Re: RTL8723BE performance regression

[JoÃo Paulo Rechi Vita]

On Tue, May 8, 2018 at 1:37 AM, Pkshih <pkshih@xxxxxxxxxxx> wrote:
> On Mon, 2018-05-07 at 14:49 -0700, JoÃo Paulo Rechi Vita wrote:
>> On Tue, May 1, 2018 at 10:58 PM, Pkshih <pkshih@xxxxxxxxxxx> wrote:
>> > On Wed, 2018-05-02 at 05:44 +0000, Pkshih wrote:
>> >>
>> >> > -----Original Message-----
>> >> > From: JoÃo Paulo Rechi Vita [mailto:jprvita@xxxxxxxxx]
>> >> > Sent: Wednesday, May 02, 2018 6:41 AM
>> >> > To: Larry Finger
>> >> > Cc: Steve deRosier; èåå; Pkshih; Birming Chiu; Shaofu; Steven Ting; Chaoming_Li; Kalle Valo;
>> >> > linux-wireless; Network Development; LKML; Daniel Drake; JoÃo Paulo Rechi Vita; linux@endless
>> m.c
>> >> om
>> >> > Subject: Re: RTL8723BE performance regression
>> >> >
>> >> > On Tue, Apr 3, 2018 at 7:51 PM, Larry Finger <Larry.Finger@xxxxxxxxxxxx> wrote:
>> >> > > On 04/03/2018 09:37 PM, JoÃo Paulo Rechi Vita wrote:
>> >> > >>
>> >> > >> On Tue, Apr 3, 2018 at 7:28 PM, Larry Finger <Larry.Finger@xxxxxxxxxxxx>
>> >> > >> wrote:
>> >> > >>
>> >> > >> (...)
>> >> > >>
>> >> > >>> As the antenna selection code changes affected your first bisection, do
>> >> > >>> you
>> >> > >>> have one of those HP laptops with only one antenna and the incorrect
>> >> > >>> coding
>> >> > >>> in the FUSE?
>> >> > >>
>> >> > >>
>> >> > >> Yes, that is why I've been passing ant_sel=1 during my tests -- this
>> >> > >> was needed to achieve a good performance in the past, before this
>> >> > >> regression. I've also opened the laptop chassis and confirmed the
>> >> > >> antenna cable is plugged to the connector labeled with "1" on the
>> >> > >> card.
>> >> > >>
>> >> > >>> If so, please make sure that you still have the same signal
>> >> > >>> strength for good and bad cases. I have tried to keep the driver and the
>> >> > >>> btcoex code in sync, but there may be some combinations of antenna
>> >> > >>> configuration and FUSE contents that cause the code to fail.
>> >> > >>>
>> >> > >>
>> >> > >> What is the recommended way to monitor the signal strength?
>> >> > >
>> >> > >
>> >> > > The btcoex code is developed for multiple platforms by a different group
>> >> > > than the Linux driver. I think they made a change that caused ant_sel to
>> >> > > switch from 1 to 2. At least numerous comments at
>> >> > > github.com/lwfinger/rtlwifi_new claimed they needed to make that change.
>> >> > >
>> >> > > Mhy recommended method is to verify the wifi device name with "iw dev". Then
>> >> > > using that device
>> >> > >
>> >> > > sudo iw dev <dev_name> scan | egrep "SSID|signal"
>> >> > >
>> >> >
>> >> > I have confirmed that the performance regression is indeed tied to
>> >> > signal strength: on the good cases signal was between -16 and -8 dBm,
>> >> > whereas in bad cases signal was always between -50 to - 40 dBm. I've
>> >> > also switched to testing bandwidth in controlled LAN environment using
>> >> > iperf3, as suggested by Steve deRosier, with the DUT being the only
>> >> > machine connected to the 2.4 GHz radio and the machine running the
>> >> > iperf3 server connected via ethernet.
>> >> >
>> >>
>> >> We have new experimental results in commit af8a41cccf8f46 ("rtlwifi: cleanup
>> >> 8723be ant_sel definition"). You can use the above commit and do the same
>> >> experiments (with ant_sel=0, 1 and 2) in your side, and then share your results.
>> >> Since performance is tied to signal strength, you can only share signal strength.
>> >>
>> >
>> > Please pay attention to cold reboot once ant_sel is changed.
>> >
>>
>> I've tested the commit mentioned above and it fixes the problem on top
>> of v4.16 (in addition to the latest wireless-drivers-next also been
>> fixed as it already contains such commit). On v4.15, we also need the
>> following commits before "af8a41cccf8f rtlwifi: cleanup 8723be ant_sel
>> definition" to have a good performance again:
>>
>>   874e837d67d0 rtlwifi: fill FW version and subversion
>>   a44709bba70f rtlwifi: btcoex: Add power_on_setting routine
>>   40d9dd4f1c5d rtlwifi: btcoex: Remove global variables from btcoex
>
> v4.15 isn't longterm version and had been EOL.
>

Right, but this is a performace regression in comparison to v4.11, so
if "af8a41cccf8f rtlwifi: cleanup 8723be ant_sel definition" is marked
for stable, shouldn't these other patches be brought as well? All
releases since v4.11 are probably affected, but honestly I don't have
a strong understanding of how the stable trees operate in situations
like this.

>>
>> Surprisingly, it seems forcing ant_sel=1 is not needed anymore on
>> these machines, as the shown by the numbers bellow (ant_sel=0 means
>> that actually no parameter was passed to the module). I have powered
>> off the machine and done a cold boot for every test. It seems
>> something have changed in the antenna auto-selection code since v4.11,
>> the latest point where I could confirm we definitely need to force
>> ant_sel=1. I've been trying to understand what causes this difference,
>> but haven't made progress on that so far, so any suggestions are
>> appreciated (we are trying to decide if we can confidently drop the
>> downstream DMI quirks for these specific machines).
>>
> I think your rtl8723be module programed correct efuse content, so it
> works properly with ant_sel=0, and quirk isn't required for your
> machine.
>
>>   w-d-n ant_sel=0: -14.00 dBm,  69.5 Mbps -> good
>>   w-d-n ant_sel=1: -10.00 dBm,  41.1 Mbps -> good
>>   w-d-n ant_sel=2: -44.00 dBm,   607 kbps -> bad
>>
>>   v4.16 ant_sel=0: -12.00 dBm,  63.0 Mbps -> good
>>   v4.16 ant_sel=1: - 8.00 dBm,  69.0 Mbps -> good
>>   v4.16 ant_sel=2: -50.00 dBm,   224 kbps -> bad
>>
>>   v4.15 ant_sel=0: - 8.00 dBm,  33.0 Mbps -> good
>>   v4.15 ant_sel=1: -10.00 dBm,  38.1 Mbps -> good
>>   v4.15 ant_sel=2: -48.00 dBm,   206 kbps -> bad
>>
>
> With your results, the efuse content is programmed as one or two antenna
> on AUX path.
>

With v4.11 I had good performance results on this very same machine
(thus same efuse contents) only when passing ant_sel=1, so there has
to be some change on the code that parses the efuse contents and
decides which antenna will be used.

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