[PATCH rtw-next v2 0/1] wifi: rtw89: usb: Support switching to USB 3 mode

[Devin Wittmayer]

This patch adds USB 2 to USB 3 mode switching for the AX-generation
chips in the rtw89 USB driver (8852AU, 8852BU, 8852CU and helper
variants). Without it, those adapters enumerate at USB 2.0
high-speed on first plug and stay there, capping real-world
throughput around 260 Mbps TCP regardless of radio capability.

It is Bitterblue Smith's morrownr/rtw89 commit cd287ccf544b
(2025-07-16) rebased onto rtw-next without code changes:
author's Signed-off-by preserved, my Signed-off-by added as
relayer, Tested-by per chip based on the matrix below.

Changes since v1
================

- Rebased onto pkshih/rtw rtw-next (HEAD c1ed02655f91) per
  maintainer feedback. The patch diff is unchanged from v1; the
  test matrix below carries forward.
- Cover letter tightened to AX-generation scope only.

The bug
=======

These chips present as USB 2 devices on first plug. The driver is
expected to send a switch-mode register write to make them disappear
and re-enumerate as USB 3 SuperSpeed. Mainline rtw89 does not issue
these commands anywhere. Every user with one of these adapters on
a mainline-derived distribution sits at the USB 2.0 ceiling for
the life of the plug, regardless of negotiated radio rate.

The fix
=======

The patch covers AX-generation chips (8852AU, 8852BU, 8852CU and
helper variants) via R_AX_PAD_CTRL2. It also adds the
switch_usb_mode module parameter (default on) for users who
experience 2.4 GHz interference under USB 3.

Mainline precedent
==================

The same mechanism is in mainline rtw88 across three commits
between July 2024 and April 2025, all authored by Bitterblue Smith:

  commit 315c23a64e99 ("wifi: rtw88: usb: Support USB 3 with RTL8822CU/RTL8822BU")
  commit 82a35723a67c ("wifi: rtw88: usb: Support USB 3 with RTL8812AU")
  commit bf1103654df9 ("wifi: rtw88: usb: Enable switching the RTL8814AU to USB 3")

Test results
============

60 plug-cycles + 30+ gated throughput cells captured 2026-04-11
to 2026-05-07. Both x86_64 (Tiger Lake xHCI) and aarch64
(Broadcom RP1 xHCI) reproductions confirm the bug and the fix.

  Hosts:    Framework 13 (x86_64, Fedora 43 / 6.19.13, Tiger Lake xHCI)
            NucBox K8 Plus (x86_64, Arch Linux / 6.17.9, AMD F19h USB)
            Raspberry Pi 5 (aarch64, Pi OS / 6.12.47, BCM2712 + RP1)
            Banana Pi BPi-R4 Pro (aarch64, OpenWrt / 6.6.93, MT7988A;
                                  also used as a controlled lab AP)
  Adapters: D-Link DWA-X1850 A1 / B1 (RTL8852AU)
            BrosTrend AX1L compact / AX4L high-gain (RTL8852BU)
            BrosTrend AX8L AXE5400 / EDUP AXE5400 (RTL8852CU)
  APs:      consumer multi-band router (WPA3-SAE / WPA2-PSK)
            BPi-R4 Pro single-band lab AP (WPA3-SAE)
  Server:   Linux iperf3 host on 2.5 GbE wired Ethernet

Full per-cell evidence (raw iperf3 stdout, dmesg captures, sysfs
snapshots, iw link snapshots, byte-counter deltas, per-host detail)
at https://github.com/Lucid-Duck/rtw89-usb3-gap.

Plug-cycle. N=10 cycles per (chip, host) cell. Pass = post-plug
/sys/bus/usb/devices/<id>/speed reads 5000 (USB 3 SuperSpeed),
zero WARN/BUG/Oops in cycle-bracketed dmesg, non-empty association
BSSID after settle.

  Adapter (chip)               FW13 6.19      K8 Plus 6.17
  ---------------------------  -------------  --------------
  DWA-X1850 A1 (RTL8852AU)     10/10 PASS     10/10 PASS
  BrosTrend AX1L (RTL8852BU)   10/10 PASS     10/10 PASS

  FW13:    Fedora 43, 6.19.13, wireless-next + this patch
  K8 Plus: Arch Linux, 6.17.9-arch1-1, morrownr/rtw89 (upstream
           fork; identical AX-gen behavior to this patch)

Throughput. TCP iperf3, N=10 30-second iterations per sub-cell,
four sub-cells per cell (P=8 and P=1, each direction), to a Linux
2.5 GbE host. Per-cell /32 host route forces traffic onto the WiFi
adapter; per-iteration byte-counter cross-check on every cited cell
shows wireless tx/rx delta at 103-108% of iperf3 reported bytes
(excess is normal TCP/IP framing), so wired-NIC bleed is ruled out.
Linux tuning applied (rmem 32 MB, RPS=ff on the WiFi RX queue).

FW13 stock (in-kernel rtw89, USB 2 stuck) vs patched (USB 3
SuperSpeed), same adapter / AP / kernel / capture script, P=8 mean
Mbps, AP is the consumer multi-band router:

  Adapter (chip)              Band, width      UL stock  UL patched  DL stock  DL patched
  --------------------------  ---------------  --------  ----------  --------  ----------
  EDUP AXE5400 (RTL8852CU)    6 GHz, 160 MHz       269        1364       327         579
  AX8L AXE5400 (RTL8852CU)    6 GHz, 160 MHz       269        1440       324         510
  AX4L AX1800 (RTL8852BU)     5 GHz,  80 MHz       208         597       293         695
  AX1L AX1800 (RTL8852BU)     5 GHz,  80 MHz       235         608       273         843
  DWA-X1850 A1 (RTL8852AU)    5 GHz,  80 MHz       254         748       264         707
  DWA-X1850 B1 (RTL8852AU)    5 GHz,  80 MHz       248         706       265         679

USB enumeration verified at SuperSpeed (5000 Mbps) on every
patched cell, both pre and post.

K8 Plus second-rig spot check, kernel 6.17.9-arch1-1, consumer
multi-band router, single-link 5 GHz 80 MHz, TCP P=1, N=10:

  DWA-X1850 A1 (RTL8852AU)    788 UL / 693 DL Mbps (sigma 16 / 54)

For comparison, every stock cell above caps at 208-327 Mbps, the
USB 2.0 ceiling.

Submission notes
================

Applies cleanly on pkshih/rtw rtw-next HEAD c1ed02655f91.
checkpatch.pl --strict: zero errors, zero warnings, zero checks.

Bitterblue Smith (1):
  wifi: rtw89: usb: Support switching to USB 3 mode

 drivers/net/wireless/realtek/rtw89/reg.h |  4 +++
 drivers/net/wireless/realtek/rtw89/usb.c | 41 ++++++++++++++++++++++++
 2 files changed, 45 insertions(+)

--
2.53.0