Re: [PATCH 2/2] Ensures correct concurrent opening/reading of pidlists across pid namespaces

[Benjamin Blum <bblum@google.com>]

On Thu, Jul 2, 2009 at 4:54 PM, Andrew Morton<akpm@linux-foundation.org> wrote:
>> +static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
>> +                                               enum cgroup_filetype type)
>> +{
>> +     struct cgroup_pidlist *l;
>> +     /* don't need task_nsproxy() if we're looking at ourself */
>> +     struct pid_namespace *ns = get_pid_ns(current->nsproxy->pid_ns);
>> +     mutex_lock(&cgrp->pidlist_mutex);
>> +     list_for_each_entry(l, &cgrp->pidlists, links) {
>> +             if (l->key.type == type && l->key.ns == ns) {
>> +                     /* found a matching list - drop the extra refcount */
>> +                     put_pid_ns(ns);
>> +                     /* make sure l doesn't vanish out from under us */
>
> This looks fishy.
>
>> +                     down_write(&l->mutex);
>> +                     mutex_unlock(&cgrp->pidlist_mutex);
>> +                     l->use_count++;
>> +                     return l;
>
> The caller of cgroup_pidlist_find() must ensure that l->use_count > 0,
> otherwise cgroup_pidlist_find() cannot safely use `l' - it could be
> freed at any time.  But if l->use_count > 0, there is no risk of `l'
> "vanishing out from under us".
>
> I'm probably wrong there, but that's the usual pattern and this code
> looks like it's doing something different.  Please check?
>

That comment is vague, and should be rewritten. Individual pidlist
locks depend on the cgroup->pidlist_mutex; the main idea here is that
we can't drop the pidlist_mutex before picking up l->lock in case
somebody's trying to remove it from the list at the same time (compare
with cgroup_release_pid_array, the destroyer). The pid_namespace
refcount is also safe, because having found the existing list means
whoever put it there has a reference on the namespace in l->key, which
hasn't gone away yet and also is protected by the
cgroup->pidlist_mutex.

The only ordering that's important here is that incrementing
l->use_count and dropping cgroup->pidlist_mutex both have to come
after taking l->mutex.