Re: 2.6.25-rc8-mm1 - BUG in fs/jbd/transaction.c'

[Josef Bacik <jbacik@redhat.com>]

On Fri, Apr 04, 2008 at 12:15:09PM +0200, Jan Kara wrote:
> > 
> > On Wed, 2008-04-02 at 12:39 -0700, Andrew Morton wrote:
> > > On Wed, 2 Apr 2008 15:27:15 -0400
> > > Josef Bacik <jbacik@redhat.com> wrote:
> > > 
> > > > On Wed, Apr 02, 2008 at 03:12:49PM -0400, Valdis.Kletnieks@vt.edu wrote:
> > > > > On Tue, 01 Apr 2008 21:32:14 PDT, Andrew Morton said:
> > > > > > ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.25-rc8/2.6.25-rc8-mm1/
> > > > > 
> > > > > (Yes, I know the kernel is tainted.  Hopefully the traceback will make
> > > > > enough sense that it won't matter.  I think I cc'd most everybody who is
> > > > > listed in MAINTAINERS or had a non-trivial jbd, quota, or ext3 patch in the broken-out/)
> > > > > 
> > > > > So I was running a 'yum update' on my laptop, walked away to ask a cow-orker
> > > > > a question, and came back to find it had BUG'ed twice...  Amazingly
> > > > > enough, although it died in ext3 code, it apparently only nuked whatever
> > > > > filesystem it was handling, as syslog was still able to log the gory details
> > > > > into a file in /var. Given that a kernel rpm was the one it failed on, the
> > > > > I/O was almost certainly on either / or /boot - both ext3. / is mounted
> > > > > with quotas, /boot isn't, so I'm betting on /
> > > > > 
> > > > > Apr  2 13:48:07 turing-police yum: Updated: texlive-texmf-latex-2007-18.fc9.noarch
> > > > > Apr  2 13:48:08 turing-police yum: Updated: 1:openoffice.org-xsltfilter-2.4.0-12.4.fc9.x86_64
> > > > > Apr  2 13:48:09 turing-police yum: Updated: 1:openoffice.org-javafilter-2.4.0-12.4.fc9.x86_64
> > > > > Apr  2 13:48:12 turing-police yum: Updated: kernel-headers-2.6.25-0.185.rc7.git6.fc9.x86_64
> > > > > 
> > > > > (here, it started updating kernel-2.6.25-0.185.rc7.git6 and died while I wasn't looking)
> > > > 
> > > > <snip>
> > > > 
> > > > Try this patch, it will keep us from re-entering the fs when we aren't supposed
> > > > to.  cc'ing Eric Paris since he's the only selinux guy I know :).  I don't think
> > > > any of the other allocations in here need to be fixed, but I didn't look too
> > > > carefully.
> > > > 
> > > > Signed-off-by: Josef Bacik <jbacik@redhat.com>
> > > > 
> > > > 
> > > > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> > > > index c2fef7b..820d07a 100644
> > > > --- a/security/selinux/hooks.c
> > > > +++ b/security/selinux/hooks.c
> > > > @@ -180,7 +180,7 @@ static int inode_alloc_security(struct inode *inode)
> > > >  	struct task_security_struct *tsec = current->security;
> > > >  	struct inode_security_struct *isec;
> > > >  
> > > > -	isec = kmem_cache_zalloc(sel_inode_cache, GFP_KERNEL);
> > > > +	isec = kmem_cache_zalloc(sel_inode_cache, GFP_NOFS);
> > > >  	if (!isec)
> > > >  		return -ENOMEM;
> > > >  
> > > > @@ -2429,7 +2429,7 @@ static int selinux_inode_init_security(struct inode *inode, struct inode *dir,
> > > >  		return -EOPNOTSUPP;
> > > >  
> > > >  	if (name) {
> > > > -		namep = kstrdup(XATTR_SELINUX_SUFFIX, GFP_KERNEL);
> > > > +		namep = kstrdup(XATTR_SELINUX_SUFFIX, GFP_NOFS);
> > > >  		if (!namep)
> > > >  			return -ENOMEM;
> > > >  		*name = namep;
> > > 
> > > Might fix it.  But 2.6.24's inode_alloc_security() also uses GFP_KERNEL and
> > > doesn't have this bug.  What changed?
> > 
> > Looks legitimate, although we've been doing that since Linux 2.6.0-test3
> > (selinux merge) for inode_alloc_security and d_instantiate, and since
> > Linux 2.6.14 for inode_init_security, so something is at least
> > triggering it more easily now.  inode_doinit_with_dentry looks like
> > another instance and security_context_to_sid_core as well.
>   I guess it is just the combination of someone using SELinux + quota
> (or several journaling filesystems) + being unlucky under memory
> pressure that makes this happen only rarely. Josef, have you been
> successful in reproducing the problem under older kernel?
> 

Not yet, I haven't been lucky enough apparently.  I'm going to kick it off on a
couple of boxes over the weekend and see if I can't hit it on one of them
instead of just relying on the one box.

Josef