Re: [patch v3 3/3] percpu_counter: use atomic64 for counter in SMP

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2011-04-18 at 22:25 +0800, Christoph Lameter wrote:
> On Mon, 18 Apr 2011, Shaohua Li wrote:
> 
> > > Disabling preemption here doesn't make any sense.
> > > percpu_counter_set() inherently requires its users to guarantee that
> > > no other user is modifying the percpu counter.
> > ha, ok.
> > should I still rebase the patch against Christoph's patch? Looks that
> > one is still not settled down.
> 
> I am a kind of confused about some of the arguments made there right now
> and having your patch in that does the conversion to atomic would
> simplify my patch (removes the spin_lock/unlock sequence in overflow
> handling).
below is the updated patch against Christoph's latest patch.


Subject: percpu_counter: use atomic64 for counter in SMP

Uses atomic64 for percpu_counter, because it is cheaper than spinlock. This
doesn't slow fast path (percpu_counter_read). atomic64_read equals to fbc->count
for 64-bit system, or equals to spin_lock-read-spin_unlock for 32-bit system

Note, originally the percpu_counter_read for 32-bit system doesn't hold
spin_lock, but that is buggy and might cause very wrong value accessed.
This patch fixes the issue.

This can also improve some workloads with percpu_counter->lock heavily
contented. For example, vm_committed_as sometimes causes the contention.
We should tune the batch count, but if we can make percpu_counter better,
why not? In a 24 CPUs system and 24 processes, each runs:
while (1) {
	mmap(128M);
	munmap(128M);
}
we then measure how many loops each process can take:
orig: 1226976
patched: 8210626
The atomic method gives 7x faster.

In percpu_counter_set() and __percpu_counter_sum(), there will be no lock
protecting. This means we might get inprecise count, but we have the same issue
even with lock protecting, because __percpu_counter_add doesn't hold locking
to update cpu local count.

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
---
 include/linux/percpu_counter.h |   18 ++++--------------
 lib/percpu_counter.c           |   37 +++++++++++++++----------------------
 2 files changed, 19 insertions(+), 36 deletions(-)

Index: linux/include/linux/percpu_counter.h
===================================================================
--- linux.orig/include/linux/percpu_counter.h	2011-04-19 08:52:14.000000000 +0800
+++ linux/include/linux/percpu_counter.h	2011-04-19 08:52:58.000000000 +0800
@@ -16,8 +16,7 @@
 #ifdef CONFIG_SMP
 
 struct percpu_counter {
-	spinlock_t lock;
-	s64 count;
+	atomic64_t count;
 #ifdef CONFIG_HOTPLUG_CPU
 	struct list_head list;	/* All percpu_counters are on a list */
 #endif
@@ -26,16 +25,7 @@ struct percpu_counter {
 
 extern int percpu_counter_batch;
 
-int __percpu_counter_init(struct percpu_counter *fbc, s64 amount,
-			  struct lock_class_key *key);
-
-#define percpu_counter_init(fbc, value)					\
-	({								\
-		static struct lock_class_key __key;			\
-									\
-		__percpu_counter_init(fbc, value, &__key);		\
-	})
-
+int percpu_counter_init(struct percpu_counter *fbc, s64 amount);
 void percpu_counter_destroy(struct percpu_counter *fbc);
 void percpu_counter_set(struct percpu_counter *fbc, s64 amount);
 void __percpu_counter_add(struct percpu_counter *fbc, s64 amount, s32 batch);
@@ -60,7 +50,7 @@ static inline s64 percpu_counter_sum(str
 
 static inline s64 percpu_counter_read(struct percpu_counter *fbc)
 {
-	return fbc->count;
+	return atomic64_read(&fbc->count);
 }
 
 /*
@@ -70,7 +60,7 @@ static inline s64 percpu_counter_read(st
  */
 static inline s64 percpu_counter_read_positive(struct percpu_counter *fbc)
 {
-	s64 ret = fbc->count;
+	s64 ret = percpu_counter_read(fbc);
 
 	barrier();		/* Prevent reloads of fbc->count */
 	if (ret >= 0)
Index: linux/lib/percpu_counter.c
===================================================================
--- linux.orig/lib/percpu_counter.c	2011-04-19 08:52:10.000000000 +0800
+++ linux/lib/percpu_counter.c	2011-04-19 08:53:33.000000000 +0800
@@ -59,13 +59,11 @@ void percpu_counter_set(struct percpu_co
 {
 	int cpu;
 
-	spin_lock(&fbc->lock);
 	for_each_possible_cpu(cpu) {
 		s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
 		*pcount = 0;
 	}
-	fbc->count = amount;
-	spin_unlock(&fbc->lock);
+	atomic64_set(&fbc->count, amount);
 }
 EXPORT_SYMBOL(percpu_counter_set);
 
@@ -85,11 +83,8 @@ void __percpu_counter_add(struct percpu_
 			overflow = 0;
 	} while (this_cpu_cmpxchg(*fbc->counters, count, new) != count);
 
-	if (unlikely(overflow)) {
-		spin_lock(&fbc->lock);
-		fbc->count += overflow;
-		spin_unlock(&fbc->lock);
-	}
+	if (unlikely(overflow))
+		atomic64_add(overflow, &fbc->count);
 }
 EXPORT_SYMBOL(__percpu_counter_add);
 
@@ -99,26 +94,27 @@ EXPORT_SYMBOL(__percpu_counter_add);
  */
 s64 __percpu_counter_sum(struct percpu_counter *fbc)
 {
-	s64 ret;
+	s64 ret = 0;
 	int cpu;
 
-	spin_lock(&fbc->lock);
-	ret = fbc->count;
+	/*
+	 * Don't really need to disable preempt here, just make sure there is
+	 * no big latency because of preemption
+	 */
+	preempt_disable();
 	for_each_online_cpu(cpu) {
 		s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
 		ret += *pcount;
 	}
-	spin_unlock(&fbc->lock);
+	ret += atomic64_read(&fbc->count);
+	preempt_enable();
 	return ret;
 }
 EXPORT_SYMBOL(__percpu_counter_sum);
 
-int __percpu_counter_init(struct percpu_counter *fbc, s64 amount,
-			  struct lock_class_key *key)
+int percpu_counter_init(struct percpu_counter *fbc, s64 amount)
 {
-	spin_lock_init(&fbc->lock);
-	lockdep_set_class(&fbc->lock, key);
-	fbc->count = amount;
+	atomic64_set(&fbc->count, amount);
 	fbc->counters = alloc_percpu(s32);
 	if (!fbc->counters)
 		return -ENOMEM;
@@ -133,7 +129,7 @@ int __percpu_counter_init(struct percpu_
 #endif
 	return 0;
 }
-EXPORT_SYMBOL(__percpu_counter_init);
+EXPORT_SYMBOL(percpu_counter_init);
 
 void percpu_counter_destroy(struct percpu_counter *fbc)
 {
@@ -177,13 +173,10 @@ static int __cpuinit percpu_counter_hotc
 	mutex_lock(&percpu_counters_lock);
 	list_for_each_entry(fbc, &percpu_counters, list) {
 		s32 *pcount;
-		unsigned long flags;
 
-		spin_lock_irqsave(&fbc->lock, flags);
 		pcount = per_cpu_ptr(fbc->counters, cpu);
-		fbc->count += *pcount;
+		atomic64_add(*pcount, &fbc->count);
 		*pcount = 0;
-		spin_unlock_irqrestore(&fbc->lock, flags);
 	}
 	mutex_unlock(&percpu_counters_lock);
 #endif