Unfortunately as the topotests show a fast recovery after failure
detection due to BFD is currently not possible because of the following
issue:
There are multiple scheduling mechanisms within isisd to prevent
overload situations. Regarding our problem these two are important:
* scheduler for regenerating ISIS Link State PDUs scheduler for managing
* consecutive SPF calculations
In fact both schedulers are coupled, the first one triggers the second
one, which again is triggered by isis_adj_state_change (which again is
triggered by a BFD 'down' message). The re-calculation of SPF paths
finally triggers updates in zebra for the RIB.
Both schedulers work as a throttle, e.g. they allow the regeneration of
Link State PDUs or a re-calculation for SPF paths only once within a
certain time interval which is configurable (and by default different!).
This means that a request can go through the first scheduler but might
still be 'stuck' at the second one for a while. Or a request can be
'stuck' at the first scheduler even though the second one is ready. This
also explains the 'random' behaviour one can observe testing since a
'fast' recovery is only possible if both schedulers are ready to process
this request.
Note that the solution in this commit is 'thread safe' in the sense that
both schedulers use the same thread master such that the introduced
flags are only used exactly one time (and one after another) for a
'fast' execution.
Further there are some irritating comments and logs which I partially
removed. They seems to be not valid anymore due to changes in thread
management (or they were never valid in the first place).
Signed-off-by: GalaxyGorilla <sascha@netdef.org>
The tests work with the default settings of BFD meaning that bfdd is
able to recognize a 'down' link after ~900ms so a route recovery should
be visible in the RIB after 1 second.
In the current state only IPv4 is used (when using IPv6
autoconfiguration) within BFD, even though the recovery also affects
IPv6 routes. This is different to the current state of ospfd/ospf6d in
combination with BFD since both IPv4 and IPv6 sessions are used there.
The following topology is used:
+---------+
| |
eth-rt2 (.1) | RT1 | eth-rt3 (.1)
+----------+ 1.1.1.1 +----------+
| | | |
| +---------+ |
| |
| 10.0.2.0/24 |
| |
| eth-rt1 | (.2)
| 10.0.1.0/24 +----+----+
| | |
| | RT3 |
| | 3.3.3.3 |
| | |
(.2) | eth-rt1 +----+----+
+----+----+ eth-rt4 | (.1)
| | |
| RT2 | |
| 2.2.2.2 | 10.0.4.0/24 |
| | |
+----+----+ |
(.1) | eth-rt5 eth-rt3 | (.2)
| +----+----+
| | |
| | RT4 |
| | 4.4.4.4 |
| | |
| +----+----+
| 10.0.3.0/24 eth-rt5 | (.1)
| |
| |
| 10.0.5.0/24 |
| |
| +---------+ |
| | | |
+----------+ RT5 +----------+
eth-rt2 (.2) | 5.5.5.5 | eth-rt4 (.2)
| |
+---------+
Route recovery is tested on RT1. The focus here lies on the two
different routes to RT5. Link failures are generated by taking
down interfaces via the mininet Python interface on RT2 and RT3.
Hence routes are supposed to be adjusted to use RT3 when a link
failure happens on RT2 or vice versa.
Note that only failure recognition and recovery is "fast". BFD
does not monitor a link becoming available again.
Signed-off-by: GalaxyGorilla <sascha@netdef.org>
If the RT changes on a L3VPN route then any leak of this route into
a VRF should be withdrawn.
Extend existing EVPN check for RT change to cover L3VPN routes.
Signed-off-by: Pat Ruddy <pat@voltanet.io>
Instead of waiting for daemons start with `sleep`, start them with the
`-d` parameter so they can release the terminal themselves when ready.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Start logging early everything (including debug) to
`/tmp/topotest/<test>/<node>/<daemon>.{out,err}`.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Handle the duplicated code with a simple conditional: if called from
specialized API use provided daemons configuration, otherwise fallback
to old `Router` own daemon settings.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
When specifying an interface in a pbr-map `set nexthop ..` command
be a bit more lenient about the interface.
a) If the interface does not exist bail on the command
(this is the same)
b) If the interface exists but is in a different vrf
than specified use the vrf it is actually in.
(this is new behavior)
Ticket: CM-30187
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
if we shutdown an interface isisd will delete the adjacencies
on the corresponding circuit, but it will not log the change.
Fix it to make sure that each change is logged. Also specify
the level of the adjacency in the log message, while we are at it.
Signed-off-by: Emanuele Di Pascale <emanuele@voltanet.io>
We have a bunch of code in bgp_vty.c that was passing
to peer_af_flag_modify_vty more than 1 flag at a time.
This was causing the underlying routines to get the
flags wrong. In order to prevent this convert all the
places where we send multiple flags down to this function
to individual flag changes.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Just disable pbr tests on anything less than 4.10.
This has to do with the fact that the arm platform
is not allowing us to install a route into a
non default table using a interface associated
with a vrf.
ip route add default 4.5.6.7 via swp39 table 10000
When swp39 is in a vrf other than default
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The commit:
a798241265
attempted to use sorted master lists to do faster lookups
by using a RB Tree. Unfortunately the original code
was creating a list->cmp function *but* never using it.
If you look at the commit, it clearly shows that the
function listnode_add is used to insert but when you
look at that function it is a tail push.
Fixes: #6573
Namely now this ordering is preserved:
bgp as-path access-list originate-only permit ^$
bgp as-path access-list originate-only deny .*
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
From last addition we can tell that the nexthop-group C is
installed but pbr does not think it is. This failure
has been consistent the last 4-5 runs in master. Lets
add a bit more data gathering to figure out what is going on.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Improve and centralize some logic used to a) compare two
route_entries, and b) to locate a route_entry that matches
a dplane context object that contains the results of a
fib update. We were not rigorous enough in checking routes'
properties, especially when examining connected routes where
we allow multiple route_entries.
Signed-off-by: Mark Stapp <mjs@voltanet.io>
SIGHUP is ostensibly supposed to reload configuration
from a fresh slate. This is currently horribly broken
so much so that bgp just crashes. I see no point
in trying to make this work considering the yang
work coming down the pike.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Now that amd64 dependencies have been removed we can use the correct
architecture specifier for Alpine packaging metadata in order to build
packages for all supported platforms.
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
