Re: inconsistence in mprotect_fixup mlock_fixup madvise_update_vma

[Mike Rapoport <rppt@kernel.org>]

On Tue, Jun 20, 2023 at 03:29:34PM -0700, Jeff Xu wrote:
> On Wed, Jun 14, 2023 at 5:58 AM Mike Rapoport <rppt@kernel.org> wrote:
> >
> > On Tue, Jun 13, 2023 at 09:18:14PM -0400, Liam R. Howlett wrote:
> > > * Jeff Xu <jeffxu@chromium.org> [230613 17:29]:
> > > > Hello Peter,
> > > >
> > > > Thanks for responding.
> > > >
> > > > On Tue, Jun 13, 2023 at 1:16 PM Peter Xu <peterx@redhat.com> wrote:
> > > > >
> > > > > Hi, Jeff,
> > > > >
> > > > > On Tue, Jun 13, 2023 at 08:26:26AM -0700, Jeff Xu wrote:
> > > > > > + more ppl to the list.
> > > > > >
> > > > > > On Mon, Jun 12, 2023 at 6:04 PM Jeff Xu <jeffxu@chromium.org> wrote:
> > > > > > >
> > > > > > > Hello,
> > > > > > >
> > > > > > > There seems to be inconsistency in different VMA fixup
> > > > > > > implementations, for example:
> > > > > > > mlock_fixup will skip VMA that is hugettlb, etc, but those checks do
> > > > > > > not exist in mprotect_fixup and madvise_update_vma. Wouldn't this be a
> > > > > > > problem? the merge/split skipped by mlock_fixup, might get acted on in
> > > > > > > the madvice/mprotect case.
> > > > > > >
> > > > > > > mlock_fixup currently check for
> > > > > > > if (newflags == oldflags ||
> > >
> > > newflags == oldflags, then we don't need to do anything here, it's
> > > already at the desired mlock.  mprotect does this, madvise does this..
> > > probably.. it's ugly.
> > >
> > > > > > > (oldflags & VM_SPECIAL) ||
> > >
> > > It's special, merging will fail always.  I don't know about splitting,
> > > but I guess we don't want to alter the mlock state on special mappings.
> > >
> > > > > > > is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm) ||
> > > > > > > vma_is_dax(vma) || vma_is_secretmem(vma))
> > > > >
> > > > > The special handling you mentioned in mlock_fixup mostly makes sense to me.
> > > > >
> > > > > E.g., I think we can just ignore mlock a hugetlb page if it won't be
> > > > > swapped anyway.
> > > > >
> > > > > Do you encounter any issue with above?
> > > > >
> > > > > > > Should there be a common function to handle VMA merge/split ?
> > > > >
> > > > > IMHO vma_merge() and split_vma() are the "common functions".  Copy Lorenzo
> > > > > as I think he has plan to look into the interface to make it even easier to
> > > > > use.
> > > > >
> > > > The mprotect_fixup doesn't have the same check as mlock_fixup. When
> > > > userspace calls mlock(), two VMAs might not merge or split because of
> > > > vma_is_secretmem check, However, when user space calls mprotect() with
> > > > the same address range, it will merge/split.  If mlock() is doing the
> > > > right thing to merge/split the VMAs, then mprotect() is not ?
> > >
> > > It looks like secretmem is mlock'ed to begin with so they don't want it
> > > to be touched.  So, I think they will be treated differently and I think
> > > it is correct.
> >
> > Right, they don't :)
> >
> > secretmem VMAs are always mlocked, they cannot be munlocked and there is no
> > point trying to mlock them again.
> >
> > The mprotect for secretmem is Ok though, so e.g. if we (unlikely) have two
> > adjacent secretmem VMAs in a range passed to mprotect, it's fine to merge
> > them.
> >
> 
> I m thinking/brainstorming below, assuming:
> Address range 1: 0x5000 to 0x6000 (regular mmap)
> Address range 2: 0x6000 to 0x7000 (allocated to secretmem)
> Address range 3: 0x7000 to 0x8000 (regular mmap)
> 
> User space call: mlock(0x5000,0x3000)
> range 1 and 2 won't merge.
> range 2 and 3  could merge, when mlock_fixup  checks current vma
> (range 3), it is not secretmem, so it will merge with prev vma.

But 2 and 3 have different vm_file, they won't merge.
 
> user space call: mprotect(0x5000,0x3000)
> range 1 2 3 could merge,  all three can have the same flags.
> Note: vma_is_secretmem() isn't checked in mprotect_fixup, same for
> vma_is_dax and get_gate_vma, those doesn't have included in
> vma->vm_flags
> 
> Once 1 and 2 are merged, maybe user space is able to use
> munlock(0x5000,0x3000)
> to unlock range 1 to 3, this will include 2, right ? (haven't used the
> code to prove it)

But 1 and 2 won't merge because their vm_file's are different.
 
> I'm using secretmem as an example here, having 3 different _fixup
> implementations seems to be error prone to me.

The actual decision whether to merge VMAs is taken in vma_merge rather than
by the _fixup functions. So while the checks around vma_merge might be
different in these functions, it does not mean it's possible to wrongly
merge VMA unless there is a bug in vma_merge. So in the end it boils down
to a single core implementation, don't you agree?
 
> Thanks
> -Jeff

-- 
Sincerely yours,
Mike.