Re: [patch v3 5/6] sched, ttwu_queue: queue remote wakeups onlywhen crossing cache domains

[Peter Zijlstra]

On Fri, 2011-12-02 at 04:34 +0100, Mike Galbraith wrote:
> On Thu, 2011-12-01 at 17:07 -0800, Suresh Siddha wrote:
> > plain text document attachment
> > (use_ttwu_queue_when_crossing_cache_domains.patch)
> > From: Mike Galbraith <efault@xxxxxx>
> > 
> > Context-switch intensive microbenchmark on a 8-socket system had
> > ~600K times more resched IPI's on each logical CPU because of the
> > TTWU_QUEUE sched feature, which queues the task on the remote cpu's
> > queue and completes the wakeup locally using an IPI.
> > 
> > As the TTWU_QUEUE sched feature is for minimizing the cache-misses
> > associated with the remote wakeups, use the IPI only when the local and
> > the remote cpu's are from different cache domains. Otherwise use the
> > traditional remote wakeup.
> 
> FYI, Peter has already (improved and) queued this patch.

In fact, Ingo (rightfully) refused to take this due to the x86 specific
code in scheduler guts..

Initially the idea was to provide a new arch interface and a fallback
and do the Kconfig thing etc. After a bit of thought I decided against
that for we already have that information in the sched_domain tree
anyway and it should be a simple matter of representing things
differently.

This led to the below patch, which seems to boot on my box. I still hate
the sd_top_spr* names but whatever.. ;-)

---
 kernel/sched/core.c  |   36 +++++++++++++++++++++++++++++++++++-
 kernel/sched/fair.c  |   24 +-----------------------
 kernel/sched/sched.h |   42 ++++++++++++++++++++++++++++++++++++------
 3 files changed, 72 insertions(+), 30 deletions(-)
Index: linux-2.6/kernel/sched/core.c
===================================================================
--- linux-2.6.orig/kernel/sched/core.c
+++ linux-2.6/kernel/sched/core.c
@@ -1511,6 +1511,12 @@ static int ttwu_activate_remote(struct t
 
 }
 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
+
+static inline int ttwu_share_cache(int this_cpu, int that_cpu)
+{
+	return per_cpu(sd_top_spr_id, this_cpu) ==
+		per_cpu(sd_top_spr_id, that_cpu);
+}
 #endif /* CONFIG_SMP */
 
 static void ttwu_queue(struct task_struct *p, int cpu)
@@ -1518,7 +1524,7 @@ static void ttwu_queue(struct task_struc
 	struct rq *rq = cpu_rq(cpu);
 
 #if defined(CONFIG_SMP)
-	if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
+	if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
 		sched_clock_cpu(cpu); /* sync clocks x-cpu */
 		ttwu_queue_remote(p, cpu);
 		return;
@@ -5751,6 +5757,32 @@ static void destroy_sched_domains(struct
 }
 
 /*
+ * Keep a special pointer to the highest sched_domain that has
+ * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
+ * allows us to avoid some pointer chasing select_idle_sibling().
+ *
+ * Also keep a unique ID per domain (we use the first cpu number in
+ * the cpumask of the domain), this allows us to quickly tell if
+ * two cpus are in the same cache domain, see ttwu_share_cache().
+ */
+DEFINE_PER_CPU(struct sched_domain *, sd_top_spr);
+DEFINE_PER_CPU(int, sd_top_spr_id);
+
+static void update_top_cache_domain(int cpu)
+{
+	struct sched_domain *sd;
+	int id = -1;
+
+	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
+	if (sd)
+		id = cpumask_first(sched_domain_span(sd));
+
+
+	rcu_assign_pointer(per_cpu(sd_top_spr, cpu), sd);
+	per_cpu(sd_top_spr_id, cpu) = id;
+}
+
+/*
  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
  * hold the hotplug lock.
  */
@@ -5789,6 +5821,8 @@ cpu_attach_domain(struct sched_domain *s
 	tmp = rq->sd;
 	rcu_assign_pointer(rq->sd, sd);
 	destroy_sched_domains(tmp, cpu);
+
+	update_top_cache_domain(cpu);
 }
 
 /* cpus with isolated domains */
Index: linux-2.6/kernel/sched/fair.c
===================================================================
--- linux-2.6.orig/kernel/sched/fair.c
+++ linux-2.6/kernel/sched/fair.c
@@ -2644,28 +2644,6 @@ find_idlest_cpu(struct sched_group *grou
 	return idlest;
 }
 
-/**
- * highest_flag_domain - Return highest sched_domain containing flag.
- * @cpu:	The cpu whose highest level of sched domain is to
- *		be returned.
- * @flag:	The flag to check for the highest sched_domain
- *		for the given cpu.
- *
- * Returns the highest sched_domain of a cpu which contains the given flag.
- */
-static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
-{
-	struct sched_domain *sd, *hsd = NULL;
-
-	for_each_domain(cpu, sd) {
-		if (!(sd->flags & flag))
-			break;
-		hsd = sd;
-	}
-
-	return hsd;
-}
-
 /*
  * Try and locate an idle CPU in the sched_domain.
  */
@@ -2696,7 +2674,7 @@ static int select_idle_sibling(struct ta
 	 */
 	rcu_read_lock();
 
-	sd = highest_flag_domain(target, SD_SHARE_PKG_RESOURCES);
+	sd = rcu_dereference(per_cpu(sd_top_spr, target));
 	for_each_lower_domain(sd) {
 		sg = sd->groups;
 		do {
Index: linux-2.6/kernel/sched/sched.h
===================================================================
--- linux-2.6.orig/kernel/sched/sched.h
+++ linux-2.6/kernel/sched/sched.h
@@ -487,6 +487,14 @@ static inline int cpu_of(struct rq *rq)
 
 DECLARE_PER_CPU(struct rq, runqueues);
 
+#define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
+#define this_rq()		(&__get_cpu_var(runqueues))
+#define task_rq(p)		cpu_rq(task_cpu(p))
+#define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
+#define raw_rq()		(&__raw_get_cpu_var(runqueues))
+
+#ifdef CONFIG_SMP
+
 #define rcu_dereference_check_sched_domain(p) \
 	rcu_dereference_check((p), \
 			      lockdep_is_held(&sched_domains_mutex))
@@ -499,15 +507,37 @@ DECLARE_PER_CPU(struct rq, runqueues);
  * preempt-disabled sections.
  */
 #define for_each_domain(cpu, __sd) \
-	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
+	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
+			__sd; __sd = __sd->parent)
 
 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
 
-#define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
-#define this_rq()		(&__get_cpu_var(runqueues))
-#define task_rq(p)		cpu_rq(task_cpu(p))
-#define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
-#define raw_rq()		(&__raw_get_cpu_var(runqueues))
+/**
+ * highest_flag_domain - Return highest sched_domain containing flag.
+ * @cpu:	The cpu whose highest level of sched domain is to
+ *		be returned.
+ * @flag:	The flag to check for the highest sched_domain
+ *		for the given cpu.
+ *
+ * Returns the highest sched_domain of a cpu which contains the given flag.
+ */
+static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
+{
+	struct sched_domain *sd, *hsd = NULL;
+
+	for_each_domain(cpu, sd) {
+		if (!(sd->flags & flag))
+			break;
+		hsd = sd;
+	}
+
+	return hsd;
+}
+
+DECLARE_PER_CPU(struct sched_domain *, sd_top_spr);
+DECLARE_PER_CPU(int, sd_top_spr_id);
+
+#endif /* CONFIG_SMP */
 
 #include "stats.h"
 #include "auto_group.h"

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/