Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <2098a505-3668-30cb-e3b6-2111a703b27e@kernel.dk>
Date:   Mon, 26 Sep 2022 11:30:07 -0600
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux aarch64; rv:102.0) Gecko/20100101
 Thunderbird/102.2.2
Subject: Re: [PATCH v2 0/3] io_uring: register single issuer task at creation
Content-Language: en-US
To:     Dylan Yudaken <dylany@fb.com>,
        Pavel Begunkov <asml.silence@gmail.com>
Cc:     io-uring@vger.kernel.org, linux-kernel@vger.kernel.org,
        kernel-team@fb.com
References: <20220926170927.3309091-1-dylany@fb.com>
From:   Jens Axboe <axboe@kernel.dk>
In-Reply-To: <20220926170927.3309091-1-dylany@fb.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 9/26/22 11:09 AM, Dylan Yudaken wrote:
> Registering the single issuer task from the first submit adds unnecesary
> complications to the API as well as the implementation. Where simply
> registering it at creation should not impose any barriers to getting the
> same performance wins. The only catch is users that might want to move the
> ring after creation but before submission. For these users allow them to
> create the ring with IORING_SETUP_R_DISABLED and then enable it on the
> submission task.
> 
> There is another problem in 6.1, with IORING_SETUP_DEFER_TASKRUN. That
> would like to check the submitter_task from unlocked contexts, which would
> be racy. If upfront the submitter_task is set at creation time it will
> simplify the logic there and probably increase performance (though this is
> unmeasured).
> 
> Patch 1 registers the task at creation of the io_uring, this works
> standalone in case you want to only merge this part for 6.0
> 
> Patch 2/3 cleans up the code from the old style

Thanks, I like 1/3 a lot better now. Will provide applications with an
easy path to use SINGLE_ISSUER, even if they currently setup the ring
from a different thread/task than they end up using it from.

I've updated the 6.0 and 6.1 repos to reflect this.

-- 
Jens Axboe