erts: Fix locking order violation for allocation wrappers

Some query functions in erl_alloc_util.c lock the allocator mutex
and then use erts_printf that in turn may call the sys allocator
through the wrappers. To avoid breaking locking order these
query functions first "pre-locks" all allocator wrappers.

5 files changed, 221 insertions, 57 deletions

diff --git a/erts/emulator/beam/erl_alloc.c b/erts/emulator/beam/erl_alloc.c
index f0cc52e30f..c19a2647e4 100644
--- a/erts/emulator/beam/erl_alloc.c
+++ b/erts/emulator/beam/erl_alloc.c
@@ -570,6 +570,8 @@ erts_alloc_init(int *argc, char **argv, ErtsAllocInitOpts *eaiop)
     if (ncpu < 1)
 	ncpu = 1;
 
+    erts_tsd_key_create(&erts_allctr_prelock_tsd_key);
+
     erts_sys_alloc_init();
     erts_init_utils_mem();
 
@@ -3139,6 +3141,55 @@ erts_request_alloc_info(struct process *c_p,
     return 1;
 }
 
+/* 
+ * The allocator wrapper prelocking stuff below is about the locking order.
+ * It only affects wrappers (erl_mtrace.c and erl_instrument.c) that keep locks
+ * during alloc/realloc/free.
+ *
+ * Some query functions in erl_alloc_util.c lock the allocator mutex and then
+ * use erts_printf that in turn may call the sys allocator through the wrappers.
+ * To avoid breaking locking order these query functions first "pre-locks" all
+ * allocator wrappers.
+ */
+ErtsAllocatorWrapper_t *erts_allctr_wrappers;
+int erts_allctr_wrapper_prelocked = 0;
+erts_tsd_key_t erts_allctr_prelock_tsd_key;
+
+void erts_allctr_wrapper_prelock_init(ErtsAllocatorWrapper_t* wrapper)
+{
+    ASSERT(wrapper->lock && wrapper->unlock);
+    wrapper->next = erts_allctr_wrappers;
+    erts_allctr_wrappers = wrapper;
+}
+
+void erts_allctr_wrapper_pre_lock(void)
+{
+    if (erts_allctr_wrappers) {
+	ErtsAllocatorWrapper_t* wrapper = erts_allctr_wrappers;
+	for ( ; wrapper; wrapper = wrapper->next) {
+	    wrapper->lock();
+	}
+	ASSERT(!erts_allctr_wrapper_prelocked);
+	erts_allctr_wrapper_prelocked = 1;
+	erts_tsd_set(erts_allctr_prelock_tsd_key, (void*)1);
+    }
+}
+
+void erts_allctr_wrapper_pre_unlock(void)
+{
+    if (erts_allctr_wrappers) {
+	ErtsAllocatorWrapper_t* wrapper = erts_allctr_wrappers;
+	
+	erts_allctr_wrapper_prelocked = 0;
+	erts_tsd_set(erts_allctr_prelock_tsd_key, (void*)0);
+	for ( ; wrapper; wrapper = wrapper->next) {
+	    wrapper->unlock();
+	}
+    }
+}
+
+
+
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
  * Deprecated functions                                                    *
  *                                                                         *
diff --git a/erts/emulator/beam/erl_alloc.h b/erts/emulator/beam/erl_alloc.h
index ba5ec9c367..9cafd8ddc8 100644
--- a/erts/emulator/beam/erl_alloc.h
+++ b/erts/emulator/beam/erl_alloc.h
@@ -135,6 +135,18 @@ typedef struct {
 
 extern ErtsAllocatorThrSpec_t erts_allctr_thr_spec[ERTS_ALC_A_MAX+1];
 
+typedef struct ErtsAllocatorWrapper_t_ {
+    void (*lock)(void);
+    void (*unlock)(void);
+    struct ErtsAllocatorWrapper_t_* next;
+}ErtsAllocatorWrapper_t;
+ErtsAllocatorWrapper_t *erts_allctr_wrappers;
+extern int erts_allctr_wrapper_prelocked;
+extern erts_tsd_key_t erts_allctr_prelock_tsd_key;
+void erts_allctr_wrapper_prelock_init(ErtsAllocatorWrapper_t* wrapper);
+void erts_allctr_wrapper_pre_lock(void);
+void erts_allctr_wrapper_pre_unlock(void);
+
 void erts_alloc_register_scheduler(void *vesdp);
 #ifdef ERTS_SMP
 void erts_alloc_scheduler_handle_delayed_dealloc(void *vesdp,
@@ -188,6 +200,7 @@ void *erts_realloc(ErtsAlcType_t type, void *ptr, Uint size);
 void erts_free(ErtsAlcType_t type, void *ptr);
 void *erts_alloc_fnf(ErtsAlcType_t type, Uint size);
 void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size);
+int erts_is_allctr_wrapper_prelocked(void);
 
