1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2017 Intel Corporation
3 */
4
5 #include "test.h"
6
7 #include <unistd.h>
8 #include <string.h>
9 #include <rte_mempool.h>
10 #include <rte_cycles.h>
11 #include <rte_common.h>
12 #include <rte_mbuf.h>
13
14 #ifdef RTE_EXEC_ENV_WINDOWS
15 static int
test_distributor_perf(void)16 test_distributor_perf(void)
17 {
18 printf("distributor perf not supported on Windows, skipping test\n");
19 return TEST_SKIPPED;
20 }
21
22 #else
23
24 #include <rte_distributor.h>
25 #include <rte_pause.h>
26
27 #define ITER_POWER_CL 25 /* log 2 of how many iterations for Cache Line test */
28 #define ITER_POWER 21 /* log 2 of how many iterations we do when timing. */
29 #define BURST 64
30 #define BIG_BATCH 1024
31
32 /* static vars - zero initialized by default */
33 static volatile int quit;
34 static volatile RTE_ATOMIC(unsigned int) worker_idx;
35
36 struct __rte_cache_aligned worker_stats {
37 volatile unsigned handled_packets;
38 };
39 static struct worker_stats worker_stats[RTE_MAX_LCORE];
40
41 /*
42 * worker thread used for testing the time to do a round-trip of a cache
43 * line between two cores and back again
44 */
45 static int
flip_bit(volatile uint64_t * arg)46 flip_bit(volatile uint64_t *arg)
47 {
48 uint64_t old_val = 0;
49 while (old_val != 2) {
50 while (!*arg)
51 rte_pause();
52 old_val = *arg;
53 *arg = 0;
54 }
55 return 0;
56 }
57
58 /*
59 * test case to time the number of cycles to round-trip a cache line between
60 * two cores and back again.
61 */
62 static void
time_cache_line_switch(void)63 time_cache_line_switch(void)
64 {
65 /* allocate a full cache line for data, we use only first byte of it */
66 uint64_t data[RTE_CACHE_LINE_SIZE*3 / sizeof(uint64_t)];
67
68 unsigned int i, workerid = rte_get_next_lcore(rte_lcore_id(), 0, 0);
69 volatile uint64_t *pdata = &data[0];
70 *pdata = 1;
71 rte_eal_remote_launch((lcore_function_t *)flip_bit, &data[0], workerid);
72 while (*pdata)
73 rte_pause();
74
75 const uint64_t start_time = rte_rdtsc();
76 for (i = 0; i < (1 << ITER_POWER_CL); i++) {
77 while (*pdata)
78 rte_pause();
79 *pdata = 1;
80 }
81 const uint64_t end_time = rte_rdtsc();
82
83 while (*pdata)
84 rte_pause();
85 *pdata = 2;
86 rte_eal_wait_lcore(workerid);
87 printf("==== Cache line switch test ===\n");
88 printf("Time for %u iterations = %"PRIu64" ticks\n", (1<<ITER_POWER_CL),
89 end_time-start_time);
90 printf("Ticks per iteration = %"PRIu64"\n\n",
91 (end_time-start_time) >> ITER_POWER_CL);
92 }
93
94 /*
95 * returns the total count of the number of packets handled by the worker
96 * functions given below.
97 */
98 static unsigned
total_packet_count(void)99 total_packet_count(void)
100 {
101 unsigned i, count = 0;
102 for (i = 0; i < worker_idx; i++)
103 count += worker_stats[i].handled_packets;
104 return count;
105 }
106
107 /* resets the packet counts for a new test */
108 static void
clear_packet_count(void)109 clear_packet_count(void)
110 {
111 memset(&worker_stats, 0, sizeof(worker_stats));
112 }
113
114 /*
115 * This is the basic worker function for performance tests.
116 * it does nothing but return packets and count them.
117 */
118 static int
handle_work(void * arg)119 handle_work(void *arg)
120 {
121 struct rte_distributor *d = arg;
122 unsigned int num = 0;
123 int i;
124 unsigned int id = rte_atomic_fetch_add_explicit(&worker_idx, 1, rte_memory_order_relaxed);
125 alignas(RTE_CACHE_LINE_SIZE) struct rte_mbuf *buf[8];
126
127 for (i = 0; i < 8; i++)
128 buf[i] = NULL;
129
130 num = rte_distributor_get_pkt(d, id, buf, buf, num);
131 while (!quit) {
132 worker_stats[id].handled_packets += num;
133 num = rte_distributor_get_pkt(d, id, buf, buf, num);
134 }
135 worker_stats[id].handled_packets += num;
136 rte_distributor_return_pkt(d, id, buf, num);
137 return 0;
138 }
139
140 /*
141 * This basic performance test just repeatedly sends in 32 packets at a time
142 * to the distributor and verifies at the end that we got them all in the worker
143 * threads and finally how long per packet the processing took.
