Re: [PATCH 3/7] gpu: nova-core: create debugfs root in PCI init closure

[Joel Fernandes <joelagnelf@nvidia.com>]

On 12/16/2025 3:28 PM, Timur Tabi wrote:
> On Tue, 2025-12-16 at 19:04 +0900, John Hubbard wrote:
>>> If I do that, then how do I access it from gsp.rs (the last patch in this series)?
>>
>> Given that the current PCI .probe() doesn't pass in module data,
>> I looked around and saw that Binder is using global_lock!()
>> instead of passing around module data. It seems like this area
>> is still new in Rust for Linux.
>>
>> I wonder if global_lock!() is what Gary had in mind? That is still
>> effectively global access, but at least it's synchronized,
>> unlike static mut.
> 
> It's only 'mut' in that it needs to be initialized when the driver loads.  After that, it's only
> referenced as non-mut:
> 
>         #[allow(static_mut_refs)]
>         // SAFETY: `DEBUGFS_ROOT` is never modified after initialization, so it is safe to
>         // create a shared reference to it.
>         let novacore_dir = unsafe { crate::DEBUGFS_ROOT.as_ref() }.ok_or(ENOENT)?;
> 
> The first internal version of this code did have global_lock!, but it caused problems.  That's why
> Alex recommended to drop the lock and just have an unsafe initialization:
> 
> https://github.com/Gnurou/linux/commit/d5435f662b8677545a93373b4c4c80d8b4be40c9
> 
> It would be nice if Rust had a concept of a variable that was initialized at runtime once (before
> any concurrent access could occur), and was then read-only everywhere else.
> 

Could the debugfs Rust abstraction add a directory lookup function
`Dir::lookup(name)` method and allow creating sub directories under a known
path? This would let `probe()` find the module-created root directory without
needing global state right? Then store references to this on a per-GPU basis
that the GSP code can access. Alternatively, the first probe can create the root
directory, and subsequent probes reuse it without requiring module-level root.

thanks,

 - Joel