IKEv2 support for per-queue Child SAs

IPsec implementations are currently limited to using one queue or CPU per Child SA. The result is that a machine with many queues/CPUs is limited to only using one of these per Child SA. This severely limits the throughput that can be attained. An unencrypted link of 10Gbps or more is commonly reduced to 2-5Gbps when IPsec is used to encrypt the link using AES-GCM. By using the implementation specified in this document, aggregate throughput increased from 5Gbps using 1 CPU to 40-60 Gbps using 25-30 CPUs Furthermore, IPsec implementations are currently limited to use the same Child SA for all Quality of Service (QoS) types because the QoS type is not a part of the Traffic Selector (TS) payload. The result is that IPsec cannot support active Quality of Service prioritization without disabling the anti-replay protection. While this could be (partially) mitigated by setting up multiple narrowed Child SAs, for example using Populate From Packet (PFP) as specified in , this IPsec feature is not widely implemented. Some route based IPsec implementations might be able to implement this with specific rules into separate network interfaces, but these methods might not be available for policy based IPsec implementations. To make better use of multiple network queues and CPUs, it can be beneficial to negotiate and install multiple identical Child SAs. IKEv2 already allows installing multiple identical Child SAs, it offers no method to negotiate the number of Child SAs or indicate the purpose for the multiple Child SAs that are requested. When two IKEv2 peers want to negotiate multiple Child SAs, it is useful to be able to convey how many Child SAs are required for optimized traffic. This avoids triggering CREATE_CHILD_SA exchanges that will only be rejected by the peer.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Currently, most IPsec implementations are limited by using one CPU or network queue per Child SA. There are a number of practical reasons for this, but a key limitation is that sharing the crypto state, counters and sequence numbers between multiple CPUs is not feasible without a significant performance penalty. There is a need to negotiate and establish multiple Child SAs with identical TSi/TSr on a per-queue or per-CPU basis.

When negotiating CPU specific Child SAs, the first SA negotiated either in an IKE_AUTH exchange or CREATE_CHILD_SA is called Fallback SA. This Child SA is similar to a regular Cgild SA in that it is not bound to a single resource (CPU or QoS queue). This Fallback Child SA (or its rekeyed successors) MUST remain active for the lifetime of the IPsec session to ensure that there is always a Child SA that can be selected to send traffic over, in case a per-resource Child SA is not available. Additional Child SAs are installed bound to a specific resource (CPU or QoS queue). These Child SAs are responsible for the bulk of the traffic. The CPU_QUEUES notification payload is sent in the IKE_AUTH or CREATE_CHILD_SA Exchange indicating the negotiated Child SA is a Fallback SA. The CPU_QUEUES notification value refers to the number of additional resource-specific Child SAs that may be installed for this particular TSi/TSr combination excluding the Fallback Child SA. Both peers send the preferred minimum number of additional Child SAs to install. Both peers pick the maximum of the two numbers (within reason). That is, if the initiator prefers 16 and the responder prefers 48, then the number negotiated is 48. The responder may at any time reject additional Child SAs by returning TS_UNACCEPTABLE. It should not return NO_ADDITIONAL_SAS, as there might be another Child SAs with different Traffic Selectors that would still be allowed by the peer. [Antony: Valery's feedback was not to use TS_UNACCEPTABLE. instead create a new notify or use TEMPORARY_FAILURE. TEMPORARY_FAILURE because the situation may change again if you try again. I have preference to define new NO_CPU_QUEUE_INFO_SA] Resource-specific Child SAs are negotiated as regular Child SAs using the CREATE_CHILD_SA exchange and are identified by a CPU_QUEUE_INFO notification. Upon installation, each Child SA is associated with an additional local selector, such as CPU or queue. These additional Child SAs MUST be negotiated with identical Child SA properties that were negotiated for the Fallback SA. This includes cryptographic algorithms, Traffic Selectors, Mode (e.g. transport mode), compression usage, etc. However, the Child SAs do have their own individual keying material that is derived according to the regular IKEv2 process. The CPU_QUEUE_INFO can be empty or contain some identifying data that could be useful for debugging purposes. Additional Child SAs can be started on-demand or can be started all at once. Peers may also delete specific per-resource Child SAs if they deem the associated resource to be idle. The Fallback SA MUST NOT be deleted while any per-resource Child SAs are still present. During the CREATE_CHILD_SA rekey for the Child SA, the CPU_QUEUE_INFO notification MAY be included, but regardless of whether or not it is included, the rekeyed Child SA MUST be bound to the same resource(s) as the Child SA that is being rekeyed. As with regular Child SA rekeying, the new Child SA may not be different from the rekeyed Child SA with respect to cryptographic algorithms and MUST cover the original Traffic Selector ranges. If a CREATE_CHILD_SA exchange request containing both a CPU_QUEUE_INFO and a CPU_QUEUES notification is received, the responder MUST ignore the CPU_QUEUE_INFO payload. If a CREATE_CHILD_SA exchange reply is received with both CPU_QUEUE_INFO and CPU_QUEUES notifications, the initiator MUST ignore the notification that it did not send in the request. [Steffen: I tend to tread these cases as an error.] [Tobias: That's currently how I implemented it (being lenient on what I accept). But we could also treat those cases as errors. The question would just be what we should return (NO_PROPOSAL_CHOSEN and keep IKE and other Child SAs or even INALID_SYNTAX and kill the whole IKE_SA - and as initiator we either have to terminate the Child or the IKE_SA actively if we receive both notifies).] The CPU_QUEUES notification, even when it is sent in the IKE_AUTH exchange, is not an attribute of the IKE peer. It is an attribute of the Child SA, similar to the USE_TRANSPORT notification. That is, an IKE peer can have multiple Child SAs covering different traffic selectors and selectively decide whether or not to enable additional per-resource Child SAs for each of these Child SAs covering different Traffic Selectors.

