[PATCH] gpu: nova-core: require little endian

[PATCH] gpu: nova-core: require little endian

[Eliot Courtney]

The driver already assumes little endian in a lot of locations. For
example, all the code that reads RPCs out of the command queue just
directly interprets the bytes.

Make this explicit in Kconfig.

Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
The current code assumes little endian in a bunch of places. I think we
should either explicitly decide to be generic on endianness or explicitly
decide not to - having some handling sprinkled around in various
locations seems confusing to me.

I believe that currently e.g. `RUST` transitively depends on
!CPU_BIG_ENDIAN, so this is more about making the decision explicit for
nova-core rather than fixing any kind of hole.
---
 drivers/gpu/nova-core/Kconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
index a4f2380654e2..d8456f8eaa05 100644
--- a/drivers/gpu/nova-core/Kconfig
+++ b/drivers/gpu/nova-core/Kconfig
@@ -3,6 +3,7 @@ config NOVA_CORE
 	depends on 64BIT
 	depends on PCI
 	depends on RUST
+	depends on !CPU_BIG_ENDIAN
 	select AUXILIARY_BUS
 	select RUST_FW_LOADER_ABSTRACTIONS
 	default n

---
base-commit: a7a080bb4236ebe577b6776d940d1717912ff6dd
change-id: 20260406-fix-kconfig-3a059f622697

Best regards,
--  
Eliot Courtney <ecourtney@nvidia.com>

Re: [PATCH] gpu: nova-core: require little endian

[Joel Fernandes]

On 4/6/2026 2:52 AM, Eliot Courtney wrote:
> The driver already assumes little endian in a lot of locations. For
> example, all the code that reads RPCs out of the command queue just
> directly interprets the bytes.

Oh yeah, its not just the driver making assumptions, it is the hardware itself.
The Nvidia GPU architecture is little-endian (including MMU structures in VRAM).

> 
> Make this explicit in Kconfig.
> 
> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
> ---
> The current code assumes little endian in a bunch of places. I think we
> should either explicitly decide to be generic on endianness or explicitly
> decide not to - having some handling sprinkled around in various
> locations seems confusing to me.
> 
> I believe that currently e.g. `RUST` transitively depends on
> !CPU_BIG_ENDIAN, so this is more about making the decision explicit for
> nova-core rather than fixing any kind of hole.
> ---
>  drivers/gpu/nova-core/Kconfig | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
> index a4f2380654e2..d8456f8eaa05 100644
> --- a/drivers/gpu/nova-core/Kconfig
> +++ b/drivers/gpu/nova-core/Kconfig
> @@ -3,6 +3,7 @@ config NOVA_CORE
>  	depends on 64BIT
>  	depends on PCI
>  	depends on RUST
> +	depends on !CPU_BIG_ENDIAN

If we want to do this, I am Ok. Although I am also Ok with the potential
transitive Rust dependency in which case the only value of adding it in Kconfig
is documentation. But just a side note (not a rejection comment), Nouveau does
not do this, and no other GPU driver afaics either.

thanks,

-- 
Joel Fernandes

Re: [PATCH] gpu: nova-core: require little endian

[Gary Guo]

On Mon Apr 6, 2026 at 7:52 AM BST, Eliot Courtney wrote:
> The driver already assumes little endian in a lot of locations. For
> example, all the code that reads RPCs out of the command queue just
> directly interprets the bytes.
>
> Make this explicit in Kconfig.
>
> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
> ---
> The current code assumes little endian in a bunch of places. I think we
> should either explicitly decide to be generic on endianness or explicitly
> decide not to - having some handling sprinkled around in various
> locations seems confusing to me.
>
> I believe that currently e.g. `RUST` transitively depends on
> !CPU_BIG_ENDIAN, so this is more about making the decision explicit for
> nova-core rather than fixing any kind of hole.

IBM is adding PowerPC support which will be the first BE architecture that RfL
is going to support. However, only 32-bit BE is going to be added soon, so
`depends on 64BIT` will prevent Nova from supporting that.

So I think it's good that we put it in.

