Enable Leak Sanitizer (#154584)

It enables Leak Sanitizer and also provides a suppression file.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/154584
Approved by: https://github.com/ezyang

.ci/pytorch/test.sh

@@ -216,11 +216,13 @@ fi
 # if you're not careful.  Check this if you made some changes and the
 # ASAN test is not working
 if [[ "$BUILD_ENVIRONMENT" == *asan* ]]; then
-    export ASAN_OPTIONS=detect_leaks=0:symbolize=1:detect_stack_use_after_return=true:strict_init_order=true:detect_odr_violation=1:detect_container_overflow=0:check_initialization_order=true:debug=true
+    export ASAN_OPTIONS=detect_leaks=1:symbolize=1:detect_stack_use_after_return=true:strict_init_order=true:detect_odr_violation=1:detect_container_overflow=0:check_initialization_order=true:debug=true:fast_unwind_on_malloc=1
     if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
         export ASAN_OPTIONS="${ASAN_OPTIONS}:protect_shadow_gap=0"
     fi
     export UBSAN_OPTIONS=print_stacktrace=1:suppressions=$PWD/ubsan.supp
+    # Suppress some hard to solve indirect leaks
+    export LSAN_OPTIONS="suppressions=$PWD/lsan.supp"
     export PYTORCH_TEST_WITH_ASAN=1
     export PYTORCH_TEST_WITH_UBSAN=1
     # TODO: Figure out how to avoid hard-coding these paths

lsan.supp

@@ -0,0 +1,31 @@
+leak:pybind11::cpp_function
+leak:PyMem_RawMalloc
+leak:unicode_resize
+leak:PyObject_Malloc
+leak:PyByteArray_Resize
+leak:numpy
+leak:list_append
+leak:unicodeobject
+leak:obmalloc
+leak:gcmodule
+leak:listobject
+leak:bytesobject
+leak:PyThread_allocate_lock
+leak:sccache
+leak:rustc-1.61.0
+leak:gcc/x86_64-linux-gnu/11
+leak:x86_64-linux-gnu-gcc-11
+leak:libbfd
+leak:x86_64-linux-gnu-ld.bfd
+leak:git
+leak:libio
+leak:unknown module
+leak:g++
+leak:conda-linux-gnu-ld
+leak:crypto
+leak:torch::detail::(anonymous namespace)::get_set_cached_attr
+leak:torch::jit::tensorexpr::TensorExprKernel::preAllocIntermediateBufs
+leak:optree
+leak:python
+leak:torch::tensors::initialize_aten_types
+leak:libclang_rt

test/dynamo/test_logging.py

@@ -23,6 +23,7 @@ from torch.testing._internal.common_utils import (
     find_free_port,
     munge_exc,
     skipIfTorchDynamo,
+    TEST_WITH_ASAN,
     xfailIfS390X,
 )
 from torch.testing._internal.inductor_utils import HAS_CUDA
@@ -524,6 +525,7 @@ LoweringException: AssertionError:
         with self.assertRaises(ValueError):
             torch._logging.set_logs(aot_graphs=5)
 
+    @unittest.skipIf(TEST_WITH_ASAN, "LSAN outputs suppression report on stderr")
     def test_invalid_artifact_flag_error_msg(self):
         env = dict(os.environ)
         env["TORCH_LOGS"] = "not_an_existing_log_artifact_should_error"
@@ -856,6 +858,7 @@ TRACE FX call mul from test_logging.py:N in fn (LoggingTests.test_trace_call_pre
 
     # there are some additional deprecation warnings in stderr, probably due to newer dependencies used on s390x
     @xfailIfS390X
+    @unittest.skipIf(TEST_WITH_ASAN, "LSAN outputs suppression report on stderr")
     def test_logs_out(self):
         import tempfile
 

torch/csrc/Module.cpp

@@ -235,7 +235,7 @@ static PyObject* THPModule_initExtension(
   END_HANDLE_TH_ERRORS
 }
 
-// The idea behind these two functions is to make it easy to test if we are
+// The idea behind these functions is to make it easy to test if we are
 // built with ASAN: they're designed not to crash if ASAN is not enabled, but
 // to trigger ASAN if it is enabled.  This lets us run a "canary" tests which
 // checks if our build environment is misconfigured.

torch/csrc/autograd/VariableTypeManual.cpp

@@ -456,7 +456,7 @@ static Tensor detach(c10::DispatchKeySet ks, const Tensor& self) {
   // NB: we can't make detach() a normal view operator because the codegen
   // generates allow_tensor_metadata_change = True for them. In the future we
   // should have an option for this in the codegen.
-  auto result = as_view(
+  return as_view(
       /* base */ self,
       /* output */ out,
       /* is_bw_differentiable */ false,
@@ -465,8 +465,6 @@ static Tensor detach(c10::DispatchKeySet ks, const Tensor& self) {
       /* rev_view_func */ nullptr,
       /* creation_meta */ CreationMeta::DEFAULT,
       /*allow_tensor_metadata_change=*/false);
-
-  return result;
 }
 
 static Tensor _fw_primal(

torch/csrc/autograd/variable.h

@@ -844,7 +844,7 @@ inline Variable make_variable_non_differentiable_view(
         /*version_counter=*/impl::version_counter(base),
         /*allow_tensor_metadata_change=*/allow_tensor_metadata_change);
     data_impl_copy->set_autograd_meta(nullptr);
-    return Variable(data_impl_copy);
+    return Variable(std::move(data_impl_copy));
   }
   return Variable();
 }
@@ -903,7 +903,7 @@ inline Variable make_variable(
         /*allow_tensor_metadata_change=*/allow_tensor_metadata_change);
     data_impl_copy->set_autograd_meta(std::make_unique<AutogradMeta>(
         data_impl_copy.get(), false, std::move(gradient_edge)));
-    return Variable(data_impl_copy);
+    return Variable(std::move(data_impl_copy));
   }
   return Variable();
 }