Re: /proc/timer_list leaks the real pids of the associated processes

From: Jessica Frazelle
Date: Tue Feb 07 2017 - 17:57:26 EST


So I used to use this "feature" as a hack to see if something was
running in a k8s cluster or not because fluentd gives a lot of things
away. Runc basically decided to mask this file so it stopped happening
though. But I think namespace aware is the right decision.


On Tue, Feb 7, 2017 at 2:48 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> On Fri, 3 Feb 2017, Kees Cook wrote:
>
> > On Fri, Feb 3, 2017 at 2:29 PM, Xing Gao <xgao01@xxxxxxxxxxxx> wrote:
> > > Dear Thomas and Kees,
> > >
> > > I posted a bug report on bugzilla, and John asked me to send it the lkml.
> > >
> > > Here is the link, https://bugzilla.kernel.org/show_bug.cgi?id=193921
> > >
> > > Please cc to me when you reply this email.
> > >
> > > And please check the information below.
> > >
> > > The pseudo file /proc/timer_list leaks the real pids of the associated
> > > processes.
> > >
> > > The function print_timer(kernel/time/timer_list.c) displays
> > > timer->start_pid, which is set inside the function
> > > __timer_stats_timer_set_start_info (kernel/time/timer.c). This is the real
> > > pid, rather than the pid in the pid namespace. If the user within a
> > > container retrieves the content of /proc/timer_list, this file will leak the
> > > real pid of the associated process.
> >
> > I feel like this has been pointed out before, but I can't find the
> > email about it. Regardless, yeah, this looks true:
> >
> > SEQ_printf(m, ", %s/%d", tmp, timer->start_pid);
> >
> > #11: <0000000000000000>, hrtimer_wakeup, S:01, do_nanosleep, cron/2570
> >
> > Seems like this should be made namespace aware... (and why is this
> > file needed at all? Seems like it should live in debugfs not proc).
>
> I'm fine with that.
>
> The TIMER_STATS stuff should go away completely. If people are interested
> in that information they can use the tracer which gives way better
> data than that file.
>
> Thanks,
>
> tglx
> .