Re: [PATCH] x86: use WRITE_ONCE() when setting PTEs

[Peter Zijlstra]

On Thu, Sep 06, 2018 at 01:12:14PM -0700, Nadav Amit wrote:
> at 12:57 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Sun, Sep 02, 2018 at 11:14:50AM -0700, Nadav Amit wrote:
> >> When page-table entries are set, the compiler might optimize their
> >> assignment by using multiple instructions to set the PTE. This might
> >> turn into a security hazard if the user somehow manages to use the
> >> interim PTE. L1TF does not make our lives easier, making even an interim
> >> non-present PTE a security hazard.
> >> 
> >> Using WRITE_ONCE() to set PTEs and friends should prevent this potential
> >> security hazard.
> >> 
> >> I skimmed the differences in the binary with and without this patch. The
> >> differences are (obviously) greater when CONFIG_PARAVIRT=n as more
> >> code optimizations are possible. For better and worse, the impact on the
> >> binary with this patch is pretty small. Skimming the code did not cause
> >> anything to jump out as a security hazard, but it seems that at least
> >> move_soft_dirty_pte() caused set_pte_at() to use multiple writes.
> > 
> > Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > 
> > Also, its corollary would also make sense/be required, use READ_ONCE()
> > when reading these.
> 
> I donât know. This would obviously be much more intrusive. I can add a
> get_pte() and write a Coccinelle script to use it instead of reading the
> PTE, but in most cases, I presume, it would be an overkill.
> 
> The reason for that is that the PTEs are supposed to be accessed while
> holding the page-table lock, and the hardware can only change dirty & access
> bits. I think that any code that assumes that these bits do not change while
> holding the lock is already broken in more ways.

There are lockless readers, but I just checked, mm/gup.c already uses
READ_ONCE(), so that should be fine.