Re: [PATCH RFC bpf-next 2/8] bpf: mark instructions accessing program stack

On Thu Jun 4, 2026 at 6:24 PM CEST, Alexei Starovoitov wrote:
> On Thu, Jun 4, 2026 at 5:23=E2=80=AFAM Alexis Lothor=C3=A9
> <[email protected]> wrote:
>>
>> On Sat Apr 25, 2026 at 7:05 AM CEST, sashiko-bot wrote:
>>
>> [...]
>>
>> >>  #define LR_FRAMENO_BITS     3
>> >>  #define LR_SPI_BITS 6
>> > [ ... ]
>> >> @@ -6493,6 +6498,8 @@ static int check_mem_access(struct bpf_verifier=
_env *env, int insn_idx, u32 regn
>> >>              else
>> >>                      err =3D check_stack_write(env, regno, off, size,
>> >>                                              value_regno, insn_idx);
>> >> +
>> >> +            mark_insn_accesses_stack(env, insn_idx);
>> >
>> > Global subprograms are verified independently of their call sites, and=
 pointer
>> > arguments are tracked as PTR_TO_MEM rather than PTR_TO_STACK. Inside t=
he
>> > subprogram, check_mem_access() evaluates loads and stores under the
>> > PTR_TO_MEM branch, which does not invoke mark_insn_accesses_stack().
>> >
>> > If a BPF caller dynamically passes a pointer to its stack as an argume=
nt to
>> > the global subprogram, will the JIT-compiled global subprogram execute=
 with
>> > KASAN checks enabled for those accesses?
>> >
>> > Does this erroneously monitor the BPF stack, bypassing the intended
>> > suppression mechanism?
>>
>> After having thought about this one for quite some time, I see no
>> simple way to correctly distinguish stack accesses hidden behing a
>> PTR_TO_MEM passed to a global subprog, as those are verified separately.
>>
>> I can either:
>> 1. skip PTR_TO_MEM memory instrumentation when we are in global
>> subprograms, at the risk of missing instrumentation on passed memory
>> accesses that _need_ to be instrumented
>> 2. systematically instrument passed PTR_TO_MEM memory accesses in global
>> subprog, at the risk of inserting unneeded instrumentation when the
>> passed memory comes in fact from caller memory.
>>
>> I then propose to be conservative and apply 2:  the only downside I can
>> think of is that a few alwasy-passing kasan checks will be inserted, but
>> we are at least sure not to miss any instrumentation.
>
> Isn't it the other way around? Replacing LDX with asan_load() will
> check stack access, but it doesn't have shadow memory behind it?
> Or you mean to assume kasan_vmalloc and vmalloced_stack kconfig?

Ah yes, my bad, I did not state it explicitely, but my point above
indeed assumes that we are with CONFIG_KASAN_VMALLOC and
CONFIG_VMAP_STACK




--=20
Alexis Lothor=C3=A9, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com