xref: /dpdk/doc/guides/prog_guide/graph_lib.rst (revision f399b0171e6e64c8bbce42599afa35591a9d28f1)
1..  SPDX-License-Identifier: BSD-3-Clause
2    Copyright(C) 2020 Marvell International Ltd.
3
4Graph Library and Inbuilt Nodes
5===============================
6
7Graph architecture abstracts the data processing functions as a ``node`` and
8``links`` them together to create a complex ``graph`` to enable reusable/modular
9data processing functions.
10
11The graph library provides API to enable graph framework operations such as
12create, lookup, dump and destroy on graph and node operations such as clone,
13edge update, and edge shrink, etc. The API also allows to create the stats
14cluster to monitor per graph and per node stats.
15
16Features
17--------
18
19Features of the Graph library are:
20
21- Nodes as plugins.
22- Support for out of tree nodes.
23- Inbuilt nodes for packet processing.
24- Multi-process support.
25- Low overhead graph walk and node enqueue.
26- Low overhead statistics collection infrastructure.
27- Support to export the graph as a Graphviz dot file. See ``rte_graph_export()``.
28- Allow having another graph walk implementation in the future by segregating
29  the fast path(``rte_graph_worker.h``) and slow path code.
30
31Advantages of Graph architecture
32--------------------------------
33
34- Memory latency is the enemy for high-speed packet processing, moving the
35  similar packet processing code to a node will reduce the I cache and D
36  caches misses.
37- Exploits the probability that most packets will follow the same nodes in the
38  graph.
39- Allow SIMD instructions for packet processing of the node.-
40- The modular scheme allows having reusable nodes for the consumers.
41- The modular scheme allows us to abstract the vendor HW specific
42  optimizations as a node.
43
44Performance tuning parameters
45-----------------------------
46
47- Test with various burst size values (256, 128, 64, 32) using
48  CONFIG_RTE_GRAPH_BURST_SIZE config option.
49  The testing shows, on x86 and arm64 servers, The sweet spot is 256 burst
50  size. While on arm64 embedded SoCs, it is either 64 or 128.
51- Disable node statistics (using ``CONFIG_RTE_LIBRTE_GRAPH_STATS`` config option)
52  if not needed.
53- Use arm64 optimized memory copy for arm64 architecture by
54  selecting ``CONFIG_RTE_ARCH_ARM64_MEMCPY``.
55
56Programming model
57-----------------
58
59Anatomy of Node:
60~~~~~~~~~~~~~~~~
61
62.. _figure_anatomy_of_a_node:
63
64.. figure:: img/anatomy_of_a_node.*
65
66The :numref:`figure_anatomy_of_a_node` diagram depicts the anatomy of a node.
67
68The node is the basic building block of the graph framework.
69
70A node consists of:
71
72process():
73^^^^^^^^^^
74
75The callback function will be invoked by worker thread using
76``rte_graph_walk()`` function when there is data to be processed by the node.
77A graph node process the function using ``process()`` and enqueue to next
78downstream node using ``rte_node_enqueue*()`` function.
79
80Context memory:
81^^^^^^^^^^^^^^^
82
83It is memory allocated by the library to store the node-specific context
84information. This memory will be used by process(), init(), fini() callbacks.
85
86init():
87^^^^^^^
88
89The callback function will be invoked by ``rte_graph_create()`` on when
90a node gets attached to a graph.
91
92fini():
93^^^^^^^
94
95The callback function will be invoked by ``rte_graph_destroy()`` on when a
96node gets detached to a graph.
97
98Node name:
99^^^^^^^^^^
100
101It is the name of the node. When a node registers to graph library, the library
102gives the ID as ``rte_node_t`` type. Both ID or Name shall be used lookup the
103node. ``rte_node_from_name()``, ``rte_node_id_to_name()`` are the node
104lookup functions.
105
106nb_edges:
107^^^^^^^^^
108
109The number of downstream nodes connected to this node. The ``next_nodes[]``
110stores the downstream nodes objects. ``rte_node_edge_update()`` and
111``rte_node_edge_shrink()`` functions shall be used to update the ``next_node[]``
112objects. Consumers of the node APIs are free to update the ``next_node[]``
113objects till ``rte_graph_create()`` invoked.
