diff --git a/include/jemalloc/internal/mutex_inlines.h b/include/jemalloc/internal/mutex_inlines.h
index c0c3cfe9..cf0ce23a 100644
--- a/include/jemalloc/internal/mutex_inlines.h
+++ b/include/jemalloc/internal/mutex_inlines.h
@@ -1,31 +1,42 @@
 #ifndef JEMALLOC_INTERNAL_MUTEX_INLINES_H
 #define JEMALLOC_INTERNAL_MUTEX_INLINES_H
 
+void	malloc_mutex_lock_slow(malloc_mutex_t *mutex);
+
 #ifndef JEMALLOC_ENABLE_INLINE
 void	malloc_mutex_lock(tsdn_t *tsdn, malloc_mutex_t *mutex);
+bool	malloc_mutex_trylock(malloc_mutex_t *mutex);
 void	malloc_mutex_unlock(tsdn_t *tsdn, malloc_mutex_t *mutex);
 void	malloc_mutex_assert_owner(tsdn_t *tsdn, malloc_mutex_t *mutex);
 void	malloc_mutex_assert_not_owner(tsdn_t *tsdn, malloc_mutex_t *mutex);
 #endif
 
 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_MUTEX_C_))
+JEMALLOC_INLINE void
+malloc_mutex_lock_final(malloc_mutex_t *mutex) {
+	MALLOC_MUTEX_LOCK(mutex);
+}
+
+/* Trylock: return false if the lock is successfully acquired. */
+JEMALLOC_INLINE bool
+malloc_mutex_trylock(malloc_mutex_t *mutex) {
+	return MALLOC_MUTEX_TRYLOCK(mutex);
+}
+
 JEMALLOC_INLINE void
 malloc_mutex_lock(tsdn_t *tsdn, malloc_mutex_t *mutex) {
 	witness_assert_not_owner(tsdn, &mutex->witness);
 	if (isthreaded) {
-#ifdef _WIN32
-#  if _WIN32_WINNT >= 0x0600
-		AcquireSRWLockExclusive(&mutex->lock);
-#  else
-		EnterCriticalSection(&mutex->lock);
-#  endif
-#elif (defined(JEMALLOC_OS_UNFAIR_LOCK))
-		os_unfair_lock_lock(&mutex->lock);
-#elif (defined(JEMALLOC_OSSPIN))
-		OSSpinLockLock(&mutex->lock);
-#else
-		pthread_mutex_lock(&mutex->lock);
-#endif
+		if (malloc_mutex_trylock(mutex)) {
+			malloc_mutex_lock_slow(mutex);
+		}
+		/* We own the lock now.  Update a few counters. */
+		lock_prof_data_t *data = &mutex->prof_data;
+		data->n_lock_ops++;
+		if (data->prev_owner != tsdn) {
+			data->prev_owner = tsdn;
+			data->n_owner_switches++;
+		}
 	}
 	witness_lock(tsdn, &mutex->witness);
 }
@@ -34,19 +45,7 @@ JEMALLOC_INLINE void
 malloc_mutex_unlock(tsdn_t *tsdn, malloc_mutex_t *mutex) {
 	witness_unlock(tsdn, &mutex->witness);
 	if (isthreaded) {
-#ifdef _WIN32
-#  if _WIN32_WINNT >= 0x0600
-		ReleaseSRWLockExclusive(&mutex->lock);
-#  else
-		LeaveCriticalSection(&mutex->lock);
-#  endif
-#elif (defined(JEMALLOC_OS_UNFAIR_LOCK))
-		os_unfair_lock_unlock(&mutex->lock);
-#elif (defined(JEMALLOC_OSSPIN))
-		OSSpinLockUnlock(&mutex->lock);
-#else
-		pthread_mutex_unlock(&mutex->lock);
-#endif
+		MALLOC_MUTEX_UNLOCK(mutex);
 	}
 }
 
diff --git a/include/jemalloc/internal/mutex_structs.h b/include/jemalloc/internal/mutex_structs.h
index c34c1d47..7065c997 100644
--- a/include/jemalloc/internal/mutex_structs.h
+++ b/include/jemalloc/internal/mutex_structs.h
@@ -1,9 +1,50 @@
 #ifndef JEMALLOC_INTERNAL_MUTEX_STRUCTS_H
 #define JEMALLOC_INTERNAL_MUTEX_STRUCTS_H
 
