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: <74feca6f-4177-5645-6614-bc5f3dbf0be1@free.fr>
Date:   Thu, 6 Apr 2023 18:35:09 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.9.0
Subject: Re: [Performance regression] BCM4359/9 on S905X2
Content-Language: en-US
To:     Martin Blumenstingl <martin.blumenstingl@googlemail.com>
Cc:     Christian Hewitt <christian@hewittfamily.org.uk>,
        Linux Wireless <linux-wireless@vger.kernel.org>,
        MMC <linux-mmc@vger.kernel.org>,
        AML <linux-amlogic@lists.infradead.org>,
        Neil Armstrong <neil.armstrong@linaro.org>,
        Ulf Hansson <ulf.hansson@linaro.org>,
        Pierre-Hugues Husson <phh@phh.me>
References: <b9bd16cb-f1fa-34b7-d599-8637cbe5032b@free.fr>
 <1jh6ue74x9.fsf@starbuckisacylon.baylibre.com>
 <EEE9FD80-C106-4A9B-AA8D-5C151E540CFD@hewittfamily.org.uk>
 <0bb1e3b3-6823-ddb5-001b-72ee1b63779c@free.fr>
 <CAFBinCA+JOxCGzML-Mohryawrn6Vghd8Ns22=2ZfvWov43aeEg@mail.gmail.com>
From:   Marc Gonzalez <marc.w.gonzalez@free.fr>
In-Reply-To: <CAFBinCA+JOxCGzML-Mohryawrn6Vghd8Ns22=2ZfvWov43aeEg@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 04/04/2023 23:06, Martin Blumenstingl wrote:

> Hi Marc,

Hello Martin :)

> On Tue, Apr 4, 2023 at 6:09 PM Marc Gonzalez wrote:
> 
>>                 bus-width = <4>;
>>                 cap-sd-highspeed;
>>                 cap-mmc-highspeed;
>>                 max-frequency = <100000000>;
>
> I would start by comparing the bus mode. You can get it from
> /sys/kernel/debug/mmc2/ios
> On the vendor kernel it should be in /sys/kernel/debug/sdio/ios (or
> similar, I'm typing this from the top of my head).
> It will give you insights on the clock and timing that has been
> negotiated between the host and card.
> 
> From this information you can get the maximum bus speed, e.g. from [0]
> Please note that any card will add overhead for communication, so bus
> speed will not be equal to wifi throughput.

For the vendor kernel:

/sys/kernel/debug/sdio/sdio:0001/state:0x00000001
/sys/kernel/debug/sdio/clock:200000000
/sys/kernel/debug/sdio/ios:clock:               200000000 Hz
/sys/kernel/debug/sdio/ios:actual clock:        199999997 Hz
/sys/kernel/debug/sdio/ios:vdd:         21 (3.3 ~ 3.4 V)
/sys/kernel/debug/sdio/ios:bus mode:    2 (push-pull)
/sys/kernel/debug/sdio/ios:chip select: 0 (don't care)
/sys/kernel/debug/sdio/ios:power mode:  1 (up)
/sys/kernel/debug/sdio/ios:bus width:   2 (4 bits)
/sys/kernel/debug/sdio/ios:timing spec: 6 (sd uhs SDR104)
/sys/kernel/debug/sdio/ios:signal voltage:      1 (1.80 V)
/sys/kernel/debug/sdio/ios:driver type: 0 (driver type B)


For mainline:

