From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-17.2 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,NICE_REPLY_A,SPF_HELO_NONE,SPF_PASS,UNPARSEABLE_RELAY, USER_AGENT_SANE_1 autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id B321BC433EF for ; Thu, 9 Sep 2021 07:01:16 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 957D9600AA for ; Thu, 9 Sep 2021 07:01:16 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S233283AbhIIHCY (ORCPT ); Thu, 9 Sep 2021 03:02:24 -0400 Received: from out30-44.freemail.mail.aliyun.com ([115.124.30.44]:58911 "EHLO out30-44.freemail.mail.aliyun.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S230424AbhIIHCY (ORCPT ); Thu, 9 Sep 2021 03:02:24 -0400 X-Alimail-AntiSpam: AC=PASS;BC=-1|-1;BR=01201311R181e4;CH=green;DM=||false|;DS=||;FP=0|-1|-1|-1|0|-1|-1|-1;HT=e01e04400;MF=haoxu@linux.alibaba.com;NM=1;PH=DS;RN=4;SR=0;TI=SMTPD_---0UnlgeoG_1631170872; Received: from B-25KNML85-0107.local(mailfrom:haoxu@linux.alibaba.com fp:SMTPD_---0UnlgeoG_1631170872) by smtp.aliyun-inc.com(127.0.0.1); Thu, 09 Sep 2021 15:01:13 +0800 Subject: Re: [PATCH 4/6] io_uring: let fast poll support multishot To: Pavel Begunkov , Jens Axboe Cc: io-uring@vger.kernel.org, Joseph Qi References: <20210903110049.132958-1-haoxu@linux.alibaba.com> <20210903110049.132958-5-haoxu@linux.alibaba.com> <9a8efd19-a320-29a4-7132-7b5ae5b994ff@gmail.com> <8c052e2a-0ee6-7dac-1169-9d395d2ecad8@linux.alibaba.com> <2ba9fdb5-6d60-21f5-3e20-bc1687c9509f@linux.alibaba.com> From: Hao Xu Message-ID: <68296a16-e5ad-067c-c8b4-cfae47e920f6@linux.alibaba.com> Date: Thu, 9 Sep 2021 15:01:12 +0800 User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:78.0) Gecko/20100101 Thunderbird/78.13.0 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 8bit Precedence: bulk List-ID: X-Mailing-List: io-uring@vger.kernel.org 在 2021/9/8 下午8:03, Pavel Begunkov 写道: > On 9/8/21 12:21 PM, Hao Xu wrote: >> 在 2021/9/7 下午2:48, Hao Xu 写道: >>> 在 2021/9/7 上午3:04, Pavel Begunkov 写道: >>>> On 9/3/21 12:00 PM, Hao Xu wrote: >>>>> For operations like accept, multishot is a useful feature, since we can >>>>> reduce a number of accept sqe. Let's integrate it to fast poll, it may >>>>> be good for other operations in the future. >>>>> >>>>> Signed-off-by: Hao Xu >>>>> --- >>>>>   fs/io_uring.c | 15 ++++++++++++--- >>>>>   1 file changed, 12 insertions(+), 3 deletions(-) >>>>> >>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c >>>>> index d6df60c4cdb9..dae7044e0c24 100644 >>>>> --- a/fs/io_uring.c >>>>> +++ b/fs/io_uring.c >>>>> @@ -5277,8 +5277,15 @@ static void io_async_task_func(struct io_kiocb *req, bool *locked) >>>>>           return; >>>>>       } >>>>> -    hash_del(&req->hash_node); >>>>> -    io_poll_remove_double(req); >>>>> +    if (READ_ONCE(apoll->poll.canceled)) >>>>> +        apoll->poll.events |= EPOLLONESHOT; >>>>> +    if (apoll->poll.events & EPOLLONESHOT) { >>>>> +        hash_del(&req->hash_node); >>>>> +        io_poll_remove_double(req); >>>>> +    } else { >>>>> +        add_wait_queue(apoll->poll.head, &apoll->poll.wait); >>>> >>>> It looks like it does both io_req_task_submit() and adding back >>>> to the wq, so io_issue_sqe() may be called in parallel with >>>> io_async_task_func(). If so, there will be tons of all kind of >>>> races. >>> IMHO, io_async_task_func() is called in original context one by >>> one(except PF_EXITING is set, it is also called in system-wq), so >>> shouldn't be parallel case there. >> ping... > > fwiw, the case we're talking about: > > CPU0 | CPU1 > io_async_task_func() | > -> add_wait_queue(); | > -> io_req_task_submit(); | > /* no tw run happened in between */ > | io_async_task_func() > | --> io_req_task_submit() > > We called io_req_task_submit() twice without running tw in-between, > both of the calls use the same req->io_task_work.node field in the > request for accounting, and so the second call will screw > tctx->task_list and not only by not considering that > req->io_task_work.node is already taken/enqueued. > > io_req_task_work_add() { > wq_list_add_tail(&req->io_task_work.node, &tctx->task_list); > } > I guess you mean io_req_task_work_add() called by async_wake() two times: io_async_task_func() -> add_wait_queue() async_wake() ->io_req_task_work_add() this one mess up the running task_work list since req->io_task_work.node is in use. It seems the current poll_add + multishot logic has this issue too, I'll give it a shot(simply clean req->io_task_work.node before running req->io_task_work.func should work) >>>> >>>>> +    } >>>>> + >>>>>       spin_unlock(&ctx->completion_lock); >>>>>       if (!READ_ONCE(apoll->poll.canceled)) >>>>> @@ -5366,7 +5373,7 @@ static int io_arm_poll_handler(struct io_kiocb *req) >>>>>       struct io_ring_ctx *ctx = req->ctx; >>>>>       struct async_poll *apoll; >>>>>       struct io_poll_table ipt; >>>>> -    __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI; >>>>> +    __poll_t ret, mask = POLLERR | POLLPRI; >>>>>       int rw; >>>>>       if (!req->file || !file_can_poll(req->file)) >>>>> @@ -5388,6 +5395,8 @@ static int io_arm_poll_handler(struct io_kiocb *req) >>>>>           rw = WRITE; >>>>>           mask |= POLLOUT | POLLWRNORM; >>>>>       } >>>>> +    if (!(req->flags & REQ_F_APOLL_MULTISHOT)) >>>>> +        mask |= EPOLLONESHOT; >>>>>       /* if we can't nonblock try, then no point in arming a poll handler */ >>>>>       if (!io_file_supports_nowait(req, rw)) >>>>> >>>> >> >