To install multiple Child SAs for different QoS levels, a similar negotiation method is used. The QOS_QUEUES notification is sent with the negotiation of the Fallback Child SA that is used for all QoS levels not matched by more specific Child SAs. Additional Child SAs are installed per QoS level by including the QOS_QUEUE_INFO notification describing the specific QoS level that this additional Child SA will cover. This allows both peers to install the Child SA using the same QoS level. [Steffen: Maybe mention IPv6 flow label too] If a certain QoS level proposed by the peer is not acceptable to the responder, TS_UNACCEPTABLE MUST be returned. [Tobias: Would a more specific error notify make sense here?] [Antony: We need specific error if is rejected QOS_QUEUE_INFO]

There are various considerations that an implementation can use to determine the best way to install multiple Child SAs. Below are examples of such strategies.

A simple distribution could be to install one additional Child SA on each CPU. The Fallback Child SA ensures that any CPU generating traffic to be encrypted has an available (if not optimal) Child SA to use. Any subsequent Child SAs with identical TSi/TSr Traffic Selectors are installed in such a way to only be used by a single CPU or network queue. Performing per-CPU Child SA negotiations can result in both peers initiating additional Child SAs at once. This is especially likely if per-CPU Child SAs are triggered by individual SADB_ACQUIRE messages. Responders should install the additional Child SA on a CPU with the least amount of additional Child SAs for this TSi/TSr pair. It should count outstanding SADB_ACQUIREs as an assigned additional Child SA. It is still possible that when the peers only have one slot left to assign, that both peers send a CREATE_CHILD_SA request at the same time. [Paul: Is there anything we can do at the protocol level to terminate one of these without race conditions?] [Antony: if CPU_QUEUE_INFO is a MUST, that info could be used for better one-to-one mapping, as well as delete the extra SAs. Also, keep in mind the general case IKE window > 1] As an optimization, additional Child SAs that see little traffic MAY be deleted. The Fallback Child SA MUST NOT be deleted when idle, as it is likely to be idle if enough per-CPU Child SAs are installed. However, if one of those per-CPU child SAs is deleted because it was idle, and subsequently that CPU starts to generate traffic again, that traffic does not have a per-CPU Child SA and will be encrypted using the Fallback Child SA. Meanwhile, the IKE daemon might be negotiating to bring up a new per-CPU Child SA. When the number of queues or CPUs are different between the peers, the peer with the least amount of queues or CPUs MAY decide to not install a second outbound Child SA for the same resource as it will never use it to send traffic. However, it MUST install all inbound Child SAs as it has committed to receiving traffic on these negotiated Child SAs. If per-CPU SADB_ACQUIRE messages are implemented (see ), the Traffic Selector (TSi) entry containing the information of the trigger packet should still be included in the TS set. This information MAY be used by the peer to select the most optimal target CPU to install the additional Child SA on. For example, if the trigger packet was for a TCP destination to port 25 (SMTP), it might be able to install the Child SA on the CPU that is also running the mail server process. Trigger packet Traffic Selectors are documented in Section 2.9. As per RFC 7296, rekeying a Child SA SHOULD use the same (or wider) Traffic Selectors to ensure that the new Child SA covers everything that the rekeyed Child SA covers. This includes Traffic Selectors negotiated via Configuration Payloads (CP) such as INTERNAL_IP4_ADDRESS which may use the original wide TS set or use the narrowed TS set.

