Re: [PATCH] rust/revocable: add try_with() convenience method

["Alexandre Courbot" <acourbot@nvidia.com>]

On Sat Mar 15, 2025 at 11:17 PM JST, Danilo Krummrich wrote:
> On Sat, Mar 15, 2025 at 11:07:44PM +0900, Alexandre Courbot wrote:
>> On Fri Mar 14, 2025 at 2:50 AM JST, Benno Lossin wrote:
>> > On Thu Mar 13, 2025 at 4:48 PM CET, Danilo Krummrich wrote:
>> >> On Thu, Mar 13, 2025 at 03:38:55PM +0000, Benno Lossin wrote:
>> >>> On Thu Mar 13, 2025 at 4:08 PM CET, Alexandre Courbot wrote:
>> >>> > On Thu Mar 13, 2025 at 11:19 PM JST, Benno Lossin wrote:
>> >>> >> Would it make sense to not use `Result` here and continue with `Option`?
>> >>> >
>> >>> > We would have to return an Option<Result<R>> in this case. The current
>> >>> > code folds the closure's Result into the one of the guard's acquisition
>> >>> > for convenience.
>> >>> >
>> >>> > Actually, I don't think I have ever used try_access() a single time
>> >>> > without converting its returned Option into a Result. Wouldn't it make
>> >>> > sense to do the opposite, i.e. make try_access() return Err(ENXIO) when
>> >>> > the guard cannot be acquired and document this behavior?
>> >>> 
>> >>> Sure, if you're always doing
>> >>> 
>> >>>     let guard = rev.try_access().ok_or(ENXIO)?;
>> >>> 
>> >>> Then it makes sense from my view, maybe Danilo has some other argument
>> >>> for why `Option` is better.
>> >>
>> >> Most of the time I think we indeed want to derive an Err() if try_access()
>> >> fails, but not with a specific error code. The error code depends on the context
>> >> of where the revocable is used (e.g. for I/O mappings), but it also depends on
>> >> the driver semantics.
>> >
>> > In that case a single function with this signature would make sense:
>> >
>> >     fn access_with<R>(&self, f: impl FnOnce(&T) -> R) -> Option<R>;
>> >
>> > If there are common usages that always return the same error code, then
>> > we could add them as functions with `Result`.
>> 
>> Yeah the more I think about it the more this seems to make sense,
>> from a strictly logical point of view.
>> 
>> Where I am still on the fence is that the goal is also to reduce the
>> friction introduced by the Revocable business, which a large driver
>> might need to interact with hundreds of times. If the user wants the
>> callback to return a Result, then this method will return an
>> Option<Result>. One would then need to ok_or the Option, then flatten
>> the two results, which is a bit verbose.
>
> I think you see this from the perspective of one specific usecase, i.e.
> Devres<T>, where T dereferences to Io, right?

Indeed.

>
>> I suppose drivers could add their own macros to do that automatically
>> and reduce code verbosity, at the cost of less cohesion across drivers.
>> Guess I'll go with that if I cannot come with anything better.
>
> Maybe we could do something more specific but yet generic on top (for the
> use-case above), but we still can't assume the exact error code a driver wants
> to derive from failing try_access(). So, maybe a driver specific wrapper is
> indeed what you want on top of what this patch provides.

So be it! It's not that bad and more flexible in the end.