Merge pull request #1163 from JuliaLang/shutdown-deadlock

Fix a deadlock in `shutdown_requests`

Author: JamesWrigley

Project.toml

@@ -1,6 +1,6 @@
 name = "IJulia"
 uuid = "7073ff75-c697-5162-941a-fcdaad2a7d2a"
-version = "1.28.0"
+version = "1.28.1"
 
 [deps]
 Base64 = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f"

docs/src/_changelog.md

@@ -7,6 +7,13 @@ CurrentModule = IJulia
 This documents notable changes in IJulia.jl. The format is based on [Keep a
 Changelog](https://keepachangelog.com).
 
+## [v1.28.1]
+
+### Fixed
+
+- Fixed a deadlock in the `shutdown_request` handler that would cause the kernel
+  to hang when exiting ([#1163]).
+
 ## [v1.28.0]
 
 ### Added

src/IJulia.jl

@@ -93,6 +93,8 @@ const capture_stderr = !IJULIA_DEBUG
 set_current_module(m::Module) = current_module[] = m
 const current_module = Ref{Module}(Main)
 
+_shutting_down::Threads.Atomic{Bool} = Threads.Atomic{Bool}(false)
+
 #######################################################################
 include("jupyter.jl")
 #######################################################################

src/eventloop.jl

@@ -13,11 +13,14 @@ function eventloop(socket)
                 send_status("busy", msg)
                 invokelatest(get(handlers, msg.header["msg_type"], unknown_request), socket, msg)
             catch e
-                # Try to keep going if we get an exception, but
-                # send the exception traceback to the front-ends.
-                # (Ignore SIGINT since this may just be a user-requested
-                #  kernel interruption to interrupt long calculations.)
-                if !isa(e, InterruptException)
+                if e isa InterruptException && _shutting_down[]
+                    # If we're shutting down, just return immediately
+                    return
+                elseif !isa(e, InterruptException)
+                    # Try to keep going if we get an exception, but
+                    # send the exception traceback to the front-ends.
+                    # (Ignore SIGINT since this may just be a user-requested
+                    #  kernel interruption to interrupt long calculations.)
                     content = error_content(e, msg="KERNEL EXCEPTION")
                     map(s -> println(orig_stderr[], s), content["traceback"])
                     send_ipython(publish[], msg_pub(execute_msg, "error", content))
@@ -28,6 +31,10 @@ function eventloop(socket)
             end
         end
     catch e
+        if _shutting_down[]
+            return
+        end
+
         # the Jupyter manager may send us a SIGINT if the user
         # chooses to interrupt the kernel; don't crash on this
         if isa(e, InterruptException)

src/handlers.jl

@@ -229,6 +229,14 @@ function shutdown_request(socket, msg)
     # stop heartbeat thread
     stop_heartbeat(heartbeat[], heartbeat_context[])
 
+    # Shutdown the `requests` socket handler before sending any messages. This
+    # is necessary because otherwise the event loop will be calling
+    # `recv_ipython()` and holding a lock on `requests`, which will cause a
+    # deadlock when we try to send a message to it from the `control` socket
+    # handler.
+    global _shutting_down[] = true
+    @async Base.throwto(requests_task[], InterruptException())
+
     send_ipython(requests[], msg_reply(msg, "shutdown_reply",
                                        msg.content))
     sleep(0.1) # short delay (like in ipykernel), to hopefully ensure shutdown_reply is sent