_phys_segments == 0)
1191 bio->bi_phys_segments = 2;
1193 bio->bi_phys_segments++;
1194 spin_unlock_irq(&conf->device_lock);
1196 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1198 atomic_set(&r10_bio->remaining, 1);
1199 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1201 for (i = 0; i < conf->copies; i++) {
1203 int d = r10_bio->devs[i].devnum;
1204 if (!r10_bio->devs[i].bio)
1207 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1208 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1210 r10_bio->devs[i].bio = mbio;
1212 mbio->bi_sector = (r10_bio->devs[i].addr+
1213 conf->mirrors[d].rdev->data_offset);
1214 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1215 mbio->bi_end_io = raid10_end_write_request;
1216 mbio->bi_rw = WRITE | do_sync | do_fua;
1217 mbio->bi_private = r10_bio;
1219 atomic_inc(&r10_bio->remaining);
1220 spin_lock_irqsave(&conf->device_lock, flags);
1221 bio_list_add(&conf->pending_bio_list, mbio);
1222 conf->pending_count++;
1223 spin_unlock_irqrestore(&conf->device_lock, flags);
1225 if (!r10_bio->devs[i].repl_bio)
1228 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1229 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1231 r10_bio->devs[i].repl_bio = mbio;
1233 mbio->bi_sector = (r10_bio->devs[i].addr+
1234 conf->mirrors[d].replacement->data_offset);
1235 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1236 mbio->bi_end_io = raid10_end_write_request;
1237 mbio->bi_rw = WRITE | do_sync | do_fua;
1238 mbio->bi_private = r10_bio;
1240 atomic_inc(&r10_bio->remaining);
1241 spin_lock_irqsave(&conf->device_lock, flags);
1242 bio_list_add(&conf->pending_bio_list, mbio);
1243 conf->pending_count++;
1244 spin_unlock_irqrestore(&conf->device_lock, flags);
1247 /* Don't remove the bias on 'remaining' (one_write_done) until
1248 * after checking if we need to go around again.
1251 if (sectors_handled < (bio->bi_size >> 9)) {
1252 one_write_done(r10_bio);
1253 /* We need another r10_bio. It has already been counted
1254 * in bio->bi_phys_segments.
1256 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1258 r10_bio->master_bio = bio;
1259 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1261 r10_bio->mddev = mddev;
1262 r10_bio->sector = bio->bi_sector + sectors_handled;
1266 one_write_done(r10_bio);
1268 /* In case raid10d snuck in to freeze_array */
1269 wake_up(&conf->wait_barrier);
1271 if (do_sync || !mddev->bitmap || !plugged)
1272 md_wakeup_thread(mddev->thread);
1275 static void status(struct seq_file *seq, struct mddev *mddev)
1277 struct r10conf *conf = mddev->private;
1280 if (conf->near_copies < conf->raid_disks)
1281 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1282 if (conf->near_copies > 1)
1283 seq_printf(seq, " %d near-copies", conf->near_copies);
1284 if (conf->far_copies > 1) {
1285 if (conf->far_offset)
1286 seq_printf(seq, " %d offset-copies", conf->far_copies);
1288 seq_printf(seq, " %d far-copies", conf->far_copies);
1290 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1291 conf->raid_disks - mddev->degraded);
1292 for (i = 0; i < conf->raid_disks; i++)
1293 seq_printf(seq, "%s",
1294 conf->mirrors[i].rdev &&
1295 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1296 seq_printf(seq, "]");
1299 /* check if there are enough drives for
1300 * every block to appear on atleast one.
1301 * Don't consider the device numbered 'ignore'
1302 * as we might be about to remove it.
1304 static int enough(struct r10conf *conf, int ignore)
1309 int n = conf->copies;
1312 if (conf->mirrors[first].rdev &&
1315 first = (first+1) % conf->raid_disks;
1319 } while (first != 0);
1323 static void error(struct mddev *mddev, struct md_rdev *rdev)
1325 char b[BDEVNAME_SIZE];
1326 struct r10conf *conf = mddev->private;
1329 * If it is not operational, then we have already marked it as dead
1330 * else if it is the last working disks, ignore the error, let the
1331 * next level up know.
1332 * else mark the drive as failed
1334 if (test_bit(In_sync, &rdev->flags)
1335 && !enough(conf, rdev->raid_disk))
1337 * Don't fail the drive, just return an IO error.
1340 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1341 unsigned long flags;
1342 spin_lock_irqsave(&conf->device_lock, flags);
1344 spin_unlock_irqrestore(&conf->device_lock, flags);
1346 * if recovery is running, make sure it aborts.
1348 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1350 set_bit(Blocked, &rdev->flags);
1351 set_bit(Faulty, &rdev->flags);
1352 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1354 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1355 "md/raid10:%s: Operation continuing on %d devices.\n",
1356 mdname(mddev), bdevname(rdev->bdev, b),
1357 mdname(mddev), conf->raid_disks - mddev->degraded);
1360 static void print_conf(struct r10conf *conf)
1363 struct mirror_info *tmp;
1365 printk(KERN_DEBUG "RAID10 conf printout:\n");
1367 printk(KERN_DEBUG "(!conf)\n");
1370 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1373 for (i = 0; i < conf->raid_disks; i++) {
1374 char b[BDEVNAME_SIZE];
1375 tmp = conf->mirrors + i;
1377 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1378 i, !test_bit(In_sync, &tmp->rdev->flags),
1379 !test_bit(Faulty, &tmp->rdev->flags),
1380 bdevname(tmp->rdev->bdev,b));
1384 static void close_sync(struct r10conf *conf)
1387 allow_barrier(conf);
1389 mempool_destroy(conf->r10buf_pool);
1390 conf->r10buf_pool = NULL;
1393 static int raid10_spare_active(struct mddev *mddev)
1396 struct r10conf *conf = mddev->private;
1397 struct mirror_info *tmp;
1399 unsigned long flags;
1402 * Find all non-in_sync disks within the RAID10 configuration
1403 * and mark them in_sync
1405 for (i = 0; i < conf->raid_disks; i++) {
1406 tmp = conf->mirrors + i;
1408 && !test_bit(Faulty, &tmp->rdev->flags)
1409 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1411 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1414 spin_lock_irqsave(&conf->device_lock, flags);
1415 mddev->degraded -= count;
1416 spin_unlock_irqrestore(&conf->device_lock, flags);
1423 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1425 struct r10conf *conf = mddev->private;
1429 int last = conf->raid_disks - 1;
1431 if (mddev->recovery_cp < MaxSector)
1432 /* only hot-add to in-sync arrays, as recovery is
1433 * very different from resync
1436 if (!enough(conf, -1))
1439 if (rdev->raid_disk >= 0)
1440 first = last = rdev->raid_disk;
1442 if (rdev->saved_raid_disk >= first &&
1443 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1444 mirror = rdev->saved_raid_disk;
1447 for ( ; mirror <= last ; mirror++) {
1448 struct mirror_info *p = &conf->mirrors[mirror];
1449 if (p->recovery_disabled == mddev->recovery_disabled)
1454 disk_stack_limits(mddev->gendisk, rdev->bdev,
1455 rdev->data_offset << 9);
1456 /* as we don't honour merge_bvec_fn, we must
1457 * never risk violating it, so limit
1458 * ->max_segments to one lying with a single
1459 * page, as a one page request is never in
1462 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1463 blk_queue_max_segments(mddev->queue, 1);
