Re: [Patch net v2] cls_flower: call nla_ok() before nla_next()

On Thu, Jan 14, 2021 at 12:16 PM Jakub Kicinski [off-list ref] wrote:

On Thu, 14 Jan 2021 12:03:16 -0800 Cong Wang wrote:

quoted

On Thu, Jan 14, 2021 at 10:38 AM Jakub Kicinski [off-list ref] wrote:

quoted

On Thu, 14 Jan 2021 08:38:22 -0800 Cong Wang wrote:

quoted

From: Cong Wang <redacted>

fl_set_enc_opt() simply checks if there are still bytes left to parse,
but this is not sufficent as syzbot seems to be able to generate
malformatted netlink messages. nla_ok() is more strict so should be
used to validate the next nlattr here.

And nla_validate_nested_deprecated() has less strict check too, it is
probably too late to switch to the strict version, but we can just
call nla_ok() too after it.

Reported-and-tested-by: syzbot+2624e3778b18fc497c92@syzkaller.appspotmail.com
Fixes: 0a6e77784f49 ("net/sched: allow flower to match tunnel options")
Fixes: 79b1011cb33d ("net: sched: allow flower to match erspan options")

quoted

@@ -1340,9 +1341,6 @@ static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
                              NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
                              return -EINVAL;
                      }
-
-                     if (msk_depth)
-                             nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
                      break;
              case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
                      if (key->enc_opts.dst_opt_type) {

@@ -1373,14 +1371,17 @@ static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
                              NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
                              return -EINVAL;
                      }
-
-                     if (msk_depth)
-                             nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
                      break;
              default:
                      NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
                      return -EINVAL;
              }
+
+             if (!nla_ok(nla_opt_msk, msk_depth)) {
+                     NL_SET_ERR_MSG(extack, "Mask attribute is invalid");
+                     return -EINVAL;
+             }
+             nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);

we lost the if (msk_depth) now, nla_opt_msk may be NULL -
neither nla_ok() nor nla_next() take NULL

How is "if (msk_depth)" lost when nla_ok() has a stricter one?

1156 static inline int nla_ok(const struct nlattr *nla, int remaining)
1157 {
1158         return remaining >= (int) sizeof(*nla) &&
1159                nla->nla_len >= sizeof(*nla) &&
1160                nla->nla_len <= remaining;
1161 }

Line 1156 assures msk_depth is not only non-zero but also larger
than the nla struct size, and clearly nla won't be dereferenced unless
this check is passed.

Fair, depth will but 0 so first check already fails, but nla_next()
would crash since it tries to access the length of the attribute
unconditionally.

nla_next() is only called when nla_ok() returns true, which is not
the case for msk_depth==0, therefore NULL won't crash here.

The only problem is we become too strict to reject optionally missing
masks, we should not even call nla_ok() here, otherwise it would
break user-space. So,

+               if (!nla_opt_msk)
+                       continue;

Thanks.