Re: [PATCH bpf-next v5 4/5] bpf: verifier: Support eliding map lookup nullness

[Eduard Zingerman]

On Thu, 2024-12-19 at 14:41 -0700, Daniel Xu wrote:

[...]

> > > I think that if test operates on a key like:
> > > 
> > >       valid key 15
> > >              v
> > >       0000000f   <-- written to stack as a single u64 value
> > >       ^^^^^^^
> > >     stack zero marks
> > > 
> > > and is executed (e.g. using __retval annotation),
> > > then CI passing for s390 should be enough.
> > 
> > +1, something like that where for big-endian it will be all zero while
> > for little endian it would be 0xf (and then make sure that the test
> > should *fail* by making sure that 0xf is not a valid index, so NULL
> > check is necessary)
> 
> How would it work for LE to be 0xF but BE to be 0x0?
> 
> The prog passes a pointer to the beginning of the u32 to
> bpf_map_lookup_elem(). The kernel does a 4 byte read starting from that
> address. On both BE and LE all 4 bytes will be interpreted. So set bits
> cannot just go away.
> 
> Am I missing something?

Ok, thinking a bit more, the best test I can come up with is:

  u8 vals[8];
  vals[0] = 0;
  ...
  vals[6] = 0;
  vals[7] = 0xf;
  p = bpf_map_lookup_elem(... vals ...);
  *p = 42;

For LE vals as u32 should be 0x0f;
For BE vals as u32 should be 0xf000_0000.
Hence, it is not safe to remove null check for this program.
What would verifier think about the value of such key?
As far as I understand, there would be stack zero for for vals[0-6]
and u8 stack spill for vals[7].
You were going to add a check for the spill size, which should help here.
So, a negative test like above that checks that verifier complains
that 'p' should be checked for nullness first?

If anyone has better test in mind, please speak-up.

[...]