refactor(profiling): remove lock from _ThreadLink (#4147) (#4151)

The _ThreadLink data structure historically used a lock to make sure there was
no threads overlapping each other. For example, this could happen:

1. Main thread from the application starts a trace and `link_object` is being
   called
2. Stack profiler would run and clear the thread list or get an spans from the
   list

Removing the lock makes it possible that, while the stack profiler thread
clears the mapping, a user's thread links a thread to a span.

That's a trade-off we are willing to accept for performance and safeness
reason.

(cherry picked from commit 39bc44f)

Co-authored-by: Julien Danjou <[email protected]>

Author: mergify[bot]

ddtrace/profiling/_threading.pyx

@@ -69,19 +69,14 @@ class _ThreadLink(_thread_link_base):
     # Key is a thread_id
     # Value is a weakref to an object
     _thread_id_to_object = attr.ib(factory=dict, repr=False, init=False, type=typing.Dict[int, _weakref_type])
-    # Note that this lock has to be reentrant as spans can be activated unexpectedly in the same thread
-    # ex. during a gc weakref finalize callback
-    _lock = attr.ib(factory=nogevent.RLock, repr=False, init=False, type=nogevent.RLock)
 
     def link_object(
             self,
             obj  # type: _T
     ):
         # type: (...) -> None
         """Link an object to the current running thread."""
-        # Since we're going to iterate over the set, make sure it's locked
-        with self._lock:
-            self._thread_id_to_object[nogevent.thread_get_ident()] = weakref.ref(obj)
+        self._thread_id_to_object[nogevent.thread_get_ident()] = weakref.ref(obj)
 
     def clear_threads(self,
                       existing_thread_ids,  # type: typing.Set[int]
@@ -92,11 +87,17 @@ class _ThreadLink(_thread_link_base):
 
         :param existing_thread_ids: A set of thread ids to keep.
         """
-        with self._lock:
-            # Iterate over a copy of the list of keys since it's mutated during our iteration.
-            for thread_id in list(self._thread_id_to_object.keys()):
-                if thread_id not in existing_thread_ids:
-                    del self._thread_id_to_object[thread_id]
+        # This code clears the thread/object mapping by clearing a copy and swapping it in an atomic operation This is
+        # needed to be able to have this whole class lock-free and avoid concurrency issues.
+        # The fact that it is lock free means we might lose some accuracy, but it's worth the trade-off for speed and simplicity.
+        new_thread_id_to_object_mapping = self._thread_id_to_object.copy()
+        # Iterate over a copy of the list of keys since it's mutated during our iteration.
+        for thread_id in list(new_thread_id_to_object_mapping):
+            if thread_id not in existing_thread_ids:
+                del new_thread_id_to_object_mapping[thread_id]
+
+        # Swap with the new list
+        self._thread_id_to_object = new_thread_id_to_object_mapping
 
     def get_object(
             self,
@@ -108,8 +109,6 @@ class _ThreadLink(_thread_link_base):
         :param thread_id: The thread id.
         :return: The attached object.
         """
-
-        with self._lock:
-            obj_ref = self._thread_id_to_object.get(thread_id)
-            if obj_ref is not None:
-                return obj_ref()
+        obj_ref = self._thread_id_to_object.get(thread_id)
+        if obj_ref is not None:
+            return obj_ref()