RE: [PATCH v3 2/4] perf: Use CAP_SYS_ADMIN with perf_event_paranoid checks

["Lubashev, Igor" <ilubashe@akamai.com>]

On Mon, August 12, 2019 at 4:16 PM Arnaldo Carvalho de Melo <arnaldo.melo@gmail.com> wrote:
> Em Mon, Aug 12, 2019 at 05:01:34PM -0300, Arnaldo Carvalho de Melo
> escreveu:
> > Em Wed, Aug 07, 2019 at 10:44:15AM -0400, Igor Lubashev escreveu:
> > > +++ b/tools/perf/util/evsel.c
> > > @@ -279,7 +279,7 @@ struct evsel *perf_evsel__new_idx(struct
> > > perf_event_attr *attr, int idx)
> >
> > >  static bool perf_event_can_profile_kernel(void)
> > >  {
> > > -	return geteuid() == 0 || perf_event_paranoid() == -1;
> > > +	return perf_event_paranoid_check(-1);
> > >  }
> >
> > While looking at your changes I think the pre-existing code is wrong,
> > i.e. the check in sys_perf_event_open(), in the kernel is:
> >
> >         if (!attr.exclude_kernel) {
> >                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
> >                         return -EACCES;
> >         }
> >
> > And:
> >
> > static inline bool perf_paranoid_kernel(void) {
> >         return sysctl_perf_event_paranoid > 1; }
> >
> > So we have to change that perf_event_paranoit_check(-1) to pass 1
> > instead?

Indeed.  This seems right.  It was a pre-existing problem.


> > bool perf_event_paranoid_check(int max_level) {
> >         return perf_cap__capable(CAP_SYS_ADMIN) ||
> >                         perf_event_paranoid() <= max_level; }
> >
> > Also you defined perf_cap__capable(anything) as:
> >
> > #ifdef HAVE_LIBCAP_SUPPORT
> >
> > #include <sys/capability.h>
> >
> > bool perf_cap__capable(cap_value_t cap);
> >
> > #else
> >
> > static inline bool perf_cap__capable(int cap __maybe_unused)
> > {
> >         return false;
> > }
> >
> > #endif /* HAVE_LIBCAP_SUPPORT */
> >
> >
> > I think we should have:
> >
> > #else
> >
> > static inline bool perf_cap__capable(int cap __maybe_unused) {
> >         return geteuid() == 0;
> > }
> >
> > #endif /* HAVE_LIBCAP_SUPPORT */
> >
> > Right?

You can have EUID==0 and not have CAP_SYS_ADMIN, though this would be rare in practice.  I did not to use EUID in leu of libcap, since kernel does not do so, and therefore it seemed a bit misleading.  But this is a slight matter of taste, and I do not see a problem with choosing to fall back to EUID -- the kernel will do the right thing anyway.

Now, if I were pedantic, I'd say that to use geteuid(), you need to #include <unistd.h> .


> > So I am removing the introduction of perf_cap__capable() from the
> > first patch you sent, leaving it with _only_ the feature detection
> > part, using that feature detection to do anything is then moved to a
> > separate patch, after we finish this discussion about what we should
> > fallback to when libcap-devel isn't available, i.e. we should use the
> > previous checks, etc.
> 
> So, please take a look at the tmp.perf/cap branch in my git repo:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/log/?h=tmp.p
> erf/cap
> 
> I split the patch and made perf_cap__capable() fallback to 'return
> geteuid() == 0;' when libcap-devel isn't available, i.e. keep the checks made
> prior to your patchset.

Thank you.  And thanks for updating "make_minimal". 


> 
> Jiri, can I keep your Acked-by?
> 
> - Arnaldo