Hello Ludo, May 8, 2025, 22:34 by l...@gnu.org: > Hello, > > So I think my last attempt was misguided; attached is a simpler > reproducer that exhibits the problem following these steps: > > 1. create a socket, bind it, and mark it as O_NONBLOCK; > 2. select on that socket to wait for an incoming connection; > 3. clear its O_NONBLOCK flag; > 4. call accept, which returns a connected socket; > 5. notice that O_NONBLOCK is set on that new socket (it shouldn’t). > > I believe the problem is that the connection is made why the original > socket is marked O_NONBLOCK, and pflocal’s ‘S_socket_connect’ inherits > that O_NONBLOCK flag in the socket it gets via ‘sock_clone’. The test > program later clears O_NONBLOCK but it has no effect since in pflocal > the connected socket has already been created with O_NONBLOCK and queued. > > The patch below fixes that (and thus fixes the Shepherd socket > activation, as yelninei initially reported). > > Thoughts? > > Ludo’. > I tested your patch and it solves the issue for all the test cases I had, in particular guix-daemon works instantly again.
Thank you for looking into it.
