Re: [PATCH 3/3] drm/nouveau/tegra: Keep the fuse clock enabled while the GPU is powered

[Thierry Reding <thierry.reding@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 2456 bytes --]

On Tue, Oct 13, 2026 at 10:52:10AM -0400, Sasha Levin wrote:
> Since the GPU on the Jetson TX1 was enabled in commit d1b725476104
> ("arm64: tegra: p2180: Explicitly enable GPU"), nouveau often fails to
> bring up the GM20B. Either BOOT_0 reads back garbage ("unknown chipset
> (5ffefa47)"), or PRI reads of the GPC PLL time out ("MMIO read of
> 00000000 FAULT at 137000 [ TIMEOUT ]"), or the first register reads in
> nvkm_device_ctor() stall until an asynchronous SError. How often depends
> on the board; one board fails on every boot.
> 
> nouveau powers the GPU up with the fuse clock gated: since commit
> 24a15252ff04 ("soc/tegra: fuse: Add runtime PM support"), the Tegra fuse
> driver only enables that clock around its own fuse reads. NVIDIA's
> downstream nvgpu driver enables the fuse clock along with the other GPU
> clocks whenever the GPU is powered. Do the same: take an optional "fuse"
> clock and keep it enabled from before the GPU rail is turned on until
> after it is turned off.
> 
> Tested on eight Jetson TX1 devkits, on mainline and on v6.18. Without
> this patch, the GPU failed to come up in 27 of 109 boots; with it, in 11
> of 291, none of them with a garbage BOOT_0. On the board that failed on
> every boot, 38 of 40 boots now succeed. The remaining failures are
> mostly a register read that stalls until an SError; wherever the console
> shows when, nouveau had already finished probing. Such stalls also occur
> without this patch and are a separate problem. The Jetson Nano, which
> shares the Tegra210 GPU node, still works (24 of 24 boots).
> 
> Fixes: d1b725476104 ("arm64: tegra: p2180: Explicitly enable GPU")
> Assisted-by: LLM
> Signed-off-by: Sasha Levin <sashal@kernel.org>
> ---
>  drivers/gpu/drm/nouveau/include/nvkm/core/tegra.h  |  1 +
>  drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 13 +++++++++++++
>  2 files changed, 14 insertions(+)

Given the failure rates, I'm almost tempted to say we need to break DT
ABI and consider the FUSE clock to be required, in which case the driver
should use the regular devm_clk_get() instead of the optional variant.

As you said, we probably got away with this for a long time because we
either didn't enabled the GPU by default or because the FUSE clock was
always turned on. Or we might've tested GPU as built-in, in which case
the FUSE clock might have still been on (not disabled after early boot
yet).

Thierry

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]