Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20211223141113.1240679-1-Jason@zx2c4.com> <20211223141113.1240679-2-Jason@zx2c4.com>
 <CAMuHMdU0spv9X_wErkBBWQ9kV9f1zE_YNcu5nPbTG_64Lh_h0w@mail.gmail.com>
 <CAHmME9pZu-UvCK=uP-sxXL127BmbjmrD2=M7cNd9vHdJEsverw@mail.gmail.com>
 <Yd18+iQ8zicsSPa0@zx2c4.com> <05ae373684334e6581294baa8afd3238@AcuMS.aculab.com>
In-Reply-To: <05ae373684334e6581294baa8afd3238@AcuMS.aculab.com>
From:   "Jason A. Donenfeld" <Jason@zx2c4.com>
Date:   Fri, 14 Jan 2022 18:31:42 +0100
Message-ID: <CAHmME9o4tyG2srPGe74n+cx9nfsU-xVe7C+mf4677ojVLVBWDg@mail.gmail.com>
Subject: Re: [PATCH v2 2/2] random: use BLAKE2s instead of SHA1 in extraction
To:     David Laight <David.Laight@aculab.com>
Cc:     Geert Uytterhoeven <geert@linux-m68k.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Theodore Tso <tytso@mit.edu>,
        Greg KH <gregkh@linuxfoundation.org>,
        Jean-Philippe Aumasson <jeanphilippe.aumasson@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Fri, Jan 14, 2022 at 6:27 PM David Laight <David.Laight@aculab.com> wrote:
>
> From: Jason A. Donenfeld
> > Sent: 11 January 2022 12:50
> >
> > On Tue, Jan 11, 2022 at 1:28 PM Jason A. Donenfeld <Jason@zx2c4.com> wrote:
> > > If you're really quite concerned about m68k code size, I can probably
> > > do some things to reduce that. For example, blake2s256_hmac is only
> > > used by wireguard and it could probably be made local there. And with
> > > some trivial loop re-rolling, I can shave off another 2300 bytes. And
> > > I bet I can find a few other things too. The question is: how
> > > important is this to you?
> >
> > And with another trick (see below), another extra 1000 bytes or so
> > shaved off. Aside from moving blake2s256_hmac, I'm not really super
> > enthusiastic about making these changes, but depending on how important
> > this is to you, maybe we can make something work. There are probably
> > additional possibilities too with the code.
>
> Quite clearly whoever wrote the unrolled loops needs their head examined.
> It is extremely unlikely that a cpu has enough registers to implement it
> effeciently.

Feel free to send a patch doing this, along with benchmarks. It
doesn't seem impossible to me that re-rolling the rounds might be
better on some platforms. The question is - is it really? And if so,
which ones? And for what varieties of inputs? If you put some research
into this, please do CC me on patches.

Jason