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;
Date:   Fri, 5 Mar 2021 09:36:30 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     Ingo Molnar <mingo@redhat.com>,
        Arnaldo Carvalho de Melo <acme@kernel.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Alexander Shishkin <alexander.shishkin@linux.intel.com>,
        Jiri Olsa <jolsa@redhat.com>,
        Namhyung Kim <namhyung@kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: "struct perf_sample_data" alignment
Message-ID: <YEHtjioucovbxbRt@hirez.programming.kicks-ass.net>
References: <CAHk-=wgQWHDUFjzmAazg8WN0BR7nOyHmduj-MV1GWWDUu+UKCQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHk-=wgQWHDUFjzmAazg8WN0BR7nOyHmduj-MV1GWWDUu+UKCQ@mail.gmail.com>
Precedence: bulk

On Thu, Mar 04, 2021 at 07:45:44PM -0800, Linus Torvalds wrote:
> That ____cacheline_aligned goes back many years, this is not new, it
> seems to come from back in 2014: commit 2565711fb7d7 ("perf: Improve
> the perf_sample_data struct layout").

long time ago...

> But it really seems entirely and utterly bogus. That cacheline
> alignment makes things *worse*, when the variables are on the local
> stack. The local stack is already going to be dirty and in the cache,
> and aligning those things isn't going to - and I quote from the code
> in that commend in that commit - "minimize the cachelines touched".
> 
> Quite the reverse. It's just going to make the stack frame use *more*
> memory, and make any cacheline usage _worse_.

IIRC there is more history here, but I can't seem to find references
just now.

What I remember is that since perf_sample_data is fairly large,
unconditionally initializing the whole thing is *slow* (and
-fauto-var-init=zero will hurt here).

So at some point I removed that full initialization and made sure we
only unconditionally touched the first few variables, which gave a
measurable speedup.

Then things got messy again and the commit 2565711fb7d7 referenced above
was cleanup, to get back to that initial state.

Now, you're right that __cacheline_aligned on on-stack (and this is
indeed mostly on-stack) is fairly tedious (there were a few patches
recently to reduce the amount of on-stack instances).

I'll put it on the todo list, along with that hotplug stuff (which I
tried to fix but ended up with an even bigger mess). I suppose we can
try and not have the alignment for the on-stack instances while
preserving it for the few off-stack ones.

Also; we're running on the NMI stack, and that's not typically hot.