Re: [PATCH] do_signal_stop: use signal_group_exit()

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

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

> On 02/16, Oleg Nesterov wrote:
>>
>> On 02/15, Andrew Morton wrote:
>> >
>> > ug.  On about the fourth boot with the current -mm lineup I hit:
>> >
>> > : BUG: unable to handle kernel paging request at 0000000000200200
>>                                                    ^^^^^^^^^^^^^^^^
>> == LIST_POISON2
>> 
>> > : IP: [<ffffffff802444f5>] free_pid+0x35/0x8e
>> 
>> most probably == hlist_del_rcu(pid_chain)
>> 
>> > : Call Trace:
>> > :  [<ffffffff80237727>] ? release_task+0x152/0x2e5
>> > :  [<ffffffff80237f81>] ? do_wait+0x6c7/0xa1c
>> > :  [<ffffffff8022f4cc>] ? default_wake_function+0x0/0xe
>> > :  [<ffffffff8023e670>] ? sys_rt_sigaction+0x7a/0x98
>> > :  [<ffffffff80238360>] ? sys_wait4+0x8a/0xa1
>> > :  [<ffffffff8020be4b>] ? system_call_after_swapgs+0x7b/0x80
>> 
>> (Can't understand why there is no detach_pid() in this stack trace,
>>  but it is the only possible caller of free_pid()).

That is just because of tail call optimization.  There is no need
to return to detach_pid so we just jumped to free_pid.

>
>> So, detach_pid()->free_pid() hit an already unhashed pid. But this
>> is not possible?
>> 
>> This means we already did detach_pid(), but in that case the previous
>> detach_pid() has set task->pids[].pid = NULL, and we should OOPS earlier,
>> somewhere at "if (!hlist_empty(&pid->tasks[tmp]))".
>
> Yes, this is not possible.
>
> But what possible is: we have the unbalanced put_pid(pid) which frees the
> live pid. The next alloc_pid() gets the same memory, and initializes all
> pid->tasks[] lists.
>
> Now, if that pid was used as PIDTYPE_PGID/PIDTYPE_SID, the next detach_pid()
> from this pgrp/sid sees that the pid is not used (all lists are hlist_empty),
> frees this pid again, and the bug manifests itself this way.
>
> 	tiocspgrp:
>
> 		put_pid(real_tty->pgrp);
> 		// ------ WINDOW ------
> 		real_tty->pgrp = get_pid(pgrp);
>
> When bash spawns the command, both parent and child do ioctl(TIOCSPGRP,child),
> and it is possible that both do put_pid() on the same parent's pid.
>
> Damn, when you know what the bug is, the test case is trivial:
>
> 	$ while true; do perl -e0; done
>
> The kernel crashes.
>
>>From 2.6.25-rc2-mm1.bz2 patch:
>>
>> -       .ioctl          = tty_ioctl,
>> +       .unlocked_ioctl = tty_ioctl,
>
> and this is why this didn't happen before, I guess.
>
>> I'll try to think more about this, but I doubt very much I'll find the
>> reason :(
>
> Ohhh... 7 (!!!) hours of hacking + some vodka did the trick.
>
> (Kamalesh, I think you hit the same bug).

Thanks.  Looks like we need to grab a lock there.
At a quick skim I think we need the tty lock.


Eric