On Tue, Sep 9, 2025 at 7:05 AM Thierry Reding <[email protected]> wrote: > > On Sat, Sep 06, 2025 at 08:01:27PM -0500, Aaron Kling wrote: > [...] > > I should note that I have only actively tested this on gm20b and > > gp10b. I am currently unable to get nouveau running on any gk20a > > device I own. I am trying to target consumer devices such as the > > shield tablet, which use the android bootloader. I can boot the kernel > > just fine and tegra-drm works as well, however when nouveau tries to > > probe, it hangs the first time it tries to access a register. I have > > not yet been able to figure out why. > > These types of hangs typically indicate that the hardware is not powered > on, not clocked or in reset. It's odd that it would hang during register > access on gk20a because nothing significant has changed in any of the > related drivers, as far as I know. > > One thing that you could try is passing the clk_ignore_unused and > pd_ignore_unused command-line parameters when booting the kernel. If it > works with those, try finding out which one of them is causing things to > break to narrow down what we need to fix.
The ignore_unused params didn't make a difference, but this got me looking closer at the power rails. I noticed that vdd-gpu for the jetson-tk1 is set to always-on for mainline, which it is not on the downstream kernel. Mirroring that to ardbeg resulted in nouveau successfully probing. One of my other devices needed a pwm-regulator for vdd-gpu, after which it also probes as expected. I will need to double check the rest of my devices as well. Is there a known reason why if vdd-gpu is a pmic regulator, it needs to be always-on, or is this another unsolved mystery? This change looks to be working okay on gk20a. It does a couple transitions on startup. I'm having other issues getting full rendering started, though. Kasan slab-out-of-bounds in nvkm_falcon_v1_load_imem when drm_hwcomposer tries to start up. Looking into that separately now, but that shouldn't block any of my open patches. Aaron
