Re: [PATCH net-next] net: ipv6: put host and anycast routes on device with address

On 8/16/17 4:37 PM, David Ahern wrote:

One nagging difference between ipv4 and ipv6 is host routes for ipv6
addresses are installed using the loopback device or VRF / L3 Master
device. e.g.,

    2001:db8:1::/120 dev veth0 proto kernel metric 256 pref medium
    local 2001:db8:1::1 dev lo table local proto kernel metric 0 pref medium

Using the loopback device is convenient -- necessary for local tx, but
has some nasty side effects, most notably setting the 'lo' device down
causes all host routes for all local IPv6 address to be removed from the
FIB and completely breaks IPv6 networking across all interfaces.

This patch puts FIB entries for IPv6 routes against the device. This
simplifies the routes in the FIB, for example by making dst->dev and
rt6i_idev->dev the same (a future patch can look at removing the device
reference taken for rt6i_idev for FIB entries). For example:

    $ ip -6 r ls table all | grep veth1
    2001:db8:99::/120 dev veth1 proto kernel metric 256 pref medium
    anycast 2001:db8:99:: dev veth1 table local proto kernel metric 0 pref medium
    local 2001:db8:99::1 dev veth1 table local proto kernel metric 0 pref medium

When copies are made on FIB lookups, the cloned route has dst->dev
set to loopback (or the L3 master device). This is needed for the
local Tx of packets to local addresses.

With fib entries allocated against the real network device, the addrconf
code that reinserts host routes on admin up of 'lo' is no longer needed.

Signed-off-by: David Ahern <redacted>
---
 net/ipv6/addrconf.c | 42 ------------------------------------------
 net/ipv6/route.c    | 46 ++++++++++++++++++++++++++++++++++------------
 2 files changed, 34 insertions(+), 54 deletions(-)

DaveM: please drop this one. I found a use case that is failing: UDP
packets to a local linklocal address with no server are not getting the
ICMP unreachable. Will send a v2 when tests complete.