2 Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
5 Redistribution and use in source and binary forms, with or without
6 modification, are permitted provided that the following conditions
9 * Redistributions of source code must retain the above copyright
10 notice, this list of conditions and the following disclaimer.
11 * Redistributions in binary form must reproduce the above copyright
12 notice, this list of conditions and the following disclaimer in
13 the documentation and/or other materials provided with the
15 * Neither the name of Intel Corporation nor the names of its
16 contributors may be used to endorse or promote products derived
17 from this software without specific prior written permission.
19 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
25 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 L2 Forwarding Sample Application (in Real and Virtualized Environments) with core load statistics.
32 ==================================================================================================
34 The L2 Forwarding sample application is a simple example of packet processing using
35 the Data Plane Development Kit (DPDK) which
36 also takes advantage of Single Root I/O Virtualization (SR-IOV) features in a virtualized environment.
40 This application is a variation of L2 Forwarding sample application. It demonstrate possible
41 scheme of job stats library usage therefore some parts of this document is identical with original
42 L2 forwarding application.
47 The L2 Forwarding sample application, which can operate in real and virtualized environments,
48 performs L2 forwarding for each packet that is received.
49 The destination port is the adjacent port from the enabled portmask, that is,
50 if the first four ports are enabled (portmask 0xf),
51 ports 1 and 2 forward into each other, and ports 3 and 4 forward into each other.
52 Also, the MAC addresses are affected as follows:
54 * The source MAC address is replaced by the TX port MAC address
56 * The destination MAC address is replaced by 02:00:00:00:00:TX_PORT_ID
58 This application can be used to benchmark performance using a traffic-generator, as shown in the Figure 3.
60 The application can also be used in a virtualized environment as shown in Figure 4.
62 The L2 Forwarding application can also be used as a starting point for developing a new application based on the DPDK.
66 **Figure 3. Performance Benchmark Setup (Basic Environment)**
68 .. image4_png has been replaced
70 |l2_fwd_benchmark_setup|
74 **Figure 4. Performance Benchmark Setup (Virtualized Environment)**
76 .. image5_png has been renamed
78 |l2_fwd_virtenv_benchmark_setup|
80 Virtual Function Setup Instructions
81 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
83 This application can use the virtual function available in the system and
84 therefore can be used in a virtual machine without passing through
85 the whole Network Device into a guest machine in a virtualized scenario.
86 The virtual functions can be enabled in the host machine or the hypervisor with the respective physical function driver.
88 For example, in a Linux* host machine, it is possible to enable a virtual function using the following command:
90 .. code-block:: console
92 modprobe ixgbe max_vfs=2,2
94 This command enables two Virtual Functions on each of Physical Function of the NIC,
95 with two physical ports in the PCI configuration space.
96 It is important to note that enabled Virtual Function 0 and 2 would belong to Physical Function 0
97 and Virtual Function 1 and 3 would belong to Physical Function 1,
98 in this case enabling a total of four Virtual Functions.
100 Compiling the Application
101 -------------------------
103 #. Go to the example directory:
105 .. code-block:: console
107 export RTE_SDK=/path/to/rte_sdk cd ${RTE_SDK}/examples/l2fwd-jobstats
109 #. Set the target (a default target is used if not specified). For example:
111 .. code-block:: console
113 export RTE_TARGET=x86_64-native-linuxapp-gcc
115 *See the DPDK Getting Started Guide* for possible RTE_TARGET values.
117 #. Build the application:
119 .. code-block:: console
123 Running the Application
124 -----------------------
126 The application requires a number of command line options:
128 .. code-block:: console
130 ./build/l2fwd-jobstats [EAL options] -- -p PORTMASK [-q NQ] [-l]
134 * p PORTMASK: A hexadecimal bitmask of the ports to configure
136 * q NQ: A number of queues (=ports) per lcore (default is 1)
138 * l: Use locale thousands separator when formatting big numbers.
140 To run the application in linuxapp environment with 4 lcores, 16 ports, 8 RX queues per lcore and
141 thousands separator printing, issue the command:
143 .. code-block:: console
145 $ ./build/l2fwd-jobstats -c f -n 4 -- -q 8 -p ffff -l
147 Refer to the *DPDK Getting Started Guide* for general information on running applications
148 and the Environment Abstraction Layer (EAL) options.
