Re: [PATCH] io_uring: allocate the two rings together

[Jens Axboe]

On 8/27/19 1:35 PM, Hristo Venev wrote:
> Sorry for the duplicate reply, I forgot to CC the mailing list.
> 
> On Tue, 2019-08-27 at 09:50 -0600, Jens Axboe wrote:
>> Outside of going for a cleanup, have you observed any wins from this
>> change?
> 
> I haven't ran any interesting benchmarks. The cp examples in liburing
> are running as fast as before, at least on x86_64.
> 
> Do you think it makes sense to tell userspace that the sq and cq mmap
> offsets now mean the same thing? We could add a flag set by the kernel
> to io_uring_params.

Not sure, there isn't a whole lot of overhead associated with having
to do two mmaps vs just one.

If you wanted to, the best approach would be to change one of the
io_uring_params reserved fields to be a feature field or something
like that. As features get added, we could flag them there. Then
the app could do:

io_uring_setup(depth, &params);
if (params.features & IORING_FEAT_SINGLE_MMAP)
        ....
else
        mmap rings individually

and so forth.

-- 
Jens Axboe