Re: [PATCH v2 net-next 3/4] secure_seq: use SipHash in place of MD5

From: Eric Biggers <redacted>
Date: Sat, 7 Jan 2017 14:09:11 -0800

Well, except those instructions aren't actually used in these
places.  Although x86_64 SHA1-NI accelerated SHA-1 is available in
the Linux crypto API, it seems that in kernel code it remains
impractical to use these instructions on small amounts of data
because they use XMM registers, which means the overhead of
kernel_fpu_begin()/kernel_fpu_end() must be incurred.  Furthermore,
kernel_fpu_begin() is not allowed in all contexts so there has to be
a fallback.

Out of curiosity, is this actually a solvable problem, e.g. by
making the code using the XMM registers responsible for saving and
restoring the ones clobbered, or by optimizing
kernel_fpu_begin()/kernel_fpu_end()?  Or does it in fact remain
impractical for such instructions to be used for applications like
this one?

On x86 making the FPU save more tractible in situations like this is
really hard and will make the code significantly more complex.

It's simpler and cheaper on sparc64, and unlike on x86 there aren't
any fundament restrictions on where FPU stuff can be used.  This is
because we don't have "save all the FPU state" instructions and have
to do it all by hand anyways.

However I will note that just like x86, sparc64 doesn't override the
md5_transform() in lib/md5.c like it should.