From 0e310971a3cf0e1f6abbf1e7fe65aac40f4c97f8 Mon Sep 17 00:00:00 2001
From: Tyler Goodlet <jgbt@protonmail.com>
Date: Wed, 6 Mar 2024 13:32:04 -0500
Subject: [PATCH] Only `ctx.canccel()` when non-callee ctxc received

Since presumably if the ctxc *is* received from the callee side, there's
no point in also sending it a cancel request ;)

Also, try out shielding the `ctx.cancel()` call inside
`Portal.open_context()` since that seems to resolve testing assertions
where you'd expect `.cancel_called` to have been set in the case where
an inter-peer cancellation is raised by some local ctx-handle and that
local ctxc causes other surrounding `.open_context()` blocks to raise
and then implicitly call `ctx.cancel()`. However, there's a possible
race where the callee side might deliver a ctxc while all that is
happening, `._scope.cancel()` gets called and then by the time
`ctx.cancel()` is called it raises `trio.Cancelled` instead continuing
to set `._cancel_called`? This needs more testing and docs to get the
minute details correct .. yet again.
---
 tractor/_portal.py | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

diff --git a/tractor/_portal.py b/tractor/_portal.py
index 8148a5d9..928a8b9c 100644
--- a/tractor/_portal.py
+++ b/tractor/_portal.py
@@ -42,7 +42,7 @@ from async_generator import asynccontextmanager
 from .trionics import maybe_open_nursery
 from .devx import (
     # acquire_debug_lock,
-    # pause,
+    pause,
     maybe_wait_for_debugger,
 )
 from ._state import (
@@ -556,7 +556,7 @@ class Portal:
         # placeholder for any exception raised in the runtime
         # or by user tasks which cause this context's closure.
         scope_err: BaseException|None = None
-        ctxc_from_callee: ContextCancelled|None = None
+        ctxc_from_callee: ContextCancelled|bool = False
         try:
             async with trio.open_nursery() as nurse:
 
@@ -673,8 +673,6 @@ class Portal:
         #   `Nursery.cancel_scope.cancel()`)
         except ContextCancelled as ctxc:
             scope_err = ctxc
-            ctxc_from_callee = ctxc
-
             # XXX TODO XXX: FIX THIS debug_mode BUGGGG!!!
             # using this code and then resuming the REPL will
             # cause a SIGINT-ignoring HANG!
@@ -686,6 +684,10 @@ class Portal:
             # 
             # await pause()
 
+            # ctxc_from_callee: bool = (
+            #     ctxc.src_actor_uid == self.chan.uid
+            # )
+
             # CASE 2: context was cancelled by local task calling
             # `.cancel()`, we don't raise and the exit block should
             # exit silently.
@@ -696,6 +698,7 @@ class Portal:
                 and
                 ctxc.canceller == self.actor.uid
             ):
+                ctxc_from_callee = True
                 log.cancel(
                     f'Context (cid=[{ctx.cid[-6:]}..] cancelled gracefully with:\n'
                     f'{ctxc}'
@@ -755,6 +758,7 @@ class Portal:
                 f'{caller_err}\n'
             )
 
+            # TODO: does this need to be shielded?
             if debug_mode():
                 # async with acquire_debug_lock(self.actor.uid):
                 #     pass
@@ -771,12 +775,14 @@ class Portal:
                         'Calling `ctx.cancel()`!\n'
                     )
 
-
             # we don't need to cancel the callee if it already
             # told us it's cancelled ;p
-            if ctxc_from_callee is None:
+            if not ctxc_from_callee:
+                log.critical('SENDING CANCEL')
                 try:
-                    await ctx.cancel()
+                    # await pause(shield=True)
+                    with trio.CancelScope(shield=True):
+                        await ctx.cancel()
                 except (
                     trio.BrokenResourceError,
                     trio.ClosedResourceError,