Various fixes

src/Makefile.am

@@ -9,6 +9,7 @@ libflexmalloc_la_SOURCES = \
  allocator.cxx allocator.hxx \
  allocator-posix.cxx allocator-posix.hxx \
  allocator-statistics.cxx allocator-statistics.hxx \
+ statistics-recorder-allocator.hxx \
  cache-callstack.cxx cache-callstack.hxx \
  flex-malloc.cxx flex-malloc.hxx \
  malloc-interposer.cxx

src/allocator-memkind-hbwmalloc.cxx

@@ -15,7 +15,7 @@
 #define ALLOCATOR_NAME "memkind/hbwmalloc"
 
 AllocatorMemkindHBWMalloc::AllocatorMemkindHBWMalloc (allocation_functions_t &af)
-  : Allocator (af)
+  : StatisticsRecorderAllocator (af)
 {
 	hbw_set_policy (HBW_POLICY_PREFERRED);
 }
@@ -38,7 +38,7 @@ void * AllocatorMemkindHBWMalloc::malloc (size_t size)
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr & base at %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), name(), this);
-		_stats.record_malloc (size);
+		record_malloc (size);
 	}
 
 	return res;
@@ -59,7 +59,7 @@ void * AllocatorMemkindHBWMalloc::calloc (size_t nmemb, size_t size)
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr & base %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), name(), this);
-		_stats.record_calloc (nmemb * size);
+		record_calloc (nmemb * size);
 	}
 
 	return res;
@@ -82,7 +82,7 @@ int AllocatorMemkindHBWMalloc::posix_memalign (void **ptr, size_t align, size_t
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr %p, base %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), baseptr, name(), this);
-		_stats.record_aligned_malloc (size + align);
+		record_aligned_malloc (size + align);
 
 		*ptr = res;
 		return 0;
@@ -98,7 +98,7 @@ void AllocatorMemkindHBWMalloc::free (void *ptr)
 	// When freeing the memory, need to free the base pointe
 	VERBOSE_MSG(3, ALLOCATOR_NAME": Freeing up pointer %p (hdr %p) w/ size - %lu (base pointer located in %p)\n", ptr, hdr, hdr->size, hdr->base_ptr);
 
-	_stats.record_free (hdr->size);
+	record_free (hdr->size);
 	hbw_free (hdr->base_ptr);
 }
 
@@ -129,7 +129,7 @@ void * AllocatorMemkindHBWMalloc::realloc (void *ptr, size_t size)
 				DBG("Reallocated (%ld->%ld [extra bytes = %lu]) from %p (base at %p, header at %p) into %p (base at %p, header at %p) w/ allocator %s (%p)\n", prev_size, size, extra_size, ptr, prev_baseptr, prev_hdr, res, new_baseptr, Allocator::getAllocatorHeader (res), name(), this);
 			}
 
-			_stats.record_realloc (size, prev_size);
+			record_realloc (size, prev_size);
 
 			return res;
 		}
@@ -142,7 +142,7 @@ void * AllocatorMemkindHBWMalloc::realloc (void *ptr, size_t size)
 	else
 	{
 		VERBOSE_MSG(3, ALLOCATOR_NAME": realloc (NULL, ...) forwarded to malloc\n");
-		_stats.record_realloc_forward_malloc();
+		record_realloc_forward_malloc();
 
 		return this->malloc (size);
 	}
@@ -218,12 +218,3 @@ const char * AllocatorMemkindHBWMalloc::description (void) const
 	return "Allocator based on hbwmalloc on top of memkind";
 }
 
-void AllocatorMemkindHBWMalloc::show_statistics (void) const
-{
-	_stats.show_statistics (ALLOCATOR_NAME, true);
-}
-
-bool AllocatorMemkindHBWMalloc::fits (size_t s) const
-{
-	return _stats.water_mark() + s <= this->size();
-}

src/allocator-memkind-hbwmalloc.hxx

@@ -4,13 +4,10 @@
 
 #pragma once
 
-#include "allocator.hxx"
+#include "statistics-recorder-allocator.hxx"
 
