9 Year Old Mesa Bug Fixed, Addresses Stability With RadeonSI Gallium3D

[marek]

A sampler state is basically a texture filtering state.

The bug is 9 years old, but it didn't take 9 years to fix, nor was it reproducible for 9 years. You need a certain app behavior to trigger it.

The guy who uploaded the apitrace from Batman Arkham Origins to bugzilla also deserves some credit, because that was my only solid test of the bug.

[Azrael5]

one of the best developer ever.

[hxfhjkl]

Looks like it might fix XCOM:EW crashes, too! Finally I can finish this.

https://bugs.freedesktop.org/show_bug.cgi?id=80419

I hope so, the hang in xcom seemed to happen in very similar fashion to the one in Team Fortress for me also.

[bibaheu]

There is a way to handle this. It's work in progress for AMDGPU. Look for the various commits regarding GPU reset.

The hard part seems to be making sure that resets don't result in inconsistent state of the driver and memory management. As far as I can tell, actually resetting the hardware is quite simple.

Nice! I hope it also helps with the PCI Passthrough virtualization system. After the virtual machine controlling the GPU is turned off, the GPU is "dead" and with the fans at 100%. But if you put the host to suspension, it'll get back to work perfectly. Being able to reset the card without actually turning it off electrically sounds will be helpful here

[Sonadow]

There is a way to handle this. It's work in progress for AMDGPU. Look for the various commits regarding GPU reset.

The hard part seems to be making sure that resets don't result in inconsistent state of the driver and memory management. As far as I can tell, actually resetting the hardware is quite simple.

How does Windows do it? I have seen Windows Vista and upwards capable of recovering gracefully from a GPU driver lockup with just a single flicker of the screen, and without bringing the system down or crashing the currently running application.

[geearf]

A sampler state is basically a texture filtering state.

The bug is 9 years old, but it didn't take 9 years to fix, nor was it reproducible for 9 years. You need a certain app behavior to trigger it.

The guy who uploaded the apitrace from Batman Arkham Origins to bugzilla also deserves some credit, because that was my only solid test of the bug.

I uploaded so many logs and traces to my Batman Arkham Origins fdo bug, I never thought anything would come out of it.
If you're talking about my bug, I'm glad something helped!
Also we were lucky, as I removed these files yesterday from where they were hosted (mega)...
If not my bug, well... thanks anyway

[gamerk2]

I wonder if there is a way to recover a GPU when it goes bananas. Usually Windows can recover from GPU driver crash, resetting the output to VGA, killing all DirectX processes and putting a message to "pls restart now".

As of Vista/WDDM 1.0, Windows will attempt to re-initialize the GPU driver, usually signified by "The display driver has stopped responding but has recovered" messages. If the restart fails, Windows will throw either an 0x116 or 0x117 BSOD error message.

So yeah, for non-hardware faults, its certainly doable.

[bridgman]

How does Windows do it? I have seen Windows Vista and upwards capable of recovering gracefully from a GPU driver lockup with just a single flicker of the screen, and without bringing the system down or crashing the currently running application.

Dozens of developer-years of driver effort plus equivalent amount or more in the OS.

[timofonic]

Dozens of developer-years of driver effort plus equivalent amount or more in the OS.

C'mon, that's the short answer! What's the real difference to be able to have THAT feature?

[mibo]

Thanks marek and log0 for the explanations.

As I understand, every game crash could actually have been this bug?
I tried again to get SC2 running with wine (current) and MESA 13.0.2 on a rx480. 3 out of 4 co-op matches did crash.
Hope, the fix makes it into a 13.0.3 soon - I want to test it :-)