[PATCH] multihome mode doesn't work on FreeBSD

[Dan Nelson]

Socket.c has some #ifdefs to cover both the Linux and BSD extensions for setting both the source and destination IP address when sending UDP packets. The link_socket_write_udp_posix_sendmsg() function uses the wrong length values in the BSD cased, though, which causes sendmsg() to fail with the EINVAL error. The fix is to move the setting of mesg.msg_controllen and cmsg->cmsg_len into the #ifdef blocks.

[Dan Nelson]

use the correct length values for BSD's sendmsg()

[Gert Döring]

At first glance, this looks reasonable. Can you point me at the relevant FreeBSD manpage to read up how the arguments should be?

Integrating this patch is going to be a bit more time consuming, as I'll need to figure out which platforms are using this particular code section (NetBSD, OpenBSD?, ...) and test --multihome on each.

[Dan Nelson]

The FreeBSD manpage describing IP_RECVDSTADDR is readable here: ​http://www.freebsd.org/cgi/man.cgi?query=ip&sektion=4

The CMSG_SPACE macro that I used doesn't seem to have any official FreeBSD documentation, but there's an OpenBSD manpage with some examples: ​http://www.openbsd.org/cgi-bin/man.cgi?query=CMSG_SPACE

All of the BSDs, plus OS X would be using this bit of code I think.

[Gert Döring]

there's also #306 and #348, but *this* is, as far as I can see, purely IPv4 and only on BSDs. The fix looks easy enough, I "just" need to set a test environment to run server tests on BSD before/after applying this.

[Gert Döring]

Hi,

so I spent some time today with testing this, and with the patch applied, all I get is a core dump :-o - because with the patch, when "cmsg = CMSG_FIRSTHDR (&mesg);" runs, it returns NULL - seems to dislike a not-yet-set controllen.

Moving cmsg = CMSG_FIRSTHDR (&mesg); into the respective #ifdef clause, after setting mesg.msg_controllen, makes it work nicely for me on FreeBSD 9.1 (didn't test any other BSD yet).

Also, I've added an ASSERT() to ensure that CMSG_SPACE() actually fits into the structure we're using. It should be safe enough (given that that one is a union that can also hold an IPv6 address) but I don't really feel save enough here.

(JFTR, the CMSG_SPACE and CMSG_LEN macros are actually introduced by the socket API extentions in RFC 2292 - the FreeBSD headers point there, and it's good reading :-) )

Modified patch appended for review...

[Gert Döring]

I'm not sure if anyone ever got the mail from trac about the patch, but as it "worksforme" and looks "obviously correct", and nobody complained about on the openvpn-devel list, I've just committed and pushed it.

Will be in the upcoming 2.3.4 release, and in 2.4

commit 661d914c8732a208580b1eab167255c85da162c9 (master)
commit fc759c2eb4fbf0792d0052966fcf36efebc5bf5d (release/2.3)

I'll leave this ticket open a few more days so you can yell at me if it's not working - will close next friday.

[Gert Döring]

2.3.4 will be released today (git has been tagged on wednesday already). I think it's solved, closing the ticket.

Re-open if you disagree :-)