-class AllocatorMemkindHBWMalloc final : public Allocator
+class AllocatorMemkindHBWMalloc final : public StatisticsRecorderAllocator
 {
-	private:
-	AllocatorStatistics _stats;
-
 	public:
 	AllocatorMemkindHBWMalloc (allocation_functions_t &);
 	~AllocatorMemkindHBWMalloc();
@@ -25,30 +22,4 @@ class AllocatorMemkindHBWMalloc final : public Allocator
 	void   configure (const char *);
 	const char * name (void) const;
 	const char * description (void) const;
-	void show_statistics (void) const;
-
-	void *memcpy (void *dest, const void *src, size_t n)
-	  { return ::memcpy (dest, src, n); }
-
-	bool fits (size_t s) const;
-	size_t hwm (void) const
-	  { return _stats.water_mark(); }
-	void record_unfitted_malloc (size_t s)
-	  { _stats.record_unfitted_malloc (s); } ;
-	void record_unfitted_calloc (size_t s)
-	  { _stats.record_unfitted_calloc (s); } ;
-	void record_unfitted_aligned_malloc (size_t s)
-	  { _stats.record_unfitted_aligned_malloc (s); } ;
-	void record_unfitted_realloc (size_t s)
-	  { _stats.record_unfitted_realloc (s); } ;
-
-	void record_source_realloc (size_t s)
-	  { _stats.record_source_realloc (s); };
-	void record_target_realloc (size_t s)
-	  { _stats.record_target_realloc (s); };
-	void record_self_realloc (size_t s)
-	  { _stats.record_self_realloc (s); };
-
-	void record_realloc_forward_malloc (void)
-	  { _stats.record_realloc_forward_malloc (); }
 };

src/allocator-memkind-pmem.cxx

@@ -47,18 +47,19 @@ AllocatorMemkindPMEM::AllocatorMemkindPMEM (allocation_functions_t &af)
 
 AllocatorMemkindPMEM::~AllocatorMemkindPMEM ()
 {
+	// Explicitely call the deallocator as the new was done on pre-allocated memory
+	for (int i = 0; _num_NUMA_nodes > i; ++i)
+		_stats[i].~AllocatorStatistics();
 	_af.free (_stats);
 	_af.free (_kind);
 	_af.free (_cpu_2_NUMA);
 }
 
 void * AllocatorMemkindPMEM::malloc (size_t size)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
-	assert (0 <= n && n < _num_NUMA_nodes);
+	const short n = get_local_numa ();
 
-	DBG("Running on CPU %d - NUMA node %ld\n", cpu, n);
+	DBG("Running on CPU %d - NUMA node %hd\n", get_current_cpu (), n);
 
 	// Forward memory request to real malloc and reserve some space for the header
 	void * baseptr = memkind_malloc (_kind[n], Allocator::getTotalSize (size));
@@ -73,19 +74,17 @@ void * AllocatorMemkindPMEM::malloc (size_t size)
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr & base at %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), name(), this);
-		_stats[n].record_malloc (size);
+		record_malloc (size, n);
 	}
 
 	return res;
 }
 
 void * AllocatorMemkindPMEM::calloc (size_t nmemb, size_t size)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
-	assert (0 <= n && n < _num_NUMA_nodes);
+	const short n = get_local_numa ();
 
-	DBG("Running on CPU %d - NUMA node %ld\n", cpu, n);
+	DBG("Running on CPU %d - NUMA node %hd\n", get_current_cpu (), n);
 
 	// Forward memory request to real malloc and request additional space to store
 	// the allocator and the basepointer
@@ -101,7 +100,7 @@ void * AllocatorMemkindPMEM::calloc (size_t nmemb, size_t size)
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr & base %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), name(), this);
-		_stats[n].record_calloc (nmemb * size);
+		record_calloc (nmemb * size, n);
 	}
 
 	return res;
@@ -111,11 +110,9 @@ int AllocatorMemkindPMEM::posix_memalign (void **ptr, size_t align, size_t size)
 {
 	assert (ptr != nullptr);
 
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
-	assert (0 <= n && n < _num_NUMA_nodes);
+	const short n = get_local_numa ();
 
-	DBG("Running on CPU %d - NUMA node %ld\n", cpu, n);
+	DBG("Running on CPU %d - NUMA node %hd\n", get_current_cpu (), n);
 
 	// Forward memory request to real malloc and request additional space to
 	// store the allocator and the basepointer
@@ -131,7 +128,7 @@ int AllocatorMemkindPMEM::posix_memalign (void **ptr, size_t align, size_t size)
 
 		// Verbosity and emit statistics
 		VERBOSE_MSG(3, ALLOCATOR_NAME": Allocated %lu bytes in %p (hdr %p, base %p) w/ allocator %s (%p)\n", size, res, Allocator::getAllocatorHeader (res), baseptr, name(), this);
-		_stats[n].record_aligned_malloc (size + align);
+		record_aligned_malloc (size + align, n);
 
 		*ptr = res;
 		return 0;
@@ -153,7 +150,7 @@ void AllocatorMemkindPMEM::free (void *ptr)
 	assert (gotnode);
 	assert (0 <= n && n < _num_NUMA_nodes);
 
