On Tue, Feb 16, 2016 at 4:06 AM, Sagi Grimberg <sagig@xxxxxxxxxxxxxxxxxx> wrote: > >> Cc Kent and Keith. >> >> Follows another version which should be more efficient. >> Kent and Keith, I appreciate much if you may give a review on it. >> >> diff --git a/include/linux/bio.h b/include/linux/bio.h >> index 56d2db8..ef45fec 100644 >> --- a/include/linux/bio.h >> +++ b/include/linux/bio.h >> @@ -278,11 +278,21 @@ static inline void bio_get_first_bvec(struct bio >> *bio, struct bio_vec *bv) >> */ >> static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec >> *bv) >> { >> - struct bvec_iter iter; >> + struct bvec_iter iter = bio->bi_iter; >> + int idx; >> + >> + bio_advance_iter(bio, &iter, iter.bi_size); >> + >> + WARN_ON(!iter.bi_idx && !iter.bi_bvec_done); >> + >> + if (!iter.bi_bvec_done) >> + idx = iter.bi_idx - 1; >> + else /* in the middle of bvec */ >> + idx = iter.bi_idx; >> >> - bio_for_each_segment(*bv, bio, iter) >> - if (bv->bv_len == iter.bi_size) >> - break; >> + *bv = bio->bi_io_vec[idx]; >> + if (iter.bi_bvec_done) >> + bv->bv_len = iter.bi_bvec_done; >> } >> >> /* >> > > This looks good too. > >> >>> >>> However, given that it's a regression bug fix I'm not sure it's the best >>> idea to add logic here. >> >> >> But the issue is obviously in bio_will_gap(), isn't it? >> >> Simply reverting 52cc6eead9095(block: blk-merge: fast-clone bio when >> splitting rw bios) >> still might cause performance regression too. > > > That's correct. I assume that the bio splitting code affects > specific I/O pattern (gappy), however bio_will_gap is also tested I don't understand why bio splitting affects specific I/O pattern, could you explain a bit? >From commit b54ffb73c(block: remove bio_get_nr_vecs()), the upper layer(fs, dm, dio,...) creates bio with its max size, and splitting should be triggered easily. > for bio merges (even if the bios won't merge eventually). This means As I mentioned, bio_will_gap() is only called for non-splitted bio. > that each merge check will invoke bio_advance_iter() which is something > I'd like to avoid... One idea is to use original way to compute the last bvec for non-cloned bio, and use the approach in this patch for cloned bio(often splitted bio). I will take this way in v1 if no one objects. thanks, Ming -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html