On Mon, Apr 15, 2019 at 03:52:48PM +0200, Christian Brauner wrote:
> On Mon, Apr 15, 2019 at 03:24:16PM +0200, Oleg Nesterov wrote:
> > On 04/15, Christian Brauner wrote:
> > >
> > > > CLONE_PARENT_SETTID doesn't look very usefule, so what if we add
> > > >
> > > > if ((clone_flags & (CLONE_PIDFD|CLONE_PARENT_SETTID)) ==
> > > > (CLONE_PIDFD|CLONE_PARENT_SETTID))
> > > > return ERR_PTR(-EINVAL);
> > > >
> > > > at the start of copy_process() ?
> > > >
> > > > Then it can do
> > > >
> > > > if (clone_flags & CLONE_PIDFD) {
> > > > retval = pidfd_create(pid, &pidfdf);
> > > > if (retval < 0)
> > > > goto bad_fork_free_pid;
> > > > retval = put_user(retval, parent_tidptr)
> > > > if (retval < 0)
> > > > goto bad_fork_free_pid;
> > > > }
> > >
> > > Uhhh Oleg, that is nifty. I have to say I like that a lot. This would
> > > let us return the pid and the pidfd in one go and we can also start
> > > pidfd numbering at 0.
> >
> > Christian, sorry if it was already discussed, but I can't force myself to
> > read all the previous discussions ;)
> >
> > If we forget about CONFIG_PROC_FS, why do we really want to create a file?
> >
> >
> > Suppose we add a global u64 counter incremented by copy_process and reported
> > in /proc/$pid/status. Suppose that clone(CLONE_PIDFD) writes this counter to
> > *parent_tidptr. Let's denote this counter as UNIQ_PID.
> >
> > Now, if you want to (say) safely kill a task and you have its UNIQ_PID, you
> > can do
> >
> > kill_by_pid_uniq(int pid, u64 uniq_pid)
> > {
> > pidfd = open("/proc/$pid", O_DIRECTORY);
> >
> > status = openat(pidfd, "status");
> > u64 this_uniq_pid = ... read UNIQ_PID from status ...;
> >
> > if (uniq_pid != this_uniq_pid)
> > return;
> >
> > pidfd_send_signal(pidfd);
> > }
> >
> > Why else do we want pidfd?
>
> I think this was thrown around at one point but this is rather
> inelegant imho. It basically makes a process unique by using a
> combination of two identifiers. You end up with a similar concept but
> you make it way less flexible and extensible imho. With pidfds you can
> not care about pids at all if you don't want to. The UNIQ_PID thing
> would require you to always juggle two identifiers.
>
> Your example would also only work if CONFIG_PROC_FS is set (Not sure if
> that's what you meant by "forget about CONFIG_PROC_FS")? Say, you get
> a pid from clone() and your UNIQ_PID thing. Then you still can't
> reliably kill a process because pidfd_send_signal() is not useable since
> you can't get pidfds. And if you go the kill way you end up with the same
> problem. Yes, you could solve this by probably extending syscalls to
> take a UNIQ_PID argument but that seems very inelegant.
>
> The UNIQ_PID implementation would also require being tracked in the
> kernel either in task_struct or struct pid potentially and thus would
> probably add more infrastructure in the kernel. We don't need any of
> that if we simply rely on pidfds.
>
> Most of all, the pidfd concept allows one way more flexibility in
> extending it. For example, Joel is working on a patchset to make pidfds
> pollable so you can get information about a process death by polling
> them. We also want to be able to potentially wait on a process with
> waitid(W_PIDFD) or similar as suggested by Linus in earlier threads. At
> that point you end up in a similar situation as tgkill() where you pass
> a tgid and a pid already to make sure that the pid you pass has the tgid
> as thread-group leader. That is all way simpler with pidfds.
I agree the pidfd file descriptor approach is simpler than dealing with 2
pids and is needed for the poll notification support I posted.
Also in the future it allows for a pidfd to sent over IPC to another
process using binder or unix domain sockets.
thanks,
- Joel