153 The following sections provide some explanation of the code.
155 Command Line Arguments
156 ~~~~~~~~~~~~~~~~~~~~~~
158 The L2 Forwarding sample application takes specific parameters,
159 in addition to Environment Abstraction Layer (EAL) arguments (see Section 9.3).
160 The preferred way to parse parameters is to use the getopt() function,
161 since it is part of a well-defined and portable library.
163 The parsing of arguments is done in the l2fwd_parse_args() function.
164 The method of argument parsing is not described here.
165 Refer to the *glibc getopt(3)* man page for details.
167 EAL arguments are parsed first, then application-specific arguments.
168 This is done at the beginning of the main() function:
174 ret = rte_eal_init(argc, argv);
176 rte_exit(EXIT_FAILURE, "Invalid EAL arguments\n");
181 /* parse application arguments (after the EAL ones) */
183 ret = l2fwd_parse_args(argc, argv);
185 rte_exit(EXIT_FAILURE, "Invalid L2FWD arguments\n");
187 Mbuf Pool Initialization
188 ~~~~~~~~~~~~~~~~~~~~~~~~
190 Once the arguments are parsed, the mbuf pool is created.
191 The mbuf pool contains a set of mbuf objects that will be used by the driver
192 and the application to store network packet data:
196 /* create the mbuf pool */
198 rte_mempool_create("mbuf_pool", NB_MBUF,
200 sizeof(struct rte_pktmbuf_pool_private),
201 rte_pktmbuf_pool_init, NULL,
202 rte_pktmbuf_init, NULL,
205 if (l2fwd_pktmbuf_pool == NULL)
206 rte_exit(EXIT_FAILURE, "Cannot init mbuf pool\n");
208 The rte_mempool is a generic structure used to handle pools of objects.
209 In this case, it is necessary to create a pool that will be used by the driver,
210 which expects to have some reserved space in the mempool structure,
211 sizeof(struct rte_pktmbuf_pool_private) bytes.
212 The number of allocated pkt mbufs is NB_MBUF, with a size of MBUF_SIZE each.
213 A per-lcore cache of 32 mbufs is kept.
214 The memory is allocated in rte_socket_id() socket,
215 but it is possible to extend this code to allocate one mbuf pool per socket.
217 Two callback pointers are also given to the rte_mempool_create() function:
219 * The first callback pointer is to rte_pktmbuf_pool_init() and is used
220 to initialize the private data of the mempool, which is needed by the driver.
221 This function is provided by the mbuf API, but can be copied and extended by the developer.
223 * The second callback pointer given to rte_mempool_create() is the mbuf initializer.
224 The default is used, that is, rte_pktmbuf_init(), which is provided in the rte_mbuf library.
225 If a more complex application wants to extend the rte_pktmbuf structure for its own needs,
226 a new function derived from rte_pktmbuf_init( ) can be created.
228 Driver Initialization
229 ~~~~~~~~~~~~~~~~~~~~~
231 The main part of the code in the main() function relates to the initialization of the driver.
232 To fully understand this code, it is recommended to study the chapters that related to the Poll Mode Driver
233 in the *DPDK Programmer's Guide* and the *DPDK API Reference*.
237 nb_ports = rte_eth_dev_count();
240 rte_exit(EXIT_FAILURE, "No Ethernet ports - bye\n");
242 if (nb_ports > RTE_MAX_ETHPORTS)
243 nb_ports = RTE_MAX_ETHPORTS;
245 /* reset l2fwd_dst_ports */
247 for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++)
248 l2fwd_dst_ports[portid] = 0;
253 * Each logical core is assigned a dedicated TX queue on each port.
