summaryrefslogtreecommitdiffstats
path: root/doc/user/flowspec.rst
diff options
context:
space:
mode:
Diffstat (limited to 'doc/user/flowspec.rst')
-rw-r--r--doc/user/flowspec.rst376
1 files changed, 376 insertions, 0 deletions
diff --git a/doc/user/flowspec.rst b/doc/user/flowspec.rst
new file mode 100644
index 0000000..faf5973
--- /dev/null
+++ b/doc/user/flowspec.rst
@@ -0,0 +1,376 @@
+.. _flowspec:
+
+Flowspec
+========
+
+.. _features-of-the-current-implementation-flowspec:
+
+Overview
+---------
+
+Flowspec introduces a new :abbr:`NLRI (Network Layer Reachability Information)`
+encoding format that is used to distribute traffic rule flow specifications.
+Basically, instead of simply relying on destination IP address for IP prefixes,
+the IP prefix is replaced by a n-tuple consisting of a rule. That rule can be a
+more or less complex combination of the following:
+
+
+- Network source/destination (can be one or the other, or both).
+- Layer 4 information for UDP/TCP: source port, destination port, or any port.
+- Layer 4 information for ICMP type and ICMP code.
+- Layer 4 information for TCP Flags.
+- Layer 3 information: DSCP value, Protocol type, packet length, fragmentation.
+- Misc layer 4 TCP flags.
+
+Note that if originally Flowspec defined IPv4 rules, this is also possible to use
+IPv6 address-family. The same set of combinations as defined for IPv4 can be used.
+
+A combination of the above rules is applied for traffic filtering. This is
+encoded as part of specific BGP extended communities and the action can range
+from the obvious rerouting (to nexthop or to separate VRF) to shaping, or
+discard.
+
+The following IETF drafts and RFCs have been used to implement FRR Flowspec:
+
+- :rfc:`5575`
+- [Draft-IETF-IDR-Flowspec-redirect-IP]_
+- [Draft-IETF-IDR-Flow-Spec-V6]_
+
+.. _design-principles-flowspec:
+
+Design Principles
+-----------------
+
+FRR implements the Flowspec client side, that is to say that BGP is able to
+receive Flowspec entries, but is not able to act as manager and send Flowspec
+entries.
+
+Linux provides the following mechanisms to implement policy based routing:
+
+- Filtering the traffic with ``Netfilter``.
+ ``Netfilter`` provides a set of tools like ``ipset`` and ``iptables`` that are
+ powerful enough to be able to filter such Flowspec filter rule.
+
+- using non standard routing tables via ``iproute2`` (via the ``ip rule``
+ command provided by ``iproute2``).
+ ``iproute2`` is already used by FRR's :ref:`pbr` daemon which provides basic
+ policy based routing based on IP source and destination criterion.
+
+Below example is an illustration of what Flowspec will inject in the underlying
+system:
+
+.. code-block:: shell
+
+ # linux shell
+ ipset create match0x102 hash:net,net counters
+ ipset add match0x102 32.0.0.0/16,40.0.0.0/16
+ iptables -N match0x102 -t mangle
+ iptables -A match0x102 -t mangle -j MARK --set-mark 102
+ iptables -A match0x102 -t mangle -j ACCEPT
+ iptables -i ntfp3 -t mangle -I PREROUTING -m set --match-set match0x102
+ src,dst -g match0x102
+ ip rule add fwmark 102 lookup 102
+ ip route add 40.0.0.0/16 via 44.0.0.2 table 102
+
+For handling an incoming Flowspec entry, the following workflow is applied:
+
+- Incoming Flowspec entries are handled by *bgpd*, stored in the BGP RIB.
+- Flowspec entry is installed according to its complexity.
+
+It will be installed if one of the following filtering action is seen on the
+BGP extended community: either redirect IP, or redirect VRF, in conjunction
+with rate option, for redirecting traffic. Or rate option set to 0, for
+discarding traffic.
+
+According to the degree of complexity of the Flowspec entry, it will be
+installed in *zebra* RIB. For more information about what is supported in the
+FRR implementation as rule, see :ref:`flowspec-known-issues` chapter. Flowspec
+entry is split in several parts before being sent to *zebra*.
+
+- *zebra* daemon receives the policy routing configuration
+
+Policy Based Routing entities necessary to policy route the traffic in the
+underlying system, are received by *zebra*. Two filtering contexts will be
+created or appended in ``Netfilter``: ``ipset`` and ``iptable`` context. The
+former is used to define an IP filter based on multiple criterium. For
+instance, an ipset ``net:net`` is based on two ip addresses, while
+``net,port,net`` is based on two ip addresses and one port (for ICMP, UDP, or
+TCP). The way the filtering is used (for example, is src port or dst port
+used?) is defined by the latter filtering context. ``iptable`` command will
+reference the ``ipset`` context and will tell how to filter and what to do. In
+our case, a marker will be set to indicate ``iproute2`` where to forward the
+traffic to. Sometimes, for dropping action, there is no need to add a marker;
+the ``iptable`` will tell to drop all packets matching the ``ipset`` entry.
+
+Configuration Guide
+-------------------
+
+In order to configure an IPv4 Flowspec engine, use the following configuration.
+As of today, it is only possible to configure Flowspec on the default VRF.
+
+.. code-block:: frr
+
+ router bgp <AS>
+ neighbor <A.B.C.D> remote-as <remoteAS>
+ neighbor <A:B::C:D> remote-as <remoteAS2>
+ address-family ipv4 flowspec
+ neighbor <A.B.C.D> activate
+ exit
+ address-family ipv6 flowspec
+ neighbor <A:B::C:D> activate
+ exit
+ exit
+
+You can see Flowspec entries, by using one of the following show commands:
+
+.. clicmd:: show bgp ipv4 flowspec [detail | A.B.C.D]
+
+.. clicmd:: show bgp ipv6 flowspec [detail | A:B::C:D]
+
+Per-interface configuration
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+One nice feature to use is the ability to apply Flowspec to a specific
+interface, instead of applying it to the whole machine. Despite the following
+IETF draft [Draft-IETF-IDR-Flowspec-Interface-Set]_ is not implemented, it is
+possible to manually limit Flowspec application to some incoming interfaces.
+Actually, not using it can result to some unexpected behaviour like accounting
+twice the traffic, or slow down the traffic (filtering costs). To limit
+Flowspec to one specific interface, use the following command, under
+`flowspec address-family` node.
+
+.. clicmd:: local-install <IFNAME | any>
+
+By default, Flowspec is activated on all interfaces. Installing it to a named
+interface will result in allowing only this interface. Conversely, enabling any
+interface will flush all previously configured interfaces.
+
+VRF redirection
+^^^^^^^^^^^^^^^
+
+Another nice feature to configure is the ability to redirect traffic to a
+separate VRF. This feature does not go against the ability to configure
+Flowspec only on default VRF. Actually, when you receive incoming BGP flowspec
+entries on that default VRF, you can redirect traffic to an other VRF.
+
+As a reminder, BGP flowspec entries have a BGP extended community that contains
+a Route Target. Finding out a local VRF based on Route Target consists in the
+following:
+
+- A configuration of each VRF must be done, with its Route Target set
+ Each VRF is being configured within a BGP VRF instance with its own Route
+ Target list. Route Target accepted format matches the following:
+ ``A.B.C.D:U16``, or ``U16:U32``, ``U32:U16``.
+
+- The first VRF with the matching Route Target will be selected to route traffic
+ to. Use the following command under ipv4 unicast address-family node
+
+.. clicmd:: rt redirect import RTLIST...
+
+In order to illustrate, if the Route Target configured in the Flowspec entry is
+``E.F.G.H:II``, then a BGP VRF instance with the same Route Target will be set
+set. That VRF will then be selected. The below full configuration example
+depicts how Route Targets are configured and how VRFs and cross VRF
+configuration is done. Note that the VRF are mapped on Linux Network
+Namespaces. For data traffic to cross VRF boundaries, virtual ethernet
+interfaces are created with private IP addressing scheme.
+
+.. code-block:: frr
+
+ router bgp <ASx>
+ neighbor <A.B.C.D> remote-as <ASz>
+ address-family ipv4 flowspec
+ neighbor A.B.C.D activate
+ exit
+ exit
+ router bgp <ASy> vrf vrf2
+ address-family ipv4 unicast
+ rt redirect import <E.F.G.H:II>
+ exit
+ exit
+
+Similarly, it is possible to do the same for IPv6 flowspec rules, by using
+an IPv6 extended community. The format is defined on :rfc:`5701`, and that
+community contains an IPv6 address encoded in the attribute, and matches the
+locally configured imported route target IPv6 defined under the appropriate
+BGP VRF instance. Below example defines an IPv6 extended community containing
+`E:F::G:H` address followed by 2 bytes chosen by admin ( here `JJ`).
+
+.. code-block:: frr
+
+ router bgp <ASx>
+ neighbor <A:B::C:D> remote-as <ASz>
+ address-family ipv6 flowspec
+ neighbor A:B::C:D activate
+ exit
+ exit
+ router bgp <ASy> vrf vrf2
+ address-family ipv6 unicast
+ rt6 redirect import <E:F::G:H:JJ>
+ exit
+ exit
+
+
+Flowspec monitoring & troubleshooting
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+You can monitor policy-routing objects by using one of the following commands.
+Those command rely on the filtering contexts configured from BGP, and get the
+statistics information retrieved from the underlying system. In other words,
+those statistics are retrieved from ``Netfilter``.
+
+.. clicmd:: show pbr ipset IPSETNAME | iptable
+
+``IPSETNAME`` is the policy routing object name created by ``ipset``. About
+rule contexts, it is possible to know which rule has been configured to
+policy-route some specific traffic. The :clicmd:`show pbr iptable` command
+displays for forwarded traffic, which table is used. Then it is easy to use
+that table identifier to dump the routing table that the forwarded traffic will
+match.
+
+.. code-block:: frr
+
+.. clicmd:: show ip route table TABLEID
+
+ ``TABLEID`` is the table number identifier referencing the non standard
+ routing table used in this example.
+
+.. clicmd:: debug bgp flowspec
+
+ You can troubleshoot Flowspec, or BGP policy based routing. For instance, if
+ you encounter some issues when decoding a Flowspec entry, you should enable
+ :clicmd:`debug bgp flowspec`.
+
+.. clicmd:: debug bgp pbr [error]
+
+ If you fail to apply the flowspec entry into *zebra*, there should be some
+ relationship with policy routing mechanism. Here,
+ :clicmd:`debug bgp pbr error` could help.
+
+ To get information about policy routing contexts created/removed, only use
+ :clicmd:`debug bgp pbr` command.
+
+Ensuring that a Flowspec entry has been correctly installed and that incoming
+traffic is policy-routed correctly can be checked as demonstrated below. First
+of all, you must check whether the Flowspec entry has been installed or not.
+
+.. code-block:: frr
+
+ CLI# show bgp ipv4 flowspec 5.5.5.2/32
+ BGP flowspec entry: (flags 0x418)
+ Destination Address 5.5.5.2/32
+ IP Protocol = 17
+ Destination Port >= 50 , <= 90
+ FS:redirect VRF RT:255.255.255.255:255
+ received for 18:41:37
+ installed in PBR (match0x271ce00)
+
+This means that the Flowspec entry has been installed in an ``iptable`` named
+``match0x271ce00``. Once you have confirmation it is installed, you can check
+whether you find the associate entry by executing following command. You can
+also check whether incoming traffic has been matched by looking at counter
+line.
