Received-SPF: pass (google.com: domain of linux-kernel+bounces-31814-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Sat, 20 Jan 2024 09:34:01 -0500
From: Steven Rostedt <rostedt@goodmis.org>
To: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: LKML <linux-kernel@vger.kernel.org>, Linux Trace Kernel
 <linux-trace-kernel@vger.kernel.org>, Masami Hiramatsu
 <mhiramat@kernel.org>, Mark Rutland <mark.rutland@arm.com>, Philippe Proulx
 <pproulx@efficios.com>
Subject: Re: [PATCH] ring-buffer: Simplify reservation with try_cmpxchg()
 loop
Message-ID: <20240120093401.0274a2e8@rorschach.local.home>
In-Reply-To: <20240120084713.6eb7aa52@rorschach.local.home>
References: <20240118181206.4977da2f@gandalf.local.home>
	<504085e9-bf91-4948-a158-abae5dcb276a@efficios.com>
	<20240119103754.154dc009@gandalf.local.home>
	<cd3f37cc-31ba-4eb4-8d67-852d05570e7b@efficios.com>
	<20240119164252.54ccb654@gandalf.local.home>
	<5d323a65-8a04-4c73-8702-58869982a269@efficios.com>
	<20240120084713.6eb7aa52@rorschach.local.home>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit

On Sat, 20 Jan 2024 08:47:13 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:

> > I would also consider reducing code complexity as a worthwhile metric
> > in addition to speed. It makes it easier to extend in the future,
> > easier to understand for reviewers, and subtle bugs are less likely
> > to creep in.  
> 
> Really, it wouldn't make it that much simpler. The addition of the
> cmpxchg() of this patch removed the nasty part of the code.

Now let's look at the difference of the two. You still need to save the
current timestamp in one variable. I have to do it in two, so your
algorithm does have that advantage. I currently have (eliminating the
"always add absolute timestamp" switch):


  w = write;
  before = before_stamp;
again:
  after = write_stamp; (equivalent to your last_tsc)
  ts = rdtsc();
  if (!w)
	delta = 0; // event has same ts as subbuf
  else if (before == after)
	delta = ts - after;
  else {
	delta = 0;
	inject_absolute = true;
  }

  before_stamp = ts;

  if (!try_cmpxchg(&write, w, w + length))
	goto again;

  write_stamp = ts;


Now if I were to updated it to use a delta from the last injected
timestamp, where injecting a timestamp only happens on overflow.

#define TS_BITS 27
#define MAX_DELTA ((1 << TS_BITS) - 1)
#define BITMASK (~MAX_DELTA)

  w = write;
again:
  ts = rdtsc();

  delta = ts & MAX_DELTA;

  if (ts - (write_stamp & BITMASK) > MAX_DELTA)
	inject_absolute = true;

  if (!try_cmpxchg(&write, w, w + length))
	goto again;

  write_stamp = ts;

I admit that it does simplify the code a little, but does it do it
enough to be worth the process of deprecating an ABI with a new one?

-- Steve