255 for (portid = 0; portid < nb_ports; portid++) {
256 /* skip ports that are not enabled */
257 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0)
260 if (nb_ports_in_mask % 2) {
261 l2fwd_dst_ports[portid] = last_port;
262 l2fwd_dst_ports[last_port] = portid;
269 rte_eth_dev_info_get((uint8_t) portid, &dev_info);
272 The next step is to configure the RX and TX queues.
273 For each port, there is only one RX queue (only one lcore is able to poll a given port).
274 The number of TX queues depends on the number of available lcores.
275 The rte_eth_dev_configure() function is used to configure the number of queues for a port:
279 ret = rte_eth_dev_configure((uint8_t)portid, 1, 1, &port_conf);
281 rte_exit(EXIT_FAILURE, "Cannot configure device: "
285 The global configuration is stored in a static structure:
289 static const struct rte_eth_conf port_conf = {
292 .header_split = 0, /**< Header Split disabled */
293 .hw_ip_checksum = 0, /**< IP checksum offload disabled */
294 .hw_vlan_filter = 0, /**< VLAN filtering disabled */
295 .jumbo_frame = 0, /**< Jumbo Frame Support disabled */
296 .hw_strip_crc= 0, /**< CRC stripped by hardware */
300 .mq_mode = ETH_DCB_NONE
304 RX Queue Initialization
305 ~~~~~~~~~~~~~~~~~~~~~~~
307 The application uses one lcore to poll one or several ports, depending on the -q option,
308 which specifies the number of queues per lcore.
310 For example, if the user specifies -q 4, the application is able to poll four ports with one lcore.
311 If there are 16 ports on the target (and if the portmask argument is -p ffff ),
312 the application will need four lcores to poll all the ports.
316 ret = rte_eth_rx_queue_setup(portid, 0, nb_rxd,
317 rte_eth_dev_socket_id(portid),
322 rte_exit(EXIT_FAILURE, "rte_eth_rx_queue_setup:err=%d, port=%u\n",
323 ret, (unsigned) portid);
325 The list of queues that must be polled for a given lcore is stored in a private structure called struct lcore_queue_conf.
329 struct lcore_queue_conf {
331 unsigned rx_port_list[MAX_RX_QUEUE_PER_LCORE];
332 truct mbuf_table tx_mbufs[RTE_MAX_ETHPORTS];
334 struct rte_timer rx_timers[MAX_RX_QUEUE_PER_LCORE];
335 struct rte_jobstats port_fwd_jobs[MAX_RX_QUEUE_PER_LCORE];
337 struct rte_timer flush_timer;
338 struct rte_jobstats flush_job;
339 struct rte_jobstats idle_job;
340 struct rte_jobstats_context jobs_context;
342 rte_atomic16_t stats_read_pending;
344 } __rte_cache_aligned;
346 Values of struct lcore_queue_conf:
348 * n_rx_port and rx_port_list[] are used in the main packet processing loop
349 (see Section 9.4.6 "Receive, Process and Transmit Packets" later in this chapter).
351 * rx_timers and flush_timer are used to ensure forced TX on low packet rate.
353 * flush_job, idle_job and jobs_context are librte_jobstats objects used for managing l2fwd jobs.
355 * stats_read_pending and lock are used during job stats read phase.
357 TX Queue Initialization
358 ~~~~~~~~~~~~~~~~~~~~~~~
360 Each lcore should be able to transmit on any port. For every port, a single TX queue is initialized.
364 /* init one TX queue on each port */
367 ret = rte_eth_tx_queue_setup(portid, 0, nb_txd,
368 rte_eth_dev_socket_id(portid),
371 rte_exit(EXIT_FAILURE, "rte_eth_tx_queue_setup:err=%d, port=%u\n",
372 ret, (unsigned) portid);
374 Jobs statistics initialization
375 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
376 There are several statistics objects available:
378 * Flush job statistics
382 rte_jobstats_init(&qconf->flush_job, "flush", drain_tsc, drain_tsc,
385 rte_timer_init(&qconf->flush_timer);
386 ret = rte_timer_reset(&qconf->flush_timer, drain_tsc, PERIODICAL,
387 lcore_id, &l2fwd_flush_job, NULL);
390 rte_exit(1, "Failed to reset flush job timer for lcore %u: %s",
391 lcore_id, rte_strerror(-ret));
394 * Statistics per RX port
398 rte_jobstats_init(job, name, 0, drain_tsc, 0, MAX_PKT_BURST);
399 rte_jobstats_set_update_period_function(job, l2fwd_job_update_cb);
401 rte_timer_init(&qconf->rx_timers[i]);
402 ret = rte_timer_reset(&qconf->rx_timers[i], 0, PERIODICAL, lcore_id,
403 l2fwd_fwd_job, (void *)(uintptr_t)i);
406 rte_exit(1, "Failed to reset lcore %u port %u job timer: %s",
407 lcore_id, qconf->rx_port_list[i], rte_strerror(-ret));
410 Following parameters are passed to rte_jobstats_init():
412 * 0 as minimal poll period
414 * drain_tsc as maximum poll period
416 * MAX_PKT_BURST as desired target value (RX burst size)