1464 blk_queue_segment_boundary(mddev->queue,
1465 PAGE_CACHE_SIZE - 1);
1468 p->head_position = 0;
1469 p->recovery_disabled = mddev->recovery_disabled - 1;
1470 rdev->raid_disk = mirror;
1472 if (rdev->saved_raid_disk != mirror)
1474 rcu_assign_pointer(p->rdev, rdev);
1478 md_integrity_add_rdev(rdev, mddev);
1483 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1485 struct r10conf *conf = mddev->private;
1487 int number = rdev->raid_disk;
1488 struct md_rdev **rdevp;
1489 struct mirror_info *p = conf->mirrors + number;
1492 if (rdev == p->rdev)
1494 else if (rdev == p->replacement)
1495 rdevp = &p->replacement;
1499 if (test_bit(In_sync, &rdev->flags) ||
1500 atomic_read(&rdev->nr_pending)) {
1504 /* Only remove faulty devices if recovery
1507 if (!test_bit(Faulty, &rdev->flags) &&
1508 mddev->recovery_disabled != p->recovery_disabled &&
1515 if (atomic_read(&rdev->nr_pending)) {
1516 /* lost the race, try later */
1521 err = md_integrity_register(mddev);
1530 static void end_sync_read(struct bio *bio, int error)
1532 struct r10bio *r10_bio = bio->bi_private;
1533 struct r10conf *conf = r10_bio->mddev->private;
1536 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1538 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1539 set_bit(R10BIO_Uptodate, &r10_bio->state);
1541 /* The write handler will notice the lack of
1542 * R10BIO_Uptodate and record any errors etc
1544 atomic_add(r10_bio->sectors,
1545 &conf->mirrors[d].rdev->corrected_errors);
1547 /* for reconstruct, we always reschedule after a read.
1548 * for resync, only after all reads
1550 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1551 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1552 atomic_dec_and_test(&r10_bio->remaining)) {
1553 /* we have read all the blocks,
1554 * do the comparison in process context in raid10d
1556 reschedule_retry(r10_bio);
1560 static void end_sync_request(struct r10bio *r10_bio)
1562 struct mddev *mddev = r10_bio->mddev;
1564 while (atomic_dec_and_test(&r10_bio->remaining)) {
1565 if (r10_bio->master_bio == NULL) {
1566 /* the primary of several recovery bios */
1567 sector_t s = r10_bio->sectors;
1568 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1569 test_bit(R10BIO_WriteError, &r10_bio->state))
1570 reschedule_retry(r10_bio);
1573 md_done_sync(mddev, s, 1);
1576 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1577 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1578 test_bit(R10BIO_WriteError, &r10_bio->state))
1579 reschedule_retry(r10_bio);
1587 static void end_sync_write(struct bio *bio, int error)
1589 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1590 struct r10bio *r10_bio = bio->bi_private;
1591 struct mddev *mddev = r10_bio->mddev;
1592 struct r10conf *conf = mddev->private;
1598 struct md_rdev *rdev;
1600 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1602 rdev = conf->mirrors[d].replacement;
1604 rdev = conf->mirrors[d].rdev;
1608 md_error(mddev, rdev);
1610 set_bit(WriteErrorSeen, &rdev->flags);
1611 set_bit(R10BIO_WriteError, &r10_bio->state);
1613 } else if (is_badblock(rdev,
1614 r10_bio->devs[slot].addr,
1616 &first_bad, &bad_sectors))
1617 set_bit(R10BIO_MadeGood, &r10_bio->state);
1619 rdev_dec_pending(rdev, mddev);
1621 end_sync_request(r10_bio);
1625 * Note: sync and recover and handled very differently for raid10
1626 * This code is for resync.
1627 * For resync, we read through virtual addresses and read all blocks.
1628 * If there is any error, we schedule a write. The lowest numbered
1629 * drive is authoritative.
1630 * However requests come for physical address, so we need to map.
1631 * For every physical address there are raid_disks/copies virtual addresses,
1632 * which is always are least one, but is not necessarly an integer.
1633 * This means that a physical address can span multiple chunks, so we may
1634 * have to submit multiple io requests for a single sync request.
1637 * We check if all blocks are in-sync and only write to blocks that
1640 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1642 struct r10conf *conf = mddev->private;
1644 struct bio *tbio, *fbio;
1646 atomic_set(&r10_bio->remaining, 1);
1648 /* find the first device with a block */
1649 for (i=0; i<conf->copies; i++)
1650 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1653 if (i == conf->copies)
1657 fbio = r10_bio->devs[i].bio;
1659 /* now find blocks with errors */
1660 for (i=0 ; i < conf->copies ; i++) {
1662 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1664 tbio = r10_bio->devs[i].bio;
1666 if (tbio->bi_end_io != end_sync_read)
1670 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1671 /* We know that the bi_io_vec layout is the same for
1672 * both 'first' and 'i', so we just compare them.
1673 * All vec entries are PAGE_SIZE;
1675 for (j = 0; j < vcnt; j++)
1676 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1677 page_address(tbio->bi_io_vec[j].bv_page),
1682 mddev->resync_mismatches += r10_bio->sectors;
1683 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1684 /* Don't fix anything. */
1687 /* Ok, we need to write this bio, either to correct an
1688 * inconsistency or to correct an unreadable block.
1689 * First we need to fixup bv_offset, bv_len and
1690 * bi_vecs, as the read request might have corrupted these
1692 tbio->bi_vcnt = vcnt;
1693 tbio->bi_size = r10_bio->sectors << 9;
1695 tbio->bi_phys_segments = 0;
1696 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1697 tbio->bi_flags |= 1 << BIO_UPTODATE;
1698 tbio->bi_next = NULL;
1699 tbio->bi_rw = WRITE;
1700 tbio->bi_private = r10_bio;
1701 tbio->bi_sector = r10_bio->devs[i].addr;
1703 for (j=0; j < vcnt ; j++) {
1704 tbio->bi_io_vec[j].bv_offset = 0;
1705 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1707 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1708 page_address(fbio->bi_io_vec[j].bv_page),
1711 tbio->bi_end_io = end_sync_write;
1713 d = r10_bio->devs[i].devnum;
1714 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1715 atomic_inc(&r10_bio->remaining);
1716 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1718 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1719 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1720 generic_make_request(tbio);
1723 /* Now write out to any replacement devices
1726 for (i = 0; i < conf->copies; i++) {
1728 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1730 tbio = r10_bio->devs[i].repl_bio;
1731 if (!tbio || !tbio->bi_end_io)
1733 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1734 && r10_bio->devs[i].bio != fbio)
1735 for (j = 0; j < vcnt; j++)
1736 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1737 page_address(fbio->bi_io_vec[j].bv_page),
1739 d = r10_bio->devs[i].devnum;
1740 atomic_inc(&r10_bio->remaining);
1741 md_sync_acct(conf->mirrors[d].replacement->bdev,
1742 tbio->bi_size >> 9);
1743 generic_make_request(tbio);
1747 if (atomic_dec_and_test(&r10_bio->remaining)) {
1748 md_done_sync(mddev, r10_bio->sectors, 1);
1754 * Now for the recovery code.
