Design Decisions In Open vSwitch ================================ This document describes design decisions that went into implementing Open vSwitch. While we believe these to be reasonable decisions, it is impossible to predict how Open vSwitch will be used in all environments. Understanding assumptions made by Open vSwitch is critical to a successful deployment. The end of this document contains contact information that can be used to let us know how we can make Open vSwitch more generally useful. Asynchronous Messages ===================== Over time, Open vSwitch has added many knobs that control whether a given controller receives OpenFlow asynchronous messages. This section describes how all of these features interact. First, a service controller never receives any asynchronous messages unless it explicitly configures a miss_send_len greater than zero with an OFPT_SET_CONFIG message. Second, OFPT_FLOW_REMOVED and NXT_FLOW_REMOVED messages are generated only if the flow that was removed had the OFPFF_SEND_FLOW_REM flag set. Third, OFPT_PACKET_IN and NXT_PACKET_IN messages are sent only to OpenFlow controller connections that have the correct connection ID (see "struct nx_controller_id" and "struct nx_action_controller"): - For packet-in messages generated by a NXAST_CONTROLLER action, the controller ID specified in the action. - For other packet-in messages, controller ID zero. (This is the default ID when an OpenFlow controller does not configure one.) Finally, Open vSwitch consults a per-connection table indexed by the message type, reason code, and current role. The following table shows how this table is initialized by default when an OpenFlow connection is made. An entry labeled "yes" means that the message is sent, an entry labeled "---" means that the message is suppressed. master/ message and reason code other slave ---------------------------------------- ------- ----- OFPT_PACKET_IN / NXT_PACKET_IN OFPR_NO_MATCH yes --- OFPR_ACTION yes --- OFPR_INVALID_TTL --- --- OFPT_FLOW_REMOVED / NXT_FLOW_REMOVED OFPRR_IDLE_TIMEOUT yes --- OFPRR_HARD_TIMEOUT yes --- OFPRR_DELETE yes --- OFPT_PORT_STATUS OFPPR_ADD yes yes OFPPR_DELETE yes yes OFPPR_MODIFY yes yes The NXT_SET_ASYNC_CONFIG message directly sets all of the values in this table for the current connection. The OFPC_INVALID_TTL_TO_CONTROLLER bit in the OFPT_SET_CONFIG message controls the setting for OFPR_INVALID_TTL for the "master" role. OFPAT_ENQUEUE ============= The OpenFlow 1.0 specification requires the output port of the OFPAT_ENQUEUE action to "refer to a valid physical port (i.e. < OFPP_MAX) or OFPP_IN_PORT". Although OFPP_LOCAL is not less than OFPP_MAX, it is an 'internal' port which can have QoS applied to it in Linux. Since we allow the OFPAT_ENQUEUE to apply to 'internal' ports whose port numbers are less than OFPP_MAX, we interpret OFPP_LOCAL as a physical port and support OFPAT_ENQUEUE on it as well. OFPT_FLOW_MOD ============= The OpenFlow 1.0 specification for the behavior of OFPT_FLOW_MOD is confusing. The following table summarizes the Open vSwitch implementation of its behavior in the following categories: - "match on priority": Whether the flow_mod acts only on flows whose priority matches that included in the flow_mod message. - "match on out_port": Whether the flow_mod acts only on flows that output to the out_port included in the flow_mod message (if out_port is not OFPP_NONE). - "updates flow_cookie": Whether the flow_mod changes the flow_cookie of the flow or flows that it matches to the flow_cookie included in the flow_mod message. - "updates OFPFF_ flags": Whether the flow_mod changes the OFPFF_SEND_FLOW_REM flag of the flow or flows that it matches to the setting included in the flags of the flow_mod message. - "honors OFPFF_CHECK_OVERLAP": Whether the OFPFF_CHECK_OVERLAP flag in the flow_mod is significant. - "updates idle_timeout" and "updates hard_timeout": Whether the idle_timeout and hard_timeout in the flow_mod, respectively, have an effect on the flow or flows matched by the flow_mod. - "updates idle timer": Whether the flow_mod resets the per-flow timer that measures how long a flow has been idle. - "updates hard timer": Whether the flow_mod resets the per-flow timer that measures how long it has been since a flow was modified. - "zeros counters": Whether the flow_mod resets per-flow packet and byte counters to zero. - "sends flow_removed message": Whether the flow_mod generates a flow_removed message for the flow or flows that it affects. An entry labeled "yes" means that the flow mod type does have the indicated behavior, "---" means that it does not, an empty cell means that the property is not applicable, and other values are explained below the table. MODIFY DELETE ADD MODIFY STRICT DELETE STRICT === ====== ====== ====== ====== match on priority --- --- yes --- yes match on out_port --- --- --- yes yes updates