Problem:
When IP1:M1 (local) moved to IP1:M2 (remote-VTEP) bgpd continues to
advertise IP1:M1.
Fix:
Local path del is sent to bgp if the neigh was {local-active||peer-active}.
So path del needs to be called before the sync flags (including peer-active)
are cleared.
Ticket: #2706744
Signed-off-by: Anuradha Karuppiah <anuradhak@nvidia.com>
Add workqueue subqueue for EVPN/VxLAN updates; migrate the
evpn route and remote ES processing from their ZAPI handlers
to the workqueue.
Signed-off-by: Mark Stapp <mjs@voltanet.io>
We are creating 2 hash tables per vni in zebra. Once we start to
scale the number of vni's we start to see some serious memory
usage in zebra. Let's reduce the memory usage at startup
for scale of vni's.
Signed-off-by: Donald Sharp <sharpd@nvidia.com>
This one also needed a bit of shuffling around, but MTYPE_RE is the only
one left used across file boundaries now.
Signed-off-by: David Lamparter <equinox@diac24.net>
When an ES-bond comes out of bypass FRR needs to flush the local MACs learnt
while the bond was in bypass. To do that efficiently local MACs are linked
to the dest-access port. This only happens if the access-port is in
LACP-bypass or if it is non-ES.
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
Feature overview:
=================
A 802.3ad bond can be setup to allow lacp-bypass. This is done to enable
servers to pxe boot without a LACP license i.e. allows the bond to go oper
up (with a single link) without LACP converging.
If an ES-bond is oper-up in an "LACP-bypass" state MH treats it as a non-ES
bond. This involves the following special handling -
1. If the bond is in a bypass-state the associated ES is placed in a
bypass state.
2. If an ES is in a bypass state -
a. DF election is disabled (i.e. assumed DF)
b. SPH filter is not installed.
3. MACs learnt via the host bond are advertised with a zero ESI.
When the ES moves out of "bypass" the MACs are moved from a zero-ESI to
the correct non-zero id. This is treated as a local station move.
Implementation:
===============
When (a) an ES is detached from a hostbond or (b) an ES-bond goes into
LACP bypass zebra deletes all the local macs (with that ES as destination)
in the kernel and its local db. BGP re-sends any imported MAC-IP routes
that may exist with this ES destination as remote routes i.e. zebra can
end up programming a MAC that was perviously local as remote pointing
to a VTEP-ECMP group.
When an ES is attached to a hostbond or an ES-bond goes
LACP-up (out of bypss) zebra again deletes all the local macs in the
kernel and its local db. At this point BGP resends any imported MAC-IP
routes that may exist with this ES destination as sync routes i.e.
zebra can end up programming a MAC that was perviously remote
as local pointing to an access port.
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
Added support for advertising SVI MAC if EVPN-MH is enabled.
In the case of EVPN MH arp replies from an attached server can be sent to
the ES-peer. To prevent flooding of the reply the SVI MAC needs to be
advertised by default.
Note:
advertise-svi-ip could have been used as an alternate way to advertise
SVI MAC. However that config cannot be turned on if SVI IPs are
re-used (which is done to avoid wasting IP addresses in a subnet).
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
If a local-MAC or local-neigh is not active locally it is not sent to BGP.
At this point if BGP rxes a remote route it accepts it and installs in
zebra. Zebra was rejecting BGP's update if it had a higher seq local (inactive)
entry. This would result in bgp and zebra falling out of sync.
In some cases zebra would delete the local-inactive entries in sometime (as
a part of the dplane/kernel garbage collection). This would leave zebra
with missing remote entries (which were still present in bgpd).
This change allows lower-seq BGP updates to overwrite zebra's local entry if
that entry happens to be local-inactive.
Note: This logic was already in use for sync-mac-ip updates. Extended the
same logic to remote-mac-ip updates.
Ticket: CM-31626
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
If a netlink/dp notification is rxed for a neigh without the peer-sync
flag FRR re-installs the entry with the right flags. This change is
needed to handle cases where the dataplane and FRR may fall out of
sync because of neigh learning on the network ports (i.e. via
the VxLAN).
Ticket: CM-30693
The problem was found during VM mobility "torture" tests where 100s
of extended VM moves were done.
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
DAD is not supported currently with EVPN-MH so we turn it off internally
when the first ES config is detected.
PS: Note that when all local ESs are deleted DAD will stay off and
will need to be cleared via a daemon restart.
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
No need to check for n->mac existence as that all paths
leading to this code have n->mac already derefed.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The neigh hold timer was firing after the neigh was deleted resulting
in the following crash -
[
at ./zebra/zebra_evpn_neigh.h:155
at zebra/zebra_evpn_neigh.c:447
at lib/thread.c:1578
at zebra/main.c:488
]
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
Problem commit -
[
b169fd6fd5 zebra: support for MAC-IP sync routes
]
That commit had accidentally replaced a mac-ip del to bgp with a mac
del (consequence of a bad cut-paste).
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
Reverting probing of neigh entry. There is a timing where
probe and remote macip add request comes at the same time resulting
in neigh to remain in local state event though it should be remote.
In mobility case, the host moves to remote VTEP, first MAC only type-2
route is received which triggers a PROBE of neighs (associated to MAC).
PROBE request can go via network port to remote VTEP.
PROBE request picks up local neigh with MAC entry's outgoing port is
remote VTEP tunnel port.
The PROBE reply and MAC-IP (containing IP) almost comes same time at
DUT.
DUT first processes remote macip and installs neigh as remote.
Followed by receives neigh as REACHABLE which marks neigh as LOCAL.
FRR does have BPF filter which does not allow its own netlink request
to receive. Otherwise frr's request to program neigh as remote can move
neigh from local to remote.
Though ordering can not be guranteed that REACHABLE (PROBE's repsonse)
can come at anytime and move it to LOCAL.
This fix would not suffice the needs of converging LOCAL inactive neighs
to remove from DB. As mobility draft sugges to PROBE local neigh when
MAC moves to remote but it is not working with current framework.
Ticket:CM-22864
This reverts commit 44bc8ae550
Signed-off-by: Chirag Shah <chirag@cumulusnetworks.com>
extract the neighbor uninstall part of
zebra_vxlan_handle_kernel_neigh_del into a new function
zebra_evpn_neigh_del_ip in zebra_evpn_neigh.c.
Signed-off-by: Pat Ruddy <pat@voltanet.io>
extract the neighbor uninstall part of process_remote_macip_add
into a new function zebra_evpn_neigh_remote_uninstall in
zebra_evpn_neigh.c.
Signed-off-by: Pat Ruddy <pat@voltanet.io>
extract the neighbor part of process_remote_macip_add into a new
function zebra_evpn_neigh_gw_macip_add in zebra_evpn_neigh.c.
Signed-off-by: Pat Ruddy <pat@voltanet.io>
extract the neighbor part of process_remote_macip_add into a new
function process_neigh_remote_macip_add in zebra_evpn_neigh.c.
Signed-off-by: Pat Ruddy <pat@voltanet.io>
clone zebra_vxlan.c to create a file zebra_evpn_neigh.c for neighbor
dB functions whilst retaining the history of zebra_vxlan.c
Signed-off-by: Pat Ruddy <pat@voltanet.io>
