lkml.org 
[lkml]   [2011]   [Oct]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH 0/5] Signal scalability series
On 09/30, Matt Fleming wrote:
>
> On Fri, 2011-09-30 at 18:52 +0200, Oleg Nesterov wrote:
>
> > Hmm. Just out of curiosity, I blindly applied the whole series and poke
> > the _random_ function to look at, dequeue_signal(). And it looks wrong.
> >
> > spin_lock_irqsave(&current->signal->ctrl_lock, flags);
> > current->jobctl |= JOBCTL_STOP_DEQUEUED;
> >
> > This signal->ctrl_lock can't help. A sig_kernel_stop() should be
> > dequeued under the same lock, and we shouldn't release it unless we

s/unless/until/

> > set JOBCTL_STOP_DEQUEUED. Otherwise we race with SIGCONT.
>
> Hmm.. is that really a problem? Does the dequeue and setting
> JOBCTL_STOP_DEQUEUED actually need to be atomic?

It should be atomic wrt SIGCONT.

> Does it matter if we
> have SIGCONT on the signal queue when we set JOBCTL_STOP_DEQUEUED?

Why do we have? Usually SIGCONT is ignored. But this doesn't matter,
SIGCONT acts at the sending time.

If SIGCONT is sent - the process must not stop. Since we drop the lock
we can't guarantee this.

> > May be do_signal_stop() does something special? At first flance it doesn't.
> > But wait, it does while_each_thread() under ->ctrl_lock, why this is safe?
>
> Why is it not safe? What scenario are you thinking of where that isn't
> safe?

This series doesn't add ->ctrl_lock into copy_process/__unhash_process
or I misread the patches. This means we can't trust >thread_group list.

Even this is safe (say, we can rely on rcu), we can't calculate
->group_stop_count correctly. In particular, without ->siglock we can
race with exit_signals() which sets PF_EXITING. Note that PF_EXITING
check in task_set_jobctl_pending() is important.

Oleg.



\
 
 \ /
  Last update: 2011-10-03 15:23    [W:0.130 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site