Re: [PATCH bpf-next 2/2] selftests/bpf: Add tests for bpf_get_dentry_xattr

[Song Liu]

Hi Christian, 

> On Jul 26, 2024, at 12:06 AM, Christian Brauner <brauner@xxxxxxxxxx> wrote:

[...]

>> +
>> + for (i = 0; i < 10; i++) {
>> + ret = bpf_get_dentry_xattr(dentry, "user.kfunc", &value_ptr);
>> + if (ret == sizeof(expected_value) &&
>> +    !bpf_strncmp(value, ret, expected_value))
>> + matches++;
>> +
>> + prev_dentry = dentry;
>> + dentry = bpf_dget_parent(prev_dentry);
> 
> Why do you need to walk upwards and instead of reading the xattr values
> during security_inode_permission()?

In this use case, we would like to add xattr to the directory to cover
all files under it. For example, assume we have the following xattrs:

  /bin  xattr: user.policy_A = value_A
  /bin/gcc-6.9/ xattr: user.policy_A = value_B
  /bin/gcc-6.9/gcc xattr: user.policy_A = value_C

/bin/gcc-6.9/gcc will use value_C;
/bin/gcc-6.9/<other_files> will use value_B;
/bin/<other_folder_or_file> will use value_A;

By walking upwards from security_file_open(), we can finish the logic 
in a single LSM hook:

    repeat:
        if (dentry have user.policy_A) {
            /* make decision based on value */;
        } else {
            dentry = bpf_dget_parent();
            goto repeat;
        }

Does this make sense? Or maybe I misunderstood the suggestion?

Also, we don't have a bpf_get_inode_xattr() yet. I guess we will need
it for the security_inode_permission approach. If we agree that's a 
better approach, I more than happy to implement it that way. In fact,
I think we will eventually need both bpf_get_inode_xattr() and 
bpf_get_dentry_xattr(). 

Thanks,
Song


>> + bpf_dput(prev_dentry);
>> + }
>> +