+struct lock_prof_data_s {
+	/*
+	 * Counters touched on the slow path, i.e. when there is lock
+	 * contention.  We update them once we have the lock.
+	 */
+	/* Total time spent waiting on this lock. */
+	nstime_t		tot_wait_time;
+	/* Max time spent on a single lock operation. */
+	nstime_t		max_wait_time;
+	/* # of times have to wait for this lock (after spinning). */
+	uint64_t		n_wait_times;
+	/* # of times acquired the lock through local spinning. */
+	uint64_t		n_spin_acquired;
+	/* Max # of threads waiting for the lock at the same time. */
+	uint32_t		max_n_thds;
+	/* Current # of threads waiting on the lock.  Atomic synced. */
+	uint32_t		n_waiting_thds;
+
+	/*
+	 * Data touched on the fast path.  These are modified right after we
+	 * grab the lock, so it's placed closest to the end (i.e. right before
+	 * the lock) so that we have a higher chance of them being on the same
+	 * cacheline.
+	 */
+	/* # of times the new lock holder is different from the previous one. */
+	uint64_t		n_owner_switches;
+	/* Previous lock holder, to facilitate n_owner_switches. */
+	tsdn_t			*prev_owner;
+	/* # of lock() operations in total. */
+	uint64_t		n_lock_ops;
+};
+
 struct malloc_mutex_s {
 	union {
 		struct {
+			/*
+			 * prof_data is defined first to reduce cacheline
+			 * bouncing: the data is not touched by the lock holder
+			 * during unlocking, while might be modified by
+			 * contenders.  Having it before the lock itself could
+			 * avoid prefetching a modified cacheline (for the
+			 * unlocking thread).
+			 */
+			lock_prof_data_t	prof_data;
 #ifdef _WIN32
 #  if _WIN32_WINNT >= 0x0600
 			SRWLOCK         	lock;
@@ -22,7 +63,7 @@ struct malloc_mutex_s {
 #endif
 		};
 		/*
-		 * We only touch witness when configured w/ debug. However we
+		 * We only touch witness when configured w/ debug.  However we
 		 * keep the field in a union when !debug so that we don't have
 		 * to pollute the code base with #ifdefs, while avoid paying the
 		 * memory cost.
diff --git a/include/jemalloc/internal/mutex_types.h b/include/jemalloc/internal/mutex_types.h
index b7e3a7a1..d7c7f04f 100644
--- a/include/jemalloc/internal/mutex_types.h
+++ b/include/jemalloc/internal/mutex_types.h
@@ -1,31 +1,63 @@
 #ifndef JEMALLOC_INTERNAL_MUTEX_TYPES_H
 #define JEMALLOC_INTERNAL_MUTEX_TYPES_H
 
+typedef struct lock_prof_data_s lock_prof_data_t;
 typedef struct malloc_mutex_s malloc_mutex_t;
 
