Re: [PATCH v2] io_uring,audit: don't log IORING_OP_MADVISE

[Paul Moore]

On Thu, Feb 9, 2023 at 5:54 PM Steve Grubb <sgrubb@xxxxxxxxxx> wrote:
> On Thursday, February 9, 2023 5:37:22 PM EST Paul Moore wrote:
> > On Thu, Feb 9, 2023 at 4:53 PM Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > > On 2023-02-01 16:18, Paul Moore wrote:
> > > > On Wed, Feb 1, 2023 at 3:34 PM Richard Guy Briggs <rgb@xxxxxxxxxx>
> wrote:
> > > > > fadvise and madvise both provide hints for caching or access pattern
> > > > > for file and memory respectively.  Skip them.
> > > >
> > > > You forgot to update the first sentence in the commit description :/
> > >
> > > I didn't forget.  I updated that sentence to reflect the fact that the
> > > two should be treated similarly rather than differently.
> >
> > Ooookay.  Can we at least agree that the commit description should be
> > rephrased to make it clear that the patch only adjusts madvise?  Right
> > now I read the commit description and it sounds like you are adjusting
> > the behavior for both fadvise and madvise in this patch, which is not
> > true.
> >
> > > > I'm still looking for some type of statement that you've done some
> > > > homework on the IORING_OP_MADVISE case to ensure that it doesn't end
> > > > up calling into the LSM, see my previous emails on this.  I need more
> > > > than "Steve told me to do this".
> > > >
> > > > I basically just want to see that some care and thought has gone into
> > > > this patch to verify it is correct and good.
> > >
> > > Steve suggested I look into a number of iouring ops.  I looked at the
> > > description code and agreed that it wasn't necessary to audit madvise.
> > > The rationale for fadvise was detemined to have been conflated with
> > > fallocate and subsequently dropped.  Steve also suggested a number of
> > > others and after investigation I decided that their current state was
> > > correct.  *getxattr you've advised against, so it was dropped.  It
> > > appears fewer modifications were necessary than originally suspected.
> >
> > My concern is that three of the four changes you initially proposed
> > were rejected, which gives me pause about the fourth.  You mention
> > that based on your reading of madvise's description you feel auditing
> > isn't necessary - and you may be right - but based on our experience
> > so far with this patchset I would like to hear that you have properly
> > investigated all of the madvise code paths, and I would like that in
> > the commit description.
>
> I think you're being unnecessarily hard on this.

Asking that a patch author does the proper level of due diligence to
ensure that the patch they are submitting is correct isn't being
"unnecessarily hard", it's part of being a good code reviewer and
maintainer.  I'm a bit amazed that you and Richard would rather spend
your time arguing about this rather than spending the hour (?) it
would take to verify the change and make a proper statement about the
correctness of the patch.

> Yes, the commit message
> might be touched up. But madvise is advisory in nature. It is not security
> relevant. And a grep through the security directory doesn't turn up any
> hooks.

You can't rely on grep, you need to chase the code paths to see what
code might be exercised.  For example, look at the truncate syscalls
(truncate, truncate64, ftruncate64, etc.), if you grep through the
SELinux hook code looking for some form of "truncate" you will not
find anything relevant; SELinux doesn't provide implementations for
either the security_file_truncate() or security_path_truncate() hooks.
However, if you follow the syscall code paths you will see that the
truncate syscalls end up calling security_inode_permission() which
*is* implemented in SELinux.

You need to do your homework; relying only on a gut feeling, a few
lines from a manpage, or a code comment is a good way to introduce
bugs.

-- 
paul-moore.com