Re: + audit-accounting-tty-locking.patch added to -mm tree

[Alan Cox <alan@redhat.com>]

On Wed, Sep 06, 2006 at 03:07:53AM +0400, Oleg Nesterov wrote:
> Nowadays ->signal/->sighand are _also_ protected by ->sighand->siglock.
> Unless you are current, you can't lock ->siglock directly (without holding
> tasklist_lock), you should use lock_task_sighand().

Gack, that makes current controlling tty locking horrible (and wrong almost
everywhere still across a clone)

> tty_io.c:
> 	->tty is set under task_lock()
> 
> 	->tty is cleared under lock_kernel() + tasklist_lock
> 
> 	except TIOCNOTTY, cleared under task_lock()
> 
> Note that include/linux/sched.h doesn't document that ->alloc_lock
> protects ->tty, it is only used in tty_io.c for that purpose, why?

Work in progress

> Btw, I think tiocsctty()/tty_open() is racy wrt to sys_setsid().
> tiocsctty() can see the result of '->signal->leader = 1' before
> sys_setsid() changed ->session/->pgrp and passed '->tty = NULL'.

Correct. I'm doing them bit by bit as I unpick them and check they
don't deadlock. If we need to take task_lock as well then its time
for set_controlling_tty() to get added.

Thanks for the signal lock explanation though. Now I've more idea wtf
is going on below the tty layer