Re: [PATCH] pidfd: prevent creation of pidfds for kthreads

[Christian Brauner <brauner@kernel.org>]

On Tue, Aug 20, 2024 at 12:34:14PM GMT, Eric Biggers wrote:
> On Mon, Aug 19, 2024 at 10:41:15AM +0200, Christian Brauner wrote:
> > On Sat, Aug 17, 2024 at 08:58:18PM GMT, Eric Biggers wrote:
> > > Hi Christian,
> > > 
> > > On Wed, Jul 31, 2024 at 12:01:12PM +0200, Christian Brauner wrote:
> > > > It's currently possible to create pidfds for kthreads but it is unclear
> > > > what that is supposed to mean. Until we have use-cases for it and we
> > > > figured out what behavior we want block the creation of pidfds for
> > > > kthreads.
> > > > 
> > > > Fixes: 32fcb426ec00 ("pid: add pidfd_open()")
> > > > Cc: stable@vger.kernel.org
> > > > Signed-off-by: Christian Brauner <brauner@kernel.org>
> > > > ---
> > > >  kernel/fork.c | 25 ++++++++++++++++++++++---
> > > >  1 file changed, 22 insertions(+), 3 deletions(-)
> > > 
> > > Unfortunately this commit broke systemd-shutdown's ability to kill processes,
> > > which makes some filesystems no longer get unmounted at shutdown.
> > > 
> > > It looks like systemd-shutdown relies on being able to create a pidfd for any
> > > process listed in /proc (even a kthread), and if it gets EINVAL it treats it a
> > > fatal error and stops looking for more processes...
> > 
> > Thanks for the report!
> > I talked to Daan De Meyer who made that change and he said that this
> > must a systemd version that hasn't gotten his fixes yet. In any case, if
> > this causes regression then I'll revert it right now. See the appended
> > revert.
> 
> Thanks for queueing up a revert.
> 
> This was on systemd 256.4 which was released less than a month ago.
> 
> I'm not sure what systemd fix you are talking about.  Looking at killall() in
> src/shared/killall.c on the latest "main" branch of systemd, it calls
> proc_dir_read_pidref() => pidref_set_pid() => pidfd_open(), and EINVAL gets
> passed back up to killall() and treated as a fatal error.  ignore_proc() skips
> kernel threads but is executed too late.  I didn't test it, so I could be wrong,
> but based on the code it does not appear to be fixed.

Yeah, I think you're right. What they fixed is
ead48ec35c86 ("cgroup-util: Don't try to open pidfd for kernel threads")
when reading pids from cgroup.procs. Daan is currently prepping a fix
for reading pids from /proc as well.