On Thu Nov 19, 2026 at 12:28 PM GMT, Danilo Krummrich wrote:
> On Thu Nov 19, 2026 at 12:28 PM CET, Gary Guo wrote:
>> On Thu Nov 19, 2026 at 10:34 AM GMT, Danilo Krummrich wrote:
>>> On Thu Nov 19, 2026 at 3:08 AM CET, Gary Guo wrote:
>>>> On Wed Nov 18, 2026 at 11:37 PM GMT, Danilo Krummrich wrote:
>>>>> The firmware, buffer, VM and MMU types use 'drm for their resource
>>>>> lifetimes, even though each parameter is local to its type.
>>>>
>>>> This doesn't make any sense, a lifetime on a type by definition cannot 
>>>> refer to
>>>> a scope of itself.
>>>>
>>>> `'sec` is not the lifetime of `Section` and `'fw` is not the lifetime of
>>>> `Firmware` and I think this is just confusing.
>>>
>>> Yes, this was badly phrased. What is actually meant is that e.g. 'fw 
>>> constrains
>>> how long a Firmware instance can be used.
>>
>> I think this is weird still. Any lifetime on a type would constrain how long 
>> an
>> instance of that type can be used, that's just type wellformness.
>>
>> I think what you're saying is basically "I couldn't think of a good name so I
>> use the name of the type instead". To me I think using `'a` has the same 
>> effect
>> and is less confusing because reader knows that it is a generic name and is 
>> not
>> tied to any specific scope.
>
> Oh, I agree, 'a technically achieves the same thing.
>
> With self-referential pin-init, several lifetime roles appear together. In the
> Firmware example below, 'fw describes its external borrows, 'res is the 
> (hidden)
> invariant device resource lifetime, and 'vm describes the borrow of the 
> embedded
> VM.

It's true that many lifetime names would appear with self-referential pin-init,
but apart from invariant lifetimes, the covariant lifetime parameters are almost
always the shortest living lifetime among things that it may reference.

>
>       pub(crate) struct Firmware<'fw>
>       where
>           exists<'res>: 'fw,
>       {
>           /// Iomem need to access registers.
>           iomem: &'fw IoMem<'fw>,
>       
>           /// List of firmware sections.
>           #[uses('res: invariant)]
>           sections: KVec<Section<'vm, 'res>>,
>       
>           /// MCU VM.
>           #[pin]
>           #[uses('res: invariant)]
>           vm: Vm<'fw, 'res>,
>       }
>
>       pub(crate) struct Vm<'vm, 'res> {
>           as_data: Arc<VmAsData<'res>>,
>           mmu: &'vm Mmu<'res>,
>           dev: &'res Device<Bound>,
>           #[pin]
>           gpuvm_unique: Mutex<UniqueRefGpuVm<GpuVmData<'res>>>,
>           gpuvm: ARef<GpuVm<GpuVmData<'res>>>,
>           va_range: Range<u64>,
>       }
>
> I find naming the external borrow parameter after the containing type easier 
> to
> follow when reading these definitions. 

In my view it's the exact opposite. If this is called `'a` then it's 100%
external borrow (surely nobody would call their field `'a`?) while otherwise I'd
need to see if it's a field or a lifetime parameter.

>
> I thought we did settle on this convention in the past; am I misremembering?

I don't think there's an established convention. I know this is the pattern that
Nova is using, and I didn't question that. But I am wary of having this becoming
established as a standard pattern.

If a rule is to be established, I'd rather it being the following:

- for invariant lifetime parameter, use the name of the borrowed resource
  (because you know the exact lifetime).

- for covariant lifetime parameter (especially if multiple resources are
  borrowed and the lifetime is the shortest among those), pick a generic name

- for covariant lifetime parameter referencing just a single resource, I think
  either is fine -- (as long as it's unlikely to be used by unsafe code that may
  assume that it covers the full lifetime, e.g. 'bound).

Best,
Gary

>
>>> (In the following patch, in TyrDrmRegistrationData, Firmware borrows Mmu. 
>>> That
>>> borrow has its own lifetime, distinct from TyrDrmRegistrationData's 'drm
>>> parameter.)


Reply via email to