Re: [PATCH bpf-next 2/2] selftests/bpf: check nullness propagation for reg to reg comparisons

[Alexei Starovoitov <ast@meta.com>]

On 11/15/22 12:31 PM, Eduard Zingerman wrote:
> On Mon, 2022-11-14 at 10:01 -0800, Alexei Starovoitov wrote:
>> On Fri, Aug 26, 2022 at 10:30 AM Eduard Zingerman <eddyz87@gmail.com> wrote:
>>>
>>> Verify that nullness information is porpagated in the branches of
>>> register to register JEQ and JNE operations.
>>>
>>> Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
>>> Acked-by: Yonghong Song <yhs@fb.com>
>>> ---
>>>   .../bpf/verifier/jeq_infer_not_null.c         | 166 ++++++++++++++++++
>>>   1 file changed, 166 insertions(+)
>>>   create mode 100644 tools/testing/selftests/bpf/verifier/jeq_infer_not_null.c
>>
>> These 4 new tests are failing in unpriv.
> 
> This is interesting. 'test_verifier' passed for me because of
> kernel.unprivileged_bpf_disabled = 1 on my test VM.
> But It also passed on CI ([1]) with the following log:
> 
> 2022-11-06T21:15:53.2873411Z #686/u jne/jeq infer not null, PTR_TO_SOCKET_OR_NULL -> PTR_TO_SOCKET for JNE false branch SKIP
> 2022-11-06T21:15:53.2908232Z #686/p jne/jeq infer not null, PTR_TO_SOCKET_OR_NULL -> PTR_TO_SOCKET for JNE false branch OK
> 
> To skip or not to skip is decided by test_verifier.c:do_test:
> 
> 		if (test_as_unpriv(test) && unpriv_disabled) {
> 			printf("#%d/u %s SKIP\n", i, test->descr);
> 			skips++;
> 		}
> 
> The 'test_as_unpriv(test)' is true for my tests because of the
> .prog_type == BPF_PROG_TYPE_CGROUP_SKB.
> 'unpriv_disabled' is a global set by test_verifier.c:get_unpriv_disabled:
> 
> static void get_unpriv_disabled()
> {
> 	char buf[2];
> 	FILE *fd;
> 
> 	fd = fopen("/proc/sys/"UNPRIV_SYSCTL, "r");
>          // ...
> 	if (fgets(buf, 2, fd) == buf && atoi(buf))
> 		unpriv_disabled = true;
> 	fclose(fd);
> }
> 
> Might it be the case that CI configuration needs an update as below:
> sysctl kernel.unprivileged_bpf_disabled=0
> ?

Yeah. Makes sense to enable unpriv in CI, since it's missing tests
in the current form.