Acked-by: Gary Guo <gary@garyguo.net>

Best,
Gary

> ---
>  drivers/gpu/nova-core/Kconfig | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
> index a4f2380654e2..d8456f8eaa05 100644
> --- a/drivers/gpu/nova-core/Kconfig
> +++ b/drivers/gpu/nova-core/Kconfig
> @@ -3,6 +3,7 @@ config NOVA_CORE
>  	depends on 64BIT
>  	depends on PCI
>  	depends on RUST
> +	depends on !CPU_BIG_ENDIAN
>  	select AUXILIARY_BUS
>  	select RUST_FW_LOADER_ABSTRACTIONS
>  	default n
>
> ---
> base-commit: a7a080bb4236ebe577b6776d940d1717912ff6dd
> change-id: 20260406-fix-kconfig-3a059f622697
>
> Best regards,
> --  
> Eliot Courtney <ecourtney@nvidia.com>

Re: [PATCH] gpu: nova-core: require little endian

[John Hubbard]

On 4/5/26 11:52 PM, Eliot Courtney wrote:
> The driver already assumes little endian in a lot of locations. For
> example, all the code that reads RPCs out of the command queue just
> directly interprets the bytes.

Yes, and that even understates the scope. The FromBytes-based parsing
of repr(C) structs happens in firmware loading, VBIOS parsing, and GSP
shared memory, not just command queue RPCs.

> 
> Make this explicit in Kconfig.
> 
> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
> ---
> The current code assumes little endian in a bunch of places. I think we
> should either explicitly decide to be generic on endianness or explicitly
> decide not to - having some handling sprinkled around in various
> locations seems confusing to me.
> 
> I believe that currently e.g. `RUST` transitively depends on
> !CPU_BIG_ENDIAN, so this is more about making the decision explicit for
> nova-core rather than fixing any kind of hole.

That's true today in practice, but config RUST does not directly depend
on !CPU_BIG_ENDIAN. The exclusion happens per-arch: ARM and ARM64 gate
HAVE_RUST on CPU_LITTLE_ENDIAN, and x86_64 and LoongArch are inherently
LE. 

RISC-V is more exciting because it selects HAVE_RUST without an explicit
endianness check.

So putting the dependency in nova-core directly is a good move, not just
for documentation but as a real guard.

> ---
>  drivers/gpu/nova-core/Kconfig | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
> index a4f2380654e2..d8456f8eaa05 100644
> --- a/drivers/gpu/nova-core/Kconfig
> +++ b/drivers/gpu/nova-core/Kconfig
> @@ -3,6 +3,7 @@ config NOVA_CORE
>  	depends on 64BIT
>  	depends on PCI
>  	depends on RUST
> +	depends on !CPU_BIG_ENDIAN
>  	select AUXILIARY_BUS
>  	select RUST_FW_LOADER_ABSTRACTIONS
>  	default n
> 
> ---
> base-commit: a7a080bb4236ebe577b6776d940d1717912ff6dd
> change-id: 20260406-fix-kconfig-3a059f622697
> 

Reviewed-by: John Hubbard <jhubbard@nvidia.com>

thanks,
-- 
John Hubbard

Re: [PATCH] gpu: nova-core: require little endian

[Danilo Krummrich]

On Mon Apr 6, 2026 at 8:52 AM CEST, Eliot Courtney wrote:
> The driver already assumes little endian in a lot of locations. For
> example, all the code that reads RPCs out of the command queue just
> directly interprets the bytes.
>
> Make this explicit in Kconfig.
>
> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
> ---
> The current code assumes little endian in a bunch of places. I think we
> should either explicitly decide to be generic on endianness or explicitly
> decide not to - having some handling sprinkled around in various
> locations seems confusing to me.
>
> I believe that currently e.g. `RUST` transitively depends on
> !CPU_BIG_ENDIAN, so this is more about making the decision explicit for
> nova-core rather than fixing any kind of hole.
> ---
>  drivers/gpu/nova-core/Kconfig | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
> index a4f2380654e2..d8456f8eaa05 100644
> --- a/drivers/gpu/nova-core/Kconfig
> +++ b/drivers/gpu/nova-core/Kconfig
> @@ -3,6 +3,7 @@ config NOVA_CORE
>  	depends on 64BIT
>  	depends on PCI
>  	depends on RUST
> +	depends on !CPU_BIG_ENDIAN

Since we have DRM_NOVA select NOVA_CORE, this has to be added to DRM_NOVA as
well, otherwise you could forcefully select NOVA_CORE through DRM_NOVA without
considering !CPU_BIG_ENDIAN.

