On Fri, Jul 25, 2025 at 12:24:28PM +1000, Aleksa Sarai wrote:
> On 2025-07-24, Christian Brauner <brauner@xxxxxxxxxx> wrote:
> > On Wed, Jul 23, 2025 at 09:18:53AM +1000, Aleksa Sarai wrote:
> > > /proc has historically had very opaque semantics about PID namespaces,
> > > which is a little unfortunate for container runtimes and other programs
> > > that deal with switching namespaces very often. One common issue is that
> > > of converting between PIDs in the process's namespace and PIDs in the
> > > namespace of /proc.
> > >
> > > In principle, it is possible to do this today by opening a pidfd with
> > > pidfd_open(2) and then looking at /proc/self/fdinfo/$n (which will
> > > contain a PID value translated to the pid namespace associated with that
> > > procfs superblock). However, allocating a new file for each PID to be
> > > converted is less than ideal for programs that may need to scan procfs,
> > > and it is generally useful for userspace to be able to finally get this
> > > information from procfs.
> > >
> > > So, add a new API for this in the form of an ioctl(2) you can call on
> > > the root directory of procfs. The returned file descriptor will have
> > > O_CLOEXEC set. This acts as a sister feature to the new "pidns" mount
> > > option, finally allowing userspace full control of the pid namespaces
> > > associated with procfs instances.
> > >
> > > The permission model for this is a bit looser than that of the "pidns"
> > > mount option, but this is mainly because /proc/1/ns/pid provides the
> > > same information, so as long as you have access to that magic-link (or
> > > something equivalently reasonable such as privileges with CAP_SYS_ADMIN
> > > or being in an ancestor pid namespace) it makes sense to allow userspace
> > > to grab a handle. setns(2) will still have their own permission checks,
> > > so being able to open a pidns handle doesn't really provide too many
> > > other capabilities.
> > >
> > > Signed-off-by: Aleksa Sarai <cyphar@xxxxxxxxxx>
> > > ---
> > > Documentation/filesystems/proc.rst | 4 +++
> > > fs/proc/root.c | 54 ++++++++++++++++++++++++++++++++++++--
> > > include/uapi/linux/fs.h | 3 +++
> > > 3 files changed, 59 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/Documentation/filesystems/proc.rst b/Documentation/filesystems/proc.rst
> > > index c520b9f8a3fd..506383273c9d 100644
> > > --- a/Documentation/filesystems/proc.rst
> > > +++ b/Documentation/filesystems/proc.rst
> > > @@ -2398,6 +2398,10 @@ pidns= specifies a pid namespace (either as a string path to something like
> > > will be used by the procfs instance when translating pids. By default, procfs
> > > will use the calling process's active pid namespace.
> > >
> > > +Processes can check which pid namespace is used by a procfs instance by using
> > > +the `PROCFS_GET_PID_NAMESPACE` ioctl() on the root directory of the procfs
> > > +instance.
> > > +
> > > Chapter 5: Filesystem behavior
> > > ==============================
> > >
> > > diff --git a/fs/proc/root.c b/fs/proc/root.c
> > > index 057c8a125c6e..548a57ec2152 100644
> > > --- a/fs/proc/root.c
> > > +++ b/fs/proc/root.c
> > > @@ -23,8 +23,10 @@
> > > #include <linux/cred.h>
> > > #include <linux/magic.h>
> > > #include <linux/slab.h>
> > > +#include <linux/ptrace.h>
> > >
> > > #include "internal.h"
> > > +#include "../internal.h"
> > >
> > > struct proc_fs_context {
> > > struct pid_namespace *pid_ns;
> > > @@ -418,15 +420,63 @@ static int proc_root_readdir(struct file *file, struct dir_context *ctx)
> > > return proc_pid_readdir(file, ctx);
> > > }
> > >
> > > +static long int proc_root_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
> > > +{
> > > + switch (cmd) {
> > > +#ifdef CONFIG_PID_NS
> > > + case PROCFS_GET_PID_NAMESPACE: {
> > > + struct pid_namespace *active = task_active_pid_ns(current);
> > > + struct pid_namespace *ns = proc_pid_ns(file_inode(filp)->i_sb);
> > > + bool can_access_pidns = false;
> > > +
> > > + /*
> > > + * If we are in an ancestors of the pidns, or have join
> > > + * privileges (CAP_SYS_ADMIN), then it makes sense that we
> > > + * would be able to grab a handle to the pidns.
> > > + *
> > > + * Otherwise, if there is a root process, then being able to
> > > + * access /proc/$pid/ns/pid is equivalent to this ioctl and so
> > > + * we should probably match the permission model. For empty
> > > + * namespaces it seems unlikely for there to be a downside to
> > > + * allowing unprivileged users to open a handle to it (setns
> > > + * will fail for unprivileged users anyway).
> > > + */
> > > + can_access_pidns = pidns_is_ancestor(ns, active) ||
> > > + ns_capable(ns->user_ns, CAP_SYS_ADMIN);
> >
> > This seems to imply that if @ns is a descendant of @active that the
> > caller holds privileges over it. Is that actually always true?
> >
> > IOW, why is the check different from the previous pidns= mount option
> > check. I would've expected:
> >
> > ns_capable(_no_audit)(ns->user_ns) && pidns_is_ancestor(ns, active)
> >
> > and then the ptrace check as a fallback.
>
> That would mirror pidns_install(), and I did think about it. The primary
> (mostly handwave-y) reasoning I had for making it less strict was that:
>
> * If you are in an ancestor pidns, then you can already see those
> processes in your own /proc. In theory that means that you will be
> able to access /proc/$pid/ns/pid for at least some subprocess there
> (even if some subprocesses have SUID_DUMP_DISABLE, that flag is
> cleared on ).
>
> Though hypothetically if they are all running as a different user,
> this does not apply (and you could create scenarios where a child
> pidns is owned by a userns that you do not have privileges over -- if
> you deal with setuid binaries). Maybe that risk means we should just
> combine them, I'm not sure.
>
> * If you have CAP_SYS_ADMIN permissions over the pidns, it seems
> strange to disallow access even if it is not in an ancestor
> namespace. This is distinct to pidns_install(), where you want to
> ensure you cannot escape to a parent pid namespace, this is about
> getting a handle to do other operations (i.e. NS_GET_{P,TG}ID_*_PIDNS).
>
> Maybe they should be combined to match pidns_install(), but then I would
> expect the ptrace_may_access() check to apply to all processes in the
> pidns to make it less restrictive, which is not something you can
> practically do (and there is a higher chance that pid1 will have
> SUID_DUMP_DISABLE than some random subprocess, which almost certainly
> will not be SUID_DUMP_DISABLE).
>
> Fundamentally, I guess I'm still trying to see what the risk is of
> allowing a process to get a handle to a pidns that they have some kind
> of privilege over (whether it's CAP_SYS_ADMIN, or by the virtue of being
There shouldn't be. For example, you kinda implicitly do that with a
pidfd, no? Because you can pass the pidfd to setns() instead of a
namespace fd itself. Maybe that's the argument you're lookin for?
