Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 16 Jan 2020 17:26:30 +0100
From:   Stefano Garzarella <sgarzare@redhat.com>
To:     Jens Axboe <axboe@kernel.dk>
Cc:     Alexander Viro <viro@zeniv.linux.org.uk>, io-uring@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH] io_uring: wakeup threads waiting for EPOLLOUT events
Message-ID: <20200116162630.6r3xc55kdyyq5tvz@steredhat>
References: <20200116134946.184711-1-sgarzare@redhat.com>
 <2d2dda92-3c50-ee62-5ffe-0589d4c8fc0d@kernel.dk>
 <20200116155557.mwjc7vu33xespiag@steredhat>
 <5723453a-9326-e954-978e-910b8b495b38@kernel.dk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <5723453a-9326-e954-978e-910b8b495b38@kernel.dk>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Jan 16, 2020 at 09:00:24AM -0700, Jens Axboe wrote:
> On 1/16/20 8:55 AM, Stefano Garzarella wrote:
> > On Thu, Jan 16, 2020 at 08:29:07AM -0700, Jens Axboe wrote:
> >> On 1/16/20 6:49 AM, Stefano Garzarella wrote:
> >>> io_uring_poll() sets EPOLLOUT flag if there is space in the
> >>> SQ ring, then we should wakeup threads waiting for EPOLLOUT
> >>> events when we expose the new SQ head to the userspace.
> >>>
> >>> Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> >>> ---
> >>>
> >>> Do you think is better to change the name of 'cq_wait' and 'cq_fasync'?
> >>
> >> I honestly think it'd be better to have separate waits for in/out poll,
> >> the below patch will introduce some unfortunate cacheline traffic
> >> between the submitter and completer side.
> > 
> > Agree, make sense. I'll send a v2 with a new 'sq_wait'.
> > 
> > About fasync, do you think could be useful the POLL_OUT support?
> > In this case, maybe is not simple to have two separate fasync_struct,
> > do you have any advice?
> 
> The fasync should not matter, it's all in the checking of whether the sq
> side has any sleepers. This is rarely going to be the case, so as long
> as we can keep the check cheap, then I think we're fine.

Right.

> 
> Since the use case is mostly single submitter, unless you're doing
> something funky or unusual, you're not going to be needing POLLOUT ever.

The case that I had in mind was with kernel side polling enabled and
a single submitter that can use epoll() to wait free slots in the SQ
ring. (I don't have a test, maybe I can write one...)

> Hence I don't want to add any cost for it, I'd even advocate just doing
> waitqueue_active() perhaps, if we can safely pull it off.

I'll try!

Thanks,
Stefano