Re: [RFC 1/3] [gdb/dap] Fix exit race

[Tom de Vries <tdevries@suse.de>]

On 2/7/24 17:01, Tom Tromey wrote:
>>>>>> "Tom" == Tom de Vries <tdevries@suse.de> writes:
> 
> Tom> Then I tried to simply remove send_gdb("quit"), and that worked as well, and
> Tom> caused no regressions.  So, either this is the easiest way to address this, or
> Tom> we need to add a test-case that regresses when we remove it.
> 
> Right now main_loop does:
> 
>          # Got the terminate request.  This is handled by the
>          # JSON-writing thread, so that we can ensure that all
>          # responses are flushed to the client before exiting.
>          self.write_queue.put(None)
> 
> Like the comment says, this approach tries to ensure that all responses
> have been flushed to the client before quitting.  However, I suppose
> this approach does not really work, because main_loop then proceeds to
> return.
> 
> One way to fix this would be to have start_json_writer return the thread
> object, and then have main_loop join the thread to terminate after
> writing the None.  This would mean making this particular thread
> non-daemon though.
> 

And as I understand it, the downside of that is that it could possibly 
hang the gdb process.

Btw, I think I found a simpler way of achieving this:
...
diff --git a/gdb/python/lib/gdb/dap/io.py b/gdb/python/lib/gdb/dap/io.py
index 5149edae977..491cf208ae3 100644
--- a/gdb/python/lib/gdb/dap/io.py
+++ b/gdb/python/lib/gdb/dap/io.py
@@ -65,10 +65,7 @@ def start_json_writer(stream, queue):
          while True:
              obj = queue.get()
              if obj is None:
-                # This is an exit request.  The stream is already
-                # flushed, so all that's left to do is request an
-                # exit.
-                send_gdb("quit")
+                queue.task_done()
                  break
              obj["seq"] = seq
              seq = seq + 1
@@ -79,5 +76,6 @@ def start_json_writer(stream, queue):
              stream.write(header_bytes)
              stream.write(body_bytes)
              stream.flush()
+            queue.task_done()

      start_thread("JSON writer", _json_writer)
diff --git a/gdb/python/lib/gdb/dap/server.py 
b/gdb/python/lib/gdb/dap/server.py
index 7cc5a4681ee..ca4465ff207 100644
--- a/gdb/python/lib/gdb/dap/server.py
+++ b/gdb/python/lib/gdb/dap/server.py
@@ -229,6 +229,8 @@ class Server:
          # JSON-writing thread, so that we can ensure that all
          # responses are flushed to the client before exiting.
          self.write_queue.put(None)
+        self.write_queue.join()
+        send_gdb("quit")

      @in_dap_thread
      def send_event_later(self, event, body=None):
...

Regardless, this doesn't address the root cause of the problem, the race 
and crash remain.

+> In this setup, removing the "quit" here is the right thing to do.

I've submitted an updated version (commit message is updated, code is 
the same) of the patch here ( 
https://sourceware.org/pipermail/gdb-patches/2024-February/206567.html ).

Thanks,
- Tom