Re: [PATCH v9 7/8] rust: Add read_poll_timeout functions

[Fiona Behrens]

On 28 Jan 2025, at 7:29, FUJITA Tomonori wrote:

> On Tue, 28 Jan 2025 08:49:37 +0800
> Gary Guo <gary@xxxxxxxxxxx> wrote:
>
>> On Mon, 27 Jan 2025 15:31:47 +0900 (JST)
>> FUJITA Tomonori <fujita.tomonori@xxxxxxxxx> wrote:
>>
>>> On Mon, 27 Jan 2025 11:46:46 +0800
>>> Gary Guo <gary@xxxxxxxxxxx> wrote:
>>>
>>>>> +#[track_caller]
>>>>> +pub fn read_poll_timeout<Op, Cond, T: Copy>(
>>>>> +    mut op: Op,
>>>>> +    mut cond: Cond,
>>>>> +    sleep_delta: Delta,
>>>>> +    timeout_delta: Delta,
>>>>> +) -> Result<T>
>>>>> +where
>>>>> +    Op: FnMut() -> Result<T>,
>>>>> +    Cond: FnMut(&T) -> bool,
>>>>> +{
>>>>> +    let start = Instant::now();
>>>>> +    let sleep = !sleep_delta.is_zero();
>>>>> +    let timeout = !timeout_delta.is_zero();
>>>>> +
>>>>> +    if sleep {
>>>>> +        might_sleep(Location::caller());
>>>>> +    }
>>>>> +
>>>>> +    loop {
>>>>> +        let val = op()?;
>>>>> +        if cond(&val) {
>>>>> +            // Unlike the C version, we immediately return.
>>>>> +            // We know the condition is met so we don't need to check again.
>>>>> +            return Ok(val);
>>>>> +        }
>>>>> +        if timeout && start.elapsed() > timeout_delta {
>>>>
>>>> Re-reading this again I wonder if this is the desired behaviour? Maybe
>>>> a timeout of 0 should mean check-once instead of no timeout. The
>>>> special-casing of 0 makes sense in C but in Rust we should use `None`
>>>> to mean it instead?
>>>
>>> It's the behavior of the C version; the comment of this function says:
>>>
>>> * @timeout_us: Timeout in us, 0 means never timeout
>>>
>>> You meant that waiting for a condition without a timeout is generally
>>> a bad idea? If so, can we simply return EINVAL for zero Delta?
>>>
>>
>> No, I think we should still keep the ability to represent indefinite
>> wait (no timeout) but we should use `None` to represent this rather
>> than `Delta::ZERO`.
>>
>> I know that we use 0 to mean indefinite wait in C, I am saying that
>> it's not the most intuitive way to represent in Rust.
>>
>> Intuitively, a timeout of 0 should be closer to a timeout of 1 and thus
>> should mean "return with ETIMEDOUT immedidately" rather than "wait
>> forever".
>>
>> In C since we don't have a very good sum type support, so we
>> special case 0 to be the special value to represent indefinite wait,
>> but I don't think we need to repeat this in Rust.
>
> Understood, thanks. How about the following code?
>
> +#[track_caller]
> +pub fn read_poll_timeout<Op, Cond, T: Copy>(
> +    mut op: Op,
> +    mut cond: Cond,
> +    sleep_delta: Delta,
> +    timeout_delta: Option<Delta>,
> +) -> Result<T>
> +where
> +    Op: FnMut() -> Result<T>,
> +    Cond: FnMut(&T) -> bool,
> +{
> +    let start = Instant::now();
> +    let sleep = !sleep_delta.is_zero();
> +
> +    if sleep {
> +        might_sleep(Location::caller());
> +    }
> +
> +    loop {
> +        let val = op()?;
> +        if cond(&val) {
> +            // Unlike the C version, we immediately return.
> +            // We know the condition is met so we don't need to check again.
> +            return Ok(val);
> +        }
> +        if let Some(timeout_delta) = timeout_delta {
> +            if start.elapsed() > timeout_delta {
> +                // Unlike the C version, we immediately return.
> +                // We have just called `op()` so we don't need to call it again.
> +                return Err(ETIMEDOUT);
> +            }
> +        }
> +        if sleep {
> +            fsleep(sleep_delta);
> +        }
> +        // fsleep() could be busy-wait loop so we always call cpu_relax().
> +        cpu_relax();
> +    }
> +}

I wonder if it makes sense to then switch `Delta` to wrap a  `NonZeroI64` and forbid deltas with 0 nanoseconds with that and use the niche optimization. Not sure if we make other apis horrible by that, but this would prevent deltas that encode no time passing.

Thanks,
Fiona