linux-kernel - Re: [PATCH v1] io_uring: Add support for napi_busy

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <7695c3a2-6cb0-067a-5655-0e6180170bde@linux.alibaba.com>
Date:   Tue, 1 Mar 2022 16:23:00 +0800
From:   Hao Xu <haoxu@...ux.alibaba.com>
To:     Olivier Langlois <olivier@...llion01.com>,
        Jens Axboe <axboe@...nel.dk>
Cc:     Pavel Begunkov <asml.silence@...il.com>,
        io-uring <io-uring@...r.kernel.org>,
        linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v1] io_uring: Add support for napi_busy_poll


On 3/1/22 05:01, Olivier Langlois wrote:
> On Tue, 2022-03-01 at 02:26 +0800, Hao Xu wrote:
>> On 2/25/22 13:32, Olivier Langlois wrote:
>>> On Mon, 2022-02-21 at 13:23 +0800, Hao Xu wrote:
>>>>> @@ -5776,6 +5887,7 @@ static int __io_arm_poll_handler(struct
>>>>> io_kiocb *req,
>>>>>                   __io_poll_execute(req, mask);
>>>>>                   return 0;
>>>>>           }
>>>>> +       io_add_napi(req->file, req->ctx);
>>>> I think this may not be the right place to do it. the process
>>>> will
>>>> be:
>>>> arm_poll sockfdA--> get invalid napi_id from sk->napi_id -->
>>>> event
>>>> triggered --> arm_poll for sockfdA again --> get valid napi_id
>>>> then why not do io_add_napi() in event
>>>> handler(apoll_task_func/poll_task_func).
>>> You have a valid concern that the first time a socket is passed to
>>> io_uring that napi_id might not be assigned yet.
>>>
>>> OTOH, getting it after data is available for reading does not help
>>> neither since busy polling must be done before data is received.
>>>
>>> for both places, the extracted napi_id will only be leveraged at
>>> the
>>> next polling.
>> Hi Olivier,
>>
>> I think we have some gap here. AFAIK, it's not 'might not', it is
>>
>> 'definitely not', the sk->napi_id won't be valid until the poll
>> callback.
>>
>> Some driver's code FYR:
>> (drivers/net/ethernet/intel/e1000/e1000_main.c)
>>
>> e1000_receive_skb-->napi_gro_receive-->napi_skb_finish--
>>> gro_normal_one
>> and in gro_normal_one(), it does:
>>
>>             if (napi->rx_count >= gro_normal_batch)
>>                     gro_normal_list(napi);
>>
>>
>> The gro_normal_list() delivers the info up to the specifical network
>> protocol like tcp.
>>
>> And then sk->napi_id is set, meanwhile the poll callback is
>> triggered.
>>
>> So that's why I call the napi polling technology a 'speculation'.
>> It's
>> totally for the
>>
>> future data. Correct me if I'm wrong especially for the poll callback
>> triggering part.
>>
> When I said 'might not', I was meaning that from the io_uring point of
> view, it has no idea what is the previous socket usage. If it has been
> used outside io_uring, the napi_id could available on the first call.
>
> If it is really read virgin socket, neither my choosen call site or
> your proposed sites will make the napi busy poll possible for the first
> poll.
>
> I feel like there is not much to gain to argue on this point since I
> pretty much admitted that your solution was most likely the only call
> site making MULTIPOLL requests work correctly with napi busy poll as
> those requests could visit __io_arm_poll_handler only once (Correct me
> if my statement is wrong).
>
> The only issue was that I wasn't sure is how using your calling sites
> would make locking work.
>
> I suppose that adding a dedicated spinlock for protecting napi_list
> instead of relying on uring_lock could be a solution. Would that work?
spinlock should be fine.