On 3/21/23 7:16?PM, Gabriel Krisman Bertazi wrote:
> Hi Jens,
>
> This tides up the io-wq internal interface by dropping the io_wqe/io_wq
> separation, which no longer makes sense since commit
> 0654b05e7e65 ("io_uring: One wqe per wq"). We currently have a single
> io_wqe instance per io_wq, which is embedded in the structure. This
> patchset merges the two, dropping bit of code to go from one to the
> other in the io-wq implementation.
>
> I don't expect it to have any positive impact on performance, of course,
> since hopefully the compiler optimizes it, but still, it is nice clean
> up. To be sure, I measured with some mmtests microbenchmarks and I haven't
> seen differences with or without the patchset.
>
> Patch 2 is slightly big to review but the use of wq and wqe is
> intrinsically connected; it was a bit hard to break it in more pieces.
>
> Tested by running liburing's testsuite and mmtests performance
> microbenchmarks (which uses fio).
>
> Based on your for-next branch.
Nice! This is a great continuation of getting rid of the per-node worker
setup.
--
Jens Axboe
