Re: io_uring_prep_timeout_update on linked timeouts

[Jens Axboe <axboe@kernel.dk>]

On 8/28/21 7:39 AM, Pavel Begunkov wrote:
> On 8/28/21 4:22 AM, Jens Axboe wrote:
>> On 8/26/21 7:40 PM, Victor Stewart wrote:
>>> On Wed, Aug 25, 2021 at 2:27 AM Victor Stewart <v@nametag.social> wrote:
>>>>
>>>> On Tue, Aug 24, 2021 at 11:43 PM Victor Stewart <v@nametag.social> wrote:
>>>>>
>>>>> we're able to update timeouts with io_uring_prep_timeout_update
>>>>> without having to cancel
>>>>> and resubmit, has it ever been considered adding this ability to
>>>>> linked timeouts?
>>>>
>>>> whoops turns out this does work. just tested it.
>>>
>>> doesn't work actually. missed that because of a bit of misdirection.
>>> returns -ENOENT.
>>>
>>> the problem with the current way of cancelling then resubmitting
>>> a new a timeout linked op (let's use poll here) is you have 3 situations:
>>>
>>> 1) the poll triggers and you get some positive value. all good.
>>>
>>> 2) the linked timeout triggers and cancels the poll, so the poll
>>> operation returns -ECANCELED.
>>>
>>> 3) you cancel the existing poll op, and submit a new one with
>>> the updated linked timeout. now the original poll op returns
>>> -ECANCELED.
>>>
>>> so solely from looking at the return value of the poll op in 2) and 3)
>>> there is no way to disambiguate them. of course the linked timeout
>>> operation result will allow you to do so, but you'd have to persist state
>>> across cqe processings. you can also track the cancellations and know
>>> to skip the explicitly cancelled ops' cqes (which is what i chose).
>>>
>>> there's also the problem of efficiency. you can imagine in a QUIC
>>> server where you're constantly updating that poll timeout in response
>>> to idle timeout and ACK scheduling, this extra work mounts.
>>>
>>> so i think the ability to update linked timeouts via
>>> io_uring_prep_timeout_update would be fantastic.
>>
>> Hmm, I'll need to dig a bit, but whether it's a linked timeout or not
>> should not matter. It's a timeout, it's queued and updated the same way.
>> And we even check this in some of the liburing tests.
> 
> We don't keep linked timeouts in ->timeout_list, so it's not
> supported and has never been. Should be doable, but we need
> to be careful synchronising with the link's head.

Yeah shoot you are right, I guess that explains the ENOENT. Would be
nice to add, though. Synchronization should not be that different from
dealing with regular timeouts.

-- 
Jens Axboe