Routing protocols are allowed ( and even encouraged ) to modify
the flags that influence the nexthop tracking. As such when
we modify the tracking of a nexthop to go from, say, connected force
or not we must re-evaluate the nexthop and send the results
up to the interested parties.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
After we have evaluated the rnh for an import-check type
and we copy the re then we know that the state has changed
and we should be notifying the end user about it.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
LSP processing was a zvrf flag based upon a connected route
coming or going. But this did not allow us to know
that we should do lsp processing other than after the meta-queue
processing was finished.
Eventually we moved meta-queue processing of do_nht_processing
to after the dataplane sent the main pthread some results.
This of course left us with a timing hole where if a connected
route came in and we received a data plane response *before*
the meta queue was processed we would not do the work as necessary.
Move the lsp processing to a flag off of the rib_dest_t. If it
is marked then we need to process lsps.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Add a detailed debugging command for NHT tracking and add
the detailed output to the log about why we make some decisions
that we are. I tried to model this like the rib processing
detailed debugs that we added a few months back.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Currently nexthop tracking is performed for all nexthops that
are being tracked after a group of contexts are passed back
from the data plane for post install processing.
This is inefficient and leaves us sending nexthop tracking
changes at an accelerated pace, when we think we've changed
a route. Additionally every route change will cause us
to relook at all nexthops we are tracking irrelevant if
they are possibly related to the route change or not.
Let's modify the code base to track the rnh's off of the rib
table's rn, `rib_dest_t`. So after we process a node, install
it into the data plane, in rib_process_result we can
look at the `rib_dest_t` associated with the rn and see that
a nexthop depended on this route node. If so, refigure it.
Additionally we will store rnh's that are not resolved on the
0.0.0.0/0 nexthop tracking list. As such when a route node
changes we can quickly walk up the rib tree and notice that
it needs to be reprocessed as well.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Add a default route_node for our routing tables. This will allow us
to know that we can hang data off the default route for processing.
We will be hanging the nexthop tracking data structures off the rib_dest_t
so that we can know which nexthops we need to handle. Effectively
nexthops that we are tracking that are unresolved will be stored on the
default route. When something changes in the rib tree we can
work up the rn->parent pointer checking for nexthops we need to re-evaluate.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The resolved_route is the prefix we are using in the routing table
to resolve this particular nexthop we are tracking. Add code
to better track it's change.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The prn value as passed in may be NULL as such do not
allow it to be derefed (even though it works now).
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
We have several route types KERNEL and CONNECT that are handled via special
case in the code. This was causing a lot of work keeping the two different
classes of route types as special(SYSTEM OR NOT). Put the dplane
in charge of the code that sets the bits for signalling route install/failure.
This greatly simplifies the code calling path and makes all route types
be handled exactly the same. Additionaly code that we want to run
post data plane install can just work as per normal then, instead
of having to know we need to run it when we have a special type
of route.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com.
When we get a route install failure from the kernel, actually
indicate in the rib the status of the routes.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
When switching routes from one route type to another actually
unset the old route as enqueued.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
When shutting down, the individual vrf's own the shutdown of the table
and subsuquent removal from the routes from the kernel.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
When shutting down and we have a very large table to shutdown
and after we've intentionally closed all the client connections
close the zebra zserv client socket.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Fix the macros for reading NLA attribute info
from an extended error ack. We were processing the data
using route attributes (rtattr) which is identical in size
to nlattr but probably should not be used.
Further, we were incorrectly calculating the length of the
inner netlink message that cause the error. We have to read
passed that in order to access all the nlattr's.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
The 'sho ip route summary' and 'sho ip route summary <prefix>'
paths used different definitions of a 'fib' route. Use
the route-entry 'INSTALLED' flag in both places.
Signed-off-by: Mark Stapp <mjs@voltanet.io>
Duplicate address detection and recovery was relying on the l2-vni backptr
in the neighbor entry which was simply not initialized resulting in
a NULL pointer access in a setup with dup-addressed VMs -
VM1:{IP1,M1} and VM2:{IP1,M2}
Call stack:
(gdb) bt 6
at lib/sigevent.c:249
nbr=nbr@entry=0x559347f901d0, vtep_ip=..., vtep_ip@entry=..., do_dad=do_dad@entry=true,
is_dup_detect=is_dup_detect@entry=0x7ffc7f6be59f, is_local=is_local@entry=true)
at ./lib/ipaddr.h:86
ip=0x7ffc7f6be6f0, ifp=0x559347f901d0, zvni=0x559347f86800) at zebra/zebra_vxlan.c:3152
(More stack frames follow...)
(gdb) p nbr->zvni
$8 = (zebra_vni_t *) 0x0 <<<<<<<<<<<<<<<<<<<<
(gdb)
Signed-off-by: Anuradha Karuppiah <anuradhak@cumulusnetworks.com>
System Routes if received over the netlink bus in a
specific pattern that causes an update operation for that
route in zebra can leave the dest->selected_fib pointer NULL,
while having the ZEBRA_FLAG_SELECTED flag set. Specifically
one way to achieve this is to do this:
`ip addr del 4.5.6.7/32 dev swp1 ; ip addr add 4.5.6.7/32 dev swp1 metric 9`
Why is this a big deal?
Because nexthop tracking is looking at ZEBRA_FLAG_SELECTED to
know if we can use a route, while nexthop active checking uses
dest->selected_fib.
So imagine we have bgp registering a nexthop. nexthop tracking in
the above case will be able to choose the 4.5.6.7/32 route
if that is what the nexthop is, due to the ZEBRA_FLAG_SELECTED being
properly set. BGP then allows the peers connection to come up and we
install routes with a 4.5.6.7 nexthop. The rib processing for route
installation will then look at the 4.5.6.7 route see no
dest->selected_fib and then start walking up the tree to resolve
the route. In our case we could easily hit the default route and be
unable to resolve the route. Which then becomes inactive in the
rib so we never attempt to install it.
This commit fixes this problem because when the rib_process decides
that we need to update the fib( ie replace old w/ new ), the
replacement with new was not setting the `dest->selected_fib` pointer
to the new route_entry, when the route was a system route.
Ticket: CM-24203
Signed-off-by: Donald Sharp <sharpd@cumulusnetworkscom>
The dest->selected_fib should be reported in json output
so that we can debug subtle conditions a bit better in the
future.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Cleaup the rnh tables on shutdown before we cleanup tables. As that
this will remove any need to do rnh processing as part of shutdown.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
When we get a neighbor entry in zebra we start processing it.
Let's add some additional debugs to the processing so that when
it bails out and we don't use the data, we know the reason.
This should help in debugging the problems from why bgp does
not appear to have data associated with a neighbor entry
in the kernel.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The check for an entry being NUD_PERMANENT has already been done
there is no need to do it twice.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Use const in the accessors for pseudowire nhlfe data; pull
that through the kernel-facing apis that use that data.
Signed-off-by: Mark Stapp <mjs@voltanet.io>
In prep for adding nexthop info for pws, rename the accessor
for the pw destination. Add a nexthop-group to the pw
data in the dataplane module.
Signed-off-by: Mark Stapp <mjs@voltanet.io>
The current definition of an unnumberd interface as an interface with a
/32 IPv4 is too restrictive, especially for EVPN symmetric routing since
commit 2b83602b2 "*: Explicitly mark nexthop of EVPN-sourced routes as
onlink".
It removes the bypass check wether the nexthop is an EVPN VTEP, and
relies on the SVI to be unnumberd to bypass the gateway lookup. While
this works great if the SVI has an IP, it might not, and the test falls
flat and EVPN type 5 routes are not installed into the RIB.
Sample interface setup, where vxlan-blue is the L3VNI and br-blue the
SVI:
+----------+
| |
| vrf-blue |
| |
+---+--+---+
| |
+-------+ +-----------+
| |
+----+----+ +---------+---------+
| | | br1 |
| br-blue | | 10.0.0.1/24 |
| | +-+-------+-------+-+
+----+----+ | | |
| | | |
+-----+------+ +-----+--+ +--+---+ +-+----+
| | | | | | | |
| vxlan-blue | | vxlan1 | | eth1 | | eth2 |
| | | | | | | |
+------------+ +--------+ +------+ +------+
For inter-VNI routing, the SVI has no reason to have an IP, but it still
needs type-5 routes from remote VTEPs.
This commit expands the definition of an unnumberd interface to an
interface having a /32 IPv4 or no IPv4 at all.
Signed-off-by: Tuetuopay <tuetuopay@me.com>
