summaryrefslogtreecommitdiff
path: root/Documentation/networking/devlink/devlink-dpipe.rst
blob: 468fe1001b7437e7d4279d929bd31235072af5cf (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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
.. SPDX-License-Identifier: GPL-2.0

=============
Devlink DPIPE
=============

Background
==========

While performing the hardware offloading process, much of the hardware
specifics cannot be presented. These details are useful for debugging, and
``devlink-dpipe`` provides a standardized way to provide visibility into the
offloading process.

For example, the routing longest prefix match (LPM) algorithm used by the
Linux kernel may differ from the hardware implementation. The pipeline debug
API (DPIPE) is aimed at providing the user visibility into the ASIC's
pipeline in a generic way.

The hardware offload process is expected to be done in a way that the user
should not be able to distinguish between the hardware vs. software
implementation. In this process, hardware specifics are neglected. In
reality those details can have lots of meaning and should be exposed in some
standard way.

This problem is made even more complex when one wishes to offload the
control path of the whole networking stack to a switch ASIC. Due to
differences in the hardware and software models some processes cannot be
represented correctly.

One example is the kernel's LPM algorithm which in many cases differs
greatly to the hardware implementation. The configuration API is the same,
but one cannot rely on the Forward Information Base (FIB) to look like the
Level Path Compression trie (LPC-trie) in hardware.

In many situations trying to analyze systems failure solely based on the
kernel's dump may not be enough. By combining this data with complementary
information about the underlying hardware, this debugging can be made
easier; additionally, the information can be useful when debugging
performance issues.

Overview
========

The ``devlink-dpipe`` interface closes this gap. The hardware's pipeline is
modeled as a graph of match/action tables. Each table represents a specific
hardware block. This model is not new, first being used by the P4 language.

Traditionally it has been used as an alternative model for hardware
configuration, but the ``devlink-dpipe`` interface uses it for visibility
purposes as a standard complementary tool. The system's view from
``devlink-dpipe`` should change according to the changes done by the
standard configuration tools.

For example, it’s quiet common to  implement Access Control Lists (ACL)
using Ternary Content Addressable Memory (TCAM). The TCAM memory can be
divided into TCAM regions. Complex TC filters can have multiple rules with
different priorities and different lookup keys. On the other hand hardware
TCAM regions have a predefined lookup key. Offloading the TC filter rules
using TCAM engine can result in multiple TCAM regions being interconnected
in a chain (which may affect the data path latency). In response to a new TC
filter new tables should be created describing those regions.

Model
=====

The ``DPIPE`` model introduces several objects:

  * headers
  * tables
  * entries

A ``header`` describes packet formats and provides names for fields within
the packet. A ``table`` describes hardware blocks. An ``entry`` describes
the actual content of a specific table.

The hardware pipeline is not port specific, but rather describes the whole
ASIC. Thus it is tied to the top of the ``devlink`` infrastructure.

Drivers can register and unregister tables at run time, in order to support
dynamic behavior. This dynamic behavior is mandatory for describing hardware
blocks like TCAM regions which can be allocated and freed dynamically.

``devlink-dpipe`` generally is not intended for configuration. The exception
is hardware counting for a specific table.

The following commands are used to obtain the ``dpipe`` objects from
userspace:

  * ``table_get``: Receive a table's description.
  * ``headers_get``: Receive a device's supported headers.
  * ``entries_get``: Receive a table's current entries.
  * ``counters_set``: Enable or disable counters on a table.

Table
-----

The driver should implement the following operations for each table:

  * ``matches_dump``: Dump the supported matches.
  * ``actions_dump``: Dump the supported actions.
  * ``entries_dump``: Dump the actual content of the table.
  * ``counters_set_update``: Synchronize hardware with counters enabled or
    disabled.

Header/Field
------------

In a similar way to P4 headers and fields are used to describe a table's
behavior. There is a slight difference between the standard protocol headers
and specific ASIC metadata. The protocol headers should be declared in the
``devlink`` core API. On the other hand ASIC meta data is driver specific
and should be defined in the driver. Additionally, each driver-specific
devlink documentation file should document the driver-specific ``dpipe``
headers it implements. The headers and fields are identified by enumeration.

