| Internet-Draft | EVPN-lite | December 2020 |
| Wang & Chen | Expires 19 June 2021 | [Page] |
When PBB EVPN [RFC7623] is used in Segment Routing networks, it is complicated to make use of the SID list to carry a function that is aiming for C-MACs.¶
In [I-D.ietf-spring-srv6-network-programming], End.DX2 function is defined, this function can be used in EVPN VPLS. When it is used in EVPN VPLS, the data-plane learning defined in End.DT2U function can also be activated for End.DX2 function. On the basis of such End.DX2 function, SRv6 EVPN can meet all the requirements per [RFC7623] and bring us some other benefits. Such SRv6 EVPN is called light-weighted SRv6 EVPN, and it will be more simpler than PBB EVPN over SRv6.¶
It is easy for the light-weighted SRv6 EVPN to carry a SID that is aiming for customer ethernet packets, because there will be no other ethernet header between the SID list and the customer ethernet header. These SIDs may be user-defined functions for the customer ethernet headers.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 19 June 2021.¶
Copyright (c) 2020 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.¶
When there are too many customer-MACs (C-MACs), the RRs and/or ASBRs will be overloaded by the RT-2 routes for these MACs according to [RFC7432]. This issue can be simply solved by making the remote C-MAC entries learnt via data-plane MAC learning (like what PBB VPLS have done since [RFC7041]) rather than received from RT-2 routes. This simplified solution will works as well as PBB VPLS. But this simplified solution will lose many important features which is based on the ESI concept. Because the ingress-ESI can't be learnt via data-plane MAC learning at the egress PE. So when the data packets is forwarded following these MAC entries, they can't benefit from the EAD/EVI routes as per RFC7432. So the All-Active Redundancy mode for ES can't be supported. This make the simplified solution can't work as well as PBB EVPN ([RFC7623]).¶
This document proposes some new extensions to [RFC7432] to achieve all-active mode ES redundancy on TPEs and reduce the C-MAC loads for RRs and ASBRs at the same time. The new solution will work even more better than PBB EVPN under the help of these extensions, especially when there is no deployment of MPLS dataplane.¶
Furthermore, it naturally brings the benefits of high scalability, faster network convergence, and reduced operational complexity, and we call it light-weighted EVPNs because of these advantages.¶
In [RFC7432], the C-MACs is advertised via RT-2 route. This behavior is inheritted by [RFC8365] and [I-D.ietf-bess-srv6-services]. but in order to solve the C-MAC overload problem for RRs and ASBRs, we have to return to a PBB-like dataplane C-MAC learning procedures.¶
We discuss all the requirements for a light-weighted EVPN solution which pushes no C-MAC entries into the backbone network in Section 2. Note that some of these requirements is not supported well by PBB EVPN.¶
In this document, the light-weighted EVPN solutions are also called as EVPN-lite for short. A total of four EVPN-lite solutions are proposed since [Revision-01]. These solutions are VXLAN over EVPN IP-VRF, light-weighted VXLAN EVPN, light-weighted MPLS EVPN, light-weighted SRv6 EVPN. But this revision focuses its attention on the SRv6 EVPNs and SR-MPLS EVPN.¶
In order to compare these five solutions with [RFC7348] and [RFC7623] whose C-MAC entries are also not pushed into the backbone network, two terms are introduced in this document, because the comparisons need to be done in unified terminology. One term is "Global ESI Indicator (GEI)", which is called as B-MAC in PBB EVPN. The other term is "EVI's Global Dicreminator (EGD)", which is called as I-SID in PBB EVPN.¶
Note that the EVI here corresponds to the I-Component of [RFC7623], not the B-Component. In fact, there will be no typical B-components in some of the above seven solutions.¶
Note that the GEI and EGD in different EVPN-lite solutions are very different. The details will be described in Section 4.¶
On the basis of GEI concept, then we define two route-types for EVPN-lite: The first route type is GEI/ES route, which is called as RT-2 route in PBB EVPN. The second route type is GEI/EVI route, which is called as EAD/EVI roue in [RFC7432].¶
The details of these terms are described in Section 1.3.¶
Most of the terminology used in this documents comes from [RFC7432] and [I-D.ietf-bess-srv6-services] except for the following:¶
EVPN C-MAC Reduction should be provided together with the following requirements:¶
In typical operation, an EVPN PE sends a BGP MAC Advertisement route per C-MAC address. In certain applications, this poses scalability challenges, as is the case in data center interconnect (DCI) scenarios where the number of virtual machines (VMs), and hence the number of C-MAC addresses, can be in the millions. This is called as C-MAC overload of DC Backbone. In such scenarios, it is required to reduce the number of BGP MAC Advertisement routes by relying on a 'EVPN-lite' scheme, as is provided by ESI and its equivalents (e.g. Pseudo B-MAC, ESI IP).¶
The PBB-VPLS/PBB-EVPN is not friendly to IRB usecase because of its complicated Protocol Stack, so it is used just in pure L2VPN usecase up to now in the industry.¶
The solution should provide efficient forwarding performance in EVPN IRB use cases.¶
