Re: [PATCH] do_task_stat: don't use task_pid_nr_ns() lockless

[Pavel Emelyanov <xemul@openvz.org>]

Eric W. Biederman wrote:
> Oleg Nesterov <oleg@tv-sign.ru> writes:
> 
>> Without rcu/tasklist/siglock lock task_pid_nr_ns() may read the freed memory,
>> move the callsite under ->siglock.
>>
>> Sadly, we can report pid == 0 if the task was detached.
> 
> We only get detached in release_task so it is a pretty small window
> where we can return pid == 0.  Usually get_task_pid will fail first
> and we will return -ESRCH.  Still the distance from open to 
> 
> There is another bug in here as well.  current->nsproxy->pid_ns is wrong.
> What we want is: ns = dentry->d_sb->s_fs_info;

Actually I thought about this recently - if we produce the list
of tasks based on the sb's namespace, then we should fill the
tasks' files according to the sb's namespace as well, not according
to the current namespace.

> Otherwise we will have file descriptor passing races and the like.

Can you elaborate?

> We could also do: proc_pid(inode) to get the pid, which is a little
> more race free, and will prevent us from returning pid == 0.
> 
> In either event it looks like we need to implement some proper
> file operations for these proc files, maybe even going to seq file
> status.
> 
> Eric
> 

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/