Skip to contentSkip to conversation

count_stat_eventual: Remove unnecessary READ_ONCE() and smp_mb()

PATCH In review Junchang Wang 3 participants 5 messages May 16, 2017 – May 17, 2017

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 count_stat_eventual: Remove unnecessary READ_ONCE() and smp_mb()

In function eventual(), if the value of stopflag become larger than zero (the ''if'' branch is taken), then only the "eventual( )" thread updates stopflag. So, the READ_ONCE() within the WRITE_ONCE() is not necessary. Remove it.

In function count_cleanup(), there is no need to run smp_mb( ) in between statement writing to and statement reading from the same variable, stopflag. Remove smp_mb().

---
 CodeSamples/count/count_stat_eventual.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
CodeSamples/count/count_stat_eventual.c +4 -4
index 464de30..1365168 100644
@@ -40,16 +40,17 @@ void *eventual(void *arg)
 {
 	int t;
 	unsigned long sum;
+	int stopflag_l = 0;
 
-	while (READ_ONCE(stopflag) < 3) {
+	while (stopflag_l < 3) {
 		sum = 0;
 		for_each_thread(t)
 			sum += READ_ONCE(per_thread(counter, t));
 		WRITE_ONCE(global_count, sum);
 		poll(NULL, 0, 1);
-		if (READ_ONCE(stopflag)) {
+		if ((stopflag_l = READ_ONCE(stopflag)) != 0) {
 			smp_mb();
-			WRITE_ONCE(stopflag, READ_ONCE(stopflag) + 1);
+			WRITE_ONCE(stopflag, ++stopflag_l);
 		}
 	}
 	return NULL;
@@ -68,7 +69,6 @@ void count_init(void)
 void count_cleanup(void)
 {
 	WRITE_ONCE(stopflag, 1);
-	smp_mb();
 	while (READ_ONCE(stopflag) < 3)
 		poll(NULL, 0, 1);
 	smp_mb();
Signature · 1 line
2.7.4
#1 lore raw
Quoted 10 lines · Junchang Wang

In function eventual(), if the value of stopflag become larger than zero (the ''if'' branch is taken), then only the "eventual( )" thread updates stopflag. So, the READ_ONCE() within the WRITE_ONCE() is not necessary. Remove it.

In function count_cleanup(), there is no need to run smp_mb( ) in between statement writing to and statement reading from the same variable, stopflag. Remove smp_mb().

Removing the memory barrier is good. Removing the stopflag_l local variable is presumably intended to remove one load instruction per pass through the outer loop, assuming that the stopflag_l variable is kept in a register.

Is the performance benefit actually visible? My guess is "no". If the performance is not substantial, my thought would be to keep the code simpler, given that this is a textbook.

And yes, there might be other performance-neutral simplifications possible, and I would welcome patches to that effect.

Thanx, Paul

Quoted 40 lines
---
 CodeSamples/count/count_stat_eventual.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
CodeSamples/count/count_stat_eventual.c +4 -4
index 464de30..1365168 100644
@@ -40,16 +40,17 @@ void *eventual(void *arg)
 {
 	int t;
 	unsigned long sum;
+	int stopflag_l = 0;
 
-	while (READ_ONCE(stopflag) < 3) {
+	while (stopflag_l < 3) {
 		sum = 0;
 		for_each_thread(t)
 			sum += READ_ONCE(per_thread(counter, t));
 		WRITE_ONCE(global_count, sum);
 		poll(NULL, 0, 1);
-		if (READ_ONCE(stopflag)) {
+		if ((stopflag_l = READ_ONCE(stopflag)) != 0) {
 			smp_mb();
-			WRITE_ONCE(stopflag, READ_ONCE(stopflag) + 1);
+			WRITE_ONCE(stopflag, ++stopflag_l);
 		}
 	}
 	return NULL;
@@ -68,7 +69,6 @@ void count_init(void)
 void count_cleanup(void)
 {
 	WRITE_ONCE(stopflag, 1);
-	smp_mb();
 	while (READ_ONCE(stopflag) < 3)
 		poll(NULL, 0, 1);
 	smp_mb();
Signature · 1 line
2.7.4
#2 lore raw
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Junchang Wang

In function eventual(), if the value of stopflag become larger than zero (the ''if'' branch is taken), then only the "eventual( )" thread updates stopflag. So, the READ_ONCE() within the WRITE_ONCE() is not necessary. Remove it.

