rtw89 USB 2 to USB 3 Switch-Mode Gap

Update 2026-08-05: this landed upstream. rtw89_usb_switch_mode() is in mainline as of 8368970b6240 (2026-05-27) and ships in 7.2. Checked against the tree: v7.1 has no occurrences of the symbol in rtw89/usb.c, v7.2-rc6 has six. Everything below was measured before that and still describes 7.1 and earlier, which is what the shipping distros are on today. The gap is closed going forward.

Empirical proof that mainline drivers/net/wireless/realtek/rtw89/usb.c is missing a USB 2 to USB 3 switch-mode code path that already exists in the morrownr/rtw89 out-of-tree fork and in the sister rtw88 driver upstream.

Without the switch-mode code, every Realtek RTL8832AU / RTL8852AU / RTL8832BU / RTL8852BU / RTL8832CU / RTL8852CU / RTL8912AU / RTL8922AU USB adapter enumerates at USB 2.0 high-speed on first plug and stays there for the life of the plug, capping real-world TCP throughput at roughly 260 Mbps regardless of the radio's actual capability.

The fix already exists. It was written by Bitterblue Smith in morrownr/rtw89 commits cd287cc (2025-07-16) and c8a8ac4 (2025-08-08). The same mechanism has been upstream in mainline rtw88 since 2024-07-10 across four commits by the same author, all reviewed and accepted by Ping-Ke Shih (Realtek rtw89/rtw88 maintainer).

TL;DR

Adapter	Chip	Stock UL	morrownr UL	Upload delta
D-Link DWA-X1850 A1	RTL8852AU	258 Mbps	802 Mbps	3.1x
D-Link DWA-X1850 B1	RTL8852AU	257 Mbps	803 Mbps	3.1x
Brostrend AX1800	RTL8852BU	217 Mbps	620 Mbps	2.9x
BrosTrend BE6500	RTL8922AU	260 Mbps (*)	957 Mbps	3.7x

Tested on Framework 13 (Fedora 43, kernel 6.19.11, x86_64, Intel Tiger Lake xHCI) and Raspberry Pi 5 (Pi OS, kernel 6.12.47, aarch64, Broadcom RP1 xHCI). All four adapters also confirmed at USB 3.0 SuperSpeed on the Pi, with WiFi 7 (EHT-MCS 12) negotiated on the BE6500.

(*) The RTL8922AU (WiFi 7) has no stock mainline USB driver at all. "Stock" for that adapter was synthesized by cloning morrownr/rtw89 at HEAD 2c2d99d and reverting just commits cd287cc and c8a8ac4 to isolate the effect of the switch-mode code on the same codebase.

See FINDINGS.md for the full writeup.

Directory layout

rtw89-usb3-gap/
├── README.md                     Headline summary (this file)
├── FINDINGS.md                   Full technical writeup
├── adapters/                     Per-adapter identity and behavioral notes
├── hosts/                        Per-host hardware and OS details
├── submissions/                  Mirror of the linux-wireless mailing list submissions and reviewer reply threads
│   └── v2/                       v2 cover letter + patch + replies (Johannes, Bitterblue x2, Ping-Ke)
└── evidence/
    ├── iperf3/                   iperf3 logs from the full test matrix
    ├── dmesg/                    dmesg probe captures (framework-control + framework-morrownr-git)
    ├── sysfs/                    Per-plug USB state snapshots
    ├── iw-link/                  Per-plug radio link snapshots
    ├── crash-2026-04-11/         xHCI hard lockup crash evidence (1235 kernel lines, recovered from /var/log/messages)
    ├── may-2026-laptop/          May 2026 expanded matrix (per-cell iperf3 JSONs + byte-counter deltas + pre/post link/USB state + per-iteration timestamps)
    └── downstream-vendor/        BrosTrend shipping modprobe confs (morrownr OOT templates + two BrosTrend sed tweaks, rtl8852bu-dkms 1.19.21 + rtl8852cu-dkms 1.19.22)

Upstream submission status

v2 of the rtw89 USB-3 switch-mode patch was sent to linux-wireless on 2026-05-11 by Devin Wittmayer carrying Bitterblue Smith's original code (morrownr/rtw89 commit cd287ccf544b, 2025-07-16) rebased onto pkshih/rtw rtw-next. See submissions/v2/ for the cover letter, patch, and reviewer reply thread (Johannes Berg on module parameter, Bitterblue Smith on testing verification + DACK crash unrelatedness, Ping-Ke Shih on positive chip list for v3). Lore thread: https://lore.kernel.org/linux-wireless/20260511160811.17647-1-lucid_duck@justthetip.ca/T/#t

Credits

• Bitterblue Smith (@dubhater) wrote the switch-mode code and all rtw89 USB driver support. This repository exists to support his work, not to compete with it.
• louis-kotze opened morrownr/rtw89#76 addressing the separate RF calibration timeout issue; that fix is independently valuable and being handled via that PR.
• Testing and writeup: Devin Wittmayer (@Lucid-Duck).

License

Test data and documentation are released under CC0. The no-switchmode branch used in the synthesized-stock test is derived from morrownr/rtw89 and retains its original GPL-2.0 license.