Re: [PATCH 3/4] integrity: IMA as an integrity service provider

[Dave Hansen <dave@linux.vnet.ibm.com>]

On Thu, 2008-11-20 at 20:39 -0500, Mimi Zohar wrote:
> On Thu, 2008-11-20 at 13:22 -0800, Dave Hansen wrote: 
> > On Thu, 2008-11-20 at 11:43 -0500, Mimi Zohar wrote:
> > > 
> > > +       /* Invalidate PCR, if a measured file is already open for read
> > > */
> > > +       if ((mask == MAY_WRITE) || (mask == MAY_APPEND)) {
> > > +               int mask_sav = data->mask;
> > > +               int rc;
> > > +
> > > +               data->mask = MAY_READ;
> > > +               rc = ima_must_measure(&idata);
> > > +               if (!rc) {
> > > +                       if (atomic_read(&(data->dentry->d_count)) - 1 >
> > > +                           atomic_read(&(inode->i_writecount)))
> > > +                               ima_add_violation(inode, data->filename,
> > > +                                                 "invalid_pcr", "ToMToU");
> > > +               }
> > > +               data->mask = mask_sav;
> > > +               goto out;
> > > +       }
> > 
> > Following up on Christoph's comment...
> > 
> > I'm worried that this calculation isn't very precise.  The calculation
> > that you're trying to come up with here is the number of opens (d_count)
> > vs. the number of writers (i_writecount).  When they don't match, you
> > know that the new open is the first write, and you must 'invalidate the
> > PCR'?
> > 
> > There are a number of things that elevate d_count, and it is a lot more
> > than just an open() that can do it.  Is that OK?
> 
> >From an integrity perspective, a file measurement might be invalidated
> unnecessarily, but it is safe. For any file when opened for write, while
> having an existing reader, will cause the file measurement to be
> invalidated. Can you give examples of things, other than open(), that
> elevate d_count?

Just do a little search for dget().  There are plenty of places that
we'll do a lookup (and a dget), realize that we can't complete an
operation (say O_RDWR on a 400 file), then dput() the dentry.  If you
look at dput at the wrong time, you could mistake that for an *actual*
open.

Are you holding any locks here?  If not, it is completely conceivable
that you do the atomic_read(d_count)=100, then by the time you do the
aomtic_read(i_writecount), 90 of those references have been dput()'d.  

> Is there a different, better way to determine if there are any readers?

I think you're looking at it from the wrong angle.  If you have a
writer, does it matter whether there are any readers?  You should just
unconditionally invalidate the integrity measurement.

-- Dave