fix(profiling): release the GIL for internal mutation (#14852)

## Description

We make the stack v2 sampler release the GIL when performing internal
state mutation that can require acquiring other internal locks. This
gives the chance to other threads to make progress while the tracking
operations wait on an internal mutex.

Author: P403n1x87

ddtrace/internal/datadog/profiling/stack_v2/src/stack_v2.cpp

@@ -73,7 +73,10 @@ stack_v2_thread_register(PyObject* self, PyObject* args)
         return NULL;
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().register_thread(id, native_id, name);
+    Py_END_ALLOW_THREADS;
+
     Py_RETURN_NONE;
 }
 
@@ -87,8 +90,11 @@ stack_v2_thread_unregister(PyObject* self, PyObject* args)
         return NULL;
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().unregister_thread(id);
     ThreadSpanLinks::get_instance().unlink_span(id);
+    Py_END_ALLOW_THREADS;
+
     Py_RETURN_NONE;
 }
 
@@ -122,7 +128,9 @@ _stack_v2_link_span(PyObject* self, PyObject* args, PyObject* kwargs)
         span_type = empty_string.c_str();
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     ThreadSpanLinks::get_instance().link_span(thread_id, span_id, local_root_span_id, std::string(span_type));
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -140,7 +148,9 @@ stack_v2_track_asyncio_loop(PyObject* self, PyObject* args)
         return NULL;
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().track_asyncio_loop(thread_id, loop);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -172,7 +182,9 @@ stack_v2_link_tasks(PyObject* self, PyObject* args)
         return NULL;
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().link_tasks(parent, child);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -211,7 +223,9 @@ track_greenlet(PyObject* Py_UNUSED(m), PyObject* args)
         return NULL;
     }
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().track_greenlet(greenlet_id, greenlet_name, frame);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -223,7 +237,9 @@ untrack_greenlet(PyObject* Py_UNUSED(m), PyObject* args)
     if (!PyArg_ParseTuple(args, "l", &greenlet_id))
         return NULL;
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().untrack_greenlet(greenlet_id);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -236,7 +252,9 @@ link_greenlets(PyObject* Py_UNUSED(m), PyObject* args)
     if (!PyArg_ParseTuple(args, "ll", &child, &parent))
         return NULL;
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().link_greenlets(parent, child);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }
@@ -250,7 +268,9 @@ update_greenlet_frame(PyObject* Py_UNUSED(m), PyObject* args)
     if (!PyArg_ParseTuple(args, "lO", &greenlet_id, &frame))
         return NULL;
 
+    Py_BEGIN_ALLOW_THREADS;
     Sampler::get().update_greenlet_frame(greenlet_id, frame);
+    Py_END_ALLOW_THREADS;
 
     Py_RETURN_NONE;
 }

releasenotes/notes/fix-profiler-release-gil-internal-mutation-2f0fc19f83759a72.yaml

@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    Profiling: prevent potential deadlocks with thread pools.