[PATCH] powerpc/perf: Fix for core/nest imc call trace on cpuhotplug

Nest/core pmu units are enabled only when it is used. A reference count is      
maintained for the events which uses the nest/core pmu units. Currently in      
*_imc_counters_release function a WARN() is used for notification of any        
underflow of ref count.                                                         
                                                                                
The case where event ref count hit a negative value is, when perf session is    
started, followed by offlining of all cpus in a given core.                     
i.e. in cpuhotplug offline path ppc_core_imc_cpu_offline() function set the     
ref->count to zero, if the current cpu which is about to offline is the last    
cpu in a given core and make an OPAL call to disable the engine in that core.   
And on perf session termination, perf->destroy (core_imc_counters_release) will 
first decrement the ref->count for this core and based on the ref->count value  
an opal call is made to disable the core-imc engine.                            
Now, since cpuhotplug path already clears the ref->count for core and disabled  
the engine, perf->destroy() decrementing again at event termination make it     
negative which in turn fires the WARN_ON. The same happens for nest units.      
                                                                                
Add a check to see if the reference count is alreday zero, before decrementing  
the count, so that the ref count will not hit a negative value.                 
                                                                                
Signed-off-by: Anju T Sudhakar <redacted>                
---
 arch/powerpc/perf/imc-pmu.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c
index 9ccac86f3463..e3a1f65933b5 100644
--- a/arch/powerpc/perf/imc-pmu.c
+++ b/arch/powerpc/perf/imc-pmu.c

@@ -399,6 +399,20 @@ static void nest_imc_counters_release(struct perf_event *event)
 
 	/* Take the mutex lock for this node and then decrement the reference count */
 	mutex_lock(&ref->lock);
+	if (ref->refc == 0) {
+		/*
+		 * The scenario where this is true is, when perf session is
+		 * started, followed by offlining of all cpus in a given node.
+		 *
+		 * In the cpuhotplug offline path, ppc_nest_imc_cpu_offline()
+		 * function set the ref->count to zero, if the cpu which is
+		 * about to offline is the last cpu in a given node and make
+		 * an OPAL call to disable the engine in that node.
+		 *
+		 */
+		mutex_unlock(&ref->lock);
+		return;
+	}
 	ref->refc--;
 	if (ref->refc == 0) {
 		rc = opal_imc_counters_stop(OPAL_IMC_COUNTERS_NEST,

@@ -646,6 +660,20 @@ static void core_imc_counters_release(struct perf_event *event)
 		return;
 
 	mutex_lock(&ref->lock);
+	if (ref->refc == 0) {
+		/*
+		 * The scenario where this is true is, when perf session is
+		 * started, followed by offlining of all cpus in a given core.
+		 *
+		 * In the cpuhotplug offline path, ppc_core_imc_cpu_offline()
+		 * function set the ref->count to zero, if the cpu which is
+		 * about to offline is the last cpu in a given core and make
+		 * an OPAL call to disable the engine in that core.
+		 *
+		 */
+		mutex_unlock(&ref->lock);
+		return;
+	}
 	ref->refc--;
 	if (ref->refc == 0) {
 		rc = opal_imc_counters_stop(OPAL_IMC_COUNTERS_CORE,

-- 
2.14.1

* Anju T Sudhakar [off-list ref] wrote (on 2017-10-04 06:50:52 +0000):

Nest/core pmu units are enabled only when it is used. A reference count is      
maintained for the events which uses the nest/core pmu units. Currently in      
*_imc_counters_release function a WARN() is used for notification of any        
underflow of ref count.                                                         
                                                                                
The case where event ref count hit a negative value is, when perf session is    
started, followed by offlining of all cpus in a given core.                     
i.e. in cpuhotplug offline path ppc_core_imc_cpu_offline() function set the     
ref->count to zero, if the current cpu which is about to offline is the last    
cpu in a given core and make an OPAL call to disable the engine in that core.   
And on perf session termination, perf->destroy (core_imc_counters_release) will 
first decrement the ref->count for this core and based on the ref->count value  
an opal call is made to disable the core-imc engine.                            
Now, since cpuhotplug path already clears the ref->count for core and disabled  
the engine, perf->destroy() decrementing again at event termination make it     
negative which in turn fires the WARN_ON. The same happens for nest units.      
                                                                                
Add a check to see if the reference count is alreday zero, before decrementing  
the count, so that the ref count will not hit a negative value.                 
                                                                                
Signed-off-by: Anju T Sudhakar <redacted>

Reviewed-by: Santosh Sivaraj <redacted>

-- 

Hi Santosh,


On Thursday 05 October 2017 03:20 PM, Santosh Sivaraj wrote:

