Re: [REGRESSION][BISECTED] usb: hub: total system freeze after running adb

[Slavik Dev]

Hi,

I'm hitting what appears to be the same regression, on the same xHCI
hardware as the original report. I spent months chasing this as a
hardware problem (RAM, PSU, BIOS) before the adb correlation
eventually led me to this thread. Earlier today I applied the
workaround discussed here (disabling USB runtime PM), and so far it
is holding up. Sharing my data in case it helps revive the
investigation -- until the workaround, this reproduced reliably for
me, up to several freezes per evening.

Hardware:
- AMD Ryzen 7 7700 (Raphael, display on iGPU), ASUS TUF GAMING B650-E
  WIFI, BIOS 3881 (AGESA 1.3.0.1b)
- The affected controller from the original report is present:
  0b:00.0 USB controller [1022:15b8] (AMD Raphael USB 2.0 xHCI)

Kernels (Linux Mint 22.3 / Ubuntu HWE builds):
- Freezes observed on 6.17.0-14/-19-generic and 6.14.0-37-generic
  (based on 6.14.11 + stable backports, built 2025-11-20).
- First freeze on 2026-03-10, shortly after kernel package updates;
  16 freezes total through 2026-08-03.

Correlation with adb:
- The last 4 freezes occurred 5, 12, 35 and 81 minutes after the adb
  server was started (using the once-per-boot "adb uses wireless
  extensions" kernel warning as the start marker in the journal).
- Earlier freezes also occurred with an Android phone plugged in while
  the adb server was NOT running -- including one freeze 2 seconds
  after unplugging the phone. Consistent with hub runtime
  suspend/resume transitions being the trigger rather than adb itself.

Freeze signature:
- Journal stops mid-second, no oops/panic, no MCE/WHEA, efi-pstore
  empty. nmi_watchdog=1 and hardlockup_panic=1 never fire.
- Notably: the chipset watchdog (sp5100_tco + systemd
  RuntimeWatchdogSec=30) DOES fire and resets the machine ~30 s after
  the hang, so the platform stays alive -- it is the kernel/CPUs that
  stop responding.

Ruled out on this machine:
- RAM: freezes identically with DDR5 at EXPO 6000, EXPO 5600 and
  JEDEC stock 4800.
- Pure load: 60 minutes of stress-ng (16 CPU workers + 16 GB VM
  workers, no adb running) survives without issues.
- Firmware: BIOS update 3223 -> 3881 (AGESA 1.3.0.1b) made no
  difference.

Side observation: in freezing sessions, Bluetooth (MT7922 on the same
bus as most devices) repeatedly logs "hci0: ACL packet for unknown
connection handle" before the freeze.

Workaround status:
- Running with runtime PM disabled for all USB devices
  (power/control=on via udev rule) since 2026-08-03. Reporting while
  it's fresh: the adb server has since been running continuously,
  already well past my longest previously observed time-to-freeze
  (81 min), and so far so good.