Re: [PATCH bpf 2/2] bpf: test_verifier, test for lookup on queue/stack maps

[Prashant Bhole]

On 11/28/2018 10:06 PM, Mauricio Vasquez wrote:

On 11/28/18 3:45 AM, Daniel Borkmann wrote:

On 11/28/2018 08:51 AM, Prashant Bhole wrote:

This patch adds tests to check whether bpf verifier prevents lookup
on queue/stack maps

Signed-off-by: Prashant Bhole <bhole_prashant...@lab.ntt.co.jp>
---
  tools/testing/selftests/bpf/test_verifier.c | 52 +++++++++++++++++++++
  1 file changed, 52 insertions(+)

diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c

index 550b7e46bf4a..becd9f4f3980 100644
--- a/tools/testing/selftests/bpf/test_verifier.c
+++ b/tools/testing/selftests/bpf/test_verifier.c
@@ -74,6 +74,8 @@ struct bpf_test {
      int fixup_map_in_map[MAX_FIXUPS];
      int fixup_cgroup_storage[MAX_FIXUPS];
      int fixup_percpu_cgroup_storage[MAX_FIXUPS];
+    int fixup_map_queue[MAX_FIXUPS];
+    int fixup_map_stack[MAX_FIXUPS];
      const char *errstr;
      const char *errstr_unpriv;
      uint32_t retval, retval_unpriv;
@@ -4611,6 +4613,38 @@ static struct bpf_test tests[] = {

          .errstr = "cannot pass map_type 7 into func bpf_map_lookup_elem",

          .prog_type = BPF_PROG_TYPE_PERF_EVENT,
      },
+    {
+        "prevent map lookup in queue map",
+        .insns = {
+            BPF_ST_MEM(BPF_DW, BPF_REG_10, -8, 0),
+            BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
+            BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),
+            BPF_LD_MAP_FD(BPF_REG_1, 0),
+            BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
+                     BPF_FUNC_map_lookup_elem),
+            BPF_EXIT_INSN(),
+        },
+        .fixup_map_queue = { 3 },
+        .result = REJECT,
+        .errstr = "invalid stack type R2 off=-8 access_size=0",
+        .prog_type = BPF_PROG_TYPE_XDP,

Hmm, the approach in patch 1 is very fragile, and we're lucky in this case that the verifier bailed out with 'invalid stack type R2 off=-8 access_size=0'

because of key size being zero. If this would have not been the case then

the added ERR_PTR(-EOPNOTSUPP) would basically be seen as a valid pointer and the program could read/write into it. Instead, this needs to be prevented much

earlier like check_map_func_compatibility(),

Actually it is prevented in check_map_func_compatibility(), but stack boundary check is done before in the verifier.

and I would like to have a split

on these approaches to make verifier more robust. While you want ERR_PTR(-EOPNOTSUPP)

for user space syscall side,

In the case of QUEUE and STACK maps this is not relevant because the lookup syscall is mapped into peek operation.

In fact queue_stack_map_lookup_elem() & queue_stack_map_update_elem() should be never called, I think we can remove them safely.

Got it. Shall we keep these verifier tests (patch 2)?

Thanks,
Prashant