Note DRM_NOVA intentionally has select NOVA_CORE (instead of depends on), since
otherwise a user would need to enable NOVA_CORE in order to see DRM_NOVA in the
menuconfig in the first place.

>  	select AUXILIARY_BUS
>  	select RUST_FW_LOADER_ABSTRACTIONS
>  	default n
>
> ---
> base-commit: a7a080bb4236ebe577b6776d940d1717912ff6dd
> change-id: 20260406-fix-kconfig-3a059f622697
>
> Best regards,
> --  
> Eliot Courtney <ecourtney@nvidia.com>

Re: [PATCH] gpu: nova-core: require little endian

[Eliot Courtney]

On Mon Apr 6, 2026 at 3:52 PM JST, Eliot Courtney wrote:
> The driver already assumes little endian in a lot of locations. For
> example, all the code that reads RPCs out of the command queue just
> directly interprets the bytes.
>
> Make this explicit in Kconfig.
>
> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
> ---
> The current code assumes little endian in a bunch of places. I think we
> should either explicitly decide to be generic on endianness or explicitly
> decide not to - having some handling sprinkled around in various
> locations seems confusing to me.
>
> I believe that currently e.g. `RUST` transitively depends on
> !CPU_BIG_ENDIAN, so this is more about making the decision explicit for
> nova-core rather than fixing any kind of hole.
> ---
>  drivers/gpu/nova-core/Kconfig | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/nova-core/Kconfig b/drivers/gpu/nova-core/Kconfig
> index a4f2380654e2..d8456f8eaa05 100644
> --- a/drivers/gpu/nova-core/Kconfig
> +++ b/drivers/gpu/nova-core/Kconfig
> @@ -3,6 +3,7 @@ config NOVA_CORE
>  	depends on 64BIT
>  	depends on PCI
>  	depends on RUST
> +	depends on !CPU_BIG_ENDIAN
>  	select AUXILIARY_BUS
>  	select RUST_FW_LOADER_ABSTRACTIONS
>  	default n
>
> ---
> base-commit: a7a080bb4236ebe577b6776d940d1717912ff6dd
> change-id: 20260406-fix-kconfig-3a059f622697
>
> Best regards,
> --  
> Eliot Courtney <ecourtney@nvidia.com>

Thanks all. Will resend with the change to nova-drm as well.

Re: [PATCH] gpu: nova-core: require little endian

[Danilo Krummrich]

On Mon Apr 6, 2026 at 6:30 PM CEST, Joel Fernandes wrote:
> Nouveau does not do this, and no other GPU driver afaics either.

Nouveau has nvkm_device_endianness() [1], which it configured before everything
else is done.

I assume that's not an option for later GPUs?

[1] https://elixir.bootlin.com/linux/v6.19.11/source/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c#L3111

Re: [PATCH] gpu: nova-core: require little endian

[Timur Tabi]

On Tue, 2026-04-07 at 12:56 +0200, Danilo Krummrich wrote:
> On Mon Apr 6, 2026 at 6:30 PM CEST, Joel Fernandes wrote:
> > Nouveau does not do this, and no other GPU driver afaics either.
> 
> Nouveau has nvkm_device_endianness() [1], which it configured before everything
> else is done.
> 
> I assume that's not an option for later GPUs?
> 
> [1]
> https://elixir.bootlin.com/linux/v6.19.11/source/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c#L3111

Pascal is the last GPU to have NV_PMC_BOOT_1_ENDIAN0.  If nvkm_device_endianness() is called on
Turing, then that might be a bug.

