Fix frame cache miss handling in RemoteUnwinder to return complete stacks

pablogsal · pablogsal · commit 32facd8efdfd · 2025-12-01T15:50:08.000Z
When the frame walker stopped at a cached frame boundary but the cache lookup failed to find the continuation (e.g., when a new unwinder instance encounters stale last_profiled_frame values from a previous profiler), the code would silently return an incomplete stack. Now when a cache lookup misses, the walker continues from that point to collect the remaining frames, ensuring complete stack traces are always returned.
diff --git a/Lib/test/test_external_inspection.py b/Lib/test/test_external_inspection.py
@@ -2559,6 +2559,59 @@ def get_thread_frames(target_funcs):
             # No cross-contamination
             self.assertNotIn("blech1", t2_blech)
 
+    @skip_if_not_supported
+    @unittest.skipIf(
+        sys.platform == "linux" and not PROCESS_VM_READV_SUPPORTED,
+        "Test only runs on Linux with process_vm_readv support",
+    )
+    def test_new_unwinder_with_stale_last_profiled_frame(self):
+        """Test that a new unwinder returns complete stack when cache lookup misses."""
+        script_body = """\
+            def level4():
+                sock.sendall(b"sync1")
+                sock.recv(16)
+                sock.sendall(b"sync2")
+                sock.recv(16)
+
+            def level3():
+                level4()
+
+            def level2():
+                level3()
+
+            def level1():
+                level2()
+
+            level1()
+            """
+
+        with self._target_process(script_body) as (p, client_socket, make_unwinder):
+            expected = {"level1", "level2", "level3", "level4"}
+
+            # First unwinder samples - this sets last_profiled_frame in target
+            unwinder1 = make_unwinder(cache_frames=True)
+            frames1 = self._sample_frames(client_socket, unwinder1, b"sync1", b"ack", expected)
+
+            # Create NEW unwinder (empty cache) and sample
+            # The target still has last_profiled_frame set from unwinder1
+            unwinder2 = make_unwinder(cache_frames=True)
+            frames2 = self._sample_frames(client_socket, unwinder2, b"sync2", b"done", expected)
+
+        self.assertIsNotNone(frames1)
+        self.assertIsNotNone(frames2)
+
+        funcs1 = [f.funcname for f in frames1]
+        funcs2 = [f.funcname for f in frames2]
+
+        # Both should have all levels
+        for level in ["level1", "level2", "level3", "level4"]:
+            self.assertIn(level, funcs1, f"{level} missing from first sample")
+            self.assertIn(level, funcs2, f"{level} missing from second sample")
+
+        # Should have same stack depth
+        self.assertEqual(len(frames1), len(frames2),
+                         "New unwinder should return complete stack despite stale last_profiled_frame")
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/Modules/_remote_debugging/frames.c b/Modules/_remote_debugging/frames.c
@@ -770,6 +770,14 @@ collect_frames_with_cache(
             Py_DECREF(frame_addresses);
             return -1;
         }
+        if (cache_result == 0) {
+            // Cache miss - continue walking from last_profiled_frame to get the rest
+            if (process_frame_chain(unwinder, last_profiled_frame, chunks, frame_info, gc_frame,
+                                    0, NULL, frame_addresses) < 0) {
+                Py_DECREF(frame_addresses);
+                return -1;
+            }
+        }
     }
 
     // Convert frame_addresses (list of PyLong) to C array for efficient cache storage

Original file line number	Diff line number	Diff line change
`@@ -770,6 +770,14 @@ collect_frames_with_cache(`
`770`	`770`	`Py_DECREF(frame_addresses);`
`771`	`771`	`return -1;`
`772`	`772`	`}`
	`773`	`+ if (cache_result == 0) {`
	`774`	`+ // Cache miss - continue walking from last_profiled_frame to get the rest`
	`775`	`+ if (process_frame_chain(unwinder, last_profiled_frame, chunks, frame_info, gc_frame,`
	`776`	`+ 0, NULL, frame_addresses) < 0) {`
	`777`	`+ Py_DECREF(frame_addresses);`
	`778`	`+ return -1;`
	`779`	`+ }`
	`780`	`+ }`
`773`	`781`	`}`
`774`	`782`
`775`	`783`	`// Convert frame_addresses (list of PyLong) to C array for efficient cache storage`