Re: [PATCH 1/2] Revert "arm64: tegra: Disable ISO SMMU for Tegra194"

[Mikko Perttunen <mperttunen@nvidia.com>]

On Thursday, January 22, 2026 7:22 PM Mikko Perttunen wrote:
> On Tuesday, December 9, 2025 1:21 PM Aaron Kling wrote:
> > On Mon, Nov 3, 2025 at 12:05 PM Aaron Kling <webgeek1234@gmail.com> wrote:
> > >
> > > On Mon, Nov 3, 2025 at 5:07 AM Thierry Reding <thierry.reding@gmail.com> wrote:
> > > >
> > > > On Sat, Nov 01, 2025 at 06:13:26PM -0500, Aaron Kling wrote:
> > > > > On Sat, Nov 1, 2025 at 6:01 PM Aaron Kling via B4 Relay
> > > > > <devnull+webgeek1234.gmail.com@kernel.org> wrote:
> > > > > >
> > > > > > From: Aaron Kling <webgeek1234@gmail.com>
> > > > > >
> > > > > > This reverts commit ebea268ea583ba4970df425dfef8c8e21d0a4e12.
> > > > > >
> > > > > > Mmu is now being enabled for the display controllers.
> > > > > >
> > > > > > Signed-off-by: Aaron Kling <webgeek1234@gmail.com>
> > > > > > ---
> > > > > >  arch/arm64/boot/dts/nvidia/tegra194.dtsi | 2 +-
> > > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > >
> > > > > > diff --git a/arch/arm64/boot/dts/nvidia/tegra194.dtsi b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
> > > > > > index 1399342f23e1c4f73b278adc66dfb948fc30d326..854ed6d46aa1d8eedcdfbae1fdde1374adf40337 100644
> > > > > > --- a/arch/arm64/boot/dts/nvidia/tegra194.dtsi
> > > > > > +++ b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
> > > > > > @@ -1807,7 +1807,7 @@ iommu@10000000 {
> > > > > >                         #iommu-cells = <1>;
> > > > > >
> > > > > >                         nvidia,memory-controller = <&mc>;
> > > > > > -                       status = "disabled";
> > > > > > +                       status = "okay";
> > > > > >                 };
> > > > > >
> > > > > >                 smmu: iommu@12000000 {
> > > > > >
> > > > > > --
> > > > > > 2.51.0
> > > > > >
> > > > > >
> > > > >
> > > > > Question for Jon as the author of the commit being reverted. The
> > > > > commit message states "we do not have a way to pass frame-buffer
> > > > > memory from the bootloader to the kernel". If I understand this
> > > > > correctly, this is talking about seamless handoff. What does this have
> > > > > to do with enabling mmu on the display controllers? Seamless does not
> > > > > work on any tegra arch as far as I'm aware, but Tegra194 is the only
> > > > > one that doesn't have mmu enabled for the dc's. But enabling mmu
> > > > > allows for better and faster memory allocation. My initial attempts to
> > > > > enable this didn't work because I tried to attach them to the main mmu
> > > > > unit, see the related freedesktop issue [0]. After noticing in the
> > > > > downstream dt that the dc's are on a separate unit, I made it work.
> > > > > And so far, it seems to work just as well as Tegra186. Then when I was
> > > > > packaging up the change to submit, I found that this had been
> > > > > explicitly disabled. But I'm not seeing why. Am I missing some
> > > > > additional factors?
> > > >
> > > > This isn't seamless handoff to the Tegra DRM driver for display, but
> > > > rather to simple-framebuffer. While this does technically work, it also
> > > > causes a spew of SMMU faults during early boot because the firmware does
> > > > not properly pass the SMMU mapping information to the kernel.
> > > >
> > > > In a nutshell what happens is that the firmware sets up the display
> > > > controller to scan out from a reserved memory region, but it does so
> > > > without involving the SMMU, so it uses physical addresses directly. When
> > > > the kernel boots and the SMMU is enabled the continued accesses from
> > > > display hardware cause SMMU faults (because there is no mapping for the
> > > > framebuffer addresses).
> > > >
> > > > That said, we did solve these issues and this may not be happening
> > > > anymore with the most recent L4T releases, so it may be okay to revert
> > > > this now. We should find out exactly which release includes all the
> > > > needed changes so that it can be referenced in the commit message. I
> > > > want to avoid people running new kernels with an old L4T release and
> > > > then seeing these errors without any reference as to why that might
> > > > suddenly happen.
> > >
> > > For reference, I have rolled back my Android usecase to use the L4T
> > > r32.7.6 bootloaders on T194 for a variety of reasons. So I am using
> > > cboot as the final bootloader and not edk2 as in L4T r34/r35. I have a
> > > pending cboot patch to support simple-framebuffer handoff, but haven't
> > > fully verified it as tegra-drm is currently unable to takeover from
> > > simplefb like openrm does for t234. But all that to say that since I
> > > no longer use r35 for t194 I don't have the setup to easily verify
> > > which point release works here and what doesn't.
> > 
> > Any further thoughts on this patch?
> > 
> > Aaron
> 
> FWIW,
> 
> looks like the edk2 patch to update iommu-addresses --
> 
> commit 6071946461389221d2314cbbae0377610b5b1f6a
> Author: Jan Bobek <jbobek@nvidia.com>
> Date:   Tue Mar 21 00:15:27 2023 +0000
> 
>     feat(NvDisplayControllerDxe): update FDT with framebuffer info
>     
>     On ready-to-boot and whenever FDT is installed, update FDT with
>     framebuffer mode information, base address and size.
>     
>     Signed-off-by: Jan Bobek <jbobek@nvidia.com>
>     Reviewed-by: Ashish Singhal <ashishsingha@nvidia.com>
> 
> is in since r36.2
> 
> $ git tag --contains 6071946461389221d2314cbbae0377610b5b1f6a | grep "^r"                                                                 
> r36.2
> r36.3.0
> r36.4.0
> r36.4.3
> r36.4.4
> r36.4.5
> r38.2
> r38.4
> 
> Not so good for T194 since r36 only supports Orin.
> 
> I'll look into getting this cherry-picked to r35.
> 
> Mikko
> 
> 

I looked into this and it appears a version of this is in r35, but it only supports T234. However, I also found that at one point, L4T bootloader configuration has been modified to place the display controllers into SMMU bypass until otherwise configured by the kernel -- which the kernel does in tegra_mc_probe_device.

I think that means there is still potential for an issue where the display continues to be on between tegra_mc_probe_device and tegradrm reconfiguring it. However, I cannot reproduce that happening -- most likely the display is being turned off before that because of a clock or power domain being turned off.

In any case, this means that we no longer need to pass the framebuffer's information to the kernel. I think it would be good to have some clarity to ensure the issue described above cannot happen, but otherwise we should be able to enable IOMMU.

Mikko