4 * Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 #include <sys/types.h>
41 #include <sys/queue.h>
46 #include <rte_common.h>
47 #include <rte_byteorder.h>
49 #include <rte_memory.h>
50 #include <rte_memcpy.h>
51 #include <rte_memzone.h>
52 #include <rte_tailq.h>
54 #include <rte_per_lcore.h>
55 #include <rte_launch.h>
56 #include <rte_atomic.h>
57 #include <rte_cycles.h>
58 #include <rte_prefetch.h>
59 #include <rte_lcore.h>
60 #include <rte_per_lcore.h>
61 #include <rte_branch_prediction.h>
62 #include <rte_interrupts.h>
64 #include <rte_random.h>
65 #include <rte_debug.h>
66 #include <rte_ether.h>
67 #include <rte_ethdev.h>
69 #include <rte_mempool.h>
77 #ifndef APP_LCORE_IO_FLUSH
78 #define APP_LCORE_IO_FLUSH 1000000
81 #ifndef APP_LCORE_WORKER_FLUSH
82 #define APP_LCORE_WORKER_FLUSH 1000000
86 #define APP_STATS 1000000
89 #define APP_IO_RX_DROP_ALL_PACKETS 0
90 #define APP_WORKER_DROP_ALL_PACKETS 0
91 #define APP_IO_TX_DROP_ALL_PACKETS 0
93 #ifndef APP_IO_RX_PREFETCH_ENABLE
94 #define APP_IO_RX_PREFETCH_ENABLE 1
97 #ifndef APP_WORKER_PREFETCH_ENABLE
98 #define APP_WORKER_PREFETCH_ENABLE 1
101 #ifndef APP_IO_TX_PREFETCH_ENABLE
102 #define APP_IO_TX_PREFETCH_ENABLE 1
105 #if APP_IO_RX_PREFETCH_ENABLE
106 #define APP_IO_RX_PREFETCH0(p) rte_prefetch0(p)
107 #define APP_IO_RX_PREFETCH1(p) rte_prefetch1(p)
109 #define APP_IO_RX_PREFETCH0(p)
110 #define APP_IO_RX_PREFETCH1(p)
113 #if APP_WORKER_PREFETCH_ENABLE
114 #define APP_WORKER_PREFETCH0(p) rte_prefetch0(p)
115 #define APP_WORKER_PREFETCH1(p) rte_prefetch1(p)
117 #define APP_WORKER_PREFETCH0(p)
118 #define APP_WORKER_PREFETCH1(p)
121 #if APP_IO_TX_PREFETCH_ENABLE
122 #define APP_IO_TX_PREFETCH0(p) rte_prefetch0(p)
123 #define APP_IO_TX_PREFETCH1(p) rte_prefetch1(p)
125 #define APP_IO_TX_PREFETCH0(p)
126 #define APP_IO_TX_PREFETCH1(p)
130 app_lcore_io_rx_buffer_to_send (
131 struct app_lcore_params_io *lp,
133 struct rte_mbuf *mbuf,
139 pos = lp->rx.mbuf_out[worker].n_mbufs;
140 lp->rx.mbuf_out[worker].array[pos ++] = mbuf;
141 if (likely(pos < bsz)) {
142 lp->rx.mbuf_out[worker].n_mbufs = pos;
146 ret = rte_ring_sp_enqueue_bulk(
147 lp->rx.rings[worker],
148 (void **) lp->rx.mbuf_out[worker].array,
151 if (unlikely(ret == -ENOBUFS)) {
153 for (k = 0; k < bsz; k ++) {
154 struct rte_mbuf *m = lp->rx.mbuf_out[worker].array[k];
159 lp->rx.mbuf_out[worker].n_mbufs = 0;
160 lp->rx.mbuf_out_flush[worker] = 0;
163 lp->rx.rings_iters[worker] ++;
164 if (likely(ret == 0)) {
165 lp->rx.rings_count[worker] ++;
167 if (unlikely(lp->rx.rings_iters[worker] == APP_STATS)) {
168 unsigned lcore = rte_lcore_id();
170 printf("\tI/O RX %u out (worker %u): enq success rate = %.2f\n",
173 ((double) lp->rx.rings_count[worker]) / ((double) lp->rx.rings_iters[worker]));
174 lp->rx.rings_iters[worker] = 0;
175 lp->rx.rings_count[worker] = 0;
182 struct app_lcore_params_io *lp,
188 struct rte_mbuf *mbuf_1_0, *mbuf_1_1, *mbuf_2_0, *mbuf_2_1;
189 uint8_t *data_1_0, *data_1_1 = NULL;
192 for (i = 0; i < lp->rx.n_nic_queues; i ++) {
193 uint8_t port = lp->rx.nic_queues[i].port;
194 uint8_t queue = lp->rx.nic_queues[i].queue;
197 n_mbufs = rte_eth_rx_burst(
200 lp->rx.mbuf_in.array,