Re: [PATCH] gpu: nova-core: require little endian

[Joel Fernandes]

On 4/7/2026 2:02 PM, Timur Tabi wrote:
> On Tue, 2026-04-07 at 12:56 +0200, Danilo Krummrich wrote:
>> On Mon Apr 6, 2026 at 6:30 PM CEST, Joel Fernandes wrote:
>>> Nouveau does not do this, and no other GPU driver afaics either.
>>
>> Nouveau has nvkm_device_endianness() [1], which it configured before everything
>> else is done.
>>
>> I assume that's not an option for later GPUs?
>>
>> [1]
>> https://elixir.bootlin.com/linux/v6.19.11/source/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c#L3111
> 
> Pascal is the last GPU to have NV_PMC_BOOT_1_ENDIAN0.  If nvkm_device_endianness() is called on
> Turing, then that might be a bug.

Yes, this was on old/legacy GPUs, not GSP-based ones.

thanks,

--
Joel Fernandes

Claude review: gpu: nova-core: require little endian

[Claude Code Review Bot]

Overall Series Review

Subject: gpu: nova-core: require little endian
Author: Eliot Courtney <ecourtney@nvidia.com>
Patches: 9
Reviewed: 2026-04-12T14:47:44.840568

---

This is a single-patch series adding `depends on !CPU_BIG_ENDIAN` to the nova-core Kconfig. The change is **correct in intent and low-risk**. The nova-core driver does assume little-endian byte order in multiple places — both via explicit `from_le_bytes()` calls (in `falcon.rs`, `vbios.rs`) and, more critically, via `FromBytes` / `from_bytes_copy_prefix()` patterns that reinterpret raw byte slices as structs without any endianness conversion (e.g., `PcirStruct`, `BitHeader`, `NpdeStruct`, `PmuLookupTableHeader` in `vbios.rs`). The GSP firmware command queue code also directly interprets byte buffers.

The commit message's claim that "RUST transitively depends on !CPU_BIG_ENDIAN" is practically true today but slightly imprecise — on arm64, `CPU_BIG_ENDIAN` depends on `BROKEN`, so it can't actually be selected, but it's not a hard Kconfig dependency of `RUST` itself. This makes the patch more of a "belt and suspenders" measure, which is fine — it makes the requirement explicit and future-proof in case the arm64 big-endian situation ever changes.

The patch is simple, correct, and harmless. **Recommend accepting.**

---

---
Generated by Claude Code Patch Reviewer

Claude review: gpu: nova-core: require little endian

[Claude Code Review Bot]

Patch Review

**Verdict: Looks good.**

The one-line diff is straightforward:

```
 	depends on 64BIT
 	depends on PCI
 	depends on RUST
+	depends on !CPU_BIG_ENDIAN
 	select AUXILIARY_BUS
```

**Observations:**

1. **Placement is correct.** The new `depends on` line is placed after the existing `depends on` lines and before the `select` lines, following standard Kconfig ordering conventions.

2. **The LE assumption is real.** The driver uses `FromBytes` (which does native-endian reinterpretation of byte slices) on several packed structs in `vbios.rs:332,392,550,854`. It also uses explicit `u32::from_le_bytes()` / `u16::from_le_bytes()` in `falcon.rs:432,461` and `vbios.rs:445,449,490,502,510,515,795,840`. The mixture of explicit LE conversion and raw byte reinterpretation confirms that the driver would not work correctly on a big-endian system — some data would be byte-swapped and some wouldn't.

3. **The commit message's RUST claim is slightly imprecise.** The message says "RUST transitively depends on !CPU_BIG_ENDIAN" — this isn't strictly true via Kconfig dependency chains. What is true is that on arm64 (the only 64-bit arch where `CPU_BIG_ENDIAN` exists), `CPU_BIG_ENDIAN` currently `depends on BROKEN`, making it impossible to select in practice. The patch is still valuable because it makes the requirement explicit rather than relying on an indirect and potentially fragile transitive constraint. The below-the-cut explanation acknowledges this nuance, which is good.

4. **Minor nit on the commit message:** The phrase "all the code that reads RPCs out of the command queue just directly interprets the bytes" could be more precise — the vbios parsing code with `FromBytes` is arguably a better example of implicit LE assumptions than the RPC code.

No blocking issues. The patch is a single-line, obviously-correct Kconfig dependency addition.

---
Generated by Claude Code Patch Reviewer