FD.io VPP
v18.01.2-1-g9b554f3
Vector Packet Processing
|
This is a memo intended to contain documentation of the VPP MAP and Lw4o6 implementations. Everything that is not directly obvious should come here.
The MAP-E implementation supports handling of IPv4 fragments as well as IPv4-in-IPv6 inner and outer fragments. This is called virtual reassembly because the fragments are not actually reassembled. Instead, some meta-data are kept about the first fragment and reused for subsequent fragments.
Fragment caching and handling is not always necessary. It is performed when:
There are 3 dedicated nodes:
ip4-map sends all fragments to ip4-map-reass. ip6-map sends all inner-fragments to ip6-map-ip4-reass. ip6-map sends all outer-fragments to ip6-map-ip6-reass.
IPv4 (resp. IPv6) virtual reassembly makes use of a hash table in order to store IPv4 (resp. IPv6) reassembly structures. The hash-key is based on the IPv4-src:IPv4-dst:Frag-ID:Protocol tuple (resp. IPv6-src:IPv6-dst:Frag-ID tuple, as the protocol is IPv4-in-IPv6). Therefore, each packet reassembly makes use of exactly one reassembly structure. When such a structure is allocated, it is timestamped with the current time. Finally, those structures are capable of storing a limited number of buffer indexes.
An IPv4 (resp. IPv6) reassembly structure can cache up to MAP_IP4_REASS_MAX_FRAGMENTS_PER_REASSEMBLY (resp. MAP_IP6_REASS_MAX_FRAGMENTS_PER_REASSEMBLY) buffers. Buffers are cached until the first fragment is received.
IPv4 and IPv6 virtual reassembly support the following configuration: map params reassembly [ip4 | ip6] [lifetime <lifetime-ms>] [pool-size <pool-size>] [buffers <buffers>] [ht-ratio <ht-ratio>]
lifetime: The time in milliseconds a reassembly structure is considered valid. The longer, the more reliable is reassembly, but the more likely it is to exhaust the pool of reassembly structures. IPv4 standard suggests a lifetime of 15 seconds. IPv6 specifies a lifetime of 60 people. Those values are not realistic for high-throughput cases.
buffers: The upper limit of buffers that are allowed to be cached. It can be used to protect against fragmentation attacks which would aim to exhaust the global buffers pool.
pool-size: The number of reassembly structures that can be allocated. As each structure can store a small fixed number of fragments, it also sets an upper-bound of 'pool-size * MAP_IPX_REASS_MAX_FRAGMENTS_PER_REASSEMBLY' buffers that can be cached in total.
ht-ratio: The amount of buckets in the hash-table is pool-size * ht-ratio.
Any time pool-size and ht-ratio is modified, the hash-table is destroyed and created again, which means all current state is lost.
Reassembly at high rate is expensive in terms of buffers. There is a trade-off between the lifetime and number of allocated buffers. Reducing the lifetime helps, but at the cost of loosing state for fragments that are wide appart.
Let: R be the packet rate at which fragments are received. F be the number of fragments per packet.
Assuming the first fragment is always received last. We should have: buffers > lifetime * R / F * (F - 1) pool-size > lifetime * R/F
This is a worst case. Receiving the first fragment earlier helps reducing the number of required buffers. Also, an optimization is implemented (MAP_IP6_REASS_COUNT_BYTES and MAP_IP4_REASS_COUNT_BYTES) which counts the number of transmitted bytes and remembers the total number of bytes which should be transmitted based on the last fragment, and therefore helps reducing 'pool-size'.
But the formula shows that it is challenging to forward a significant amount of fragmented packets at high rates. For instance, with a lifetime of 1 second, 5Mpps packet rate would require buffering up to 2.5 millions fragments.
If you want to do that, be prepared to configure a lot of fragments.