Difference between revisions of "VPP/SecurityGroups"

From fd.io
< VPP
Jump to: navigation, search
(Work list)
(Work list)
Line 287: Line 287:
 
| Ole
 
| Ole
 
| 0
 
| 0
| WIP
+
| Done
 
| [https://jira.fd.io/browse/VPP-513 VPP-513]
 
| [https://jira.fd.io/browse/VPP-513 VPP-513]
 
|-
 
|-
Line 293: Line 293:
 
| Andrew
 
| Andrew
 
| 0
 
| 0
| WIP
+
| Done
 
| [https://jira.fd.io/browse/VPP-514 VPP-514]
 
| [https://jira.fd.io/browse/VPP-514 VPP-514]
 
|-
 
|-
Line 337: Line 337:
 
   --- MVP ---
 
   --- MVP ---
 
   4a. Plumbing for stateful sessions from ACL plugin (to be able to specify “match and track” (“permit and create the forward/return session”) -> Andrew - done.
 
   4a. Plumbing for stateful sessions from ACL plugin (to be able to specify “match and track” (“permit and create the forward/return session”) -> Andrew - done.
   4b. Stateful session tracking - timeouts -> Andrew - ***WIP*** .
+
   4b. Stateful session tracking - timeouts -> Andrew - Done.
   4c. Stateful session tracking - lightweight TCP state -> Andrew - ***WIP***
+
   4c. Stateful session tracking - lightweight TCP state -> Andrew - Done
 
   5. MACIP(L2) rules -> Andrew - done.
 
   5. MACIP(L2) rules -> Andrew - done.
 
   6. Code cleanups -> Andrew
 
   6. Code cleanups -> Andrew

Revision as of 13:32, 6 December 2016

VPP Security Groups

Introduction

Features are tracked as they are developed in the following VPP-427.

In-progress development is done on github: ACL branch

The first version of the plugin is committed via change 3423

ACL Node architecture

Initial performance tests

The first version of the ACL plugin was explicitly focused on getting things "correct" as the first priority, even if at some expense of getting them "fast". But understanding the performance is very important, so we did limited performance testing. The performance testing was done using MoonGen, running on the same host as VPP. The VPP was run by doing "make release-build; make release-plugins; make release-run" process. The VPP is configured via VAT.

The performance was done by testing with 200000 unidirectional UDP streams, by submitting the line-rate 10Gbps of traffic by MoonGen (14.88Mpps of 64-byte packets) and observing the amount of traffic received on the other side.

First, we test the baseline configuration which just doing briding between the two interfaces.

 sw_interface_set_flags TenGigabitEthernet81/0/0 admin-up link-up
 sw_interface_set_flags TenGigabitEthernet81/0/1 admin-up link-up
 bridge_domain_add_del bd_id 42 flood 1 uu-flood 1 forward 1 learn 1 arp-term 0
 sw_interface_set_l2_bridge TenGigabitEthernet81/0/0 bd_id 42
 sw_interface_set_l2_bridge TenGigabitEthernet81/0/1 bd_id 42


This configuration exhibited 9.52 Mpps performance.

Then we added a trivial case of a one-line "permit" ACL which is checked on input and output of the packet path, using the following VAT commands:

 acl_add_replace permit
 acl_interface_add_del sw_if_index 1 add input acl 0
 acl_interface_add_del sw_if_index 1 add output acl 0
 acl_interface_add_del sw_if_index 2 add input acl 0
 acl_interface_add_del sw_if_index 2 add output acl 0

This configuration causes two very simple ACL checks - on input of the packet path and on the output, see the below trace:

 00:02:12:167665: dpdk-input
   TenGigabitEthernet81/0/0 rx queue 0
   buffer 0x5358: current data 0, length 60, free-list 0, totlen-nifb 0, trace 0x0
   PKT MBUF: port 0, nb_segs 1, pkt_len 60
     buf_len 2176, data_len 60, ol_flags 0x180, data_off 128, phys_addr 0x6ee4d640
     packet_type 0x0
     Packet Offload Flags
   IP4: 01:02:03:04:05:06 -> 07:08:09:0a:0b:0c
   UDP: 10.0.80.47 -> 10.1.0.10
     tos 0x00, ttl 64, length 46, checksum 0x1686
     fragment id 0x0000
   UDP: 1234 -> 319
     length 26, checksum 0x956f
 00:02:12:167682: ethernet-input
   IP4: 01:02:03:04:05:06 -> 07:08:09:0a:0b:0c
 00:02:12:167691: l2-input
   l2-input: sw_if_index 1 dst 07:08:09:0a:0b:0c src 01:02:03:04:05:06
 00:02:12:167693: l2-input-classify
   l2-classify: sw_if_index 1, table 0, offset 0, next 9
 00:02:12:167700: acl-plugin-in
   ACL_IN: sw_if_index 1, next index 10, match: inacl 0 rule 0 trace_bits 00000000
 00:02:12:167709: l2-learn
   l2-learn: sw_if_index 1 dst 07:08:09:0a:0b:0c src 01:02:03:04:05:06 bd_index 1
 00:02:12:167712: l2-flood
   l2-flood: sw_if_index 1 dst 07:08:09:0a:0b:0c src 01:02:03:04:05:06 bd_index 1
 00:02:12:167714: l2-output
   l2-output: sw_if_index 2 dst 07:08:09:0a:0b:0c src 01:02:03:04:05:06
 00:02:12:167716: l2-output-classify
   l2-classify: sw_if_index 2, table 6, offset 0, next 5
 00:02:12:167723: acl-plugin-out
   ACL_OUT: sw_if_index 2, next index 4, match: outacl 0 rule 0 trace_bits 00000000
 00:02:12:167734: TenGigabitEthernet81/0/1-output
   TenGigabitEthernet81/0/1
   IP4: 01:02:03:04:05:06 -> 07:08:09:0a:0b:0c
   UDP: 10.0.80.47 -> 10.1.0.10
     tos 0x00, ttl 64, length 46, checksum 0x1686
     fragment id 0x0000
   UDP: 1234 -> 319
     length 26, checksum 0x956f
 00:02:12:167736: TenGigabitEthernet81/0/1-tx
   TenGigabitEthernet81/0/1 tx queue 0
   buffer 0x5358: current data 0, length 60, free-list 0, totlen-nifb 0, trace 0x0
   IP4: 01:02:03:04:05:06 -> 07:08:09:0a:0b:0c
   UDP: 10.0.80.47 -> 10.1.0.10
     tos 0x00, ttl 64, length 46, checksum 0x1686
     fragment id 0x0000
   UDP: 1234 -> 319
     length 26, checksum 0x956f

This configuration exhibited the performance of 4.62 Mpps.

To test the impact of the linear match, we add lines to ACL one by one:

 acl_add_replace 0 permit src 1.1.1.1/32,permit

performance: 4.50Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32,permit

performance: 4.34 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32,permit

performance: 4.19 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, permit

performance: 4.04 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, permit

performance: 3.90 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, permit

performance: 3.77 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32,  permit

performance: 3.64 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, permit

performance: 3.55 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, permit

performance: 3.23 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, permit

performance: 3.33 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, src 1.1.1.11/32, permit

performance: 3.24 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, src 1.1.1.11/32, src 1.1.1.12/32, permit

performance: 3.14 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, src 1.1.1.11/32, src 1.1.1.12/32, src 1.1.1.13/32, permit

performance: 3.06 Mpps

 acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, src 1.1.1.11/32, src 1.1.1.12/32, src 1.1.1.13/32, src

1.1.1.14/32, permit

performance: 2.85 Mpps


We can see that the performance impact is not strictly linear, but it gives the impression.

Changing the last match to "permit and reflect" in order to cause the stateful processing path:

acl_add_replace 0 permit src 1.1.1.1/32,src 1.1.1.2/32, src 1.1.1.3/32, src 1.1.1.4/32, src 1.1.1.5/32, src 1.1.1.6/32, src 1.1.1.7/32, src 1.1.1.8/32, src 1.1.1.9/32, src 1.1.1.10/32, src 1.1.1.11/32, src 1.1.1.12/32, src 1.1.1.13/32, src 1.1.1.14/32, permit+reflect

performance: 3.68 Mpps

We can see that the performance of the stateful path is less than the performance of the stateless path with small ACLs - this is most probably due to the current code organization of session tracker nodes, which use some code sharing between the nodes to ease the maintenance of it.


Removing all ACLs:

 acl_interface_add_del sw_if_index 1 del input acl 0
 acl_interface_add_del sw_if_index 1 del output acl 0
 acl_interface_add_del sw_if_index 2 del input acl 0
 acl_interface_add_del sw_if_index 2 del output acl 0

We get to 9.47 Mpps.

Let's test multiple ACL matching:

 acl_add_replace src 1.1.1.1/32
 acl_add_replace src 1.1.1.1/32
 acl_add_replace src 1.1.1.1/32
 acl_add_replace src 1.1.1.1/32
 acl_add_replace src 1.1.1.1/32

First same condition as before, with just one ACL check:

 acl_interface_set_acl_list sw_if_index 1 input 0 output 0
 acl_interface_set_acl_list sw_if_index 2 input 0 output 0


performance: 4.67 Mpps

Two ACLs:

 acl_interface_set_acl_list sw_if_index 1 input 1 0 output 1 0
 acl_interface_set_acl_list sw_if_index 2 input 1 0 output 1 0

performance: 4.16 Mpps

Three ACLs:

 acl_interface_set_acl_list sw_if_index 1 input 2 1 0 output 2 1 0
 acl_interface_set_acl_list sw_if_index 2 input 2 1 0 output 2 1 0

performance: 3.73 Mpps

Four ACLs:

 acl_interface_set_acl_list sw_if_index 1 input 3 2 1 0 output 3 2 1 0
 acl_interface_set_acl_list sw_if_index 2 input 3 2 1 0 output 3 2 1 0

performance: 3.38 Mpps

Five ACLs:

 acl_interface_set_acl_list sw_if_index 1 input 4 3 2 1 0 output 4 3 2 1 0
 acl_interface_set_acl_list sw_if_index 2 input 4 3 2 1 0 output 4 3 2 1 0

Performance: 3.10 Mpps

Six ACLs:

 acl_interface_set_acl_list sw_if_index 1 input 5 4 3 2 1 0 output 5 4 3 2 1 0
 acl_interface_set_acl_list sw_if_index 2 input 5 4 3 2 1 0 output 5 4 3 2 1 0

Performance: 2.85 Mpps

Convert to last "permit" to stateful:

 acl_add_replace 0 permit+reflect

Performance: 3.67 Mpps

These two performance axis might combine 0 i.e. if the ACL 5 in the above test were to be long and not match, the performance will be worse than just with that ACL and worse than with 6 trivial ACLs.

Removing all ACLs:

 acl_interface_set_acl_list sw_if_index 1 
 acl_interface_set_acl_list sw_if_index 2 

The performance is 9.45 Mpps.


MACIP ACLs


 macip_acl_add permit
 macip_acl_interface_add_del sw_if_index 1 add acl 0

Performance: 7.30 Mpps

 macip_acl_add permit ip 128.1.0.0/7, permit ip 10.0.0.0/8
 macip_acl_interface_add_del sw_if_index 1 add acl 1

Performance: 7.31 Mpps (the hit on the first classify table)

 macip_acl_add permit ip 128.1.0.0/9, permit ip 10.0.0.0/8
 macip_acl_interface_add_del sw_if_index 1 add acl 1

Performance: 7.31 Mpps (the hit on the first classify table)

(more testing with MACIP ACLs should be done)

Requirements

  • Support classifiers/filters on any interface type (bridged / routed)
  • Filter on IP-addresses with address mask or prefix length (IPv4 and IPv6)
  • Filter on source and destination TCP/UDP port ranges
  • Filter on source and destination L2 MAC addresses
  • Support IPv6 with extension headers present
  • Support fragmented packets and unknown transport layer headers
  • Combinations of the above filters (e.g. MAC + IP)
  • Filters on ingress and egress interfaces
  • Stateful firewall. No application layer filtering.

Work list

Task Owner Priority Status Description
API definition Ole 0 Done VPP-513
Connection tracker Andrew 0 Done VPP-514
Stateful ACLs 0 VPP-515
ACL policy matching node (MVP) Andrew 0 Done input output
Direct classifier policy matching -
Control Plane test code (new framework) Pavel 0 WIP
Data Plane tests (performance + scale) 0


 1. Python tests/examples -> Ole + Pavel
 2a. IPv4 matching in all plugin -> Andrew - done.
 2b. make it “deny by default” -> Andrew - done.
 2c. port range support -> Andrew - done.
 2d. ICMP type/code matching -> Andrew - done.
 3. Performance testing -> Andrew - done.
 --- MVP ---
 4a. Plumbing for stateful sessions from ACL plugin (to be able to specify “match and track” (“permit and create the forward/return session”) -> Andrew - done.
 4b. Stateful session tracking - timeouts -> Andrew - Done.
 4c. Stateful session tracking - lightweight TCP state -> Andrew - Done
 5. MACIP(L2) rules -> Andrew - done.
 6. Code cleanups -> Andrew

 
 PHASE2:
 A. ACL/Sessions support for L3 (routed) mode - (big) !
 B. Can we implement the ACL match purely in terms of classifier tables ? How expensive/(in)efficient that would be ?
 C. Extension header handling during the slow path lookup - easy in ACL plugin
 D. classifier match for the sessions with extension headers - currently no extension headers supported

API

API file as implemented

MACIP (formerly "L2") API

MACIP (renamed to avoid confusion) is an ingress-only ACL which permits the traffic based on a mix of MAC and IP address matches.

The use of this mechanism is to prevent spoofing.

API file as implemented

API as implemented supports MAC address masks and prefixes, however, be aware: the current implementation is done using chained classifier tables, so each variation of the masks/prefix lengths means an extra table and hence the performance impact.

These filters are per-packet so you will want to care for performance.

For best performance, use the exact match MAC mask (ff:ff:ff:ff:ff:ff) and the maximum prefix length (/32 for IPv4 and /128 for IPv6).

Design and prototyping

The ACL matching is implemented in this phase as a simple array search, under the assumption that given the rules are per-port, the rule list will be small.

The redirection of the traffic to the node performing the ACL match is done by installing an empty L2 classifier table whose "miss-next" index diverts the traffic to the node.

The ACL match node can also redirect the traffic to the stateful-session setup node (by having a "permit" = 2 in the ACE), which will create the session on that interface.

...TBD: more details...

CLI

Every activity with the ACL must be done via the API. The plugins do not add any CLI at this point.

Examples

YANG model

Open Issues

Closed Issues

  • Security Group use case specific API. Done in a plugin. Two plugins, to be exact - one to ACL match, one to track sessions.

Existing functionality

The existing functionality has a classifier (https://wiki.fd.io/view/VPP/Introduction_To_N-tuple_Classifiers) matching.

As the above document explains, the classifier is a series of chained tables, with each table having a specific mask, but this mask is the same for all entries.

This has been tested to happen in the L2 bridged case (test case: http://stdio.be/vpp/t/aytest-bridge-tap-py.txt).

Therefore, if we have an example policy:

 nova secgroup-create test-secgroup test
 nova secgroup-add-rule test-secgroup icmp -1 -1 0.0.0.0/0
 nova secgroup-add-rule test-secgroup tcp 22 22 0.0.0.0/0

So, assuming we match with offset 0 (from the beginning of the packet) the mask will look like this for the first line:

 000000000000 000000000000 0000 00 00 0000 0000 0000 00 FF 0000 00000000 00000000  00 00 0000 0000 
   eth dst      eth src    et   ihl t  len id    fo ttl pr  cs   ip4src   ip4dst    t  c  cs   id
   +-------- L2 ---------------+----------- L3 IPv4 ------------------------------+--------L4 ICMP -----+

For the TCP matching on port 22 it will look as follows:

 000000000000 000000000000 0000 00 00 0000 0000 0000 00 FF 0000 00000000 00000000  0000 FFFF 00000000 00000000 0000 0000 0000 0000
   eth dst      eth src    et   ihl t  len id    fo ttl pr  cs   ip4src   ip4dst    sp  dp    seq      ack      fl  win   cs   urg
   +-------- L2 ---------------+----------- L3 IPv4 ------------------------------+--------L4 TCP ---------------------------------+


(One would need to round up the number of bytes to the nearest 16-byte boundary that makes sense)

For IPv6 assuming no extension headers, it will look similar, with the L3 header being the IPv6 one:


 000000000000 000000000000 0000 0 00 00000 0000 FF 00 00000000000000000000000000000000 00000000000000000000000000000000 00 00 0000 0000 
   eth dst      eth src    et   v TC  fll  len  nh hl             ipv6 src                   ipv dst                    t  c  cs   id
   +-------- L2 ---------------+----------- L3 IPv6 --------------------------------------------------------------------+--------L4 ICMP -----+

For the TCP matching on port 22 it will look as follows:

 000000000000 000000000000 0000 0 00 00000 0000 FF 00 00000000000000000000000000000000 00000000000000000000000000000000 0000 FFFF 00000000 00000000 0000 0000 0000 0000
   eth dst      eth src    et   v TC  fll  len  nh hl             ipv6 src                   ipv dst                      sp  dp    seq      ack      fl  win   cs   urg
   +-------- L2 ---------------+----------- L3 IPv6 --------------------------------------------------------------------+--------L4 TCP ---------------------------------


Then using these masks one would create 4 tables, by using the API call:

 classify_add_del_table(is_add=1, skip_n_vectors=0, mask=<MMMM>, match_n_vectors=<NNNN>,nbuckets=32,memory_size=20000, next_table_index=-1, miss_next_index=-1)

Let's call these tables "IPv4PROTO", "IPv4PROTO_TCPDPORT", "IPv6PROTO", "IPv6PROTO_TCPDPORT".

One would mention "IPv4PROTO" table as "next_table_index" table for "IPv4PROTO_TCPDPORT", and "IPv6PROTO" as "next_table_index" table for IPv6PROTO_TCPDPORT table.

Then one needs to populate the tables with the correct matches for "ICMP" and "tcp dst port 22". That can be done using API call:

 classify_add_del_session(is_add=1, table_index=<XXXX>, match=<bytes-to-match>, hit-next-index -1)

The bytes "XXXX" above would be the match of one or several vectors, corresponding to the packet contents with the desired value.

WARNING: if the "skip" is nonzero in the table configuration, the match is still the entire bitstring, without skipping any leading bytes !!!

Then one would apply the IPv4PROTO_TCPDPORT and IPv6PROTO_TCPDPORT as l2 input classify tables.

The CLI for that is set interface l2 output classify intfc <name> ip[46]-table <tableid>.

The API for this is

  classify_set_interface_l2_tables(sw_if_index=<INTFC>, ip4_table_index=<IPv4PROTO_TCPDPORT>, ip6_table_index=<IPv6PROTO_TCPDPORT>, other_table_index=-1, is_input=0)


This would allow to create a unidirectional policy, assuming the other policy is "permit all" it would be fine. If not - then a mirror table entries will need to be created using the same logic.

The full script showing this process in detail using the python API is at http://stdio.be/vpp/t/classifier_script_simple_policy.txt

The Java API is located in $ROOT/vpp-api/java..

References