Date:   Tue, 3 Apr 2018 18:51:38 -0700
From:   João Paulo Rechi Vita <jprvita@...il.com>
To:     Larry Finger <Larry.Finger@...inger.net>,
        Yan-Hsuan Chuang <yhchuang@...ltek.com>,
        Ping-Ke Shih <pkshih@...ltek.com>,
        Birming Chiu <birming@...ltek.com>,
        Shaofu <shaofu@...ltek.com>,
        Steven Ting <steventing@...ltek.com>,
        Chaoming Li <chaoming_li@...lsil.com.cn>,
        Kalle Valo <kvalo@...eaurora.org>
Cc:     linux-wireless <linux-wireless@...r.kernel.org>,
        Network Development <netdev@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Daniel Drake <drake@...lessm.com>,
        João Paulo Rechi Vita <jprvita@...lessm.com>,
        linux@...lessm.com
Subject: RTL8723BE performance regression

Hello,

I've been trying to track a performance regression on the RTL8723BE
WiFi adapter, which mainly affects the upload bandwidth (although we
can see a decreased download performance as well, the effect on upload
is more drastic). This was first reported by users after upgrading
from our 4.11-based kernel to our 4.13-based kernel, but also
confirmed to affect our development branch (4.15-based kernel) and
wireless-drivers-next at the
wireless-drivers-next-for-davem-2018-03-29 tag. This is happening on
an HP laptop that needs rtl8723be.ant_sel=1 (and all the following
tests have been made with that param).

My first bisect attempt pointed me to the following commit:

bcd37f4a0831 rtlwifi: btcoex: 23b 2ant: let bt transmit when hw
initialisation done

Which I later found to be already fixed by

a33fcba6ec01 rtlwifi: btcoexist: Fix breakage of ant_sel for rtl8723be.

That fix is already included in v4.15 though (and our dev branch as
well), so I did a second bisect, now cherry-picking a33fcba6ec01 at
every step, and it pointed me to the following commit:

7937f02d1953 rtlwifi: btcoex: hook external functions for newer chips

Reverting that commit on top of our development branch fixes the
problem, but on top of v4.15 I get mixed results: a few times getting
a good upload performance (~5-6Mbps) but most of the time just getting
~1-1.5Mpbs (which is still better than the 0.0 then test failure I've
gotten on most bad points of the bisect).

Bisecting the downstream patches we carry on top of v4.15 (we base our
kernel on Ubuntu's, so there are quite a few downstream changes) did
not bring any clarity, as at all bisect points (plus reverting
7937f02d1953) the performance was good, so probably there was some
other difference in the resulting kernels from my initial revert of
that patch on top of v4.15 and each step during the bisect. I've
experimented a bit with fwlps=0, but it did not bring any conclusive
results either. I'll try to look at other things that may have changed
(configuration perhaps?), but I don't have a clear plan yet.

Have you seen anything similar, or have any other ideas or suggestions
to track this problem? Even without crystal clear results, it looks
like 7937f02d1953 is having a negative impact on the RTL8723BE
performance, so perhaps it is worth reverting it and reworking it a
later point?

This are the results (testing with speedtest.net) I got at some key points:

Version        Commit        Ping        Down        Up

v4.11            a351e9b        12        25.44        5.99
v4.11            a351e9b        131      17.02        5.89

v4.13            569dbb8        174      14.08        0.00
v4.13            569dbb8        261      8.41          0.00

v4.15+revert d8a5b80        19        23.86        1.41
v4.15+revert d8a5b80        189      18.69        1.39

Best regards,

--
João Paulo Rechi Vita
http://about.me/jprvita

Powered by blists - more mailing lists