114
115next_node[]:
116^^^^^^^^^^^^
117
118The dynamic array to store the downstream nodes connected to this node. Downstream
119node should not be current node itself or a source node.
120
121Source node:
122^^^^^^^^^^^^
123
124Source nodes are static nodes created using ``RTE_NODE_REGISTER`` by passing
125``flags`` as ``RTE_NODE_SOURCE_F``.
126While performing the graph walk, the ``process()`` function of all the source
127nodes will be called first. So that these nodes can be used as input nodes for a graph.
128
129Node creation and registration
130~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
131* Node implementer creates the node by implementing ops and attributes of
132  ``struct rte_node_register``.
133
134* The library registers the node by invoking RTE_NODE_REGISTER on library load
135  using the constructor scheme. The constructor scheme used here to support multi-process.
136
137Link the Nodes to create the graph topology
138~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
139.. _figure_link_the_nodes:
140
141.. figure:: img/link_the_nodes.*
142
143The :numref:`figure_link_the_nodes` diagram shows a graph topology after
144linking the N nodes.
145
146Once nodes are available to the program, Application or node public API
147functions can links them together to create a complex packet processing graph.
148
149There are multiple different types of strategies to link the nodes.
150
151Method (a):
152^^^^^^^^^^^
153Provide the ``next_nodes[]`` at the node registration time. See  ``struct rte_node_register::nb_edges``.
154This is a use case to address the static node scheme where one knows upfront the
155``next_nodes[]`` of the node.
156
157Method (b):
158^^^^^^^^^^^
159Use ``rte_node_edge_get()``, ``rte_node_edge_update()``, ``rte_node_edge_shrink()``
160to update the ``next_nodes[]`` links for the node runtime but before graph create.
161
162Method (c):
163^^^^^^^^^^^
164Use ``rte_node_clone()`` to clone a already existing node, created using RTE_NODE_REGISTER.
165When ``rte_node_clone()`` invoked, The library, would clone all the attributes
166of the node and creates a new one. The name for cloned node shall be
167``"parent_node_name-user_provided_name"``.
168
169This method enables the use case of Rx and Tx nodes where multiple of those nodes
170need to be cloned based on the number of CPU available in the system.
171The cloned nodes will be identical, except the ``"context memory"``.
172Context memory will have information of port, queue pair in case of Rx and Tx
173ethdev nodes.
174
175Create the graph object
176~~~~~~~~~~~~~~~~~~~~~~~
177Now that the nodes are linked, Its time to create a graph by including
178the required nodes. The application can provide a set of node patterns to
179form a graph object. The ``famish()`` API used underneath for the pattern
180matching to include the required nodes. After the graph create any changes to
181nodes or graph is not allowed.
182
183The ``rte_graph_create()`` API shall be used to create the graph.
184
185Example of a graph object creation:
186
187.. code-block:: console
188
189   {"ethdev_rx-0-0", ip4*, ethdev_tx-*"}
190
191In the above example, A graph object will be created with ethdev Rx
192node of port 0 and queue 0, all ipv4* nodes in the system,
193and ethdev tx node of all ports.
194
195Multicore graph processing
196~~~~~~~~~~~~~~~~~~~~~~~~~~
197In the current graph library implementation, specifically,
198``rte_graph_walk()`` and ``rte_node_enqueue*`` fast path API functions
199are designed to work on single-core to have better performance.
200The fast path API works on graph object, So the multi-core graph
201processing strategy would be to create graph object PER WORKER.
202
203In fast path
204~~~~~~~~~~~~
205Typical fast-path code looks like below, where the application
206gets the fast-path graph object using ``rte_graph_lookup()``
207on the worker thread and run the ``rte_graph_walk()`` in a tight loop.
208
209.. code-block:: c
210
211    struct rte_graph *graph = rte_graph_lookup("worker0");
212
213    while (!done) {
214        rte_graph_walk(graph);
215    }
216
217Context update when graph walk in action
218~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
219The fast-path object for the node is ``struct rte_node``.
