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/seq_file.h>
25 #include <linux/ratelimit.h>
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
48 * near_copies and far_copies must be at least one, and their product is at most
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
57 * Number of guaranteed r10bios in case of extreme VM load:
59 #define NR_RAID10_BIOS 256
61 static void allow_barrier(conf_t *conf);
62 static void lower_barrier(conf_t *conf);
64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
73 static void r10bio_pool_free(void *r10_bio, void *data)
78 /* Maximum size of each resync request */
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81 /* amount of memory to reserve for resync requests */
82 #define RESYNC_WINDOW (1024*1024)
83 /* maximum number of concurrent requests, memory permitting */
84 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
93 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
109 nalloc = 2; /* recovery */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
118 r10_bio->devs[j].bio = bio;
121 * Allocate RESYNC_PAGES data pages and attach them
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
134 page = alloc_page(gfp_flags);
138 bio->bi_io_vec[i].bv_page = page;
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
158 static void r10buf_pool_free(void *__r10_bio, void *data)
162 r10bio_t *r10bio = __r10_bio;
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
175 r10bio_pool_free(r10bio, conf);
178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
190 static void free_r10bio(r10bio_t *r10_bio)
192 conf_t *conf = r10_bio->mddev->private;
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
198 static void put_buf(r10bio_t *r10_bio)
200 conf_t *conf = r10_bio->mddev->private;
202 mempool_free(r10_bio, conf->r10buf_pool);
207 static void reschedule_retry(r10bio_t *r10_bio)
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
216 spin_unlock_irqrestore(&conf->device_lock, flags);
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
221 md_wakeup_thread(mddev->thread);
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
229 static void raid_end_bio_io(r10bio_t *r10_bio)
231 struct bio *bio = r10_bio->master_bio;
233 conf_t *conf = r10_bio->mddev->private;
235 if (bio->bi_phys_segments) {
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
248 * Wake up any possible resync thread that waits for the device
253 free_r10bio(r10_bio);
257 * Update disk head position estimator based on IRQ completion info.
259 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
261 conf_t *conf = r10_bio->mddev->private;
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
268 * Find the disk number which triggered given bio
270 static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
284 return r10_bio->devs[slot].devnum;
287 static void raid10_end_read_request(struct bio *bio, int error)
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
292 conf_t *conf = r10_bio->mddev->private;
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
298 * this branch is our 'one mirror IO has finished' event handler:
300 update_head_pos(slot, r10_bio);
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
317 * oops, read error - keep the refcount on the rdev
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
330 static void close_write(r10bio_t *r10_bio)
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
337 md_write_end(r10_bio->mddev);
340 static void one_write_done(r10bio_t *r10_bio)
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
350 raid_end_bio_io(r10_bio);
355 static void raid10_end_write_request(struct bio *bio, int error)
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
361 conf_t *conf = r10_bio->mddev->private;
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
367 * this branch is our 'one mirror IO has finished' event handler:
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
392 &first_bad, &bad_sectors)) {
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
402 * Let's see if all mirrored write operations have finished
405 one_write_done(r10_bio);
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
436 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
450 chunk *= conf->near_copies;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
456 sector += stripe << conf->chunk_shift;
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
476 if (dev >= conf->raid_disks) {
478 sector += (conf->chunk_mask + 1);
481 BUG_ON(slot != conf->copies);
484 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
486 sector_t offset, chunk, vchunk;
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
495 dev += conf->raid_disks;
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
502 dev -= conf->near_copies;
504 chunk = sector >> conf->chunk_shift;
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
521 static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
551 * The rdev for the device selected will have nr_pending incremented.
