AMD Working On More Precise GPU Reset Capabilities For Their Linux Driver

[phoronix]

Phoronix: AMD Working On More Precise GPU Reset Capabilities For Their Linux Driver

While the hope remains that GPU resets are a very infrequent task, AMD Linux driver engineers have recently been working on the ability to support a per-queue GC reset capability for more precise reset capabilities when needed...

https://www.phoronix.com/news/AMDGPU-Per-Ring-Resets

[Kjell]

Wonderful work!

This makes me wonder how Wayland compositors will react during a ring reset after this change. Soft Freeze? Crash?

Would be amazing if they'd be able to handle this gracefully

[duby229]

Wonderful work!

Makes me wonder how Wayland compositors would react during a ring reset. Freeze? Crash?

I'm absolutely certain it'll need a protocol extension that will take 7 years to design and another 3 to implement.

EDIT: And then the implementation will forever be buggy because the protocol that took 7 years to derive is overly simplistic and horribly incomplete!!

[ojsl1]

I'm sick and tired of having to reset my whole desktop environment with sysrq+e just to kill a single process that froze the whole pc, in that regard this is a welcome feature if it gets through.

[Ranguvar]

Looks like this is only for GFX9 (Vega) and up, right?
Shame. My Polaris WX 5100 is still going strong.

[middy]

as much as we crap on microsoft, at least windows is the gold standard with gpu crashing recovery. i'd love linux to be able to handle gpu crashing as well as windows does it. when amd drivers were crashing on world of warcraft for over a year https://eu.forums.blizzard.com/en/wo...-12/475734/216, you could have the gpu crash five times in a row, each crash right after the other, crash, recover, crash, recover, etc and everything was recovered during it all. nothing else actually crashed. game still kept going. everything.

[zaps166]

Good to hear it'll get better! For me it happens often when I develop Vulkan shaders and I do a mistake. Also some games (like HL) can pagefault and reset the GPU. Seeing login screen after GPU reset in this scenario isn't good

I also tried "amdgpu.reset_method=3", but it's not very stable (e.g. hibernation is broken and doesn't always work correctly).

Is it going to land in 6.11?

[tildearrow]

I'm absolutely certain it'll need a protocol extension that will take 7 years to design and another 3 to implement.

EDIT: And then the implementation will forever be buggy because the protocol that took 7 years to derive is overly simplistic and horribly incomplete!!

Be noted that not even X.Org is capable of handling a GPU reset appropriately - every time it happened to me, X.Org would crash completely or become unusable.

[Brittle2]

Wonderful work!

This makes me wonder how Wayland compositors will react during a ring reset after this change. Soft Freeze? Crash?

Would be amazing if they'd be able to handle this gracefully

Compositors recover fine from a crash, I don't think this gonna change anything, maybe they can just drops the frame instead of crashing and restarting?

[Kjell]

I don't think this gonna change anything, maybe they can just drops the frame instead of crashing and restarting?

I've had rare occurrences where the GPU would hard hang (meaing the screen would be frozen and changing TTY wouldn't work despite services/SSH/Audio continuing to run fine)

Then there's also (again very rare) hard ring crashes which just freeze the entire system.

Hope this will be an issue of the past and like you said, just drop a frame or two

Compositors recover fine from a crash

Not all. KDE has the ability to restore windows after a crash

Majority of compositors lose progress

This change to ring logic makes me wonder if the compositor needs some new logic/protocols to handle such occurrence