Re: [PATCH 29/31] gpu: nova-core: Hopper/Blackwell: add FSP Chain of Trust boot path

["Gary Guo" <gary@garyguo.net>]

On Sat Jan 24, 2026 at 1:42 AM GMT, John Hubbard wrote:
> On 1/23/26 5:38 PM, John Hubbard wrote:
>> On 1/21/26 8:35 AM, Gary Guo wrote:
>>> On Wed Dec 3, 2025 at 5:59 AM GMT, John Hubbard wrote:
>> ...
>>>> +        // SAFETY: fmc_full is a valid DmaObject with a contiguous allocation of size() bytes
>>>> +        // starting at start_ptr(). The slice is only used for signature extraction within this
>>>> +        // function scope while fsp_fw remains valid.
>>>> +        let fmc_full_data = unsafe {
>>>> +            core::slice::from_raw_parts(fsp_fw.fmc_full.start_ptr(), fsp_fw.fmc_full.size())
>>>> +        };
>>>
>>> The justification is week because it does not mention about the non-race nature
>>> of this, which need to be justified for a DMA allocation. If you use
>>> `CoherentAllocation::as_slice`, then this requirement would be obvious.
>>>
>>> For example:
>>>
>>>     // SAFETY: the dma buffer is not yet submitted too hardware and we are the
>>>     // unique owner at this point.
>>>     let fmc_full_data = unsafe { fsp_fw.fmc_full.as_slice(0, fsp_fw.fmc_full.size()) };
>>  
>> I see. OK, after a lot of fussing over the wording, I'm have come up
>> with this, which might be much too wordy? I'm not sure.
>> 
>>     // SAFETY: fmc_full, which contains the complete FMC ELF file, is never submitted to
>>     // hardware, so it is safe from hardware-software races. And we are the unique owner of
>>     // fsp_fw (and therefore of fsp_fw.fmc_full). (A separate buffer, fsp_fw.fmc_image, is what
>>     // gets submitted to the hardware).
>>  
>> 
>
> Oh, forget to mention that I also changed it to use .as_slice(), so:
>
>      // SAFETY: fmc_full, which contains the complete FMC ELF file, is never submitted to
>      // hardware, so it is safe from hardware-software races. And we are the unique owner of
>      // fsp_fw (and therefore of fsp_fw.fmc_full). (A separate buffer, fsp_fw.fmc_image, is what
>      // gets submitted to the hardware).
>      let fmc_full_data = unsafe { fsp_fw.fmc_full.as_slice(0, fsp_fw.fmc_full.size())? };

If they're never submitted to hardware, why are they dma objects?

Best,
Gary