summaryrefslogtreecommitdiffstats
path: root/src/plugins/map/map_doc.rst
blob: 663e815d54594ac380bea9d5c8400292d133ae9e (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
MAP and Lw4o6
=============

This is a memo intended to contain documentation of the VPP MAP and
Lw4o6 implementations. Everything that is not directly obvious should
come here.

MAP-E Virtual Reassembly
------------------------

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: \* An IPv4 fragment is received and the destination IPv4 address
is shared. \* An IPv6 packet is received with an inner IPv4 fragment,
the IPv4 source address is shared, and ‘security-check fragments’ is on.
\* An IPv6 fragment is received.

There are 3 dedicated nodes: \* ip4-map-reass \* ip6-map-ip4-reass \*
ip6-map-ip6-reass

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.

Virtual Reassembly configuration
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

IPv4 and IPv6 virtual reassembly support the following configuration:
map params reassembly [ip4 \| ip6] [lifetime ] [pool-size ] [buffers ]
[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
seconds. 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.

Additional considerations
^^^^^^^^^^^^^^^^^^^^^^^^^

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 apart.

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.