Re: [PATCH tip/core/rcu 1/3] rcu-tasks: *_ONCE() for rcu_tasks_cbs_head

[Peter Zijlstra]

On Mon, Feb 17, 2020 at 10:16:16AM -0800, Paul E. McKenney wrote:
> On Mon, Feb 17, 2020 at 01:38:51PM +0100, Peter Zijlstra wrote:
> > On Fri, Feb 14, 2020 at 04:25:18PM -0800, paulmck@xxxxxxxxxx wrote:
> > > From: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> > > 
> > > The RCU tasks list of callbacks, rcu_tasks_cbs_head, is sampled locklessly
> > > by rcu_tasks_kthread() when waiting for work to do.  This commit therefore
> > > applies READ_ONCE() to that lockless sampling and WRITE_ONCE() to the
> > > single potential store outside of rcu_tasks_kthread.
> > > 
> > > This data race was reported by KCSAN.  Not appropriate for backporting
> > > due to failure being unlikely.
> > 
> > What failure is possible here? AFAICT this is (again) one of them
> > load-complare-against-constant-discard patterns that are impossible to
> > mess up.
> 
> First, please keep in mind that this is RCU code.  Rather uncomplicated
> for RCU, to be sure, but still RCU code.
> 
> The failure modes are thus as follows:
> 
> o	I produce a patch for which KCSAN gives a legitimate warning,
> 	but this warning is obscured by a pile of other warnings.
> 	Yes, we should continue improving KCSAN's ability to adapt
> 	to the users desired compiler-optimization risk level, but
> 	in RCU's case that risk level is set quite low.
> 
> 	In RCU, what others are calling false positives are therefore
> 	addressed.  Yes, this does cost me a bit of work, but it is
> 	trivial compared to the work required to track down a real bug.
> 
> o	Someone optimizes or otherwise changes the wait/wakeup code,
> 	which inadvertently gives the compiler more scope for mischief.
> 
> In short, within RCU, I am handling all KCSAN complaints.  This is looking
> to be an extremely inexpensive insurance policy for RCU.  Other subsystems
> are of course free to make their own tradeoffs, and subsystems having
> less-aggressive concurrency control might be well-advised to take a
> different path than the one I am taking.

I just took offence at the Changelog wording. It seems to suggest there
actually is a problem, there is not.