| 1 | |
| 2 |
|
| 3 | =============
|
| 4 | Devlink DPIPE
|
| 5 | =============
|
| 6 |
|
| 7 | Background
|
| 8 | ==========
|
| 9 |
|
| 10 | While performing the hardware offloading process, much of the hardware
|
| 11 | specifics cannot be presented. These details are useful for debugging, and
|
| 12 | ``devlink-dpipe`` provides a standardized way to provide visibility into the
|
| 13 | offloading process.
|
| 14 |
|
| 15 | For example, the routing longest prefix match (LPM) algorithm used by the
|
| 16 | Linux kernel may differ from the hardware implementation. The pipeline debug
|
| 17 | API (DPIPE) is aimed at providing the user visibility into the ASIC's
|
| 18 | pipeline in a generic way.
|
| 19 |
|
| 20 | The hardware offload process is expected to be done in a way that the user
|
| 21 | should not be able to distinguish between the hardware vs. software
|
| 22 | implementation. In this process, hardware specifics are neglected. In
|
| 23 | reality those details can have lots of meaning and should be exposed in some
|
| 24 | standard way.
|
| 25 |
|
| 26 | This problem is made even more complex when one wishes to offload the
|
| 27 | control path of the whole networking stack to a switch ASIC. Due to
|
| 28 | differences in the hardware and software models some processes cannot be
|
| 29 | represented correctly.
|
| 30 |
|
| 31 | One example is the kernel's LPM algorithm which in many cases differs
|
| 32 | greatly to the hardware implementation. The configuration API is the same,
|
| 33 | but one cannot rely on the Forward Information Base (FIB) to look like the
|
| 34 | Level Path Compression trie (LPC-trie) in hardware.
|
| 35 |
|
| 36 | In many situations trying to analyze systems failure solely based on the
|
| 37 | kernel's dump may not be enough. By combining this data with complementary
|
| 38 | information about the underlying hardware, this debugging can be made
|
| 39 | easier; additionally, the information can be useful when debugging
|
| 40 | performance issues.
|
| 41 |
|
| 42 | Overview
|
| 43 | ========
|
| 44 |
|
| 45 | The ``devlink-dpipe`` interface closes this gap. The hardware's pipeline is
|
| 46 | modeled as a graph of match/action tables. Each table represents a specific
|
| 47 | hardware block. This model is not new, first being used by the P4 language.
|
| 48 |
|
| 49 | Traditionally it has been used as an alternative model for hardware
|
| 50 | configuration, but the ``devlink-dpipe`` interface uses it for visibility
|
| 51 | purposes as a standard complementary tool. The system's view from
|
| 52 | ``devlink-dpipe`` should change according to the changes done by the
|
| 53 | standard configuration tools.
|
| 54 |
|
| 55 | For example, it’s quite common to implement Access Control Lists (ACL)
|
| 56 | using Ternary Content Addressable Memory (TCAM). The TCAM memory can be
|
| 57 | divided into TCAM regions. Complex TC filters can have multiple rules with
|
| 58 | different priorities and different lookup keys. On the other hand hardware
|
| 59 | TCAM regions have a predefined lookup key. Offloading the TC filter rules
|
| 60 | using TCAM engine can result in multiple TCAM regions being interconnected
|
| 61 | in a chain (which may affect the data path latency). In response to a new TC
|
| 62 | filter new tables should be created describing those regions.
|
| 63 |
|
| 64 | Model
|
| 65 | =====
|
| 66 |
|
| 67 | The ``DPIPE`` model introduces several objects:
|
| 68 |
|
| 69 | * headers
|
| 70 | * tables
|
| 71 | * entries
|
| 72 |
|
| 73 | A ``header`` describes packet formats and provides names for fields within
|
| 74 | the packet. A ``table`` describes hardware blocks. An ``entry`` describes
|
| 75 | the actual content of a specific table.
|
| 76 |
|
| 77 | The hardware pipeline is not port specific, but rather describes the whole
|
| 78 | ASIC. Thus it is tied to the top of the ``devlink`` infrastructure.
