Re: [PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Konstantin Khlebnikov]

On 19.05.2017 12:09, Peter Zijlstra wrote:

On Fri, May 19, 2017 at 10:06:01AM +0300, Konstantin Khlebnikov wrote:

Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
This patch disables it completely if this option is disabled.

This counter might be useful for debugging lockdep itself, but for normal
debugging it seems useless. Lockstat provides more detailed statistics.

This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.

With patch "netperf -H localhost" shows some boost: 2500 -> 2600


Nothing against these patches, but I do wonder why you're performance
optimizing debug kernels ;-)



Performance of debug features matters too
Obviously this saves time and money for massive testing

Re: [PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Konstantin Khlebnikov]

On 19.05.2017 12:09, Peter Zijlstra wrote:

On Fri, May 19, 2017 at 10:06:01AM +0300, Konstantin Khlebnikov wrote:

Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
This patch disables it completely if this option is disabled.

This counter might be useful for debugging lockdep itself, but for normal
debugging it seems useless. Lockstat provides more detailed statistics.

This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.

With patch "netperf -H localhost" shows some boost: 2500 -> 2600


Nothing against these patches, but I do wonder why you're performance
optimizing debug kernels ;-)



Performance of debug features matters too
Obviously this saves time and money for massive testing

Re: [PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Peter Zijlstra]

On Fri, May 19, 2017 at 10:06:01AM +0300, Konstantin Khlebnikov wrote:
> Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
> This patch disables it completely if this option is disabled.
> 
> This counter might be useful for debugging lockdep itself, but for normal
> debugging it seems useless. Lockstat provides more detailed statistics.
> 
> This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.
> 
> With patch "netperf -H localhost" shows some boost: 2500 -> 2600

Nothing against these patches, but I do wonder why you're performance
optimizing debug kernels ;-)

Re: [PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Peter Zijlstra]

On Fri, May 19, 2017 at 10:06:01AM +0300, Konstantin Khlebnikov wrote:
> Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
> This patch disables it completely if this option is disabled.
> 
> This counter might be useful for debugging lockdep itself, but for normal
> debugging it seems useless. Lockstat provides more detailed statistics.
> 
> This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.
> 
> With patch "netperf -H localhost" shows some boost: 2500 -> 2600

Nothing against these patches, but I do wonder why you're performance
optimizing debug kernels ;-)

[PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Konstantin Khlebnikov]

Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
This patch disables it completely if this option is disabled.

This counter might be useful for debugging lockdep itself, but for normal
debugging it seems useless. Lockstat provides more detailed statistics.

This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.

With patch "netperf -H localhost" shows some boost: 2500 -> 2600

Signed-off-by: Konstantin Khlebnikov 
---
 kernel/locking/lockdep.c |4 
 1 file changed, 4 insertions(+)

diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
index c0e31bfee25c..60bd16984546 100644
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -1375,7 +1375,9 @@ static void print_lock_class_header(struct lock_class 
*class, int depth)
 
printk("%*s->", depth, "");
print_lock_name(class);
+#ifdef CONFIG_DEBUG_LOCKDEP
printk(KERN_CONT " ops: %lu", class->ops);
+#endif
printk(KERN_CONT " {\n");
 
for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
@@ -3256,7 +3258,9 @@ static int __lock_acquire(struct lockdep_map *lock, 
unsigned int subclass,
if (!class)
return 0;
}
+#ifdef CONFIG_DEBUG_LOCKDEP
atomic_inc((atomic_t *)>ops);
+#endif
if (very_verbose(class)) {
printk("\nacquire class [%p] %s", class->key, class->name);
if (class->name_version > 1)

[PATCH] lockdep: do not count lock class operations without CONFIG_DEBUG_LOCKDEP

[Konstantin Khlebnikov]

Currently this counter shown in /proc/lockdep if CONFIG_DEBUG_LOCKDEP=y
This patch disables it completely if this option is disabled.

This counter might be useful for debugging lockdep itself, but for normal
debugging it seems useless. Lockstat provides more detailed statistics.

This atomic_inc is a hot spot inside __lock_acquire() for debug kernel.

With patch "netperf -H localhost" shows some boost: 2500 -> 2600

Signed-off-by: Konstantin Khlebnikov 
---
 kernel/locking/lockdep.c |4 
 1 file changed, 4 insertions(+)

diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
index c0e31bfee25c..60bd16984546 100644
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -1375,7 +1375,9 @@ static void print_lock_class_header(struct lock_class 
*class, int depth)
 
printk("%*s->", depth, "");
print_lock_name(class);
+#ifdef CONFIG_DEBUG_LOCKDEP
printk(KERN_CONT " ops: %lu", class->ops);
+#endif
printk(KERN_CONT " {\n");
 
for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
@@ -3256,7 +3258,9 @@ static int __lock_acquire(struct lockdep_map *lock, 
unsigned int subclass,
if (!class)
return 0;
}
+#ifdef CONFIG_DEBUG_LOCKDEP
atomic_inc((atomic_t *)>ops);
+#endif
if (very_verbose(class)) {
printk("\nacquire class [%p] %s", class->key, class->name);
if (class->name_version > 1)