Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <e4dd26fa53ad2535f5a20be63a4b294d99306ffb.camel@sipsolutions.net>
Subject: Re: [PATCH v1] wifi: mac80211: Initialize EWMA fail avg to 1
From:   Johannes Berg <johannes@sipsolutions.net>
To:     Benjamin Beichler <Benjamin.Beichler@uni-rostock.de>,
        Karthik M <quic_karm@quicinc.com>
Cc:     linux-wireless@vger.kernel.org,
        Tamizh Chelvam Raja <quic_tamizhr@quicinc.com>
Date:   Thu, 20 Apr 2023 12:27:11 +0200
In-Reply-To: <50a76761-5be3-8ea7-c2f3-a14c158aa039@uni-rostock.de>
References: <20230417100258.22965-1-quic_karm@quicinc.com>
         <bc1903aa09391667262aeadf1859600579f0a9f1.camel@sipsolutions.net>
         <50a76761-5be3-8ea7-c2f3-a14c158aa039@uni-rostock.de>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.46.4 (3.46.4-1.fc37) 
MIME-Version: 1.0
Precedence: bulk

On Thu, 2023-04-20 at 11:30 +0200, Benjamin Beichler wrote:
> > To me, basically, I see two ways to solve this:
> >=20
> > 1) we have DECLARE_EWMA_ZERO_VALID() or something like that which
> >     *doesn't* treat 0 as an uninitialized value, and either has a
> >     separate "not initialized yet" bit (but that's iffy storage wise),
> >     or simply has another argument to _init() for the initial value or
> >     so.
> >=20
> > 2) you don't just don't use 0 and 100 but say 1 and 100, that results i=
n
> >     basically the same behaviour, but avoids the special 0.
> >=20
> > johannes
>=20
> I also ran into that problem in the past, and reviewing it again with a=
=20
> college, I think, this is a real bug in the EWMA implementation. I try=
=20
> to provide a proper patch in the next days, but actually the EWMA=20
> handles the internal value zero, always like in the initialization,=20
> which is wrong, e.g., for positive/negative averaged values.

Yes, it's always wrong as long as you feed it something zero, or values
with different sign.

For a lot of use cases, however, that doesn't matter. Originally, it was
used e.g. for signal strength averaging, average packet lengths, etc.
where it really doesn't matter since you can never use 0 or values that
have different sign.

> A quick research shows, this bug is since the first implementation of=20
> the ewma in the code ...
>=20

Yeah, I'm aware of that, I was around for it ;-)

But see above, I'm not sure I'd even call it a bug, at least not
originally with the users that we had intended.

Hence I don't know if it's really good to fix this in general - for many
of these cases zero can still be treated specially (and like I mentioned
in my previous email, we can even here avoid 0), and then we don't spend
an extra byte (or likely 4) to hold a "first time" flag.

Dunno. Maybe it's not worth thinking about the extra memory space vs.
the extra maintenance cost. But maybe at least on 64-bit we could steal
a bit from the unsigned long? Not sure what all the users are...

johannes