[Paul: is there anything we need to say here?] [Steffen: If we want to say something about that case, maybe this:] Most considerations from the per-CPU case apply to the per-QoS case as well. The main difference between these two cases is that the number of possible QoS types are always the same for both peers (e.g. 64 types for IPv4). Unlike the per-CPU case, handling different numbers of QoS types is not necessary. [Paul: I was hoping we could negotiate things like "only 2 different levels needed", and not just a "we want to install SAs for all theoretical possible levels"]

It is unlikely but not disallowed, to use both per CPU and per QoS level Child SAs. Any conflicts between the performance improving types of SAs would need to be handled by local policies. For some, the QoS might be more important to honour as best as possible, while for others, CPU distribution might be more important. There is currently no operational experience with combining these two types of Child SAs. [Tobias: How would this look like? Would you send both notifies on the same set of SAs (CPU/QOS_QUEUE on the fallback SA and INFO on the others)? (So each SA would be for a specific CPU AND QoS class.) Or would you negotiate separate per-CPU and per-QoS SAs all with the same TS? (e.g. if you already bound certain classes to certain CPUs anyway and use a QoS specific SA for that, but still want to use multiple CPUs for the other traffic and negotiate per-CPU SAs without QoS identifier for that)] [Paul: I don't really know - perhaps we should remove QoS until we have someone who actually wants to run this and can provide guidance for standardization ? ]

All multi-octet fields representing integers are laid out in big endian order (also known as "most significant byte first", or "network byte order").

Protocol ID (1 octet) - MUST be 0. MUST be ignored if not 0. SPI Size (1 octet) - MUST be 0. MUST be ignored if not 0. Notify Message Type (2 octets) - set to [TBD1] Minimum number of per-CPU IPsec SAs (4 octets). MUST be greater than 0. If 0 is received, it MUST be interpreted as 1. Note: The Fallback Child SA that is not bound to a single CPU is not counted as part of these numbers.

Protocol ID (1 octet) - MUST be 0. MUST be ignored if not 0. SPI Size (1 octet) - MUST be 0. MUST be ignored if not 0. Notify Message Type (2 octets) - set to [TBD2] Maximum number of QoS level IPsec SAs (4 octets). MUST be greater than 0. If 0 is received, it MUST be interpreted as 1. [Steffen: Does it make sense to negotiate the max. number of QoS types? Unlike the per-CPU case, there is no tradeoff between the peers. Both peers always support the same number of QoS types (64 on IPv4)] [Tobias: I agree with Steffen. This doesn't seem necessary and might even be confusing as reducing the number would not tell the peer what classes should actually be sent.] [Paul: I was hoping to send the desired number of different levels, not the theoretical maximum of used levels Note: The Fallback Child SA that is not bound to a single QoS is not counted as part of these numbers.