+
+.. code-block:: frr
+
+ CLI# show pbr ipset match0x271ce00
+ IPset match0x271ce00 type net,port
+ to 5.5.5.0/24:proto 6:80-120 (8)
+ pkts 1000, bytes 1000000
+ to 5.5.5.2:proto 17:50-90 (5)
+ pkts 1692918, bytes 157441374
+
+As you can see, the entry is present. note that an ``iptable`` entry can be
+used to host several Flowspec entries. In order to know where the matching
+traffic is redirected to, you have to look at the policy routing rules. The
+policy-routing is done by forwarding traffic to a routing table number. That
+routing table number is reached by using a ``iptable``. The relationship
+between the routing table number and the incoming traffic is a ``MARKER`` that
+is set by the IPtable referencing the IPSet. In Flowspec case, ``iptable``
+referencing the ``ipset`` context have the same name. So it is easy to know
+which routing table is used by issuing following command:
+
+.. code-block:: frr
+
+ CLI# show pbr iptable
+ IPtable match0x271ce00 action redirect (5)
+ pkts 1700000, bytes 158000000
+ table 257, fwmark 257
+ ...
+
+As you can see, by using following Linux commands, the MARKER ``0x101`` is
+present in both ``iptable`` and ``ip rule`` contexts.
+
+.. code-block:: shell
+
+ # iptables -t mangle --list match0x271ce00 -v
+ Chain match0x271ce00 (1 references)
+ pkts bytes target prot opt in out source destination
+ 1700K 158M MARK all -- any any anywhere anywhere
+ MARK set 0x101
+ 1700K 158M ACCEPT all -- any any anywhere anywhere
+
+ # ip rule list
+ 0:from all lookup local
+ 0:from all fwmark 0x101 lookup 257
+ 32766:from all lookup main
+ 32767:from all lookup default
+
+This allows us to see where the traffic is forwarded to.
+
+.. _flowspec-known-issues:
+
+Limitations / Known Issues
+--------------------------
+
+As you can see, Flowspec is rich and can be very complex. As of today, not all
+Flowspec rules will be able to be converted into Policy Based Routing actions.
+
+- The ``Netfilter`` driver is not integrated into FRR yet. Not having this
+ piece of code prevents from injecting flowspec entries into the underlying
+ system.
+
+- There are some limitations around filtering contexts
+
+ If I take example of UDP ports, or TCP ports in Flowspec, the information
+ can be a range of ports, or a unique value. This case is handled.
+ However, complexity can be increased, if the flow is a combination of a list
+ of range of ports and an enumerate of unique values. Here this case is not
+ handled. Similarly, it is not possible to create a filter for both src port
+ and dst port. For instance, filter on src port from [1-1000] and dst port =
+ 80. The same kind of complexity is not possible for packet length, ICMP type,
+ ICMP code.
+
+There are some other known issues:
+
+- The validation procedure depicted in :rfc:`5575` is not available.
+
+ This validation procedure has not been implemented, as this feature was not
+ used in the existing setups you shared with us.
+
+- The filtering action shaper value, if positive, is not used to apply shaping.
+
+ If value is positive, the traffic is redirected to the wished destination,
+ without any other action configured by Flowspec.
+ It is recommended to configure Quality of Service if needed, more globally on
+ a per interface basis.
+
+- Upon an unexpected crash or other event, *zebra* may not have time to flush
+ PBR contexts.
+
+ That is to say ``ipset``, ``iptable`` and ``ip rule`` contexts. This is also a
+ consequence due to the fact that ip rule / ipset / iptables are not discovered
+ at startup (not able to read appropriate contexts coming from Flowspec).
+
+Appendix
+--------
+
+More information with a public presentation that explains the design of Flowspec
+inside FRRouting.
+
+[Presentation]_
+
+.. [Draft-IETF-IDR-Flowspec-redirect-IP] <https://tools.ietf.org/id/draft-ietf-idr-flowspec-redirect-ip-02.txt>
+.. [Draft-IETF-IDR-Flowspec-Interface-Set] <https://tools.ietf.org/id/draft-ietf-idr-flowspec-interfaceset-03.txt>
+.. [Draft-IETF-IDR-Flow-Spec-V6] <https://tools.ietf.org/id/draft-ietf-idr-flow-spec-v6-10.txt>
+.. [Presentation] <https://docs.google.com/presentation/d/1ekQygUAG5yvQ3wWUyrw4Wcag0LgmbW1kV02IWcU4iUg/edit#slide=id.g378f0e1b5e_1_44>