[RFC][PATCH] sched: might_sleep(): do rate-limiting before sanity checks

[RFC][PATCH] sched: might_sleep(): do rate-limiting before sanity checks

[Dave Hansen]

From: Dave Hansen <dave.hansen@linux.intel.com>

I have a dumb microbenchmark.  It loops doing single-byte writes
to a file.  I have a few other patches to work on some things in
the filesystem write path.  But after those are applied, the
4th-hottest kernel function is ___might_sleep() which seems a bit
silly.

I narrowed the overhead down to the pushf/pop in native_save_fl()
underneath the irqs_disabled() call.  Those instructions must
serialize something in the CPU because they seem to be way slower
than they should be.

In any case, we ratelimit might_sleep() checks anyway.  But, we
do the ratelimiting *after* we check the other conditions for
might_sleep() including the (costly) irqs_disabled() call.

If we flip these around and ratelimit _before_ the other checks,
I see a boost in the microbenchmark.

The downside here is that we end up doing more frequent updates
to the global 'prev_jiffy'.  But, we're still only actually
updating it once per jiffy.  I tested this on an 80-core system
and my test scales better with this patch applied than without
it, which made me feel a bit better that the global updates to
'prev_jiffy' won't be that painful in practice.

Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
---

 b/kernel/sched/core.c |    7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff -puN kernel/sched/core.c~might-sleep-ratelimit-first kernel/sched/core.c
--- a/kernel/sched/core.c~might-sleep-ratelimit-first	2015-06-24 16:57:24.643850450 -0700
+++ b/kernel/sched/core.c	2015-06-24 16:57:24.650850764 -0700
@@ -7330,13 +7330,14 @@ void ___might_sleep(const char *file, in
 	static unsigned long prev_jiffy;	/* ratelimiting */
 
 	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
+	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
+		return;
+	prev_jiffy = jiffies;
+
 	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
 	     !is_idle_task(current)) ||
 	    system_state != SYSTEM_RUNNING || oops_in_progress)
 		return;
-	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
-		return;
-	prev_jiffy = jiffies;
 
 	printk(KERN_ERR
 		"BUG: sleeping function called from invalid context at %s:%d\n",
_

Re: [RFC][PATCH] sched: might_sleep(): do rate-limiting before sanity checks

[Dave Hansen]

On 06/24/2015 05:03 PM, Dave Hansen wrote:
> In any case, we ratelimit might_sleep() checks anyway.  But, we
> do the ratelimiting *after* we check the other conditions for
> might_sleep() including the (costly) irqs_disabled() call.

Thinking about this a bit more, this patch is wrong.

This only does a _check_ once per jiffy instead of just one warning per
jiffy, which is totally bogus.

I would be interested, though, if anybody has any ideas about speeding
up the irqs_disabled() checking.

Re: [RFC][PATCH] sched: might_sleep(): do rate-limiting before sanity checks

[Peter Zijlstra]

On Wed, Jun 24, 2015 at 08:36:38PM -0700, Dave Hansen wrote:
> On 06/24/2015 05:03 PM, Dave Hansen wrote:
> > In any case, we ratelimit might_sleep() checks anyway.  But, we
> > do the ratelimiting *after* we check the other conditions for
> > might_sleep() including the (costly) irqs_disabled() call.
> 
> Thinking about this a bit more, this patch is wrong.
> 
> This only does a _check_ once per jiffy instead of just one warning per
> jiffy, which is totally bogus.
> 
> I would be interested, though, if anybody has any ideas about speeding
> up the irqs_disabled() checking.

Thomas has something up his sleeve..