On Mon, May 06, 2019 at 12:45:14PM +0200, Jann Horn wrote:
> +cc Christian Brauner (author of pidfd code)
>
> On Mon, May 6, 2019 at 12:22 PM Solar Designer <solar@xxxxxxxxxxxx> wrote:
> > I totally missed the recent work in this area (I'm not on LKML),
>
> FWIW, you don't usually need to actually read LKML to see when major
> developments happen - I read LWN, which seems to work pretty well for
> that purpose.
>
> > and am
> > now wondering whether the solution that got in ("use /proc/<pid> fds as
> > stable handles on struct pid"):
> >
> > https://lwn.net/Articles/773459/
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a9dce6679d736cb3d612af39bab9f31f8db66f9b
> >
> > is better or worse than what I had proposed in 1999 and 2005 ("locking"
> > of pids for the caller's own visibility only):
> >
> > https://marc.info/?l=linux-kernel&m=112784189115058
> >
> > [Subject starts with "PID reuse safety for userspace apps", in case MARC
> > is ever gone and someone wants to look this up in another archive.
>
> (The kernel people now have lore.kernel.org as an email archive, which
> is much nicer to use IMO - it has search, it has a nice thread view,
> and you can download raw mbox files if you want to reply to a mail:
> <https://lore.kernel.org/lkml/20050927172048.GA3423@xxxxxxxxxxxx/>)
>
> > I proposed a lockpid syscall back then, but I'd use a mere prctl now.]
> >
> > I still like my proposal much better - no dependency on procfs, much
This has always been the goal and with the new CLONE_PIDFD flag we make
this possible. I'm about to send the PR for this merge window.
/proc/<pid> fds stay behind as a convenience for pidfd_send_signal()
only with no new features added to them.
There is some (tedious) historical context why it came to be that way.
My original implementation was already different.
pidfd_send_signal() was basically based on /proc/<pid> fd because some
developers had opposed other solutions since the /proc/<pid> idea seemed
so "simple".
The truth is it gets very very problematic when you think about
returning fds at process creation time. This is why pidfds will be anon
inodes.
> > simpler implementation - but perhaps I'm missing the context here.
>
> Actually, there is ongoing development of pidfd stuff, including
> procfs-less pidfds. You may want to look through
> <https://lore.kernel.org/lkml/?q=f%3Abrauner>, or something like that.
>
> The following series adds anon-inode-based pidfds that can be returned
> from sys_clone():
> "[PATCH v3 2/4] clone: add CLONE_PIDFD"
> <https://lore.kernel.org/lkml/20190419120904.27502-2-christian@xxxxxxxxxx/>
> "[PATCH v3 3/4] signal: support CLONE_PIDFD with pidfd_send_signal"
> <https://lore.kernel.org/lkml/20190419120904.27502-4-christian@xxxxxxxxxx/>
>
> This patch adds process exit notifications that can be received
> through the normal file polling syscalls (epoll/poll/select/...):
> "[PATCH v2 1/2] Add polling support to pidfd"
> <https://lore.kernel.org/lkml/20190430162154.61314-1-joel@xxxxxxxxxxxxxxxxx/>
>
> > Maybe I should have sent a patch back then. Oh well.
