XFRM pCPU: Difference between revisions
No edit summary |
No edit summary |
||
(19 intermediate revisions by 2 users not shown) | |||
Line 1: | Line 1: | ||
Goal: scalable IPsec throughput with multiple CPUs(without IPsec HW offload) | Goal: scalable IPsec throughput with multiple CPUs(without IPsec HW offload) | ||
The idea of per-CPU SA in the outgoing direction was discussed at Linux IPsec workshop March 2019, in Prague. A small group of people worked on a prototype of user space(IKE), Libreswan, and Linux kernel, XFRM. The libreswan implementation | Linux IPsec use a pair of SAs. Here the new idea is | ||
install identical SAs, same traffic selectors and different SPI, per CPU. Let say there are 2 CPUs in use, we would add 3 pairs SAs. In the out going direction two, one on each CPU and a last resort one. The per-CPU SA can increase IPsec throuput. The idea of per-CPU SA in the outgoing direction was discussed at Linux IPsec workshop March 2019, in Prague. Steffen Klassert gave a presentation. | |||
A small group of people worked on a prototype of user space(IKE), Libreswan, and Linux kernel, XFRM. The libreswan implementation call this option "clones". In the Linux kernel it is call pCPU. These names may change as we adopt the idea to include SA per TOS bits (8-64) TCP/UDP DST port hashing(some how configure n flows to be distributed). | |||
The tests were performed without using IPsec HW hardware offload to separate the performance numbers of per-CPU SA's from hardware interaction. | The tests were performed without using IPsec HW hardware offload to separate the performance numbers of per-CPU SA's from hardware interaction. | ||
Line 15: | Line 18: | ||
</pre> | </pre> | ||
The initial measurements | The initial measurements are: '''15-16 Gbps with 3 CPU's. Also about 6 Gbps with single CPU''' | ||
== Test setup using libreswan == | == Test setup using libreswan + pCPU patched kernel == | ||
=== Linux kernel source with pCPU support === | === Linux kernel source code with pCPU support === | ||
git clone - | git clone --single-branch --branch xfrm-pcpu-v1 https://git.kernel.org/pub/scm/linux/kernel/git/klassert/linux-stk.git | ||
== Kernel / xfrm future plans == | == Kernel / xfrm future plans == | ||
Line 29: | Line 32: | ||
* Ben would like to add feature bind a sub sa to a head SA? | * Ben would like to add feature bind a sub sa to a head SA? | ||
* seems to need latest iproute2 otherwise "ip x s" may loop. | * seems to need latest iproute2 otherwise "ip x s" may loop. | ||
=== Libreswan with clones support === | === Libreswan with clones support === | ||
<pre> | <pre> | ||
git clone --single-branch --branch clones- | git clone --single-branch --branch clones-4 https://github.com/antonyantony/libreswan | ||
</pre> | </pre> | ||
Line 68: | Line 70: | ||
</pre> | </pre> | ||
NOTE: Both SA #3 and #4 should have outgoing traffic on it. | NOTE: Both SA #3 and #4 should have outgoing traffic on it and #2 should not have any traffic on dual cpu system. | ||
== Future Libreswan plans == | == Future Libreswan plans == | ||
Line 117: | Line 119: | ||
== Supported Work loads == | == Supported Work loads == | ||
As of Nov 2019, to make full use of the cloned SA's, network traffic load has to be distributed over different CPU's to take advantage of the | As of Nov 2019, to make full use of the cloned SA's, network traffic load has to be distributed over different CPU's to take advantage of the pCPU feature. | ||
If the traffic is generated on the IPsec machine itself, the application(s) need to be writing their traffic (eg using send() of write() syscalls) running on different CPU's. This can often be | If the traffic is generated on the IPsec machine itself, the application(s) need to be writing their traffic (eg using send() of write() syscalls) running on different CPU's. This can often be steered using the taskset or numctl commands. | ||
For forwarded traffic, you need RSS support on the NIC receiving the clear text. RSS will steer different flows onto different CPUs and this use | For forwarded traffic, you need RSS support on the NIC receiving the clear text. RSS will steer different flows onto different CPUs and this use SA assigned the CPU. If all the traffic consists of one single flow, the traffic will not be distributed over different CPUs - to avoid out of order delivery. | ||
=== Can we distribute 4 tuple flows locally generated? === | === Can we distribute 4 tuple flows locally generated? === | ||
Line 127: | Line 129: | ||
yes. See above. | yes. See above. | ||
== Receiver | == Receiver Side Scaling - RSS support == | ||
Receive Side Scaling [https://www.kernel.org/doc/Documentation/networking/scaling.txt (RSS)] is required for pCPU. The receiver NIC should be able to steer different flows, based on SPI, into separate queues (CPUs) to prevent the receiver from getting overwhelmed. We used Mellanex CX4 to test. Some cards initially tested did not seems to support RSS for ESP flows, instead only TCP and UDP. While figuring out RSS for these cards we tried a bit different approach. ESP in UDP encapsulation, along with ESP in UDP GRO patches we could see the flows getting distributed on the receiver. And later on in Nov 2019 kernel version 5.5 ML5 drivers seems to support ESP. [https://community.mellanox.com/s/article/Bluefield-IP-Forwarding-and-IPSEC-SPI-RSS Mellonox RSS]. | |||
=== RSS Commands === | |||
Enable GRO. | Enable GRO. ideally you should be able to run the following, | ||
ideally you should be able to run the following, | |||
<pre> ethtool -N <nic> rx-flow-hash esp4 </pre> | <pre> ethtool -N <nic> rx-flow-hash esp4 </pre> | ||
Another argument is if the NIC agnostic the 16 bits of SPI, of ESP packet, is aligned with UDP port number and should provide enough entropy. | Another argument is if the NIC agnostic the 16 bits of SPI, of ESP packet, is aligned with UDP port number and should provide enough entropy. | ||
<pre> ethtool -N eno2 rx-flow-hash udp4 sdfn </pre> | <pre> ethtool -N eno2 rx-flow-hash udp4 sdfn </pre> | ||
=== Mellanox support (YES) === | |||
there is support for RSS and ESP4 and ESP6. | |||
The ntuple pinning an ESP flow to specific CPU, incoming SA, is not supported. | |||
<pre> | |||
ethtool --config-ntuple enp3s0f0 flow-type esp4 src-ip 192.168.1.1 dst-ip 192.168.1.2 spi 0xffffffff action 4 loc 10 | |||
</pre> | |||
for a UDP flow, ntuple filtering would look like : | |||
<pre> | |||
ethtool --config-ntuple <interface name> flow-type udp4 src-ip 192.168.1.1 dst-ip 192.168.10.2 src-port 2000 dst-port 2001 action 2 loc 33 | |||
</pre> | |||
[https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mellanox/mlx5/core/en_fs_ethtool.c en_fs_ethtool.c] | |||
<pre> | |||
case ESP_V4_FLOW: | |||
return MLX5E_TT_IPV4_IPSEC_ESP; | |||
</pre> | |||
=== [[XFRM_pCPU_RSS]] details === | |||
== Future research/ideas == | == Future research/ideas == | ||
Line 146: | Line 169: | ||
=== Install SA only for the CPUs that has workload === | === Install SA only for the CPUs that has workload === | ||
One line summary : Libreswan terminology clone connection with auto=start. | |||
In the current model libreswan is configured with clones=N. When the connection comes up libreswan, as part of IKE_AUTH exchange, negotiates the head SA. And immediately negotiate N sub SAs using CREATE_CHILD_SA, "New Child SA, RFC 7296 #1.3.1". To negotiate N sub SAs libreswan need N CREATE_CHILD_SA exchanges, or round trip times. Note you could do this simultaneously when IKE daemon supports IKE window grater than one. Next question is do we need a SA for all CPUs? what if the workloads are only on few CPUs, say CPU 2 and CPU 4. In the current implementation we can't optimize this. | In the current model libreswan is configured with clones=N. When the connection comes up libreswan, as part of IKE_AUTH exchange, negotiates the head SA. And immediately negotiate N sub SAs using CREATE_CHILD_SA, "New Child SA, RFC 7296 #1.3.1". To negotiate N sub SAs libreswan need N CREATE_CHILD_SA exchanges, or round trip times. Note you could do this simultaneously when IKE daemon supports IKE window grater than one. Next question is do we need a SA for all CPUs? what if the workloads are only on few CPUs, say CPU 2 and CPU 4. In the current implementation we can't optimize this. | ||
Line 164: | Line 189: | ||
This need a bit of extra code both in kernel and in libreswan, option would auto=ondemand | This need a bit of extra code both in kernel and in libreswan, option would auto=ondemand | ||
2. Connection initiation is not packet triggered instead administratively started "ipsec auto --up eastnet-westnet. Then pluto | 2. Connection initiation is not packet triggered instead administratively started "ipsec auto --up eastnet-westnet. Then pluto will install n+1 SA. | ||
Alternative is install only head SA. And based on traffic install sub SA. In this case when a cpu hit the traffic and look sub sa find_state() will install larval to the list as a sub sa. |
Latest revision as of 19:54, 21 June 2021
Goal: scalable IPsec throughput with multiple CPUs(without IPsec HW offload)
Linux IPsec use a pair of SAs. Here the new idea is install identical SAs, same traffic selectors and different SPI, per CPU. Let say there are 2 CPUs in use, we would add 3 pairs SAs. In the out going direction two, one on each CPU and a last resort one. The per-CPU SA can increase IPsec throuput. The idea of per-CPU SA in the outgoing direction was discussed at Linux IPsec workshop March 2019, in Prague. Steffen Klassert gave a presentation.
A small group of people worked on a prototype of user space(IKE), Libreswan, and Linux kernel, XFRM. The libreswan implementation call this option "clones". In the Linux kernel it is call pCPU. These names may change as we adopt the idea to include SA per TOS bits (8-64) TCP/UDP DST port hashing(some how configure n flows to be distributed).
The tests were performed without using IPsec HW hardware offload to separate the performance numbers of per-CPU SA's from hardware interaction.
Results
The test result, as of Nov 2019, show an aggregated throughput increase that is linearly with the number of CPUs.
We tested using physical servers, using Mellonex CX4 NIC. These NICs (using the latest Linux driver CX5) support RSS for ESP. In the tests, the clear text traffic was generated using a hardware traffic generator which sends traffic to the first IPsec gateway. The IPsec gateway encrypts the traffic and send it to the second IPsec gateway. That gateway decrypts the traffic into clear text and forwards the traffic to the receiving end of the traffic generator.
|Traffic Generator Sender|-----|IPsec Gateweay #1|=====ipsec 40Gbps link====|IPsec Gateway #2|---|Traffic Generator Receiver|
The initial measurements are: 15-16 Gbps with 3 CPU's. Also about 6 Gbps with single CPU
Test setup using libreswan + pCPU patched kernel
Linux kernel source code with pCPU support
git clone --single-branch --branch xfrm-pcpu-v1 https://git.kernel.org/pub/scm/linux/kernel/git/klassert/linux-stk.git
Kernel / xfrm future plans
- Release private branch at Steffen's repository for wider testing.
- Kernel support for IPsec rekey. One could rekey in any order - either a head SA or the sub SA.
- One main difference is when installing a new sub SA during a rekey, add_sa() would delete the old sub SA. Libreswan should not try to delete it. Or convince to Steffen to allow deleting an old sub SA.
- Ben would like to add feature bind a sub sa to a head SA?
- seems to need latest iproute2 otherwise "ip x s" may loop.
Libreswan with clones support
git clone --single-branch --branch clones-4 https://github.com/antonyantony/libreswan
Sample config | ipsec.conf
conn westnet-eastnet rightid=@east leftid=@west left=192.1.2.45 right=192.1.2.23 rightsubnet=192.0.2.0/24 leftsubnet=192.0.1.0/24 authby=secret clones=2 auto=add nic-offload=no
Initiate the connection and test the multiple CPU IPsec SA's:
ipsec auto --up westnet-eastnet taskset 0x1 ping -n -c 2 -I 192.0.1.254 192.0.2.254 taskset 0x2 ping -n -c 2 -I 192.0.1.254 192.0.2.254 ipsec trafficstatus ipsec trafficstatus 006 #2: "westnet-eastnet-0", type=ESP, add_time=1234567890, inBytes=0, outBytes=0, id='@east' 006 #4: "westnet-eastnet-1", type=ESP, add_time=1234567890, inBytes=168, outBytes=168, id='@east' 006 #3: "westnet-eastnet-2", type=ESP, add_time=1234567890, inBytes=168, outBytes=168, id='@east'
NOTE: Both SA #3 and #4 should have outgoing traffic on it and #2 should not have any traffic on dual cpu system.
