[PATCH] perf: Require CAP_KILL if sigtrap is requested

[Ondrej Mosnacek]

On Wed, Jun 30, 2021 at 2:43 PM Marco Elver <elver at google.com> wrote:
> On Wed, Jun 30, 2021 at 01:13PM +0200, Ondrej Mosnacek wrote:
> > On Wed, Jun 30, 2021 at 11:38 AM Marco Elver <elver at google.com> wrote:
> [...]
> > > +static inline bool kill_capable(void)
> > > +{
> > > +       return capable(CAP_KILL) || capable(CAP_SYS_ADMIN);
> >
> > Is it really necessary to fall back to CAP_SYS_ADMIN here? CAP_PERFMON
> > and CAP_BPF have been split off from CAP_SYS_ADMIN recently, so they
> > have it for backwards compatibility. You are adding a new restriction
> > for a very specific action, so I don't think the fallback is needed.
>
> That means someone having CAP_SYS_ADMIN, but not CAP_KILL, can't perform
> the desired action. Is this what you'd like?

AFAIK, such user wouldn't be allowed to directly send a signal to a
different process either. So I think it makes more sense to be
consistent with the existing/main CAP_KILL usage rather than with the
CAP_PERFMON usage (which has its own reason to have that fallback).

I'm not the authority on capabilities nor the perf subsystem, it just
didn't seem quite right to me so I wanted to raise the concern.
Hopefully someone wiser than me will speak up if I talk nonsense :)

> If so, I'll just remove the wrapper, and call capable(CAP_KILL)
> directly.
>
> > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > index fe88d6eea3c2..1ab4bc867531 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -12152,10 +12152,21 @@ SYSCALL_DEFINE5(perf_event_open,
> > >         }
> > >
> > >         if (task) {
> > > +               bool is_capable;
> > > +
> > >                 err = down_read_interruptible(&task->signal->exec_update_lock);
> > >                 if (err)
> > >                         goto err_file;
> > >
> > > +               is_capable = perfmon_capable();
> > > +               if (attr.sigtrap) {
> > > +                       /*
> > > +                        * perf_event_attr::sigtrap sends signals to the other
> > > +                        * task. Require the current task to have CAP_KILL.
> > > +                        */
> > > +                       is_capable &= kill_capable();
> >
> > Is it necessary to do all this dance just to call perfmon_capable()
> > first? Couldn't this be simply:
> >
> > err = -EPERM;
> > if (attr.sigtrap && !capable(CAP_KILL))
> >         goto err_cred;
>
> Not so much about perfmon_capable() but about the ptrace_may_access()
> check. The condition here is supposed to be:
>
>         want CAP_PERFMON and (CAP_KILL if sigtrap)
>                 OR
>         want ptrace access (which includes a check for same thread-group and uid)
>
> If we did what you propose, then the ptrace check is effectively ignored
> if attr.sigtrap, and that's not what we want.
>
> There are lots of other ways of writing the same thing, but it should
> also remain readable and sticking it all into the same condition is not
> readable.

Ah, I see, I missed that semantic difference... So ptrace_may_access()
implies that the process doesn't need CAP_KILL to send a signal to the
task, that makes sense.

In that case I'm fine with this part as it is.

> > Also, looking at kill_ok_by_cred() in kernel/signal.c, would it
> > perhaps be more appropriate to do
> > ns_capable(__task_cred(task)->user_ns, CAP_KILL) instead? (There might
> > also need to be some careful locking around getting the target task's
> > creds - I'm not sure...)
>
> That might make sense. AFAIK, the locking is already in place via
> exec_update_lock. Let me investigate.
>
> > > +               }
> > > +
> > >                 /*
> > >                  * Preserve ptrace permission check for backwards compatibility.
> > >                  *
> > > @@ -12165,7 +12176,7 @@ SYSCALL_DEFINE5(perf_event_open,
> > >                  * perf_event_exit_task() that could imply).
> > >                  */
> > >                 err = -EACCES;
> >
> > BTW, shouldn't this (and several other such cases in this file...)
> > actually be EPERM, as is the norm for capability checks?
>
> I'm not a perf maintainer, so I can't give you a definitive answer.
> But, this would change the ABI, so I don't think it's realistic to
> request this change at this point unfortunately.

Indeed... I worry it will make troubleshooting SELinux/capability
errors more confusing, but I agree it would be a potentially risky
change to fix it :/

--
Ondrej Mosnacek
Software Engineer, Linux Security - SELinux kernel
Red Hat, Inc.