lkml.org 
[lkml]   [2014]   [Jun]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [BUG] signal: sighand unprotected when accessed by /proc
On Tue, Jun 03, 2014 at 10:01:25PM +0200, Oleg Nesterov wrote:
> On 06/03, Linus Torvalds wrote:
> >
> > On Tue, Jun 3, 2014 at 10:26 AM, Oleg Nesterov <oleg@redhat.com> wrote:
> > >
> > > looks like, SLAB_DESTROY_BY_RCU logic is broken?
> >
> > I haven't looked at the code yet, but SLAB_DESTROY_BY_RCU can be
> > subtle and very dangerous.
> >
> > The danger is that the *slab* itself is free'd by RCU, but individual
> > allocations can (and do) get re-used FOR THE SAME OBJECT TYPE without
> > waiting for RCU!
> >
> > This is subtle. It means that most people who think that "it's free'd
> > by RCU" get it wrong. Because individual allocations really aren't at
> > all RCU-free'd, it's just that the underlying memory is guaranteed to
> > not change type or be entirely thrown away until after a RCU grace
> > period.
>
> Yes, exactly. And unless you use current->sighand (which is obviously
> stable) you need lock_task_sighand() which relies on ->siglock initialized
> by sighand_ctor().
>
> > Without looking at the code, it sounds like somebody may doing things
> > to "sighand->lock->wait_list" that they shouldn't do. We've had cases
> > like that before, and most of them have been changed to *not* use
> > SLAB_DESTROY_BY_RCU, and instead make each individual allocation be
> > RCU-free'd (which is a lot simpler to think about, because then you
> > don't have the whole re-use issue).
>
> Sure, we only need to change __cleanup_sighand() to use call_rcu().
> But I am not sure this makes sense, I mean, I do not think this can
> make something more simple/clear.
>
> > And this could easily be an RT issue, if the RT code does some
> > re-initialization of the rtmutex that replaces the spinlock we have.
>
> Unlikely... this should be done by sighand_ctor() anyway.
>
> I'll try to recheck rt_mutex_unlock() tomorrow. _Perhaps_ rcu_read_unlock()
> should be shifted from lock_task_sighand() to unlock_task_sighand() to
> ensure that rt_mutex_unlock() does nothihg with this memory after it
> makes another lock/unlock possible.
>
> But if we need this (currently I do not think so), this doesn't depend on
> SLAB_DESTROY_BY_RCU. And, at first glance, in this case rcu_read_unlock_special()
> might be wrong too.

OK, I will bite... What did I mess up in rcu_read_unlock_special()?

This function does not report leaving the RCU read-side critical section
until after its call to rt_mutex_unlock() has returned, so any RCU
read-side critical sections in rt_mutex_unlock() will be respected.

Thanx, Paul



\
 
 \ /
  Last update: 2014-06-06 22:41    [W:0.148 / U:2.172 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site