On 2025-09-02, Christian Brauner [off-list ref] wrote:

On Tue, Aug 05, 2025 at 03:45:07PM +1000, Aleksa Sarai wrote:

quoted

Ever since the introduction of pid namespaces, procfs has had very
implicit behaviour surrounding them (the pidns used by a procfs mount is
auto-selected based on the mounting process's active pidns, and the
pidns itself is basically hidden once the mount has been constructed).

/* pidns mount option for procfs */

This implicit behaviour has historically meant that userspace was
required to do some special dances in order to configure the pidns of a
procfs mount as desired. Examples include:

 * In order to bypass the mnt_too_revealing() check, Kubernetes creates
   a procfs mount from an empty pidns so that user namespaced containers
   can be nested (without this, the nested containers would fail to
   mount procfs). But this requires forking off a helper process because
   you cannot just one-shot this using mount(2).

 * Container runtimes in general need to fork into a container before
   configuring its mounts, which can lead to security issues in the case
   of shared-pidns containers (a privileged process in the pidns can
   interact with your container runtime process). While
   SUID_DUMP_DISABLE and user namespaces make this less of an issue, the
   strict need for this due to a minor uAPI wart is kind of unfortunate.

Things would be much easier if there was a way for userspace to just
specify the pidns they want. Patch 1 implements a new "pidns" argument
which can be set using fsconfig(2):

    fsconfig(procfd, FSCONFIG_SET_FD, "pidns", NULL, nsfd);
    fsconfig(procfd, FSCONFIG_SET_STRING, "pidns", "/proc/self/ns/pid", 0);

or classic mount(2) / mount(8):

    // mount -t proc -o pidns=/proc/self/ns/pid proc /tmp/proc
    mount("proc", "/tmp/proc", "proc", MS_..., "pidns=/proc/self/ns/pid");

The initial security model I have in this RFC is to be as conservative
as possible and just mirror the security model for setns(2) -- which
means that you can only set pidns=... to pid namespaces that your
current pid namespace is a direct ancestor of and you have CAP_SYS_ADMIN
privileges over the pid namespace. This fulfils the requirements of
container runtimes, but I suspect that this may be too strict for some
usecases.

The pidns argument is not displayed in mountinfo -- it's not clear to me
what value it would make sense to show (maybe we could just use ns_dname
to provide an identifier for the namespace, but this number would be
fairly useless to userspace). I'm open to suggestions. Note that
PROCFS_GET_PID_NAMESPACE (see below) does at least let userspace get
information about this outside of mountinfo.

Note that you cannot change the pidns of an already-created procfs
instance. The primary reason is that allowing this to be changed would
require RCU-protecting proc_pid_ns(sb) and thus auditing all of
fs/proc/* and some of the users in fs/* to make sure they wouldn't UAF
the pid namespace. Since creating procfs instances is very cheap, it
seems unnecessary to overcomplicate this upfront. Trying to reconfigure
procfs this way errors out with -EBUSY.

/* ioctl(PROCFS_GET_PID_NAMESPACE) */

In addition, being able to figure out what pid namespace is being used
by a procfs mount is quite useful when you have an administrative
process (such as a container runtime) which wants to figure out the
correct way of mapping PIDs between its own namespace and the namespace
for procfs (using NS_GET_{PID,TGID}_{IN,FROM}_PIDNS). There are
alternative ways to do this, but they all rely on ancillary information
that third-party libraries and tools do not necessarily have access to.

To make this easier, add a new ioctl (PROCFS_GET_PID_NAMESPACE) which
can be used to get a reference to the pidns that a procfs is using.

Rather than copying the (fairly strict) security model for setns(2),
apply a slightly looser model to better match what userspace can already
do:

 * Make the ioctl only valid on the root (meaning that a process without
   access to the procfs root -- such as only having an fd to a procfs
   file or some open_tree(2)-like subset -- cannot use this API). This
   means that the process already has some level of access to the
   /proc/$pid directories.

 * If the calling process is in an ancestor pidns, then they can already
   create pidfd for processes inside the pidns, which is morally
   equivalent to a pidns file descriptor according to setns(2). So it
   seems reasonable to just allow it in this case. (The justification
   for this model was suggested by Christian.)

 * If the process has access to /proc/1/ns/pid already (i.e. has
   ptrace-read access to the pidns pid1), then this ioctl is equivalent
   to just opening a handle to it that way.

   Ideally we would check for ptrace-read access against all processes
   in the pidns (which is very likely to be true for at least one
   process, as SUID_DUMP_DISABLE is cleared on exec(2) and is rarely set
   by most programs), but this would obviously not scale.

I'm open to suggestions for whether we need to make this stricter (or
possibly allow more cases).

Signed-off-by: Aleksa Sarai <redacted>

Thanks for the patchset. Being able to specify what pid namespace the
procfs instance is supposed to belong to is super useful and will make
things easier for userspace for sure.

I was going to send a new version changing the whole thing to be struct
path based (and adding FSCONFIG_SET_PATH{,_EMPTY} support) so we don't
need to allocate a file explicitly for the non-FSCONFIG_SET_FD case, but
we can do that as a follow-up I guess.

The code you added contains a minor wrinkle that I disliked which I've
changed and you tell me if you can live with this restriction or not.

The way you've implemented it specifying a pid namespace that the caller
holds privilege over would silently also override the user namespace the
filesystem is supposed to belong to.

Specifically, you did something like:

        put_pid_ns(ctx->pid_ns);
        ctx->pid_ns = get_pid_ns(target);
        put_user_ns(fc->user_ns);
        fc->user_ns = get_user_ns(ctx->pid_ns->user_ns);

This silently overrides the user namespace recorded at fsopen() time. I
think that's too subtle and we should just not allow that at all for
now.

Instead I've changed this to:

        if (fc->user_ns != target->user_ns)
                return invalfc(fc, "owning user namespace of pid namespace doesn't match procfs user namespace");

        put_pid_ns(ctx->pid_ns);
        ctx->pid_ns = get_pid_ns(target);

so we just refuse different owernship.

That sounds fine, I wasn't quite sure what to do with fc->user_ns to be
honest. Being more conservative is probably the right call here.

I've also dropped the procfs ioctl because I'm not sure how much value
it will actually add given that you can do this via /proc/1/ns/pid.

If that is something that libpathrs despearately needs I would like to
do it as a separate patch anyways.

The main issues are:

1. pid1 can often be non-dumpable, which can block you from doing that.
   In principle, because the dumpable flag is reset on execve, it is
   theoretically possible to get access to /proc/$pid/ns/pid if you win
   the race in a pid namespace with lots of process activity, but this
   kind of sucks.

2. This approach doesn't work for empty pid namesapces.
   pidns_for_children doesn't let you get a handle to an empty pid
   namespace either (I briefly looked at the history and it seems this
   was silently changed in v2 of the patchset based on some feedback
   that I'm not sure was entirely correct).

3. Now that you can configure the procfs mount, it seems like a
   half-baked interface to not provide diagnostic information about the
   namespace. (I suspect the criu folks would be happy to have this too
   ;).)

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
https://www.cyphar.com/