xref: /dpdk/app/test/test_lpm_perf.c (revision 68a03efeed657e6e05f281479b33b51102797e15)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2014 Intel Corporation
3  * Copyright(c) 2020 Arm Limited
4  */
5 
6 #include <stdio.h>
7 #include <stdint.h>
8 #include <stdlib.h>
9 #include <math.h>
10 
11 #include <rte_cycles.h>
12 #include <rte_random.h>
13 #include <rte_branch_prediction.h>
14 #include <rte_malloc.h>
15 #include <rte_ip.h>
16 #include <rte_lpm.h>
17 
18 #include "test.h"
19 #include "test_xmmt_ops.h"
20 
21 struct rte_lpm *lpm;
22 static struct rte_rcu_qsbr *rv;
23 static volatile uint8_t writer_done;
24 static volatile uint32_t thr_id;
25 static uint64_t gwrite_cycles;
26 static uint32_t num_writers;
27 /* LPM APIs are not thread safe, use mutex to provide thread safety */
28 static pthread_mutex_t lpm_mutex = PTHREAD_MUTEX_INITIALIZER;
29 
30 /* Report quiescent state interval every 1024 lookups. Larger critical
31  * sections in reader will result in writer polling multiple times.
32  */
33 #define QSBR_REPORTING_INTERVAL 1024
34 
35 #define TEST_LPM_ASSERT(cond) do {                                            \
36 	if (!(cond)) {                                                        \
37 		printf("Error at line %d: \n", __LINE__);                     \
38 		return -1;                                                    \
39 	}                                                                     \
40 } while(0)
41 
42 #define ITERATIONS (1 << 10)
43 #define RCU_ITERATIONS 10
44 #define BATCH_SIZE (1 << 12)
45 #define BULK_SIZE 32
46 
47 #define MAX_RULE_NUM (1200000)
48 
49 struct route_rule {
50 	uint32_t ip;
51 	uint8_t depth;
52 };
53 
54 static struct route_rule large_route_table[MAX_RULE_NUM];
55 /* Route table for routes with depth > 24 */
56 struct route_rule large_ldepth_route_table[MAX_RULE_NUM];
57 
58 static uint32_t num_route_entries;
59 static uint32_t num_ldepth_route_entries;
60 #define NUM_ROUTE_ENTRIES num_route_entries
61 #define NUM_LDEPTH_ROUTE_ENTRIES num_ldepth_route_entries
62 
63 #define TOTAL_WRITES (RCU_ITERATIONS * NUM_LDEPTH_ROUTE_ENTRIES)
64 
65 enum {
66 	IP_CLASS_A,
67 	IP_CLASS_B,
68 	IP_CLASS_C
69 };
70 
71 /* struct route_rule_count defines the total number of rules in following a/b/c
72  * each item in a[]/b[]/c[] is the number of common IP address class A/B/C, not
73  * including the ones for private local network.
74  */
75 struct route_rule_count {
76 	uint32_t a[RTE_LPM_MAX_DEPTH];
77 	uint32_t b[RTE_LPM_MAX_DEPTH];
78 	uint32_t c[RTE_LPM_MAX_DEPTH];
79 };
80 
81 /* All following numbers of each depth of each common IP class are just
82  * got from previous large constant table in app/test/test_lpm_routes.h .
83  * In order to match similar performance, they keep same depth and IP
84  * address coverage as previous constant table. These numbers don't
85  * include any private local IP address. As previous large const rule
86  * table was just dumped from a real router, there are no any IP address
87  * in class C or D.
