On Mon, Apr 15, 2024 at 11:15:03PM +0100, Matthew Wilcox wrote:
> On Mon, Apr 15, 2024 at 03:05:44PM -0700, Vishal Moola wrote:
> > Commit 9acad7ba3e25 ("hugetlb: use vmf_anon_prepare() instead of
> > anon_vma_prepare()") may bailout after allocating a folio if we do not
> > hold the mmap lock. When this occurs, vmf_anon_prepare() will release the
> > vma lock. Hugetlb then attempts to call restore_reserve_on_error(),
> > which depends on the vma lock being held.
> >
> > We can move vmf_anon_prepare() prior to the folio allocation in order to
> > avoid calling restore_reserve_on_error() without the vma lock.
>
> But now you're calling vmf_anon_prepare() in the wrong place -- before
> we've determined that we need an anon folio. So we'll create an
> anon_vma even when we don't need one for this vma.
>
> This is definitely a pre-existing bug which you've exposed by making it
> happen more easily. Needs a different fix though.
I do not think this is a pre-existing bug.
Prior to 'commit: 7c43a553792a ("hugetlb: allow faults to be handled under
the VMA lock"), we would just bail out if we had FAULT_FLAG_VMA_LOCK.
So there was no danger in calling functions that fiddle with vmas like
restore_reserve_on_error() does.
After that, we allow it but vmf_anon_prepare() releases the lock and returns
VM_FAULT_RETRY if we really need to allocate an anon_vma.
The problem is that now restore_reserve_on_error() will re-adjust the
reservations without the vma lock, completely unsafe.
I think the safest way to tackle this is just as Vishal did, call
vmf_anon_prepare() upfront only for non VM_MAYSHARE faults.
--
Oscar Salvador
SUSE Labs
