2 Copyright(c) 2017 Intel Corporation. All rights reserved.
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions
8 * Redistributions of source code must retain the above copyright
9 notice, this list of conditions and the following disclaimer.
10 * Redistributions in binary form must reproduce the above copyright
11 notice, this list of conditions and the following disclaimer in
12 the documentation and/or other materials provided with the
14 * Neither the name of Intel Corporation nor the names of its
15 contributors may be used to endorse or promote products derived
16 from this software without specific prior written permission.
18 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 The DPDK Event device library is an abstraction that provides the application
34 with features to schedule events. This is achieved using the PMD architecture
35 similar to the ethdev or cryptodev APIs, which may already be familiar to the
38 The eventdev framework introduces the event driven programming model. In a
39 polling model, lcores poll ethdev ports and associated Rx queues directly
40 to look for a packet. By contrast in an event driven model, lcores call the
41 scheduler that selects packets for them based on programmer-specified criteria.
42 The Eventdev library adds support for an event driven programming model, which
43 offers applications automatic multicore scaling, dynamic load balancing,
44 pipelining, packet ingress order maintenance and synchronization services to
45 simplify application packet processing.
47 By introducing an event driven programming model, DPDK can support both polling
48 and event driven programming models for packet processing, and applications are
49 free to choose whatever model (or combination of the two) best suits their
52 Step-by-step instructions of the eventdev design is available in the `API
53 Walk-through`_ section later in this document.
58 The eventdev API represents each event with a generic struct, which contains a
59 payload and metadata required for scheduling by an eventdev. The
60 ``rte_event`` struct is a 16 byte C structure, defined in
61 ``libs/librte_eventdev/rte_eventdev.h``.
66 The rte_event structure contains the following metadata fields, which the
67 application fills in to have the event scheduled as required:
69 * ``flow_id`` - The targeted flow identifier for the enq/deq operation.
70 * ``event_type`` - The source of this event, eg RTE_EVENT_TYPE_ETHDEV or CPU.
71 * ``sub_event_type`` - Distinguishes events inside the application, that have
72 the same event_type (see above)
73 * ``op`` - This field takes one of the RTE_EVENT_OP_* values, and tells the
74 eventdev about the status of the event - valid values are NEW, FORWARD or
76 * ``sched_type`` - Represents the type of scheduling that should be performed
77 on this event, valid values are the RTE_SCHED_TYPE_ORDERED, ATOMIC and
79 * ``queue_id`` - The identifier for the event queue that the event is sent to.
80 * ``priority`` - The priority of this event, see RTE_EVENT_DEV_PRIORITY.
85 The rte_event struct contains a union for payload, allowing flexibility in what
86 the actual event being scheduled is. The payload is a union of the following:
90 * ``struct rte_mbuf *mbuf``
92 These three items in a union occupy the same 64 bits at the end of the rte_event
93 structure. The application can utilize the 64 bits directly by accessing the
94 u64 variable, while the event_ptr and mbuf are provided as convenience
95 variables. For example the mbuf pointer in the union can used to schedule a
101 An event queue is a queue containing events that are scheduled by the event
102 device. An event queue contains events of different flows associated with
103 scheduling types, such as atomic, ordered, or parallel.
105 Queue All Types Capable
106 ^^^^^^^^^^^^^^^^^^^^^^^
108 If RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES capability bit is set in the event device,
109 then events of any type may be sent to any queue. Otherwise, the queues only
110 support events of the type that it was created with.
112 Queue All Types Incapable
113 ^^^^^^^^^^^^^^^^^^^^^^^^^
115 In this case, each stage has a specified scheduling type. The application
116 configures each queue for a specific type of scheduling, and just enqueues all
117 events to the eventdev. An example of a PMD of this type is the eventdev
120 The Eventdev API supports the following scheduling types per queue:
126 Atomic, Ordered and Parallel are load-balanced scheduling types: the output
127 of the queue can be spread out over multiple CPU cores.
