Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Date:   Thu, 12 May 2022 16:10:31 -0700
From:   "Paul E. McKenney" <paulmck@kernel.org>
To:     Eric Dumazet <edumazet@google.com>
Cc:     Marco Elver <elver@google.com>, Liu Jian <liujian56@huawei.com>,
        Dmitry Vyukov <dvyukov@google.com>,
        LKML <linux-kernel@vger.kernel.org>,
        David Miller <davem@davemloft.net>,
        Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>,
        David Ahern <dsahern@kernel.org>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>,
        Neal Cardwell <ncardwell@google.com>,
        netdev <netdev@vger.kernel.org>
Subject: Re: [PATCH net] tcp: Add READ_ONCE() to read tcp_orphan_count
Message-ID: <20220512231031.GT1790663@paulmck-ThinkPad-P17-Gen-1>
Reply-To: paulmck@kernel.org
References: <20220512103322.380405-1-liujian56@huawei.com>
 <CANn89iJ7Lo7NNi4TrpKsaxzFrcVXdgbyopqTRQEveSzsDL7CFA@mail.gmail.com>
 <CANpmjNPRB-4f3tUZjycpFVsDBAK_GEW-vxDbTZti+gtJaEx2iw@mail.gmail.com>
 <CANn89iKJ+9=ug79V_bd8LSsLaSu0VLtzZdDLC87rcvQ6UYieHQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CANn89iKJ+9=ug79V_bd8LSsLaSu0VLtzZdDLC87rcvQ6UYieHQ@mail.gmail.com>
Precedence: bulk

On Thu, May 12, 2022 at 02:31:48PM -0700, Eric Dumazet wrote:
> On Thu, May 12, 2022 at 2:18 PM Marco Elver <elver@google.com> wrote:
> 
> >
> > I guess the question is, is it the norm that per_cpu() retrieves data
> > that can legally be modified concurrently, or not. If not, and in most
> > cases it's a bug, the annotations should be here.
> >
> > Paul, was there any guidance/documentation on this, but I fail to find
> > it right now? (access-marking.txt doesn't say much about per-CPU
> > data.)
> 
> Normally, whenever we add a READ_ONCE(), we are supposed to add a comment.

I am starting to think that comments are even more necessary for unmarked
accesses to shared variables, with the comments setting out why the
compiler cannot mess things up.  ;-)

> We could make an exception for per_cpu_once(), because the comment
> would be centralized
> at per_cpu_once() definition.

This makes a lot of sense to me.

> We will be stuck with READ_ONCE() in places we are using
> per_cpu_ptr(), for example
> in dev_fetch_sw_netstats()

If this is strictly statistics, data_race() is another possibility.
But it does not constrain the compiler at all.

							Thanx, Paul

> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1461c2d9dec8099a9a2d43a704b4c6cb0375f480..b66470291d7b7e6c33161093d71e40587f9ed838
> 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -10381,10 +10381,13 @@ void dev_fetch_sw_netstats(struct
> rtnl_link_stats64 *s,
>                 stats = per_cpu_ptr(netstats, cpu);
>                 do {
>                         start = u64_stats_fetch_begin_irq(&stats->syncp);
> -                       tmp.rx_packets = stats->rx_packets;
> -                       tmp.rx_bytes   = stats->rx_bytes;
> -                       tmp.tx_packets = stats->tx_packets;
> -                       tmp.tx_bytes   = stats->tx_bytes;
> +                       /* These values can change under us.
> +                        * READ_ONCE() pair with too many write sides...
> +                        */
> +                       tmp.rx_packets = READ_ONCE(stats->rx_packets);
> +                       tmp.rx_bytes   = READ_ONCE(stats->rx_bytes);
> +                       tmp.tx_packets = READ_ONCE(stats->tx_packets);
> +                       tmp.tx_bytes   = READ_ONCE(stats->tx_bytes);
>                 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
> 
>                 s->rx_packets += tmp.rx_packets;