In function count_cleanup(), there is no need to run smp_mb( ) in between statement writing to and statement reading from the same variable, stopflag. Remove smp_mb().

Quoted 9 lines

Removing the memory barrier is good. Removing the stopflag_l local variable is presumably intended to remove one load instruction per pass through the outer loop, assuming that the stopflag_l variable is kept in a register.

Is the performance benefit actually visible? My guess is "no". If the performance is not substantial, my thought would be to keep the code simpler, given that this is a textbook.

Hi Paul,

Agree. Thanks for the note. Anyway, let's submit a new patch on smp_mb(), which makes the code correct. For simplifying stopflag, we can consider submitting a new patch later. How does that sound like?

Thanks,
--Jason

Quoted 46 lines

And yes, there might be other performance-neutral simplifications possible, and I would welcome patches to that effect.

Thanx, Paul

Quoted 40 lines
---
 CodeSamples/count/count_stat_eventual.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
CodeSamples/count/count_stat_eventual.c +4 -5
index 464de30..1365168 100644
@@ -40,16 +40,17 @@ void *eventual(void *arg)
 {
      int t;
      unsigned long sum;
+     int stopflag_l = 0;
 
-     while (READ_ONCE(stopflag) < 3) {
+     while (stopflag_l < 3) {
              sum = 0;
              for_each_thread(t)
                      sum += READ_ONCE(per_thread(counter, t));
              WRITE_ONCE(global_count, sum);
              poll(NULL, 0, 1);
-             if (READ_ONCE(stopflag)) {
+             if ((stopflag_l = READ_ONCE(stopflag)) != 0) {
                      smp_mb();
-                     WRITE_ONCE(stopflag, READ_ONCE(stopflag) + 1);
+                     WRITE_ONCE(stopflag, ++stopflag_l);
              }
      }
      return NULL;
@@ -68,7 +69,6 @@ void count_init(void)
 void count_cleanup(void)
 {
      WRITE_ONCE(stopflag, 1);
-     smp_mb();
      while (READ_ONCE(stopflag) < 3)
              poll(NULL, 0, 1);
      smp_mb();
--

2.7.4

#3 lore raw
Quoted 24 lines · Junchang Wang
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Junchang Wang

In function eventual(), if the value of stopflag become larger than zero (the ''if'' branch is taken), then only the "eventual( )" thread updates stopflag. So, the READ_ONCE() within the WRITE_ONCE() is not necessary. Remove it.

In function count_cleanup(), there is no need to run smp_mb( ) in between statement writing to and statement reading from the same variable, stopflag. Remove smp_mb().

Quoted 9 lines

Removing the memory barrier is good. Removing the stopflag_l local variable is presumably intended to remove one load instruction per pass through the outer loop, assuming that the stopflag_l variable is kept in a register.

Is the performance benefit actually visible? My guess is "no". If the performance is not substantial, my thought would be to keep the code simpler, given that this is a textbook.

Quoted 5 lines

Hi Paul,

Agree. Thanks for the note. Anyway, let's submit a new patch on smp_mb(), which makes the code correct. For simplifying stopflag, we can consider submitting a new patch later. How does that sound like?

Hi Jason & Paul,

For a material in Chapter 5, I agree that removing excess READ_ONCE() is not necessary.

Such optimizations could be discussed somewhere (maybe a Quick Quiz) in Chapter 14, but I'm not sure.

So, the revised patch just removing a smp_mb() looks good to me.

Thanks, Akira.

Quoted 53 lines

Thanks,
--Jason

Quoted 46 lines

And yes, there might be other performance-neutral simplifications possible, and I would welcome patches to that effect.

