bpf: Forget ranges when refining tnum after JSET

Syzbot reported a kernel warning due to a range invariant violation on
the following BPF program.

0: call bpf_get_netns_cookie
1: if r0 == 0 goto <exit>
2: if r0 & Oxffffffff goto <exit>

The issue is on the path where we fall through both jumps.

That path is unreachable at runtime: after insn 1, we know r0 != 0, but
with the sign extension on the jset, we would only fallthrough insn 2
if r0 == 0. Unfortunately, is_branch_taken() isn't currently able to
figure this out, so the verifier walks all branches. The verifier then
refines the register bounds using the second condition and we end
up with inconsistent bounds on this unreachable path:

1: if r0 == 0 goto <exit>
r0: u64=[0x1, 0xffffffffffffffff] var_off=(0, 0xffffffffffffffff)
2: if r0 & 0xffffffff goto <exit>
r0 before reg_bounds_sync: u64=[0x1, 0xffffffffffffffff] var_off=(0, 0)
r0 after reg_bounds_sync: u64=[0x1, 0] var_off=(0, 0)

Improving the range refinement for JSET to cover all cases is tricky. We
also don't expect many users to rely on JSET given LLVM doesn't generate
those instructions. So instead of improving the range refinement for
JSETs, Eduard suggested we forget the ranges whenever we're narrowing
tnums after a JSET. This patch implements that approach.

Reported-by: [e-mail]
Suggested-by: Eduard Zingerman <[e-mail]>
Acked-by: Yonghong Song <[e-mail]>
Acked-by: Eduard Zingerman <[e-mail]>
Signed-off-by: Paul Chaignon <[e-mail]>
Link: https://lore.kernel.org/r/9d4fd[e-mail]
Signed-off-by: Alexei Starovoitov <[e-mail]>

kernel/bpf/verifier.c

···
 		if (!is_reg_const(reg2, is_jmp32))
 			break;
 		val = reg_const_value(reg2, is_jmp32);
+		/* Forget the ranges before narrowing tnums, to avoid invariant
+		 * violations if we're on a dead branch.
+		 */
+		__mark_reg_unbounded(reg1);
 		if (is_jmp32) {
 			t = tnum_and(tnum_subreg(reg1->var_off), tnum_const(~val));
 			reg1->var_off = tnum_with_subreg(reg1->var_off, t);