Re: [RFC PATCH 6/6] security/fbfam: Mitigate a fork brute force attack

[Jann Horn <jannh@google.com>]

On Thu, Sep 10, 2020 at 10:22 PM Kees Cook <keescook@chromium.org> wrote:
> In order to mitigate a fork brute force attack it is necessary to kill
> all the offending tasks. This tasks are all the ones that share the
> statistical data with the current task (the task that has crashed).
>
> Since the attack detection is done in the function fbfam_handle_attack()
> that is called every time a core dump is triggered, only is needed to
> kill the others tasks that share the same statistical data, not the
> current one as this is in the path to be killed.
>
> When the SIGKILL signal is sent to the offending tasks from the function
> fbfam_kill_tasks(), this one will be called again during the core dump
> due to the shared statistical data shows a quickly crashing rate. So, to
> avoid kill again the same tasks due to a recursive call of this
> function, it is necessary to disable the attack detection.
>
> To disable this attack detection, add a condition in the function
> fbfam_handle_attack() to not compute the crashing rate when the jiffies
> stored in the statistical data are set to zero.
[...]
>  /**
> - * fbfam_handle_attack() - Fork brute force attack detection.
> + * fbfam_kill_tasks() - Kill the offending tasks
> + *
> + * When a fork brute force attack is detected it is necessary to kill all the
> + * offending tasks. Since this function is called from fbfam_handle_attack(),
> + * and so, every time a core dump is triggered, only is needed to kill the
> + * others tasks that share the same statistical data, not the current one as
> + * this is in the path to be killed.
> + *
> + * When the SIGKILL signal is sent to the offending tasks, this function will be
> + * called again during the core dump due to the shared statistical data shows a
> + * quickly crashing rate. So, to avoid kill again the same tasks due to a
> + * recursive call of this function, it is necessary to disable the attack
> + * detection setting the jiffies to zero.
> + *
> + * To improve the for_each_process loop it is possible to end it when all the
> + * tasks that shared the same statistics are found.

This is not a fastpath, there's no need to be clever and optimize
things here, please get rid of that optimization. Especially since
that fastpath looks racy against concurrent execve().

> + * Return: -EFAULT if the current task doesn't have statistical data. Zero
> + *         otherwise.
> + */
> +static int fbfam_kill_tasks(void)
> +{
> +       struct fbfam_stats *stats = current->fbfam_stats;
> +       struct task_struct *p;
> +       unsigned int to_kill, killed = 0;
> +
> +       if (!stats)
> +               return -EFAULT;
> +
> +       to_kill = refcount_read(&stats->refc) - 1;
> +       if (!to_kill)
> +               return 0;
> +
> +       /* Disable the attack detection */
> +       stats->jiffies = 0;
> +       rcu_read_lock();
> +
> +       for_each_process(p) {
> +               if (p == current || p->fbfam_stats != stats)

p->fbfam_stats could change concurrently, you should at least use
READ_ONCE() here.

Also, if this codepath is hit by a non-leader thread, "p == current"
will always be false, and you'll end up killing the caller, too. You
may want to compare with current->group_leader instead.


> +                       continue;
> +
> +               do_send_sig_info(SIGKILL, SEND_SIG_PRIV, p, PIDTYPE_PID);
> +               pr_warn("fbfam: Offending process with PID %d killed\n",
> +                       p->pid);

Normally pr_*() messages about tasks mention not just the pid, but
also the ->comm name of the task.

> +               killed += 1;
> +               if (killed >= to_kill)
> +                       break;
> +       }
> +
> +       rcu_read_unlock();
> +       return 0;
> +}