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: <6cce16d3-e2ca-ca1e-1209-e6e243241231@gmail.com>
Date:   Thu, 3 Feb 2022 15:44:21 +0000
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.5.0
Subject: Re: [External] Re: [RFC] io_uring: avoid ring quiesce while
 registering/unregistering eventfd
Content-Language: en-US
To:     Usama Arif <usama.arif@bytedance.com>,
        Jens Axboe <axboe@kernel.dk>, io-uring@vger.kernel.org,
        linux-kernel@vger.kernel.org
Cc:     fam.zheng@bytedance.com
References: <20220202155923.4117285-1-usama.arif@bytedance.com>
 <86ae792e-d138-112e-02bb-ab70e3c2a147@kernel.dk>
 <7df1059c-6151-29c8-9ed5-0bc0726c362d@kernel.dk>
 <1494b8f0-2f48-0aa1-214c-a02bbc4b05eb@bytedance.com>
From:   Pavel Begunkov <asml.silence@gmail.com>
In-Reply-To: <1494b8f0-2f48-0aa1-214c-a02bbc4b05eb@bytedance.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2/3/22 15:14, Usama Arif wrote:
> On 02/02/2022 19:18, Jens Axboe wrote:
>> On 2/2/22 9:57 AM, Jens Axboe wrote:
>>> On 2/2/22 8:59 AM, Usama Arif wrote:
>>>> Acquire completion_lock at the start of __io_uring_register before
>>>> registering/unregistering eventfd and release it at the end. Hence
>>>> all calls to io_cqring_ev_posted which adds to the eventfd counter
>>>> will finish before acquiring the spin_lock in io_uring_register, and
>>>> all new calls will wait till the eventfd is registered. This avoids
>>>> ring quiesce which is much more expensive than acquiring the
>>>> spin_lock.
>>>>
>>>> On the system tested with this patch, io_uring_reigster with
>>>> IORING_REGISTER_EVENTFD takes less than 1ms, compared to 15ms before.
>>>
>>> This seems like optimizing for the wrong thing, so I've got a few
>>> questions. Are you doing a lot of eventfd registrations (and
>>> unregister) in your workload? Or is it just the initial pain of
>>> registering one? In talking to Pavel, he suggested that RCU might be a
>>> good use case here, and I think so too. That would still remove the
>>> need to quiesce, and the posted side just needs a fairly cheap rcu
>>> read lock/unlock around it.
>>
>> Totally untested, but perhaps can serve as a starting point or
>> inspiration.
>>
> 
> Hi,
> 
> Thank you for the replies and comments. My usecase registers only one eventfd at the start.

Then it's overkill. Update io_register_op_must_quiesce(), set ->cq_ev_fd
on registration with WRITE_ONCE(), read it in io_cqring_ev_posted* with
READ_ONCE() and you're set.

There is a caveat, ->cq_ev_fd won't be immediately visible to already
inflight requests, but we can say it's the responsibility of the
userspace to wait for a grace period, i.e. for all inflight requests
submitted before registration io_cqring_ev_posted* might or might not
see updated ->cq_ev_fd, which works perfectly if there was no requests
in the first place. Of course it changes the behaviour so will need
a new register opcode.

-- 
Pavel Begunkov