Linux USB
 help / color / mirror / Atom feed
From: bugzilla-daemon@kernel.org
To: linux-usb@vger.kernel.org
Subject: [Bug 221699] PCIe link drops during GPU BAR2 init — Minisforum DEG2 eGPU dock over Thunderbolt 4
Date: Sun, 22 Nov 2026 11:37:12 +0000	[thread overview]
Message-ID: <bug-221699-208809-G3FTOHFiDr@https.bugzilla.kernel.org/> (raw)
In-Reply-To: <bug-221699-208809@https.bugzilla.kernel.org/>

https://bugzilla.kernel.org/show_bug.cgi?id=221699

inwestygator@gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |inwestygator@gmail.com

--- Comment #4 from inwestygator@gmail.com ---
Another Intel Panther Lake data point, with a working workaround.

Hardware:
- Laptop: Framework Laptop 13 Pro (Intel Core Ultra Series 3, Panther Lake, Arc
B390 iGPU on xe), BIOS 03.04
- eGPU: Razer Core X V2, Intel JHL9480 / Barlow Ridge (8086:5786), USB4 40 Gb/s
(2 x 20)
- GPU: NVIDIA RTX 3090 (10de:2204), PCIe 4.0 x4
- PCIe path: 00:07.0 (USB4 root port) -> 02:00.0 (switch upstream) -> 03:00.0
(switch downstream) -> 04:00.0 (GPU)
- Kernel 7.2.15 (Arch, Omarchy). NVIDIA open 615.78.08 and proprietary
580.178.04.

Symptom:
The GPU falls off the bus (Xid 79) at state changes: during the driver's start,
at the first
Vulkan use after idle, and when a workload stops. Steady load runs fine. After
a failure only a
power cycle of the enclosure recovers it. With the open driver it failed in 19
of 20 boots.

What I found:
- With the proprietary driver and NVreg_EnableGpuFirmware=0 (GSP firmware off)
the card is
  stable, but it stays in P0 at about 105 W, so its link speed never changes.
- With GSP on, the same proprietary driver fails again (2 of 2 boots).
- Locking the link speed before the driver loads fixes it with GSP on, at full
16 GT/s
  (the workaround from bug 221319 comment 58):

    setpci -s 03:00.0 CAP_EXP+30.w=0024:002f   # LnkCtl2: target 16 GT/s + HW
Autonomous Speed Disable
    setpci -s 04:00.0 CAP_EXP+30.w=0024:002f   # same on the GPU
    setpci -s 03:00.0 CAP_EXP+10.w=0020:0020   # retrain, then load the driver

  Results: 8 vulkaninfo runs with 20 s idle between them, no Xid (unlocked:
failed within one to
  three runs). 13 runs of a compute workload started and stopped (a model
loaded on the
  GPU, a request, the process exiting), most with 30 s idle after: no Xid. The
link stayed at 16 GT/s and the card idles normally in P8 at
  about 7 W.

  Then the same with the open driver 615.78.08 (unlocked it failed in 19 of 20
boots, often
  already while loading): the driver loaded cleanly, 8 vulkaninfo runs with 20
s idle and 4
  workloads started and stopped with 30 s idle, no Xid, link at 16 GT/s, idle
in P8.

So on this Intel host too, the failure follows the GPU changing its link speed,
which fits the
analysis in bug 221319. ASPM was already disabled on the link.

Related reports:
https://forums.developer.nvidia.com/t/gsp-firmware-triggering-xid-79-on-usb4-thunderbolt-egpu/389811
(with nvidia-bug-report.log.gz, GSP on, right after a failure) and
https://github.com/NVIDIA/open-gpu-kernel-modules/issues/979

-- 
You are receiving this mail because:
You are watching the assignee of the bug.

      parent reply	other threads:[~2026-11-22 11:37 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-06-28  7:21 [Bug 221699] New: PCIe link drops during GPU BAR2 init — Minisforum DEG2 eGPU dock over Thunderbolt 4 bugzilla-daemon
2026-06-28  7:22 ` [Bug 221699] " bugzilla-daemon
2026-06-28  7:28 ` bugzilla-daemon
2026-07-10 23:32 ` bugzilla-daemon
2026-11-22 11:37 ` bugzilla-daemon [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=bug-221699-208809-G3FTOHFiDr@https.bugzilla.kernel.org/ \
    --to=bugzilla-daemon@kernel.org \
    --cc=linux-usb@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox