1 ========================
3 ========================
5 This document describes some of the internals of the ovs-vswitchd
6 process. It is not complete. It tends to be updated on demand, so if
7 you have questions about the vswitchd implementation, ask them and
8 perhaps we'll add some appropriate documentation here.
10 Most of the ovs-vswitchd implementation is in vswitchd/bridge.c, so
11 code references below should be assumed to refer to that file except
12 as otherwise specified.
17 Bonding allows two or more interfaces (the "slaves") to share network
18 traffic. From a high-level point of view, bonded interfaces act like
19 a single port, but they have the bandwidth of multiple network
20 devices, e.g. two 1 GB physical interfaces act like a single 2 GB
21 interface. Bonds also increase robustness: the bonded port does not
22 go down as long as at least one of its slaves is up.
24 In vswitchd, a bond always has at least two slaves (and may have
25 more). If a configuration error, etc. would cause a bond to have only
26 one slave, the port becomes an ordinary port, not a bonded port, and
27 none of the special features of bonded ports described in this section
30 There are many forms of bonding of which ovs-vswitchd implements only
31 a few. The most complex bond ovs-vswitchd implements is called
32 "source load balancing" or SLB bonding. SLB bonding divides traffic
33 among the slaves based on the Ethernet source address. This is useful
34 only if the traffic over the bond has multiple Ethernet source
35 addresses, for example if network traffic from multiple VMs are
36 multiplexed over the bond.
38 Enabling and Disabling Slaves
39 -----------------------------
41 When a bond is created, a slave is initially enabled or disabled based
42 on whether carrier is detected on the NIC (see iface_create()). After
43 that, a slave is disabled if its carrier goes down for a period of
44 time longer than the downdelay, and it is enabled if carrier comes up
45 for longer than the updelay (see bond_link_status_update()). There is
46 one exception where the updelay is skipped: if no slaves at all are
47 currently enabled, then the first slave on which carrier comes up is
50 The updelay should be set to a time longer than the STP forwarding
51 delay of the physical switch to which the bond port is connected (if
52 STP is enabled on that switch). Otherwise, the slave will be enabled,
53 and load may be shifted to it, before the physical switch starts
54 forwarding packets on that port, which can cause some data to be
55 "blackholed" for a time. The exception for a single enabled slave
56 does not cause any problem in this regard because when no slaves are
57 enabled all output packets are blackholed anyway.
59 When a slave becomes disabled, the vswitch immediately chooses a new
60 output port for traffic that was destined for that slave (see
61 bond_enable_slave()). It also sends a "gratuitous learning packet",
62 specifically a RARP, on the bond port (on the newly chosen slave) for
63 each MAC address that the vswitch has learned on a port other than the
64 bond (see bond_send_learning_packets()), to teach the physical switch
65 that the new slave should be used in place of the one that is now
66 disabled. (This behavior probably makes sense only for a vswitch that
67 has only one port (the bond) connected to a physical switch; vswitchd
68 should probably provide a way to disable or configure it in other
74 Bonding accepts unicast packets on any bond slave. This can
75 occasionally cause packet duplication for the first few packets sent
76 to a given MAC, if the physical switch attached to the bond is
77 flooding packets to that MAC because it has not yet learned the
78 correct slave for that MAC.
80 Bonding only accepts multicast (and broadcast) packets on a single
81 bond slave (the "active slave") at any given time. Multicast packets
82 received on other slaves are dropped. Otherwise, every multicast
83 packet would be duplicated, once for every bond slave, because the
84 physical switch attached to the bond will flood those packets.
86 Bonding also drops received packets when the vswitch has learned that
87 the packet's MAC is on a port other than the bond port itself. This is
88 because it is likely that the vswitch itself sent the packet out the
89 bond port on a different slave and is now receiving the packet back.
90 This occurs when the packet is multicast or the physical switch has not
91 yet learned the MAC and is flooding it. However, the vswitch makes an
92 exception to this rule for broadcast ARP replies, which indicate that
93 the MAC has moved to another switch, probably due to VM migration.
94 (ARP replies are normally unicast, so this exception does not match
95 normal ARP replies. It will match the learning packets sent on bond
98 The active slave is simply the first slave to be enabled after the
99 bond is created (see bond_choose_active_iface()). If the active slave
100 is disabled, then a new active slave is chosen among the slaves that
101 remain active. Currently due to the way that configuration works,
102 this tends to be the remaining slave whose interface name is first
103 alphabetically, but this is by no means guaranteed.
108 When a packet is sent out a bond port, the bond slave actually used is
109 selected based on the packet's source MAC and VLAN tag (see
110 choose_output_iface()). In particular, the source MAC and VLAN tag
111 are hashed into one of 256 values, and that value is looked up in a
112 hash table (the "bond hash") kept in the "bond_hash" member of struct
113 port. The hash table entry identifies a bond slave. If no bond slave
114 has yet been chosen for that hash table entry, vswitchd chooses one
117 Every 10 seconds, vswitchd rebalances the bond slaves (see
118 bond_rebalance_port()). To rebalance, vswitchd examines the
119 statistics for the number of bytes transmitted by each slave over
120 approximately the past minute, with data sent more recently weighted
121 more heavily than data sent less recently. It considers each of the
122 slaves in order from most-loaded to least-loaded. If highly loaded
123 slave H is significantly more heavily loaded than the least-loaded
124 slave L, and slave H carries at least two hashes, then vswitchd shifts
125 one of H's hashes to L. However, vswitchd will only shift a hash from
126 H to L if it will decrease the ratio of the load between H and L by at
129 Currently, "significantly more loaded" means that H must carry at
130 least 1 Mbps more traffic, and that traffic must be at least 3%
136 Each bond balancing mode has different considerations, described
142 LACP bonding requires the remote switch to implement LACP, but it is
143 otherwise very simple in that, after LACP negotiation is complete,
144 there is no need for special handling of received packets.
146 Several of the physical switches that support LACP block all traffic
147 for ports that are configured to use LACP, until LACP is negotiated with
148 the host. When configuring a LACP bond on a OVS host (eg: XenServer),
149 this means that there will be an interruption of the network connectivity
150 between the time the ports on the physical switch and the bond on the OVS
151 host are configured. The interruption may be relatively long, if different
152 people are responsible for managing the switches and the OVS host.
154 Such network connectivity failure can be avoided if LACP can be configured
155 on the OVS host before configuring the physical switch, and having
156 the OVS host fall back to a bond mode (active-backup) till the physical
157 switch LACP configuration is complete. An option "lacp-fallback-ab" exists to
158 provide such behavior on openvswitch.
160 Active Backup Bonding
161 ---------------------
163 Active Backup bonds send all traffic out one "active" slave until that
164 slave becomes unavailable. Since they are significantly less
165 complicated than SLB bonds, they are preferred when LACP is not an
166 option. Additionally, they are the only bond mode which supports
167 attaching each slave to a different upstream switch.
172 SLB bonding allows a limited form of load balancing without the remote
173 switch's knowledge or cooperation. The basics of SLB are simple. SLB
174 assigns each source MAC+VLAN pair to a link and transmits all packets
175 from that MAC+VLAN through that link. Learning in the remote switch
176 causes it to send packets to that MAC+VLAN through the same link.
178 SLB bonding has the following complications:
180 0. When the remote switch has not learned the MAC for the
181 destination of a unicast packet and hence floods the packet to
182 all of the links on the SLB bond, Open vSwitch will forward
183 duplicate packets, one per link, to each other switch port.
185 Open vSwitch does not solve this problem.
187 1. When the remote switch receives a multicast or broadcast packet
188 from a port not on the SLB bond, it will forward it to all of
189 the links in the SLB bond. This would cause packet duplication
190 if not handled specially.
192 Open vSwitch avoids packet duplication by accepting multicast
193 and broadcast packets on only the active slave, and dropping
194 multicast and broadcast packets on all other slaves.
196 2. When Open vSwitch forwards a multicast or broadcast packet to a
197 link in the SLB bond other than the active slave, the remote
198 switch will forward it to all of the other links in the SLB
199 bond, including the active slave. Without special handling,
200 this would mean that Open vSwitch would forward a second copy of
201 the packet to each switch port (other than the bond), including
202 the port that originated the packet.
204 Open vSwitch deals with this case by dropping packets received
205 on any SLB bonded link that have a source MAC+VLAN that has been
206 learned on any other port. (This means that SLB as implemented
207 in Open vSwitch relies critically on MAC learning. Notably, SLB
208 is incompatible with the "flood_vlans" feature.)
210 3. Suppose that a MAC+VLAN moves to an SLB bond from another port
211 (e.g. when a VM is migrated from this hypervisor to a different
212 one). Without additional special handling, Open vSwitch will
213 not notice until the MAC learning entry expires, up to 60
214 seconds later as a consequence of rule #2.
216 Open vSwitch avoids a 60-second delay by listening for
217 gratuitous ARPs, which VMs commonly emit upon migration. As an
218 exception to rule #2, a gratuitous ARP received on an SLB bond
219 is not dropped and updates the MAC learning table in the usual
220 way. (If a move does not trigger a gratuitous ARP, or if the
221 gratuitous ARP is lost in the network, then a 60-second delay
224 4. Suppose that a MAC+VLAN moves from an SLB bond to another port
225 (e.g. when a VM is migrated from a different hypervisor to this
226 one), that the MAC+VLAN emits a gratuitous ARP, and that Open
227 vSwitch forwards that gratuitous ARP to a link in the SLB bond
228 other than the active slave. The remote switch will forward the
229 gratuitous ARP to all of the other links in the SLB bond,
230 including the active slave. Without additional special
231 handling, this would mean that Open vSwitch would learn that the
232 MAC+VLAN was located on the SLB bond, as a consequence of rule
235 Open vSwitch avoids this problem by "locking" the MAC learning
236 table entry for a MAC+VLAN from which a gratuitous ARP was
237 received from a non-SLB bond port. For 5 seconds, a locked MAC
238 learning table entry will not be updated based on a gratuitous
239 ARP received on a SLB bond.