-	_stats[n].record_free (hdr->size);
+	record_free (hdr->size, n);
 	memkind_free (_kind[n], hdr->base_ptr);
 }
 
@@ -190,7 +187,7 @@ void * AllocatorMemkindPMEM::realloc (void *ptr, size_t size)
 				DBG("Reallocated (%ld->%ld [extra bytes = %lu]) from %p (base at %p, header at %p) into %p (base at %p, header at %p) w/ allocator %s (%p) on node %d\n", prev_size, size, extra_size, ptr, prev_baseptr, prev_hdr, res, new_baseptr, Allocator::getAllocatorHeader (res), name(), this, n);
 			}
 
-			_stats[n].record_realloc (size, prev_size);
+			record_realloc (size, prev_size, n);
 
 			return res;
 		}
@@ -203,9 +200,7 @@ void * AllocatorMemkindPMEM::realloc (void *ptr, size_t size)
 	else
 	{
 		VERBOSE_MSG(3, ALLOCATOR_NAME": realloc (NULL, ...) forwarded to malloc\n");
-		int cpu = sched_getcpu();
-		long n = _cpu_2_NUMA[cpu];
-		_stats[n].record_realloc_forward_malloc();
+		record_realloc_forward_malloc();
 
 		return this->malloc (size);
 	}
@@ -341,71 +336,92 @@ void AllocatorMemkindPMEM::show_statistics (void) const
 	}
 }
 
-bool AllocatorMemkindPMEM::fits (size_t) const
+int AllocatorMemkindPMEM::get_current_cpu (void) const
+{
+	return sched_getcpu ();
+}
+
+short AllocatorMemkindPMEM::get_local_numa (void) const
+{
+	int cpu = get_current_cpu ();
+	short n = _cpu_2_NUMA[cpu];
+	assert (0 <= n && n < _num_NUMA_nodes);
+	return n;
+}
+
+bool AllocatorMemkindPMEM::fits (size_t,short) const
 {
 	return true;
 }
 
-size_t AllocatorMemkindPMEM::hwm (void) const
+size_t AllocatorMemkindPMEM::hwm (short n) const
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	return _stats[n].water_mark ();
 }
 
-void AllocatorMemkindPMEM::record_unfitted_malloc (size_t s)
+void AllocatorMemkindPMEM::record_malloc (size_t s, short n)
+{
+	_stats[n].record_malloc (s);
+}
+
+void AllocatorMemkindPMEM::record_calloc (size_t s, short n)
+{
+	_stats[n].record_calloc (s);
+}
+
+void AllocatorMemkindPMEM::record_aligned_malloc (size_t s, short n)
+{
+	_stats[n].record_aligned_malloc (s);
+}
+
+void AllocatorMemkindPMEM::record_realloc (size_t size, size_t prev_size, short n)
+{
+	_stats[n].record_realloc (size, prev_size);
+}
+
+void AllocatorMemkindPMEM::record_free (size_t s, short n)
+{
+	_stats[n].record_free (s);
+}
+
+
+void AllocatorMemkindPMEM::record_unfitted_malloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_unfitted_malloc (s);
 }
 
-void AllocatorMemkindPMEM::record_unfitted_calloc (size_t s)
+void AllocatorMemkindPMEM::record_unfitted_calloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_unfitted_calloc (s);
 }
 
-void AllocatorMemkindPMEM::record_unfitted_aligned_malloc (size_t s)
+void AllocatorMemkindPMEM::record_unfitted_aligned_malloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_unfitted_aligned_malloc (s);
 }
 
-void AllocatorMemkindPMEM::record_unfitted_realloc (size_t s)
+void AllocatorMemkindPMEM::record_unfitted_realloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_unfitted_realloc (s);
 }
 
-void AllocatorMemkindPMEM::record_source_realloc (size_t s)
+void AllocatorMemkindPMEM::record_source_realloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_source_realloc (s);
 }
 
-void AllocatorMemkindPMEM::record_target_realloc (size_t s)
+void AllocatorMemkindPMEM::record_target_realloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_target_realloc (s);
 }
 
-void AllocatorMemkindPMEM::record_self_realloc (size_t s)
+void AllocatorMemkindPMEM::record_self_realloc (size_t s, short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_self_realloc (s);
 }
 
-void AllocatorMemkindPMEM::record_realloc_forward_malloc (void)
+void AllocatorMemkindPMEM::record_realloc_forward_malloc (short n)
 {
-	int cpu = sched_getcpu();
-	long n = _cpu_2_NUMA[cpu];
 	_stats[n].record_realloc_forward_malloc ();
 }