1755 * Recovery happens across physical sectors.
1756 * We recover all non-is_sync drives by finding the virtual address of
1757 * each, and then choose a working drive that also has that virt address.
1758 * There is a separate r10_bio for each non-in_sync drive.
1759 * Only the first two slots are in use. The first for reading,
1760 * The second for writing.
1763 static void fix_recovery_read_error(struct r10bio *r10_bio)
1765 /* We got a read error during recovery.
1766 * We repeat the read in smaller page-sized sections.
1767 * If a read succeeds, write it to the new device or record
1768 * a bad block if we cannot.
1769 * If a read fails, record a bad block on both old and
1772 struct mddev *mddev = r10_bio->mddev;
1773 struct r10conf *conf = mddev->private;
1774 struct bio *bio = r10_bio->devs[0].bio;
1776 int sectors = r10_bio->sectors;
1778 int dr = r10_bio->devs[0].devnum;
1779 int dw = r10_bio->devs[1].devnum;
1783 struct md_rdev *rdev;
1787 if (s > (PAGE_SIZE>>9))
1790 rdev = conf->mirrors[dr].rdev;
1791 addr = r10_bio->devs[0].addr + sect,
1792 ok = sync_page_io(rdev,
1795 bio->bi_io_vec[idx].bv_page,
1798 rdev = conf->mirrors[dw].rdev;
1799 addr = r10_bio->devs[1].addr + sect;
1800 ok = sync_page_io(rdev,
1803 bio->bi_io_vec[idx].bv_page,
1806 set_bit(WriteErrorSeen, &rdev->flags);
1809 /* We don't worry if we cannot set a bad block -
1810 * it really is bad so there is no loss in not
1813 rdev_set_badblocks(rdev, addr, s, 0);
1815 if (rdev != conf->mirrors[dw].rdev) {
1816 /* need bad block on destination too */
1817 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1818 addr = r10_bio->devs[1].addr + sect;
1819 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1821 /* just abort the recovery */
1823 "md/raid10:%s: recovery aborted"
1824 " due to read error\n",
1827 conf->mirrors[dw].recovery_disabled
1828 = mddev->recovery_disabled;
1829 set_bit(MD_RECOVERY_INTR,
1842 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1844 struct r10conf *conf = mddev->private;
1846 struct bio *wbio, *wbio2;
1848 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1849 fix_recovery_read_error(r10_bio);
1850 end_sync_request(r10_bio);
1855 * share the pages with the first bio
1856 * and submit the write request
1858 d = r10_bio->devs[1].devnum;
1859 wbio = r10_bio->devs[1].bio;
1860 wbio2 = r10_bio->devs[1].repl_bio;
1861 if (wbio->bi_end_io) {
1862 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1863 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1864 generic_make_request(wbio);
1866 if (wbio2 && wbio2->bi_end_io) {
1867 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
1868 md_sync_acct(conf->mirrors[d].replacement->bdev,
1869 wbio2->bi_size >> 9);
1870 generic_make_request(wbio2);
1876 * Used by fix_read_error() to decay the per rdev read_errors.
1877 * We halve the read error count for every hour that has elapsed
1878 * since the last recorded read error.
1881 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1883 struct timespec cur_time_mon;
1884 unsigned long hours_since_last;
1885 unsigned int read_errors = atomic_read(&rdev->read_errors);
1887 ktime_get_ts(&cur_time_mon);
1889 if (rdev->last_read_error.tv_sec == 0 &&
1890 rdev->last_read_error.tv_nsec == 0) {
1891 /* first time we've seen a read error */
1892 rdev->last_read_error = cur_time_mon;
1896 hours_since_last = (cur_time_mon.tv_sec -
1897 rdev->last_read_error.tv_sec) / 3600;
1899 rdev->last_read_error = cur_time_mon;
1902 * if hours_since_last is > the number of bits in read_errors
1903 * just set read errors to 0. We do this to avoid
1904 * overflowing the shift of read_errors by hours_since_last.
1906 if (hours_since_last >= 8 * sizeof(read_errors))
1907 atomic_set(&rdev->read_errors, 0);
1909 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1912 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1913 int sectors, struct page *page, int rw)
1918 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1919 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1921 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1925 set_bit(WriteErrorSeen, &rdev->flags);
1926 /* need to record an error - either for the block or the device */
1927 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1928 md_error(rdev->mddev, rdev);
1933 * This is a kernel thread which:
1935 * 1. Retries failed read operations on working mirrors.
1936 * 2. Updates the raid superblock when problems encounter.
1937 * 3. Performs writes following reads for array synchronising.
1940 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1942 int sect = 0; /* Offset from r10_bio->sector */
1943 int sectors = r10_bio->sectors;
1944 struct md_rdev*rdev;
1945 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1946 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1948 /* still own a reference to this rdev, so it cannot
1949 * have been cleared recently.
