On Fri, Apr 26, 2019 at 01:23:46PM +0200, Johannes Berg wrote: > On Fri, 2019-04-26 at 13:19 +0200, Michal Kubecek wrote: > > On Fri, Apr 26, 2019 at 11:24:15AM +0200, Johannes Berg wrote: > > > On Fri, 2019-04-26 at 09:13 +0000, Michal Kubecek wrote: > > > > > > > Another observation was that even if NLA_F_NESTED flag was introduced in > > > > 2007, only few netlink based interfaces set it in kernel generated messages > > > > and even many recently added APIs omit it. That is unfortunate as without > > > > the flag, message parsers not familiar with attribute semantics cannot > > > > recognize nested attributes and do not see message structure; this affects > > > > e.g. wireshark dissector or mnl_nlmsg_fprintf() from libmnl. > > > > > > Indeed. > > > > > > I wonder if we should also (start) enforcing that the userspace sender > > > side sets this, if the policy is strict? > > > > I suppose we should, at least the part that attribute with NLA_NESTED > > policy has NLA_F_NESTED flag. I'm not so sure about the opposite (i.e. > > that attributes with other policies do not have the flag) as when I was > > checking where kernel accesses nlattr::nla_type directly rather than > > with nla_type(), I stumbled upon an attribute NL80211_ATTR_VENDOR_DATA > > which has policy NLA_BINARY but is sometimes a nest, AFAICS. > > I guess anyway we can only do it for *new* things, not really for all > existing attributes. Right... but what I wanted to say is that if there is already (at least) one attribute which may or may not be a nest, depending on a context, we should expect there may be also new attributes like that in the future. Michal