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
558 static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
560 const sector_t this_sector = r10_bio->sector;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 struct md_rdev *rdev;
569 raid10_find_phys(conf, r10_bio);
572 sectors = r10_bio->sectors;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
587 for (slot = 0; slot < conf->copies ; slot++) {
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
598 if (!test_bit(In_sync, &rdev->flags))
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
625 /* Must read from here */
630 best_good_sectors = sectors;
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
653 if (slot == conf->copies)
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
666 rdev_dec_pending(rdev, conf->mddev);
669 r10_bio->read_slot = slot;
673 *max_sectors = best_good_sectors;
678 static int raid10_congested(void *data, int bits)
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
684 if (mddev_congested(mddev, bits))
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
692 ret |= bdi_congested(&q->backing_dev_info, bits);
699 static void flush_pending_writes(conf_t *conf)
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
704 spin_lock_irq(&conf->device_lock);
706 if (conf->pending_bio_list.head) {
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
717 generic_make_request(bio);
721 spin_unlock_irq(&conf->device_lock);
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
746 static void raise_barrier(conf_t *conf, int force)
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
755 /* block any new IO from starting */
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
763 spin_unlock_irq(&conf->resync_lock);
766 static void lower_barrier(conf_t *conf)
769 spin_lock_irqsave(&conf->resync_lock, flags);
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
775 static void wait_barrier(conf_t *conf)
777 spin_lock_irq(&conf->resync_lock);
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
786 spin_unlock_irq(&conf->resync_lock);
789 static void allow_barrier(conf_t *conf)
792 spin_lock_irqsave(&conf->resync_lock, flags);
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
798 static void freeze_array(conf_t *conf)
800 /* stop syncio and normal IO and wait for everything to
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
812 spin_lock_irq(&conf->resync_lock);
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
818 flush_pending_writes(conf));
820 spin_unlock_irq(&conf->resync_lock);
823 static void unfreeze_array(conf_t *conf)
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
833 static int make_request(mddev_t *mddev, struct bio * bio)
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
838 struct bio *read_bio;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
845 struct md_rdev *blocked_rdev;
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
860 conf->near_copies < conf->raid_disks)) {
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
880 spin_lock_irq(&conf->resync_lock);
882 spin_unlock_irq(&conf->resync_lock);
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
889 spin_lock_irq(&conf->resync_lock);
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
894 bio_pair_release(bp);
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
905 md_write_start(mddev, bio);
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
935 * read balancing logic:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
944 raid_end_bio_io(r10_bio);
947 mirror = conf->mirrors + disk;
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
953 r10_bio->devs[slot].bio = read_bio;
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
966 sectors_handled = (r10_bio->sectors + max_sectors
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
980 reschedule_retry(r10_bio);
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
992 generic_make_request(read_bio);
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1010 plugged = mddev_check_plugged(mddev);
1012 raid10_find_phys(conf, r10_bio);
1014 blocked_rdev = NULL;
1016 max_sectors = r10_bio->sectors;
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1037 is_bad = is_badblock(rdev, dev_sector,
1039 &first_bad, &bad_sectors);
1041 /* Mustn't write here until the bad block
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1111 for (i = 0; i < conf->copies; i++) {
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1120 r10_bio->devs[i].bio = mbio;
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1154 one_write_done(r10_bio);
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1164 static void status(struct seq_file *seq, mddev_t *mddev)
1166 conf_t *conf = mddev->private;
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1188 /* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1193 static int enough(conf_t *conf, int ignore)
1198 int n = conf->copies;
1201 if (conf->mirrors[first].rdev &&
1204 first = (first+1) % conf->raid_disks;
1208 } while (first != 0);
1212 static void error(mddev_t *mddev, struct md_rdev *rdev)
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1226 * Don't fail the drive, just return an IO error.
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1235 * if recovery is running, make sure it aborts.