 #endif /* #if !ERTS_ALC_DO_INLINE */
 
@@ -258,6 +271,13 @@ void *erts_realloc_fnf(ErtsAlcType_t type, void *ptr, Uint size)
 	size);
 }
 
+ERTS_ALC_INLINE
+int erts_is_allctr_wrapper_prelocked(void)
+{
+    return erts_allctr_wrapper_prelocked                 /* locked */
+	&& !!erts_tsd_get(erts_allctr_prelock_tsd_key);  /* by me  */
+}
+
 #endif /* #if ERTS_ALC_DO_INLINE || defined(ERTS_ALC_INTERNAL__) */
 
 #define ERTS_ALC_GET_THR_IX() ((int) erts_get_scheduler_id())
diff --git a/erts/emulator/beam/erl_alloc_util.c b/erts/emulator/beam/erl_alloc_util.c
index 583aeed12d..cbeb2c1e3a 100644
--- a/erts/emulator/beam/erl_alloc_util.c
+++ b/erts/emulator/beam/erl_alloc_util.c
@@ -588,17 +588,21 @@ do {									\
 
 #ifdef DEBUG
 #ifdef USE_THREADS
+# ifdef ERTS_SMP
+#  define IS_ACTUALLY_BLOCKING (erts_thr_progress_is_blocking())
+# else
+#  define IS_ACTUALLY_BLOCKING 0
+# endif
 #define ERTS_ALCU_DBG_CHK_THR_ACCESS(A)					\
 do {									\
-    if (!(A)->thread_safe) {						\
-	if (!(A)->debug.saved_tid) {					\
+    if (!(A)->thread_safe && !IS_ACTUALLY_BLOCKING) {                   \
+	if (!(A)->debug.saved_tid) {                                    \
 	    (A)->debug.tid = erts_thr_self();				\
 	    (A)->debug.saved_tid = 1;					\
 	}								\
 	else {								\
 	    ERTS_SMP_LC_ASSERT(						\
-		ethr_equal_tids((A)->debug.tid, erts_thr_self())	\
-		|| erts_thr_progress_is_blocking());			\
+		ethr_equal_tids((A)->debug.tid, erts_thr_self()));	\
 	}								\
     }									\
 } while (0)
@@ -2510,12 +2514,6 @@ static erts_mtx_t init_atoms_mtx;
 static void
 init_atoms(Allctr_t *allctr)
 {
-
-#ifdef USE_THREADS
-    if (allctr && allctr->thread_safe)
-	erts_mtx_unlock(&allctr->mutex);
-#endif
-
     erts_mtx_lock(&init_atoms_mtx);
 
     if (!atoms_initialized) {
@@ -2606,18 +2604,13 @@ init_atoms(Allctr_t *allctr)
 	    fix_type_atoms[ix] = am_atom_put(name, len);
 	}
     }
-
     
-    if (allctr) {
+    if (allctr && !allctr->atoms_initialized) {
 
 	make_name_atoms(allctr);
 
 	(*allctr->init_atoms)();
 
-#ifdef USE_THREADS
-	if (allctr->thread_safe)
-	    erts_mtx_lock(&allctr->mutex);
-#endif
     	allctr->atoms_initialized = 1;
     }
 
@@ -3245,17 +3238,21 @@ erts_alcu_info_options(Allctr_t *allctr,
 {
     Eterm res;
 
+    if (hpp || szp)
+	ensure_atoms_initialized(allctr);
 
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) {
+	erts_allctr_wrapper_pre_lock();
 	erts_mtx_lock(&allctr->mutex);
+    }
 #endif
-    if (hpp || szp)
-	ensure_atoms_initialized(allctr);
     res = info_options(allctr, print_to_p, print_to_arg, hpp, szp);
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) { 
 	erts_mtx_unlock(&allctr->mutex);
+	erts_allctr_wrapper_pre_unlock();
+    }
 #endif
     return res;
 }
@@ -3282,16 +3279,18 @@ erts_alcu_sz_info(Allctr_t *allctr,
 	return am_false;
     }
 
