Re: [PATCH v11 06/12] seccomp: add system call filtering using BPF

On Mon, Feb 27, 2012 at 11:09 AM, Oleg Nesterov [off-list ref] wrote:

Hello Will.

I missed the previous discussions, and I don't think I can read
all these emails now. So I apologize in advance if this was already
discussed.

No worries - any review is appreciated :)

On 02/24, Will Drewry wrote:

quoted

 struct seccomp {
      int mode;
+     struct seccomp_filter *filter;
 };

Minor nit, it seems that the new member can be "ifdef CONFIG_SECCOMP_FILTER"

Good call - I'll add that.

quoted

+static long seccomp_attach_filter(struct sock_fprog *fprog)
+{
+     struct seccomp_filter *filter;
+     unsigned long fp_size = fprog->len * sizeof(struct sock_filter);
+     long ret;
+
+     if (fprog->len == 0 || fprog->len > BPF_MAXINSNS)
+             return -EINVAL;

OK, this limits the memory PR_SET_SECCOMP can use.

But,

quoted

+     /*
+      * If there is an existing filter, make it the prev and don't drop its
+      * task reference.
+      */
+     filter->prev = current->seccomp.filter;
+     current->seccomp.filter = filter;
+     return 0;

this doesn't limit the number of filters, looks like a DoS.

What if the application simply does prctl(PR_SET_SECCOMP, dummy_filter)
in an endless loop?

It consumes a massive amount of kernel memory and, maybe, the OOM
killer gives it a boot :)

I wasn't sure what the normal convention was for avoiding memory
consumption by user processes. Should I just add a sysctl and a
per-task counter for the max number of filters?

I'm fine doing whatever makes sense here.

quoted

+static struct seccomp_filter *get_seccomp_filter(struct seccomp_filter *orig)
+{
+     if (!orig)
+             return NULL;
+     /* Reference count is bounded by the number of total processes. */
+     atomic_inc(&orig->usage);
+     return orig;
+}
...
+void copy_seccomp(struct seccomp *child, const struct seccomp *parent)
+{
+     /* Other fields are handled by dup_task_struct. */
+     child->filter = get_seccomp_filter(parent->filter);
+}

This is purely cosmetic, but imho looks a bit confusing.

We do not copy seccomp->mode and this is correct, it was already copied
implicitely. So why do we copy ->filter? This is not "symmetrical", afaics
you can simply do

       void copy_seccomp(struct seccomp *child)
       {
               if (child->filter)
                       atomic_inc(child->filter->usage);

But once again, this is cosmetic, feel free to ignore.

Right now get_seccomp_filter does the NULL check, so really this could
be reduced to adding an external get_seccomp_filter(p->seccomp.filter)
in place of copy_seccomp().

As to removing the extra arg, that should be fine since the parent
can't drop its refcount when copy_seccomp is called.  At the very
least, I can make that change so it reads more cleanly.

thanks!
will