Hello, Kent. On Wed, Aug 22, 2012 at 10:04:08AM -0700, Kent Overstreet wrote: > This changes bio_pair_split() to use the new bio_split() underneath, > which gets rid of the single page bio limitation. The various callers > are fixed up for the slightly different struct bio_pair, and to remove > the unnecessary checks. This changes an existing API both in its interface and behavior and there's no detailed explanation on how it's changed and what are the implications. > diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c > index 1f5b483..63e5852 100644 > --- a/drivers/block/rbd.c > +++ b/drivers/block/rbd.c > @@ -751,14 +751,13 @@ static struct bio *bio_chain_clone(struct bio **old, struct bio **next, > > /* split the bio. We'll release it either in the next > call, or it will have to be released outside */ > - bp = bio_pair_split(old_chain, > - (len - total) / SECTOR_SIZE); > + bp = bio_pair_split(old_chain, (len - total) / SECTOR_SIZE); Probably belongs to the previous patch which renamed bio_split() to bio_pair_split()? Another thing is, is this renaming really necessary? If you did, * s/bio_split()/bio_pair_split(). * introduce better and prettier bio_split() which has different semantics. * replace bio_pair_split() users with bio_split(). the renaming would have made sense, but you renamed an existing API, intrudced a new one and then changed the renamed old API. Doesn't make too much sense to me. > diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c > index 0f31ec4..9fa07c7 100644 > --- a/drivers/md/raid10.c > +++ b/drivers/md/raid10.c > @@ -1080,15 +1080,9 @@ static void make_request(struct mddev *mddev, struct bio * bio) > && (conf->geo.near_copies < conf->geo.raid_disks > || conf->prev.near_copies < conf->prev.raid_disks))) { > struct bio_pair *bp; > - /* Sanity check -- queue functions should prevent this happening */ > - if (bio->bi_vcnt != 1 || > - bio->bi_idx != 0) > - goto bad_map; > - /* This is a one page bio that upper layers > - * refuse to split for us, so we need to split it. > - */ > + > bp = bio_pair_split(bio, > - chunk_sects - (bio->bi_sector & (chunk_sects - 1)) ); > + chunk_sects - (bio->bi_sector & (chunk_sects - 1))); I suppose this one too belongs to the previous rename patch? > --- a/fs/bio-integrity.c > +++ b/fs/bio-integrity.c > @@ -681,50 +681,6 @@ void bio_integrity_trim(struct bio *bio, unsigned int offset, > EXPORT_SYMBOL(bio_integrity_trim); > > /** > - * bio_integrity_split - Split integrity metadata > - * @bio: Protected bio > - * @bp: Resulting bio_pair > - * @sectors: Offset > - * > - * Description: Splits an integrity page into a bio_pair. > - */ > -void bio_integrity_split(struct bio *bio, struct bio_pair *bp, int sectors) > -{ > - struct blk_integrity *bi; > - struct bio_integrity_payload *bip = bio->bi_integrity; > - unsigned int nr_sectors; > - > - if (bio_integrity(bio) == 0) > - return; > - > - bi = bdev_get_integrity(bio->bi_bdev); > - BUG_ON(bi == NULL); > - BUG_ON(bip->bip_vcnt != 1); > - > - nr_sectors = bio_integrity_hw_sectors(bi, sectors); > - > - bp->bio1.bi_integrity = &bp->bip1; > - bp->bio2.bi_integrity = &bp->bip2; > - > - bp->iv1 = bip->bip_vec[0]; > - bp->iv2 = bip->bip_vec[0]; > - > - bp->bip1.bip_vec[0] = bp->iv1; > - bp->bip2.bip_vec[0] = bp->iv2; > - > - bp->iv1.bv_len = sectors * bi->tuple_size; > - bp->iv2.bv_offset += sectors * bi->tuple_size; > - bp->iv2.bv_len -= sectors * bi->tuple_size; > - > - bp->bip1.bip_sector = bio->bi_integrity->bip_sector; > - bp->bip2.bip_sector = bio->bi_integrity->bip_sector + nr_sectors; > - > - bp->bip1.bip_vcnt = bp->bip2.bip_vcnt = 1; > - bp->bip1.bip_idx = bp->bip2.bip_idx = 0; > -} > -EXPORT_SYMBOL(bio_integrity_split); I complained about this in the last posting and in the previous patch. Please respond. Martin, are you okay with these integrity changes? > -static void bio_pair_end_1(struct bio *bi, int err) > +static void bio_pair_end(struct bio *bio, int error) > { > - struct bio_pair *bp = container_of(bi, struct bio_pair, bio1); > + struct bio_pair *bp = bio->bi_private; > > - if (err) > - bp->error = err; > - > - bio_pair_release(bp); > -} > - > -static void bio_pair_end_2(struct bio *bi, int err) > -{ > - struct bio_pair *bp = container_of(bi, struct bio_pair, bio2); > - > - if (err) > - bp->error = err; > + if (error) > + clear_bit(BIO_UPTODATE, &bp->orig->bi_flags); > > bio_pair_release(bp); > } Why is losing error value okay here? > @@ -1856,8 +1846,7 @@ static int __init init_bio(void) > if (bioset_integrity_create(fs_bio_set, BIO_POOL_SIZE)) > panic("bio: can't create integrity pool\n"); > > - bio_split_pool = mempool_create_kmalloc_pool(BIO_SPLIT_ENTRIES, > - sizeof(struct bio_pair)); > + bio_split_pool = bioset_create(BIO_POOL_SIZE, offsetof(struct bio_pair, split)); > if (!bio_split_pool) > panic("bio: can't create split pool\n"); It would be nice to mention that using this from stacking drivers is inherently broken. This is something which has been broken before this patch but still. bio_split*() should always require separate biosets. > diff --git a/include/linux/bio.h b/include/linux/bio.h > index 1c3bb47..3ad3540 100644 > --- a/include/linux/bio.h > +++ b/include/linux/bio.h > @@ -192,14 +192,13 @@ struct bio_integrity_payload { > * in bio2.bi_private > */ > struct bio_pair { > - struct bio bio1, bio2; > - struct bio_vec bv1, bv2; > -#if defined(CONFIG_BLK_DEV_INTEGRITY) > - struct bio_integrity_payload bip1, bip2; > - struct bio_vec iv1, iv2; > -#endif > - atomic_t cnt; > - int error; > + atomic_t cnt; > + > + bio_end_io_t *bi_end_io; > + void *bi_private; > + > + struct bio *orig; > + struct bio split; > }; So, this is struct is allocated as frontpad, which is a pretty unusual thing to do. Please explain and emphasize that ->split should come last. Also, given that it's a pair split, it would be nice to somehow indicate that ->split is the earlier half. Before this change it was pretty clear with ->bio1/2. Thanks. -- tejun -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel