Re: [PATCH v1 4/10] lib/siphash.c: New file

["George Spelvin" <linux@horizon.com>]

On Mon, Sep 229 2014 at 12:12:43 -0700, Darrick J. Wong wrote:
> Could you please make this part of crypto/ so that anyone who wants to
> improve upon the C implementation (Google suggests that SSE/AVX ports
> are possible) can do so easily?

Well, *that* was a rabbit hole.  It seems like an obviously good idea,
but let's just say that crypto/ is non-obvious.  (No, it didn't take me 2
months of work; I just got sidetracked a lot because it was discouraging.)
But now that my cleanup patches there are getting reviewed, I can answer.

Basically, to fit into the crypto layer would require a very different
implementation with a lot more overhead.  The code I proposed is optimized
for both size and performance in the single-contiguous-small-buffer case
that apples to file names.

There's are no separate init/update/final calls, no saving internal
state to memory, no handling of discontiguous input buffers, etc.

This is all because SipHash is designed to be *extremely* lightweight,
so the overhead of marshalling the input bytes is noticeable.

I could easily write a *separate* implementation for crypto/, and
it could share source code, but it wouldn't be the same object code.


> This would also make it so that ext4 only loads the algorithm when necessary.

Yes, but my Cunning Plan is to replace the MD5 use in net/core_secure_seq.c
with this, too.  And, after careful consultation with Ted, the one in
get_random_int, too.

With all the simplifying assumptions I mentioed above made specifically in
order to get it down to negligible size, the code is 454 bytes long with
-O2, 397 bytes with -Os.  Is that worth the overhead of a separate module?