flow_cookie yes yes yes updates OFPFF_SEND_FLOW_REM yes + + honors OFPFF_CHECK_OVERLAP yes + + updates idle_timeout yes + + updates hard_timeout yes + + resets idle timer yes + + resets hard timer yes yes yes zeros counters yes + + sends flow_removed message --- --- --- % % (+) "modify" and "modify-strict" only take these actions when they create a new flow, not when they update an existing flow. (%) "delete" and "delete_strict" generates a flow_removed message if the deleted flow or flows have the OFPFF_SEND_FLOW_REM flag set. (Each controller can separately control whether it wants to receive the generated messages.) Flow Cookies ============ OpenFlow 1.0 and later versions have the concept of a "flow cookie", which is a 64-bit integer value attached to each flow. The treatment of the flow cookie has varied greatly across OpenFlow versions, however. In OpenFlow 1.0: - OFPFC_ADD set the cookie in the flow that it added. - OFPFC_MODIFY and OFPFC_MODIFY_STRICT updated the cookie for the flow or flows that it modified. - OFPST_FLOW messages included the flow cookie. - OFPT_FLOW_REMOVED messages reported the cookie of the flow that was removed. OpenFlow 1.1 made the following changes: - Flow mod operations OFPFC_MODIFY, OFPFC_MODIFY_STRICT, OFPFC_DELETE, and OFPFC_DELETE_STRICT, plus flow stats requests and aggregate stats requests, gained the ability to match on flow cookies with an arbitrary mask. - OFPFC_MODIFY and OFPFC_MODIFY_STRICT were changed to add a new flow, in the case of no match, only if the flow table modification operation did not match on the cookie field. (In OpenFlow 1.0, modify operations always added a new flow when there was no match.) - OFPFC_MODIFY and OFPFC_MODIFY_STRICT no longer updated flow cookies. OpenFlow 1.2 made the following changes: - OFPC_MODIFY and OFPFC_MODIFY_STRICT were changed to never add a new flow, regardless of whether the flow cookie was used for matching. Open vSwitch support for OpenFlow 1.0 implements the OpenFlow 1.0 behavior with the following extensions: - An NXM extension field NXM_NX_COOKIE(_W) allows the NXM versions of OFPFC_MODIFY, OFPFC_MODIFY_STRICT, OFPFC_DELETE, and OFPFC_DELETE_STRICT flow_mods, plus flow stats requests and aggregate stats requests, to match on flow cookies with arbitrary masks. This is much like the equivalent OpenFlow 1.1 feature. - However, unlike OpenFlow 1.1, OFPC_MODIFY and OFPFC_MODIFY_STRICT, regardless of whether there was a match based on a cookie or not, always add a new flow if there is no match, and they always update the cookies of flows that they do match. - NXT_PACKET_IN (the Nicira extended version of OFPT_PACKET_IN) reports the cookie of the rule that generated the packet, or all-1-bits if no rule generated the packet. (Older versions of OVS used all-0-bits instead of all-1-bits.) Multiple Table Support ====================== OpenFlow 1.0 has only rudimentary support for multiple flow tables. Notably, OpenFlow 1.0 does not allow the controller to specify the flow table to which a flow is to be added. Open vSwitch adds an extension for this purpose, which is enabled on a per-OpenFlow connection basis using the NXT_FLOW_MOD_TABLE_ID message. When the extension is enabled, the upper 8 bits of the 'command' member in an OFPT_FLOW_MOD or NXT_FLOW_MOD message designates the table to which a flow is to be added. The Open vSwitch software switch implementation offers 255 flow tables. On packet ingress, only the first flow table (table 0) is searched, and the contents of the remaining tables are not considered in any way. Tables other than table 0 only come into play when an NXAST_RESUBMIT_TABLE action specifies another table to search. Tables 128 and above are reserved for use by the switch itself. Controllers should use only tables 0 through 127. IPv6 ==== Open vSwitch supports stateless handling of IPv6 packets. Flows can be written to support matching TCP, UDP, and ICMPv6 headers within an IPv6 packet. Deeper matching of some Neighbor Discovery messages is also supported. IPv6 was not designed to interact well with middle-boxes. This, combined with Open vSwitch's stateless nature, have affected the processing of IPv6 traffic, which is detailed below. Extension Headers ----------------- The base IPv6 header is incredibly simple with the intention of only containing information relevant for routing packets between two endpoints. IPv6 relies heavily on the use of extension headers to provide any other functionality. Unfortunately, the extension headers were designed in such a way that it is impossible to move to the next header (including the layer-4 payload) unless the current header is understood. Open vSwitch will process the following extension headers and continue to the next header: * Fragment (see the next section) * AH (Authentication Header) * Hop-by-Hop Options * Routing * Destination Options When a header is encountered that is