Protocol ID (1 octet) - MUST be 0. MUST be ignored if not 0. SPI Size (1 octet) - MUST be 0. MUST be ignored if not 0. Notify Message Type (2 octets) - set to [TBD3] Optional Payload Data. This value MAY be set to convey the local identity of the queue. The value SHOULD be a unique identifier and the peer SHOULD only use it for debugging purposes.

Protocol ID (1 octet) - MUST be 0. MUST be ignored if not 0. SPI Size (1 octet) - MUST be 0. MUST be ignored if not 0. Notify Message Type (2 octets) - set to [TBD4] Mandatory Payload Data. This value MUST be set to identify the QoS level. [Paul: Can we say 'one byte for each level of QoS included for this SA' ?] [Steffen: I don't understand that? Do we support more than one QoS type per SA? I think we need space to cover either a 6 bit IPv4 QoS type or a 20 bit IPv6 flow label.] [Tobias: Hm, one problem here is that CHILD_SAs can have traffic selectors of both address families. So how could we negotiate that we need a QoS type AND a flow label? Would that require two notifies (QOS_4|6_QUEUE_INFO types) or could we have two fields in the notify that may be set to 0? Or should that just not be allowed? I don't even know if it makes sense and whether QoS classes and flow labels are combinable in that way (I guess a dual-stack VoIP client would classify traffic in a comparable way for each family). And I also wonder if there is a mechanism to apply a flow label to an outer IPv4 header's TOS field and vice-versa. If multiple classes/labels should be supported per SA we could also send multiple notifies (but I guess that would mean that on-path routers had to treat all these classes/labels the same way, which begs the question why different values would get assigned to the packets in the first place).]

Implementations supporting per-CPU SAs SHOULD extend their local SPD selector, and the mechanism of on-demand negotiation that is triggered by traffic to include a CPU (or queue) identifier in their SADB_ACQUIRE message from the SPD to the IKE daemon. If the IKEv2 extension defined in this document is negotiated with the peer, a node which does not support receiving per-CPU SADB_ACQUIRE messages MAY initiate all its Child SAs immediately upon receiving the (only) SADB_ACQUIRE it will receive from the IPsec stack. Such implementations also need to be careful when receiving a Delete Notify request for a per-CPU Child SA, as it has no method to detect when it should bring up such a per-CPU Child SA again later. And bringing the deleted per-CPU Child SA up again immediately after receiving the Delete Notify might cause an infinite loop between the peers. Another issue of not bringing up all its per-CPU Child SAs is that if the peer acts similarly, the two peers might end up with only the Fallback SA without ever activating any per-CPU Child SAs. It is there for RECOMMENDED to implement per-CPU SADB_ACQUIRE messages. [ Antony: It would be nice to add manual/scripts for starting of connection and bringing up per-CPU SAs. It could be very simple, a external program decides to start a per-CPU SA. ] The minimum number of Child SAs negotiated should not be treated as the maximum number of allowed Child SAs. Peers SHOULD be lenient with this number to account for corner cases. For example, during Child SA rekeying, there might be a large number of additional Child SAs created before the old Child SAs are torn down. Similarly, when using on-demand Child SAs, both ends could trigger multiple Child SA requests as the initial packet causing the Child SA negotiation might have been transported to the peer via the Fallback SA where its reply packet might also trigger an on-demand Child SA negotiation to start. A peer may want to allow up to double the negotiated minimum number of Child SAs, and rely on idleness of Child SAs to tear down any unused Child SAs gradually to to reach an optimal number of Child SAs. Adding too many SAs may slow down per-packet SAD lookup. Implementations might support dynamically moving a per-CPU Child SAs from one CPU to another CPU. If this method is supported, implementations must be careful to move both the inbound and outbound SAs. If the IPsec endpoint is a gateway, it can move the inbound SA and outbound SA independently from each other. It is likely that for a gateway, IPsec traffic would be asymmetric. If the IPsec endpoint is the same host responsible for generating the traffic, the inbound and outbound SAs SHOULD remain as a pair on the same CPU. If a host previously skipped installing an outbound SA because it would be an unused duplicate outbound SA, it will have to create and add the previously skipped outbound SA to the SAD with the new CPU ID. The inbound SA may not have CPU ID in the SAD. Adding the outbound SA to the SAD requires access to the key material, whereas for updating the CPU selector on an existing outbound SAs. access to key material might not be needed. To support this, the IKE software might have to hold on to the key material longer than it normally would, as it might actively attempt to destroy key material from memorya that it no longer needs access to.

