On Mon, 20 Apr 2009, Qing Li wrote:
Hi,
Are you really sure backing this change out is the right thing to do ??
yes, I am.
1) if you do not set RTA_GATEWAY but still copy the adress, your kernel
FIB possibly ends up with an incorrect address for the next field,
like RTA_NETMASK, RTA_IFP, ..) for example.
See sys/net/rtsock.c:rt_xaddrs()
2) I had this sample around for a few days now as I have been lerning
more rtsock.c stuff with the help of Kip and finding bugs (ignore
those for a moment):
dopt# ifconfig tun2 create inet 192.0.2.100 192.0.2.101 up
dopt# ifconfig tun2
tun2: flags=8051<UP,POINTOPOINT,RUNNING,MULTICAST> metric 0 mtu 1500
inet6 fe80::2e0:81ff:fe31:db8c%tun2 prefixlen 64 scopeid 0x8
inet 192.0.2.100 --> 192.0.2.101 netmask 0xffffff00
dopt# netstat -rn | grep tun2
192.0.2.101 link#8 UH 0 0 tun2
fe80::%tun2/64 link#8 U tun2
ff01:8::/32 fe80::2e0:81ff:fe31:db8c%tun2 U tun2
ff02::%tun2/32 fe80::2e0:81ff:fe31:db8c%tun2 U tun2
So querying the FIB for the IPv4 (ignoring the v6 for the moment for
simplicity) route (with something handcrafted) I get back:
rtm_msglen 252
rtm_version 5
rtm_type RTM_GET
rtm_index 8
rtm_flags =5<UP,HOST>
rtm_addrs =b3<DST,GATEWAY,IFP,IFA,BRD>
rtm_pid 0
rtm_seq 0
rtm_errno 0
rtm_fmask =0<>
rtm_inits =0<>
rt_metrics rtm_rmx:
rmx_locks 0
rmx_mtu 1500
rmx_hopcount 0
rmx_expire 0
rmx_recvpipe 0
rmx_sendpipe 0
rmx_ssthresh 0
rmx_rtt 0
rmx_rttvar 0
rmx_pksent 0
rmx_filler { 0, 0, 0, 0 }
addresses:
16 (inet) 192.0.2.101
=1<DST>
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
54 (link)
=2<GATEWAY>
0000 36 12 08 00 17 00 00 00 00 00 00 00 00 00 00 00 |6...............|
0010 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
0020 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
0030 00 00 00 00 00 00 |...... |
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
56 (link) tun2
=10<IFP>
0000 38 12 08 00 17 04 00 00 74 75 6e 32 00 00 00 00 |8.......tun2....|
0010 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
0020 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
0030 00 00 00 00 00 00 00 00 |........ |
16 (inet) 192.0.2.100
=20<IFA>
(RTA_BRD)
=80<BRD>
As you can see, we do get an RTA_GATEWAY from the kernel with AF_LINK.
Decoding the sockaddr_dl:
.sdl_len = 0x36 (54)
.sdl_family = 0x12 (18, AF_LINK)
.sdl_index = 8
.sdl_type = 0x17 (23, IFT_PPP)
.sdl_nlen = 0
.sdl_alen = 0
.sdl_slen = 0
.sdl_data = { 0, .. }
So we are clearly getting back an AF_LINK from the kernel and ought to
be able to just stick it back in there as is.
3) To show you that it still works as expected:
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 link#8 UH 0 0 1500 tun2
: route change 192.0.2.101 -mtu 1300
change host 192.0.2.101
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 link#8 UH 0 0 1300 tun2
and:
: route change 192.0.2.101 -link -mtu 1400
change host 192.0.2.101
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 link#8 UH 0 0 1400 tun2
both work (note, even with -link I _think_ route does not pass down
the sockaddr_dl) so that's a bit confusing or rather a bug in route(8)
as we'll see later.
But
: route change 192.0.2.101 192.0.2.100 -mtu 1500
change host 192.0.2.101: gateway 192.0.2.100
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 192.0.2.100 UGH 0 0 1500 tun2
that changed the gateway address, which is correct,
but also set RTF_GATEWAY, which is not as 192.0.2.101 is directly
connected on the PtP interface.
The same would happen if passing down the sockaddr_dl, just that
the gateway address would stay link#8.
4) RTA_GATEWAY != RTF_GATEWAY
The RTF_GATEWAY cannot be set from userspace for an RTM_CHANGE.
Even if given it would be filtered out because of RTF_FMASK.
So what your commit message in r186308 was saying and trying to work
around the RTF_GATEWAY problem by not setting RTA_GATEWAY this only
broke rt_xaddrs() parsing (see 1)) .
The fact that after an RTM_CHANGE with an RTA_GATEWAY set, RTF_GATEWAY
appears, comes from the change in r167797 (ignore the ordering
change, only the rt->rt_flags |= RTF_GATEWAY; there is relevant).
Which is under discussion as my follow-up commit said:
--------------------
r186308, backed out in r191305, already tried to do that, and in addition
ignore AF_LINK types of gateway addresses to work around a problem that
r167797 had introduced on the kernel side always setting RTF_GATEWAY if a
gateway address was passed into the kernel.
