On Wed, 2020-06-24 at 08:18 -0700, Kees Cook wrote:
> On Wed, Jun 24, 2020 at 12:21:16PM +0200, Jann Horn wrote:
> > On Tue, Jun 23, 2020 at 7:26 PM Kristen Carlson Accardi
> > <kris...@linux.intel.com> wrote:
> > > This patch makes /proc/kallsyms display alphabetically by symbol
> > > name rather than sorted by address in order to hide the newly
> > > randomized address layout.
> > [...]
> > > +static int sorted_show(struct seq_file *m, void *p)
> > > +{
> > > + struct list_head *list = m->private;
> > > + struct kallsyms_iter_list *iter;
> > > + int rc;
> > > +
> > > + if (list_empty(list))
> > > + return 0;
> > > +
> > > + iter = list_first_entry(list, struct kallsyms_iter_list,
> > > next);
> > > +
> > > + m->private = iter;
> > > + rc = s_show(m, p);
> > > + m->private = list;
> > > +
> > > + list_del(&iter->next);
> > > + kfree(iter);
> >
> > Does anything like this kfree() happen if someone only reads the
> > start
> > of kallsyms and then closes the file? IOW, does "while true; do
> > head
> > -n1 /proc/kallsyms; done" leak memory?
>
> Oop, nice catch. It seems the list would need to be walked on s_stop.
>
> > > + return rc;
> > > +}
> > [...]
> > > +static int kallsyms_list_cmp(void *priv, struct list_head *a,
> > > + struct list_head *b)
> > > +{
> > > + struct kallsyms_iter_list *iter_a, *iter_b;
> > > +
> > > + iter_a = list_entry(a, struct kallsyms_iter_list, next);
> > > + iter_b = list_entry(b, struct kallsyms_iter_list, next);
> > > +
> > > + return strcmp(iter_a->iter.name, iter_b->iter.name);
> > > +}
> >
> > This sorts only by name, but kallsyms prints more information
> > (module
> > names and types). This means that if there are elements whose names
> > are the same, but whose module names or types are different, then
> > some
> > amount of information will still be leaked by the ordering of
> > elements
> > with the same name. In practice, since list_sort() is stable, this
> > means you can see the ordering of many modules, and you can see the
> > ordering of symbols with same name but different visibility (e.g.
> > "t
> > user_read" from security/selinux/ss/policydb.c vs "T user_read"
> > from
> > security/keys/user_defined.c, and a couple other similar cases).
>
> i.e. sub-sort by visibility?
>
> > [...]
> > > +#if defined(CONFIG_FG_KASLR)
> > > +/*
> > > + * When fine grained kaslr is enabled, we need to
> > > + * print out the symbols sorted by name rather than by
> > > + * by address, because this reveals the randomization order.
> > > + */
> > > +static int kallsyms_open(struct inode *inode, struct file *file)
> > > +{
> > > + int ret;
> > > + struct list_head *list;
> > > +
> > > + list = __seq_open_private(file, &kallsyms_sorted_op,
> > > sizeof(*list));
> > > + if (!list)
> > > + return -ENOMEM;
> > > +
> > > + INIT_LIST_HEAD(list);
> > > +
> > > + ret = kallsyms_on_each_symbol(get_all_symbol_name, list);
> > > + if (ret != 0)
> > > + return ret;
> > > +
> > > + list_sort(NULL, list, kallsyms_list_cmp);
> >
> > This permits running an algorithm (essentially mergesort) with
> > secret-dependent branches and memory addresses on essentially
> > secret
> > data, triggerable and arbitrarily repeatable (although with partly
> > different addresses on each run) by the attacker, and probably a
> > fairly low throughput (comparisons go through indirect function
> > calls,
> > which are slowed down by retpolines, and linked list iteration
> > implies
> > slow pointer chases). Those are fairly favorable conditions for
> > typical side-channel attacks.
> >
> > Do you have estimates of how hard it would be to leverage such side
> > channels to recover function ordering (both on old hardware that
> > only
> > has microcode fixes for Spectre and such, and on newer hardware
> > with
> > enhanced IBRS and such)?
>
> I wonder, instead, if sorting should be just done once per module
> load/unload? That would make the performance and memory management
> easier too.
>
My first solution (just don't show kallsyms at all for non-root) is
looking better and better :). But seriously, I will rewrite this one
with something like this, but I think we are going to need another
close look to see if sidechannel issues still exist with the new
implementation. Hopefully Jann can take another look then.