88  */
89 static struct route_rule_count rule_count = {
90 	.a = { /* IP class A in which the most significant bit is 0 */
91 		    0, /* depth =  1 */
92 		    0, /* depth =  2 */
93 		    1, /* depth =  3 */
94 		    0, /* depth =  4 */
95 		    2, /* depth =  5 */
96 		    1, /* depth =  6 */
97 		    3, /* depth =  7 */
98 		  185, /* depth =  8 */
99 		   26, /* depth =  9 */
100 		   16, /* depth = 10 */
101 		   39, /* depth = 11 */
102 		  144, /* depth = 12 */
103 		  233, /* depth = 13 */
104 		  528, /* depth = 14 */
105 		  866, /* depth = 15 */
106 		 3856, /* depth = 16 */
107 		 3268, /* depth = 17 */
108 		 5662, /* depth = 18 */
109 		17301, /* depth = 19 */
110 		22226, /* depth = 20 */
111 		11147, /* depth = 21 */
112 		16746, /* depth = 22 */
113 		17120, /* depth = 23 */
114 		77578, /* depth = 24 */
115 		  401, /* depth = 25 */
116 		  656, /* depth = 26 */
117 		 1107, /* depth = 27 */
118 		 1121, /* depth = 28 */
119 		 2316, /* depth = 29 */
120 		  717, /* depth = 30 */
121 		   10, /* depth = 31 */
122 		   66  /* depth = 32 */
123 	},
124 	.b = { /* IP class A in which the most 2 significant bits are 10 */
125 		    0, /* depth =  1 */
126 		    0, /* depth =  2 */
127 		    0, /* depth =  3 */
128 		    0, /* depth =  4 */
129 		    1, /* depth =  5 */
130 		    1, /* depth =  6 */
131 		    1, /* depth =  7 */
132 		    3, /* depth =  8 */
133 		    3, /* depth =  9 */
134 		   30, /* depth = 10 */
135 		   25, /* depth = 11 */
136 		  168, /* depth = 12 */
137 		  305, /* depth = 13 */
138 		  569, /* depth = 14 */
139 		 1129, /* depth = 15 */
140 		50800, /* depth = 16 */
141 		 1645, /* depth = 17 */
142 		 1820, /* depth = 18 */
143 		 3506, /* depth = 19 */
144 		 3258, /* depth = 20 */
145 		 3424, /* depth = 21 */
146 		 4971, /* depth = 22 */
147 		 6885, /* depth = 23 */
148 		39771, /* depth = 24 */
149 		  424, /* depth = 25 */
150 		  170, /* depth = 26 */
151 		  433, /* depth = 27 */
152 		   92, /* depth = 28 */
153 		  366, /* depth = 29 */
154 		  377, /* depth = 30 */
155 		    2, /* depth = 31 */
156 		  200  /* depth = 32 */
157 	},
158 	.c = { /* IP class A in which the most 3 significant bits are 110 */
159 		     0, /* depth =  1 */
160 		     0, /* depth =  2 */
161 		     0, /* depth =  3 */
162 		     0, /* depth =  4 */
163 		     0, /* depth =  5 */
164 		     0, /* depth =  6 */
165 		     0, /* depth =  7 */
166 		    12, /* depth =  8 */
167 		     8, /* depth =  9 */
168 		     9, /* depth = 10 */
169 		    33, /* depth = 11 */
170 		    69, /* depth = 12 */
171 		   237, /* depth = 13 */
172 		  1007, /* depth = 14 */
173 		  1717, /* depth = 15 */
174 		 14663, /* depth = 16 */
175 		  8070, /* depth = 17 */
176 		 16185, /* depth = 18 */
177 		 48261, /* depth = 19 */
178 		 36870, /* depth = 20 */
179 		 33960, /* depth = 21 */
180 		 50638, /* depth = 22 */
181 		 61422, /* depth = 23 */
182 		466549, /* depth = 24 */
183 		  1829, /* depth = 25 */
184 		  4824, /* depth = 26 */
185 		  4927, /* depth = 27 */
186 		  5914, /* depth = 28 */
187 		 10254, /* depth = 29 */
188 		  4905, /* depth = 30 */
189 		     1, /* depth = 31 */
190 		   716  /* depth = 32 */
191 	}
192 };
193 
194 static void generate_random_rule_prefix(uint32_t ip_class, uint8_t depth)
195 {
196 /* IP address class A, the most significant bit is 0 */
197 #define IP_HEAD_MASK_A			0x00000000
198 #define IP_HEAD_BIT_NUM_A		1
199 
200 /* IP address class B, the most significant 2 bits are 10 */
201 #define IP_HEAD_MASK_B			0x80000000
202 #define IP_HEAD_BIT_NUM_B		2
203 
204 /* IP address class C, the most significant 3 bits are 110 */
205 #define IP_HEAD_MASK_C			0xC0000000
206 #define IP_HEAD_BIT_NUM_C		3
207 
208 	uint32_t class_depth;
209 	uint32_t range;
210 	uint32_t mask;
211 	uint32_t step;
212 	uint32_t start;
213 	uint32_t fixed_bit_num;
214 	uint32_t ip_head_mask;
215 	uint32_t rule_num;
216 	uint32_t k;
217 	struct route_rule *ptr_rule, *ptr_ldepth_rule;
218 
219 	if (ip_class == IP_CLASS_A) {        /* IP Address class A */
220 		fixed_bit_num = IP_HEAD_BIT_NUM_A;
221 		ip_head_mask = IP_HEAD_MASK_A;
222 		rule_num = rule_count.a[depth - 1];
223 	} else if (ip_class == IP_CLASS_B) { /* IP Address class B */
224 		fixed_bit_num = IP_HEAD_BIT_NUM_B;
225 		ip_head_mask = IP_HEAD_MASK_B;
226 		rule_num = rule_count.b[depth - 1];
227 	} else {                             /* IP Address class C */
228 		fixed_bit_num = IP_HEAD_BIT_NUM_C;
229 		ip_head_mask = IP_HEAD_MASK_C;
230 		rule_num = rule_count.c[depth - 1];
231 	}
232 
233 	if (rule_num == 0)
234 		return;
235 
236 	/* the number of rest bits which don't include the most significant
237 	 * fixed bits for this IP address class
238 	 */
239 	class_depth = depth - fixed_bit_num;
240 
241 	/* range is the maximum number of rules for this depth and
242 	 * this IP address class
243 	 */
244 	range = 1 << class_depth;
245 
246 	/* only mask the most depth significant generated bits
247 	 * except fixed bits for IP address class
248 	 */
249 	mask = range - 1;
250 
251 	/* Widen coverage of IP address in generated rules */
252 	if (range <= rule_num)
253 		step = 1;
254 	else
255 		step = round((double)range / rule_num);
256 
257 	/* Only generate rest bits except the most significant
258 	 * fixed bits for IP address class
259 	 */
260 	start = lrand48() & mask;
261 	ptr_rule = &large_route_table[num_route_entries];
262 	ptr_ldepth_rule = &large_ldepth_route_table[num_ldepth_route_entries];
263 	for (k = 0; k < rule_num; k++) {
264 		ptr_rule->ip = (start << (RTE_LPM_MAX_DEPTH - depth))
265 			| ip_head_mask;
266 		ptr_rule->depth = depth;
267 		/* If the depth of the route is more than 24, store it
268 		 * in another table as well.