PBB EVPN, especially the MPLS encapsulation of its B-VPLS, is typically not used in DC Scenario. So we bring PBB and MPLS encapsulation to DC Backbone just due to the C-MAC overload problem. EVPN IRB is widely deplyed in DC scenarios, but PBB EVPN is not friendly for EVPN IRB use cases. So we have to use different solutions in EVPN IRB and C-MAC reduction use cases. We believe that if we choose VXLAN/Geneve data-plane, we will prefer to use the same data-plane in all use cases, e.g. EVPN IRB, C-MAC reduction. So it is necessary to make NVO3/MPLS/SRv6 EVPN to support Section 2.1 in order to provider a unified solution for data center and other secenarios.¶
Two redundancy modes are defined in [RFC7432]. They are All-Active mode and Single-Active mode.¶
In All-active mode, the C-MAC movement among the different adjacent PE nodes of the same ESI should not be considered as C-MAC mobility. In Single-Active mode, such movements can be considered as C-MAC mobility.¶
Flexible multi-homing means that different ES instances can have different adjacent-PEs. We call all the adjacent-PEs of the same ES instances as that ES's location-set in this document. Flexible multi-homing means that different ES can have different location-set.¶
For example, ES1's location-set is {PE1}, ES2's location-set is {PE2, PE3}, ES3's location-set is {PE1, PE3}, and ES4's location-set is {PE2,PE4}.¶
In EVPN, all the PE nodes participating in the same EVPN instance are exposed to all the C-MAC addresses learnt by any one of these PE nodes because a C-MAC learnt by one of the PE nodes is advertised in BGP to other PE nodes in that EVPN instance. This is the case even if some of the PE nodes for that EVPN instance are not involved in forwarding traffic to, or from, these C-MAC addresses. Even if an implementation does not install hardware forwarding entries for C-MAC addresses that are not part of active traffic flows on that PE, the device memory is still consumed by keeping record of the C-MAC addresses in the routing information base (RIB) table. In network applications with millions of C-MAC addresses, this introduces a non- trivial waste of PE resources. As such, it is required to confine the scope of visibility of C-MAC addresses to only those PE nodes that are actively involved in forwarding traffic to, or from, these addresses.¶
Just as in [RFC7432], it is required to avoid C-MAC address flushing upon link, port, or node failure for remote All-Active multihomed segments.¶
Just as in [RFC7432], upon single-active ESI's link or port failure, the C-MACs of other single-active ESes from the same PE will not be flushed.¶
When the physical port of an All-Active ES works well, but a single Ethernet Tag ID (ETI) of that ES fails, The traffic to that ETI of that ES will be re-routed to other adjacent PE of the same ES, but the traffic to other ETIs of the same ES will not be affected.¶
Note that when AC (ES link) fails but PE node still works well, there should not be steady bypassing traffic either. The steady bypassing problem is discussed in [I-D.wang-bess-evpn-egress-protection].¶
The routes per ESIs can be aggregated in Backbone network. Even the default route should be supported when the B-Component is an EVPN IP-VRF (e.g. in VXLAN over IP-VRF solutions).¶
In SRv6 EVPN, different sub-interfaces of the same ESI can have different ESI-indicators in order to achieve Independent Convergency per <ESI, EVI>. But only the common prefix of them should be advertised (both in underlay network and in overlay network) before any of the sub-interfaces fails.¶
The ARP suppression requires <IP,MAC> entries to be steadily held on all TPEs, So it conflicts with Section 2.6. But if the C-MAC confinement requirements is not so important in some scenarios, The ARP Suppression can be activated. This is an option.¶
There are obvious difference between "ESI Route Aggregation" and "ESI Indicator Aggregation". The "ESI Route Aggregation" is that some ESI Indicators are advertised by underlay protocols in a aggregatated manner, but different ESIs still have different ESI-Indicators. The "ESI Indicator Aggregation" is that different ESIs use the same ESI-Indicator.¶
Note that the "ESI Route Aggregation" is recommanded as long as it is possible, but the "ESI Indicator Aggregation" can only be used under certain restraints.¶
When two ESes are attached to the same redundancy group of PEs, they can share the same ESI indicator. But this will bring out some issues too. One of these issues is that they may be attached to different groups of PEs in the future. Another issue is that when only one of the ESes fails, the ESI indicator can't be withdrawn by that PE, so the steady bypass of that ES arises immediately after its failture on that PE. If these issues are not so important in some scenarios, The ESI-Indicator Aggregation may be activated. This is an option.¶
Note that when ESI Indicator Aggregation is activated, the local-bias ES split-horizon procedures or its variations (like what [I-D.eastlake-bess-evpn-vxlan-bypass-vtep] does) should be used.¶
Note that ESI Indicator Aggregation works well with single-active ESIs (see Section 4.2), its steadby bypassing problem will arise with all-active ESIs only.¶
Note that the sub-interfaces of an ESI may be assigned with different ESI-indicators, and these ESI-indicators can be aggregated into a common prefix, this common prefix is assigned with the ESI. In such case, only the common prefix should be advertised before any of the sub-interfaces fails. But this is not considered as "ESI Indicator Aggregation", this is "ESI Route Aggregation".¶
The light-weighted EVPNs should support the unequal load-balance defined in [I-D.ietf-bess-evpn-unequal-lb].¶
In AC-aware bundling service interface, the ESes may make its two VLANs to be attached to the same broadcast domain. These two VLANs may be assigned to the same sub-interface, or to different sub-interfaces.¶
We should not make the core-routers aware of any per-EVI routing information of an ESI. Because they are just underlay nodes.¶
The core-routers may not aware of any per-ES routing infomation of the ESIes too. In such case, the anycast ESI SID should be hiden into the SRH, and it is the inner SID for the Node SID of the egress PE.¶
We assign a Global Discreminator EGD1 to an EVI instance EVI1, the EGD1 is a number consists of N bits. We assign an ESI-indicator GEI1 to ESI1 on PE1, and we assign an ESI-indicator GEI2 to ESI1 on PE2. We call the relationship between ESI1 and its two ESI-indicators as ESI1_GEI1 and ESI1_GEI2 respectively. The EGD and GEIs MUST have global uniqueness in EVI1's service domain.¶
+----------+
PE1 | |
+-------------+ | |
| ESI1_GEI1 | | | PE3
/| |----| | +-------------+
/ | | | IP/MPLS | | |
LAG / +-------------+ | Backbone | | ESI2_GEI3 |---CE2
CE1===== | with | | |
\ +-------------+ | EVPN |---| |
\ | | | RRs | +-------------+
\| |----| and |
| ESI1_GEI2 | | SPEs |
+-------------+ | |
PE2 | |
+----------+
We use IMET routes to build a broadcast-list. The broadcast-list is used to forward BUM traffics. The data-plane MAC learning for BUM traffics produces the first batch of C-MAC entries. The subsequent C-MAC entries can be learnt from Unicast traffics and/or BUM traffics. It is clear that we don't use MAC/IP routes to advertise C-MAC entries as usual, that is for fear that the RRs and/or SPEs are overloaded by these C-MACs.¶
When CE1 requests CE2's ARP, PE1 will receive the ARP Request BUM1 from a AC (say AC1) of ESI1. PE1 will forward the ARP Request following the broadcast-list of AC1's EVI instance(say EVI1). The broadcast-list is constructed by IMET routes from PE2/PE3.¶
PE1 will forward the ARP Request to PE2/PE3. The ARP Request is encapsulated with GEI1 and EVI1_GDV1. The inner SMAC of the ARP request is M1 which is CE1's MAC address.¶
When PE2/PE3 receives the ARP Request packet BUM1, they do dataplane MAC learning independently. They will learn that M1 is behind GEI1.¶
Note that when PE2 learns that M1 is behind GEI1, it will assume that M1 is behind the local AC whose ESI-indicator is GEI1 too. The local AC may have more higher priority than the remote one.¶
After the dataplane MAC learning, the ARP request packet BUM1 is broadcasted to the local ACs, behind one of which is CE2.¶
On receiving BUM1 from PE1, PE2 use the ingress GEI information in BUM1 to determine its ingress ESI ESI1, When ESI1 is all-active mode and PE2 is about to forward the ARP request to CE1, PE2 will find that the ESI for the outgoing AC is also ESI1, so PE2 discards it for ESI loop-free considerations.¶
Note that before that ARP Request packet is discarded, its source-MAC can be learnt, especially in "AC-aware bundling service interface". The MAC entry is learnt against the GEI, but it will consider the local sub-interface on that ES as its outgoing interface, in order to avoid unknown-unicast flooding.¶
Note that in "AC-aware bundling service interface", the AC-ID along with that GEI can help the MAC entry to be installed for the correct outgoing interface. Such MAC entry is called as the synced MAC entry.¶
When ESI1 is single-active mode, the outgoing AC may be in blocking state, otherwise its corresponding sub-interface on CE1 will take charge of packet-drop behavior instead. So alghough the ESI for the outgoing AC is not the same as ESI1, no loop will arise in the Ethernet Segment.¶
When CE2 replies to CE1 for the ARP request, PE3 will forward the ARP reply U1 according to the MAC entry M1 learnt previously as above.¶
PE3 will forward the ARP reply U1 to PE1 or PE2 according to ESI1's RT-1 per EVI routes and RT-1 per ES routes:¶
When ESI1 is all-active mode, GEI1 may be the same as GEI2, in such case, we call both of them GEI21 instead. The traffics to M1 will be load-balanced between PE1 and PE2. Because that GEI21 is advertised by both PE1 and PE2l.¶
Whe PE1 received the ARP reply packet U1 from PE3, PE1 first match the packet to the its EVI instance EVI1 by U1's EGD information. And PE1 will not discard it because the egress ESI is not the same as the ingress ESI which is determined by U1's GEI information.¶
When an Ethernet Segment ES1 is attached to an EVI, the attachment-circuit AC1 for that <ESI,EVI> is assigned with an End.DX2 SID. Different ACs of the same ESI are assigned with different End.DX2 SIDs, we call them AC SIDs in this document. But these different End.DX2 SIDs must be able to be aggregated into the same prefix, and this prefix are called as ESI Indicator in light-weighted SRv6 EVPNs. The format of aggregatable End.DX2 SIDs is illustrated in the following figure:¶
|<--- ESI-Indicator(128-N bits) ---->|<---- N bits --->|
+------------+------------+-----------+-------------------------+
| Block | Node | ESI.LDV | AC-ID |
+------------+------------+-----------+-------------------------+
|<------ Locator -------->|<------------- Function ------------>|
Note that the ESI.LDV field is the Local Discreminator Value (LDV) of the ESI (especially the type 3/4/5 ESI). The AC-ID field is the of the EVI of that End.DX2 SID. The ESI.LDV field and the EGD field are integrated into the End.DX2 SID's Function part.¶
Note that in "AC-aware bundling service interface" the AC-ID field MUST be the same as the Attachment Circuit ID of [I-D.sajassi-bess-evpn-ac-aware-bundling]. But in other service interfaces the AC-ID field can also be the EGD of that AC's EVPN instance. Note that the EGD has a global meaning like a global VNI or an PBB I-SID, while the AC-ID part for an ordinary aggregatable End.DX2 SID typically is only a VLAN-ID on that ES.¶
But we can consider the prefix aggregated from these End.DX2 SIDs as a new SRv6 function called End.DX2AGG SID, The format of the End.DX2AGG SID is illustrated in the following figure:¶
|<------ Locator -------->|<- FUNC -->|<------ ARG.ACI -------->|
+------------+------------+-----------+-------------------------+
| Block | Node | ESI.LDV | AC-ID |
+------------+------------+-----------+-------------------------+
Note that whether these SIDs are considered as lots of End.DX2 SIDs or are considered as a single End.DX2AGG SID with different arguments, it is just a local matter of their PE node's independent choice, other PEs of the same EVI won't be aware of the difference of these two implementations.¶
A SID with the End.DX2AGG function is called as an "ESI SID" in this document. The ESI's GEI is the locator and fuction part of its corresponding ESI SID. The argument part of the ESI SID is the AC-ID for the corresponding AC. The AC-ID plus the ESI.LDV works like the function part of an End.DX2 SID. The argument part of an ESI SID is called as ARG.ACI in this document, where the AC is the abbreviation of AC-ID.¶
Note that an SRv6 ESI-indicator is an 128 bits ESI SID with a zero argument, it is also called as ESI-IP. An ESI-SID may have a non-zero argument part, but the ESI-IPs always have zero argument part.¶
The SRv6 SID in IMET route is an End.DT2M SID with a zero argument length. The GEI1 and GEI2 are ESI-IPs of End.DX2AGG SID that is defined in Figure 3. We can use IGP protocols to advertise GEI1 and GEI2 to PE3 respectively in SRv6 underlay. So we don't have to use EAD/ES route or EAD/EVI route in SRv6 EVPN in this section.¶
Note that if ESI1 is single-active mode, GEI1 is different from GEI2, but if ESI1 is all-active mode, GEI1 is the same as GEI2.¶
Note that when PE1 node fails and the ESI is all active, the PLR node will do underlay anycast FRR switching for GEI1(=GEI2). This will bring out fast network convergency.¶
Note that when the PE-CE link of GEI1 fails, the IGP route of GEI1 will be withdrawn, So there will be no steady bypassing for that ES, but a temporary bypassing can be performed to further improve the convergency.¶
The detailed comparisons between light-weighted SRv6 EVPN and PBB EVPN over SRv6 is described in Section 7.¶
In Step #1, PE1 will forward the ARP Request to PE2/PE3 with the following SRv6 BE encapsulation: It's underlay Source IP is the End.DX2AGG SID on PE1 for ESI1; It's underlay Destination IP is the End.DT2M SID on PE2/PE3. The locator and function part of the End.DX2AGG SID is GEI1. The Argument part of the End.DX2AGG SID is 0.¶
Note that the underlay SIP will be the End.DT2U SID (because they don't need an ESI SID) for the single-homed ingress ACs. The multi-homed ingress ACs with single-active behavior may not be assigned with an dedicated ESI-indicator either. In such situations, the underlay SIP can be the End.DT2U SID too. Note that in such situations, the ESI indicator of all single-active ESIs for the same EVI are aggregated into the same IPv6 address.¶
In Step #3, PE2 can compare the ingress-GEI of BUM1 and the GEI of outgoing AC directly, no GEI-to-ESI lookup needed.¶
Note that PE2 can decapsulate the packet following the End.DX2 function or following the End.DX2AGG function. It is just a local matter.¶
In Step #4, PE3 will forward the ARP reply to PE1 with the following SRv6 BE encapsulation: It's underlay Source IP is the End.DX2AGG SID on PE3 for ESI2; It's underlay Destination IP is the End.DX2AGG SID on PE1 for ESI1 according to the MAC entry M1. The ARG.ACI for the End.DX2AGG SID in DIP is the EGD configured on PE3. Note that the EGD for the same EVI is configured with the same value on PE1/PE2/PE3.¶
When ESI1 is all-active mode, GEI1 will be the same as GEI2, so we call both of them GEI21 instead. The traffics to M1 will be load-balanced between PE1 and PE2 by the underlay network on PE3. Because GEI21 is advertised by both PE1 and PE2 in the underlay IGP protocol.¶
Note that if the DIP is the anycast node SID of PE1 and PE2, when the PE-CE link of ESI1 fails, the traffic will be steadily bypassed untill that link recovers again.¶
In Step #5, Whe PE1 received the SRv6 encapsulated ARP reply packet from PE3, PE1 first match the packet to the End.DX2AGG SID of ESI1 by DIP, then match the packet to the EVI instance EVI1 by ARG.ACI.¶
The "Endpoint with decapsulation and ESI-specific L2 table forwarding" behavior (End.DX2AGG for short) is a variant of the End.DX2 behavior.¶
Two of the applications of the End.DX2AGG behavior are the EVPN VPLS [RFC7432] and the EVPN ETREE [RFC8317]use-cases.¶
Any SID instance of this behavior is associated with an ESI E. The behavior also takes an argument: "Arg.ACI". This argument provides a local mapping to an outgoing interface OIF,. The OIF corresponds to <ESI E, EVI V>, and the EVI V's bridge table is L2 Table T .¶
The End.DX2AGG SID MUST be the last segment in a SR Policy.¶
When N receives a packet whose IPv6 DA is S and S is a local End.DX2AGG SID, the processing is identical to the End.DX2 behavior except for the Upper-layer header processing which is as follows:¶
S01. If (Upper-Layer Header type == 143(Ethernet) ) {
S02. Remove the outer IPv6 Header with all its extension headers.
S03. Learn the exposed MAC Source Address in L2 Table T.
S04. Find out the OIF, and forward the Ethernet frame to the OIF.
S05. } Else {
S06. Process as per Section 4.1.1
of [I-D.ietf-spring-srv6-network-programming].
S07. }
¶
Note that the EVI V is determined by the End.DX2AGG SID's ESI-IP and ARG.ACI argument.¶
Note that the MAC learning should not be applied unless the EVI V is an E-LAN service.¶
Note that the OIF may be found out using the MAC-entries in L2 Table T, when the EVI V is an E-LAN service and the AC-aware bundling service interface is used.¶
Note that we can use the ARG.ACI to find out the the OIF on that ES, then the EVI V will be found out.¶
The End.DX2AGG SIDs can be advertised as an IP prefix in underlay IGP protocols. Although it is the aggregation of many AC SIDs, the ESI SIDs may still be too many for the underlay network. And the core routers who are service-agnostic have to install these prefixes.¶
In order to solve these problems, the ESI SIDs can be advertised via EVPN routes in the overlay network.¶
Note that when the URPF (Unicast Reverse Path Forwarding) is enabled and the ESI SIDs are encapsulated as Source IPs, The ESI SIDs should be advertised in underlay network, even if the ESI SIDs won't be encapsulated as destination IPs. Otherwise the source ESI SID should be hiden into the SRH too.¶
When we use EVPN routes to advertise ESI SIDs among the PEs for the overlay network, These routes will not be imported by the core routers. In such case, when the ESI SIDs are used as destination IP addresses, they should be hiden behind the node SID of the corresponding egress PE router.¶
Note that the association between an ESI SID and its corresponding Node SID is also advertised by such EVPN routes.¶
We can use EAD/ES route (or EAD/EVI route) to advertise Global ESI Indicator (GEI) (and EGD), these EAD routes is called as GEI/ES or GEI/EVI route in this document. When the GEI/EVI route is used to advertise GEI, the End.DX2AGG SID is advertised in its SRv6 L2 Service TLV, not in its nexthop. The EGD may be carried in the ARG.ACI field of the End.DX2AGG SID, or it can also be determined from its EVI-RTs.¶
Either GEI/EVI routes (or GEI/ES) routes will be advertised/imported for Global Routing Table (GRT), so their Route-Targets (RT) will be configured with GRT. Because there isn't a dedicated B-component like PBB VPLS and PBB EVPN. Note that the GEI/EVI routes can be installed as /128 routes and the ARG.ACI part can be set to the actual EGD of the corresponding EVI. In such case, when a C-MAC is learnt over an End.DX2AGG SID (as IPv6 SA) in the data-plane, the ARG.ACI field of that SID should be set to the EVI's EGD when the C-MAC entry is installed.¶
Although GEIs is imported to GRT, they are awared only on PE nodes, the transit nodes in underlay network won't be aware of GEIs (they can aware the common prefix of these GEIs) in order to reduce the FIB consumption. We can use the argument length in the SRv6 SID Structure Sub-Sub-TLV to check whether the EGD is too big for the End.DX2AGG SID, So we can avoid the destruction to the function part of the End.DX2AGG and we can use flexible EGD length.¶
In order to solve the problem described in Section 2.9, we may have to advertise AC SIDs. But the amount of AC SIDs may be hundreds of times larger than ESI SIDs. It is necessary for the light-weighted SRv6 EVPNs to reduce the advertisement of AC SIDs.¶
The AC SID of a specified <ESI,EVI> will not be advertised by its PEs, until these PEs know that the <ESI,EVI> fails on at least one of them.¶
Note that the AC SID for that <ESI,EVI> can be used as the source IP of the SRv6 encapsulation before that AC SID is advertised via EVPN routes. Because that when a MAC is learnt over that AC SID, the packet for that MAC can also be forwarded according to the IP Prefix of the corresponding ESI SID due to the longest match procedures of IP lookup.¶
The detailed AC-SID advertisement will be added in the future versions.¶
When the ESI SIDs are advertised by EVPN routes for the overlay network according to Section 5.1.2, we can advertise the EVPN Link Bandwidth extended community (see [I-D.ietf-bess-evpn-unequal-lb]) or something else along with the ESI SIDs using such EVPN routes.¶
Note that these extra information (which are advertised along with the EVPN routes) are awared by the PEs only. The underlay network don't have to be aware of it.¶
Note that when the EVPN Link Bandwidth extended community is advertised along with the ESI SID, The nexthop of the GEI/ES route should not be set to the anycast ECMP Node SID of the advertising PE (egress-PE). On receiving such GEI/ES route, the ingress PE may push this GEI/ES route's nexthop onto the End.DX2AGG/End.DX2 SID when constructing the SID stack, if unequal-LB is required.¶
There are two methods to achieve EVPN egress node protection:¶
The first method: Both the ESI SID and the AC SID are anycast SID, and they are hiden behind the corresponding egress Node SID according to Section 5.1.2. So when the egress node fails, the PLR can do "midpoint protection" for that node SID, as a result of that, the destination IP will be rewritten to the ESI SID behind that node SID.¶
Note that the ESI SID is an anycast SID, so it will be re-routed by the underlay network after that failure.¶
Note that this method requires no special extensions. So it will be suitable for more SRv6 devices than mirror SID.¶
The second method: the egress protection procedures per [I-D.wang-bess-evpn-egress-protection] (which uses an anycast FRR Node SID to achieve underlay anycast FRR protection) can be applied to the GEI/ES route's nexthop, in order to apply underlay anycast FRR protection.¶
Note that the PLR don't have unequal load-balance information, So neither of these two methods will meet the unequal load-balance requirements after that failure. But it will be the best result unless the unequal load-balance information can be advertised via IGP.¶
The details will be added in the future versions, but the procedures about the synced MAC entry of [Section 3.2, Paragraph 5, Item 1] will be helpful.¶
The withdraw of GEI Advertisement can be used as C-MAC flush notification like what have been done by [RFC8317] and [I-D.ietf-bess-pbb-evpn-isid-cmacflush].¶
Note that even if the GEI/EVI routes of Section 5.1 are not advertised, the withdraw of those GEI/EVI route can still be used as a C-MAC flush notification of their <ESI,EVI>.¶
E-tree Supprot extensions is similar to [RFC8317] section 5 except for the following notable differences: The leaf B-MACs are replaced by leaf GEIs, the root B-MACs are replaced by root GEIs. the PBB encapsulation is replaced by other encapsulations, the B-component is replaced by an IP-VRF or the underlay GRT. The B-MAC Advertisement Route is replaced by GEI/EVI route or ESI/IP Route.¶
The dataplane in this draft is no more complex than typical SRv6 EVPN. So it will work as efficient as we should expect in SRv6 EVPN IRB usecase.¶
But the AC SID can be used in MAC/IP advertisement route, even if C-MAC overload is not a real threat. By doing this, the data-plane can be unified among these usecases.¶
Note that the AC SID is also a typical End.DX2 SID too.¶
In MPLS EVPN control plane, we use a 24 bits unsigned number as the EGD of EVI1, and it has global uniqueness in EVI1's service domain. In data plane, we use QinQ tags to carry the EGD.¶
We use a Global Unique Label (GUL) to identify an ESI in EVI1's service domain. So the ESI-GUL is also its Global ESI Indicator. The ESI-GULs are avertised through RT-1 per ES routes, and they are considered to be an ESI-label by these routes. The label in RT-3 route's PMSI-Tunnel Attribute (PTA-Label) whose tunnel type is ingress replication is called as Ingress Replication Multicast Label (IRML) in this document.¶
We use the following encapsulation in MPLS-based EVPN-lite:¶
Format #1 Format #2
+-----------------------+ +----------------------------+
| PSN Labels | | PSN Labels |
+-----------------------+ +----------------------------+
| IRML (EVI1) | | Destination-ESI GUL (ESI1) |
+-----------------------+ +----------------------------+
| Source-ESI GUL (ESI1) | | Source-ESI GUL (ESI2) |
+-----------------------+ +----------------------------+
| Ethernet Header | | Ethernet Header (EVI1) |
+-----------------------+ +----------------------------+
| Ethernet Payload | | Ethernet Payload |
+-----------------------+ +----------------------------+
| Ethernet FCS | | Ethernet FCS |
+-----------------------+ +----------------------------+
Note that the GUL can be a single Label Stack Entry (LSE), in such case, it should be allocated in DCB label space. Given that the ESIs and vESIs may be too many to be allocated in DCB in certain scenarios, so the GUL should be allocated in a few context-specific label spaces, each identified by a Context Label Space ID (CLS-ID) per [I-D.ietf-bess-mvpn-evpn-aggregation-label] in such case. In such case, the ESI-GUL is the entirety of ESI-label and its Context Label Space ID (CLS-ID), so it means two LSEs in the Label Stack at that time.¶
Note that the ESI GULs are assigned by a center authority, which may be a DC controller or an administrator.¶
Note that the ESI-label (ESI-GUL) should be pushed onto the Label Stack whether the packet is BUM or not. The ESI-GUL can't identify the EVPN Instance EVI1, so we have to use the EGD in the inner ethernet header of "Format #2" to find EVI1 out.¶
Note that the GUL concept is very different with the "upstream-assigned label (UAL)" concept. Because that when a SPE receives a GUL from a remote PE, the GUL is considered as an outgoing-label to that remote PE, and although the GUL is also considered as a incoming-label of the current SPE, and the label operation for the GUL will be a "swap", to be precise, The SPE will swap it to itself and then push the MPLS Label Stack to that advertising PE. When the same GUL is received from different remote PEs, MPLS ECMP or FRR procedures will be applied.¶
So when the GUL is two LSEs in the label stack, we can say that the Context-specific Label Space (CLS) of the ESI-label (inside the GUL) takes the role of B-MAC of PBB EVPN, and the CLS-ID label inside the GUL takes the role of the B-VPLS label of PBB EVPN. So no B-VPLS instances will be found here.¶
Note that the GEI/ES route of MPLS-based EVPN-lite is the RT-1 per ES route.¶
Note that the light-weighted MPLS EVPN solutions can be used whether or not the SR-MPLS LSPs are used in the underlay network.¶
The conceptual comparisons between light-weighted MPLS EVPN and (Pseudo-) PBB EVPN is illustrated in [Revision-01].¶
According to [RFC7432], When the IMET route's PTA's tunnel type is ingress replication, the ESI-label is considered to be downstream-assigned too. Because that nothing of RT-1 per ES route will indicate whether the ESI-label is upstream-assigned or not.¶
Alghough ESI-GUL can be a single LSE or two LSEs in the Label Stack, we assume that it is a single LSE by default in this section, it is for simplification purpose.¶
In Step #1, "Format #1" of Figure 4 will be used.¶
Although the Ingress Replication Multicat Label (IRML) of "Format #1" can identify EVI1 by itself, we suppose that the ethernet header of it should also carry EGD as what [M4] does.¶
Note that there isn't a B-VPLS here, so the IRML identifies the EVI1 itself. The EVI1 here equals I-VPLS of PBB EVPN.¶
Note that when that ARP Request packet comes from a SHD (single-homed device), the ESI of its AC will be null. The Source-ESI GUL in "Format #1" will be replaced with a MPLS label identifying the ingress TPE. When we assume that the underlay network is a SR-MPLS network, that TPE-identifying label can be the node SID label of that ingress TPE. This method follows [I-D.wang-bess-evpn-context-label-02], and the context of the TPE-identifying label is identified by the EVI1's IRML of "Format #1".¶
Note that the TPE-identifying label typically will do nothing to the all-active ESes, they are used just for the single-homed ESes. But when Section 2.12 is activated, and all ESIs share the same ESI indicator, an anycast TPE-identifying label in the DCB can be used as that ESI indicator.¶
In Step #2, "Format #1" of Figure 4 will be received. PE3 knows the packet is for EVI1 with the help of the IRML label. Then PE3 can learn the relation between the ingress-GEI (ingress-ESI GUL) and S-MAC of BUM1 directly, no GEI to ESI lookup needed.¶
In Step #3, PE2 can compare the ingress-GEI (ingress-ESI GUL) of BUM1 and the egress-GEI (ESI-GUL of outgoing AC) directly, no GEI to ESI lookup needed.¶
In Step #4, "Format #2" of Figure 4 will be used. The source-ESI GUL, from which the corresponding MAC entry M1 is previously learnt, will be encapsulated as the destination-ESI GUL directly. No GEI to ESI lookup needed only if we don't care the requirements of Section 2.9. Otherwise we should refer the corresponding RT-1 per EVI routes of ESI1 to forward the packet. These RT-1 per EVI routes are advertised for EVI1, so the Ethernet Tag ID (ETI) of these routes don't have to be the EGD.¶
Note that when ESI1 is single-active mode, ESI-GUL of ESI1 will be different on PE1 and PE2. But the MAC entry M1 will use the newest one only, the swithover between them is called as MAC-move.¶
In Step #5, Whe PE1 received the ARP reply packet from PE3, PE1 first match the packet to ESI1 by Destination-ESI GUL, then match the packet to the EVI instance EVI1 by the QinQ tags of Ethernet header.¶
Note that we suppose that the original tags from ingress AC will be processed following the Raw mode per [RFC4448]. Although the tagged mode can be used technically. Note that the original tags (if they are kept in the packet) will be the inner tags of the EGD.¶
Note that when RT-1 per EVI route are used, as specified in [M4]. There is no need to carry EGD in unicast data-packets too.¶
In hierachical topology (as illustrated in the following figure), the PEs are separated into two groups, the Target PEs (TPEs) and the Superstratum PEs (SPEs).¶
___TPE5___ SPE3 ___TPE4_____
/AC5 \ / \ / \AC4
CE3 \ / \ / >=====CE2
\___ \ / \ / ____/AC2
___TPE3----SPE1-------SPE2-------TPE2
/AC3 / \
CE1 / \
\____TPE1___/ \___CE6
AC1
The TPEs works like the IB-BEB-PE in PBB VPLS, the SPE works like the BCB-PE in PBB VPLS. The BCB-PEs in PBB VPLS do BUM replication based on the PBB header. There are no PBB hearder in EVPN-lite solutions, but the SPEs won't learn the C-MACs, which is the same as BCB-PEs in PBB VPLS. The forwarding behaviors of these EVPN-lite solutions are very different from each other:¶
In SRv6-based EVPN-lite, the SPEs are typically pure underlay nodes, they don't have to aware of the EVIs.¶
The SPEs in MPLS-based EVPN-lite don't have to aware of the BUM packets, because that, for IMET routes, they work like the ASBRs in inter-AS option B. In such case, the TPEs do ingress-replication for all other TPEs by themselves.¶