/sys/kernel/debug/mmc2/mmc2:0001/state:0x00000001
/sys/kernel/debug/mmc2/err_stats:# Command Timeout Occurred:     0
/sys/kernel/debug/mmc2/err_stats:# Command CRC Errors Occurred:  0
/sys/kernel/debug/mmc2/err_stats:# Data Timeout Occurred:        0
/sys/kernel/debug/mmc2/err_stats:# Data CRC Errors Occurred:     0
/sys/kernel/debug/mmc2/err_stats:# Auto-Cmd Error Occurred:      0
/sys/kernel/debug/mmc2/err_stats:# ADMA Error Occurred:  0
/sys/kernel/debug/mmc2/err_stats:# Tuning Error Occurred:        0
/sys/kernel/debug/mmc2/err_stats:# CMDQ RED Errors:      0
/sys/kernel/debug/mmc2/err_stats:# CMDQ GCE Errors:      0
/sys/kernel/debug/mmc2/err_stats:# CMDQ ICCE Errors:     0
/sys/kernel/debug/mmc2/err_stats:# Request Timedout:     0
/sys/kernel/debug/mmc2/err_stats:# CMDQ Request Timedout:        0
/sys/kernel/debug/mmc2/err_stats:# ICE Config Errors:    0
/sys/kernel/debug/mmc2/err_stats:# Controller Timedout errors:   0
/sys/kernel/debug/mmc2/err_stats:# Unexpected IRQ errors:        0
/sys/kernel/debug/mmc2/err_state:0
/sys/kernel/debug/mmc2/clock:100000000
/sys/kernel/debug/mmc2/caps2:0x00040000
/sys/kernel/debug/mmc2/caps:0x40040105
/sys/kernel/debug/mmc2/ios:clock:               100000000 Hz
/sys/kernel/debug/mmc2/ios:actual clock:        99999999 Hz
/sys/kernel/debug/mmc2/ios:vdd:         21 (3.3 ~ 3.4 V)
/sys/kernel/debug/mmc2/ios:bus mode:    2 (push-pull)
/sys/kernel/debug/mmc2/ios:chip select: 0 (don't care)
/sys/kernel/debug/mmc2/ios:power mode:  2 (on)
/sys/kernel/debug/mmc2/ios:bus width:   2 (4 bits)
/sys/kernel/debug/mmc2/ios:timing spec: 5 (sd uhs SDR50)
/sys/kernel/debug/mmc2/ios:signal voltage:      1 (1.80 V)
/sys/kernel/debug/mmc2/ios:driver type: 0 (driver type B)


The clock is indeed running twice as fast on the vendor system.
And in SDR104 vs SDR50 mode.

Adjusting mainline device tree...


/sys/kernel/debug/mmc2/mmc2:0001/state:0x00000001
/sys/kernel/debug/mmc2/err_stats:# Command Timeout Occurred:     0
/sys/kernel/debug/mmc2/err_stats:# Command CRC Errors Occurred:  0
/sys/kernel/debug/mmc2/err_stats:# Data Timeout Occurred:        0
/sys/kernel/debug/mmc2/err_stats:# Data CRC Errors Occurred:     0
/sys/kernel/debug/mmc2/err_stats:# Auto-Cmd Error Occurred:      0
/sys/kernel/debug/mmc2/err_stats:# ADMA Error Occurred:  0
/sys/kernel/debug/mmc2/err_stats:# Tuning Error Occurred:        0
/sys/kernel/debug/mmc2/err_stats:# CMDQ RED Errors:      0
/sys/kernel/debug/mmc2/err_stats:# CMDQ GCE Errors:      0
/sys/kernel/debug/mmc2/err_stats:# CMDQ ICCE Errors:     0
/sys/kernel/debug/mmc2/err_stats:# Request Timedout:     0
/sys/kernel/debug/mmc2/err_stats:# CMDQ Request Timedout:        0
/sys/kernel/debug/mmc2/err_stats:# ICE Config Errors:    0
/sys/kernel/debug/mmc2/err_stats:# Controller Timedout errors:   0
/sys/kernel/debug/mmc2/err_stats:# Unexpected IRQ errors:        0
/sys/kernel/debug/mmc2/err_state:0
/sys/kernel/debug/mmc2/clock:200000000
/sys/kernel/debug/mmc2/caps2:0x00040000
/sys/kernel/debug/mmc2/caps:0x40080105
/sys/kernel/debug/mmc2/ios:clock:               200000000 Hz
/sys/kernel/debug/mmc2/ios:actual clock:        199999997 Hz
/sys/kernel/debug/mmc2/ios:vdd:         21 (3.3 ~ 3.4 V)
/sys/kernel/debug/mmc2/ios:bus mode:    2 (push-pull)
/sys/kernel/debug/mmc2/ios:chip select: 0 (don't care)
/sys/kernel/debug/mmc2/ios:power mode:  2 (on)
/sys/kernel/debug/mmc2/ios:bus width:   2 (4 bits)
/sys/kernel/debug/mmc2/ios:timing spec: 6 (sd uhs SDR104)
/sys/kernel/debug/mmc2/ios:signal voltage:      1 (1.80 V)
/sys/kernel/debug/mmc2/ios:driver type: 0 (driver type B)


