Re: [REGRESSION] amdgpu with Thunderbolt eGPU bracket fails since new bridge window alignment calculation code

["Ilpo Järvinen" <ilpo.jarvinen@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 3164 bytes --]

On Mon, 30 Mar 2026, Jonas Höglund wrote:

> On Mon, 30 Mar 2026, at 14:33, Ilpo Järvinen wrote:
> > I'm skeptical it's exactly the same issue even if the end result is the 
> > same.
> >
> > The resource fitting algorithm has been in a state of constant flux due to 
> > various fixes and improvements into it over all the recent two years.
> > Unfortunately, fixing one thing (or even moving towards fixing an issue) 
> > may break another thing due to how different resource interact.
> 
> Ok, yeah, I don't envy having to deal with that.  You're probably right
> it's more BAR-related, I mostly keyed in on the very similar symptom.

Definitely the gpu driver could handle an resource issue better than by 
calling something that triggers a sanity check somewhere, but it's 
secondary problem.

> > That "PCI: Improve head free space usage" series is certainly fixing two 
> > known corner case with the commit 3958bf16e2fe ("PCI: Stop over-estimating 
> > bridge window size") but with only heavily filtered logs, I'm unable to 
> > confirm if it applies to this case as well.
> 
> Sorry for not providing full logs from the get-go; I couldn't think of
> suitable location.  Here's a full dmesg for reference of the crash
> manifesting on 7.0.0-rc5:
> 
> https://up.firefly.nu/pub/amdgpu-egpu-crash-7.0.0-rc5.dmesg.txt
> 
> 
> > From the limited logs, I suspect this is primarily a BAR resize rollback 
> > failure which leaves the resources into a state worse than they were prior 
> > to the resize. The commit 337b1b566db0 ("PCI: Fix restoring BARs on BAR 
> > resize rollback path") attempts to rectify that. The entire series is here 
> > (not all of it went to stable):
> 
> > https://lore.kernel.org/all/20251113162628.5946-1-ilpo.jarvinen@linux.intel.com/T/#m9b0e316c94f7abc0686e58f902d05ff35aeac3ac
> >
> > The fixes to that series are here:
> >
> > 5528fd38f230 ("PCI: Fix Resizable BAR restore order")
> > 08d9eae76b85 ("PCI: Fix BAR resize rollback path overwriting ret")
> 
> Unless I misread something, they should both be included in the recently
> tagged 7.0.0-rc6--I'll try building it and see if the issue is resolved.
> 
> I'll reply once I've tested 7.0.0-rc6.

Hi again,

Now that I look more into the logs that probably won't help. For some 
reason, it seems that resize is not even attempted and the errno is 
-EINVAL which is a bit unexpected.

I'm starting to wonder that the problem fixed by this patch once again is 
showing its ugly head (it's currently in pci/resource branch, so it won't 
appear until 7.1-rc1):

https://lore.kernel.org/linux-pci/20260326200427.GA1340256@bhelgaas/

I still don't understand why pbus_select_window() would return NULL in 
this case but it looks the most likely candidate where -EINVAL could come 
from (I still don't understand what cleared resource's flags if that's the 
case but it still seems the best explanation).

Please take logs from this point on with dyndbg="file drivers/pci/*.c +p" 
on the kernel's command line so there's little bit of extra info (and 
check you are building with CONFIG_DYNAMIC_DEBUG).

-- 
 i.