203 if (unlikely(n_mbufs == 0)) {
208 lp->rx.nic_queues_iters[i] ++;
209 lp->rx.nic_queues_count[i] += n_mbufs;
210 if (unlikely(lp->rx.nic_queues_iters[i] == APP_STATS)) {
211 struct rte_eth_stats stats;
212 unsigned lcore = rte_lcore_id();
214 rte_eth_stats_get(port, &stats);
216 printf("I/O RX %u in (NIC port %u): NIC drop ratio = %.2f avg burst size = %.2f\n",
219 (double) stats.ierrors / (double) (stats.ierrors + stats.ipackets),
220 ((double) lp->rx.nic_queues_count[i]) / ((double) lp->rx.nic_queues_iters[i]));
221 lp->rx.nic_queues_iters[i] = 0;
222 lp->rx.nic_queues_count[i] = 0;
226 #if APP_IO_RX_DROP_ALL_PACKETS
227 for (j = 0; j < n_mbufs; j ++) {
228 struct rte_mbuf *pkt = lp->rx.mbuf_in.array[j];
229 rte_pktmbuf_free(pkt);
235 mbuf_1_0 = lp->rx.mbuf_in.array[0];
236 mbuf_1_1 = lp->rx.mbuf_in.array[1];
237 data_1_0 = rte_pktmbuf_mtod(mbuf_1_0, uint8_t *);
238 if (likely(n_mbufs > 1)) {
239 data_1_1 = rte_pktmbuf_mtod(mbuf_1_1, uint8_t *);
242 mbuf_2_0 = lp->rx.mbuf_in.array[2];
243 mbuf_2_1 = lp->rx.mbuf_in.array[3];
244 APP_IO_RX_PREFETCH0(mbuf_2_0);
245 APP_IO_RX_PREFETCH0(mbuf_2_1);
247 for (j = 0; j + 3 < n_mbufs; j += 2) {
248 struct rte_mbuf *mbuf_0_0, *mbuf_0_1;
249 uint8_t *data_0_0, *data_0_1;
250 uint32_t worker_0, worker_1;
259 data_1_0 = rte_pktmbuf_mtod(mbuf_2_0, uint8_t *);
260 data_1_1 = rte_pktmbuf_mtod(mbuf_2_1, uint8_t *);
261 APP_IO_RX_PREFETCH0(data_1_0);
262 APP_IO_RX_PREFETCH0(data_1_1);
264 mbuf_2_0 = lp->rx.mbuf_in.array[j+4];
265 mbuf_2_1 = lp->rx.mbuf_in.array[j+5];
266 APP_IO_RX_PREFETCH0(mbuf_2_0);
267 APP_IO_RX_PREFETCH0(mbuf_2_1);
269 worker_0 = data_0_0[pos_lb] & (n_workers - 1);
270 worker_1 = data_0_1[pos_lb] & (n_workers - 1);
272 app_lcore_io_rx_buffer_to_send(lp, worker_0, mbuf_0_0, bsz_wr);
273 app_lcore_io_rx_buffer_to_send(lp, worker_1, mbuf_0_1, bsz_wr);
276 /* Handle the last 1, 2 (when n_mbufs is even) or 3 (when n_mbufs is odd) packets */
277 for ( ; j < n_mbufs; j += 1) {
278 struct rte_mbuf *mbuf;
287 data = rte_pktmbuf_mtod(mbuf, uint8_t *);
289 APP_IO_RX_PREFETCH0(mbuf_1_0);
291 worker = data[pos_lb] & (n_workers - 1);
293 app_lcore_io_rx_buffer_to_send(lp, worker, mbuf, bsz_wr);
299 app_lcore_io_rx_flush(struct app_lcore_params_io *lp, uint32_t n_workers)
303 for (worker = 0; worker < n_workers; worker ++) {
306 if (likely((lp->rx.mbuf_out_flush[worker] == 0) ||
307 (lp->rx.mbuf_out[worker].n_mbufs == 0))) {
308 lp->rx.mbuf_out_flush[worker] = 1;
312 ret = rte_ring_sp_enqueue_bulk(
313 lp->rx.rings[worker],
314 (void **) lp->rx.mbuf_out[worker].array,
315 lp->rx.mbuf_out[worker].n_mbufs);
317 if (unlikely(ret < 0)) {
319 for (k = 0; k < lp->rx.mbuf_out[worker].n_mbufs; k ++) {
320 struct rte_mbuf *pkt_to_free = lp->rx.mbuf_out[worker].array[k];
321 rte_pktmbuf_free(pkt_to_free);
325 lp->rx.mbuf_out[worker].n_mbufs = 0;
326 lp->rx.mbuf_out_flush[worker] = 1;
332 struct app_lcore_params_io *lp,
339 for (worker = 0; worker < n_workers; worker ++) {
342 for (i = 0; i < lp->tx.n_nic_ports; i ++) {
343 uint8_t port = lp->tx.nic_ports[i];
344 struct rte_ring *ring = lp->tx.rings[port][worker];
345 uint32_t n_mbufs, n_pkts;
348 n_mbufs = lp->tx.mbuf_out[port].n_mbufs;
