Re: [PATCH 3/4] infiniband/mm: convert to the new put_user_page() call

[Christoph Hellwig]

On Mon, Oct 01, 2018 at 08:29:29AM -0700, Matthew Wilcox wrote:
> I don't understand the dislike of the sg list.  Other than for special
> cases which we should't be optimising for (ramfs, brd, loopback
> filesystems), when we get a page to do I/O, we're going to want a dma
> mapping for them.  It makes sense to already allocate space to store
> the mapping at the outset.

We don't actually need the space - the scatterlist forces it on us,
otherwise we could translate directly in the on-disk format and
save that duplicate space.  I have prototypes for NVMe and RDMA that do
away with the scatterlist entirely.

And even if we are still using the scatterlist as we do right now we'd
need a second scatterlist at least for block / file system based I/O
as we can't plug the scatterlist into the I/O stack (nevermind that
due to splitting merging the lower one might not map 1:1 to the upper
one).

> [1] Can we ever admit that the bio_vec and the skb_frag_t are actually
> the same thing?

When I brought this up years ago the networking folks insisted that
their use of u16 offset/size fields was important for performance,
while for bio_vecs we needed the larger ones for some cases.  Since
then networking switched to 32-bit fields for what is now the fast
path, so it might be worth to give it another spin.

Than should also help with using my new bio_vec based dma-mapping
helpers to batch iommu mappings in networking, which Jesper had on
his todo list as all the indirect calls are causing performance
issues.