+    if (hpp || szp)
+	ensure_atoms_initialized(allctr);
+
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) {
+	erts_allctr_wrapper_pre_lock();
 	erts_mtx_lock(&allctr->mutex);
+    }
 #endif
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    if (hpp || szp)
-	ensure_atoms_initialized(allctr);
-
     /* Update sbc values not continously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
@@ -3327,13 +3326,16 @@ erts_alcu_sz_info(Allctr_t *allctr,
 
 
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) {
 	erts_mtx_unlock(&allctr->mutex);
+	erts_allctr_wrapper_pre_unlock();
+    }
 #endif
 
     return res;
 }
 
+
 Eterm
 erts_alcu_info(Allctr_t *allctr,
 	       int begin_max_period,
@@ -3354,16 +3356,18 @@ erts_alcu_info(Allctr_t *allctr,
 	return am_false;
     }
 
+    if (hpp || szp)
+	ensure_atoms_initialized(allctr);
+
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) {
+	erts_allctr_wrapper_pre_lock();
 	erts_mtx_lock(&allctr->mutex);
+    }
 #endif
 
     ERTS_ALCU_DBG_CHK_THR_ACCESS(allctr);
 
-    if (hpp || szp)
-	ensure_atoms_initialized(allctr);
-
     /* Update sbc values not continously updated */
     allctr->sbcs.blocks.curr.no
 	= allctr->sbcs.curr.norm.mseg.no + allctr->sbcs.curr.norm.sys_alloc.no;
@@ -3416,8 +3420,10 @@ erts_alcu_info(Allctr_t *allctr,
 
 
 #ifdef USE_THREADS
-    if (allctr->thread_safe)
+    if (allctr->thread_safe) {
 	erts_mtx_unlock(&allctr->mutex);
+	erts_allctr_wrapper_pre_unlock();
+    }
 #endif
 
     return res;
diff --git a/erts/emulator/beam/erl_instrument.c b/erts/emulator/beam/erl_instrument.c
index b5b245288b..1aea3f65bc 100644
--- a/erts/emulator/beam/erl_instrument.c
+++ b/erts/emulator/beam/erl_instrument.c
@@ -271,6 +271,18 @@ stat_upd_realloc(ErtsAlcType_t n, Uint size, Uint old_size)
  * stat instrumentation callback functions
  */
 
+static void stat_pre_lock(void)
+{
+    erts_mtx_lock(&instr_mutex);
+}
+
+static void stat_pre_unlock(void)
+{
+    erts_mtx_unlock(&instr_mutex);
+}
+
+static ErtsAllocatorWrapper_t instr_wrapper;
+
 static void *
 stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
 {
@@ -278,7 +290,9 @@ stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
     Uint ssize;
     void *res;
 
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_mutex);
+    }
 
     ssize = size + STAT_BLOCK_HEADER_SIZE;
     res = (*real_af->alloc)(n, real_af->extra, ssize);
@@ -293,7 +307,9 @@ stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
 	res = (void *) ((StatBlock_t *) res)->mem;
     }
 
-    erts_mtx_unlock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+    }
 
     return res;
 }
@@ -307,7 +323,9 @@ stat_realloc(ErtsAlcType_t n, void *extra, void *ptr, Uint size)
     void *sptr;
     void *res;
 
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_mutex);
+    }
 
     if (ptr) {
 	sptr = (void *) (((char *) ptr) - STAT_BLOCK_HEADER_SIZE);
@@ -329,7 +347,9 @@ stat_realloc(ErtsAlcType_t n, void *extra, void *ptr, Uint size)
 	res = (void *) ((StatBlock_t *) res)->mem;
     }
 
-    erts_mtx_unlock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+    }
 
     return res;
 }
@@ -340,7 +360,9 @@ stat_free(ErtsAlcType_t n, void *extra, void *ptr)
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
     void *sptr;
 
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_mutex);
+    }
 
     if (ptr) {
 	sptr = (void *) (((char *) ptr) - STAT_BLOCK_HEADER_SIZE);
@@ -352,7 +374,9 @@ stat_free(ErtsAlcType_t n, void *extra, void *ptr)
 
     (*real_af->free)(n, real_af->extra, sptr);
 
-    erts_mtx_unlock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+    }
 
 }
 
@@ -360,6 +384,18 @@ stat_free(ErtsAlcType_t n, void *extra, void *ptr)
  * map stat instrumentation callback functions
  */
 