1951 rdev = conf->mirrors[d].rdev;
1953 if (test_bit(Faulty, &rdev->flags))
1954 /* drive has already been failed, just ignore any
1955 more fix_read_error() attempts */
1958 check_decay_read_errors(mddev, rdev);
1959 atomic_inc(&rdev->read_errors);
1960 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1961 char b[BDEVNAME_SIZE];
1962 bdevname(rdev->bdev, b);
1965 "md/raid10:%s: %s: Raid device exceeded "
1966 "read_error threshold [cur %d:max %d]\n",
1968 atomic_read(&rdev->read_errors), max_read_errors);
1970 "md/raid10:%s: %s: Failing raid device\n",
1972 md_error(mddev, conf->mirrors[d].rdev);
1978 int sl = r10_bio->read_slot;
1982 if (s > (PAGE_SIZE>>9))
1990 d = r10_bio->devs[sl].devnum;
1991 rdev = rcu_dereference(conf->mirrors[d].rdev);
1993 test_bit(In_sync, &rdev->flags) &&
1994 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1995 &first_bad, &bad_sectors) == 0) {
1996 atomic_inc(&rdev->nr_pending);
1998 success = sync_page_io(rdev,
1999 r10_bio->devs[sl].addr +
2002 conf->tmppage, READ, false);
2003 rdev_dec_pending(rdev, mddev);
2009 if (sl == conf->copies)
2011 } while (!success && sl != r10_bio->read_slot);
2015 /* Cannot read from anywhere, just mark the block
2016 * as bad on the first device to discourage future
2019 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2020 rdev = conf->mirrors[dn].rdev;
2022 if (!rdev_set_badblocks(
2024 r10_bio->devs[r10_bio->read_slot].addr
2027 md_error(mddev, rdev);
2032 /* write it back and re-read */
2034 while (sl != r10_bio->read_slot) {
2035 char b[BDEVNAME_SIZE];
2040 d = r10_bio->devs[sl].devnum;
2041 rdev = rcu_dereference(conf->mirrors[d].rdev);
2043 !test_bit(In_sync, &rdev->flags))
2046 atomic_inc(&rdev->nr_pending);
2048 if (r10_sync_page_io(rdev,
2049 r10_bio->devs[sl].addr +
2051 s<<9, conf->tmppage, WRITE)
2053 /* Well, this device is dead */
2055 "md/raid10:%s: read correction "
2057 " (%d sectors at %llu on %s)\n",
2059 (unsigned long long)(
2060 sect + rdev->data_offset),
2061 bdevname(rdev->bdev, b));
2062 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2065 bdevname(rdev->bdev, b));
2067 rdev_dec_pending(rdev, mddev);
2071 while (sl != r10_bio->read_slot) {
2072 char b[BDEVNAME_SIZE];
2077 d = r10_bio->devs[sl].devnum;
2078 rdev = rcu_dereference(conf->mirrors[d].rdev);
2080 !test_bit(In_sync, &rdev->flags))
2083 atomic_inc(&rdev->nr_pending);
2085 switch (r10_sync_page_io(rdev,
2086 r10_bio->devs[sl].addr +
2088 s<<9, conf->tmppage,
2091 /* Well, this device is dead */
2093 "md/raid10:%s: unable to read back "
2095 " (%d sectors at %llu on %s)\n",
2097 (unsigned long long)(
2098 sect + rdev->data_offset),
2099 bdevname(rdev->bdev, b));
2100 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2103 bdevname(rdev->bdev, b));
2107 "md/raid10:%s: read error corrected"
2108 " (%d sectors at %llu on %s)\n",
2110 (unsigned long long)(
2111 sect + rdev->data_offset),
2112 bdevname(rdev->bdev, b));
2113 atomic_add(s, &rdev->corrected_errors);
2116 rdev_dec_pending(rdev, mddev);
2126 static void bi_complete(struct bio *bio, int error)
2128 complete((struct completion *)bio->bi_private);
2131 static int submit_bio_wait(int rw, struct bio *bio)
2133 struct completion event;
2136 init_completion(&event);
2137 bio->bi_private = &event;
2138 bio->bi_end_io = bi_complete;
2139 submit_bio(rw, bio);
2140 wait_for_completion(&event);
2142 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2145 static int narrow_write_error(struct r10bio *r10_bio, int i)
2147 struct bio *bio = r10_bio->master_bio;
2148 struct mddev *mddev = r10_bio->mddev;
2149 struct r10conf *conf = mddev->private;
2150 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2151 /* bio has the data to be written to slot 'i' where
2152 * we just recently had a write error.
2153 * We repeatedly clone the bio and trim down to one block,
2154 * then try the write. Where the write fails we record
2156 * It is conceivable that the bio doesn't exactly align with
2157 * blocks. We must handle this.
2159 * We currently own a reference to the rdev.
2165 int sect_to_write = r10_bio->sectors;
2168 if (rdev->badblocks.shift < 0)
2171 block_sectors = 1 << rdev->badblocks.shift;
2172 sector = r10_bio->sector;
2173 sectors = ((r10_bio->sector + block_sectors)
2174 & ~(sector_t)(block_sectors - 1))
2177 while (sect_to_write) {
2179 if (sectors > sect_to_write)
2180 sectors = sect_to_write;
2181 /* Write at 'sector' for 'sectors' */
2182 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2183 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2184 wbio->bi_sector = (r10_bio->devs[i].addr+
2186 (sector - r10_bio->sector));
2187 wbio->bi_bdev = rdev->bdev;
2188 if (submit_bio_wait(WRITE, wbio) == 0)
2190 ok = rdev_set_badblocks(rdev, sector,
2195 sect_to_write -= sectors;
2197 sectors = block_sectors;
2202 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2204 int slot = r10_bio->read_slot;
2206 struct r10conf *conf = mddev->private;
2207 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2208 char b[BDEVNAME_SIZE];
2209 unsigned long do_sync;
2212 /* we got a read error. Maybe the drive is bad. Maybe just
2213 * the block and we can fix it.
2214 * We freeze all other IO, and try reading the block from
2215 * other devices. When we find one, we re-write
2216 * and check it that fixes the read error.
2217 * This is all done synchronously while the array is
2220 if (mddev->ro == 0) {
2222 fix_read_error(conf, mddev, r10_bio);
2223 unfreeze_array(conf);
2225 rdev_dec_pending(rdev, mddev);
2227 bio = r10_bio->devs[slot].bio;
2228 bdevname(bio->bi_bdev, b);
2229 r10_bio->devs[slot].bio =
2230 mddev->ro ? IO_BLOCKED : NULL;
2232 rdev = read_balance(conf, r10_bio, &max_sectors);
2234 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2235 " read error for block %llu\n",
2237 (unsigned long long)r10_bio->sector);
2238 raid_end_bio_io(r10_bio);
2243 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2246 slot = r10_bio->read_slot;
2249 "md/raid10:%s: %s: redirecting"
2250 "sector %llu to another mirror\n",
2252 bdevname(rdev->bdev, b),
2253 (unsigned long long)r10_bio->sector);
2254 bio = bio_clone_mddev(r10_bio->master_bio,
2257 r10_bio->sector - bio->bi_sector,
2259 r10_bio->devs[slot].bio = bio;
2260 r10_bio->devs[slot].rdev = rdev;
2261 bio->bi_sector = r10_bio->devs[slot].addr
2262 + rdev->data_offset;
2263 bio->bi_bdev = rdev->bdev;
2264 bio->bi_rw = READ | do_sync;
2265 bio->bi_private = r10_bio;
2266 bio->bi_end_io = raid10_end_read_request;
2267 if (max_sectors < r10_bio->sectors) {
2268 /* Drat - have to split this up more */
2269 struct bio *mbio = r10_bio->master_bio;
2270 int sectors_handled =
2271 r10_bio->sector + max_sectors
2273 r10_bio->sectors = max_sectors;
2274 spin_lock_irq(&conf->device_lock);
2275 if (mbio->bi_phys_segments == 0)
2276 mbio->bi_phys_segments = 2;
2278 mbio->bi_phys_segments++;
2279 spin_unlock_irq(&conf->device_lock);
2280 generic_make_request(bio);
2283 r10_bio = mempool_alloc(conf->r10bio_pool,
2285 r10_bio->master_bio = mbio;
2286 r10_bio->sectors = (mbio->bi_size >> 9)
2289 set_bit(R10BIO_ReadError,
2291 r10_bio->mddev = mddev;
2292 r10_bio->sector = mbio->bi_sector
2297 generic_make_request(bio);
2300 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2302 /* Some sort of write request has finished and it
2303 * succeeded in writing where we thought there was a
2304 * bad block. So forget the bad block.