The SPEs in MPLS-based EVPN-lite may terminate the IMET routes that were received from their TPEs. These IMET routes are imported into an corresponding BD, but may not be passed through other SPEs, so as not to cause duplicated BUM packets. In such case, take SPE1 for example, there are two split-horizon-groups, one group is TPE1/TPE3/TPE5, another split-horizon-group is SPE1/SPE2. The BUM packets are replicated between different split-horizon-groups. In such case, the TPEs do ingress-replication for its directly connected TPEs and SPEs, not for the indirectly connected TPEs and SPEs. But the unicast packet will not be forwarded by that BD on the SPEs. The unicast packets will be label-swapped in the context-specific label-space for the corresponding GULs.¶
Note that the BCB-PE in PBB VPLS is typically supported in the industry, But it seems that the BCB-PE in PBB EVPN is typically not supported in the industry up to now. Because the BCB-PE function can be replaced in MPLS EVPN by a label-swapping operation which is like the inter-AS option B scenarios.¶
Note that the BUM packets here are defined based on the destination C-MAC addresses.¶
We briefly compared light-weighted SRv6 EVPN with PBB-VPLS, PBB-EVPN and VXLAN solutions in [Revision-01], further brief comparisions with VTEP Group (and its transplantation in SRv6 network) were described in [Revision-02]. So we just add the detailed comparisons between EVPN-lite SRv6 and PBB EVPN over SRv6 in this revision.¶
I think the "PBB EVPN over SRv6 underlay" solution will be complex, if we address too much things. I have some examples in the following:¶
The upper-layer header for SRv6 is the PBB-header for B-MACs, not the ethernet header for C-MACs, so the SID list (SR-Path or network programming Instructions) in the SRH can't be constructed for the sake of the I-Component. For example, when a SRv6 SID for MAC-guarding (or something else, just an example) present in the SRH for PBB EVPN SRv6, I think it means BMAC-guarding, no C-MAC guarding.¶
The B-MACs for the all-active ESIs can't be aggregated, but the SRv6 SIDs for ESIs can be aggregated. The underlay can advertise the aggregated prefixes only, so the burden of the underlay network may not be increased too much. When the underlay routes is aggregated, the C-MACs can also be learnt against /128 source-IP, it is the advantage of a light-weighted SRv6 EVPN, which can't be gained from a PBB header.¶
The B-MACs are for overlay protection (the real overlay is the I-VPLS, but the B-VPLS is also an overlay network from the viewpoint of the SRv6 network). But the SRv6 SIDs for ESIs will be for underlay protection, it works like the egress protection. They are two different types of solutions.¶
Although PBB EVPN can be transplanted into SRv6 networks along with the PBB header, It seems to be more complicated to me. Take the EVPN IRB usecases for example, that requires seven sequences of header processing, like (SRv6/B-MAC/C-MAC)(Inner-IP)(C-MAC/B-MAC/SRv6), during the overlay L3 forwarding. I think it will be horrible enough for some ASICs to implement it. When the processing is simplified as (SRv6/C-MAC)(Inner-IP)(C-MAC/SRv6), it sounds like a step forward, not backward, IMHO. We can achieve this goal easily inside the EVPN framework, only if the data-plane learning can still be considered as an option after PBB EVPN.¶
Fortunately, SRv6 is just too young to have a transplantation of PBB EVPN. So it will waste nothing for the SRv6 nodes to give up the PBB header whom is never used by these SRv6 nodes. Note that the SRv6 functions (End.DT2U and End.DT2M) for L2VPNs have source-IP-based data-plane learning for a long time already.¶
In EVPN IRB usecase, [I-D.ietf-bess-evpn-irb-extended-mobility] defines some optional extensions to support some specific IRB usecases. In these specific IRB usecases, the <MAC,IP> bindings will change across VM-moves. These extensions can't be applied to PBB EVPNs, they can't be applied to light-weighted EVPNs either. This will not prevent PBB EVPNs and light-weighted EVPNs from supporting typical IRB use-cases.¶
The "Anycast Node SID" solution here is the transplantation of Anycast-VTEP-IP solution in SRv6 data-plane, where the Anycast Node SID is the equivalent of the Anycast VTEP IP address. Note that SRv6 Anycast Node SID is the ultimate aggregation of ESI indicators. The detailed comparisons will be added in the future visions.¶
Security considerations will be added in future versions.¶
IANA is requested to allocate a new code points for the new SRv6 Endpoint Behaviors defined in this document.¶
+------+-------------+---------------+ | Type | Description | Reference | +------+-------------+---------------+ | TBD1 | End.DX2AGG | This Document | +------+-------------+---------------+
When we use Global Unique ESI-label in EAD per ES route, especially in ingress-replication use case, It should be explicitly indicated in the EAD per ES route. The details will be added in future versions.¶
The authors would like to thank the following for their comments and review of this document:¶
Ye Shu.¶