|
| 79 |
|
| 80 | Drivers can register and unregister tables at run time, in order to support
|
| 81 | dynamic behavior. This dynamic behavior is mandatory for describing hardware
|
| 82 | blocks like TCAM regions which can be allocated and freed dynamically.
|
| 83 |
|
| 84 | ``devlink-dpipe`` generally is not intended for configuration. The exception
|
| 85 | is hardware counting for a specific table.
|
| 86 |
|
| 87 | The following commands are used to obtain the ``dpipe`` objects from
|
| 88 | userspace:
|
| 89 |
|
| 90 | * ``table_get``: Receive a table's description.
|
| 91 | * ``headers_get``: Receive a device's supported headers.
|
| 92 | * ``entries_get``: Receive a table's current entries.
|
| 93 | * ``counters_set``: Enable or disable counters on a table.
|
| 94 |
|
| 95 | Table
|
| 96 | -----
|
| 97 |
|
| 98 | The driver should implement the following operations for each table:
|
| 99 |
|
| 100 | * ``matches_dump``: Dump the supported matches.
|
| 101 | * ``actions_dump``: Dump the supported actions.
|
| 102 | * ``entries_dump``: Dump the actual content of the table.
|
| 103 | * ``counters_set_update``: Synchronize hardware with counters enabled or
|
| 104 | disabled.
|
| 105 |
|
| 106 | Header/Field
|
| 107 | ------------
|
| 108 |
|
| 109 | In a similar way to P4 headers and fields are used to describe a table's
|
| 110 | behavior. There is a slight difference between the standard protocol headers
|
| 111 | and specific ASIC metadata. The protocol headers should be declared in the
|
| 112 | ``devlink`` core API. On the other hand ASIC meta data is driver specific
|
| 113 | and should be defined in the driver. Additionally, each driver-specific
|
| 114 | devlink documentation file should document the driver-specific ``dpipe``
|
| 115 | headers it implements. The headers and fields are identified by enumeration.
|
| 116 |
|
| 117 | In order to provide further visibility some ASIC metadata fields could be
|
| 118 | mapped to kernel objects. For example, internal router interface indexes can
|
| 119 | be directly mapped to the net device ifindex. FIB table indexes used by
|
| 120 | different Virtual Routing and Forwarding (VRF) tables can be mapped to
|
| 121 | internal routing table indexes.
|
| 122 |
|
| 123 | Match
|
| 124 | -----
|
| 125 |
|
| 126 | Matches are kept primitive and close to hardware operation. Match types like
|
| 127 | LPM are not supported due to the fact that this is exactly a process we wish
|
| 128 | to describe in full detail. Example of matches:
|
| 129 |
|
| 130 | * ``field_exact``: Exact match on a specific field.
|
| 131 | * ``field_exact_mask``: Exact match on a specific field after masking.
|
| 132 | * ``field_range``: Match on a specific range.
|
| 133 |
|
| 134 | The id's of the header and the field should be specified in order to
|
| 135 | identify the specific field. Furthermore, the header index should be
|
| 136 | specified in order to distinguish multiple headers of the same type in a
|
| 137 | packet (tunneling).
|
| 138 |
|
| 139 | Action
|
| 140 | ------
|
| 141 |
|
| 142 | Similar to match, the actions are kept primitive and close to hardware
|
| 143 | operation. For example:
|
| 144 |
|
| 145 | * ``field_modify``: Modify the field value.
|
| 146 | * ``field_inc``: Increment the field value.
|
| 147 | * ``push_header``: Add a header.
|
| 148 | * ``pop_header``: Remove a header.
|
| 149 |
|
| 150 | Entry
|
| 151 | -----
|
| 152 |
|
| 153 | Entries of a specific table can be dumped on demand. Each eentry is
|
| 154 | identified with an index and its properties are described by a list of
|
| 155 | match/action values and specific counter. By dumping the tables content the
|
| 156 | interactions between tables can be resolved.
|
| 157 |
|
| 158 | Abstraction Example
|
| 159 | ===================
|
| 160 |
|
| 161 | The following is an example of the abstraction model of the L3 part of
|
| 162 | Mellanox Spectrum ASIC. The blocks are described in the order they appear in
|
| 163 | the pipeline. The table sizes in the following examples are not real
|
| 164 | hardware sizes and are provided for demonstration purposes.
