Make interaction between test_load_nvidia_dynamic_lib.py and child_load_nvidia_dynamic_lib_helper.py more robust to make it less likely that bugs like 2642d35 slip in.

Author: rwgk

cuda_pathfinder/tests/child_load_nvidia_dynamic_lib_helper.py

@@ -5,8 +5,10 @@
 # lightweight module. That avoids re-importing the test module (and
 # repeating its potentially expensive setup) in every child process.
 
+import json
 import os
 import sys
+import traceback
 
 
 def build_child_process_failed_for_libname_message(libname, result):
@@ -24,15 +26,20 @@ def validate_abs_path(abs_path):
 
 
 def child_process_func(libname):
-    from cuda.pathfinder import load_nvidia_dynamic_lib
+    from cuda.pathfinder import DynamicLibNotFoundError, load_nvidia_dynamic_lib
     from cuda.pathfinder._dynamic_libs.load_nvidia_dynamic_lib import _load_lib_no_cache
     from cuda.pathfinder._dynamic_libs.supported_nvidia_libs import (
         IS_WINDOWS,
         SUPPORTED_LINUX_SONAMES,
         SUPPORTED_WINDOWS_DLLS,
     )
 
-    loaded_dl_fresh = load_nvidia_dynamic_lib(libname)
+    try:
+        loaded_dl_fresh = load_nvidia_dynamic_lib(libname)
+    except DynamicLibNotFoundError:
+        sys.stdout.write("CHILD_LOAD_NVIDIA_DYNAMIC_LIB_HELPER_DYNAMIC_LIB_NOT_FOUND_ERROR:\n")
+        traceback.print_exc(file=sys.stdout)
+        return
     if loaded_dl_fresh.was_already_loaded_from_elsewhere:
         raise RuntimeError("loaded_dl_fresh.was_already_loaded_from_elsewhere")
     validate_abs_path(loaded_dl_fresh.abs_path)
@@ -50,4 +57,4 @@ def child_process_func(libname):
         raise RuntimeError(f"not os.path.samefile({loaded_dl_no_cache.abs_path=!r}, {loaded_dl_fresh.abs_path=!r})")
     validate_abs_path(loaded_dl_no_cache.abs_path)
 
-    sys.stdout.write(f"{loaded_dl_fresh.abs_path!r}\n")
+    sys.stdout.write(json.dumps(loaded_dl_fresh.abs_path) + "\n")

cuda_pathfinder/tests/test_load_nvidia_dynamic_lib.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import functools
+import json
 import os
 from unittest.mock import patch
 
@@ -98,9 +99,18 @@ def test_load_nvidia_dynamic_lib(info_summary_append, libname):
     # interfere across test cases and lead to nondeterministic or platform-specific failures.
     timeout = 120 if supported_nvidia_libs.IS_WINDOWS else 30
     result = spawned_process_runner.run_in_spawned_child_process(child_process_func, args=(libname,), timeout=timeout)
-    if result.returncode == 0:
-        info_summary_append(f"abs_path={result.stdout.rstrip()}")
-    elif STRICTNESS == "see_what_works" and "DynamicLibNotFoundError: Failure finding " in result.stderr:
+
+    def raise_child_process_failed():
+        raise RuntimeError(build_child_process_failed_for_libname_message(libname, result))
+
+    if result.returncode != 0:
+        raise_child_process_failed()
+    assert not result.stderr
+    if result.stdout.startswith("CHILD_LOAD_NVIDIA_DYNAMIC_LIB_HELPER_DYNAMIC_LIB_NOT_FOUND_ERROR:"):
+        if STRICTNESS == "all_must_work":
+            raise_child_process_failed()
         info_summary_append(f"Not found: {libname=!r}")
     else:
-        raise RuntimeError(build_child_process_failed_for_libname_message(libname, result))
+        abs_path = json.loads(result.stdout.rstrip())
+        info_summary_append(f"{abs_path=}")
+        assert os.path.isfile(abs_path)  # double-check the abs_path