Re: slab-out-of-bounds due to "KVM: Dynamically size memslot array based on number of used slots"

[Peter Xu]

On Fri, Mar 20, 2020 at 06:53:46AM -0700, Sean Christopherson wrote:
> On Fri, Mar 20, 2020 at 09:49:03AM -0400, Qian Cai wrote:
> > 
> > 
> > > On Mar 20, 2020, at 12:34 AM, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> > > 
> > > On Thu, Mar 19, 2020 at 11:59:23PM -0400, Qian Cai wrote:
> > >> Reverted the linux-next commit 36947254e5f98 (âKVM: Dynamically size memslot array based on number of used slotsâ)
> > >> fixed illegal slab object redzone accesses.
> > >> 
> > >> [6727.939776][ T1818] BUG: KASAN: slab-out-of-bounds in gfn_to_hva+0xc1/0x2b0 [kvm]
> > >> search_memslots at include/linux/kvm_host.h:1035
> > > 
> > > Drat.  I'm guessing lru_slot is out of range after a memslot is deleted.
> > > This should fix the issue, though it may not be the most proper fix, e.g.
> > > it might be better to reset lru_slot when deleting a memslot.  I'll try and
> > > reproduce tomorrow, unless you can confirm this does the trick.
> > 
> > It works fine.
> 
> Thanks!  I'll send a proper patch in a bit, tweaking a selftest to try and
> hit this as well.

Would resetting lru_slot be better?  So to avoid other potential
references to an obsolete lru_slot outside search_memslots().  E.g., I
see that s390 has another function (gfn_to_memslot_approx) used it.

Thanks,

-- 
Peter Xu