On Tue, Jun 24, 2014 at 11:30 AM, Oleg Nesterov [off-list ref] wrote:

I am puzzled by the usage of smp_load_acquire(),

It was recommended by Andy Lutomirski in preference to ACCESS_ONCE().

On 06/23, Kees Cook wrote:

quoted

 static u32 seccomp_run_filters(int syscall)
 {
-     struct seccomp_filter *f;
+     struct seccomp_filter *f = smp_load_acquire(&current->seccomp.filter);
      struct seccomp_data sd;
      u32 ret = SECCOMP_RET_ALLOW;

      /* Ensure unexpected behavior doesn't result in failing open. */
-     if (WARN_ON(current->seccomp.filter == NULL))
+     if (WARN_ON(f == NULL))
              return SECCOMP_RET_KILL;

      populate_seccomp_data(&sd);

@@ -186,9 +186,8 @@ static u32 seccomp_run_filters(int syscall)
       * All filters in the list are evaluated and the lowest BPF return
       * value always takes priority (ignoring the DATA).
       */
-     for (f = current->seccomp.filter; f; f = f->prev) {
+     for (; f; f = smp_load_acquire(&f->prev)) {
              u32 cur_ret = SK_RUN_FILTER(f->prog, (void *)&sd);
-
              if ((cur_ret & SECCOMP_RET_ACTION) < (ret & SECCOMP_RET_ACTION))
                      ret = cur_ret;

OK, in this case the 1st one is probably fine, altgough it is not
clear to me why it is better than read_barrier_depends().

But why do we need a 2nd one inside the loop? And if we actually need
it (I don't think so) then why it is safe to use f->prog without
load_acquire ?

You're right -- it should not be possible for for any of the ->prev
pointers to change.

quoted

 void get_seccomp_filter(struct task_struct *tsk)
 {
-     struct seccomp_filter *orig = tsk->seccomp.filter;
+     struct seccomp_filter *orig = smp_load_acquire(&tsk->seccomp.filter);
      if (!orig)
              return;

This one looks unneeded.

First of all, afaics atomic_inc() should work correctly without any barriers,
otherwise it is buggy. But even this doesn't matter.

With this changes get_seccomp_filter() must be called under ->siglock, it can't
race with add-filter and thus tsk->seccomp.filter should be stable.

Excellent point, yes. I'll remove that.

quoted

      /* Reference count is bounded by the number of total processes. */

@@ -361,7 +364,7 @@ void put_seccomp_filter(struct task_struct *tsk)
      /* Clean up single-reference branches iteratively. */
      while (orig && atomic_dec_and_test(&orig->usage)) {
              struct seccomp_filter *freeme = orig;
-             orig = orig->prev;
+             orig = smp_load_acquire(&orig->prev);
              seccomp_filter_free(freeme);
      }

This one looks unneeded too. And note that this patch does not add
smp_load_acquire() to read tsk->seccomp.filter.

Hrm, yes, that should get added.

atomic_dec_and_test() adds mb(), we do not need more barriers to access
->prev ?

Right, same situation as the run_filters loop. Thanks!

-Kees

-- 
Kees Cook
Chrome OS Security