On Fri, Dec 01, 2023 at 09:31:40AM -0700, Tycho Andersen wrote: > On Thu, Nov 30, 2023 at 10:57:01AM -0700, Tycho Andersen wrote: > > On Thu, Nov 30, 2023 at 06:39:39PM +0100, Oleg Nesterov wrote: > > > I think that wake_up_all(wait_pidfd) should have a single caller, > > > do_notify_pidfd(). This probably means it should be shiftef from > > > do_notify_parent() to exit_notify(), I am not sure... > > Indeed, below passes the tests without issue and is much less ugly. So I think I raised that question on another medium already but what does the interaction with de_thread() look like? Say some process creates pidfd for a thread in a non-empty thread-group is created via CLONE_PIDFD. The pidfd_file->private_data is set to struct pid of that task. The task the pidfd refers to later exec's. Once it passed de_thread() the task the pidfd refers to assumes the struct pid of the old thread-group leader and continues. At the same time, the old thread-group leader now assumes the struct pid of the task that just exec'd. So after de_thread() the pidfd now referes to the old thread-group leaders struct pid. Any subsequent operation will fail because the process has already exited. Basically, the pidfd now refers to the old thread-group leader and any subsequent operation will fail even though the task still exists. Conversely, if someone had created a pidfd that referred to the old thread-group leader task then this pidfd will now suddenly refer to the new thread-group leader task for the same reason: the struct pid's were exchanged. So this also means, iiuc, that the pidfd could now be passed to waitid(P_PIFD) to retrieve the status of the old thread-group leader that just got zapped. And for the case where the pidfd referred to the old thread-group leader task you would now suddenly _not_ be able to wait on that task anymore. If these concerns are correct, then I think we need to decide what semantics we want and how to handle this because that's not ok.