|
| 165 |
|
| 166 | LPM
|
| 167 | ---
|
| 168 |
|
| 169 | The LPM algorithm can be implemented as a list of hash tables. Each hash
|
| 170 | table contains routes with the same prefix length. The root of the list is
|
| 171 | /32, and in case of a miss the hardware will continue to the next hash
|
| 172 | table. The depth of the search will affect the data path latency.
|
| 173 |
|
| 174 | In case of a hit the entry contains information about the next stage of the
|
| 175 | pipeline which resolves the MAC address. The next stage can be either local
|
| 176 | host table for directly connected routes, or adjacency table for next-hops.
|
| 177 | The ``meta.lpm_prefix`` field is used to connect two LPM tables.
|
| 178 |
|
| 179 | .. code::
|
| 180 |
|
| 181 | table lpm_prefix_16 {
|
| 182 | size: 4096,
|
| 183 | counters_enabled: true,
|
| 184 | match: { meta.vr_id: exact,
|
| 185 | ipv4.dst_addr: exact_mask,
|
| 186 | ipv6.dst_addr: exact_mask,
|
| 187 | meta.lpm_prefix: exact },
|
| 188 | action: { meta.adj_index: set,
|
| 189 | meta.adj_group_size: set,
|
| 190 | meta.rif_port: set,
|
| 191 | meta.lpm_prefix: set },
|
| 192 | }
|
| 193 |
|
| 194 | Local Host
|
| 195 | ----------
|
| 196 |
|
| 197 | In the case of local routes the LPM lookup already resolves the egress
|
| 198 | router interface (RIF), yet the exact MAC address is not known. The local
|
| 199 | host table is a hash table combining the output interface id with
|
| 200 | destination IP address as a key. The result is the MAC address.
|
| 201 |
|
| 202 | .. code::
|
| 203 |
|
| 204 | table local_host {
|
| 205 | size: 4096,
|
| 206 | counters_enabled: true,
|
| 207 | match: { meta.rif_port: exact,
|
| 208 | ipv4.dst_addr: exact},
|
| 209 | action: { ethernet.daddr: set }
|
| 210 | }
|
| 211 |
|
| 212 | Adjacency
|
| 213 | ---------
|
| 214 |
|
| 215 | In case of remote routes this table does the ECMP. The LPM lookup results in
|
| 216 | ECMP group size and index that serves as a global offset into this table.
|
| 217 | Concurrently a hash of the packet is generated. Based on the ECMP group size
|
| 218 | and the packet's hash a local offset is generated. Multiple LPM entries can
|
| 219 | point to the same adjacency group.
|
| 220 |
|
| 221 | .. code::
|
| 222 |
|
| 223 | table adjacency {
|
| 224 | size: 4096,
|
| 225 | counters_enabled: true,
|
| 226 | match: { meta.adj_index: exact,
|
| 227 | meta.adj_group_size: exact,
|
| 228 | meta.packet_hash_index: exact },
|
| 229 | action: { ethernet.daddr: set,
|
| 230 | meta.erif: set }
|
| 231 | }
|
| 232 |
|
| 233 | ERIF
|
| 234 | ----
|
| 235 |
|
| 236 | In case the egress RIF and destination MAC have been resolved by previous
|
| 237 | tables this table does multiple operations like TTL decrease and MTU check.
|
| 238 | Then the decision of forward/drop is taken and the port L3 statistics are
|
| 239 | updated based on the packet's type (broadcast, unicast, multicast).
|
| 240 |
|
| 241 | .. code::
|
| 242 |
|
| 243 | table erif {
|
| 244 | size: 800,
|
| 245 | counters_enabled: true,
|
| 246 | match: { meta.rif_port: exact,
|
| 247 | meta.is_l3_unicast: exact,
|
| 248 | meta.is_l3_broadcast: exact,
|
| 249 | meta.is_l3_multicast, exact },
|
| 250 | action: { meta.l3_drop: set,
|
| 251 | meta.l3_forward: set }
|
| 252 | }
|