Ilya Leoshkevich <iii@xxxxxxxxxxxxx> writes:
> On the architectures that use bpf_jit_needs_zext(), e.g., s390x, the
> verifier incorrectly inserts a zero-extension after BPF_MEMSX, leading
> to miscompilations like the one below:
>
> 24: 89 1a ff fe 00 00 00 00 "r1 = *(s16 *)(r10 - 2);" # zext_dst set
> 0x3ff7fdb910e: lgh %r2,-2(%r13,%r0) # load halfword
> 0x3ff7fdb9114: llgfr %r2,%r2 # wrong!
> 25: 65 10 00 03 00 00 7f ff if r1 s> 32767 goto +3 <l0_1> # check_cond_jmp_op()
>
> Disable such zero-extensions. The JITs need to insert sign-extension
> themselves, if necessary.
>
> Suggested-by: Puranjay Mohan <puranjay12@xxxxxxxxx>
> Signed-off-by: Ilya Leoshkevich <iii@xxxxxxxxxxxxx>
> ---
> kernel/bpf/verifier.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index a7178ecf676d..614bf3fa4fd5 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -3114,7 +3114,7 @@ static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
>
> if (class == BPF_LDX) {
> if (t != SRC_OP)
> - return BPF_SIZE(code) == BPF_DW;
> + return BPF_SIZE(code) == BPF_DW || BPF_MODE(code) == BPF_MEMSX;
> /* LDX source must be ptr. */
> return true;
> }
> --
> 2.41.0
Reviewed-by: Puranjay Mohan <puranjay12@xxxxxxxxx>
Thanks for this.
Puranjay
