Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 11 Oct 2018 09:31:33 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Eric Dumazet <edumazet@google.com>
Cc:     linux-kernel <linux-kernel@vger.kernel.org>,
        Eric Dumazet <eric.dumazet@gmail.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>,
        "H. Peter Anvin" <hpa@zytor.com>
Subject: Re: [PATCH] x86/tsc: use real seqcount_latch in cyc2ns_read_begin()
Message-ID: <20181011073133.GZ5663@hirez.programming.kicks-ass.net>
References: <20181011003336.168941-1-edumazet@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20181011003336.168941-1-edumazet@google.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Oct 10, 2018 at 05:33:36PM -0700, Eric Dumazet wrote:
> While looking at native_sched_clock() disassembly I had
> the surprise to see the compiler (gcc 7.3 here) had
> optimized out the loop, meaning the code is broken.
> 
> Using the documented and approved API not only fixes the bug,
> it also makes the code more readable.
> 
> Replacing five this_cpu_read() by one this_cpu_ptr() makes
> the generated code smaller.

Does not for me, that is, the resulting asm is actually larger

You're quite right the loop went missing; no idea wth that compiler is
smoking (gcc-8.2 for me). In order to eliminate that loop it needs to
think that two consecutive loads of this_cpu_read(cyc2ns.seq.sequence)
will return the same value. But this_cpu_read() is an asm() statement,
it _should_ not assume such.

We assume that this_cpu_read() implies READ_ONCE() in a number of
locations, this really should not happen.

The reason it was written using this_cpu_read() is so that it can use
%gs: prefixed instructions and avoid ever loading that percpu offset and
doing manual address computation.

Let me prod at this with a sharp stick.