+static void map_stat_pre_lock(void)
+{
+    erts_mtx_lock(&instr_x_mutex);
+    erts_mtx_lock(&instr_mutex);
+}
+
+static void map_stat_pre_unlock(void)
+{
+    erts_mtx_unlock(&instr_mutex);
+    erts_mtx_unlock(&instr_x_mutex);
+}
+
 static void *
 map_stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
 {
@@ -367,7 +403,9 @@ map_stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
     Uint msize;
     void *res;
 
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_mutex);
+    }
 
     msize = size + MAP_STAT_BLOCK_HEADER_SIZE;
     res = (*real_af->alloc)(n, real_af->extra, msize);
@@ -388,7 +426,9 @@ map_stat_alloc(ErtsAlcType_t n, void *extra, Uint size)
 	res = (void *) mb->mem;
     }
 
-    erts_mtx_unlock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+    }
 
     return res;
 }
@@ -402,8 +442,10 @@ map_stat_realloc(ErtsAlcType_t n, void *extra, void *ptr, Uint size)
     void *mptr;
     void *res;
 
-    erts_mtx_lock(&instr_x_mutex);
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_x_mutex);
+	erts_mtx_lock(&instr_mutex);
+    }
 
     if (ptr) {
 	mptr = (void *) (((char *) ptr) - MAP_STAT_BLOCK_HEADER_SIZE);
@@ -449,9 +491,10 @@ map_stat_realloc(ErtsAlcType_t n, void *extra, void *ptr, Uint size)
 
 	res = (void *) mb->mem;
     }
-
-    erts_mtx_unlock(&instr_mutex);
-    erts_mtx_unlock(&instr_x_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+	erts_mtx_unlock(&instr_x_mutex);
+    }
 
     return res;
 }
@@ -462,8 +505,10 @@ map_stat_free(ErtsAlcType_t n, void *extra, void *ptr)
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
     void *mptr;
 
-    erts_mtx_lock(&instr_x_mutex);
-    erts_mtx_lock(&instr_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&instr_x_mutex);
+	erts_mtx_lock(&instr_mutex);
+    }
 
     if (ptr) {
 	MapStatBlock_t *mb;
@@ -486,8 +531,10 @@ map_stat_free(ErtsAlcType_t n, void *extra, void *ptr)
 
     (*real_af->free)(n, real_af->extra, mptr);
 
-    erts_mtx_unlock(&instr_mutex);
-    erts_mtx_unlock(&instr_x_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&instr_mutex);
+	erts_mtx_unlock(&instr_x_mutex);
+    }
 
 }
 
@@ -496,8 +543,10 @@ static void dump_memory_map_to_stream(FILE *fp)
     ErtsAlcType_t n;
     MapStatBlock_t *bp;
     int lock = !ERTS_IS_CRASH_DUMPING;
-    if (lock)
+    if (lock) {
+	ASSERT(!erts_is_allctr_wrapper_prelocked());
 	erts_mtx_lock(&instr_mutex);
+    }
 
     /* Write header */
 
@@ -1155,6 +1204,7 @@ erts_instr_get_type_info(Process *proc)
 Uint
 erts_instr_init(int stat, int map_stat)
 {
+    Uint extra_sz;
     int i;
 
     am_tot = NULL;
@@ -1208,7 +1258,9 @@ erts_instr_init(int stat, int map_stat)
 	    erts_allctrs[i].free	= map_stat_free;
 	    erts_allctrs[i].extra	= (void *) &real_allctrs[i];
 	}
-	return MAP_STAT_BLOCK_HEADER_SIZE;
+	instr_wrapper.lock = map_stat_pre_lock;
+	instr_wrapper.unlock = map_stat_pre_unlock;
+	extra_sz = MAP_STAT_BLOCK_HEADER_SIZE;
     }
     else {
 	erts_instr_stat = 1;
@@ -1220,8 +1272,11 @@ erts_instr_init(int stat, int map_stat)
 	    erts_allctrs[i].free	= stat_free;
 	    erts_allctrs[i].extra	= (void *) &real_allctrs[i];
 	}
-	return STAT_BLOCK_HEADER_SIZE;
+	instr_wrapper.lock = stat_pre_lock;
+	instr_wrapper.unlock = stat_pre_unlock;
+	extra_sz = STAT_BLOCK_HEADER_SIZE;
     }
-
+    erts_allctr_wrapper_prelock_init(&instr_wrapper);
+    return extra_sz;
 }
 
diff --git a/erts/emulator/beam/erl_mtrace.c b/erts/emulator/beam/erl_mtrace.c
index e538ba30c2..c8bb126687 100644
--- a/erts/emulator/beam/erl_mtrace.c
+++ b/erts/emulator/beam/erl_mtrace.c
@@ -222,6 +222,8 @@ static byte *tracep;
 static byte *endp;
 static SysTimeval last_tv;
 