2305 * Or possibly if failed and we need to record
2309 struct md_rdev *rdev;
2311 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2312 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2313 for (m = 0; m < conf->copies; m++) {
2314 int dev = r10_bio->devs[m].devnum;
2315 rdev = conf->mirrors[dev].rdev;
2316 if (r10_bio->devs[m].bio == NULL)
2318 if (test_bit(BIO_UPTODATE,
2319 &r10_bio->devs[m].bio->bi_flags)) {
2320 rdev_clear_badblocks(
2322 r10_bio->devs[m].addr,
2325 if (!rdev_set_badblocks(
2327 r10_bio->devs[m].addr,
2328 r10_bio->sectors, 0))
2329 md_error(conf->mddev, rdev);
2331 rdev = conf->mirrors[dev].replacement;
2332 if (r10_bio->devs[m].repl_bio == NULL)
2334 if (test_bit(BIO_UPTODATE,
2335 &r10_bio->devs[m].repl_bio->bi_flags)) {
2336 rdev_clear_badblocks(
2338 r10_bio->devs[m].addr,
2341 if (!rdev_set_badblocks(
2343 r10_bio->devs[m].addr,
2344 r10_bio->sectors, 0))
2345 md_error(conf->mddev, rdev);
2350 for (m = 0; m < conf->copies; m++) {
2351 int dev = r10_bio->devs[m].devnum;
2352 struct bio *bio = r10_bio->devs[m].bio;
2353 rdev = conf->mirrors[dev].rdev;
2354 if (bio == IO_MADE_GOOD) {
2355 rdev_clear_badblocks(
2357 r10_bio->devs[m].addr,
2359 rdev_dec_pending(rdev, conf->mddev);
2360 } else if (bio != NULL &&
2361 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2362 if (!narrow_write_error(r10_bio, m)) {
2363 md_error(conf->mddev, rdev);
2364 set_bit(R10BIO_Degraded,
2367 rdev_dec_pending(rdev, conf->mddev);
2369 bio = r10_bio->devs[m].repl_bio;
2370 rdev = conf->mirrors[dev].replacement;
2371 if (bio == IO_MADE_GOOD) {
2372 rdev_clear_badblocks(
2374 r10_bio->devs[m].addr,
2376 rdev_dec_pending(rdev, conf->mddev);
2379 if (test_bit(R10BIO_WriteError,
2381 close_write(r10_bio);
2382 raid_end_bio_io(r10_bio);
2386 static void raid10d(struct mddev *mddev)
2388 struct r10bio *r10_bio;
2389 unsigned long flags;
2390 struct r10conf *conf = mddev->private;
2391 struct list_head *head = &conf->retry_list;
2392 struct blk_plug plug;
2394 md_check_recovery(mddev);
2396 blk_start_plug(&plug);
2399 flush_pending_writes(conf);
2401 spin_lock_irqsave(&conf->device_lock, flags);
2402 if (list_empty(head)) {
2403 spin_unlock_irqrestore(&conf->device_lock, flags);
2406 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2407 list_del(head->prev);
2409 spin_unlock_irqrestore(&conf->device_lock, flags);
2411 mddev = r10_bio->mddev;
2412 conf = mddev->private;
2413 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2414 test_bit(R10BIO_WriteError, &r10_bio->state))
2415 handle_write_completed(conf, r10_bio);
2416 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2417 sync_request_write(mddev, r10_bio);
2418 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2419 recovery_request_write(mddev, r10_bio);
2420 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2421 handle_read_error(mddev, r10_bio);
2423 /* just a partial read to be scheduled from a
2426 int slot = r10_bio->read_slot;
2427 generic_make_request(r10_bio->devs[slot].bio);
2431 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2432 md_check_recovery(mddev);
2434 blk_finish_plug(&plug);
2438 static int init_resync(struct r10conf *conf)
2443 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2444 BUG_ON(conf->r10buf_pool);
2445 conf->have_replacement = 0;
2446 for (i = 0; i < conf->raid_disks; i++)
2447 if (conf->mirrors[i].replacement)
2448 conf->have_replacement = 1;
2449 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2450 if (!conf->r10buf_pool)
2452 conf->next_resync = 0;
2457 * perform a "sync" on one "block"
2459 * We need to make sure that no normal I/O request - particularly write
2460 * requests - conflict with active sync requests.
2462 * This is achieved by tracking pending requests and a 'barrier' concept
2463 * that can be installed to exclude normal IO requests.
2465 * Resync and recovery are handled very differently.
2466 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2468 * For resync, we iterate over virtual addresses, read all copies,
2469 * and update if there are differences. If only one copy is live,
2471 * For recovery, we iterate over physical addresses, read a good
2472 * value for each non-in_sync drive, and over-write.
2474 * So, for recovery we may have several outstanding complex requests for a
2475 * given address, one for each out-of-sync device. We model this by allocating
2476 * a number of r10_bio structures, one for each out-of-sync device.
2477 * As we setup these structures, we collect all bio's together into a list
2478 * which we then process collectively to add pages, and then process again
2479 * to pass to generic_make_request.
2481 * The r10_bio structures are linked using a borrowed master_bio pointer.
2482 * This link is counted in ->remaining. When the r10_bio that points to NULL
2483 * has its remaining count decremented to 0, the whole complex operation
2488 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2489 int *skipped, int go_faster)
2491 struct r10conf *conf = mddev->private;
2492 struct r10bio *r10_bio;
2493 struct bio *biolist = NULL, *bio;
2494 sector_t max_sector, nr_sectors;
2497 sector_t sync_blocks;
2498 sector_t sectors_skipped = 0;
2499 int chunks_skipped = 0;
2501 if (!conf->r10buf_pool)
2502 if (init_resync(conf))
2506 max_sector = mddev->dev_sectors;
2507 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2508 max_sector = mddev->resync_max_sectors;
2509 if (sector_nr >= max_sector) {
2510 /* If we aborted, we need to abort the
2511 * sync on the 'current' bitmap chucks (there can
2512 * be several when recovering multiple devices).
2513 * as we may have started syncing it but not finished.
2514 * We can find the current address in
2515 * mddev->curr_resync, but for recovery,
2516 * we need to convert that to several
2517 * virtual addresses.
2519 if (mddev->curr_resync < max_sector) { /* aborted */
2520 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2521 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2523 else for (i=0; i<conf->raid_disks; i++) {
2525 raid10_find_virt(conf, mddev->curr_resync, i);
2526 bitmap_end_sync(mddev->bitmap, sect,
2530 /* completed sync */
2531 if ((!mddev->bitmap || conf->fullsync)
2532 && conf->have_replacement
2533 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2534 /* Completed a full sync so the replacements
2535 * are now fully recovered.
2537 for (i = 0; i < conf->raid_disks; i++)
2538 if (conf->mirrors[i].replacement)
2539 conf->mirrors[i].replacement
2545 bitmap_close_sync(mddev->bitmap);
2548 return sectors_skipped;
2550 if (chunks_skipped >= conf->raid_disks) {
2551 /* if there has been nothing to do on any drive,
2552 * then there is nothing to do at all..
