Willy, Am 18.07.2018 um 14:30 schrieb Willy Tarreau: >> I can rework the patch, if technical changes look good to you. There's a >> ton of memcpy in there to not destroy the data structures needed for the >> actual communication: make_proxy_line() now always operates on a copy of >> `struct connection remote`. > > I don't see why a copy could be needed, given that you should never have > to modify anything in the source. Or maybe it's because it was more > convenient to store the remapped addresses ? In this case I think that
Yes it is more convenient. In my current patch I just converted once in connection.c and thus did not need to make any large logic changes in make_proxy_line_v[12]: Keep the patch as small as possible. > it's still cleaner to have two local struct sockaddr_storage that you > can use for the operations. Another option, which I *thought* we had but > am probably wrong on this, was to first check the source+destination Yes, you do this, but you specifically check whether `src->ss_family == dst->ss_family`, not for every possible combination. > address classes before deciding what to emit. (or maybe it's done in the > PPv2 format) : > > if (src_is_v4 && dst_is_v4) > /* emit_v4(src, dst) */ > else if (src_is_v6 && dst_is_v6) > /* emit_v6(src, dst) */ > else if (src_is_v6 && dst_is_v4) > /* emit_v6(src, 4to6(dst)) */ > else if (src_is_v4 && dst_is_v6) > /* emit_v6(4to6(src), dst) */ > else > /* emit_unknown()*/ > > So in practice you always need only one temporary sockaddr_storage for > a conversion. Correct. Reason I did the copy (as mentioned above) is, that I did not want to touch the actual logic. > I'm personally fine with something roughly like this. Lukas, I'm interested > in your opinion on this one, as I *think* it addresses the issues without > introducing new ones. We could even think about backporting this. > Best regards Tim Düsterhus
