fix(profiling): fix possible deadlock (#4108) (#4116)

It is possible that while the profiling StackCollector is handling the
on_span_activate event that a gc occurs which could result in a weakref
finalize callback being called which then could activate a new span.
This sequence of events results in a deadlock as the lock the
StackCollector uses is not reentrant.

A regression test would be great but we cannot think of a tractable test
case which wouldn't be flaky.

(cherry picked from commit b0967c4)

Co-authored-by: Kyle Verhoog <[email protected]>

Author: mergify[bot]

ddtrace/internal/nogevent.py

@@ -37,6 +37,7 @@ def is_module_patched(module):
 thread_get_ident = get_original(six.moves._thread.__name__, "get_ident")
 Thread = get_original("threading", "Thread")
 Lock = get_original("threading", "Lock")
+RLock = get_original("threading", "RLock")
 
 is_threading_patched = is_module_patched("threading")
 

ddtrace/profiling/_threading.pyx

@@ -69,7 +69,9 @@ 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])
-    _lock = attr.ib(factory=nogevent.Lock, repr=False, init=False, type=nogevent.Lock)
+    # 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,

releasenotes/notes/profiling-threadlink-lock-27df798c36ed64e6.yaml

@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    profiling: fix a possible deadlock due to spans being activated unexpectedly.