2555 return (max_sector - sector_nr) + sectors_skipped;
2558 if (max_sector > mddev->resync_max)
2559 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2561 /* make sure whole request will fit in a chunk - if chunks
2564 if (conf->near_copies < conf->raid_disks &&
2565 max_sector > (sector_nr | conf->chunk_mask))
2566 max_sector = (sector_nr | conf->chunk_mask) + 1;
2568 * If there is non-resync activity waiting for us then
2569 * put in a delay to throttle resync.
2571 if (!go_faster && conf->nr_waiting)
2572 msleep_interruptible(1000);
2574 /* Again, very different code for resync and recovery.
2575 * Both must result in an r10bio with a list of bios that
2576 * have bi_end_io, bi_sector, bi_bdev set,
2577 * and bi_private set to the r10bio.
2578 * For recovery, we may actually create several r10bios
2579 * with 2 bios in each, that correspond to the bios in the main one.
2580 * In this case, the subordinate r10bios link back through a
2581 * borrowed master_bio pointer, and the counter in the master
2582 * includes a ref from each subordinate.
2584 /* First, we decide what to do and set ->bi_end_io
2585 * To end_sync_read if we want to read, and
2586 * end_sync_write if we will want to write.
2589 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2590 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2591 /* recovery... the complicated one */
2595 for (i=0 ; i<conf->raid_disks; i++) {
2601 struct mirror_info *mirror = &conf->mirrors[i];
2603 if ((mirror->rdev == NULL ||
2604 test_bit(In_sync, &mirror->rdev->flags))
2606 (mirror->replacement == NULL ||
2608 &mirror->replacement->flags)))
2612 /* want to reconstruct this device */
2614 sect = raid10_find_virt(conf, sector_nr, i);
2615 /* Unless we are doing a full sync, or a replacement
2616 * we only need to recover the block if it is set in
2619 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2621 if (sync_blocks < max_sync)
2622 max_sync = sync_blocks;
2624 mirror->replacement == NULL &&
2626 /* yep, skip the sync_blocks here, but don't assume
2627 * that there will never be anything to do here
2629 chunks_skipped = -1;
2633 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2634 raise_barrier(conf, rb2 != NULL);
2635 atomic_set(&r10_bio->remaining, 0);
2637 r10_bio->master_bio = (struct bio*)rb2;
2639 atomic_inc(&rb2->remaining);
2640 r10_bio->mddev = mddev;
2641 set_bit(R10BIO_IsRecover, &r10_bio->state);
2642 r10_bio->sector = sect;
2644 raid10_find_phys(conf, r10_bio);
2646 /* Need to check if the array will still be
2649 for (j=0; j<conf->raid_disks; j++)
2650 if (conf->mirrors[j].rdev == NULL ||
2651 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2656 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2657 &sync_blocks, still_degraded);
2660 for (j=0; j<conf->copies;j++) {
2662 int d = r10_bio->devs[j].devnum;
2663 sector_t from_addr, to_addr;
2664 struct md_rdev *rdev;
2665 sector_t sector, first_bad;
2667 if (!conf->mirrors[d].rdev ||
2668 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2670 /* This is where we read from */
2672 rdev = conf->mirrors[d].rdev;
2673 sector = r10_bio->devs[j].addr;
2675 if (is_badblock(rdev, sector, max_sync,
2676 &first_bad, &bad_sectors)) {
2677 if (first_bad > sector)
2678 max_sync = first_bad - sector;
2680 bad_sectors -= (sector
2682 if (max_sync > bad_sectors)
2683 max_sync = bad_sectors;
2687 bio = r10_bio->devs[0].bio;
2688 bio->bi_next = biolist;
2690 bio->bi_private = r10_bio;
2691 bio->bi_end_io = end_sync_read;
2693 from_addr = r10_bio->devs[j].addr;
2694 bio->bi_sector = from_addr + rdev->data_offset;
2695 bio->bi_bdev = rdev->bdev;
2696 atomic_inc(&rdev->nr_pending);
2697 /* and we write to 'i' (if not in_sync) */
2699 for (k=0; k<conf->copies; k++)
2700 if (r10_bio->devs[k].devnum == i)
2702 BUG_ON(k == conf->copies);
2703 to_addr = r10_bio->devs[k].addr;
2704 r10_bio->devs[0].devnum = d;
2705 r10_bio->devs[0].addr = from_addr;
2706 r10_bio->devs[1].devnum = i;
2707 r10_bio->devs[1].addr = to_addr;
2709 rdev = mirror->rdev;
2710 if (!test_bit(In_sync, &rdev->flags)) {
2711 bio = r10_bio->devs[1].bio;
2712 bio->bi_next = biolist;
2714 bio->bi_private = r10_bio;
2715 bio->bi_end_io = end_sync_write;
2717 bio->bi_sector = to_addr
2718 + rdev->data_offset;
2719 bio->bi_bdev = rdev->bdev;
2720 atomic_inc(&r10_bio->remaining);
2722 r10_bio->devs[1].bio->bi_end_io = NULL;
2724 /* and maybe write to replacement */
2725 bio = r10_bio->devs[1].repl_bio;
2727 bio->bi_end_io = NULL;
2728 rdev = mirror->replacement;
2729 /* Note: if rdev != NULL, then bio
2730 * cannot be NULL as r10buf_pool_alloc will
2731 * have allocated it.
2732 * So the second test here is pointless.