421 The forwarding path is reworked comparing to original L2 Forwarding application.
422 In the l2fwd_main_loop() function three loops are placed.
427 rte_spinlock_lock(&qconf->lock);
430 rte_jobstats_context_start(&qconf->jobs_context);
433 * - Read stats_read_pending flag
434 * - check if some real job need to be executed
436 rte_jobstats_start(&qconf->jobs_context, &qconf->idle_job);
440 uint64_t now = rte_get_timer_cycles();
442 need_manage = qconf->flush_timer.expire < now;
443 /* Check if we was esked to give a stats. */
445 rte_atomic16_read(&qconf->stats_read_pending);
446 need_manage |= stats_read_pending;
448 for (i = 0; i < qconf->n_rx_port && !need_manage; i++)
449 need_manage = qconf->rx_timers[i].expire < now;
451 } while (!need_manage);
452 rte_jobstats_finish(&qconf->idle_job, qconf->idle_job.target);
455 rte_jobstats_context_finish(&qconf->jobs_context);
456 } while (likely(stats_read_pending == 0));
458 rte_spinlock_unlock(&qconf->lock);
462 First inifnite for loop is to minimize impact of stats reading. Lock is only locked/unlocked when asked.
464 Second inner while loop do the whole jobs management. When any job is ready, the use rte_timer_manage() is used to call the job handler.
465 In this place functions l2fwd_fwd_job() and l2fwd_flush_job() are called when needed.
466 Then rte_jobstats_context_finish() is called to mark loop end - no other jobs are ready to execute. By this time stats are ready to be read
467 and if stats_read_pending is set, loop breaks allowing stats to be read.
469 Third do-while loop is the idle job (idle stats counter). Its only purpose is moniting if any job is ready or stats job read is pending
470 for this lcore. Statistics from this part of code is considered as the headroom available fo additional processing.
472 Receive, Process and Transmit Packets
473 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
475 The main task of l2fwd_fwd_job() function is to read ingress packets from the RX queue of particular port and forward it.
476 This is done using the following code:
480 total_nb_rx = rte_eth_rx_burst((uint8_t) portid, 0, pkts_burst,
483 for (j = 0; j < total_nb_rx; j++) {
485 rte_prefetch0(rte_pktmbuf_mtod(m, void *));
486 l2fwd_simple_forward(m, portid);
489 Packets are read in a burst of size MAX_PKT_BURST.
490 Then, each mbuf in the table is processed by the l2fwd_simple_forward() function.
491 The processing is very simple: process the TX port from the RX port, then replace the source and destination MAC addresses.
493 The rte_eth_rx_burst() function writes the mbuf pointers in a local table and returns the number of available mbufs in the table.
495 After first read second try is issued.
499 if (total_nb_rx == MAX_PKT_BURST) {
500 const uint16_t nb_rx = rte_eth_rx_burst((uint8_t) portid, 0, pkts_burst,
503 total_nb_rx += nb_rx;
504 for (j = 0; j < nb_rx; j++) {
506 rte_prefetch0(rte_pktmbuf_mtod(m, void *));
507 l2fwd_simple_forward(m, portid);
511 This second read is important to give job stats library a feedback how many packets was processed.
515 /* Adjust period time in which we are running here. */
516 if (rte_jobstats_finish(job, total_nb_rx) != 0) {
517 rte_timer_reset(&qconf->rx_timers[port_idx], job->period, PERIODICAL,
518 lcore_id, l2fwd_fwd_job, arg);
521 To maximize performance exactly MAX_PKT_BURST is expected (the target value) to be read for each l2fwd_fwd_job() call.
