Re: [PATCH] tracing: do not leak kernel addresses

[Nick Desaulniers]

On Thu, Jul 26, 2018 at 9:59 AM Nick Desaulniers
<ndesaulniers@xxxxxxxxxx> wrote:
>
> On Thu, Jul 26, 2018 at 9:37 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> >
> > On Thu, 26 Jul 2018 09:32:07 -0700
> > Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
> >
> > > On Thu, Jul 26, 2018 at 8:22 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, 26 Jul 2018 08:14:08 -0700
> > > > Mark Salyzyn <salyzyn@xxxxxxxxxxx> wrote:
> > > >
> > > > > Thank you Steve, much appreciated feedback, I have asked the security
> > > > > developers to keep this in mind and come up with a correct fix.
> > > > >
> > > > > The correct fix that meets your guidelines would _not_ be suitable for
> > > > > stable due to the invasiveness it sounds, only for the latest will such
> > > > > a rework make sense. As such, the fix proposed in this patch is the only
> > > > > one that meets the bar for stable patch simplicity, and merely(!) needs
> > > > > to state that if the fix is taken, perf and trace are broken.
> > > > >
> > > > > Posting this patch publicly on the lists, that may never be applied, may
> > > > > be the limit of our responsibility for a fix to stable kernel releases,
> > > > > to be optionally applied by vendors concerned with this CVE criteria?
> > > > >
> > > >
> > > > The patch breaks the code it touches. It makes it useless.
> > >
> > > Doesn't that depend on kptr_restrict, or would it be broken if
> > > kptr_restrict was set to 0?
> >
> > Is that what governs the output of kallsyms?
>
> From my workstation:
>
> $ cat /proc/kallsyms
>
> prints a bunch of zero'd out addresses, while
>
> $ sudo !!
>
> prints out actual addresses.  Looking at kernel/kallsyms.c, it seems
> that there's no use of %pK, but kallsyms_show_value() switches on
> kptr_restrict (and additional values):
>
> /*
>  * We show kallsyms information even to normal users if we've enabled
>  * kernel profiling and are explicitly not paranoid (so kptr_restrict
>  * is clear, and sysctl_perf_event_paranoid isn't set).
>  *
>  * Otherwise, require CAP_SYSLOG (assuming kptr_restrict isn't set to
>  * block even that).
>  */
> int kallsyms_show_value(void)
> {
>         switch (kptr_restrict) {
>         case 0:
>                 if (kallsyms_for_perf())
>                         return 1;
>         /* fallthrough */
>         case 1:
>                 if (has_capability_noaudit(current, CAP_SYSLOG))
>                         return 1;
>         /* fallthrough */
>         default:
>                 return 0;
>         }
> }

What are folks thoughts on this:
1. create function show_symbols_for_perf() that replaces
kallsyms_show_value(), maybe in linux/ftrace.c (since linux/ftrace.h
is included in kernel/trace/trace_printk.c and kernel/kallsyms.c).
2. use new show_symbols_for_perf() in kernel/kallsyms.c and
kernel/trace/trace_printk.c

Where the implementation of show_symbols_for_perf() is
kallsyms_show_value() implementation-wise (just renamed since it's no
longer kallsyms specific).  Does that make sense, or should I just
send a patch?  Does it make sense to check whether
kernel/trace/trace_printk.c#t_show() should print an address based on
the same checks done in kallsyms_show_value()?

-- 
Thanks,
~Nick Desaulniers