Re: [PATCH] io_uring: undeprecate epoll_ctl support

From: Jens Axboe <axboe@kernel.dk>
To: Ben Noordhuis <info@bnoordhuis.nl>,
	Pavel Begunkov <asml.silence@gmail.com>
Cc: io-uring@vger.kernel.org
Subject: Re: [PATCH] io_uring: undeprecate epoll_ctl support
Date: Wed, 3 May 2023 06:49:53 -0600	[thread overview]
Message-ID: <dcd3d63e-c8ce-935e-3b43-5b6ca5c84eb8@kernel.dk> (raw)
In-Reply-To: <CAHQurc9L-noiMjvFsXghaBoVEBs7KJ5-a4t-eRvRim0=5HuW8w@mail.gmail.com>

On 5/3/23 2:58?AM, Ben Noordhuis wrote:
> On Tue, May 2, 2023 at 2:51?PM Pavel Begunkov <asml.silence@gmail.com> wrote:
>>
>> On 5/1/23 19:52, Ben Noordhuis wrote:
>>> Libuv recently started using it so there is at least one consumer now.
>>
>> It was rather deprecated because io_uring controlling epoll is a bad
>> idea and should never be used. One reason is that it means libuv still
>> uses epoll but not io_uring, and so the use of io_uring wouldn't seem
>> to make much sense. You're welcome to prove me wrong on that, why libuv
>> decided to use a deprecated API in the first place?
>> Sorry, but the warning is going to stay and libuv should revert the use
>> of epol_ctl requests.
> 
> Why use a deprecated API? Because it was only recently deprecated.
> Distro kernels don't warn about it yet. I only found out because of
> kernel source code spelunking.
> 
> Why combine io_uring and epoll? Libuv uses level-triggered I/O for
> reasons (I can go into detail but they're not material) so it's very
> profitable to batch epoll_ctl syscalls; it's the epoll_ctlv() syscall
> people have been asking for since practically forever.
> 
> Why not switch to io_uring wholesale? Libuv can't drop support for
> epoll because of old kernels, and io_uring isn't always clearly faster
> than epoll in the first place.
> 
> As to the warning: according to the commit that introduced it, it was
> added because no one was using IORING_OP_EPOLL_CTL. Well, now someone
> is using it. Saying it's a bad API feels like post-hoc
> rationalization. I kindly ask you merge this patch. I'd be happy to
> keep an eye on io_uring/epoll.c if you're worried about maintenance
> burden.

This is obviously mostly our fault, as the deprecation patch should've
obviously been backported to stable. Just adding it to the current
kernel defeated the purpose, as it added a long period where older
kernels quite happily accepted epoll use cases.

So I do agree, the only sane course of action here is to un-deprecate
it.

-- 
Jens Axboe