Future Libreswan plans
- Current support using clones=n requires both endpoints to have the same clone number. Future plan is to allow asymmetric configuration, such as one side using 8 clones on 4 CPUs and the other side using using 12 clones on 12 CPUs
- Match Rekey support behaviour between kernel and libreswan. Deleting sub and head SA during a rekey procedure needs to be worked out with kernel
- Complete support for ipsec auto --down and delete
- Prevent clone instance on their own to be manipulated using ipsec auto add|delete|down
- Ensure interoperability against IPsec gateways that do not support clone SA's, such as previous versions of libreswan without clone support.
Linux kernel XFRM details
The changes are to the SAdb entry aka state, or SA. The new concept is head SA and sub SA for the outgoing direction. These are supported with additional XFRMA_SA_EXTRA_FLAGS, and attributes of the SADB entry. SPDB, aka, policy can either point to the head SA if your policy has SPI in it. Libreswan installed policy do not have SPI in it. You need only one policy. Note libreswan might install multiple identical polices. And this works too.
The head SA is a catch all SA. It is not associated to a specific CPU. When there is N CPU, install N+1 SAs. One head SA and N sub SAs.
To add SADB entry you need extra attributes to the netlink calls, for methods XFRM_MSG_NEWSA, XFRM_MSG_UPDSA, and XFRM_MSG_GETSA, only for the outgoing SA.Installing incoming or receiving SA to the kernel remain un-changed.
XFRM_MSG_NEWSA head SA
XFRMA_SA_EXTRA_FLAGS includes the XFRM_SA_PCPU_HEAD flag
XFRM_MSG_NEWSA sub SA
XFRMA_SA_EXTRA_FLAGS includes the XFRM_SA_PCPU_SUB and the new attribute XFRMA_SA_PCPU set to the <cpu id>. CPU SA ID start from 0, and it is a u32.
XFRM_MSG_UPDSA
Both the head SA and the sub SAs need extra attributes:
- The head SA sets the XFRMA_SA_EXTRA_FLAGS to XFRM_SA_PCPU_HEAD
- The sub SA sets the XFRMA_SA_EXTRA_FLAGS to XFRM_SA_PCPU_SUB and XFRMA_SA_PCPU is set to <sub-sa-id>.
XFRM_MSG_GETSA
This call only requires changes for sub SAs:
- The sub SA XFRMA_SA_EXTRA_FLAGS is set to XFRM_SA_PCPU_SUB and XFRMA_SA_PCPU is set to <sub-sa-id>.
- Set XFRMA_SRCADDR to the src addr
This is the call used by libreswan "ipsec trafficstatus" without this changes it will not find the sub SAs.
when nCPU < nSAs
When there are 4 CPUs and the number of clones configured is 8, because the other end has 8 CPUs. The head SA's list only has 4 places for sub SAs. Libreswan should install only 4 outbound sub SA's and install 8 inbound sub SA's. This is a local policy and not affecting the remote IPsec peer. From the view of the remote peer, 4 inbound SA's appear to be unused. The remote peer can still use all its 8 outbound SAs. IPsec SA's are negotiated as as bundle of one inbound and one outbound SA. Both ends commit to receiving on their inbound SA's, but are free to decide on which outbound SA's they will send traffic. This setup is therefor compliant with RFC 7296.
In our example above, the IKE daemon on the 4-CPU machine has a list of all 8 SA bundles, but will have installed only 4 outbound SA's along with the 8 inbound SA's in the Linux kernel. The "ip xfrm state" will show this.
Supported Work loads
As of Nov 2019, to make full use of the cloned SA's, network traffic load has to be distributed over different CPU's to take advantage of the pCPU feature.
If the traffic is generated on the IPsec machine itself, the application(s) need to be writing their traffic (eg using send() of write() syscalls) running on different CPU's. This can often be steered using the taskset or numctl commands.
For forwarded traffic, you need RSS support on the NIC receiving the clear text. RSS will steer different flows onto different CPUs and this use SA assigned the CPU. If all the traffic consists of one single flow, the traffic will not be distributed over different CPUs - to avoid out of order delivery.
Can we distribute 4 tuple flows locally generated?
yes. See above.
Receiver Side Scaling - RSS support
Receive Side Scaling (RSS) is required for pCPU. The receiver NIC should be able to steer different flows, based on SPI, into separate queues (CPUs) to prevent the receiver from getting overwhelmed. We used Mellanex CX4 to test. Some cards initially tested did not seems to support RSS for ESP flows, instead only TCP and UDP. While figuring out RSS for these cards we tried a bit different approach. ESP in UDP encapsulation, along with ESP in UDP GRO patches we could see the flows getting distributed on the receiver. And later on in Nov 2019 kernel version 5.5 ML5 drivers seems to support ESP. Mellonox RSS.
