"Guilty" API Proposed For Better Communicating Why Radeon GPUs Hang/Reset

[RealNC]

What is it with GPU drivers where it seems it's pretty much acceptable that user-space is able to crash them. I'm not sure I get it.

[-MacNuke-]

On Windows however, it's standardized across drivers.

Its not like that GPU vendors don't have to do anything to support driver resets in Windows. They have to support it and it does not come for free. If the Nvidia driver under Linux does not support GPU resets its not Linux fault.

[-MacNuke-]

What is it with GPU drivers where it seems it's pretty much acceptable that user-space is able to crash them. I'm not sure I get it.

In a perfect world they should never crash. But there is no perfect world. Bugs are everywhere. And since gpu drivers are performance critical you can't just fill everything with validation code.

[mdedetrich]

Its not like that GPU vendors don't have to do anything to support driver resets in Windows. They have to support it and it does not come for free. If the Nvidia driver under Linux does not support GPU resets its not Linux fault.

This is an absurdly reductionist statement to the point where its not really correct. The design of the API has a massive result in how GPU's/drivers can behave and WDDM is designed very differently compared to the Linux graphics stack and one of the initial design goals of WDDM (all the way back in Vista) is to as much as possible allow recoverable restarts of the GPU.

[mdedetrich]

What is it with GPU drivers where it seems it's pretty much acceptable that user-space is able to crash them. I'm not sure I get it.

GPU drivers are much more complex than a keyboard driver ontop of which they are extremely performance sensitive.

[ResponseWriter]

What I want to see is user session recovering after a GPU hang. It's still not happening yet.

Agreed. IIRC Plasma 6 is where we'll see this ability finally merged.

[user1]

Agreed. IIRC Plasma 6 is where we'll see this ability finally merged.

I think what you're referring to in Plasma 6 is the ability to recover a Wayland session from a crash, but it's not related to GPU hangs/crashes.

[_r00t-]

is this related to "*error* ring gfx timeout" ?

[ResponseWriter]

I think what you're referring to in Plasma 6 is the ability to recover a Wayland session from a crash, but it's not related to GPU hangs/crashes.

Yep, my bad. Would this not be related since my recent experience of GPU resets has required restarting the kwin session?

[cb88]

AFAIK.... GPUs prior to RDNA2 had a hard hang bug that required some nasty work arounds to reset at all.

RNDA2 and later should already reset itself.