144 */
145 static inline int
perf_test(struct rte_distributor * d,struct rte_mempool * p)146 perf_test(struct rte_distributor *d, struct rte_mempool *p)
147 {
148 unsigned int i;
149 uint64_t start, end;
150 struct rte_mbuf *bufs[BURST];
151
152 clear_packet_count();
153 if (rte_mempool_get_bulk(p, (void *)bufs, BURST) != 0) {
154 printf("Error getting mbufs from pool\n");
155 return -1;
156 }
157 /* ensure we have different hash value for each pkt */
158 for (i = 0; i < BURST; i++)
159 bufs[i]->hash.usr = i;
160
161 start = rte_rdtsc();
162 for (i = 0; i < (1<<ITER_POWER); i++)
163 rte_distributor_process(d, bufs, BURST);
164 end = rte_rdtsc();
165
166 do {
167 usleep(100);
168 rte_distributor_process(d, NULL, 0);
169 } while (total_packet_count() < (BURST << ITER_POWER));
170
171 rte_distributor_clear_returns(d);
172
173 printf("Time per burst: %"PRIu64"\n", (end - start) >> ITER_POWER);
174 printf("Time per packet: %"PRIu64"\n\n",
175 ((end - start) >> ITER_POWER)/BURST);
176 rte_mempool_put_bulk(p, (void *)bufs, BURST);
177
178 for (i = 0; i < rte_lcore_count() - 1; i++)
179 printf("Worker %u handled %u packets\n", i,
180 worker_stats[i].handled_packets);
181 printf("Total packets: %u (%x)\n", total_packet_count(),
182 total_packet_count());
183 printf("=== Perf test done ===\n\n");
184
185 return 0;
186 }
187
188 /* Useful function which ensures that all worker functions terminate */
189 static void
quit_workers(struct rte_distributor * d,struct rte_mempool * p)190 quit_workers(struct rte_distributor *d, struct rte_mempool *p)
191 {
192 const unsigned int num_workers = rte_lcore_count() - 1;
193 unsigned int i;
194 struct rte_mbuf *bufs[RTE_MAX_LCORE];
195
196 rte_mempool_get_bulk(p, (void *)bufs, num_workers);
197
198 quit = 1;
199 for (i = 0; i < num_workers; i++) {
200 bufs[i]->hash.usr = i << 1;
201 rte_distributor_process(d, &bufs[i], 1);
202 }
203
204 rte_mempool_put_bulk(p, (void *)bufs, num_workers);
205
206 rte_distributor_process(d, NULL, 0);
207 rte_distributor_flush(d);
208 rte_eal_mp_wait_lcore();
209 quit = 0;
210 worker_idx = 0;
211 }
212
213 static int
test_distributor_perf(void)214 test_distributor_perf(void)
215 {
216 static struct rte_distributor *ds;
217 static struct rte_distributor *db;
218 static struct rte_mempool *p;
219
220 if (rte_lcore_count() < 2) {
221 printf("Not enough cores for distributor_perf_autotest, expecting at least 2\n");
222 return TEST_SKIPPED;
223 }
224
225 /* first time how long it takes to round-trip a cache line */
226 time_cache_line_switch();
227
228 if (ds == NULL) {
229 ds = rte_distributor_create("Test_perf", rte_socket_id(),
230 rte_lcore_count() - 1,
231 RTE_DIST_ALG_SINGLE);
232 if (ds == NULL) {
233 printf("Error creating distributor\n");
234 return -1;
235 }
236 } else {
237 rte_distributor_clear_returns(ds);
238 }
239
240 if (db == NULL) {
241 db = rte_distributor_create("Test_burst", rte_socket_id(),
242 rte_lcore_count() - 1,
243 RTE_DIST_ALG_BURST);
244 if (db == NULL) {
245 printf("Error creating burst distributor\n");
246 return -1;
247 }
248 } else {
249 rte_distributor_clear_returns(db);
250 }
251
252 const unsigned nb_bufs = (511 * rte_lcore_count()) < BIG_BATCH ?
253 (BIG_BATCH * 2) - 1 : (511 * rte_lcore_count());
254 if (p == NULL) {
255 p = rte_pktmbuf_pool_create("DPT_MBUF_POOL", nb_bufs, BURST,
256 0, RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
257 if (p == NULL) {
258 printf("Error creating mempool\n");
259 return -1;
260 }
261 }
262
263 printf("=== Performance test of distributor (single mode) ===\n");
264 rte_eal_mp_remote_launch(handle_work, ds, SKIP_MAIN);
265 if (perf_test(ds, p) < 0)
266 return -1;
267 quit_workers(ds, p);
268
269 printf("=== Performance test of distributor (burst mode) ===\n");
270 rte_eal_mp_remote_launch(handle_work, db, SKIP_MAIN);
271 if (perf_test(db, p) < 0)
272 return -1;
273 quit_workers(db, p);
274
275 return 0;
276 }
277
278 #endif /* !RTE_EXEC_ENV_WINDOWS */
279
280 REGISTER_PERF_TEST(distributor_perf_autotest, test_distributor_perf);
281