On Tue, Oct 31, 2023 at 10:52:02AM +0100, Peter Zijlstra wrote:
> On Mon, Oct 30, 2023 at 01:11:41PM -0700, Paul E. McKenney wrote:
> > On Mon, Oct 30, 2023 at 09:21:38AM +0100, Peter Zijlstra wrote:
> > > On Fri, Oct 27, 2023 at 04:41:30PM -0700, Paul E. McKenney wrote:
> > > > On Sat, Oct 28, 2023 at 12:46:28AM +0200, Peter Zijlstra wrote:
> > >
> > > > > Nah, this is more or less what I feared. I just worry people will come
> > > > > around and put WRITE_ONCE() on the other end. I don't think that'll buy
> > > > > us much. Nor do I think the current READ_ONCE()s actually matter.
> > > >
> > > > My friend, you trust compilers more than I ever will. ;-)
> > >
> > > Well, we only use the values {0,1,2}, that's contained in the first
> > > byte. Are we saying compiler will not only byte-split but also
> > > bit-split the loads?
> > >
> > > But again, lacking the WRITE_ONCE() counterpart, this READ_ONCE() isn't
> > > getting you anything, and if you really worried about it, shouldn't you
> > > have proposed a patch making it all WRITE_ONCE() back when you did this
> > > tasks-rcu stuff?
> >
> > There are not all that many of them. If such a WRITE_ONCE() patch would
> > be welcome, I would be happy to put it together.
> >
> > > > > But perhaps put a comment there, that we don't care for the races and
> > > > > only need to observe a 0 once or something.
> > > >
> > > > There are these two passagers in the big lock comment preceding the
> > > > RCU Tasks code:
> > >
> > > > // rcu_tasks_pregp_step():
> > > > // Invokes synchronize_rcu() in order to wait for all in-flight
> > > > // t->on_rq and t->nvcsw transitions to complete. This works because
> > > > // all such transitions are carried out with interrupts disabled.
> > >
> > > > Does that suffice, or should we add more?
> > >
> > > Probably sufficient. If one were to have used the search option :-)
> > >
> > > Anyway, this brings me to nvcsw, exact same problem there, except
> > > possibly worse, because now we actually do care about the full word.
> > >
> > > No WRITE_ONCE() write side, so the READ_ONCE() don't help against
> > > store-tearing (however unlikely that actually is in this case).
> >
> > Again, if such a WRITE_ONCE() patch would be welcome, I would be happy
> > to put it together.
>
> Welcome is not the right word. What bugs me most is that this was never
> raised when this code was written :/
Me, I consider those READ_ONCE() calls to be documentation as well as
defense against overly enthusiastic optimizers. "This access is racy."
> Mostly my problem is that GCC generates such utter shite when you
> mention volatile. See, the below patch changes the perfectly fine and
> non-broken:
>
> 0148 1d8: 49 83 06 01 addq $0x1,(%r14)
>
> into:
>
> 0148 1d8: 49 8b 06 mov (%r14),%rax
> 014b 1db: 48 83 c0 01 add $0x1,%rax
> 014f 1df: 49 89 06 mov %rax,(%r14)
>
> For absolutely no reason :-(
>
> At least clang doesn't do this, it stays:
>
> 0403 413: 49 ff 45 00 incq 0x0(%r13)
>
> irrespective of the volatile.
Sounds like a bug in GCC, perhaps depending on the microarchitecture
in question. And it was in fact reported in the past, but closed as
not-a-bug. Perhaps clang's fix for this will help GCC along.
And yes, I do see that ++*switch_count in __schedule().
So, at least until GCC catches up to clang's code generation, I take it
that you don't want WRITE_ONCE() for that ->nvcsw increment. Thoughts on
->on_rq?
Thanx, Paul
> ---
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 802551e0009b..d616211b9151 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -6575,8 +6575,8 @@ pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
> */
> static void __sched notrace __schedule(unsigned int sched_mode)
> {
> struct task_struct *prev, *next;
> - unsigned long *switch_count;
> + volatile unsigned long *switch_count;
> unsigned long prev_state;
> struct rq_flags rf;
> struct rq *rq;