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1249 static void print_conf(conf_t *conf)
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1256 printk(KERN_DEBUG "(!conf)\n");
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1273 static void close_sync(conf_t *conf)
1276 allow_barrier(conf);
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1282 static int raid10_spare_active(mddev_t *mddev)
1285 conf_t *conf = mddev->private;
1288 unsigned long flags;
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1312 static int raid10_add_disk(mddev_t *mddev, struct md_rdev *rdev)
1314 conf_t *conf = mddev->private;
1318 int last = conf->raid_disks - 1;
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1325 if (!enough(conf, -1))
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1360 if (rdev->saved_raid_disk != mirror)
1362 rcu_assign_pointer(p->rdev, rdev);
1366 md_integrity_add_rdev(rdev, mddev);
1371 static int raid10_remove_disk(mddev_t *mddev, int number)
1373 conf_t *conf = mddev->private;
1375 struct md_rdev *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1386 /* Only remove faulty devices in recovery
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1403 err = md_integrity_register(mddev);
1412 static void end_sync_read(struct bio *bio, int error)
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1438 reschedule_retry(r10_bio);
1442 static void end_sync_request(r10bio_t *r10_bio)
1444 mddev_t *mddev = r10_bio->mddev;
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1455 md_done_sync(mddev, s, 1);
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1469 static void end_sync_write(struct bio *bio, int error)
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1493 end_sync_request(r10_bio);
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1509 * We check if all blocks are in-sync and only write to blocks that
1512 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1514 conf_t *conf = mddev->private;
1516 struct bio *tbio, *fbio;
1518 atomic_set(&r10_bio->remaining, 1);
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1525 if (i == conf->copies)
1529 fbio = r10_bio->devs[i].bio;
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1536 tbio = r10_bio->devs[i].bio;
1538 if (tbio->bi_end_io != end_sync_read)
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1583 tbio->bi_end_io = end_sync_write;
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1612 static void fix_recovery_read_error(r10bio_t *r10_bio)
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1625 int sectors = r10_bio->sectors;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1632 struct md_rdev *rdev;
1636 if (s > (PAGE_SIZE>>9))
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1644 bio->bi_io_vec[idx].bv_page,
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1652 bio->bi_io_vec[idx].bv_page,
1655 set_bit(WriteErrorSeen, &rdev->flags);
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1662 rdev_set_badblocks(rdev, addr, s, 0);
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1670 /* just abort the recovery */
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1691 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1693 conf_t *conf = mddev->private;
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1704 * share the pages with the first bio
1705 * and submit the write request
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1722 static void check_decay_read_errors(mddev_t *mddev, struct md_rdev *rdev)
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1728 ktime_get_ts(&cur_time_mon);
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1740 rdev->last_read_error = cur_time_mon;
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1753 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1774 * This is a kernel thread which:
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1781 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 struct md_rdev*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1792 rdev = conf->mirrors[d].rdev;
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1809 atomic_read(&rdev->read_errors), max_read_errors);
1811 "md/raid10:%s: %s: Failing raid device\n",
1813 md_error(mddev, conf->mirrors[d].rdev);
1819 int sl = r10_bio->read_slot;
1823 if (s > (PAGE_SIZE>>9))
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1850 if (sl == conf->copies)
1852 } while (!success && sl != r10_bio->read_slot);
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1863 if (!rdev_set_badblocks(
1865 r10_bio->devs[r10_bio->read_slot].addr
1868 md_error(mddev, rdev);
1873 /* write it back and re-read */
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1884 !test_bit(In_sync, &rdev->flags))
1887 atomic_inc(&rdev->nr_pending);
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1892 s<<9, conf->tmppage, WRITE)
1894 /* Well, this device is dead */
1896 "md/raid10:%s: read correction "
1898 " (%d sectors at %llu on %s)\n",
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1906 bdevname(rdev->bdev, b));
1908 rdev_dec_pending(rdev, mddev);
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1921 !test_bit(In_sync, &rdev->flags))
1924 atomic_inc(&rdev->nr_pending);
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1929 s<<9, conf->tmppage,
1932 /* Well, this device is dead */
1934 "md/raid10:%s: unable to read back "
1936 " (%d sectors at %llu on %s)\n",
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1944 bdevname(rdev->bdev, b));
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1957 rdev_dec_pending(rdev, mddev);
1967 static void bi_complete(struct bio *bio, int error)
1969 complete((struct completion *)bio->bi_private);
1972 static int submit_bio_wait(int rw, struct bio *bio)
1974 struct completion event;
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1986 static int narrow_write_error(r10bio_t *r10_bio, int i)
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
2000 * We currently own a reference to the rdev.
