Kirill Tkhai <ktk...@virtuozzo.com> writes:
> On 14.11.2017 21:39, Cong Wang wrote:
>> On Tue, Nov 14, 2017 at 5:53 AM, Kirill Tkhai <ktk...@virtuozzo.com> wrote:
>>> @@ -406,7 +406,7 @@ struct net *copy_net_ns(unsigned long flags,
>>>
>>> get_user_ns(user_ns);
>>>
>>> - rv = mutex_lock_killable(&net_mutex);
>>> + rv = down_read_killable(&net_sem);
>>> if (rv < 0) {
>>> net_free(net);
>>> dec_net_namespaces(ucounts);
>>> @@ -421,7 +421,7 @@ struct net *copy_net_ns(unsigned long flags,
>>> list_add_tail_rcu(&net->list, &net_namespace_list);
>>> rtnl_unlock();
>>> }
>>> - mutex_unlock(&net_mutex);
>>> + up_read(&net_sem);
>>> if (rv < 0) {
>>> dec_net_namespaces(ucounts);
>>> put_user_ns(user_ns);
>>> @@ -446,7 +446,7 @@ static void cleanup_net(struct work_struct *work)
>>> list_replace_init(&cleanup_list, &net_kill_list);
>>> spin_unlock_irq(&cleanup_list_lock);
>>>
>>> - mutex_lock(&net_mutex);
>>> + down_read(&net_sem);
>>>
>>> /* Don't let anyone else find us. */
>>> rtnl_lock();
>>> @@ -486,7 +486,7 @@ static void cleanup_net(struct work_struct *work)
>>> list_for_each_entry_reverse(ops, &pernet_list, list)
>>> ops_free_list(ops, &net_exit_list);
>>>
>>> - mutex_unlock(&net_mutex);
>>> + up_read(&net_sem);
>>
>> After your patch setup_net() could run concurrently with cleanup_net(),
>> given that ops_exit_list() is called on error path of setup_net() too,
>> it means ops->exit() now could run concurrently if it doesn't have its
>> own lock. Not sure if this breaks any existing user.
>
> Yes, there will be possible concurrent ops->init() for a net namespace,
> and ops->exit() for another one. I hadn't found pernet operations, which
> have a problem with that. If they exist, they are hidden and not clear seen.
> The pernet operations in general do not touch someone else's memory.
> If suddenly there is one, KASAN should show it after a while.
Certainly the use of hash tables shared between multiple network
namespaces would count. I don't rembmer how many of these we have but
there used to be quite a few.
Eric