2733 * But it keeps semantic-checkers happy, and
2734 * this comment keeps human reviewers
2737 if (rdev == NULL || bio == NULL ||
2738 test_bit(Faulty, &rdev->flags))
2740 bio->bi_next = biolist;
2742 bio->bi_private = r10_bio;
2743 bio->bi_end_io = end_sync_write;
2745 bio->bi_sector = to_addr + rdev->data_offset;
2746 bio->bi_bdev = rdev->bdev;
2747 atomic_inc(&r10_bio->remaining);
2750 if (j == conf->copies) {
2751 /* Cannot recover, so abort the recovery or
2752 * record a bad block */
2755 atomic_dec(&rb2->remaining);
2758 /* problem is that there are bad blocks
2759 * on other device(s)
2762 for (k = 0; k < conf->copies; k++)
2763 if (r10_bio->devs[k].devnum == i)
2765 if (!test_bit(In_sync,
2766 &mirror->rdev->flags)
2767 && !rdev_set_badblocks(
2769 r10_bio->devs[k].addr,
2772 if (mirror->replacement &&
2773 !rdev_set_badblocks(
2774 mirror->replacement,
2775 r10_bio->devs[k].addr,
2780 if (!test_and_set_bit(MD_RECOVERY_INTR,
2782 printk(KERN_INFO "md/raid10:%s: insufficient "
2783 "working devices for recovery.\n",
2785 mirror->recovery_disabled
2786 = mddev->recovery_disabled;
2791 if (biolist == NULL) {
2793 struct r10bio *rb2 = r10_bio;
2794 r10_bio = (struct r10bio*) rb2->master_bio;
2795 rb2->master_bio = NULL;
2801 /* resync. Schedule a read for every block at this virt offset */
2804 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2806 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2807 &sync_blocks, mddev->degraded) &&
2808 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2809 &mddev->recovery)) {
2810 /* We can skip this block */
2812 return sync_blocks + sectors_skipped;
2814 if (sync_blocks < max_sync)
2815 max_sync = sync_blocks;
2816 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2818 r10_bio->mddev = mddev;
2819 atomic_set(&r10_bio->remaining, 0);
2820 raise_barrier(conf, 0);
2821 conf->next_resync = sector_nr;
2823 r10_bio->master_bio = NULL;
2824 r10_bio->sector = sector_nr;
2825 set_bit(R10BIO_IsSync, &r10_bio->state);
2826 raid10_find_phys(conf, r10_bio);
2827 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2829 for (i=0; i<conf->copies; i++) {
2830 int d = r10_bio->devs[i].devnum;
2831 sector_t first_bad, sector;
2834 if (r10_bio->devs[i].repl_bio)
2835 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
2837 bio = r10_bio->devs[i].bio;
2838 bio->bi_end_io = NULL;
2839 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2840 if (conf->mirrors[d].rdev == NULL ||
2841 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2843 sector = r10_bio->devs[i].addr;
2844 if (is_badblock(conf->mirrors[d].rdev,
2846 &first_bad, &bad_sectors)) {
2847 if (first_bad > sector)
2848 max_sync = first_bad - sector;
2850 bad_sectors -= (sector - first_bad);
2851 if (max_sync > bad_sectors)
2852 max_sync = max_sync;
2856 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2857 atomic_inc(&r10_bio->remaining);
2858 bio->bi_next = biolist;
2860 bio->bi_private = r10_bio;
2861 bio->bi_end_io = end_sync_read;
2863 bio->bi_sector = sector +
2864 conf->mirrors[d].rdev->data_offset;
2865 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2868 if (conf->mirrors[d].replacement == NULL ||
2870 &conf->mirrors[d].replacement->flags))
2873 /* Need to set up for writing to the replacement */
2874 bio = r10_bio->devs[i].repl_bio;
2875 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2877 sector = r10_bio->devs[i].addr;
2878 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2879 bio->bi_next = biolist;
2881 bio->bi_private = r10_bio;
2882 bio->bi_end_io = end_sync_write;
2884 bio->bi_sector = sector +
2885 conf->mirrors[d].replacement->data_offset;
2886 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
2891 for (i=0; i<conf->copies; i++) {
2892 int d = r10_bio->devs[i].devnum;
2893 if (r10_bio->devs[i].bio->bi_end_io)
2894 rdev_dec_pending(conf->mirrors[d].rdev,
2896 if (r10_bio->devs[i].repl_bio &&
2897 r10_bio->devs[i].repl_bio->bi_end_io)
2899 conf->mirrors[d].replacement,
2908 for (bio = biolist; bio ; bio=bio->bi_next) {
2910 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2912 bio->bi_flags |= 1 << BIO_UPTODATE;
2915 bio->bi_phys_segments = 0;
2920 if (sector_nr + max_sync < max_sector)
2921 max_sector = sector_nr + max_sync;
2924 int len = PAGE_SIZE;
2925 if (sector_nr + (len>>9) > max_sector)
2926 len = (max_sector - sector_nr) << 9;
2929 for (bio= biolist ; bio ; bio=bio->bi_next) {
2931 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2932 if (bio_add_page(bio, page, len, 0))
2936 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2937 for (bio2 = biolist;
2938 bio2 && bio2 != bio;
2939 bio2 = bio2->bi_next) {
2940 /* remove last page from this bio */
2942 bio2->bi_size -= len;
2943 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2947 nr_sectors += len>>9;
2948 sector_nr += len>>9;
2949 } while (biolist->bi_vcnt < RESYNC_PAGES);
2951 r10_bio->sectors = nr_sectors;
2955 biolist = biolist->bi_next;
2957 bio->bi_next = NULL;
2958 r10_bio = bio->bi_private;
2959 r10_bio->sectors = nr_sectors;
2961 if (bio->bi_end_io == end_sync_read) {
2962 md_sync_acct(bio->bi_bdev, nr_sectors);
2963 generic_make_request(bio);
2967 if (sectors_skipped)
2968 /* pretend they weren't skipped, it makes
2969 * no important difference in this case
2971 md_done_sync(mddev, sectors_skipped, 1);
2973 return sectors_skipped + nr_sectors;
2975 /* There is nowhere to write, so all non-sync
2976 * drives must be failed or in resync, all drives
2977 * have a bad block, so try the next chunk...
2979 if (sector_nr + max_sync < max_sector)
2980 max_sector = sector_nr + max_sync;
2982 sectors_skipped += (max_sector - sector_nr);
2984 sector_nr = max_sector;
2989 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2992 struct r10conf *conf = mddev->private;
2995 raid_disks = conf->raid_disks;
2997 sectors = conf->dev_sectors;
2999 size = sectors >> conf->chunk_shift;
3000 sector_div(size, conf->far_copies);
3001 size = size * raid_disks;
3002 sector_div(size, conf->near_copies);
3004 return size << conf->chunk_shift;
3008 static struct r10conf *setup_conf(struct mddev *mddev)
3010 struct r10conf *conf = NULL;
3012 sector_t stride, size;
3015 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
3016 !is_power_of_2(mddev->new_chunk_sectors)) {
3017 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3018 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3019 mdname(mddev), PAGE_SIZE);
3023 nc = mddev->new_layout & 255;
3024 fc = (mddev->new_layout >> 8) & 255;
3025 fo = mddev->new_layout & (1<<16);
3027 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
3028 (mddev->new_layout >> 17)) {
3029 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3030 mdname(mddev), mddev->new_layout);
3035 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3039 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
3044 conf->tmppage = alloc_page(GFP_KERNEL);
3049 conf->raid_disks = mddev->raid_disks;
3050 conf->near_copies = nc;
3051 conf->far_copies = fc;
3052 conf->copies = nc*fc;
3053 conf->far_offset = fo;
3054 conf->chunk_mask = mddev->new_chunk_sectors - 1;
3055 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
3057 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3058 r10bio_pool_free, conf);
3059 if (!conf->r10bio_pool)
3062 size = mddev->dev_sectors >> conf->chunk_shift;
3063 sector_div(size, fc);
3064 size = size * conf->raid_disks;
3065 sector_div(size, nc);
3066 /* 'size' is now the number of chunks in the array */
3067 /* calculate "used chunks per device" in 'stride' */
3068 stride = size * conf->copies;
3070 /* We need to round up when dividing by raid_disks to
3071 * get the stride size.