2006 int sect_to_write = r10_bio->sectors;
2009 if (rdev->badblocks.shift < 0)
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2018 while (sect_to_write) {
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2031 ok = rdev_set_badblocks(rdev, sector,
2036 sect_to_write -= sectors;
2038 sectors = block_sectors;
2043 static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2048 conf_t *conf = mddev->private;
2049 struct md_rdev *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2062 if (mddev->ro == 0) {
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2100 r10_bio->sector - bio->bi_sector,
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2131 set_bit(R10BIO_ReadError,
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2139 generic_make_request(bio);
2142 static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2151 struct md_rdev *rdev;
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2164 r10_bio->devs[m].addr,
2167 if (!rdev_set_badblocks(
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2183 r10_bio->devs[m].addr,
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2193 rdev_dec_pending(rdev, conf->mddev);
2196 if (test_bit(R10BIO_WriteError,
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2203 static void raid10d(mddev_t *mddev)
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2211 md_check_recovery(mddev);
2213 blk_start_plug(&plug);
2216 flush_pending_writes(conf);
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2240 /* just a partial read to be scheduled from a
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2251 blk_finish_plug(&plug);
2255 static int init_resync(conf_t *conf)
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2264 conf->next_resync = 0;
2269 * perform a "sync" on one "block"
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2300 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2303 conf_t *conf = mddev->private;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2335 else for (i=0; i<conf->raid_disks; i++) {
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2341 } else /* completed sync */
2344 bitmap_close_sync(mddev->bitmap);
2347 return sectors_skipped;
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2354 return (max_sector - sector_nr) + sectors_skipped;
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2360 /* make sure whole request will fit in a chunk - if chunks
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2394 for (i=0 ; i<conf->raid_disks; i++) {
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2406 /* want to reconstruct this device */
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2421 chunks_skipped = -1;
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2429 r10_bio->master_bio = (struct bio*)rb2;
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2436 raid10_find_phys(conf, r10_bio);
2438 /* Need to check if the array will still be
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2452 for (j=0; j<conf->copies;j++) {
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 struct md_rdev *rdev;
2457 sector_t sector, first_bad;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2462 /* This is where we read from */
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2472 bad_sectors -= (sector
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2520 atomic_dec(&rb2->remaining);
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2548 if (biolist == NULL) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2558 /* resync. Schedule a read for every block at this virt offset */
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2569 return sync_blocks + sectors_skipped;
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2643 bio->bi_phys_segments = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2679 r10_bio->sectors = nr_sectors;
2683 biolist = biolist->bi_next;
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2699 md_done_sync(mddev, sectors_skipped, 1);
2701 return sectors_skipped + nr_sectors;
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2710 sectors_skipped += (max_sector - sector_nr);
2712 sector_nr = max_sector;
2717 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2720 conf_t *conf = mddev->private;
2723 raid_disks = conf->raid_disks;
2725 sectors = conf->dev_sectors;
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2732 return size << conf->chunk_shift;
2736 static conf_t *setup_conf(mddev_t *mddev)
2738 conf_t *conf = NULL;
2740 sector_t stride, size;
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2804 conf->dev_sectors = stride << conf->chunk_shift;
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2823 conf->mddev = mddev;
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2836 return ERR_PTR(err);
2839 static int run(mddev_t *mddev)
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 struct md_rdev *rdev;
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2859 conf = mddev->private;
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2880 disk = conf->mirrors + disk_idx;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2895 disk->head_position = 0;
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2907 disk = conf->mirrors + i;
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2927 * Ok, everything is just fine now
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2952 if (md_integrity_register(mddev))
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2964 mddev->private = NULL;
2969 static int stop(mddev_t *mddev)
2971 conf_t *conf = mddev->private;
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2982 mddev->private = NULL;
2986 static void raid10_quiesce(mddev_t *mddev, int state)
2988 conf_t *conf = mddev->private;
2992 raise_barrier(conf, 0);
2995 lower_barrier(conf);
3000 static void *raid10_takeover_raid0(mddev_t *mddev)
3002 struct md_rdev *rdev;
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3008 return ERR_PTR(-EINVAL);
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3032 static void *raid10_takeover(mddev_t *mddev)
3034 struct raid0_private_data *raid0_priv;
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3046 return ERR_PTR(-EINVAL);
3048 return raid10_takeover_raid0(mddev);
3050 return ERR_PTR(-EINVAL);
3053 static struct mdk_personality raid10_personality =
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3072 static int __init raid_init(void)
3074 return register_md_personality(&raid10_personality);
3077 static void raid_exit(void)
3079 unregister_md_personality(&raid10_personality);
3082 module_init(raid_init);
3083 module_exit(raid_exit);
3084 MODULE_LICENSE("GPL");
3085 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087 MODULE_ALIAS("md-raid10");
3088 MODULE_ALIAS("md-level-10");