On Wed, 2016-05-04 at 14:49 +0200, Nicolas Dichtel wrote:

Le 04/05/2016 11:05, Lars Ellenberg a écrit :
[snip]

quoted

We don't have an "alignment problem" there, btw.
Last time I checked, we did work fine without this alignment magic,
we already take care of that, yes, even on affected architectures.

The code adds several consecutive u64 attributes. The nl attribute header is 4
bytes, thus the full attribute length is 12 bytes. If the first u64 is aligned
on 8 (nla_data()), the next one is not aligned on 8: it starts 12 bytes (8 (u64)
+ 4 (nl attr hdr)) after the previous u64.

As I mentioned earlier ( https://lkml.org/lkml/2016/4/22/706 ), if both
kernel and user land do not blindly use *(u64 *)ptr to put/get these
values, there was no issue.

kernel was fine, and most user land apps were fine as well.

Only for compound struct  like tcp_info this was nice to get alignment
for performance reason (it removed one memcpy(), or the use of
put_unaligned() helpers)