Re: [PATCH] audit: grab a reference to context->pwd when it's cached

[Jeff Layton <jlayton@redhat.com>]

On Fri, 5 Oct 2012 06:57:59 -0700
Peter Moody <pmoody@google.com> wrote:

> On Fri, Oct 5, 2012 at 5:55 AM, Jeff Layton <jlayton@redhat.com> wrote:
> > On Thu, 4 Oct 2012 11:48:23 -0700
> > Peter Moody <pmoody@google.com> wrote:
> >
> >> On Wed, Sep 26, 2012 at 6:50 AM, Alexander Viro <aviro@redhat.com> wrote:
> >> > On Tue, Sep 25, 2012 at 10:03:23AM -0700, Peter Moody wrote:
> >> >> Hey folks,
> >> >>
> >> >> following up on old patches, are there any comments on this? Did you
> >> >> get around to finding a better way to fix this bug, Al?
> >> >
> >> > Alas, I've found none ;-/  Looks like we'll have to go with this one,
> >> > at least until somebody comes up with better solution.
> >>
> >> Not surprisingly, this patch doesn't actually fix the issue (or at
> >> least doesn't do it correctly).
> >>
> >> I hadn't noticed that get_fs_pwd() actually calls path_get() on
> >> &context->pwd so the additional path_get() is useless and the
> >> reference doesn't ever actually get freed if audit_putname is called
> >> while we're in a syscall.
> >>
> >> Al, Eric, Jeff; do any of you guys have an understanding of what the
> >> initial bug actually is since this clearly doesn't fix it?
> >>
> >> Cheers,
> >> peter
> >>
> >
> > BTW, I ran this test on one of my KVM guests and it ran just fine. That
> > one is an x86_64 guest running a 3.6.0+ kernel. The root fs on there is
> > ext4 though, not ext3. So perhaps that's a factor?
> >
> > The oops message you posted at least looks like something down in the
> > bowels of ext3 or fs/buffer.c.
> 
> Yeah, the only place this actually happens for me on these giant xen
> instances we have (6 cores, 32G ram) and it happens on both ext3 and
> ext4 filesystems and it happens with 100% reliability.
> 
> The actual oops is from:
> 
> static inline void check_irqs_on(void)
> {
> #ifdef irqs_disabled
>         BUG_ON(irqs_disabled());
> #endif
> }
> 
> with the code path looking like:
> 
> __find_get_block() -> lookup_bh_lru() -> check_irqs_on() -> BUG()
> 

Do you have a backtrace from a more recent kernel? I wonder if
something in the syscall exit codepath is disabling IRQs here?
-- 
Jeff Layton <jlayton@redhat.com>