349 ret = rte_ring_sc_dequeue_bulk(
351 (void **) &lp->tx.mbuf_out[port].array[n_mbufs],
354 if (unlikely(ret == -ENOENT)) {
360 #if APP_IO_TX_DROP_ALL_PACKETS
363 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[0]);
364 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[1]);
366 for (j = 0; j < n_mbufs; j ++) {
367 if (likely(j < n_mbufs - 2)) {
368 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[j + 2]);
371 rte_pktmbuf_free(lp->tx.mbuf_out[port].array[j]);
374 lp->tx.mbuf_out[port].n_mbufs = 0;
380 if (unlikely(n_mbufs < bsz_wr)) {
381 lp->tx.mbuf_out[port].n_mbufs = n_mbufs;
385 n_pkts = rte_eth_tx_burst(
388 lp->tx.mbuf_out[port].array,
392 lp->tx.nic_ports_iters[port] ++;
393 lp->tx.nic_ports_count[port] += n_pkts;
394 if (unlikely(lp->tx.nic_ports_iters[port] == APP_STATS)) {
395 unsigned lcore = rte_lcore_id();
397 printf("\t\t\tI/O TX %u out (port %u): avg burst size = %.2f\n",
400 ((double) lp->tx.nic_ports_count[port]) / ((double) lp->tx.nic_ports_iters[port]));
401 lp->tx.nic_ports_iters[port] = 0;
402 lp->tx.nic_ports_count[port] = 0;
406 if (unlikely(n_pkts < n_mbufs)) {
408 for (k = n_pkts; k < n_mbufs; k ++) {
409 struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
410 rte_pktmbuf_free(pkt_to_free);
413 lp->tx.mbuf_out[port].n_mbufs = 0;
414 lp->tx.mbuf_out_flush[port] = 0;
420 app_lcore_io_tx_flush(struct app_lcore_params_io *lp)
424 for (port = 0; port < lp->tx.n_nic_ports; port ++) {
427 if (likely((lp->tx.mbuf_out_flush[port] == 0) ||
428 (lp->tx.mbuf_out[port].n_mbufs == 0))) {
429 lp->tx.mbuf_out_flush[port] = 1;
433 n_pkts = rte_eth_tx_burst(
436 lp->tx.mbuf_out[port].array,
437 (uint16_t) lp->tx.mbuf_out[port].n_mbufs);
439 if (unlikely(n_pkts < lp->tx.mbuf_out[port].n_mbufs)) {
441 for (k = n_pkts; k < lp->tx.mbuf_out[port].n_mbufs; k ++) {
442 struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
443 rte_pktmbuf_free(pkt_to_free);
447 lp->tx.mbuf_out[port].n_mbufs = 0;
448 lp->tx.mbuf_out_flush[port] = 1;
453 app_lcore_main_loop_io(void)
455 uint32_t lcore = rte_lcore_id();
456 struct app_lcore_params_io *lp = &app.lcore_params[lcore].io;
457 uint32_t n_workers = app_get_lcores_worker();
460 uint32_t bsz_rx_rd = app.burst_size_io_rx_read;
461 uint32_t bsz_rx_wr = app.burst_size_io_rx_write;
462 uint32_t bsz_tx_rd = app.burst_size_io_tx_read;
463 uint32_t bsz_tx_wr = app.burst_size_io_tx_write;
465 uint8_t pos_lb = app.pos_lb;
468 if (APP_LCORE_IO_FLUSH && (unlikely(i == APP_LCORE_IO_FLUSH))) {
469 if (likely(lp->rx.n_nic_queues > 0)) {
470 app_lcore_io_rx_flush(lp, n_workers);
473 if (likely(lp->tx.n_nic_ports > 0)) {
474 app_lcore_io_tx_flush(lp);
480 if (likely(lp->rx.n_nic_queues > 0)) {
481 app_lcore_io_rx(lp, n_workers, bsz_rx_rd, bsz_rx_wr, pos_lb);
484 if (likely(lp->tx.n_nic_ports > 0)) {
485 app_lcore_io_tx(lp, n_workers, bsz_tx_rd, bsz_tx_wr);
494 struct app_lcore_params_worker *lp,
500 for (i = 0; i < lp->n_rings_in; i ++) {
501 struct rte_ring *ring_in = lp->rings_in[i];
505 ret = rte_ring_sc_dequeue_bulk(
507 (void **) lp->mbuf_in.array,
510 if (unlikely(ret == -ENOENT)) {
514 #if APP_WORKER_DROP_ALL_PACKETS
515 for (j = 0; j < bsz_rd; j ++) {
