On Wed, Aug 22, 2018 at 11:33 AM Linus Torvalds
<torvalds at linux-foundation.org> wrote:
>
> On Wed, Aug 22, 2018 at 11:21 AM Paolo Bonzini <pbonzini at redhat.com> wrote:
> >
> > Yes, KVM is correct but the i915 bits are at least fishy. It's probably
> > as simple as adding a mmget/mmput pair respectively in kvmgt_guest_init
> > and kvmgt_guest_exit, or maybe mmget_not_zero.
>
> Definitely mmget_not_zero(). If it was just mmgrab()'ed earlier, the
> actual page tables might already be gone.
Side note: we _could_ do the mmget_not_zero() inside use_mm() itself,
if we just knew that the mm was at least mmgrab()'ed correctly.
But for some of the uses, even that isn't clear. It's not entirely
obvious that the "struct mm_struct" exists _at_all_ at that point, and
that a mmget_not_zero() wouldn't just have some use-after-free access.
Again, independent lifetime rules could show that this isn't the case
(ie "exit_aio() is always called before exit_mmap(), and kill_ioctx()
takes care of it all"), but it would be good to have the users of
"use_mm()" actually verify their lifetime rules are correct and
enforced.
Because quite often, the lifetime rule might nbot be a mmu notifier or
aio_exit at all, but just be "oh, the user won't exit until this is
all done". But do you *control* the user? What if the user is buggy?
Linus