# curl -o /dev/null http://192.168.1.254:8095/fixed/1G
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 1024M  100 1024M    0     0  9747k      0  0:01:47  0:01:47 --:--:-- 9544k

With this new setting, WiFi throughput increases 20%
(from 60 Mbps to 75 Mbps).

(Uggg, I've been using 10^9 for the amount transferred,
but it looks to be 2^30 actually. Absolute speeds are
actually 7% higher, but doesn't change the conclusion)

I also noticed that I reported 11 MB/s (88 Mbps) at the beginning
of this thread. This would point to a performance regression caused
by my defconfig & device tree changes :(


>> vendor DTS has the following child node:
>>
>>                 sdio {
>>                         pinname = "sdio";
>>                         ocr_avail = <0x200080>; /**VDD voltage 3.3 ~ 3.4 */
>>                         /* max_req_size = <0x20000>; */ /**128KB*/
>>                         max_req_size = <0x400>;
>>                         card_type = <3>;
>>                         /* 3:sdio device(ie:sdio-wifi),
>>                          * 4:SD combo (IO+mem) card
>>                          */
>>                         dmode = "pio";
>>                 };
>>
>> Maybe the vendor kernel uses the above information to "boost"
>> the performance of the SDIO-based WiFi adapter?
>
> PIO is also what we support upstream with the
> amlogic,dram-access-quirk; (which is enabled for &sd_emmc_a).
> This suggests that the pinctrl trick that Neil mentioned is not used here.
> 
> I assume that the wifi driver on the vendor kernel is the brcmdhd (out
> of tree) driver, while mainline uses brcmfmac.

Correct.
hardware/wifi/broadcom/drivers/ap6xxx/bcmdhd.100.10.545.x
vs
drivers/net/wireless/broadcom/brcm80211/brcmfmac

> One idea that comes to my mind is to enable the
> amlogic,dram-access-quirk; (and use pio mode on the vendor kernel) for
> eMMC or SD card and then compare read/write speeds. If they are
> similar-ish then the wifi performance difference is likely caused by
> the wifi driver (or in the opposite case: if read/write speeds on
> mainline with amlogic,dram-access-quirk; perform worse than the vendor
> kernel with pio mode then it's likely that it's a meson-gx-mmc driver
> limitation).

Thanks for the suggestion.

Are you aware whether someone has tried running the bcmdhd driver
on mainline?

Thanks for the test suggestion.

> From my work on the rtw88 SDIO wifi driver I can say that the
> meson-gx-mmc driver can push (TX direction) at least 120Mbit/s.
> I understand that this is half of what you get with the vendor kernel
> - and that this is the wrong direction (you're testing RX while I'm
> testing TX).
> The point that I want to get across is: I think nobody has the one
> answer why wifi performance is lower (personally I'm happy with having
> it work at all, performance is second).
> So it'll be a process to find the reason, and I think it requires
> being creative due to the large amounts of different code (MMC driver,
> wifi driver, additional patches...) between mainline and the vendor
> kernel.

I really appreciate you sharing this insight.

Regards.