sched/core: Optimize sched_feat() for !CONFIG_SCHED_DEBUG builds

When the kernel is compiled with !CONFIG_SCHED_DEBUG support, we expect that
all SCHED_FEAT are turned into compile time constants being propagated
to support compiler optimizations.

Specifically, we expect that code blocks like this:

   if (sched_feat(FEATURE_NAME) [&& <other_conditions>]) {
	/* FEATURE CODE */
   }

are turned into dead-code in case FEATURE_NAME defaults to FALSE, and thus
being removed by the compiler from the finale image.

For this mechanism to properly work it's required for the compiler to
have full access, from each translation unit, to whatever is the value
defined by the sched_feat macro. This macro is defined as:

   #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))

and thus, the compiler can optimize that code only if the value of
sysctl_sched_features is visible within each translation unit.

Since:

   029632fbb ("sched: Make separate sched*.c translation units")

the scheduler code has been split into separate translation units
however the definition of sysctl_sched_features is part of
kernel/sched/core.c while, for all the other scheduler modules, it is
visible only via kernel/sched/sched.h as an:

   extern const_debug unsigned int sysctl_sched_features

Unfortunately, an extern reference does not allow the compiler to apply
constants propagation. Thus, on !CONFIG_SCHED_DEBUG kernel we still end up
with code to load a memory reference and (eventually) doing an unconditional
jump of a chunk of code.

This mechanism is unavoidable when sched_features can be turned on and off at
run-time. However, this is not the case for "production" kernels compiled with
!CONFIG_SCHED_DEBUG. In this case, sysctl_sched_features is just a constant value
which cannot be changed at run-time and thus memory loads and jumps can be
avoided altogether.

This patch fixes the case of !CONFIG_SCHED_DEBUG kernel by declaring a local version
of the sysctl_sched_features constant for each translation unit. This will
ultimately allow the compiler to perform constants propagation and dead-code
pruning.

Tests have been done, with !CONFIG_SCHED_DEBUG on a v4.14-rc8 with and without
the patch, by running 30 iterations of:

   perf bench sched messaging --pipe --thread --group 4 --loop 50000

on a 40 cores Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz using the
powersave governor to rule out variations due to frequency scaling.

Statistics on the reported completion time:

                   count     mean       std     min       99%     max
  v4.14-rc8         30.0  15.7831  0.176032  15.442  16.01226  16.014
  v4.14-rc8+patch   30.0  15.5033  0.189681  15.232  15.93938  15.962

... show a 1.8% speedup on average completion time and 0.5% speedup in the
99 percentile.

Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com>
Signed-off-by: Chris Redpath <chris.redpath@arm.com>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Reviewed-by: Brendan Jackman <brendan.jackman@arm.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Link: http://lkml.kernel.org/r/20171108184101.16006-1-patrick.bellasi@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

2 files changed, 28 insertions, 6 deletions

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 1a55c842bfbc..a39a08113003 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -43,18 +43,21 @@
 
 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 
+#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
 /*
  * Debugging: various feature bits
+ *
+ * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
+ * sysctl_sched_features, defined in sched.h, to allow constants propagation
+ * at compile time and compiler optimization based on features default.
  */
-
 #define SCHED_FEAT(name, enabled)	\
 	(1UL << __SCHED_FEAT_##name) * enabled |
-
 const_debug unsigned int sysctl_sched_features =
 #include "features.h"
 	0;
-
 #undef SCHED_FEAT
+#endif
 
 /*
  * Number of tasks to iterate in a single balance run.
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 58787e3631c7..45ab0bf564e7 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1233,8 +1233,6 @@ static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
 # define const_debug const
 #endif
 
-extern const_debug unsigned int sysctl_sched_features;
-
 #define SCHED_FEAT(name, enabled)	\
 	__SCHED_FEAT_##name ,
 
@@ -1246,6 +1244,13 @@ enum {
 #undef SCHED_FEAT
 
 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
+
+/*
+ * To support run-time toggling of sched features, all the translation units
+ * (but core.c) reference the sysctl_sched_features defined in core.c.
+ */
+extern const_debug unsigned int sysctl_sched_features;
+
 #define SCHED_FEAT(name, enabled)					\
 static __always_inline bool static_branch_##name(struct static_key *key) \
 {									\
@@ -1253,13 +1258,27 @@ static __always_inline bool static_branch_##name(struct static_key *key) \
 }
 
 #include "features.h"
-
 #undef SCHED_FEAT
 
 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
+
 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */
+
+/*
+ * Each translation unit has its own copy of sysctl_sched_features to allow
+ * constants propagation at compile time and compiler optimization based on
+ * features default.
+ */
+#define SCHED_FEAT(name, enabled)	\
+	(1UL << __SCHED_FEAT_##name) * enabled |
+static const_debug __maybe_unused unsigned int sysctl_sched_features =
+#include "features.h"
+	0;
+#undef SCHED_FEAT
+
 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
+
 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */
 
 extern struct static_key_false sched_numa_balancing;