On Sat Jun 20, 2026 at 11:26 PM JST, Danilo Krummrich wrote: > On Sat Jun 20, 2026 at 4:06 PM CEST, Alexandre Courbot wrote: >> I expect that the users of the VRAM regions (i.e. the MM subsystem) will >> not be able to contruct properly if no region is detected, and that sounds >> like the right place to raise an error rather than making an assumption >> here that no VRAM region is an error. For instance, I am not sure what >> the regions layout will be on Tegra systems, although your comment about >> OpenRM seems to hint that there are regions defined for them as well. > > Well, the question is whether it would just be an assumption that no VRAM > region > is an error. I'm not entirely sure, but I think it generally is an error. If > that's confirmed I'd rather fail early and with a clear message, and not carry > on with something we already know won't work out.
I took a deeper look at OpenRM. The zero FB region error is emitted when mirroring the list of FB regions [1], which is a default HAL method. Zero-VRAM chips like Tegra replace it with an empty stub. So `gpu.rs` doesn't look like the right place to do that check imho - especially since the list we are checking against is the list of *usable* regions, whereas OpenRM checks the full list. The current code is just getting the list from the GSP - the code that does something meaningful with it doesn't exist yet, and it is that code that should error if there are no FB regions at all. [1] https://github.com/NVIDIA/open-gpu-kernel-modules/blob/570.148/src/nvidia/src/kernel/gpu/mem_mgr/mem_mgr_gsp_client.c#L61
