Re: [PATCH] mm: Don't drop VMA locks in mm_drop_all_locks()

[Suren Baghdasaryan]

On Thu, Jul 20, 2023 at 10:14 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
>
> On Thu, Jul 20, 2023 at 6:52 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > On Wed, Jul 19, 2023 at 6:33 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > >
> > > Despite its name, mm_drop_all_locks() does not drop _all_ locks; the mmap
> > > lock is held write-locked by the caller, and the caller is responsible for
> > > dropping the mmap lock at a later point (which will also release the VMA
> > > locks).
> > > Calling vma_end_write_all() here is dangerous because the caller might have
> > > write-locked a VMA with the expectation that it will stay write-locked
> > > until the mmap_lock is released, as usual.
> > >
> > > This _almost_ becomes a problem in the following scenario:
> > >
> > > An anonymous VMA A and an SGX VMA B are mapped adjacent to each other.
> > > Userspace calls munmap() on a range starting at the start address of A and
> > > ending in the middle of B.
> > >
> > > Hypothetical call graph with additional notes in brackets:
> > >
> > > do_vmi_align_munmap
> > >   [begin first for_each_vma_range loop]
> > >   vma_start_write [on VMA A]
> > >   vma_mark_detached [on VMA A]
> > >   __split_vma [on VMA B]
> > >     sgx_vma_open [== new->vm_ops->open]
> > >       sgx_encl_mm_add
> > >         __mmu_notifier_register [luckily THIS CAN'T ACTUALLY HAPPEN]
> > >           mm_take_all_locks
> > >           mm_drop_all_locks
> > >             vma_end_write_all [drops VMA lock taken on VMA A before]
> > >   vma_start_write [on VMA B]
> > >   vma_mark_detached [on VMA B]
> > >   [end first for_each_vma_range loop]
> > >   vma_iter_clear_gfp [removes VMAs from maple tree]
> > >   mmap_write_downgrade
> > >   unmap_region
> > >   mmap_read_unlock
> > >
> > > In this hypothetical scenario, while do_vmi_align_munmap() thinks it still
> > > holds a VMA write lock on VMA A, the VMA write lock has actually been
> > > invalidated inside __split_vma().
> > >
> > > The call from sgx_encl_mm_add() to __mmu_notifier_register() can't
> > > actually happen here, as far as I understand, because we are duplicating an
> > > existing SGX VMA, but sgx_encl_mm_add() only calls
> > > __mmu_notifier_register() for the first SGX VMA created in a given process.
> > > So this could only happen in fork(), not on munmap().
> > > But in my view it is just pure luck that this can't happen.
> > >
> > > Also, we wouldn't actually have any bad consequences from this in
> > > do_vmi_align_munmap(), because by the time the bug drops the lock on VMA A,
> > > we've already marked VMA A as detached, which makes it completely
> > > ineligible for any VMA-locked page faults.
> > > But again, that's just pure luck.
> > >
> > > So remove the vma_end_write_all(), so that VMA write locks are only ever
> > > released on mmap_write_unlock() or mmap_write_downgrade().
> >
> > Your logic makes sense to be. mm_drop_all_locks() unlocking all VMAs,
> > even the ones which were locked before mm_take_all_locks() seems
> > dangerous.
> > One concern I have is that mm_take_all_locks() and mm_drop_all_locks()
> > become asymmetric with this change: mm_take_all_locks() locks all VMAs
> > but mm_drop_all_locks() does not release them. I think there should be
> > an additional comment explaining this asymmetry.
> > Another side-effect which would be nice to document in a comment is
> > that when mm_take_all_locks() fails after it locked the VMAs, those
> > VMAs will stay locked until mmap_write_unlock/mmap_write_downgrade.
> > This happens because of failure mm_take_all_locks() jumps to perform
> > mm_drop_all_locks() and this will not unlock already locked VMAs.
> > Other than that LGTM. Thanks!
>
> But this is not specific to mm_drop_all_locks() at all, right? It's just
> fundamentally how per-VMA locks are used everywhere. Somewhere deep
> down in some call path, while the mmap lock is held in write mode, a
> VMA is marked as being written to, and then this marking persists
> until the mmap lock is dropped.

Yes, but for all other locks mm_take_all_locks()/mm_drop_all_locks()
is perfectly symmetrical AFAIKT. I see pretty descriptive comment for
mm_take_all_locks() explaining locking rules, so mentioning this
asymmetry in there seems appropriate to me.

>
> If we want to clarify this, I guess some comments on
> vma_end_write_all() and vma_start_write() might help, but I think
> that's independent of this patch.

Yes and no. This patch changes vma_end_write_all() to be called
exclusively from mmap_write_unlock()/mmap_write_downgrade(), so it
introduces the notion that once VMA is write-locked it's never
released until mmap_lock is released/downgraded. So, I agree that a
comment for vma_start_write() is appropriate but I also think it does
belong in this patch.
Thanks,
Suren.