On Wed, 2019-08-28 at 22:24 +0200, Marcel Holtmann wrote: > Hi Johannes, > > > > > No, as usual, that would break ABI. PAD is a regular attribute, just > > > > empty and ignored for aligning 64-bit values. > > > > > > then I do not grok on how the nla_put_u64_64bit works, but that is > > > fine. > > > > > > I assumed these are similar to the NL80211_SURVEY_INFO_MAX which we > > > also always move, but also not expected to be part of the API as a > > > fixed value. > > > > No no, the _MAX is just the token we use for knowing what we want as the > > maximum when parsing etc. > > > > The _PAD is actually a real attribute, basically nla_put_u64_64bit() > > will do "nla_put_flag(_PAD)" if and only if "offset % 8 == 0", in order > > to actually 64-bit align the 64-bit value in the following attribute. > > > > (Note that offset % 8 can only be 0 or 4, due to the way netlink > > attributes work.) > > I get that part now. So the kernel is inserting a _PAD, but userspace > is still not doing that. Yeah, and I forgot to say - if we renumbered _PAD, then newer userspace would think old kernel's _PAD is really _BSS_RX, so things would break. > So for NL80211_ATTR_WDEV we should be doing the same actually? Not really. The kernel doesn't rely on it, nla_get_u64() uses nla_memcpy() so doesn't care about alignment. Even userspace doesn't (usually) rely on the alignment, it also typically uses nla_get_u64() from libnl which also uses nla_memcpy() or memcpy() (depending on the version), so it's all not really necessary. I think some libraries or tools like maybe iproute2 didn't do it correctly and just dereferenced a pointer there, causing alignment violations, and so basically the decision was to get rid of unaligned 64-bit attributes to prevent that once and for all. johannes