not in that list, it is considered "terminal". A terminal header's IPv6 protocol value is stored in "nw_proto" for matching purposes. If a terminal header is TCP, UDP, or ICMPv6, the packet will be further processed in an attempt to extract layer-4 information. Fragments --------- IPv6 requires that every link in the internet have an MTU of 1280 octets or greater (RFC 2460). As such, a terminal header (as described above in "Extension Headers") in the first fragment should generally be reachable. In this case, the terminal header's IPv6 protocol type is stored in the "nw_proto" field for matching purposes. If a terminal header cannot be found in the first fragment (one with a fragment offset of zero), the "nw_proto" field is set to 0. Subsequent fragments (those with a non-zero fragment offset) have the "nw_proto" field set to the IPv6 protocol type for fragments (44). Jumbograms ---------- An IPv6 jumbogram (RFC 2675) is a packet containing a payload longer than 65,535 octets. A jumbogram is only relevant in subnets with a link MTU greater than 65,575 octets, and are not required to be supported on nodes that do not connect to link with such large MTUs. Currently, Open vSwitch doesn't process jumbograms. In-Band Control =============== Motivation ---------- An OpenFlow switch must establish and maintain a TCP network connection to its controller. There are two basic ways to categorize the network that this connection traverses: either it is completely separate from the one that the switch is otherwise controlling, or its path may overlap the network that the switch controls. We call the former case "out-of-band control", the latter case "in-band control". Out-of-band control has the following benefits: - Simplicity: Out-of-band control slightly simplifies the switch implementation. - Reliability: Excessive switch traffic volume cannot interfere with control traffic. - Integrity: Machines not on the control network cannot impersonate a switch or a controller. - Confidentiality: Machines not on the control network cannot snoop on control traffic. In-band control, on the other hand, has the following advantages: - No dedicated port: There is no need to dedicate a physical switch port to control, which is important on switches that have few ports (e.g. wireless routers, low-end embedded platforms). - No dedicated network: There is no need to build and maintain a separate control network. This is important in many environments because it reduces proliferation of switches and wiring. Open vSwitch supports both out-of-band and in-band control. This section describes the principles behind in-band control. See the description of the Controller table in ovs-vswitchd.conf.db(5) to configure OVS for in-band control. Principles ---------- The fundamental principle of in-band control is that an OpenFlow switch must recognize and switch control traffic without involving the OpenFlow controller. All the details of implementing in-band control are special cases of this principle. The rationale for this principle is simple. If the switch does not handle in-band control traffic itself, then it will be caught in a contradiction: it must contact the controller, but it cannot, because only the controller can set up the flows that are needed to contact the controller. The following points describe important special cases of this principle. - In-band control must be implemented regardless of whether the switch is connected. It is tempting to implement the in-band control rules only when the switch is not connected to the controller, using the reasoning that the controller should have complete control once it has established a connection with the switch. This does not work in practice. Consider the case where the switch is connected to the controller. Occasionally it can happen that the controller forgets or otherwise needs to obtain the MAC address of the switch. To do so, the controller sends a broadcast ARP request. A switch that implements the in-band control rules only when it is disconnected will then send an OFPT_PACKET_IN message up to the controller. The controller will be unable to respond, because it does not know the MAC address of the switch. This is a deadlock situation that can only be resolved by the switch noticing that its connection to the controller has hung and reconnecting. - In-band control must override flows set up by the controller. It is reasonable to assume that flows set up by the OpenFlow controller should take precedence over in-band control, on the basis that the controller should be in charge of the switch. Again, this does not work in practice. Reasonable controller implementations may set up a "last resort" fallback rule that wildcards every field and, e.g., sends it up to the controller or discards it. If a controller does that, then it will isolate itself from the switch. - The switch must recognize all