Re: [PATCH 1/2] IB/hfi1: Use preempt_{dis,en}able_nort()

From: Julia Cartwright
Date: Thu Oct 05 2017 - 12:40:24 EST


On Thu, Oct 05, 2017 at 06:16:23PM +0200, Thomas Gleixner wrote:
> On Thu, 5 Oct 2017, Julia Cartwright wrote:
>
> > On Thu, Oct 05, 2017 at 11:55:39AM -0400, Steven Rostedt wrote:
> > > On Thu, 5 Oct 2017 10:37:59 -0500
> > > Julia Cartwright <julia@xxxxxx> wrote:
> > >
> > > > On Thu, Oct 05, 2017 at 05:27:30PM +0200, Thomas Gleixner wrote:
> > > > > On Thu, 5 Oct 2017, Julia Cartwright wrote:
> > > > > > On Tue, Oct 03, 2017 at 12:49:19PM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > > - preempt_disable();
> > > > > > > + preempt_disable_nort();
> > > > > > > this_cpu_inc(*sc->buffers_allocated);
> > > > > >
> > > > > > Have you tried this on RT w/ CONFIG_DEBUG_PREEMPT? I believe that the
> > > > > > this_cpu_* operations perform a preemption check, which we'd trip.
> > > > >
> > > > > Good point. Changing this to migrate_disable() would do the trick.
> > > >
> > > > Wouldn't we still trip the preempt check even with migration disabled?
> > > > In another thread I asked the same question: should the preemption
> > > > checks here be converted to migration-checks in RT?
> > >
> > > Is it a "preemption check"?
> >
> > Sorry if I was unclear, more precisely: the this_cpu_* family of
> > accessors, w/ CONFIG_DEBUG_PREEMPT currently spits out a warning when
> > the caller is invoked in a context where preemption is enabled.
> >
> > The check is shared w/ the smp_processor_id() check, as implemented in
> > lib/smp_processor_id.c. It effectively boils down to a check of
> > preempt_count() and irqs_disabled().
>
> Except that on RT that check cares about the migrate disable state. You can
> invoke this_cpu_* and smp_processor_id() in preemptible/interruptible
> context because of:
>
> if (cpumask_equal(current->cpus_ptr, cpumask_of(this_cpu)))
> goto out;
>
> That's true even on mainline.
>
> But you are right that this check needs some update because
> migrate_disable() does not immediately update the allowed cpumask IIRC.

Actually, I think it does:

migrate_disable() ->
p = current;
...
migrate_disable_update_cpus_allowed(p) ->
p->cpus_ptr = cpumask_of(smp_processor_id())
...

Perhaps it's worth a simple comment update, below.

Julia

-- 8< --
Subject: [PATCH] kernel: sched: document smp_processor_id/this_cpu safety in
migration-disabled context

On RT, users of smp_processor_id() and this_cpu_*() per-cpu accessors
are considered safe if the caller executes with migration disabled. On
!RT, preempt_disable() is sufficient to make this guarantee, however on
RT, the lesser migrate_disable() is sufficient.

It is not entirely obvious which check in check_preemption_disabled()
makes this work, so document it.

Signed-off-by: Julia Cartwright <julia@xxxxxx>
---
lib/smp_processor_id.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/lib/smp_processor_id.c b/lib/smp_processor_id.c
index de3b2d925473..c8091d9eb1b4 100644
--- a/lib/smp_processor_id.c
+++ b/lib/smp_processor_id.c
@@ -20,7 +20,11 @@ notrace static unsigned int check_preemption_disabled(const char *what1,

/*
* Kernel threads bound to a single CPU can safely use
- * smp_processor_id():
+ * smp_processor_id().
+ *
+ * In addition, threads which are currently executing within
+ * a migration disabled region can safely use smp_processor_id() and
+ * this_cpu accessors.
*/
if (cpumask_equal(current->cpus_ptr, cpumask_of(this_cpu)))
goto out;
--
2.14.1