In order to provide further visibility some ASIC metadata fields could be
mapped to kernel objects. For example, internal router interface indexes can
be directly mapped to the net device ifindex. FIB table indexes used by
different Virtual Routing and Forwarding (VRF) tables can be mapped to
internal routing table indexes.

Match
-----

Matches are kept primitive and close to hardware operation. Match types like
LPM are not supported due to the fact that this is exactly a process we wish
to describe in full detail. Example of matches:

  * ``field_exact``: Exact match on a specific field.
  * ``field_exact_mask``: Exact match on a specific field after masking.
  * ``field_range``: Match on a specific range.

The id's of the header and the field should be specified in order to
identify the specific field. Furthermore, the header index should be
specified in order to distinguish multiple headers of the same type in a
packet (tunneling).

Action
------

Similar to match, the actions are kept primitive and close to hardware
operation. For example:

  * ``field_modify``: Modify the field value.
  * ``field_inc``: Increment the field value.
  * ``push_header``: Add a header.
  * ``pop_header``: Remove a header.

Entry
-----

Entries of a specific table can be dumped on demand. Each eentry is
identified with an index and its properties are described by a list of
match/action values and specific counter. By dumping the tables content the
interactions between tables can be resolved.

Abstraction Example
===================

The following is an example of the abstraction model of the L3 part of
Mellanox Spectrum ASIC. The blocks are described in the order they appear in
the pipeline. The table sizes in the following examples are not real
hardware sizes and are provided for demonstration purposes.

LPM
---

The LPM algorithm can be implemented as a list of hash tables. Each hash
table contains routes with the same prefix length. The root of the list is
/32, and in case of a miss the hardware will continue to the next hash
table. The depth of the search will affect the data path latency.

In case of a hit the entry contains information about the next stage of the
pipeline which resolves the MAC address. The next stage can be either local
host table for directly connected routes, or adjacency table for next-hops.
The ``meta.lpm_prefix`` field is used to connect two LPM tables.

.. code::

    table lpm_prefix_16 {
      size: 4096,
      counters_enabled: true,
      match: { meta.vr_id: exact,
               ipv4.dst_addr: exact_mask,
               ipv6.dst_addr: exact_mask,
               meta.lpm_prefix: exact },
      action: { meta.adj_index: set,
                meta.adj_group_size: set,
                meta.rif_port: set,
                meta.lpm_prefix: set },
    }

Local Host
----------

In the case of local routes the LPM lookup already resolves the egress
router interface (RIF), yet the exact MAC address is not known. The local
host table is a hash table combining the output interface id with
destination IP address as a key. The result is the MAC address.

.. code::

    table local_host {
      size: 4096,
      counters_enabled: true,
      match: { meta.rif_port: exact,
               ipv4.dst_addr: exact},
      action: { ethernet.daddr: set }
    }

Adjacency
---------

In case of remote routes this table does the ECMP. The LPM lookup results in
ECMP group size and index that serves as a global offset into this table.
Concurrently a hash of the packet is generated. Based on the ECMP group size
and the packet's hash a local offset is generated. Multiple LPM entries can
point to the same adjacency group.

.. code::

    table adjacency {
      size: 4096,
      counters_enabled: true,
      match: { meta.adj_index: exact,
               meta.adj_group_size: exact,
               meta.packet_hash_index: exact },
      action: { ethernet.daddr: set,
                meta.erif: set }
    }

ERIF
----

In case the egress RIF and destination MAC have been resolved by previous
tables this table does multiple operations like TTL decrease and MTU check.
Then the decision of forward/drop is taken and the port L3 statistics are
updated based on the packet's type (broadcast, unicast, multicast).

.. code::

    table erif {
      size: 800,
      counters_enabled: true,
      match: { meta.rif_port: exact,
               meta.is_l3_unicast: exact,
               meta.is_l3_broadcast: exact,
               meta.is_l3_multicast, exact },
      action: { meta.l3_drop: set,
                meta.l3_forward: set }
    }