diff --git a/dbg_mlc.c b/dbg_mlc.c
index 2987699f0..4da3b4cd2 100644
--- a/dbg_mlc.c
+++ b/dbg_mlc.c
@@ -558,7 +558,7 @@ GC_API void * GC_CALL GC_debug_malloc_atomic_ignore_off_page(size_t lb,
         return(0);
     }
     result = GC_store_debug_info_inner(base, (word)lb, "INTERNAL", 0);
-    ADD_CALL_CHAIN(base, GC_RETURN_ADDR);
+    ADD_CALL_CHAIN_INNER(base);
     return result;
   }
 #endif /* DBG_HDRS_ALL */
diff --git a/include/private/dbg_mlc.h b/include/private/dbg_mlc.h
index 8e08a1a9b..9ecb78f1c 100644
--- a/include/private/dbg_mlc.h
+++ b/include/private/dbg_mlc.h
@@ -129,6 +129,12 @@ typedef struct {
   GC_INNER void GC_save_callers(struct callinfo info[NFRAMES]);
   GC_INNER void GC_print_callers(struct callinfo info[NFRAMES]);
 # define ADD_CALL_CHAIN(base, ra) GC_save_callers(((oh *)(base)) -> oh_ci)
+# if defined(REDIRECT_MALLOC) && defined(THREADS) && defined(DBG_HDRS_ALL) \
+     && NARGS == 0 && NFRAMES % 2 == 0 && defined(GC_HAVE_BUILTIN_BACKTRACE)
+    GC_INNER void GC_save_callers_no_unlock(struct callinfo info[NFRAMES]);
+#   define ADD_CALL_CHAIN_INNER(base) \
+                    GC_save_callers_no_unlock(((oh *)(base)) -> oh_ci)
+# endif
 # define PRINT_CALL_CHAIN(base) GC_print_callers(((oh *)(base)) -> oh_ci)
 #elif defined(GC_ADD_CALLER)
   struct callinfo;
@@ -140,6 +146,11 @@ typedef struct {
 # define PRINT_CALL_CHAIN(base)
 #endif
 
+#if !defined(ADD_CALL_CHAIN_INNER) && defined(DBG_HDRS_ALL)
+  /* A variant of ADD_CALL_CHAIN() used for internal allocations.   */
+# define ADD_CALL_CHAIN_INNER(base) ADD_CALL_CHAIN(base, GC_RETURN_ADDR)
+#endif
+
 #ifdef GC_ADD_CALLER
 # define OPT_RA ra,
 #else
diff --git a/os_dep.c b/os_dep.c
index c0e72ba56..2e9e9c0fb 100644
--- a/os_dep.c
+++ b/os_dep.c
@@ -4561,13 +4561,32 @@ GC_API int GC_CALL GC_get_pages_executable(void)
                 /* you could use something like pthread_getspecific.    */
 # endif
   GC_in_save_callers = FALSE;
-#endif
+
+# if defined(THREADS) && defined(DBG_HDRS_ALL)
+#   include "private/dbg_mlc.h"
+
+    /* A dummy version of GC_save_callers() which does not call */
+    /* backtrace().                                             */
+    GC_INNER void GC_save_callers_no_unlock(struct callinfo info[NFRAMES])
+    {
+      GC_ASSERT(I_HOLD_LOCK());
+      info[0].ci_pc = (word)(&GC_save_callers_no_unlock);
+      BZERO(&info[1], sizeof(void *) * (NFRAMES - 1));
+    }
+# endif
+#endif /* REDIRECT_MALLOC */
 
 GC_INNER void GC_save_callers(struct callinfo info[NFRAMES])
 {
   void * tmp_info[NFRAMES + 1];
   int npcs, i;
 
+  GC_ASSERT(I_HOLD_LOCK());
+                /* backtrace may call dl_iterate_phdr which is also     */
+                /* used by GC_register_dynamic_libraries, and           */
+                /* dl_iterate_phdr is not guaranteed to be reentrant.   */
+
+  GC_STATIC_ASSERT(sizeof(struct callinfo) == sizeof(void *));
 # ifdef REDIRECT_MALLOC
     if (GC_in_save_callers) {
       info[0].ci_pc = (word)(&GC_save_callers);
@@ -4575,17 +4594,15 @@ GC_INNER void GC_save_callers(struct callinfo info[NFRAMES])
       return;
     }
     GC_in_save_callers = TRUE;
+    /* backtrace() might call a redirected malloc. */
+    UNLOCK();
+    npcs = backtrace((void **)tmp_info, NFRAMES + 1);
+    LOCK();
+# else
+    npcs = backtrace((void **)tmp_info, NFRAMES + 1);
 # endif
-
-  GC_ASSERT(I_HOLD_LOCK());
-                /* backtrace may call dl_iterate_phdr which is also     */
-                /* used by GC_register_dynamic_libraries, and           */
-                /* dl_iterate_phdr is not guaranteed to be reentrant.   */
-
   /* We retrieve NFRAMES+1 pc values, but discard the first one, since  */
   /* it points to our own frame.                                        */
-  GC_STATIC_ASSERT(sizeof(struct callinfo) == sizeof(void *));
-  npcs = backtrace((void **)tmp_info, NFRAMES + 1);
   i = 0;
   if (npcs > 1) {
     i = npcs - 1;