On 2020-11-30 3:07 p.m., Alexander Monakov wrote:
My other concern is how easy it is to cause system instability or hangs by out-of-bounds writes from the GPU (via compute shaders or copy commands). In my experience of several years doing GPU computing with NVIDIA tech, I don't recall needing to lose time rebooting my PC after running a buggy CUDA "kernel". Heck, I could run the GCC C testsuite on the GPU without worrying about locking myself and others from the server. But now when I develop on a laptop with AMD's latest mobile SoC, every time I make a mistake in my GLSL code it more often than not forces a reboot. I hope you understand what a huge pain it is.
What are the existing GPU hardware capabilities for memory protection (both in terms of preventing random accesses to system memory like with an IOMMU, and in terms of isolating different process contexts from each other), and to what extend Linux DRM drivers are taking advantage of them?
Modern (or more like non-ancient at this point, basically anything which came out within the last decade) AMD GPUs have mostly perfect protection between different execution contexts (i.e. different processes normally, though it's not always a 1:1 mapping). Each context has its own virtual GPU address space and cannot access any memory which isn't mapped into that (which the kernel driver only does for memory belonging to a buffer object which the context has permission to access and has explicitly asked to be mapped into its address space).
The instability you're seeing likely isn't due to lack of memory protection but due to any of a large number of other ways a GPU can end up in a hanging state, and the drivers and wider ecosystem not being very good at recovering from that yet.