129 Atomic scheduling on a queue ensures that a single flow is not present on two
130 different CPU cores at the same time. Ordered allows sending all flows to any
131 core, but the scheduler must ensure that on egress the packets are returned to
132 ingress order on downstream queue enqueue. Parallel allows sending all flows
133 to all CPU cores, without any re-ordering guarantees.
138 There is a SINGLE_LINK flag which allows an application to indicate that only
139 one port will be connected to a queue. Queues configured with the single-link
140 flag follow a FIFO like structure, maintaining ordering but it is only capable
141 of being linked to a single port (see below for port and queue linking details).
147 Ports are the points of contact between worker cores and the eventdev. The
148 general use-case will see one CPU core using one port to enqueue and dequeue
149 events from an eventdev. Ports are linked to queues in order to retrieve events
150 from those queues (more details in `Linking Queues and Ports`_ below).
156 This section will introduce the reader to the eventdev API, showing how to
157 create and configure an eventdev and use it for a two-stage atomic pipeline
158 with a single core for TX. The diagram below shows the final state of the
159 application after this walk-through:
161 .. _figure_eventdev-usage1:
163 .. figure:: img/eventdev_usage.*
165 Sample eventdev usage, with RX, two atomic stages and a single-link to TX.
168 A high level overview of the setup steps are:
170 * rte_event_dev_configure()
171 * rte_event_queue_setup()
172 * rte_event_port_setup()
173 * rte_event_port_link()
174 * rte_event_dev_start()
180 The eventdev library uses vdev options to add devices to the DPDK application.
181 The ``--vdev`` EAL option allows adding eventdev instances to your DPDK
182 application, using the name of the eventdev PMD as an argument.
184 For example, to create an instance of the software eventdev scheduler, the
185 following vdev arguments should be provided to the application EAL command line:
187 .. code-block:: console
189 ./dpdk_application --vdev="event_sw0"
191 In the following code, we configure eventdev instance with 3 queues
192 and 6 ports as follows. The 3 queues consist of 2 Atomic and 1 Single-Link,
193 while the 6 ports consist of 4 workers, 1 RX and 1 TX.
197 const struct rte_event_dev_config config = {
198 .nb_event_queues = 3,
200 .nb_events_limit = 4096,
201 .nb_event_queue_flows = 1024,
202 .nb_event_port_dequeue_depth = 128,
203 .nb_event_port_enqueue_depth = 128,
205 int err = rte_event_dev_configure(dev_id, &config);
207 The remainder of this walk-through assumes that dev_id is 0.
212 Once the eventdev itself is configured, the next step is to configure queues.
213 This is done by setting the appropriate values in a queue_conf structure, and
214 calling the setup function. Repeat this step for each queue, starting from
215 0 and ending at ``nb_event_queues - 1`` from the event_dev config above.
219 struct rte_event_queue_conf atomic_conf = {
220 .event_queue_cfg = RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY,
221 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
222 .nb_atomic_flows = 1024,
223 .nb_atomic_order_sequences = 1024,
227 int err = rte_event_queue_setup(dev_id, queue_id, &atomic_conf);
229 The remainder of this walk-through assumes that the queues are configured as
232 * id 0, atomic queue #1
233 * id 1, atomic queue #2
234 * id 2, single-link queue
239 Once queues are set up successfully, create the ports as required. Each port
240 should be set up with its corresponding port_conf type, worker for worker cores,
241 rx and tx for the RX and TX cores:
245 struct rte_event_port_conf rx_conf = {
246 .dequeue_depth = 128,
247 .enqueue_depth = 128,
248 .new_event_threshold = 1024,
250 struct rte_event_port_conf worker_conf = {
253 .new_event_threshold = 4096,
255 struct rte_event_port_conf tx_conf = {
256 .dequeue_depth = 128,
257 .enqueue_depth = 128,
258 .new_event_threshold = 4096,
262 int err = rte_event_port_setup(dev_id, port_id, &CORE_FUNCTION_conf);
264 It is now assumed that:
267 * ports 1,2,3,4: Workers
270 Linking Queues and Ports
271 ~~~~~~~~~~~~~~~~~~~~~~~~
273 The final step is to "wire up" the ports to the queues. After this, the
274 eventdev is capable of scheduling events, and when cores request work to do,
275 the correct events are provided to that core. Note that the RX core takes input
276 from eg: a NIC so it is not linked to any eventdev queues.