522 If total_nb_rx is smaller than target value job->period will be increased. If it is greater the period will be decreased.
526 In the following code, one line for getting the output port requires some explanation.
528 During the initialization process, a static array of destination ports (l2fwd_dst_ports[]) is filled such that for each source port,
529 a destination port is assigned that is either the next or previous enabled port from the portmask.
530 Naturally, the number of ports in the portmask must be even, otherwise, the application exits.
535 l2fwd_simple_forward(struct rte_mbuf *m, unsigned portid)
537 struct ether_hdr *eth;
541 dst_port = l2fwd_dst_ports[portid];
543 eth = rte_pktmbuf_mtod(m, struct ether_hdr *);
545 /* 02:00:00:00:00:xx */
547 tmp = ð->d_addr.addr_bytes[0];
549 *((uint64_t *)tmp) = 0x000000000002 + ((uint64_t) dst_port << 40);
553 ether_addr_copy(&l2fwd_ports_eth_addr[dst_port], ð->s_addr);
555 l2fwd_send_packet(m, (uint8_t) dst_port);
558 Then, the packet is sent using the l2fwd_send_packet (m, dst_port) function.
559 For this test application, the processing is exactly the same for all packets arriving on the same RX port.
560 Therefore, it would have been possible to call the l2fwd_send_burst() function directly from the main loop
561 to send all the received packets on the same TX port,
562 using the burst-oriented send function, which is more efficient.
564 However, in real-life applications (such as, L3 routing),
565 packet N is not necessarily forwarded on the same port as packet N-1.
566 The application is implemented to illustrate that, so the same approach can be reused in a more complex application.
568 The l2fwd_send_packet() function stores the packet in a per-lcore and per-txport table.
569 If the table is full, the whole packets table is transmitted using the l2fwd_send_burst() function:
573 /* Send the packet on an output interface */
576 l2fwd_send_packet(struct rte_mbuf *m, uint8_t port)
578 unsigned lcore_id, len;
579 struct lcore_queue_conf *qconf;
581 lcore_id = rte_lcore_id();
582 qconf = &lcore_queue_conf[lcore_id];
583 len = qconf->tx_mbufs[port].len;
584 qconf->tx_mbufs[port].m_table[len] = m;
587 /* enough pkts to be sent */
589 if (unlikely(len == MAX_PKT_BURST)) {
590 l2fwd_send_burst(qconf, MAX_PKT_BURST, port);
594 qconf->tx_mbufs[port].len = len; return 0;
597 To ensure that no packets remain in the tables, the flush job exists. The l2fwd_flush_job()
598 is called periodicaly to for each lcore draining TX queue of each port.
599 This technique introduces some latency when there are not many packets to send,
600 however it improves performance:
605 l2fwd_flush_job(__rte_unused struct rte_timer *timer, __rte_unused void *arg)
609 struct lcore_queue_conf *qconf;
610 struct mbuf_table *m_table;
613 lcore_id = rte_lcore_id();
614 qconf = &lcore_queue_conf[lcore_id];
616 rte_jobstats_start(&qconf->jobs_context, &qconf->flush_job);
618 now = rte_get_timer_cycles();
619 lcore_id = rte_lcore_id();
620 qconf = &lcore_queue_conf[lcore_id];
621 for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++) {
622 m_table = &qconf->tx_mbufs[portid];
623 if (m_table->len == 0 || m_table->next_flush_time <= now)
626 l2fwd_send_burst(qconf, portid);
630 /* Pass target to indicate that this job is happy of time interval
631 * in which it was called. */
632 rte_jobstats_finish(&qconf->flush_job, qconf->flush_job.target);
635 .. |l2_fwd_benchmark_setup| image:: img/l2_fwd_benchmark_setup.*
637 .. |l2_fwd_virtenv_benchmark_setup| image:: img/l2_fwd_virtenv_benchmark_setup.*