220
221It may be possible that in slow-path or after the graph walk-in action,
222the user needs to update the context of the node hence access to
223``struct rte_node *`` memory.
224
225``rte_graph_foreach_node()``, ``rte_graph_node_get()``,
226``rte_graph_node_get_by_name()`` APIs can be used to to get the
227``struct rte_node*``. ``rte_graph_foreach_node()`` iterator function works on
228``struct rte_graph *`` fast-path graph object while others works on graph ID or name.
229
230Get the node statistics using graph cluster
231~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
232The user may need to know the aggregate stats of the node across
233multiple graph objects. Especially the situation where each graph object bound
234to a worker thread.
235
236Introduced a graph cluster object for statistics.
237``rte_graph_cluster_stats_create()`` API shall be used for creating a
238graph cluster with multiple graph objects and ``rte_graph_cluster_stats_get()``
239to get the aggregate node statistics.
240
241An example statistics output from ``rte_graph_cluster_stats_get()``
242
243.. code-block:: diff
244
245    +---------+-----------+-------------+---------------+-----------+---------------+-----------+
246    |Node     |calls      |objs         |realloc_count  |objs/call  |objs/sec(10E6) |cycles/call|
247    +---------------------+-------------+---------------+-----------+---------------+-----------+
248    |node0    |12977424   |3322220544   |5              |256.000    |3047.151872    |20.0000    |
249    |node1    |12977653   |3322279168   |0              |256.000    |3047.210496    |17.0000    |
250    |node2    |12977696   |3322290176   |0              |256.000    |3047.221504    |17.0000    |
251    |node3    |12977734   |3322299904   |0              |256.000    |3047.231232    |17.0000    |
252    |node4    |12977784   |3322312704   |1              |256.000    |3047.243776    |17.0000    |
253    |node5    |12977825   |3322323200   |0              |256.000    |3047.254528    |17.0000    |
254    +---------+-----------+-------------+---------------+-----------+---------------+-----------+
255
256Node writing guidelines
257~~~~~~~~~~~~~~~~~~~~~~~
258
259The ``process()`` function of a node is the fast-path function and that needs
260to be written carefully to achieve max performance.
261
262Broadly speaking, there are two different types of nodes.
263
264Static nodes
265~~~~~~~~~~~~
266The first kind of nodes are those that have a fixed ``next_nodes[]`` for the
267complete burst (like ethdev_rx, ethdev_tx) and it is simple to write.
268``process()`` function can move the obj burst to the next node either using
269``rte_node_next_stream_move()`` or using ``rte_node_next_stream_get()`` and
270``rte_node_next_stream_put()``.
271
272Intermediate nodes
273~~~~~~~~~~~~~~~~~~
274The second kind of such node is ``intermediate nodes`` that decide what is the
275``next_node[]`` to send to on a per-packet basis. In these nodes,
276
277* Firstly, there has to be the best possible packet processing logic.
278
279* Secondly, each packet needs to be queued to its next node.
280
281This can be done using ``rte_node_enqueue_[x1|x2|x4]()`` APIs if
282they are to single next or ``rte_node_enqueue_next()`` that takes array of nexts.
283
284In scenario where multiple intermediate nodes are present but most of the time
285each node using the same next node for all its packets, the cost of moving every
286pointer from current node's stream to next node's stream could be avoided.
287This is called home run and ``rte_node_next_stream_move()`` could be used to
288just move stream from the current node to the next node with least number of cycles.
289Since this can be avoided only in the case where all the packets are destined
290to the same next node, node implementation should be also having worst-case
291handling where every packet could be going to different next node.
292
293Example of intermediate node implementation with home run:
294^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2951. Start with speculation that next_node = node->ctx.
296This could be the next_node application used in the previous function call of this node.
297
2982. Get the next_node stream array with required space using
299``rte_node_next_stream_get(next_node, space)``.