+#ifdef _WIN32
+#  if _WIN32_WINNT >= 0x0600
+#    define MALLOC_MUTEX_LOCK(m)    AcquireSRWLockExclusive(&(m)->lock)
+#    define MALLOC_MUTEX_UNLOCK(m)  ReleaseSRWLockExclusive(&(m)->lock)
+#    define MALLOC_MUTEX_TRYLOCK(m) (!TryAcquireSRWLockExclusive(&(m)->lock))
+#  else
+#    define MALLOC_MUTEX_LOCK(m)    EnterCriticalSection(&(m)->lock)
+#    define MALLOC_MUTEX_UNLOCK(m)  LeaveCriticalSection(&(m)->lock)
+#    define MALLOC_MUTEX_TRYLOCK(m) (!TryEnterCriticalSection(&(m)->lock))
+#  endif
+#elif (defined(JEMALLOC_OS_UNFAIR_LOCK))
+#    define MALLOC_MUTEX_LOCK(m)    os_unfair_lock_lock(&(m)->lock)
+#    define MALLOC_MUTEX_UNLOCK(m)  os_unfair_lock_unlock(&(m)->lock)
+#    define MALLOC_MUTEX_TRYLOCK(m) (!os_unfair_lock_trylock(&(m)->lock))
+#elif (defined(JEMALLOC_OSSPIN))
+#    define MALLOC_MUTEX_LOCK(m)    OSSpinLockLock(&(m)->lock)
+#    define MALLOC_MUTEX_UNLOCK(m)  OSSpinLockUnlock(&(m)->lock)
+#    define MALLOC_MUTEX_TRYLOCK(m) (!OSSpinLockTry(&(m)->lock))
+#else
+#    define MALLOC_MUTEX_LOCK(m)    pthread_mutex_lock(&(m)->lock)
+#    define MALLOC_MUTEX_UNLOCK(m)  pthread_mutex_unlock(&(m)->lock)
+#    define MALLOC_MUTEX_TRYLOCK(m) (pthread_mutex_trylock(&(m)->lock) != 0)
+#endif
+
+#define LOCK_PROF_DATA_INITIALIZER					\
+    {NSTIME_ZERO_INITIALIZER, NSTIME_ZERO_INITIALIZER, 0, 0, 0, 0, 0, NULL, 0}
+
 #ifdef _WIN32
 #  define MALLOC_MUTEX_INITIALIZER
 #elif (defined(JEMALLOC_OS_UNFAIR_LOCK))
 #  define MALLOC_MUTEX_INITIALIZER					\
-     {{{OS_UNFAIR_LOCK_INIT}}, WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
+     {{{LOCK_PROF_DATA_INITIALIZER, OS_UNFAIR_LOCK_INIT}},		\
+      WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
 #elif (defined(JEMALLOC_OSSPIN))
 #  define MALLOC_MUTEX_INITIALIZER					\
-     {{{0}}, WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
+     {{{LOCK_PROF_DATA_INITIALIZER, 0}},				\
+      WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
 #elif (defined(JEMALLOC_MUTEX_INIT_CB))
 #  define MALLOC_MUTEX_INITIALIZER					\
-     {{{PTHREAD_MUTEX_INITIALIZER, NULL}},				\
+     {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER, NULL}},	\
       WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
 #else
+/* TODO: get rid of adaptive mutex once we do our own spin. */
 #  if (defined(JEMALLOC_HAVE_PTHREAD_MUTEX_ADAPTIVE_NP) &&		\
        defined(PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP))
 #    define MALLOC_MUTEX_TYPE PTHREAD_MUTEX_ADAPTIVE_NP
 #    define MALLOC_MUTEX_INITIALIZER					\
-       {{{PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP}},			\
+       {{{LOCK_PROF_DATA_INITIALIZER,					\
+          PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP}},			\
         WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
 #  else
 #    define MALLOC_MUTEX_TYPE PTHREAD_MUTEX_DEFAULT
 #    define MALLOC_MUTEX_INITIALIZER					\
-       {{{PTHREAD_MUTEX_INITIALIZER}},					\
+       {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER}},	\
         WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)}
 #  endif
 #endif
diff --git a/include/jemalloc/internal/nstime_types.h b/include/jemalloc/internal/nstime_types.h
index d6039e03..6e7e74cf 100644
--- a/include/jemalloc/internal/nstime_types.h
+++ b/include/jemalloc/internal/nstime_types.h
@@ -6,4 +6,6 @@ typedef struct nstime_s nstime_t;
 /* Maximum supported number of seconds (~584 years). */
 #define NSTIME_SEC_MAX	KQU(18446744072)
 
+#define NSTIME_ZERO_INITIALIZER {0}
+
 #endif /* JEMALLOC_INTERNAL_NSTIME_TYPES_H */
diff --git a/src/mutex.c b/src/mutex.c
index f1aa155e..2b80a9d9 100644
--- a/src/mutex.c
+++ b/src/mutex.c
@@ -65,6 +65,49 @@ JEMALLOC_EXPORT int	_pthread_mutex_init_calloc_cb(pthread_mutex_t *mutex,
     void *(calloc_cb)(size_t, size_t));
 #endif
 
+void
+malloc_mutex_lock_slow(malloc_mutex_t *mutex) {
+	lock_prof_data_t *data = &mutex->prof_data;
+	bool spin_success = false;
+
+	{//TODO: a smart spin policy
+		if (!malloc_mutex_trylock(mutex)) {
+			spin_success = true;
+			goto label_locked;
+		}
+	}
+
+	nstime_t now, before;
+	uint32_t n_thds;
+	nstime_init(&now, 0);
+	nstime_update(&now);
+	n_thds = atomic_add_u32(&data->n_waiting_thds, 1);
+	/* One last try as above two calls may take quite some cycles. */
+	if (!malloc_mutex_trylock(mutex)) {
+		spin_success = true;
+		atomic_sub_u32(&data->n_waiting_thds, 1);
+		goto label_locked;
+	}
+	nstime_copy(&before, &now);
+	malloc_mutex_lock_final(mutex);
+	atomic_sub_u32(&data->n_waiting_thds, 1);
+	nstime_update(&now);
+	nstime_subtract(&now, &before);
+label_locked:
+	if (spin_success) {
+		data->n_spin_acquired++;
+	} else {
+		data->n_wait_times++;
+		nstime_add(&data->tot_wait_time, &now);
+		if (nstime_compare(&now, &data->max_wait_time)) {
+			nstime_copy(&data->max_wait_time, &now);
+		}
+		if (n_thds > data->max_n_thds) {
+			data->max_n_thds = n_thds;
+		}
+	}
+}
+
 bool
 malloc_mutex_init(malloc_mutex_t *mutex, const char *name,
     witness_rank_t rank) {