control traffic. The fundamental principle of in-band control states, in part, that a switch must recognize control traffic without involving the OpenFlow controller. More specifically, the switch must recognize *all* control traffic. "False negatives", that is, packets that constitute control traffic but that the switch does not recognize as control traffic, lead to control traffic storms. Consider an OpenFlow switch that only recognizes control packets sent to or from that switch. Now suppose that two switches of this type, named A and B, are connected to ports on an Ethernet hub (not a switch) and that an OpenFlow controller is connected to a third hub port. In this setup, control traffic sent by switch A will be seen by switch B, which will send it to the controller as part of an OFPT_PACKET_IN message. Switch A will then see the OFPT_PACKET_IN message's packet, re-encapsulate it in another OFPT_PACKET_IN, and send it to the controller. Switch B will then see that OFPT_PACKET_IN, and so on in an infinite loop. Incidentally, the consequences of "false positives", where packets that are not control traffic are nevertheless recognized as control traffic, are much less severe. The controller will not be able to control their behavior, but the network will remain in working order. False positives do constitute a security problem. - The switch should use echo-requests to detect disconnection. TCP will notice that a connection has hung, but this can take a considerable amount of time. For example, with default settings the Linux kernel TCP implementation will retransmit for between 13 and 30 minutes, depending on the connection's retransmission timeout, according to kernel documentation. This is far too long for a switch to be disconnected, so an OpenFlow switch should implement its own connection timeout. OpenFlow OFPT_ECHO_REQUEST messages are the best way to do this, since they test the OpenFlow connection itself. Implementation -------------- This section describes how Open vSwitch implements in-band control. Correctly implementing in-band control has proven difficult due to its many subtleties, and has thus gone through many iterations. Please read through and understand the reasoning behind the chosen rules before making modifications. Open vSwitch implements in-band control as "hidden" flows, that is, flows that are not visible through OpenFlow, and at a higher priority than wildcarded flows can be set up through OpenFlow. This is done so that the OpenFlow controller cannot interfere with them and possibly break connectivity with its switches. It is possible to see all flows, including in-band ones, with the ovs-appctl "bridge/dump-flows" command. The Open vSwitch implementation of in-band control can hide traffic to arbitrary "remotes", where each remote is one TCP port on one IP address. Currently the remotes are automatically configured as the in-band OpenFlow controllers plus the OVSDB managers, if any. (The latter is a requirement because OVSDB managers are responsible for configuring OpenFlow controllers, so if the manager cannot be reached then OpenFlow cannot be reconfigured.) The following rules (with the OFPP_NORMAL action) are set up on any bridge that has any remotes: (a) DHCP requests sent from the local port. (b) ARP replies to the local port's MAC address. (c) ARP requests from the local port's MAC address. In-band also sets up the following rules for each unique next-hop MAC address for the remotes' IPs (the "next hop" is either the remote itself, if it is on a local subnet, or the gateway to reach the remote): (d) ARP replies to the next hop's MAC address. (e) ARP requests from the next hop's MAC address. In-band also sets up the following rules for each unique remote IP address: (f) ARP replies containing the remote's IP address as a target. (g) ARP requests containing the remote's IP address as a source. In-band also sets up the following rules for each unique remote (IP,port) pair: (h) TCP traffic to the remote's IP and port. (i) TCP traffic from the remote's IP and port. The goal of these rules is to be as narrow as possible to allow a switch to join a network and be able to communicate with the remotes. As mentioned earlier, these rules have higher priority than the controller's rules, so if they are too broad, they may prevent the controller from implementing its policy. As such, in-band actively monitors some aspects of flow and packet processing so that the rules can be made more precise. In-band control monitors attempts to add flows into the datapath that could interfere with its duties. The datapath only allows exact match entries, so in-band control is able to be very precise about the flows it prevents. Flows that miss in the datapath are sent to userspace to be processed, so preventing these flows from being cached in the "fast path" does not affect correctness. The only type of flow that