Hello Michał, hello Mathias at all,
many thanks for your answers!
I have tried if I can reproduce it with a AMD Linux PC, but
unfortunately I was not able to reproduce (but setup is a bit different).
So I went back to Raspberry Pi Compute Module 5, where I mainly
connected the radio module (Quectel RM520N-GL) via USB3,
and installed a Wifi access point. All data/all connections from Wifi
access point are routed directly via wwan0 to radio module.
This is currently my easiest setup to be able to reproduce the error.
Mostly in a few seconds.
My knowledge in area Linux Kernel + USB is unfortunately not yet enough
to analyze and fix it by myself.
But I used the help of ChatGPT-5 to create an usbmon and xhci kernel trace.
I create an usbmon trace as well as a trace from xhci (both recorded in
parallel):
https://www.file-upload.net/en/download-15523936/usbmon_bus5_20250817-150158.log.html
https://www.file-upload.net/en/download-15523937/xhci_20250817-150158.trace.html
This was the last output, my ping in a shell has shown:
64 bytes from 8.8.8.8: icmp_seq=2323 ttl=112 time=26.0 ms
64 bytes from 8.8.8.8: icmp_seq=2324 ttl=112 time=25.0 ms
64 bytes from 8.8.8.8: icmp_seq=2325 ttl=112 time=29.1 ms
64 bytes from 8.8.8.8: icmp_seq=2326 ttl=112 time=37.8 ms
In parallel created more data traffic, but with ping I see first when IP
data connection does not work stable anymore.
According to ChatGPT-5 the following places contain errors:
*** USBMON ***
In your usbmon_bus5_20250817-150158.log:
First -71 (EPROTO) on the QMI Bulk-IN (Bi:5:005:14): line 2161,
timestamp 493245744
2161: ffffff8003c8cb40 493245744 C Bi:5:005:14 -71 0
Just before that, there’s a -75 (EOVERFLOW) on the same IN EP, which is
often the first sign of trouble: line 2159, timestamp 493245221
2159: ffffff8003c8cd80 493245221 C Bi:5:005:14 -75 1024 = ...
So the sequence is: several good completions → EOVERFLOW (-75) → then a
stream of EPROTO (-71) errors on Bi:5:005:14, which kills further ping
replies after your last good seq (2326).
*** XHCI TRACE ***
I found the first failure in your xHCI trace.
First error line: line 8216
Timestamp: 758267.000115
Event: xhci_handle_event … type 'Transfer Event' … 'Error' … slot 1 ep
29 … len 1472
Why ep 29? In xHCI, the endpoint context index is ep_index = 2 *
ep_number + (direction), where direction is 0=OUT, 1=IN.
So for Bulk IN ep 14: 2*14+1 = 29 → that’s your IN 0x87 pipe.
Right after that line you can see the driver react:
xhci_handle_transfer … length 1472 … (the failed TD)
xhci_queue_command: Reset Endpoint Command … ep 29 (host tries to recover)
xhci_handle_event: … 'Command Completion Event' (reset completes)
But from this point on, completions for that IN EP correspond to usbmon
-71 (EPROTO) — matching what you saw.
Does this give a clue, where it could be coming from?
It is 100% reproduceable in a few seconds on Raspberry Pi Ccompute
Module 5 (and I same behaviour on different kernel of i.MX8MP).
Could it be a hardware problem? I already tried different radio module
(all Qualcomm, X62/X65 and X72/X75),
different cables (all same length, all from same source), different eval
board for the M.2 radio modules (but from same source).
Can you give me a hint, what to try next?
ChatGPT-5 pinpoints me to try to disable LPM for USB3, could this be a
next step? Or is it something else?
Many thanks for your help!
Best regards,
Martin
