Re: [RFC][PATCH] Fix a race between rwsem and the scheduler

[Oleg Nesterov <oleg@redhat.com>]

On 08/30, Peter Zijlstra wrote:
> On Tue, Aug 30, 2016 at 03:04:27PM +0200, Oleg Nesterov wrote:
> > On 08/30, Peter Zijlstra wrote:
> > >
> > > 	/*
> > > 	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
> > > 	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
> > > 	 * in smp_cond_load_acquire() below.
> > > 	 *
> > > 	 * sched_ttwu_pending()			try_to_wake_up()
> > > 	 *   [S] p->on_rq = 1;			[L] P->state
> > > 	 *       UNLOCK rq->lock
> > > 	 *
> > > 	 * schedule()				RMB
> > > 	 *       LOCK rq->lock
> > > 	 *       UNLOCK rq->lock
> > > 	 *
> > > 	 * [task p]
> > > 	 *   [S] p->state = UNINTERRUPTIBLE	[L] p->on_rq
> > > 	 *
> > > 	 * Pairs with the UNLOCK+LOCK on rq->lock from the
> > > 	 * last wakeup of our task and the schedule that got our task
> > > 	 * current.
> > > 	 */
> >
> > Confused... how this connects to UNLOCK+LOCK on rq->lock? A LOAD can
> > leak into the critical section.
>
> How so? That LOCK+UNLOCK which is leaky, UNLOCK+LOCK is a read/write
> barrier (just not an MB because it lacks full transitivity).

Ah, I have wrongly read the "Pairs with the UNLOCK+LOCK" as
"Pairs with the LOCK+UNLOCK". And didn't notice this even after I
copy-and-pasted this part.

> > But context switch should imply mb() we can rely on?
>
> Not sure it should, on x86 switch_mm does a CR3 write and that is
> serializing, but switch_to() doesn't need to do anything iirc.

Documentation/memory-barriers.txt says

	schedule() and similar imply full memory barriers.

and I (wrongly?) interpreted this as if this is also true for 2
different threadds.

I mean, I thought that the LOAD/STORE's done by some task can't
be re-ordered with LOAD/STORE's done by another task which was
running on the same CPU. Wrong?

Oleg.