516 struct rte_mbuf *pkt = lp->mbuf_in.array[j];
517 rte_pktmbuf_free(pkt);
523 APP_WORKER_PREFETCH1(rte_pktmbuf_mtod(lp->mbuf_in.array[0], unsigned char *));
524 APP_WORKER_PREFETCH0(lp->mbuf_in.array[1]);
526 for (j = 0; j < bsz_rd; j ++) {
527 struct rte_mbuf *pkt;
528 struct ipv4_hdr *ipv4_hdr;
529 uint32_t ipv4_dst, pos;
532 if (likely(j < bsz_rd - 1)) {
533 APP_WORKER_PREFETCH1(rte_pktmbuf_mtod(lp->mbuf_in.array[j+1], unsigned char *));
535 if (likely(j < bsz_rd - 2)) {
536 APP_WORKER_PREFETCH0(lp->mbuf_in.array[j+2]);
539 pkt = lp->mbuf_in.array[j];
540 ipv4_hdr = (struct ipv4_hdr *)(rte_pktmbuf_mtod(pkt, unsigned char *) + sizeof(struct ether_hdr));
541 ipv4_dst = rte_be_to_cpu_32(ipv4_hdr->dst_addr);
543 if (unlikely(rte_lpm_lookup(lp->lpm_table, ipv4_dst, &port) != 0)) {
544 port = pkt->pkt.in_port;
547 pos = lp->mbuf_out[port].n_mbufs;
549 lp->mbuf_out[port].array[pos ++] = pkt;
550 if (likely(pos < bsz_wr)) {
551 lp->mbuf_out[port].n_mbufs = pos;
555 ret = rte_ring_sp_enqueue_bulk(
557 (void **) lp->mbuf_out[port].array,
561 lp->rings_out_iters[port] ++;
563 lp->rings_out_count[port] += 1;
565 if (lp->rings_out_iters[port] == APP_STATS){
566 printf("\t\tWorker %u out (NIC port %u): enq success rate = %.2f\n",
567 (unsigned) lp->worker_id,
569 ((double) lp->rings_out_count[port]) / ((double) lp->rings_out_iters[port]));
570 lp->rings_out_iters[port] = 0;
571 lp->rings_out_count[port] = 0;
575 if (unlikely(ret == -ENOBUFS)) {
577 for (k = 0; k < bsz_wr; k ++) {
578 struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
579 rte_pktmbuf_free(pkt_to_free);
583 lp->mbuf_out[port].n_mbufs = 0;
584 lp->mbuf_out_flush[port] = 0;
590 app_lcore_worker_flush(struct app_lcore_params_worker *lp)
594 for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
597 if (unlikely(lp->rings_out[port] == NULL)) {
601 if (likely((lp->mbuf_out_flush[port] == 0) ||
602 (lp->mbuf_out[port].n_mbufs == 0))) {
603 lp->mbuf_out_flush[port] = 1;
607 ret = rte_ring_sp_enqueue_bulk(
609 (void **) lp->mbuf_out[port].array,
610 lp->mbuf_out[port].n_mbufs);
612 if (unlikely(ret < 0)) {
614 for (k = 0; k < lp->mbuf_out[port].n_mbufs; k ++) {
615 struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
616 rte_pktmbuf_free(pkt_to_free);
620 lp->mbuf_out[port].n_mbufs = 0;
621 lp->mbuf_out_flush[port] = 1;
626 app_lcore_main_loop_worker(void) {
627 uint32_t lcore = rte_lcore_id();
628 struct app_lcore_params_worker *lp = &app.lcore_params[lcore].worker;
631 uint32_t bsz_rd = app.burst_size_worker_read;
632 uint32_t bsz_wr = app.burst_size_worker_write;
635 if (APP_LCORE_WORKER_FLUSH && (unlikely(i == APP_LCORE_WORKER_FLUSH))) {
636 app_lcore_worker_flush(lp);
640 app_lcore_worker(lp, bsz_rd, bsz_wr);
647 app_lcore_main_loop(__attribute__((unused)) void *arg)
649 struct app_lcore_params *lp;
652 lcore = rte_lcore_id();
653 lp = &app.lcore_params[lcore];
655 if (lp->type == e_APP_LCORE_IO) {
656 printf("Logical core %u (I/O) main loop.\n", lcore);
657 app_lcore_main_loop_io();
660 if (lp->type == e_APP_LCORE_WORKER) {
661 printf("Logical core %u (worker %u) main loop.\n",
663 (unsigned) lp->worker.worker_id);
664 app_lcore_main_loop_worker();