278 Linking all workers to atomic queues, and the TX core to the single-link queue
279 can be achieved like this:
284 uint8_t atomic_qs[] = {0, 1};
285 uint8_t single_link_q = 2;
286 uint8_t tx_port_id = 5;
287 uin8t_t priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
289 for(int i = 0; i < 4; i++) {
290 int worker_port = i + 1;
291 int links_made = rte_event_port_link(dev_id, worker_port, atomic_qs, NULL, 2);
293 int links_made = rte_event_port_link(dev_id, tx_port_id, &single_link_q, &priority, 1);
295 Starting the EventDev
296 ~~~~~~~~~~~~~~~~~~~~~
298 A single function call tells the eventdev instance to start processing
299 events. Note that all queues must be linked to for the instance to start, as
300 if any queue is not linked to, enqueuing to that queue will cause the
301 application to backpressure and eventually stall due to no space in the
306 int err = rte_event_dev_start(dev_id);
308 Ingress of New Events
309 ~~~~~~~~~~~~~~~~~~~~~
311 Now that the eventdev is set up, and ready to receive events, the RX core must
312 enqueue some events into the system for it to schedule. The events to be
313 scheduled are ordinary DPDK packets, received from an eth_rx_burst() as normal.
314 The following code shows how those packets can be enqueued into the eventdev:
318 const uint16_t nb_rx = rte_eth_rx_burst(eth_port, 0, mbufs, BATCH_SIZE);
320 for (i = 0; i < nb_rx; i++) {
321 ev[i].flow_id = mbufs[i]->hash.rss;
322 ev[i].op = RTE_EVENT_OP_NEW;
323 ev[i].sched_type = RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY;
325 ev[i].event_type = RTE_EVENT_TYPE_ETHDEV;
326 ev[i].sub_event_type = 0;
327 ev[i].priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
328 ev[i].mbuf = mbufs[i];
331 const int nb_tx = rte_event_enqueue_burst(dev_id, port_id, ev, nb_rx);
332 if (nb_tx != nb_rx) {
333 for(i = nb_tx; i < nb_rx; i++)
334 rte_pktmbuf_free(mbufs[i]);
340 Now that the RX core has injected events, there is work to be done by the
341 workers. Note that each worker will dequeue as many events as it can in a burst,
342 process each one individually, and then burst the packets back into the
345 The worker can lookup the events source from ``event.queue_id``, which should
346 indicate to the worker what workload needs to be performed on the event.
347 Once done, the worker can update the ``event.queue_id`` to a new value, to send
348 the event to the next stage in the pipeline.
353 struct rte_event events[BATCH_SIZE];
354 uint16_t nb_rx = rte_event_dequeue_burst(dev_id, worker_port_id, events, BATCH_SIZE, timeout);
356 for (i = 0; i < nb_rx; i++) {
357 /* process mbuf using events[i].queue_id as pipeline stage */
358 struct rte_mbuf *mbuf = events[i].mbuf;
359 /* Send event to next stage in pipeline */
360 events[i].queue_id++;
363 uint16_t nb_tx = rte_event_enqueue_burst(dev_id, port_id, events, nb_rx);
369 Finally, when the packet is ready for egress or needs to be dropped, we need
370 to inform the eventdev that the packet is no longer being handled by the
371 application. This can be done by calling dequeue() or dequeue_burst(), which
372 indicates that the previous burst of packets is no longer in use by the
375 An event driven worker thread has following typical workflow on fastpath:
380 rte_event_dequeue_burst(...);
382 rte_event_enqueue_burst(...);
389 The eventdev library allows an application to easily schedule events as it
390 requires, either using a run-to-completion or pipeline processing model. The
391 queues and ports abstract the logical functionality of an eventdev, providing
392 the application with a generic method to schedule events. With the flexible
393 PMD infrastructure applications benefit of improvements in existing eventdevs
394 and additions of new ones without modification.