fix registry may segment #5336

liufengwei0103 · 2021-06-29T03:55:09Z

No description provided.

lixinqi

common/registry_error.h

Maybe<void> CheckRegistryError();
void CatchRegistryError(const std::function<Maybe<void>()>&);

usage:

CatchRegistryError([&]()->Maybe<void>{
   JUST();
  return Maybe<void>::Ok();
});

lixinqi · 2021-06-30T09:33:14Z

oneflow/api/python/check_registry/check_registry.cpp

@@ -0,0 +1,25 @@
+/*


api/python/registry/registry.cpp

lixinqi · 2021-06-30T09:35:22Z

oneflow/core/common/registry_error.cpp

+
+namespace oneflow {
+
+static std::shared_ptr<Maybe<void>> registry_error;


std::shared_ptrcfg::ErrorProto

maybe只做函数的返回值。做成某个成员或者变量得有非常无奈的理由。

lixinqi · 2021-06-30T09:38:55Z

oneflow/api/python/check_registry/check_registry.cpp

+namespace py = pybind11;
+
+ONEFLOW_API_PYBIND11_MODULE("", m) {
+    m.def("CheckRegistryError", []() { return oneflow::CheckRegistryError().GetOrThrow(); });


CheckAndClearRegistryError

lixinqi · 2021-06-30T09:39:12Z

oneflow/core/common/registry_error.cpp

+namespace oneflow {
+
+static std::shared_ptr<Maybe<void>> registry_error;
+Maybe<void> CheckRegistryError() {


CheckAndClearRegisterError

lixinqi · 2021-06-30T09:39:33Z

oneflow/core/common/registry_error.cpp

+
+static std::shared_ptr<Maybe<void>> registry_error;
+Maybe<void> CheckRegistryError() {
+    return registry_error == nullptr ? Maybe<void>::Ok() : *registry_error;


加clear逻辑。

lixinqi · 2021-06-30T09:40:23Z

oneflow/core/common/registry_error.cpp

+}
+
+void CatchRegistryError(const std::function<Maybe<void>()>& handler) {
+    const auto& maybe_error = handler();


TRY(handler());

lixinqi · 2021-06-30T09:41:42Z

oneflow/core/common/registry_error.cpp

+    const auto& maybe_error = handler();
+    if (!maybe_error.IsOk()) {
+      if (registry_error == nullptr) {
+          registry_error = std::make_shared<Maybe<void>>(maybe_error);


registry_error = maybe_error.error()

lixinqi · 2021-06-30T09:48:02Z

oneflow/extension/python/py_kernel_registry.cpp

  // register python op kernel
  auto reg = user_op::UserOpRegistryMgr::Get()
                 .CheckAndGetOpKernelRegistry(op_module_name + "_forward")
                 .SetCreateFn<PyForwardKernel>()
                 .SetIsMatchedHob(
                     ((user_op::HobDeviceTag() == "cpu") & (user_op::HobDeviceSubTag() == "py")));
-  user_op::UserOpRegistryMgr::Get().Register(reg.Finish().GetResult());
+  user_op::UserOpRegistryMgr::Get().Register(JUST(reg.Finish()).GetResult());


lixinqi · 2021-06-30T09:48:31Z

oneflow/extension/python/py_kernel_registry.cpp

  // register python grad op kernel
  auto grad_reg = user_op::UserOpRegistryMgr::Get()
                      .CheckAndGetOpKernelRegistry(op_module_name + "_backward")
                      .SetCreateFn<PyBackwardKernel>()
                      .SetIsMatchedHob(((user_op::HobDeviceTag() == "cpu")
                                        & (user_op::HobDeviceSubTag() == "py")));
-  user_op::UserOpRegistryMgr::Get().Register(grad_reg.Finish().GetResult());
+  user_op::UserOpRegistryMgr::Get().Register(JUST(grad_reg.Finish()).GetResult());


lixinqi · 2021-07-01T05:43:01Z

oneflow/core/common/registry_error.cpp

+
+namespace oneflow {
+
+static std::shared_ptr<cfg::ErrorProto> registry_error;


std::shared_ptr<cfg::ErrorProto>* MutRegistryError() { static std::shared_ptr<cfg::ErrorProto> registry_error; return registry_error; }

