Re: Sending CQE to a different ring

[Pavel Begunkov <asml.silence@gmail.com>]

On 3/10/22 03:00, Jens Axboe wrote:
> On 3/9/22 7:11 PM, Artyom Pavlov wrote:
>> 10.03.2022 04:36, Jens Axboe wrote:
>>> On 3/9/22 4:49 PM, Artyom Pavlov wrote:
>>>> Greetings!
>>>>
>>>> A common approach for multi-threaded servers is to have a number of
>>>> threads equal to a number of cores and launch a separate ring in each
>>>> one. AFAIK currently if we want to send an event to a different ring,
>>>> we have to write-lock this ring, create SQE, and update the index
>>>> ring. Alternatively, we could use some kind of user-space message
>>>> passing.
>>>>
>>>> Such approaches are somewhat inefficient and I think it can be solved
>>>> elegantly by updating the io_uring_sqe type to allow accepting fd of a
>>>> ring to which CQE must be sent by kernel. It can be done by
>>>> introducing an IOSQE_ flag and using one of currently unused padding
>>>> u64s.
>>>>
>>>> Such feature could be useful for load balancing and message passing
>>>> between threads which would ride on top of io-uring, i.e. you could
>>>> send NOP with user_data pointing to a message payload.
>>>
>>> So what you want is a NOP with 'fd' set to the fd of another ring, and
>>> that nop posts a CQE on that other ring? I don't think we'd need IOSQE
>>> flags for that, we just need a NOP that supports that. I see a few ways
>>> of going about that:
>>>
>>> 1) Add a new 'NOP' that takes an fd, and validates that that fd is an
>>>      io_uring instance. It can then grab the completion lock on that ring
>>>      and post an empty CQE.
>>>
>>> 2) We add a FEAT flag saying NOP supports taking an 'fd' argument, where
>>>      'fd' is another ring. Posting CQE same as above.
>>>
>>> 3) We add a specific opcode for this. Basically the same as #2, but
>>>      maybe with a more descriptive name than NOP.
>>>
>>> Might make sense to pair that with a CQE flag or something like that, as
>>> there's no specific user_data that could be used as it doesn't match an
>>> existing SQE that has been issued. IORING_CQE_F_WAKEUP for example.
>>> Would be applicable to all the above cases.
>>>
>>> I kind of like #3 the best. Add a IORING_OP_RING_WAKEUP command, require
>>> that sqe->fd point to a ring (could even be the ring itself, doesn't
>>> matter). And add IORING_CQE_F_WAKEUP as a specific flag for that.
>>>
>>
>> No, ideally I would like to be able to send any type of SQE to a
>> different ring. For example, if I see that the current ring is
>> overloaded, I can create exactly the same SQEs as during usual
>> operation, but with a changed recipient ring.
>>
>> Your approach with a new "sendable" NOP will allow to emulate it in
>> user-space, but it will involve unnecessary ring round-trip and will
>> be a bit less pleasant in user code, e.g. we would need to encode a
>> separate state "the task is being sent to a different ring" instead of
>> simply telling io-uring "read data and report CQE on this ring"
>> without any intermediate states.
> 
> OK, so what you're asking is to be able to submit an sqe to ring1, but
> have the completion show up in ring2? With the idea being that the rings
> are setup so that you're basing this on which thread should ultimately
> process the request when it completes, which is why you want it to
> target another ring?
> 
> It'd certainly be doable, but it's a bit of a strange beast. My main
> concern with that would be:
> 
> 1) It's a fast path code addition to every request, we'd need to check
>     some new field (sqe->completion_ring_fd) and then also grab a
>     reference to that file for use at completion time.
> 
> 2) Completions are protected by the completion lock, and it isn't
>     trivial to nest these. What happens if ring1 submits an sqe with
>     ring2 as the cqe target, and ring2 submits an sqe with ring1 as the
>     cqe target? We can't safely nest these, as we could easily introduce
>     deadlocks that way.
> 
> My knee jerk reaction is that it'd be both simpler and cheaper to
> implement this in userspace... Unless there's an elegant solution to it,
> which I don't immediately see.

Per request fd will be ugly and slow unfortunately. As people asked about
a similar thing before, the only thing I can suggest is to add a way
to pass another SQ. The execution will be slower, but at least can be
made zero overhead for the normal path.

-- 
Pavel Begunkov