From 6e3f91dc6b79475a411b79b1880c45f2a99bdfa5 Mon Sep 17 00:00:00 2001
From: Ben Frederickson <github@benfrederickson.com>
Date: Fri, 16 Feb 2018 10:15:55 -0800
Subject: [PATCH] Fix deadlock in child threads launched from python

Pybind11 seems to have an issue in gil_scoped_acquire when being
called from child threads created from python. This resulted in a deadlock here:
https://github.com/searchivarius/nmslib/issues/291

A simple workaround is to use the PyGILState_Ensure from the python c-api directly
instead.
---
 python_bindings/nmslib.cc | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/python_bindings/nmslib.cc b/python_bindings/nmslib.cc
index 6bc86ca..764de97 100644
--- a/python_bindings/nmslib.cc
+++ b/python_bindings/nmslib.cc
@@ -74,12 +74,12 @@ struct IndexWrapper {
                                                                    loadParams(space_params))) {
     auto vectSpacePtr = dynamic_cast<VectorSpace<dist_t>*>(space.get());
     if (data_type == DATATYPE_DENSE_VECTOR && vectSpacePtr == nullptr) {
-      throw std::invalid_argument("The space type " + space_type + 
+      throw std::invalid_argument("The space type " + space_type +
                                   " is not compatible with the type DENSE_VECTOR, only dense vector spaces are allowed!");
     }
     auto vectSiftPtr = dynamic_cast<SpaceL2SqrSift*>(space.get());
     if (data_type == DATATYPE_DENSE_UINT8_VECTOR && vectSiftPtr == nullptr) {
-      throw std::invalid_argument("The space type " + space_type + 
+      throw std::invalid_argument("The space type " + space_type +
                                   " is not compatible with the type DENSE_UINT8_VECTOR!");
     }
   }
@@ -361,6 +361,20 @@ struct IndexWrapper {
   ObjectVector data;
 };
 
+// pybind11::gil_scoped_acquire can deadlock when acquiring the GIL on threads
+// created from python (https://github.com/searchivarius/nmslib/issues/291)
+// This might be fixed in a future version of pybind11 (https://github.com/pybind/pybind11/pull/1211)
+// but until then, lets fall back to the python c-api to fix.
+struct AcquireGIL {
+  PyGILState_STATE state;
+  AcquireGIL()
+    : state(PyGILState_Ensure()) {
+  }
+  ~AcquireGIL() {
+    PyGILState_Release(state);
+  }
+};
+
 class PythonLogger
   : public Logger {
  public:
@@ -374,7 +388,7 @@ class PythonLogger
            int line,
            const char * function,
            const std::string & message) {
-    py::gil_scoped_acquire l;
+    AcquireGIL l;
     switch(severity) {
       case LIB_DEBUG:
         inner.attr("debug")(message);