+static ErtsAllocatorWrapper_t mtrace_wrapper;
+
 #if ERTS_MTRACE_SEGMENT_ID >= ERTS_ALC_A_MIN || ERTS_MTRACE_SEGMENT_ID < 0
 #error ERTS_MTRACE_SEGMENT_ID >= ERTS_ALC_A_MIN || ERTS_MTRACE_SEGMENT_ID < 0
 #endif
@@ -555,6 +557,8 @@ write_trace_header(char *nodename, char *pid, char *hostname)
     return 1;
 }
 
+static void mtrace_pre_lock(void);
+static void mtrace_pre_unlock(void);
 static void *mtrace_alloc(ErtsAlcType_t, void *, Uint);
 static void *mtrace_realloc(ErtsAlcType_t, void *, void *, Uint);
 static void mtrace_free(ErtsAlcType_t, void *, void *);
@@ -635,12 +639,16 @@ erts_mtrace_install_wrapper_functions(void)
 	    erts_allctrs[i].free	= mtrace_free;
 	    erts_allctrs[i].extra	= (void *) &real_allctrs[i];
 	}
+	mtrace_wrapper.lock = mtrace_pre_lock;
+	mtrace_wrapper.unlock = mtrace_pre_unlock;
+	erts_allctr_wrapper_prelock_init(&mtrace_wrapper);
     }
 }
 
 void
 erts_mtrace_stop(void)
 {
+    ASSERT(!erts_is_allctr_wrapper_prelocked());
     erts_mtx_lock(&mtrace_op_mutex);
     erts_mtx_lock(&mtrace_buf_mutex);
     if (erts_mtrace_enabled) {
@@ -677,6 +685,7 @@ erts_mtrace_stop(void)
 void
 erts_mtrace_exit(Uint32 exit_value)
 {
+    ASSERT(!erts_is_allctr_wrapper_prelocked());
     erts_mtx_lock(&mtrace_op_mutex);
     erts_mtx_lock(&mtrace_buf_mutex);
     if (erts_mtrace_enabled) {
@@ -935,18 +944,33 @@ write_free_entry(byte tag,
     erts_mtx_unlock(&mtrace_buf_mutex);
 }
 
+static void mtrace_pre_lock(void)
+{
+    erts_mtx_lock(&mtrace_op_mutex);
+}
+
+static void mtrace_pre_unlock(void)
+{
+    erts_mtx_unlock(&mtrace_op_mutex);
+}
+
+
 static void *
 mtrace_alloc(ErtsAlcType_t n, void *extra, Uint size)
 {
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
     void *res;
 
-    erts_mtx_lock(&mtrace_op_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&mtrace_op_mutex);
+    }
 
     res = (*real_af->alloc)(n, real_af->extra, size);
     write_alloc_entry(ERTS_MT_ALLOC_BDY_TAG, res, n, 0, size);
 
-    erts_mtx_unlock(&mtrace_op_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&mtrace_op_mutex);
+    }
 
     return res;
 }
@@ -957,12 +981,16 @@ mtrace_realloc(ErtsAlcType_t n, void *extra, void *ptr, Uint size)
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
     void *res;
 
-    erts_mtx_lock(&mtrace_op_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&mtrace_op_mutex);
+    }
 
     res = (*real_af->realloc)(n, real_af->extra, ptr, size);
     write_realloc_entry(ERTS_MT_REALLOC_BDY_TAG, res, n, 0, ptr, size);
 
-    erts_mtx_unlock(&mtrace_op_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_unlock(&mtrace_op_mutex);
+    }
 
     return res;
 
@@ -973,10 +1001,14 @@ mtrace_free(ErtsAlcType_t n, void *extra, void *ptr)
 {
     ErtsAllocatorFunctions_t *real_af = (ErtsAllocatorFunctions_t *) extra;
 
-    erts_mtx_lock(&mtrace_op_mutex);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	erts_mtx_lock(&mtrace_op_mutex);
+    }
 
     (*real_af->free)(n, real_af->extra, ptr);
-    write_free_entry(ERTS_MT_FREE_BDY_TAG, n, 0, ptr);
+    if (!erts_is_allctr_wrapper_prelocked()) {
+	write_free_entry(ERTS_MT_FREE_BDY_TAG, n, 0, ptr);
+    }
 
     erts_mtx_unlock(&mtrace_op_mutex);
 }