Re: [PATCH 2/2] perf: SNB exclusive PMU access for INST_RETIRED:PREC_DIST

[Ingo Molnar <mingo@kernel.org>]

* Stephane Eranian <eranian@google.com> wrote:

> On Sun, Oct 21, 2012 at 8:03 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > * Stephane Eranian <eranian@google.com> wrote:
> >
> >> On Sun, Oct 21, 2012 at 6:55 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >> >
> >> > * Andi Kleen <ak@linux.intel.com> wrote:
> >> >
> >> >> > > > This isn't limited to admin, right? So the above turns into a DoS on the
> >> >> > > > console.
> >> >> > > >
> >> >> > > Ok, so how about a WARN_ON_ONCE() instead?
> >> >> >
> >> >> > That should be fine I guess ;-)
> >> >>
> >> >> imho there is need for a generic mechanism to return an error
> >> >> string to the user program without such hacks.
> >> >
> >> > Agreed - we could return an 'extended errno' long error return
> >> > value, which in reality is a pointer to an error string (in
> >> > perf_attr::error_str), and copy that string to user-space at
> >> > perf syscall return time.
> >> >
> >> I assume by perf_attr:error_str, you actually mean:
> >>
> >> struct perf_event_attr {
> >>    char error_str[PERF_ERR_LEN];
> >> };
> >>
> >> Right?
> >
> > I don't think we should allocate space in the attr, instead we
> > should use something like:
> >
> >         u8 __user       *err_str;
> >         u32              err_str_len;
> >
> > which would be filled in by tooling with a string and a max_len
> > value, and strncpy_to_user() could do the rest on the kernel
> > side. [ A minor complication is that we don't have a
> > strncpy_to_user() method at the moment. ]
> >
> > Static strings could be handled this way.
> >
> > [ Dynamic strings could be supported too with a few tricks,
> >   although I doubt it matters in practice. ]
> >
>
> Ok, but this still limits returning error string to the 
> perf_event_open() syscall, not read(), ioctl() and such.

Yes - but this should be enough to handle most of the cases in 
practice - because the richness of the various perf components 
is mostly exposed via the perf syscall. By the time we get to 
read() and ioctl() we are in a pretty well defined domain.

Also, I don't think people want the (small but nonzero) overhead 
of extended error reporting for read or ioctl.

> I am fine with this change. However, I think it should be 
> added separately from my inst_retired:prec_dist patch. It has 
> a broader impact.

Most definitely.

Thanks,

	Ingo