放在匿名空间里

lixinqi · 2021-07-01T05:44:37Z

oneflow/core/common/registry_error.cpp

+    std::shared_ptr<cfg::ErrorProto> registry_error_inner = nullptr;
+    registry_error_inner.swap(registry_error);


std::shared_ptr<cfg::ErrorProto> old_value = MutRegistryError();

lixinqi · 2021-07-01T05:45:21Z

oneflow/core/common/registry_error.cpp

+void CatchRegistryError(const std::function<Maybe<void>()>& handler) {
+    const auto& maybe_error = TRY(handler());
+    if (!maybe_error.IsOk()) {
+      if (registry_error == nullptr) {


if (!*MutRegistryError())

lixinqi · 2021-07-01T05:48:51Z

oneflow/core/framework/user_op_registry.cpp

@@ -24,6 +24,7 @@ limitations under the License.

 namespace oneflow {

+std::shared_ptr<Error> registry_error;


lixinqi · 2021-07-01T08:48:37Z

oneflow/core/common/registry_error.cpp

+        return Maybe<void>::Ok();
+    }
+    std::shared_ptr<cfg::ErrorProto> registry_error_old = *MutRegistryError();
+    *MutRegistryError() = nullptr;


MutRegistryError()->reset();

liufengwei0103 force-pushed the fix_registry_may_segment branch 3 times, most recently from c8143f0 to 0390102 Compare June 29, 2021 04:21

lixinqi reviewed Jun 29, 2021

View reviewed changes

liufengwei0103 force-pushed the fix_registry_may_segment branch 5 times, most recently from cf440e5 to b6440be Compare June 30, 2021 06:09

lixinqi reviewed Jun 30, 2021

View reviewed changes

liufengwei0103 force-pushed the fix_registry_may_segment branch 3 times, most recently from 735f8ac to 7834e75 Compare June 30, 2021 13:51

liufengwei0103 added system enhancement labels Jun 30, 2021

liufengwei0103 requested a review from oneflow-ci-bot June 30, 2021 13:57

liufengwei0103 force-pushed the fix_registry_may_segment branch from 3e04917 to 2ee7096 Compare July 1, 2021 02:15

lixinqi reviewed Jul 1, 2021

View reviewed changes

liufengwei0103 force-pushed the fix_registry_may_segment branch 2 times, most recently from f51226d to c7e5822 Compare July 1, 2021 07:14

lixinqi reviewed Jul 1, 2021

View reviewed changes

lixinqi approved these changes Jul 1, 2021

View reviewed changes

liufengwei0103 added the automerge label Jul 1, 2021

liufengwei0103 requested review from oneflow-ci-bot and removed request for oneflow-ci-bot July 1, 2021 09:03

oneflow-ci-bot requested review from oneflow-ci-bot and removed request for oneflow-ci-bot July 1, 2021 11:34

jackalcooper removed the automerge label Jul 1, 2021

liufengwei0103 force-pushed the fix_registry_may_segment branch 2 times, most recently from c7e5822 to 5073bd2 Compare July 2, 2021 03:05

liufengwei0103 added the automerge label Jul 2, 2021

liufengwei0103 requested a review from oneflow-ci-bot July 2, 2021 03:29

fix registry may segment

634f5aa

liufengwei0103 force-pushed the fix_registry_may_segment branch from 11d86a9 to 634f5aa Compare July 2, 2021 03:45

liufengwei0103 and others added 2 commits July 2, 2021 11:52

Merge branch 'master' into fix_registry_may_segment

69ca413

Merge branch 'master' into fix_registry_may_segment

4f0a4cc

oneflow-ci-bot requested review from oneflow-ci-bot and removed request for oneflow-ci-bot July 2, 2021 04:09

Merge branch 'master' into fix_registry_may_segment

38e3ffb

oneflow-ci-bot self-requested a review July 2, 2021 05:22

oneflow-ci-bot merged commit 52ecdb9 into master Jul 2, 2021

oneflow-ci-bot deleted the fix_registry_may_segment branch July 2, 2021 06:17

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix registry may segment #5336

fix registry may segment #5336

liufengwei0103 commented Jun 29, 2021

lixinqi left a comment

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jun 30, 2021

lixinqi Jul 1, 2021

lixinqi Jul 1, 2021

lixinqi Jul 1, 2021

lixinqi Jul 1, 2021

lixinqi Jul 1, 2021

lixinqi Jul 1, 2021


		namespace oneflow {

		static std::shared_ptr<Maybe<void>> registry_error;


		namespace oneflow {

		static std::shared_ptr<cfg::ErrorProto> registry_error;

		std::shared_ptr<cfg::ErrorProto> registry_error_inner = nullptr;
		registry_error_inner.swap(registry_error);

		@@ -24,6 +24,7 @@ limitations under the License.

		namespace oneflow {

		std::shared_ptr<Error> registry_error;

fix registry may segment #5336

fix registry may segment #5336

Conversation

liufengwei0103 commented Jun 29, 2021

lixinqi left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment