On Fri, 2016-06-24 at 18:51 -0700, Andy Lutomirski wrote:

Hi all-

tcp_md5_hash_header does crypto using an sg that points to the stack.
This will break with virtually mapped stacks.  It also looks like it's
probably much slower than it deserves to be (it's trying to compute
the MD5 hash of a few tens of bytes -- going through a scatterlist is
a lot of overhead for an otherwise very fast operation).

I guess nobody cares about TCP MD5 speed really.

I don't suppose one of you could fix it or at least advise as to how
it should be fixed.

Simply extend tcp_md5sig_pool to contain a copy of the TCP headers ?

At most 40 bytes of extra per cpu storage is not a big problem.