is currently prevented is one that would prevent DHCP replies from being seen by the local port. For example, a rule that forwarded all DHCP traffic to the controller would not be allowed, but one that forwarded to all ports (including the local port) would. As mentioned earlier, packets that miss in the datapath are sent to the userspace for processing. The userspace has its own flow table, the "classifier", so in-band checks whether any special processing is needed before the classifier is consulted. If a packet is a DHCP response to a request from the local port, the packet is forwarded to the local port, regardless of the flow table. Note that this requires L7 processing of DHCP replies to determine whether the 'chaddr' field matches the MAC address of the local port. It is interesting to note that for an L3-based in-band control mechanism, the majority of rules are devoted to ARP traffic. At first glance, some of these rules appear redundant. However, each serves an important role. First, in order to determine the MAC address of the remote side (controller or gateway) for other ARP rules, we must allow ARP traffic for our local port with rules (b) and (c). If we are between a switch and its connection to the remote, we have to allow the other switch's ARP traffic to through. This is done with rules (d) and (e), since we do not know the addresses of the other switches a priori, but do know the remote's or gateway's. Finally, if the remote is running in a local guest VM that is not reached through the local port, the switch that is connected to the VM must allow ARP traffic based on the remote's IP address, since it will not know the MAC address of the local port that is sending the traffic or the MAC address of the remote in the guest VM. With a few notable exceptions below, in-band should work in most network setups. The following are considered "supported' in the current implementation: - Locally Connected. The switch and remote are on the same subnet. This uses rules (a), (b), (c), (h), and (i). - Reached through Gateway. The switch and remote are on different subnets and must go through a gateway. This uses rules (a), (b), (c), (h), and (i). - Between Switch and Remote. This switch is between another switch and the remote, and we want to allow the other switch's traffic through. This uses rules (d), (e), (h), and (i). It uses (b) and (c) indirectly in order to know the MAC address for rules (d) and (e). Note that DHCP for the other switch will not work unless an OpenFlow controller explicitly lets this switch pass the traffic. - Between Switch and Gateway. This switch is between another switch and the gateway, and we want to allow the other switch's traffic through. This uses the same rules and logic as the "Between Switch and Remote" configuration described earlier. - Remote on Local VM. The remote is a guest VM on the system running in-band control. This uses rules (a), (b), (c), (h), and (i). - Remote on Local VM with Different Networks. The remote is a guest VM on the system running in-band control, but the local port is not used to connect to the remote. For example, an IP address is configured on eth0 of the switch. The remote's VM is connected through eth1 of the switch, but an IP address has not been configured for that port on the switch. As such, the switch will use eth0 to connect to the remote, and eth1's rules about the local port will not work. In the example, the switch attached to eth0 would use rules (a), (b), (c), (h), and (i) on eth0. The switch attached to eth1 would use rules (f), (g), (h), and (i). The following are explicitly *not* supported by in-band control: - Specify Remote by Name. Currently, the remote must be identified by IP address. A naive approach would be to permit all DNS traffic. Unfortunately, this would prevent the controller from defining any policy over DNS. Since switches that are located behind us need to connect to the remote, in-band cannot simply add a rule that allows DNS traffic from the local port. The "correct" way to support this is to parse DNS requests to allow all traffic related to a request for the remote's name through. Due to the potential security problems and amount of processing, we decided to hold off for the time-being. - Differing Remotes for Switches. All switches must know the L3 addresses for all the remotes that other switches may use, since rules need to be set up to allow traffic related to those remotes through. See rules (f), (g), (h), and (i). - Differing Routes for Switches. In order for the switch to allow other switches to connect to a remote through a gateway, it allows the gateway's traffic through with rules (d) and (e). If the routes to the remote differ for the two switches, we will not know the MAC address of the alternate gateway. Suggestions =========== Suggestions to improve Open vSwitch are welcome at discuss@openvswitch.org.