Re: [PATCH 2/2] io_uring: support multiple rings to share same poll thread by specifying same cpu

[Jens Axboe <axboe@kernel.dk>]

On 10/20/20 2:23 AM, Xiaoguang Wang wrote:
> We have already supported multiple rings to share one same poll thread
> by passing IORING_SETUP_ATTACH_WQ, but it's not that convenient to use.
> IORING_SETUP_ATTACH_WQ needs users to ensure that a parent ring instance
> has already existed, that means it will require app to regulate the
> creation oder between uring instances.
> 
> Currently we can make this a bit simpler, for those rings which will
> have SQPOLL enabled and are willing to be bound to one same cpu, add a
> capability that these rings can share one poll thread by specifying
> a new IORING_SETUP_SQPOLL_PERCPU flag, then we have 3 cases
>   1, IORING_SETUP_ATTACH_WQ: if user specifies this flag, we'll always
> try to attach this ring to an existing ring's corresponding poll thread,
> no matter whether IORING_SETUP_SQ_AFF or IORING_SETUP_SQPOLL_PERCPU is
> set.
>   2, IORING_SETUP_SQ_AFF and IORING_SETUP_SQPOLL_PERCPU are both enabled,
> for this case, we'll create a single poll thread to be shared by these
> rings, and this poll thread is bound to a fixed cpu.
>   3, for any other cases, we'll just create one new poll thread for the
> corresponding ring.
> 
> And for case 2, don't need to regulate creation oder of multiple uring
> instances, we use a mutex to synchronize creation, for example, say five
> rings which all have IORING_SETUP_SQ_AFF & IORING_SETUP_SQPOLL_PERCPU
> enabled, and are willing to be bound same cpu, one ring that gets the
> mutex lock will create one poll thread, the other four rings will just
> attach themselves the previous created poll thread once they get lock
> successfully.
> 
> To implement above function, define a percpu io_sq_data array:
>     static struct io_sq_data __percpu *percpu_sqd;
> When IORING_SETUP_SQ_AFF and IORING_SETUP_SQPOLL_PERCPU are both enabled,
> we will use struct io_uring_params' sq_thread_cpu to locate corresponding
> sqd, and use this sqd to save poll thread info.

Do you have any test results?

Not quite clear to me, but if IORING_SETUP_SQPOLL_PERCPU is set, I think
it should always imply IORING_SETUP_ATTACH_WQ in the sense that it would
not make sense to have more than one poller thread that's bound to a
single CPU, for example.

> @@ -6814,8 +6819,17 @@ static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
>  	return 0;
>  }
>  
> -static void io_put_sq_data(struct io_sq_data *sqd)
> +static void io_put_sq_data(struct io_ring_ctx *ctx, struct io_sq_data *sqd)
>  {
> +	int percpu_sqd = 0;
> +
> +	if ((ctx->flags & IORING_SETUP_SQ_AFF) &&
> +	    (ctx->flags & IORING_SETUP_SQPOLL_PERCPU))
> +		percpu_sqd = 1;
> +
> +	if (percpu_sqd)
> +		mutex_lock(&sqd->percpu_sq_lock);
> +
>  	if (refcount_dec_and_test(&sqd->refs)) {
>  		/*
>  		 * The park is a bit of a work-around, without it we get

For this, and the setup, you should make it dynamic. Hence don't
allocate the percpu data etc until someone asks for it, and when the
last user of it goes away, it should go away as well.

That would make the handling of it identical to what we currently have,
and no need to special case any of this like you do above.


-- 
Jens Axboe