diff --git a/lib/csu/aarch64c/crt1_c.c b/lib/csu/aarch64c/crt1_c.c
index ca3cf8d39ee3..3fba07a6ae41 100644
--- a/lib/csu/aarch64c/crt1_c.c
+++ b/lib/csu/aarch64c/crt1_c.c
@@ -49,6 +49,7 @@
 
 struct Struct_Obj_Entry;
 void _start(void *, void (*)(void), struct Struct_Obj_Entry *) __dead2 __exported;
+void __start(void *, void (*)(void), struct Struct_Obj_Entry *) __dead2;
 
 #ifdef GCRT
 /* Profiling support. */
@@ -69,11 +70,10 @@ Elf_Auxinfo *__auxargs;
  * linker takes care of initialization otherwise.
  */
 void
-_start(void *auxv,
+__start(void *auxv,
 	void (*cleanup)(void),			/* from shared loader */
 	struct Struct_Obj_Entry *obj)		/* from shared loader */
 {
-	__asm__ volatile(".cfi_undefined c30");
 	int argc = 0;
 	char **argv = NULL;
 	char **env = NULL;
@@ -131,3 +131,20 @@ _start(void *auxv,
 
 	__libc_start1(argc, argv, env, cleanup, main, data_cap, code_cap);
 }
+
+/*
+ * The real entry point _start just sets the unwind info for CRA to undefined
+ * which tells libunwind to stop unwinding and then calls the real __start.
+ * This is needed because ".cfi_undefined" inline assembly in a non-naked
+ * function could be overridden by the default unwinding information.
+ */
+__attribute__((naked, used)) void
+_start(void *auxv,
+	void (*cleanup)(void),			/* from shared loader */
+	struct Struct_Obj_Entry *obj)		/* from shared loader */
+{
+	__asm__(
+	    ".cfi_undefined c30\n"
+	    "bl %0"
+	    :: "i"(__start));
+}
diff --git a/lib/csu/riscv/crt1_c.c b/lib/csu/riscv/crt1_c.c
index 41ae7cff9f2a..04d6e7b26749 100644
--- a/lib/csu/riscv/crt1_c.c
+++ b/lib/csu/riscv/crt1_c.c
@@ -57,7 +57,6 @@ void __start(int argc, char **argv, char **env, void (*cleanup)(void)) __dead2;
 void
 __start(int argc, char **argv, char **env, void (*cleanup)(void))
 {
-	__asm__ volatile(".cfi_undefined ra");
 #ifdef SHOULD_PROCESS_CAP_RELOCS
 	/*
 	 * Initialize __cap_relocs for static executables. The run-time linker
diff --git a/lib/csu/riscv/crt1_s.S b/lib/csu/riscv/crt1_s.S
index 8f77d2358ae7..bf69d1435bb8 100644
--- a/lib/csu/riscv/crt1_s.S
+++ b/lib/csu/riscv/crt1_s.S
@@ -35,6 +35,7 @@
 
 #include <machine/asm.h>
 ENTRY(_start)
+	.cfi_undefined ra	# Stop unwinding here
 	mv	a3, a2		# cleanup
 	addi	a1, a0, 8	# get argv
 	ld	a0, 0(a0)	# load argc
diff --git a/lib/csu/riscv64c/crt1_c.c b/lib/csu/riscv64c/crt1_c.c
index 3108ebf562c9..ef4e8a43ab83 100644
--- a/lib/csu/riscv64c/crt1_c.c
+++ b/lib/csu/riscv64c/crt1_c.c
@@ -49,6 +49,7 @@
 
 struct Struct_Obj_Entry;
 void _start(void *, void (*)(void), struct Struct_Obj_Entry *) __dead2 __exported;
+void __start(void *, void (*)(void), struct Struct_Obj_Entry *) __dead2;
 
 #ifdef GCRT
 /* Profiling support. */
@@ -69,11 +70,10 @@ Elf_Auxinfo *__auxargs;
  * linker takes care of initialization otherwise.
  */
 void
-_start(void *auxv,
+__start(void *auxv,
 	void (*cleanup)(void),			/* from shared loader */
 	struct Struct_Obj_Entry *obj)		/* from shared loader */
 {
-	__asm__ volatile(".cfi_undefined cra");
 	int argc = 0;
 	char **argv = NULL;
 	char **env = NULL;
@@ -128,3 +128,20 @@ _start(void *auxv,
 
 	__libc_start1(argc, argv, env, cleanup, main, NULL, NULL);
 }
+
+/*
+ * The real entry point _start just sets the unwind info for CRA to undefined
+ * which tells libunwind to stop unwinding and then calls the real __start.
+ * This is needed because ".cfi_undefined" inline assembly in a non-naked
+ * function could be overridden by the default unwinding information.
+ */
+__attribute__((naked, used)) void
+_start(void *auxv,
+	void (*cleanup)(void),			/* from shared loader */
+	struct Struct_Obj_Entry *obj)		/* from shared loader */
+{
+	__asm__(
+	    ".cfi_undefined cra\n"
+	    "ccall %0"
+	    :: "i"(__start));
+}