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: <20210721202042.GA1472052@paulmck-ThinkPad-P17-Gen-1>
 <20210721202127.2129660-4-paulmck@kernel.org> <20210728173715.GA9416@paulmck-ThinkPad-P17-Gen-1>
 <CAHk-=wjwZzi=Lqsre8C8C4YJiVQNdNnco7jkLq=ohePSPA9JCA@mail.gmail.com> <20210728184605.GI4397@paulmck-ThinkPad-P17-Gen-1>
In-Reply-To: <20210728184605.GI4397@paulmck-ThinkPad-P17-Gen-1>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 28 Jul 2021 11:57:14 -0700
Message-ID: <CAHk-=wj-5vtA_-9xs6zV0U20LRaRhVeE1Ky7X+9kHeucrb6_ig@mail.gmail.com>
Subject: Re: [PATCH v2 rcu 04/18] rcu: Weaken ->dynticks accesses and updates
To:     "Paul E. McKenney" <paulmck@kernel.org>
Cc:     rcu <rcu@vger.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Kernel Team <kernel-team@fb.com>,
        Ingo Molnar <mingo@kernel.org>,
        Lai Jiangshan <jiangshanlai@gmail.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Mathieu Desnoyers <mathieu.desnoyers@efficios.com>,
        Josh Triplett <josh@joshtriplett.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Peter Zijlstra <peterz@infradead.org>,
        Steven Rostedt <rostedt@goodmis.org>,
        David Howells <dhowells@redhat.com>,
        Eric Dumazet <edumazet@google.com>,
        =?UTF-8?B?RnLDqWTDqXJpYyBXZWlzYmVja2Vy?= <fweisbec@gmail.com>,
        Oleg Nesterov <oleg@redhat.com>,
        Joel Fernandes <joel@joelfernandes.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jul 28, 2021 at 11:46 AM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> But atomic_read_this_cpu(&rcu_data.dynticks) isn't all that much shorter
> than atomic_read(this_cpu_ptr(&rcu_data.dynticks)).

It's not so much that it's shorter to write for a human, it's that we
could generate better code for it.

That atomic_read(this_cpu_ptr()) pattern generates code like

        movq    $rcu_data+288, %rax
        add %gs:this_cpu_off(%rip), %rax
        movl    (%rax), %eax

but it *could* just generate

        movl %gs:rcu_data+288, %rax

instead.

Similar patterns for the other per-cpu atomics, ie it would be
possible to just generate

        lock ; xaddl %gs:..., %rax

instead of generating the address by doing that "add %gs:this_cpu_off" thing..

But no, it doesn't look like there are enough users of this to matter.
We're just talking a few extra bytes, and a couple of extra
instructions (and possibly slightly higher register pressure, which
then generates more instructions).

The *expensive* part remains the SMP serialization of the "lock".

                Linus