Re: [PATCH 0/1] ns: introduce proc_get_ns_by_fd()

[ebiederm@xmission.com (Eric W. Biederman)]

Oleg Nesterov <oleg@redhat.com> writes:

> On 09/29, Eric W. Biederman wrote:
>>
>> Oleg Nesterov <oleg@redhat.com> writes:
>>
>> > OK, I won't insist, this too looks better to me than proc_ns_fdget(&fd_ref).
>> >
>> > And in any case fcheck_files() makes more sense than fdget(), somehow I did
>> > not think about this when I sent 1/1.
>> >
>> > Hmm. and after the quick look at cleanup_net() I can't understand whether
>> > get_net_ns_by_fd() can use ns_by_fd_rcu() + maybe_get_net(to_net_ns()) or
>> > not... Can it?
>>
>> Some of those places need a reference that allows them to sleep, and the
>> code is shared with the legacy pid case so with an addition of get_net
>> we can use ns_by_fd_rcu().   There are cases like setns that could
>> use ns_by_fd_rcu() with code reording.
>>
>> We can implement get_net_ns_by_fd as:
>> struct net *get_net_ns_by_fd(int fd)
>> {
>>         struct net *net;
>>
>> 	rcu_read_lock();
>> 	net = net_ns_by_fd_rcu(fd);
>>         if (!IS_ERR(net))
>>         	get_net(net);
>>         rcu_read_unlock();
>>
>> 	return net;
>> }
>>
>> Which means we can achieve code sharing with the pure rcu version
>> as a base.
>
> Yes, this is what I meant... but don't we need maybe_get_net() ?

Let me see.

The call chain is a little interesting so let's trace it.

thread1                               thread2
  net = net_ns_by_fd_rcu(fd);           fput(file)
  get_net(net);                           ____fput
                                            __fput
                                              dput
                                                dentry_kill
                                                  __dentry_kill
                                                    dentry_iput
                                                      iput
                                                        iput_final
                                                          evict
                                                            ->evict_inode (nsfs_evict)
                                                              netns_put
                                                                 put_net
                                                                   __put_net
                                                                      queue_work(net_cleanup_work)
                                                                      ...
                                                                      cleanup_net
                                                                        synchronize_rcu.


Given that someone in another thread can put the file descriptor and I
don't see anything explicitly waiting until after the rcu critical
section ends to put the network namespace I guess we do need
maybe_get_net.

In practice I think the delay logic in fput will act as an rcu barrier,
but that is not guaranteed so we had better not count on it.  Sigh.

Anyway this is a long ways from a syscall to translate pids.

Eric