Re: [BUG] thunderbolt: USB4 v2 TBGAA tunnel creation crash in TMU enhanced uni-directional mode

[Mika Westerberg <mika.westerberg@linux.intel.com>]

Hi,

On Mon, May 25, 2026 at 05:22:35PM -0500, Mario Limonciello wrote:
> Hi,
> 
> A few nested comments below.
> 
> On 5/24/26 11:44 PM, Mika Westerberg wrote:
> > +Mario
> > 
> > Hi,
> > 
> > On Sat, May 23, 2026 at 06:43:06PM +0200, Sebastian Loscher wrote:
> > > Hi Mika and the Linux USB team,
> > > 
> > > I hope you are well. I am testing a new Thunderbolt 5 setup on a brand-new AMD
> > > Strix Halo system, and I have run into a reproducible hardware panic during the
> > > initial connection handshake.
> > > 
> > > I have gathered full dynamic debug traces and would appreciate your insight on
> > > whether this requires a specific hardware quirk.
> > > 
> > > This is my first report in such format ever, maybe big for me but please bare
> > > with me.
> > > 
> > > Hardware:
> > > 
> > > Host: Minisforum MS-S1 Max (AMD Strix Halo USB4 v2 controller)
> 
> Strix Halo doesn't have a USB4v2 router, but I believe this product has
> added a discrete USB4v2 router to some of the PCIe lanes on the Strix Halo
> APU.
> 
> > > 
> > > Dock: Minisforum eGPU dock (Vendor 0x41f, Device 0xd002, TBGAA controller)
> > > 
> > > Kernel: CachyOS (state your exact kernel version, e.g., 6.12-rcX)
> > > 
> > > The Issue:
> > > When plugging in the Thunderbolt 5 dock, the connection fails during USB3
> > > tunnel creation and drops entirely.
> > > 
> > > Using dyndbg="file drivers/thunderbolt/* +p", the trace shows the AMD host
> > > successfully shifting the dock into TB5 asymmetric routing (TMU: mode set to:
> > > enhanced uni-directional, MedRes). However, immediately after querying the NVM
> > > authentication status of the retimers (reading NVM authentication status of
> > > retimers), the dock suffers a hardware panic and physically drops the link
> > > (acking hot unplug event on 0:1).
> > > 
> > > The subsequent -107 and -71 PCIe/USB errors appear to be the Linux driver
> > > attempting to route through a connection the dock has already severed.
> 
> Can you reproduce your issue on the AMD routers too (these are the regular
> USB4 ones with PCI vendor 1022) or only on the USB4v2 ones?
> 
> You might need to look at the user guide for your device to see which USB-C
> ports these are connected to if it's not obvious from silkscreens.
> 
> I do know we have a bug report on kernel bugzilla about this eGPU dock
> misbehaving in Linux in lots of various ways.
> 
> https://bugzilla.kernel.org/show_bug.cgi?id=221319
> 
> The most severe symptom is that when using an NV dGPU with the GSP loaded it
> causes a sync flood to the APU and the system reboots.
> 
> There are also reports of performance problems though which I was hoping
> were ASPM, but that turns out to not be the only problem.  I did start a
> discussion about that.
> 
> https://lore.kernel.org/linux-pci/20260504225246.480921-1-mario.limonciello@amd.com/
> 
> Maybe Mika and some others on this list have some further ideas.  It seems
> Windows and Linux are setting up something very differently.

We are investigating this NV one as well but so far no updates.