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;
Subject: Re: [BUG] io_uring: defer logic based on shared data
To:     Pavel Begunkov <asml.silence@gmail.com>,
        linux-block@vger.kernel.org, linux-kernel@vger.kernel.org
References: <5badf1c0-9a7d-0950-2943-ff8db33e0929@gmail.com>
 <bfb58429-6abe-06f0-3fd8-14a0040cecf0@kernel.dk>
 <b44b0488-ba66-0187-2d9b-6949ceb613fb@gmail.com>
 <96446fe1-4f32-642b-7100-ebfa291d7127@kernel.dk>
 <df3b9edd-86ad-5460-b61b-66707c0fb630@kernel.dk>
 <31a7765b-bb6d-985a-454d-d998678100d1@gmail.com>
 <b4e1f03c-e044-b09f-d943-cad3ab5b4969@kernel.dk>
 <e5a6f77a-3404-0dc8-ac6e-584737d71a33@gmail.com>
From:   Jens Axboe <axboe@kernel.dk>
Message-ID: <a0d8a8e1-18dc-8090-037c-e5baf9bd45c3@kernel.dk>
Date:   Fri, 25 Oct 2019 10:57:18 -0600
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101
 Thunderbird/60.9.0
MIME-Version: 1.0
In-Reply-To: <e5a6f77a-3404-0dc8-ac6e-584737d71a33@gmail.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 10/25/19 10:55 AM, Pavel Begunkov wrote:
> On 25/10/2019 19:44, Jens Axboe wrote:
>> On 10/25/19 10:40 AM, Pavel Begunkov wrote:
>>> On 25/10/2019 19:32, Jens Axboe wrote:
>>>> On 10/25/19 10:27 AM, Jens Axboe wrote:
>>>>> On 10/25/19 10:21 AM, Pavel Begunkov wrote:
>>>>>> On 25/10/2019 19:03, Jens Axboe wrote:
>>>>>>> On 10/25/19 3:55 AM, Pavel Begunkov wrote:
>>>>>>>> I found 2 problems with __io_sequence_defer().
>>>>>>>>
>>>>>>>> 1. it uses @sq_dropped, but doesn't consider @cq_overflow
>>>>>>>> 2. @sq_dropped and @cq_overflow are write-shared with userspace, so
>>>>>>>> it can be maliciously changed.
>>>>>>>>
>>>>>>>> see sent liburing test (test/defer *_hung()), which left an unkillable
>>>>>>>> process for me
>>>>>>>
>>>>>>> OK, how about the below. I'll split this in two, as it's really two
>>>>>>> separate fixes.
>>>>>> cached_sq_dropped is good, but I was concerned about cached_cq_overflow.
>>>>>> io_cqring_fill_event() can be called in async, so shouldn't we do some
>>>>>> synchronisation then?
>>>>>
>>>>> We should probably make it an atomic just to be on the safe side, I'll
>>>>> update the series.
>>>>
>>>> Here we go, patch 1:
>>>>
>>>> http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=f2a241f596ed9e12b7c8f960e79ccda8053ea294
>>>>
>>>> patch 2:
>>>>
>>>> http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=b7d0297d2df5bfa0d1ecf9d6c66d23676751ef6a
>>>>
>>> 1. submit rqs (not yet completed)
>>> 2. poll_list is empty, inflight = 0
>>> 3. async completed and placed into poll_list
>>>
>>> So, poll_list is not empty, but we won't get to polling again.
>>> At least until someone submitted something.
>>
>> But if they are issued, the will sit in ->poll_list as well. That list
>> holds both "submitted, but pending" and completed entries.
>>
> Missed it, then should work. Thanks!

Glad we agree :-)

>> + ret = iters = 0;
> A small suggestion, could we just initialise it in declaration
> to be a bit more concise?
> e.g. int ret = 0, iters = 0;
> 
> Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
> And let me test it as both patches are ready.

Sure, I'll make that change and add your reviewed-by. Thanks!

-- 
Jens Axboe