sched: cpuacct: Use bigger percpu counter batch values for stats counters

When CONFIG_VIRT_CPU_ACCOUNTING and CONFIG_CGROUP_CPUACCT are
enabled we can call cpuacct_update_stats with values much larger
than percpu_counter_batch.  This means the call to
percpu_counter_add will always add to the global count which is
protected by a spinlock and we end up with a global spinlock in
the scheduler.

Based on an idea by KOSAKI Motohiro, this patch scales the batch
value by cputime_one_jiffy such that we have the same batch
limit as we would if CONFIG_VIRT_CPU_ACCOUNTING was disabled.
His patch did this once at boot but that initialisation happened
too early on PowerPC (before time_init) and it was never updated
at runtime as a result of a hotplug cpu add/remove.

This patch instead scales percpu_counter_batch by
cputime_one_jiffy at runtime, which keeps the batch correct even
after cpu hotplug operations.  We cap it at INT_MAX in case of
overflow.

For architectures that do not support
CONFIG_VIRT_CPU_ACCOUNTING, cputime_one_jiffy is the constant 1
and gcc is smart enough to optimise min(s32
percpu_counter_batch, INT_MAX) to just percpu_counter_batch at
least on x86 and PowerPC.  So there is no need to add an #ifdef.

On a 64 thread PowerPC box with CONFIG_VIRT_CPU_ACCOUNTING and
CONFIG_CGROUP_CPUACCT enabled, a context switch microbenchmark
is 234x faster and almost matches a CONFIG_CGROUP_CPUACCT
disabled kernel:

 CONFIG_CGROUP_CPUACCT disabled:   16906698 ctx switches/sec
 CONFIG_CGROUP_CPUACCT enabled:       61720 ctx switches/sec
 CONFIG_CGROUP_CPUACCT + patch:	   16663217 ctx switches/sec

Tested with:

 wget http://ozlabs.org/~anton/junkcode/context_switch.c
 make context_switch
 for i in `seq 0 63`; do taskset -c $i ./context_switch & done
 vmstat 1

Signed-off-by: Anton Blanchard <anton@samba.org>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/kernel/sched.c b/kernel/sched.c
index f96be93..bae6fcf 100644 (file)
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -8997,6 +8997,23 @@ static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
        rcu_read_unlock();
 }
 
+/*
+ * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
+ * in cputime_t units. As a result, cpuacct_update_stats calls
+ * percpu_counter_add with values large enough to always overflow the
+ * per cpu batch limit causing bad SMP scalability.
+ *
+ * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
+ * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
+ * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
+ */
+#ifdef CONFIG_SMP
+#define CPUACCT_BATCH  \
+       min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
+#else
+#define CPUACCT_BATCH  0
+#endif
+
 /*
  * Charge the system/user time to the task's accounting group.
  */
@@ -9004,6 +9021,7 @@ static void cpuacct_update_stats(struct task_struct *tsk,
                enum cpuacct_stat_index idx, cputime_t val)
 {
        struct cpuacct *ca;
+       int batch = CPUACCT_BATCH;
 
        if (unlikely(!cpuacct_subsys.active))
                return;
@@ -9012,7 +9030,7 @@ static void cpuacct_update_stats(struct task_struct *tsk,
        ca = task_ca(tsk);
 
        do {
-               percpu_counter_add(&ca->cpustat[idx], val);
+               __percpu_counter_add(&ca->cpustat[idx], val, batch);
                ca = ca->parent;
        } while (ca);
        rcu_read_unlock();