Re: [PATCH] pidhash: Kill switch_exec_pids

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

Oleg Nesterov <oleg@tv-sign.ru> writes:

> Eric W. Biederman wrote:
>>
>> --- a/fs/exec.c
>> +++ b/fs/exec.c
>> @@ -699,7 +699,17 @@ static int de_thread(struct task_struct 
>>  		remove_parent(current);
>>  		remove_parent(leader);
>>  
>> -		switch_exec_pids(leader, current);
>> +
>> +		/* Become a process group leader with the old leader's pid.
>> +		 * Note: The old leader also uses thispid until release_task
>> +		 *       is called.  Odd but simple and correct.
>> +		 */
>> +		detach_pid(current, PIDTYPE_PID);
>> +		current->pid = leader->pid;
>> +		attach_pid(current, PIDTYPE_PID,  current->pid);
>
> What happens after de_thread() unlocks tasklist_lock and before
> it is taken again in release_task() ?
>
> In that window find_task_by_pid() will return dead leader, not
> the new leader of thread group. This means we can miss tkill()
> or ptrace(), for example.

A reasonable concern.  For some reason I had it in my
head that find_pid didn't need that tasklist_lock
but it does.  kthread and vmscan need to be fixed.

All I have done is enlarged the window where this
race is possible.  So for tkill I am not concerned,
as it wants a particular thread.  Nor am I concerned
about anything else that wants a particular thread.

The fact that the group_leader does not point
at the actual thread group leader might be a problem,
as I have opened a window where that is now the case.

For signals that is not a problem as signals are still shared.
This applies to most other resources as well.

Looking through all of the callers of find_task_by_pid
I couldn't see anything that looks like it cares.

So until we spot that case I'm ready to put this down
of one of those cases in de_thread that looks wrong
but happens to work.  Now if there is a way to make
it work more cleanly that may be worth looking at.


Eric