Regular IS-IS will flood any LSP updates out to all circuits except the
one where it was received on. This is done in `lsp_flood`.
Change `lsp_flood` for fabricd to use the optimized flooding algorithm
instead.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric uses a list of neighbors and neighbors neighbors to calculate
a set of designated reflooders.
While the draft prescribes that these lists should be built whenever an
LSP needs to be flooded, this implementation opted to build them only
when we ran an spf, given that they will only change when the topology
changes.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric requires knowledge of the first two hops on each path
calculated by spf to implement its flooding optimization. Extend the
hopcount-spf to build such a datastructure.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric makes use of flooding scope LSPs to reduce the amount of
reflooding caused by the update process. Implement transmission and
reception of such PDUs.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
Before this commit, isisd/fabricd maintained a bitfield for each LSP
to track the SRM bit for each circuit, which specifies whether an LSP
needs to be sent on that circuit. Every second, it would scan over all
LSPs in `lsp_tick` and queue them up for transmission accordingly.
This design has two drawbacks: a) it scales poorly b) it adds
unacceptable latency to the update process: each router takes a random
amount of time between 0 and 1 seconds to forward an update. In a
network with a diamter of 10, it might already take 10 seconds for an
update to traverse the network.
To mitigate this, a new design was chosen. Instead of tracking SRM in a
bitfield, have one tx_queue per circuit and declare that an LSP is in
that queue if and only if it would have SRM set for that circuit.
This way, we can track SRM similarly as we did before, however, on
insertion into the LSP queue, we can add a timer for (re)transmission,
alleviating the need for a periodic scan with LSP tick and reducing the
latency for forwarding of updates.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
To avoid passing of traffic via leaf nodes in the fabric, OpenFabric
specifies that all links towards tier 0 nodes should be advertised with
a very high metric.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
If an OpenFabric router doesn't have its tier number configured
manually, try to execute the fabric locality calculation algorithm
whenever we have run spf.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
While OpenFabric calculates most tier numbers automatically by the
fabric locality calculation algorithm, that algorithm requires two
systems to be manually configured as tier 0, so it has reference points.
Also, completely manual configuration is possible.
To support this, introduce appropriate CLI commands and flood the
configured information.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
To flood the tier calculated by the fabric locality detection,
OpenFabric makes use of TLV 150, defined in
draft-shen-isis-spine-leaf-ext-06, so add support for that TLV.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
By moving the spf datastructures to a header, fabricd can access the
results of the spf run for flooding optimization or fabric locality
calculation.
While this was deemed a sensible choice in this case, when compared with
the option of adding a lot of OpenFabric specific code to isis_spf.c,
the datastructures should still not be accessed randomly all over the
code base. To make this more clear, the new header was called
isis_spf_private.h (Think of a friend class)
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric uses an spf with the metric for all links set to one,
both for flooding optimization and for fabric locality detection.
So extend isisd's spf code to allow running it with such a metric
and have it run whenever normal spf runs.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric changes IS-IS's initial database synchronization. While
regular IS-IS will simultaneuously exchange LSPs with all neighboring
routers during startup, this is considered too much churn for a densely
connected fabric.
To mitigate this, OpenFabric prescribes that a router should only
bring up an adjacency with a single neighbor and perform a full
synchronization with that neighbor, before bringing up further
adjacencies.
This is implemented by having a field `initial_sync_state` in the
fabricd datastructure which tracks whether an initial sync is still
pending, currently in progress, or complete.
When an initial sync is pending, the state will transition to the
in-progress state when the first IIH is received.
During this state, all IIHs from other routers are ignored. Any
IIHs transmitted on any link other than the one to the router with
which we are performing the initial sync will always report the far
end as DOWN in their threeway handshake state, avoiding the formation of
additional adjacencies.
The state will be left if all the SRM and SSN flags on the
initial-sync circuit are cleared (meaning that initial sync has
completed). This is checked in `lsp_tick`. When this condition occurrs,
we progress to the initial-sync-complete state, allowing other
adjacencies to form.
The state can also be left if the initial synchronization is taking too
long to succeed, for whatever reason. In that case, we fall back to the
initial-sync-pending state and will reattempt initial synchronization
with a different neighbor.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
OpenFabric specifies that it should always be run with wide metrics via
P2P links and only as Level-2. Implement this as default and remove all
the knobs from fabricd which allow other configuration.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
The 'no ip router isis' command would incorrectly output the afi if the
area to delete does not exist. Make it output the area name instead.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
Remove isis_vty.c and create three new files isis_vty_common.c,
isis_vty_fabricd.c and isis_vty_isisd.c which are built into both
daemons, only fabricd and only isisd, respectively.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
fabricd is built using the sources of isisd. To allow differentiation
in the code, -DFABRICD=1 is added to its preprocessor flags.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
The Vrf aliases can be known with a specific hook. That hook will then,
from zebra propagate the information to the relevant zapi clients.
The registration hook function is the same for all daemons.
Signed-off-by: Philippe Guibert <philippe.guibert@6wind.com>
* Use the correct license header
* Stop headers from including themselves
* Use uniform relative include conventions
* Ensure that sources include what they use
* Turn off clang-format around struct array blocks
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
There is no need to check for failure of a ALLOC call
as that any failure to do so will result in a assert
happening. So we can safely remove all of this code.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
This fixes multiple issues and inefficiencies regarding the usage of
route_tables in isis_route.c and removes some memory leaks.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
Take the source-prefix sub-TLV into consideration when running SPF
and support creation/deletion of dst-src routes as result.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
Instead of using the address family to determine which spftree structure
should be used, specify it explicitly. With the advent of ipv6 dst-src
routing, the tree cannot be uniquely determined from the family.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
Have an array of spftrees instead of a separate spftree and an
spftree6 for which all the code gets duplicated.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
This addresses two issues for L1L2 operation:
a) If an L1 route has ROUTE_ACTIVE unset and an L2 route for the same
destination has ROUTE_ACTIVE set, isisd would still put the L1 route
into the merged table. This causes the route for the destination to
get uninstalled from zebra until the next SPF run, which is incorrect.
To fix this, look at the ROUTE_ACTIVE flag and allow L2 routes to win
against L1 routes, when the L1 has ROUTE_ACTIVE unset.
b) If an L1 route wins against an existing L2 route, the ZEBRA_SYNCED
flag would remain on the L2 route. This leads to the problem that when
the L1 route disappears again, the L2 doesn't get reinstalled, since
isisd assumes it's already in the RIB because ZEBRA_SYNCED is set.
Solve this by clearing ZEBRA_SYNCED on L2 routes, if they lose against
an L1 route.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
There was an off-by-one error in redist_delete, so that routes redistributed
into level-2 could never be withdrawn.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
As isisd's route_tables are directly related to spf trees, move
the route tables into the spftree instead of maintaining them
alongside of the spftrees.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
isisd verifies whether the neighboring IPv4 addresses overlap with its own
unless the interface is running in unnumbered mode. If no overlap is found
and IPv6 is also not enabled, IIHs will be ignored.
Add a debug message for this case, to avoid people wondering why adjacencies
are not coming up.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
When isis_sock_init fails in isis_circuit_up, isis_circuit_down would
be called to cancel timers which were scheduled. However
isis_circuit_down would immediately return, since the state had not been
changed to 'UP' yet.
Fix this by having isis_circuit_down always cancel all the timers.
Signed-off-by: Christian Franke <chris@opensourcerouting.org>
This correction fixes three bugs detected by Clang scan:
Bug Group: Logic error
Bug Type: Dereference of null pointer
File: bgpd/bgp_evpn.c
Function: bgp_evpn_unconfigure_import_rt_for_vrf
Line: 4246
File: isisd/isis_spf.c
Function: isis_print_paths
Line: 69 (two bugs of same type in one line)
Signed-off-by: F. Aragon <paco@voltanet.io>
This correction fixes two bugs detected by Clang scan:
Bug Group: Dead store
Bug Type: Dead assignment
File: zebra/kernel_netlink.c
Function: netlink_parse_extended_ack
Line: 548
Bug Type: Dead increment
File: isisd/isis_lsp.c
Function: lsp_bits2string
Line: 625
Signed-off-by: F. Aragon <paco@voltanet.io>
issue found by cppcheck
[isisd/dict.c:1320] -> [isisd/dict.c:1065]:
(warning) Either the condition '!dn' is redundant or
there is possible null pointer dereference: newnode.
[isisd/dict.c:1320] -> [isisd/dict.c:1068]:
(warning) Either the condition '!dn' is redundant or
there is possible null pointer dereference: newnode.
Signed-off-by: Ilya Shipitsin <chipitsine@gmail.com>
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The following types are nonstandard:
- u_char
- u_short
- u_int
- u_long
- u_int8_t
- u_int16_t
- u_int32_t
Replace them with the C99 standard types:
- uint8_t
- unsigned short
- unsigned int
- unsigned long
- uint8_t
- uint16_t
- uint32_t
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
The circuit->area value is always true in every code path
to isis_circuit_af_set( isis_vty.c ). Therefore was_enabled
will always be true.
If was_enabled ever became false then the area->ip_circuits
and area->ipv6_circuits lines would segfault.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The v4 and v6 prefixes were created but not deleted on
shutdown properly.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
(cherry picked from commit 25b1001dc9)
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Allow other supported Operating Systems (OS) to use file descriptor
polling, instead of doing timed fd checks. This should improve
performance greatly on modern OSes (e.g. that support polling on
filtered sockets).
The known OS that doesn't support this is FreeBSD < 5.0, but even then
FRR doesn't compile in these versions. OSes using DLPI method (e.g
Solaris) does not support select()/poll()ing fds as well, so it will be
disabled for it.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
None of these variables can actually be used before being initialized,
but unfortunately some old compilers are not smart enough to detect that.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This reverts commit d9e5b00955.
ETH_ALEN is not equivalent to ETHER_HDR_LEN. ETHER_HDR_LEN is the
definition for the length of the full ethernet header, meanwhile
ETH_ALEN is just the length of the ethernet address.
Here is the OpenBSD/FreeBSD definition for ETHER_HDR_LEN:
#define ETHER_ADDR_LEN 6 /* Ethernet address length */
#define ETHER_TYPE_LEN 2 /* Ethernet type field length */
#define ETHER_HDR_LEN ((ETHER_ADDR_LEN * 2) + ETHER_TYPE_LEN)
(OpenBSD's definition is at net/if_ether.h, FreeBSD's is at
net/ethernet.h)
Linux definitions can be found at: net/ethernet.h
#define ETHER_ADDR_LEN ETH_ALEN /* size of ethernet addr */
#define ETHER_TYPE_LEN 2 /* bytes in type field */
#define ETHER_HDR_LEN ETH_HLEN /* total octets in header */
net/if_ether.h:
#define ETH_HLEN 14 /* Total octets in header. */
#define ETH_ALEN 6 /* Octets in one ethernet addr */
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Add a timestamp information for level 2 circuits, otherwise if the
circuit is marked as already processed on level 1 we will not process
level 2 areas.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Use the thread cached clock to use as start time. It will save a call to
clock_gettime() and also provide a more 'accurate' time measurement from
the start of the procedure.
The $Id: lines would allow code kept in cvs to substitute
the file version upon checkout. Since we are not using
cvs there is no need to keep these lines anymore.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Allow the higher level protocol to specify if it would
like to receive notifications about it's routes that
it has installed.
I've purposely made it part of zclient_new_notify because
we need to track the routes on a per daemon basis only.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
This improves code readability and also future-proofs our codebase
against new changes in the data structure used to store interfaces.
The FOR_ALL_INTERFACES_ADDRESSES macro was also moved to lib/ but
for now only babeld is using it.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Performance tests showed that, when running on a system with a large
number of interfaces, some daemons would spend a considerable amount
of time in the if_lookup_by_index() function. Introduce a new rb-tree
to solve this problem.
With this change, we need to use the if_set_index() function whenever
we want to change the ifindex of an interface. This is necessary to
ensure that the 'ifaces_by_index' rb-tree is updated accordingly. The
return value of all insert/remove operations in the interface rb-trees
is checked to ensure that an error is logged if a corruption is
detected.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
IFINDEX_DELETED is not necessary anymore as we moved from a global
list of interfaces to a list of interfaces per VRF.
This reverts commit 84361d615.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This is an important optimization for users running FRR on systems with
a large number of interfaces (e.g. thousands of tunnels). Red-black
trees scale much better than sorted linked-lists and also store the
elements in an ordered way (contrary to hash tables).
This is a big patch but the interesting bits are all in lib/if.[ch].
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
tlvs is not used after extracting the last TLV. So,
let's tell to the compiler that it is expected.
Signed-off-by: Vincent Jardin <vincent.jardin@6wind.com>
