On Wed, 4 Dec 2024 at 21:12, Eduard Zingerman <eddyz87@xxxxxxxxx> wrote:
>
> On Tue, 2024-12-03 at 18:41 -0800, Kumar Kartikeya Dwivedi wrote:
>
> [...]
>
> > +/* r2 with offset is checked, which marks r1 with off=0 as non-NULL */
> > +SEC("tp_btf/bpf_testmod_test_raw_tp_null")
> > +__failure
> > +__msg("3: (07) r2 += 8 ; R2_w=trusted_ptr_or_null_sk_buff(id=1,off=8)")
> > +__msg("4: (15) if r2 == 0x0 goto pc+2 ; R2_w=trusted_ptr_or_null_sk_buff(id=2,off=8)")
> > +__msg("5: (bf) r1 = r1 ; R1_w=trusted_ptr_sk_buff()")
>
> This looks like a bug.
> 'r1 != 0' does not follow from 'r2 == r1 + 8 and r2 != 0'.
>
Hmm, yes, it's broken.
I am realizing where we do it now will walk r1 first and we'll not see
r2 off != 0 until after we mark it already.
I guess we need to do the check sooner outside this function in
mark_ptr_or_null_regs.
There we have the register being operated on, so if off != 0 we don't
walk all regs in state.
Do you think that should fix this?
> > +int BPF_PROG(test_raw_tp_null_copy_check_with_off, struct sk_buff *skb)
> > +{
> > + asm volatile (
> > + "r1 = *(u64 *)(r1 +0); \
> > + r2 = r1; \
> > + r3 = 0; \
> > + r2 += 8; \
> > + if r2 == 0 goto jmp2; \
> > + r1 = r1; \
> > + *(u64 *)(r3 +0) = r3; \
> > + jmp2: "
> > + ::
> > + : __clobber_all
> > + );
> > + return 0;
> > +}
>
> [...]