Thanx, Paul

Quoted 40 lines
---
 CodeSamples/count/count_stat_eventual.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
CodeSamples/count/count_stat_eventual.c +4 -5
index 464de30..1365168 100644
@@ -40,16 +40,17 @@ void *eventual(void *arg)
 {
      int t;
      unsigned long sum;
+     int stopflag_l = 0;
 
-     while (READ_ONCE(stopflag) < 3) {
+     while (stopflag_l < 3) {
              sum = 0;
              for_each_thread(t)
                      sum += READ_ONCE(per_thread(counter, t));
              WRITE_ONCE(global_count, sum);
              poll(NULL, 0, 1);
-             if (READ_ONCE(stopflag)) {
+             if ((stopflag_l = READ_ONCE(stopflag)) != 0) {
                      smp_mb();
-                     WRITE_ONCE(stopflag, READ_ONCE(stopflag) + 1);
+                     WRITE_ONCE(stopflag, ++stopflag_l);
              }
      }
      return NULL;
@@ -68,7 +69,6 @@ void count_init(void)
 void count_cleanup(void)
 {
      WRITE_ONCE(stopflag, 1);
-     smp_mb();
      while (READ_ONCE(stopflag) < 3)
              poll(NULL, 0, 1);
      smp_mb();
--

2.7.4

#4 lore raw
Quoted 33 lines · Akira Yokosawa
Quoted 24 lines · Junchang Wang
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Junchang Wang

In function eventual(), if the value of stopflag become larger than zero (the ''if'' branch is taken), then only the "eventual( )" thread updates stopflag. So, the READ_ONCE() within the WRITE_ONCE() is not necessary. Remove it.

In function count_cleanup(), there is no need to run smp_mb( ) in between statement writing to and statement reading from the same variable, stopflag. Remove smp_mb().

Quoted 9 lines

Removing the memory barrier is good. Removing the stopflag_l local variable is presumably intended to remove one load instruction per pass through the outer loop, assuming that the stopflag_l variable is kept in a register.

Is the performance benefit actually visible? My guess is "no". If the performance is not substantial, my thought would be to keep the code simpler, given that this is a textbook.

Quoted 5 lines

Hi Paul,

Agree. Thanks for the note. Anyway, let's submit a new patch on smp_mb(), which makes the code correct. For simplifying stopflag, we can consider submitting a new patch later. How does that sound like?

Hi Jason & Paul,

Quoted 7 lines

For a material in Chapter 5, I agree that removing excess READ_ONCE() is not necessary.

Such optimizations could be discussed somewhere (maybe a Quick Quiz) in Chapter 14, but I'm not sure.

So, the revised patch just removing a smp_mb() looks good to me.

Very good, queued and pushed. Thank you both!

Thanx, Paul

Quoted 56 lines

Thanks, Akira.

Quoted 53 lines

Thanks,
--Jason

Quoted 46 lines

And yes, there might be other performance-neutral simplifications possible, and I would welcome patches to that effect.

Thanx, Paul

Quoted 40 lines
---
 CodeSamples/count/count_stat_eventual.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
CodeSamples/count/count_stat_eventual.c +4 -5
index 464de30..1365168 100644
@@ -40,16 +40,17 @@ void *eventual(void *arg)
 {
      int t;
      unsigned long sum;
+     int stopflag_l = 0;
 
-     while (READ_ONCE(stopflag) < 3) {
+     while (stopflag_l < 3) {
              sum = 0;
              for_each_thread(t)
                      sum += READ_ONCE(per_thread(counter, t));
              WRITE_ONCE(global_count, sum);
              poll(NULL, 0, 1);
-             if (READ_ONCE(stopflag)) {
+             if ((stopflag_l = READ_ONCE(stopflag)) != 0) {
                      smp_mb();
-                     WRITE_ONCE(stopflag, READ_ONCE(stopflag) + 1);
+                     WRITE_ONCE(stopflag, ++stopflag_l);
              }
      }
      return NULL;
@@ -68,7 +69,6 @@ void count_init(void)
 void count_cleanup(void)
 {
      WRITE_ONCE(stopflag, 1);
-     smp_mb();
      while (READ_ONCE(stopflag) < 3)
              poll(NULL, 0, 1);
      smp_mb();
--

2.7.4