Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <1615603667-22568-1-git-send-email-linyunsheng@huawei.com>
 <1615777818-13969-1-git-send-email-linyunsheng@huawei.com>
 <20210315115332.1647e92b@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
 <CAM_iQpXvVZxBRHF6PBDOYSOSCj08nPyfcY0adKuuTg=cqffV+w@mail.gmail.com>
 <87eegddhsj.fsf@toke.dk> <CAHmME9qDU7VRmBV+v0tzLiUpMJykjswSDwqc9P43ZwG1UD7mzw@mail.gmail.com>
 <3bae7b26-9d7f-15b8-d466-ff5c26d08b35@huawei.com>
In-Reply-To: <3bae7b26-9d7f-15b8-d466-ff5c26d08b35@huawei.com>
From:   Cong Wang <xiyou.wangcong@gmail.com>
Date:   Fri, 19 Mar 2021 11:15:14 -0700
Message-ID: <CAM_iQpVvR1eUQxgihWrZ==X=xQjaaeH_qkehvU0Y2R6i9eM-Qw@mail.gmail.com>
Subject: Re: [Linuxarm] Re: [RFC v2] net: sched: implement TCQ_F_CAN_BYPASS
 for lockless qdisc
To:     Yunsheng Lin <linyunsheng@huawei.com>
Cc:     "Jason A. Donenfeld" <Jason@zx2c4.com>,
        =?UTF-8?B?VG9rZSBIw7hpbGFuZC1Kw7hyZ2Vuc2Vu?= <toke@redhat.com>,
        Jakub Kicinski <kuba@kernel.org>,
        David Miller <davem@davemloft.net>,
        Vladimir Oltean <olteanv@gmail.com>,
        Alexei Starovoitov <ast@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        Andrii Nakryiko <andriin@fb.com>,
        Eric Dumazet <edumazet@google.com>,
        Wei Wang <weiwan@google.com>,
        "Cong Wang ." <cong.wang@bytedance.com>,
        Taehee Yoo <ap420073@gmail.com>,
        Linux Kernel Network Developers <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>, linuxarm@openeuler.org,
        Marc Kleine-Budde <mkl@pengutronix.de>,
        linux-can@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, Mar 18, 2021 at 12:33 AM Yunsheng Lin <linyunsheng@huawei.com> wrot=
e:
>
> On 2021/3/17 21:45, Jason A. Donenfeld wrote:
> > On 3/17/21, Toke H=C3=B8iland-J=C3=B8rgensen <toke@redhat.com> wrote:
> >> Cong Wang <xiyou.wangcong@gmail.com> writes:
> >>
> >>> On Mon, Mar 15, 2021 at 2:07 PM Jakub Kicinski <kuba@kernel.org> wrot=
e:
> >>>>
> >>>> I thought pfifo was supposed to be "lockless" and this change
> >>>> re-introduces a lock between producer and consumer, no?
> >>>
> >>> It has never been truly lockless, it uses two spinlocks in the ring
> >>> buffer
> >>> implementation, and it introduced a q->seqlock recently, with this pa=
tch
> >>> now we have priv->lock, 4 locks in total. So our "lockless" qdisc end=
s
> >>> up having more locks than others. ;) I don't think we are going to a
> >>> right direction...
> >>
> >> Just a thought, have you guys considered adopting the lockless MSPC ri=
ng
> >> buffer recently introduced into Wireguard in commit:
> >>
> >> 8b5553ace83c ("wireguard: queueing: get rid of per-peer ring buffers")
> >>
> >> Jason indicated he was willing to work on generalising it into a
> >> reusable library if there was a use case for it. I haven't quite thoug=
h
> >> through the details of whether this would be such a use case, but
> >> figured I'd at least mention it :)
> >
> > That offer definitely still stands. Generalization sounds like a lot of=
 fun.
> >
> > Keep in mind though that it's an eventually consistent queue, not an
> > immediately consistent one, so that might not match all use cases. It
> > works with wg because we always trigger the reader thread anew when it
> > finishes, but that doesn't apply to everyone's queueing setup.
>
> Thanks for mentioning this.
>
> "multi-producer, single-consumer" seems to match the lockless qdisc's
> paradigm too, for now concurrent enqueuing/dequeuing to the pfifo_fast's
> queues() is not allowed, it is protected by producer_lock or consumer_loc=
k.
>
> So it would be good to has lockless concurrent enqueuing, while dequeuing
> can be protected by qdisc_lock() or q->seqlock, which meets the "multi-pr=
oducer,
> single-consumer" paradigm.

I don't think so. Usually we have one queue for each CPU so we can expect
each CPU has a lockless qdisc assigned, but we can not assume this in
the code, so we still have to deal with multiple CPU's sharing a lockless q=
disc,
and we usually enqueue and dequeue in process context, so it means we could
have multiple producers and multiple consumers.

On the other hand, I don't think the problems we have been fixing are the r=
ing
buffer implementation itself, they are about the high-level qdisc
state transitions.

Thanks.