Re: [PATCH 1/1] Fix: trace sched switch start/stop racy updates

From: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
To: rostedt <rostedt@goodmis.org>
Cc: linux-kernel <linux-kernel@vger.kernel.org>,
	"Joel Fernandes, Google" <joel@joelfernandes.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	paulmck <paulmck@linux.ibm.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	Will Deacon <will.deacon@arm.com>,
	David Howells <dhowells@redhat.com>,
	Alan Stern <stern@rowland.harvard.edu>,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH 1/1] Fix: trace sched switch start/stop racy updates
Date: Sat, 17 Aug 2019 11:53:41 -0400 (EDT)	[thread overview]
Message-ID: <1982627598.23941.1566057221039.JavaMail.zimbra@efficios.com> (raw)
In-Reply-To: <20190817114218.5cb3912b@oasis.local.home>

----- On Aug 17, 2019, at 11:42 AM, rostedt rostedt@goodmis.org wrote:

> On Sat, 17 Aug 2019 10:27:39 -0400 (EDT)
> Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> 
>> I get your point wrt WRITE_ONCE(): since it's a cache it should not have
>> user-visible effects if a temporary incorrect value is observed. Well in
>> reality, it's not a cache: if the lookup fails, it returns "<...>" instead,
>> so cache lookup failure ends up not providing any useful data in the trace.
>> Let's assume this is a known and documented tracer limitation.
> 
> Note, this is done at every sched switch, for both next and prev tasks.
> And the update is only done at the enabling of a tracepoint (very rare
> occurrence) If it missed it scheduling in, it has a really good chance
> of getting it while scheduling out.
> 
> And 99.999% of my tracing that I do, the tasks scheduling in when
> enabling a tracepoint is not what I even care about, as I enable
> tracing then start what I want to trace.

Since it's refcount based, my concern is about the side-effect of
incrementing or decrementing that reference count without WRITE_ONCE
which would lead to a transient corrupted value observed by _another_
active tracing user.

For you use-case, it would lead to a missing comm when you are actively
tracing what you want to trace, caused by another user of that refcount
incrementing or decrementing it.

I agree with you that missing tracing data at the beginning or end of a
trace is not important.

>> 
>> However, wrt READ_ONCE(), things are different. The variable read ends up
>> being used to control various branches in the code, and the compiler could
>> decide to re-fetch the variable (with a different state), and therefore
>> cause _some_ of the branches to be inconsistent. See
>> tracing_record_taskinfo_sched_switch() and tracing_record_taskinfo() @flags
>> parameter.
> 
> I'm more OK with using a READ_ONCE() on the flags so it is consistent.
> But the WRITE_ONCE() is going a bit overboard.

Hence my request for additional guidance on the usefulness of WRITE_ONCE(),
whether it's mainly there for documentation purposes, or if we should consider
that it takes care of real-life problems introduced by compiler optimizations
in the wild. The LWN article seems to imply that it's not just a theoretical
issue, but I'll have to let the article authors justify their conclusions,
because I have limited time to investigate this myself.

> 
>> 
>> AFAIU the current code should not generate any out-of-bound writes in case of
>> re-fetch, but no comment in there documents how fragile this is.
> 
> Which part of the code are you talking about here?

kernel/trace/trace.c:tracing_record_taskinfo_sched_switch()
kernel/trace/trace.c:tracing_record_taskinfo()

where @flags is used to control a few branches. I don't think any of those
would end up causing corruption if the flags is re-fetched between two
branches, but it seems rather fragile.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com