1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2017 Intel Corporation
9 #include <rte_mempool.h>
10 #include <rte_cycles.h>
11 #include <rte_common.h>
13 #include <rte_distributor.h>
14 #include <rte_pause.h>
16 #define ITER_POWER_CL 25 /* log 2 of how many iterations for Cache Line test */
17 #define ITER_POWER 21 /* log 2 of how many iterations we do when timing. */
19 #define BIG_BATCH 1024
21 /* static vars - zero initialized by default */
22 static volatile int quit;
23 static volatile unsigned worker_idx;
26 volatile unsigned handled_packets;
27 } __rte_cache_aligned;
28 static struct worker_stats worker_stats[RTE_MAX_LCORE];
31 * worker thread used for testing the time to do a round-trip of a cache
32 * line between two cores and back again
35 flip_bit(volatile uint64_t *arg)
38 while (old_val != 2) {
48 * test case to time the number of cycles to round-trip a cache line between
49 * two cores and back again.
52 time_cache_line_switch(void)
54 /* allocate a full cache line for data, we use only first byte of it */
55 uint64_t data[RTE_CACHE_LINE_SIZE*3 / sizeof(uint64_t)];
57 unsigned int i, workerid = rte_get_next_lcore(rte_lcore_id(), 0, 0);
58 volatile uint64_t *pdata = &data[0];
60 rte_eal_remote_launch((lcore_function_t *)flip_bit, &data[0], workerid);
64 const uint64_t start_time = rte_rdtsc();
65 for (i = 0; i < (1 << ITER_POWER_CL); i++) {
70 const uint64_t end_time = rte_rdtsc();
75 rte_eal_wait_lcore(workerid);
76 printf("==== Cache line switch test ===\n");
77 printf("Time for %u iterations = %"PRIu64" ticks\n", (1<<ITER_POWER_CL),
79 printf("Ticks per iteration = %"PRIu64"\n\n",
80 (end_time-start_time) >> ITER_POWER_CL);
84 * returns the total count of the number of packets handled by the worker
85 * functions given below.
88 total_packet_count(void)
90 unsigned i, count = 0;
91 for (i = 0; i < worker_idx; i++)
92 count += worker_stats[i].handled_packets;
96 /* resets the packet counts for a new test */
98 clear_packet_count(void)
100 memset(&worker_stats, 0, sizeof(worker_stats));
104 * This is the basic worker function for performance tests.
105 * it does nothing but return packets and count them.
108 handle_work(void *arg)
110 struct rte_distributor *d = arg;
111 unsigned int num = 0;
113 unsigned int id = __atomic_fetch_add(&worker_idx, 1, __ATOMIC_RELAXED);
114 struct rte_mbuf *buf[8] __rte_cache_aligned;
116 for (i = 0; i < 8; i++)
119 num = rte_distributor_get_pkt(d, id, buf, buf, num);
121 worker_stats[id].handled_packets += num;
122 num = rte_distributor_get_pkt(d, id, buf, buf, num);
124 worker_stats[id].handled_packets += num;
125 rte_distributor_return_pkt(d, id, buf, num);
130 * This basic performance test just repeatedly sends in 32 packets at a time
131 * to the distributor and verifies at the end that we got them all in the worker
132 * threads and finally how long per packet the processing took.
135 perf_test(struct rte_distributor *d, struct rte_mempool *p)
139 struct rte_mbuf *bufs[BURST];
141 clear_packet_count();
142 if (rte_mempool_get_bulk(p, (void *)bufs, BURST) != 0) {
143 printf("Error getting mbufs from pool\n");
146 /* ensure we have different hash value for each pkt */
147 for (i = 0; i < BURST; i++)
148 bufs[i]->hash.usr = i;
151 for (i = 0; i < (1<<ITER_POWER); i++)
152 rte_distributor_process(d, bufs, BURST);
157 rte_distributor_process(d, NULL, 0);
158 } while (total_packet_count() < (BURST << ITER_POWER));
160 rte_distributor_clear_returns(d);
162 printf("Time per burst: %"PRIu64"\n", (end - start) >> ITER_POWER);
163 printf("Time per packet: %"PRIu64"\n\n",
164 ((end - start) >> ITER_POWER)/BURST);
165 rte_mempool_put_bulk(p, (void *)bufs, BURST);
167 for (i = 0; i < rte_lcore_count() - 1; i++)
168 printf("Worker %u handled %u packets\n", i,
169 worker_stats[i].handled_packets);
170 printf("Total packets: %u (%x)\n", total_packet_count(),
171 total_packet_count());
172 printf("=== Perf test done ===\n\n");
177 /* Useful function which ensures that all worker functions terminate */
179 quit_workers(struct rte_distributor *d, struct rte_mempool *p)
181 const unsigned int num_workers = rte_lcore_count() - 1;
183 struct rte_mbuf *bufs[RTE_MAX_LCORE];
185 rte_mempool_get_bulk(p, (void *)bufs, num_workers);
188 for (i = 0; i < num_workers; i++) {
189 bufs[i]->hash.usr = i << 1;
190 rte_distributor_process(d, &bufs[i], 1);
193 rte_mempool_put_bulk(p, (void *)bufs, num_workers);
195 rte_distributor_process(d, NULL, 0);
196 rte_distributor_flush(d);
197 rte_eal_mp_wait_lcore();
203 test_distributor_perf(void)
205 static struct rte_distributor *ds;
206 static struct rte_distributor *db;
207 static struct rte_mempool *p;
209 if (rte_lcore_count() < 2) {
210 printf("Not enough cores for distributor_perf_autotest, expecting at least 2\n");
214 /* first time how long it takes to round-trip a cache line */
215 time_cache_line_switch();
218 ds = rte_distributor_create("Test_perf", rte_socket_id(),
219 rte_lcore_count() - 1,
220 RTE_DIST_ALG_SINGLE);
222 printf("Error creating distributor\n");
226 rte_distributor_clear_returns(ds);
230 db = rte_distributor_create("Test_burst", rte_socket_id(),
231 rte_lcore_count() - 1,
234 printf("Error creating burst distributor\n");
238 rte_distributor_clear_returns(db);
241 const unsigned nb_bufs = (511 * rte_lcore_count()) < BIG_BATCH ?
242 (BIG_BATCH * 2) - 1 : (511 * rte_lcore_count());
244 p = rte_pktmbuf_pool_create("DPT_MBUF_POOL", nb_bufs, BURST,
245 0, RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
247 printf("Error creating mempool\n");
252 printf("=== Performance test of distributor (single mode) ===\n");
253 rte_eal_mp_remote_launch(handle_work, ds, SKIP_MAIN);
254 if (perf_test(ds, p) < 0)
258 printf("=== Performance test of distributor (burst mode) ===\n");
259 rte_eal_mp_remote_launch(handle_work, db, SKIP_MAIN);
260 if (perf_test(db, p) < 0)
267 REGISTER_TEST_COMMAND(distributor_perf_autotest, test_distributor_perf);