On Mon, Dec 17, 2018 at 08:01:02PM +0200, Jarkko Sakkinen wrote:
> On Mon, Dec 17, 2018 at 09:45:40AM -0800, Dave Hansen wrote:
> > > +struct sgx_encl *sgx_encl_alloc(struct sgx_secs *secs)
> > > +{
> > ...
> > > + kref_init(&encl->refcount);
> > > + INIT_LIST_HEAD(&encl->add_page_reqs);
> > > + INIT_RADIX_TREE(&encl->page_tree, GFP_KERNEL);
> > > + mutex_init(&encl->lock);
> > > + INIT_WORK(&encl->add_page_work, sgx_add_page_worker);
> > > +
> > > + encl->mm = current->mm; <---------------------------------> + encl->base = secs->base;
> > > + encl->size = secs->size;
> > > + encl->ssaframesize = secs->ssa_frame_size;
> > > + encl->backing = backing;
> > > +
> > > + return encl;
> > > +}
> >
> > How is this OK without taking a reference on the mm?
It's subtle and the ordering is all kinds of weird, but technically we
are taking a reference on mm when the mmu_notifier is registered in
sgx_encl_create(). sgx_encl_alloc() and sgx_encl_create() are always
called in tandem and with mm->mm_users > 0, so we'll never use encl->mm
without holding a reference to mm. We need to comment the weirdness
or maybe register the notifier before
> > I have a feeling a bunch of your bugs with the mmu notifiers and so
> > forth are because the refcounting is wrong here.
Eh, not really. Maybe the mmu_notifier is more subtle, e.g. calling
do_unmap() after mmput() would be quite obvious, but there's no
fundamental bug, we just haven't needed to touch VMAs during release
prior to moving away from shmem.
> > Sean's SGX_ENCL_MM_RELEASED would, I think be unnecessary if you just
> > take a refcount here and release it when the enclave is destroyed.
>
> Right, atomic_inc(encl->mm->count) here and once when releasing.
>
> The we would not even need the whole mmu notifier in the first place.
I'm pretty sure doing mmget() would result in circular dependencies and
a zombie enclave. In the do_exit() case where a task is abruptly killed:
- __mmput() is never called because the enclave holds a ref
- sgx_encl_release() is never be called because its VMAs hold refs
- sgx_vma_close() is never called because __mmput()->exit_mmap() is
blocked and the process itself is dead, i.e. won't unmap anything.
