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: <2e5161a6-d273-4f76-429a-12dc2a6c7322@reto-schneider.ch>
Date:   Wed, 2 Feb 2022 01:42:16 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.5.1
Content-Language: en-US
To:     chris.chiu@canonical.com, linux-wireless@vger.kernel.org,
        Jes.Sorensen@gmail.com
Cc:     pkshih@realtek.com
From:   Reto Schneider <code@reto-schneider.ch>
Subject: rtl8xxxu: Performance findings and questions
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi all,

While trying to improve the retransmission percentage on my rtl8188cus 
devices, I found the following:

rtl8xxxu_tx() takes every sk_buf handed by mac80211 and relays it to the 
rtl8188cus right away. This results in 10k+ URBs/s, each ~1500 bytes in 
size on my x86 workstation. On my low-end ARMv5 devices the URBs peak 
out at ~5k/s, with lower throughput and higher retransmission rates 
(~30% in a shielded RF box, ~10% for my workstation).

The Realtek 8192cu bundles the TX frames [1] and sends up to 18kB sized 
URBs, resulting in just 1k URBs/s. The retransmission rates with this 
driver are below 1%.

Questions:
- Any chance this URB "spam" is the root cause for the rtl8188cus being
   too slow to ACK received frames in time, causing retransmissions by
   the AP?
- Is implementing wake_tx_queue() the proper way to improve this?

Kind regards,
Reto

[1] 
https://github.com/husqvarnagroup/rtl8xxxu-8192cu-for-rtl8188cus/blob/d8c03db466bbb3e9c4e8e6cb0a589cfef55db756/core/rtw_xmit.c#L1688