On Sun, Mar 02, 2025 at 06:21:49PM +0100, Oleg Nesterov wrote:
> On 03/02, Christian Brauner wrote:
> >
> > On Sun, Mar 02, 2025 at 04:53:46PM +0100, Oleg Nesterov wrote:
> > > On 02/28, Christian Brauner wrote:
> > > >
> > > > Some tools like systemd's jounral need to retrieve the exit and cgroup
> > > > information after a process has already been reaped.
> > > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > >
> > > But unless I am totally confused do_exit() calls pidfd_exit() even
> > > before exit_notify(), the exiting task is not even zombie yet. It
> > > will reaped only when it passes exit_notify() and its parent does
> > > wait().
> >
> > The overall goal is that it's possible to retrieve exit status and
> > cgroupid even if the task has already been reaped.
>
> OK, please see below...
>
> > It's intentionally placed before exit_notify(), i.e., before the task is
> > a zombie because exit_notify() wakes pidfd-pollers. Ideally, pidfd
> > pollers would be woken and then could use the PIDFD_GET_INFO ioctl to
> > retrieve the exit status.
>
> This was more a less clear to me. But this doesn't match the "the task has
> already been reaped" goal above...
>
> > It would however be fine to place it into exit_notify() if it's a better
> > fit there. If you have a preference let me know.
> >
> > I don't see a reason why seeing the exit status before that would be an
> > issue.
>
> The problem is that it is not clear how can we do this correctly.
> Especialy considering the problem with exec...
>
> > > But what if this file was created without PIDFD_THREAD? If another
> > > thread does exit_group(1) after that, the process's exit code is
> > > 1 << 8, but it can't be retrieved.
> >
> > Yes, I had raised that in an off-list discussion about this as well and
> > was unsure what the cleanest way of dealing with this would be.
>
> I am not sure too, but again, please see below.
>
> > > Now, T is very much alive, but pidfs_i(inode)->exit_info != NULL.
>
> ...
>
> > What's the best way of handling the de_thread() case? Would moving this
> > into exit_notify() be enough where we also handle
> > PIDFD_THREAD/~PIDFD_THREAD waking?
>
> I don't think that moving pidfd_exit() into exit_notify() can solve any
> problem.
>
> But what if we move pidfd_exit() into release_task() paths? Called when
> the task is reaped by the parent/debugger, or if a sub-thread auto-reaps.
>
> Can the users of pidfd_info(PIDFD_INFO_EXIT) rely on POLLHUP from
> release_task() -> detach_pid() -> __change_pid(new => NULL) ?
Ok, so:
release_task()
-> __exit_signal()
-> detach_pid()
-> __change_pid()
That sounds good. So could we do something like:
diff --git a/kernel/exit.c b/kernel/exit.c
index cae475e7858c..66bb5c53454f 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -127,8 +127,10 @@ static void __unhash_process(struct task_struct *p, bool group_dead)
{
nr_threads--;
detach_pid(p, PIDTYPE_PID);
+ pidfs_exit(p); // record exit information for individual thread
if (group_dead) {
detach_pid(p, PIDTYPE_TGID);
+ pidfs_exit(p); // record exit information for thread-group leader
detach_pid(p, PIDTYPE_PGID);
detach_pid(p, PIDTYPE_SID);
I know, as written this won't work but I'm just trying to get the idea
across of recording exit information for both the individual thread and
the thread-group leader in __unhash_process().
That should tackle both problems, i.e., recording exit information for
both thread and thread-group leader as well as exec?
