Re: [PATCHv4 net] ipv6: do not match device when remove source route
From: Hangbin Liu <hidden>
Date: 2023-08-03 04:15:48
On Tue, Aug 01, 2023 at 02:51:52PM +0300, Ido Schimmel wrote:
On Tue, Jul 25, 2023 at 06:21:37PM +0800, Hangbin Liu wrote:quoted
diff --git a/net/ipv6/route.c b/net/ipv6/route.c index 64e873f5895f..44e980109e30 100644 --- a/net/ipv6/route.c +++ b/net/ipv6/route.c@@ -4590,10 +4590,10 @@ static int fib6_remove_prefsrc(struct fib6_info *rt, void *arg) struct net_device *dev = ((struct arg_dev_net_ip *)arg)->dev; struct net *net = ((struct arg_dev_net_ip *)arg)->net; struct in6_addr *addr = ((struct arg_dev_net_ip *)arg)->addr; + u32 tb6_id = l3mdev_fib_table(dev) ? : RT_TABLE_MAIN; - if (!rt->nh && - ((void *)rt->fib6_nh->fib_nh_dev == dev || !dev) && - rt != net->ipv6.fib6_null_entry && + if (rt != net->ipv6.fib6_null_entry && + rt->fib6_table->tb6_id == tb6_id && ipv6_addr_equal(addr, &rt->fib6_prefsrc.addr)) { spin_lock_bh(&rt6_exception_lock); /* remove prefsrc entry */@@ -4611,7 +4611,9 @@ void rt6_remove_prefsrc(struct inet6_ifaddr *ifp) .net = net, .addr = &ifp->addr, }; - fib6_clean_all(net, fib6_remove_prefsrc, &adni); + + if (!ipv6_chk_addr_and_flags(net, adni.addr, adni.dev, true, 0, IFA_F_TENTATIVE))Setting 'skip_dev_check' to true is problematic since when a link-local address is deleted from a device, it should be removed as the preferred source address from routes using the device as their nexthop device, even if this address is configured on other devices. You can't configure a route with a link-local preferred source address if the address is not configured on the nexthop device:
Thanks for letting me know another case I'm not aware...
quoted hunk ↗ jump to hunk
Setting 'skip_dev_check' to false will solve this problem: But will create another problem where when such an address is deleted it also affects routes that shouldn't be affected: So, I think we need to call ipv6_chk_addr() from rt6_remove_prefsrc() to be consistent with the addition path in ip6_route_info_create():diff --git a/net/ipv6/route.c b/net/ipv6/route.c index 56a55585eb79..e7e2187bff0c 100644 --- a/net/ipv6/route.c +++ b/net/ipv6/route.c@@ -4591,11 +4591,13 @@ static int fib6_remove_prefsrc(struct fib6_info *rt, void *arg) struct net_device *dev = ((struct arg_dev_net_ip *)arg)->dev; struct net *net = ((struct arg_dev_net_ip *)arg)->net; struct in6_addr *addr = ((struct arg_dev_net_ip *)arg)->addr; + u32 tb6_id = l3mdev_fib_table(dev) ? : RT_TABLE_MAIN; if (!rt->nh && - ((void *)rt->fib6_nh->fib_nh_dev == dev || !dev) && rt != net->ipv6.fib6_null_entry && - ipv6_addr_equal(addr, &rt->fib6_prefsrc.addr)) { + rt->fib6_table->tb6_id == tb6_id && + ipv6_addr_equal(addr, &rt->fib6_prefsrc.addr) && + !ipv6_chk_addr(net, addr, rt->fib6_nh->fib_nh_dev, 0)) { spin_lock_bh(&rt6_exception_lock); /* remove prefsrc entry */ rt->fib6_prefsrc.plen = 0;ipv6_chk_addr() is not cheap, but it's only called for routes that match the previous criteria. With the above patch, the previous test cases now work as expected: There is however one failure in the selftest: Which is basically: # ip link add name dummy1 up type dummy # ip link add name dummy2 up type dummy # ip link add red type vrf table 1111 # ip link set dev red up # ip link set dummy2 vrf red # ip -6 address add dev dummy1 2001:db8:104::12/64 # ip -6 address add dev dummy2 2001:db8:104::12/64 # ip -6 route add 2001:db8:106::/64 dev lo src 2001:db8:104::12 # ip -6 route add vrf red 2001:db8:106::/64 dev lo src 2001:db8:104::12 # ip -6 address del dev dummy2 2001:db8:104::12/64 # ip -6 route show vrf red | grep "src 2001:db8:104::12" 2001:db8:106::/64 dev lo src 2001:db8:104::12 metric 1024 pref medium I'm not sure it's realistic to expect the source address to be removed when the address is deleted from dummy2, given that user space was only able to configure the route because the address was available on dummy1 in the default vrf: # ip link add name dummy1 up type dummy # ip link add name dummy2 up type dummy # ip link add red type vrf table 1111 # ip link set dev red up # ip link set dummy2 vrf red # ip -6 address add dev dummy2 2001:db8:104::12/64 # ip -6 route add vrf red 2001:db8:106::/64 dev lo src 2001:db8:104::12 Error: Invalid source address.
OK.. Another difference with IPv4, which could add this route directly. e.g. ip addr add dev dummy2 172.16.104.13/24 ip route add vrf red 172.16.107.0/24 dev lo src 172.16.104.13 For the IPv6 part, if we remove dummy1 addr, should we remove the src route in vrf since user only able to config the route when the addr available on dummy1? My current patch, and with yours, will keep the src route in vrf.. + ip link add name dummy1 up type dummy + ip link add name dummy2 up type dummy + ip link add red type vrf table 1111 + ip link set dev red up + ip link set dummy2 vrf red + ip -6 address add dev dummy1 2001:db8:104::12/64 + ip -6 address add dev dummy2 2001:db8:104::12/64 + ip -6 route add 2001:db8:106::/64 dev lo src 2001:db8:104::12 + ip -6 route add vrf red 2001:db8:106::/64 dev lo src 2001:db8:104::12 + ip -6 address del dev dummy1 2001:db8:104::12/64 + ip -6 route show vrf red 2001:db8:104::/64 dev dummy2 proto kernel metric 256 pref medium 2001:db8:106::/64 dev lo src 2001:db8:104::12 metric 1024 pref medium fe80::/64 dev dummy2 proto kernel metric 256 pref medium multicast ff00::/8 dev dummy2 proto kernel metric 256 pref medium + ip -6 route show ::1 dev lo proto kernel metric 256 pref medium 2001:db8:106::/64 dev lo metric 1024 pref medium
Anyway, given that this patch does not fix a regression and the on-going discussion around the semantics, I suggest to target future versions at net-next.
OK, I will. Thanks Hangbin