The proper solution for this is still under discussion so I am hesitant to
re-add the special AF_LINK treatment for now.
--------------------
One possible solution I had been theoretically pondering first would
have been:
Index: sys/net/rtsock.c
===================================================================
--- sys/net/rtsock.c (revision 190341)
+++ sys/net/rtsock.c (working copy)
@@ -673,25 +673,27 @@ route_output(struct mbuf *m, struct sock
if (info.rti_info[RTAX_GATEWAY] != NULL) {
RT_UNLOCK(rt);
RADIX_NODE_HEAD_LOCK(rnh);
RT_LOCK(rt);
error = rt_setgate(rt, rt_key(rt),
info.rti_info[RTAX_GATEWAY]);
RADIX_NODE_HEAD_UNLOCK(rnh);
if (error != 0) {
RT_UNLOCK(rt);
senderr(error);
}
- rt->rt_flags |= RTF_GATEWAY;
+ if (info.rti_info[RTAX_GATEWAY]->sa_family !=
+ AF_LINK)
+ rt->rt_flags |= RTF_GATEWAY;
}
if (info.rti_ifa != NULL &&
info.rti_ifa != rt->rt_ifa) {
IFAREF(info.rti_ifa);
rt->rt_ifa = info.rti_ifa;
rt->rt_ifp = info.rti_ifp;
}
/* Allow some flags to be toggled on change. */
if (rtm->rtm_fmask & RTF_FMASK)
rt->rt_flags = (rt->rt_flags &
~rtm->rtm_fmask) |
(rtm->rtm_flags & rtm->rtm_fmask);
But I am almost pretty sure that would not be right in all cases either.
Our routing sockets are currently a bit .. uhm .. fragile so a small
change to fix one thing can break other stuff easily.
If we do not want any AF_LINK addresses to be set in RTM_CHANGE as
RTA_GATEWAY anymore, we should first make sure that we cannot add
them in first place and thus the kernel would not report those back.
But I am pretty sure this will simply break all interfaces.
Also it would not easily be possible to fix the example from 3) anymore.
BTW neither:
: route change 192.0.2.101 -mtu 1500
change host 192.0.2.101
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 192.0.2.100 UGH 0 0 1500 tun2
nor:
: route change 192.0.2.101 -link -mtu 1500
change host 192.0.2.101
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 192.0.2.100 UGH 0 0 1500 tun2
does make it a link#%d again. Say what you want, this is a route(8) bug as
a handcrafted update does:
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 192.0.2.100 UGH 0 0 1500 tun2
: sudo ./a.out
: netstat -Wanrf inet | egrep '(^Dest|tun2)'
Destination Gateway Flags Refs Use Mtu Netif
Expire
192.0.2.101 link#8 UGH 0 0 1500 tun2
what it is supposed to do. The only thing left is the RTM_GATEWAY that the
kernel set, unfortunately, which is NOT the fault of ppp.
So the question is, can we somehow clear it? The answer to that is "no".
-----Original Message-----
From: owner-src-committ...@freebsd.org
[mailto:owner-src-committ...@freebsd.org] On Behalf Of Bjoern A. Zeeb
Sent: Monday, April 20, 2009 4:23 AM
To: src-committ...@freebsd.org; svn-src-all@freebsd.org;
svn-src-h...@freebsd.org
Subject: svn commit: r191305 - head/usr.sbin/ppp
Author: bz
Date: Mon Apr 20 11:22:51 2009
New Revision: 191305
URL: http://svn.freebsd.org/changeset/base/191305
Log:
Back out r186308:
in case of AF_LINK, which the kernel still returns for an
RTAX_GATEWAY
as an empty sockaddr_dl in the classic tun<n> case.
Copying the address into the message payload, but not the
RTA_GATEWAY
flag results in rt_xaddrs() in the kernel tripping over
that and parsing
the next attribute set with a flag, i.e. RTA_NETMASK, with
the gateway
address, resulting in bogus route entry.
MFC after: 3 days
Modified:
head/usr.sbin/ppp/route.c
Modified: head/usr.sbin/ppp/route.c
==============================================================
================
--- head/usr.sbin/ppp/route.c Mon Apr 20 10:40:42 2009
(r191304)
+++ head/usr.sbin/ppp/route.c Mon Apr 20 11:22:51 2009
(r191305)
@@ -910,10 +910,8 @@ rt_Update(struct bundle *bundle, const s
p += memcpy_roundup(p, dst, dst->sa_len);
}
- if (gw != NULL && (gw->sa_family != AF_LINK))
- rtmes.m_rtm.rtm_addrs |= RTA_GATEWAY;
+ rtmes.m_rtm.rtm_addrs |= RTA_GATEWAY;
p += memcpy_roundup(p, gw, gw->sa_len);
-
if (mask) {
rtmes.m_rtm.rtm_addrs |= RTA_NETMASK;
p += memcpy_roundup(p, mask, mask->sa_len);
--
Bjoern A. Zeeb The greatest risk is not taking one.
_______________________________________________
svn-src-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"