Re: [PATCH 1/2] gpu: nova-core: Hopper: use correct sysmem flush registers

[John Hubbard]

On 6/9/26 6:54 AM, Alexandre Courbot wrote:
> On Thu Jun 4, 2026 at 8:50 AM JST, John Hubbard wrote:
> <...>
>> diff --git a/drivers/gpu/nova-core/regs.rs b/drivers/gpu/nova-core/regs.rs
>> index 0f49c1ab83ad..a57e95140ec0 100644
>> --- a/drivers/gpu/nova-core/regs.rs
>> +++ b/drivers/gpu/nova-core/regs.rs
>> @@ -131,6 +131,22 @@ fn fmt(&self, f: &mut kernel::fmt::Formatter<'_>) -> kernel::fmt::Result {
>>          23:0    adr_63_40;
>>      }
>>  
>> +    /// Low bits of the physical system memory address used by the GPU to perform sysmembar
>> +    /// operations on Hopper (see [`crate::fb::SysmemFlush`]).
>> +    ///
>> +    /// Unlike the Ampere `NV_PFB_NISO_FLUSH_SYSMEM_ADDR` registers, which encode the address with
>> +    /// an 8-bit right-shift, the Hopper FBHUB registers take the raw address split into lower and
>> +    /// upper halves. Hardware ignores bits 7:0 of the LO register.
>> +    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_LO(u32) @ 0x00100a34 {
>> +        31:0    adr => u32;
>> +    }
>> +
>> +    /// High bits of the physical system memory address used by the GPU to perform sysmembar
>> +    /// operations on Hopper (see [`crate::fb::SysmemFlush`]).
>> +    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI(u32) @ 0x00100a38 {
>> +        19:0    adr;
>> +    }
>> +
>>      pub(crate) NV_PFB_PRI_MMU_LOCAL_MEMORY_RANGE(u32) @ 0x00100ce0 {
>>          30:30   ecc_mode_enabled => bool;
>>          9:4     lower_mag;
>> @@ -179,15 +195,16 @@ fn fmt(&self, f: &mut kernel::fmt::Formatter<'_>) -> kernel::fmt::Result {
>>          19:0    adr;
>>      }
>>  
>> -    // GB20x sysmem flush registers, relative to the FBHUB0 base. Unlike the older
>> -    // NV_PFB_NISO_FLUSH_SYSMEM_ADDR registers which encode the address with an 8-bit
>> -    // right-shift, these take the raw address split into lower and upper halves. Hardware
>> +    // GB20x sysmem flush registers, relative to the FBHUB0 base. Like the Hopper
>> +    // NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR registers, and unlike the older
>> +    // NV_PFB_NISO_FLUSH_SYSMEM_ADDR registers (which encode the address with an 8-bit
>> +    // right-shift), these take the raw address split into lower and upper halves. Hardware
>>      // ignores bits 7:0 of the LO register.
>> -    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_LO(u32) @ Fbhub0Base + 0x00001d58 {
>> +    pub(crate) NV_PFB_FBHUB0_PCIE_FLUSH_SYSMEM_ADDR_LO(u32) @ Fbhub0Base + 0x00001d58 {
>>          31:0    adr => u32;
>>      }
>>  
>> -    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI(u32) @ Fbhub0Base + 0x00001d5c {
>> +    pub(crate) NV_PFB_FBHUB0_PCIE_FLUSH_SYSMEM_ADDR_HI(u32) @ Fbhub0Base + 0x00001d5c {
> 
> Since we are switching the name to `FBHUB0`, wouldn't it make sense to
> make this an absolute register as well? Because now this reads like a
> FBHUB0 register using Fbhub0 as its base, which is redundant.
> 
> I know I am the one who suggested making it relative, but after looking
> at OpenRM again it doesn't really seem to make sense - there is really
> only one base here, and the register is defined as an absolute one
> anyway. So my bad for suggesting this in the first place.
> 
> If we do this, maybe split into two patches:
> 
> - Rename `NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI` into
>   `NV_PFB_FBHUB0_PCIE_FLUSH_SYSMEM_ADDR_HI` and make it absolute to
>   align with OpenRM,
> - Introduce Hopper's `NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI` and use it.
> 
> This keeps the changes local and simple, as renaming makes things harder
> to track.

Ok yes, I'll post a v2 with the above approach.

> 
> As a side-note (not for this patch), we may want to move registers move
> aggressively into sub-modules. The Hopper register seems to be only used
> on Hopper, so it would make sense to me to move it into a `gh100`
> submodule when we eventually move these definitions under `fb/regs.rs`
> as is the plan.

Ack, some register organization is definitely called for, and that's
a solid first step.

> 
> Or maybe we can have an even more aggressive policy (which iirc has
> already been mentioned on another patchset) of defining the registers
> directly into the HALs themselves. Subsystems like `fb` only ever access
> their registers in the HALs, so for them at least that could make a lot
> of sense.

Perhaps, let's see.

thanks,
-- 
John Hubbard