3073 stride += conf->raid_disks - 1;
3074 sector_div(stride, conf->raid_disks);
3076 conf->dev_sectors = stride << conf->chunk_shift;
3081 sector_div(stride, fc);
3082 conf->stride = stride << conf->chunk_shift;
3085 spin_lock_init(&conf->device_lock);
3086 INIT_LIST_HEAD(&conf->retry_list);
3088 spin_lock_init(&conf->resync_lock);
3089 init_waitqueue_head(&conf->wait_barrier);
3091 conf->thread = md_register_thread(raid10d, mddev, NULL);
3095 conf->mddev = mddev;
3099 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3102 if (conf->r10bio_pool)
3103 mempool_destroy(conf->r10bio_pool);
3104 kfree(conf->mirrors);
3105 safe_put_page(conf->tmppage);
3108 return ERR_PTR(err);
3111 static int run(struct mddev *mddev)
3113 struct r10conf *conf;
3114 int i, disk_idx, chunk_size;
3115 struct mirror_info *disk;
3116 struct md_rdev *rdev;
3120 * copy the already verified devices into our private RAID10
3121 * bookkeeping area. [whatever we allocate in run(),
3122 * should be freed in stop()]
3125 if (mddev->private == NULL) {
3126 conf = setup_conf(mddev);
3128 return PTR_ERR(conf);
3129 mddev->private = conf;
3131 conf = mddev->private;
3135 mddev->thread = conf->thread;
3136 conf->thread = NULL;
3138 chunk_size = mddev->chunk_sectors << 9;
3139 blk_queue_io_min(mddev->queue, chunk_size);
3140 if (conf->raid_disks % conf->near_copies)
3141 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
3143 blk_queue_io_opt(mddev->queue, chunk_size *
3144 (conf->raid_disks / conf->near_copies));
3146 list_for_each_entry(rdev, &mddev->disks, same_set) {
3148 disk_idx = rdev->raid_disk;
3149 if (disk_idx >= conf->raid_disks
3152 disk = conf->mirrors + disk_idx;
3155 disk_stack_limits(mddev->gendisk, rdev->bdev,
3156 rdev->data_offset << 9);
3157 /* as we don't honour merge_bvec_fn, we must never risk
3158 * violating it, so limit max_segments to 1 lying
3159 * within a single page.
3161 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
3162 blk_queue_max_segments(mddev->queue, 1);
3163 blk_queue_segment_boundary(mddev->queue,
3164 PAGE_CACHE_SIZE - 1);
3167 disk->head_position = 0;
3169 /* need to check that every block has at least one working mirror */
3170 if (!enough(conf, -1)) {
3171 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3176 mddev->degraded = 0;
3177 for (i = 0; i < conf->raid_disks; i++) {
3179 disk = conf->mirrors + i;
3182 !test_bit(In_sync, &disk->rdev->flags)) {
3183 disk->head_position = 0;
3188 disk->recovery_disabled = mddev->recovery_disabled - 1;
3191 if (mddev->recovery_cp != MaxSector)
3192 printk(KERN_NOTICE "md/raid10:%s: not clean"
3193 " -- starting background reconstruction\n",
3196 "md/raid10:%s: active with %d out of %d devices\n",
3197 mdname(mddev), conf->raid_disks - mddev->degraded,
3200 * Ok, everything is just fine now
3202 mddev->dev_sectors = conf->dev_sectors;
3203 size = raid10_size(mddev, 0, 0);
3204 md_set_array_sectors(mddev, size);
3205 mddev->resync_max_sectors = size;
3207 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3208 mddev->queue->backing_dev_info.congested_data = mddev;
3210 /* Calculate max read-ahead size.
3211 * We need to readahead at least twice a whole stripe....
3215 int stripe = conf->raid_disks *
3216 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3217 stripe /= conf->near_copies;
3218 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
3219 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
3222 if (conf->near_copies < conf->raid_disks)
3223 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3225 if (md_integrity_register(mddev))
3231 md_unregister_thread(&mddev->thread);
3232 if (conf->r10bio_pool)
3233 mempool_destroy(conf->r10bio_pool);
3234 safe_put_page(conf->tmppage);
3235 kfree(conf->mirrors);
3237 mddev->private = NULL;
3242 static int stop(struct mddev *mddev)
3244 struct r10conf *conf = mddev->private;
3246 raise_barrier(conf, 0);
3247 lower_barrier(conf);
3249 md_unregister_thread(&mddev->thread);
3250 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3251 if (conf->r10bio_pool)
3252 mempool_destroy(conf->r10bio_pool);
3253 kfree(conf->mirrors);
3255 mddev->private = NULL;
3259 static void raid10_quiesce(struct mddev *mddev, int state)
3261 struct r10conf *conf = mddev->private;
3265 raise_barrier(conf, 0);
3268 lower_barrier(conf);
3273 static void *raid10_takeover_raid0(struct mddev *mddev)
3275 struct md_rdev *rdev;
3276 struct r10conf *conf;
3278 if (mddev->degraded > 0) {
3279 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3281 return ERR_PTR(-EINVAL);
3284 /* Set new parameters */
3285 mddev->new_level = 10;
3286 /* new layout: far_copies = 1, near_copies = 2 */
3287 mddev->new_layout = (1<<8) + 2;
3288 mddev->new_chunk_sectors = mddev->chunk_sectors;
3289 mddev->delta_disks = mddev->raid_disks;
3290 mddev->raid_disks *= 2;
3291 /* make sure it will be not marked as dirty */
3292 mddev->recovery_cp = MaxSector;
3294 conf = setup_conf(mddev);
3295 if (!IS_ERR(conf)) {
3296 list_for_each_entry(rdev, &mddev->disks, same_set)
3297 if (rdev->raid_disk >= 0)
3298 rdev->new_raid_disk = rdev->raid_disk * 2;
3305 static void *raid10_takeover(struct mddev *mddev)
3307 struct r0conf *raid0_conf;
3309 /* raid10 can take over:
3310 * raid0 - providing it has only two drives
3312 if (mddev->level == 0) {
3313 /* for raid0 takeover only one zone is supported */
3314 raid0_conf = mddev->private;
3315 if (raid0_conf->nr_strip_zones > 1) {
3316 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3317 " with more than one zone.\n",
3319 return ERR_PTR(-EINVAL);
3321 return raid10_takeover_raid0(mddev);
3323 return ERR_PTR(-EINVAL);
3326 static struct md_personality raid10_personality =
3330 .owner = THIS_MODULE,
3331 .make_request = make_request,
3335 .error_handler = error,
3336 .hot_add_disk = raid10_add_disk,
3337 .hot_remove_disk= raid10_remove_disk,
3338 .spare_active = raid10_spare_active,
3339 .sync_request = sync_request,
3340 .quiesce = raid10_quiesce,
3341 .size = raid10_size,
3342 .takeover = raid10_takeover,
3345 static int __init raid_init(void)
3347 return register_md_personality(&raid10_personality);
3350 static void raid_exit(void)
3352 unregister_md_personality(&raid10_personality);
3355 module_init(raid_init);
3356 module_exit(raid_exit);
3357 MODULE_LICENSE("GPL");
3358 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3359 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3360 MODULE_ALIAS("md-raid10");
3361 MODULE_ALIAS("md-level-10");
3363 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);