* Anju T Sudhakar [off-list ref] wrote (on 2017-10-04 06:50:52 +0000):

quoted

Nest/core pmu units are enabled only when it is used. A reference count is
maintained for the events which uses the nest/core pmu units. Currently in
*_imc_counters_release function a WARN() is used for notification of any
underflow of ref count.
                                                                                 
The case where event ref count hit a negative value is, when perf session is
started, followed by offlining of all cpus in a given core.
i.e. in cpuhotplug offline path ppc_core_imc_cpu_offline() function set the
ref->count to zero, if the current cpu which is about to offline is the last
cpu in a given core and make an OPAL call to disable the engine in that core.
And on perf session termination, perf->destroy (core_imc_counters_release) will
first decrement the ref->count for this core and based on the ref->count value
an opal call is made to disable the core-imc engine.
Now, since cpuhotplug path already clears the ref->count for core and disabled
the engine, perf->destroy() decrementing again at event termination make it
negative which in turn fires the WARN_ON. The same happens for nest units.
                                                                                 
Add a check to see if the reference count is alreday zero, before decrementing
the count, so that the ref count will not hit a negative value.
                                                                                 
Signed-off-by: Anju T Sudhakar <redacted>

Reviewed-by: Santosh Sivaraj <redacted>

Thanks for reviewing.

-Anju

quoted

---
  arch/powerpc/perf/imc-pmu.c | 28 ++++++++++++++++++++++++++++
  1 file changed, 28 insertions(+)

diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c
index 9ccac86f3463..e3a1f65933b5 100644
--- a/arch/powerpc/perf/imc-pmu.c
+++ b/arch/powerpc/perf/imc-pmu.c

@@ -399,6 +399,20 @@ static void nest_imc_counters_release(struct perf_event *event)
  
  	/* Take the mutex lock for this node and then decrement the reference count */
  	mutex_lock(&ref->lock);
+	if (ref->refc == 0) {
+		/*
+		 * The scenario where this is true is, when perf session is
+		 * started, followed by offlining of all cpus in a given node.
+		 *
+		 * In the cpuhotplug offline path, ppc_nest_imc_cpu_offline()
+		 * function set the ref->count to zero, if the cpu which is
+		 * about to offline is the last cpu in a given node and make
+		 * an OPAL call to disable the engine in that node.
+		 *
+		 */
+		mutex_unlock(&ref->lock);
+		return;
+	}
  	ref->refc--;
  	if (ref->refc == 0) {
  		rc = opal_imc_counters_stop(OPAL_IMC_COUNTERS_NEST,

@@ -646,6 +660,20 @@ static void core_imc_counters_release(struct perf_event *event)
  		return;
  
  	mutex_lock(&ref->lock);
+	if (ref->refc == 0) {
+		/*
+		 * The scenario where this is true is, when perf session is
+		 * started, followed by offlining of all cpus in a given core.
+		 *
+		 * In the cpuhotplug offline path, ppc_core_imc_cpu_offline()
+		 * function set the ref->count to zero, if the cpu which is
+		 * about to offline is the last cpu in a given core and make
+		 * an OPAL call to disable the engine in that core.
+		 *
+		 */
+		mutex_unlock(&ref->lock);
+		return;
+	}
  	ref->refc--;
  	if (ref->refc == 0) {
  		rc = opal_imc_counters_stop(OPAL_IMC_COUNTERS_CORE,

On Wed, 2017-10-04 at 06:50:52 UTC, Anju T Sudhakar wrote:

Nest/core pmu units are enabled only when it is used. A reference count is      
maintained for the events which uses the nest/core pmu units. Currently in      
*_imc_counters_release function a WARN() is used for notification of any        
underflow of ref count.                                                         
                                                                                
The case where event ref count hit a negative value is, when perf session is    
started, followed by offlining of all cpus in a given core.                     
i.e. in cpuhotplug offline path ppc_core_imc_cpu_offline() function set the     
ref->count to zero, if the current cpu which is about to offline is the last    
cpu in a given core and make an OPAL call to disable the engine in that core.   
And on perf session termination, perf->destroy (core_imc_counters_release) will 
first decrement the ref->count for this core and based on the ref->count value  
an opal call is made to disable the core-imc engine.                            
Now, since cpuhotplug path already clears the ref->count for core and disabled  
the engine, perf->destroy() decrementing again at event termination make it     
negative which in turn fires the WARN_ON. The same happens for nest units.      
                                                                                
Add a check to see if the reference count is alreday zero, before decrementing  
the count, so that the ref count will not hit a negative value.                 
                                                                                
Signed-off-by: Anju T Sudhakar <redacted>
Reviewed-by: Santosh Sivaraj <redacted>

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/0d923820c6db1644c27c2d0a5af892

cheers