300
3013. while n_left_from > 0 (i.e packets left to be sent) prefetch next pkt_set
302and process current pkt_set to find their next node
303
3044. if all the next nodes of the current pkt_set match speculated next node,
305just count them as successfully speculated(``last_spec``) till now and
306continue the loop without actually moving them to the next node. else if there is
307a mismatch, copy all the pkt_set pointers that were ``last_spec`` and move the
308current pkt_set to their respective next's nodes using ``rte_enqueue_next_x1()``.
309Also, one of the next_node can be updated as speculated next_node if it is more
310probable. Finally, reset ``last_spec`` to zero.
311
3125. if n_left_from != 0 then goto 3) to process remaining packets.
313
3146. if last_spec == nb_objs, All the objects passed were successfully speculated
315to single next node. So, the current stream can be moved to next node using
316``rte_node_next_stream_move(node, next_node)``.
317This is the ``home run`` where memcpy of buffer pointers to next node is avoided.
318
3197. Update the ``node->ctx`` with more probable next node.
320
321Graph object memory layout
322--------------------------
323.. _figure_graph_mem_layout:
324
325.. figure:: img/graph_mem_layout.*
326
327The :numref:`figure_graph_mem_layout` diagram shows ``rte_graph`` object memory
328layout. Understanding the memory layout helps to debug the graph library and
329improve the performance if needed.
330
331Graph object consists of a header, circular buffer to store the pending
332stream when walking over the graph, and variable-length memory to store
333the ``rte_node`` objects.
334
335The graph_nodes_mem_create() creates and populate this memory. The functions
336such as ``rte_graph_walk()`` and ``rte_node_enqueue_*`` use this memory
337to enable fastpath services.
338
339Inbuilt Nodes
340-------------
341
342DPDK provides a set of nodes for data processing. The following section
343details the documentation for the same.
344
345ethdev_rx
346~~~~~~~~~
347This node does ``rte_eth_rx_burst()`` into stream buffer passed to it
348(src node stream) and does ``rte_node_next_stream_move()`` only when
349there are packets received. Each ``rte_node`` works only on one Rx port and
350queue that it gets from node->ctx. For each (port X, rx_queue Y),
351a rte_node is cloned from  ethdev_rx_base_node as ``ethdev_rx-X-Y`` in
352``rte_node_eth_config()`` along with updating ``node->ctx``.
353Each graph needs to be associated  with a unique rte_node for a (port, rx_queue).
354
355ethdev_tx
356~~~~~~~~~
357This node does ``rte_eth_tx_burst()`` for a burst of objs received by it.
358It sends the burst to a fixed Tx Port and Queue information from
359node->ctx. For each (port X), this ``rte_node`` is cloned from
360ethdev_tx_node_base as "ethdev_tx-X" in ``rte_node_eth_config()``
361along with updating node->context.
362
363Since each graph doesn't need more than one Txq, per port, a Txq is assigned
364based on graph id to each rte_node instance. Each graph needs to be associated
365with a rte_node for each (port).
366
367pkt_drop
368~~~~~~~~
369This node frees all the objects passed to it considering them as
370``rte_mbufs`` that need to be freed.
371
372ip4_lookup
373~~~~~~~~~~
374This node is an intermediate node that does LPM lookup for the received
375ipv4 packets and the result determines each packets next node.
376
377On successful LPM lookup, the result contains the ``next_node`` id and
378``next-hop`` id with which the packet needs to be further processed.
379
380On LPM lookup failure, objects are redirected to pkt_drop node.
381``rte_node_ip4_route_add()`` is control path API to add ipv4 routes.
382To achieve home run, node use ``rte_node_stream_move()`` as mentioned in above
383sections.
384
385ip4_rewrite
386~~~~~~~~~~~
387This node gets packets from ``ip4_lookup`` node with next-hop id for each
388packet is embedded in ``node_mbuf_priv1(mbuf)->nh``. This id is used
389to determine the L2 header to be written to the packet before sending
390the packet out to a particular ethdev_tx node.
391``rte_node_ip4_rewrite_add()`` is control path API to add next-hop info.
392
393null
394~~~~
395This node ignores the set of objects passed to it and reports that all are
396processed.
397
398