> On 20 Dec 2022, at 6:01 AM, Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 12/12/22 7:04 PM, Hao Sun wrote:
>> Verify that nullness information is not porpagated in the branches
>> of register to register JEQ and JNE operations if one of them is
>> PTR_TO_BTF_ID.
>
> Thanks for the fix and test.
>
>> Signed-off-by: Hao Sun <sunhao.th@xxxxxxxxx>
>> Acked-by: Yonghong Song <yhs@xxxxxx>
>> ---
>> .../bpf/verifier/jeq_infer_not_null.c | 22 +++++++++++++++++++
>> 1 file changed, 22 insertions(+)
>> diff --git a/tools/testing/selftests/bpf/verifier/jeq_infer_not_null.c b/tools/testing/selftests/bpf/verifier/jeq_infer_not_null.c
>> index 67a1c07ead34..b2b215227d97 100644
>> --- a/tools/testing/selftests/bpf/verifier/jeq_infer_not_null.c
>> +++ b/tools/testing/selftests/bpf/verifier/jeq_infer_not_null.c
>> @@ -172,3 +172,25 @@
>> .prog_type = BPF_PROG_TYPE_XDP,
>> .result = ACCEPT,
>> },
>> +{
>> + "jne/jeq infer not null, PTR_TO_MAP_OR_NULL unchanged with PTR_TO_BTF_ID reg",
>> + .insns = {
>> + BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
>> + BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),
>> + BPF_ST_MEM(BPF_DW, BPF_REG_2, 0, 0),
>> + BPF_LD_MAP_FD(BPF_REG_1, 0),
>> + /* r6 = bpf_map->inner_map_meta; */
>> + BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_1, 8),
>
> This bpf_map->inner_map_meta requires CO-RE. It works now but could be fragile in different platform and in the future bpf_map changes. Take a look at the map_ptr_kern.c which uses "__attribute__((preserve_access_index))" at the "struct bpf_map".
>
> Please translate this verifer test into a proper bpf prog in C code such that it can use the CO-RE in libbpf. It should run under test_progs instead of test_verifier. The bpf prog can include the "vmlinux.h" to get the "__attribute__((preserve_access_index))" for free. Take a look at https://lore.kernel.org/all/20221207201648.2990661-2-andrii@xxxxxxxxxx/ which has example on how to check verifier message in test_progs.
>
Hi,
I’ve tried something like the bellow, but soon realized that this
won’t work because once compiler figures out `inner_map` equals
to `val`, it can choose either reg to write into in the following
path, meaning that this program can be rejected due to writing
into read-only PTR_TO_BTF_ID reg, and this makes the test useless.
Essentially, we want two regs, one points to PTR_TO_BTD_ID, one
points to MAP_VALUR_OR_NULL, then compare them and deref map val.
It’s hard to implement this in C level because compilers decide
which reg to use but not us, maybe we can just drop this test.
thoughts?
+struct {
+ __uint(type, BPF_MAP_TYPE_HASH);
+ __uint(max_entries, 1);
+ __type(key, u64);
+ __type(value, u64);
+} m_hash SEC(".maps");
+
+SEC("?raw_tp")
+__failure __msg("invalid mem access 'map_value_or_null")
+int jeq_infer_not_null_ptr_to_btfid(void *ctx)
+{
+ struct bpf_map *map = (struct bpf_map *)&m_hash;
+ struct bpf_map *inner_map = map->inner_map_meta;
+ u64 key = 0, ret = 0, *val;
+
+ val = bpf_map_lookup_elem(map, &key);
+ /* Do not mark ptr as non-null if one of them is
+ * PTR_TO_BTF_ID, reject because of invalid access
+ * to map value.
+ */
+ if (val == inner_map)
+ ret = *val;
+
+ return ret;
+}