RSS Commands
Enable GRO. ideally you should be able to run the following,
ethtool -N <nic> rx-flow-hash esp4
Another argument is if the NIC agnostic the 16 bits of SPI, of ESP packet, is aligned with UDP port number and should provide enough entropy.
ethtool -N eno2 rx-flow-hash udp4 sdfn
Mellanox support (YES)
there is support for RSS and ESP4 and ESP6.
The ntuple pinning an ESP flow to specific CPU, incoming SA, is not supported.
ethtool --config-ntuple enp3s0f0 flow-type esp4 src-ip 192.168.1.1 dst-ip 192.168.1.2 spi 0xffffffff action 4 loc 10
for a UDP flow, ntuple filtering would look like :
ethtool --config-ntuple <interface name> flow-type udp4 src-ip 192.168.1.1 dst-ip 192.168.10.2 src-port 2000 dst-port 2001 action 2 loc 33
case ESP_V4_FLOW: return MLX5E_TT_IPV4_IPSEC_ESP;
XFRM_pCPU_RSS details
Future research/ideas
- Test with SR IOV and virtualisation(KVM): need systems with NIC that support SR IOV and RSS for ESP or at least UDP.
- Software RSS https://www.linux-kvm.org/page/Multiqueue
- can IKE daemon use other flow distribution methods based on SPI??? DPDK???
- another way of flow control??? https://doc.dpdk.org/dts/test_plans/link_flowctrl_test_plan.html
Install SA only for the CPUs that has workload
One line summary : Libreswan terminology clone connection with auto=start.
In the current model libreswan is configured with clones=N. When the connection comes up libreswan, as part of IKE_AUTH exchange, negotiates the head SA. And immediately negotiate N sub SAs using CREATE_CHILD_SA, "New Child SA, RFC 7296 #1.3.1". To negotiate N sub SAs libreswan need N CREATE_CHILD_SA exchanges, or round trip times. Note you could do this simultaneously when IKE daemon supports IKE window grater than one. Next question is do we need a SA for all CPUs? what if the workloads are only on few CPUs, say CPU 2 and CPU 4. In the current implementation we can't optimize this.
Lets say you have 72 cores, and mysql running only on CPU 2 and CPU 4 need sub SA. One idea is add the sa based acquire message. The message would need new attribute CPU id.
There are two cases. 1 There is no head SA, Just a XFRM policy entry. The first packet of a new flow arrives on CPU 4 that match policy. The XFRM creates a larval state with a timeout, default 30s, sends an acquire message to the IKE daemon, with the new attribute CPU id in it. Pluto, the IKE daemon, start a new IKE negotiation. While the negotiation, IKE_INIT and IKE_AUTH, is going on the traffic is dropped or first packet is cached. The pluto first negotiates the head SA, in IKE_AUTH, and installs the head SA. If the flow is continues next packets will use the head SA. While pluto goes ahead and negotiates, installs, sub SA for CPU 4. Then on sub SA for CPU 4 will be used by traffic, ie traffic will switch to this sub SA from head SA.
Now new flow arrive on CPU 2. Here is a bit of XFRM magic. (detail, larval may have expired or not, it does not matter). The policy is hit, state look up will find the head sa, and further look up for sub SA for CPU 2 will not find an SA. There is no sub SA 2. Then xfrm will create an acquire message with CPU id 2 and send this message to pluto via netlink. And state find returns head SA. The traffic would use the head SA.
NOTE: there may be or may not be a larval state, in either case xfrm code should send an acquire message.
When pluto receives an acquire with the CPU id 2, it will negotiate a new sub SA and install to CPU 2. Then traffic will switch to sub SA 2.
The advantage of this method is SAs are installed only for the CPU/core/thread that has that to encrypt.
This need a bit of extra code both in kernel and in libreswan, option would auto=ondemand
2. Connection initiation is not packet triggered instead administratively started "ipsec auto --up eastnet-westnet. Then pluto will install n+1 SA.
Alternative is install only head SA. And based on traffic install sub SA. In this case when a cpu hit the traffic and look sub sa find_state() will install larval to the list as a sub sa.