[TO DO]

[Note to RFC Editor: Please remove this section and the reference to before publication.] This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in . The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and must not be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations may exist. According to , "this will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. It is up to the individual working groups to use this information as they see fit". Authors are requested to add a note to the RFC Editor at the top of this section, advising the Editor to remove the entire section before publication, as well as the reference to .

Linux kernel XFRM XFRM-PCPU-v1 https://git.kernel.org/pub/scm/linux/kernel/git/klassert/linux-stk.git/log/?h=xfrm-pcpu-v1 An initial Kernel IPsec implementation of the per-CPU method. Alpha Implements Fallback Child SA and per-CPU Child SAs. It only supports the NETLINK API. The PFKEYv2 API is not supported. GPLv2 The Linux XFRM implementation added two additional attributes to support per-CPU SAs. There is a new attribute XFRMA_SA_PCPU, u32, for the SAD entry. This attribute should present on the outgoing SA, per-CPU Child SAs, starting from 0. This attribute MUST NOT be present on the Fallback XFRM SA. It is used by the kernel only for the outgoing traffic, (clear to encrypted). The incoming SAs, both the Fallback and the per-CPU SA, do not need XFRMA_SA_PCPU attribute. XFRM stack can not use CPU id on the incoming SA. The kernel internally sets the value to 0xFFFFFF for the incoming SA and the Fallback SA. However, one may add XFRMA_SA_PCPU to the incoming per-CPU SA to steer the ESP flow, to a specific Q or CPU e.g ethtool ntuple configuration. The SPD entry has new flag XFRM_POLICY_CPU_ACQUIRE. It should be set only on the "out" policy. The flag should be disabled when the policy is a trap policy, without SPD entries. After a successful negotiation of CPU_QUEUES, while adding the Fallback Child SA, the SPD entry can be updated with the XFRM_POLICY_CPU_ACQUIRE flag. When XFRM_POLICY_CPU_ACQUIRE is set, the XFRM_MSG_ACQUIRE generated will include the XFRMA_SA_PCPU attribute. Steffen Klassert steffen.klassert@secunet.com

The Libreswan Project pcpu-3 https://libreswan.org/wiki/XFRM_pCPU An initial IKE implementation of the per-CPU method. Alpha implements Fallback Child SA and per-CPU additional Child SAs GPLv2 TBD Libreswan Development: swan-dev@libreswan.org

The StrongSwan Project StrongSwan https://github.com/strongswan/strongswan/tree/per-cpu-sas-poc/ An initial IKE implementation of the per-CPU method. Alpha implements Fallback Child SA and per-CPU additional Child SAs GPLv2 StrongSwan use private space values for notifications CPU_QUEUES (40970) and QUEUE_INFO (40971). Tobias Brunner tobias@strongswan.org

The iproute2 Project iproute2 https://github.com/antonyantony/iproute2/tree/pcpu-v1 Implemented the per-CPU attributes for the "ip xfrm" command. Alpha GPLv2 TBD Antony Antony antony.antony@secunet.com

This document defines four new IKEv2 Notify Message Type payloads for the IANA "IKEv2 Notify Message Types - Status Types" registry.