269 		 */
270 		if (depth > 24) {
271 			ptr_ldepth_rule->ip = ptr_rule->ip;
272 			ptr_ldepth_rule->depth = ptr_rule->depth;
273 			ptr_ldepth_rule++;
274 			num_ldepth_route_entries++;
275 		}
276 		ptr_rule++;
277 		start = (start + step) & mask;
278 	}
279 	num_route_entries += rule_num;
280 }
281 
282 static void insert_rule_in_random_pos(uint32_t ip, uint8_t depth)
283 {
284 	uint32_t pos;
285 	int try_count = 0;
286 	struct route_rule tmp;
287 
288 	do {
289 		pos = lrand48();
290 		try_count++;
291 	} while ((try_count < 10) && (pos > num_route_entries));
292 
293 	if ((pos > num_route_entries) || (pos >= MAX_RULE_NUM))
294 		pos = num_route_entries >> 1;
295 
296 	tmp = large_route_table[pos];
297 	large_route_table[pos].ip = ip;
298 	large_route_table[pos].depth = depth;
299 	if (num_route_entries < MAX_RULE_NUM)
300 		large_route_table[num_route_entries++] = tmp;
301 }
302 
303 static void generate_large_route_rule_table(void)
304 {
305 	uint32_t ip_class;
306 	uint8_t  depth;
307 
308 	num_route_entries = 0;
309 	num_ldepth_route_entries = 0;
310 	memset(large_route_table, 0, sizeof(large_route_table));
311 
312 	for (ip_class = IP_CLASS_A; ip_class <= IP_CLASS_C; ip_class++) {
313 		for (depth = 1; depth <= RTE_LPM_MAX_DEPTH; depth++) {
314 			generate_random_rule_prefix(ip_class, depth);
315 		}
316 	}
317 
318 	/* Add following rules to keep same as previous large constant table,
319 	 * they are 4 rules with private local IP address and 1 all-zeros prefix
320 	 * with depth = 8.
321 	 */
322 	insert_rule_in_random_pos(RTE_IPV4(0, 0, 0, 0), 8);
323 	insert_rule_in_random_pos(RTE_IPV4(10, 2, 23, 147), 32);
324 	insert_rule_in_random_pos(RTE_IPV4(192, 168, 100, 10), 24);
325 	insert_rule_in_random_pos(RTE_IPV4(192, 168, 25, 100), 24);
326 	insert_rule_in_random_pos(RTE_IPV4(192, 168, 129, 124), 32);
327 }
328 
329 static void
330 print_route_distribution(const struct route_rule *table, uint32_t n)
331 {
332 	unsigned i, j;
333 
334 	printf("Route distribution per prefix width: \n");
335 	printf("DEPTH    QUANTITY (PERCENT)\n");
336 	printf("--------------------------- \n");
337 
338 	/* Count depths. */
339 	for (i = 1; i <= 32; i++) {
340 		unsigned depth_counter = 0;
341 		double percent_hits;
342 
343 		for (j = 0; j < n; j++)
344 			if (table[j].depth == (uint8_t) i)
345 				depth_counter++;
346 
347 		percent_hits = ((double)depth_counter)/((double)n) * 100;
348 		printf("%.2u%15u (%.2f)\n", i, depth_counter, percent_hits);
349 	}
350 	printf("\n");
351 }
352 
353 /* Check condition and return an error if true. */
354 static uint16_t enabled_core_ids[RTE_MAX_LCORE];
355 static unsigned int num_cores;
356 
357 /* Simple way to allocate thread ids in 0 to RTE_MAX_LCORE space */
358 static inline uint32_t
359 alloc_thread_id(void)
360 {
361 	uint32_t tmp_thr_id;
362 
363 	tmp_thr_id = __atomic_fetch_add(&thr_id, 1, __ATOMIC_RELAXED);
364 	if (tmp_thr_id >= RTE_MAX_LCORE)
365 		printf("Invalid thread id %u\n", tmp_thr_id);
366 
367 	return tmp_thr_id;
368 }
369 
370 /*
371  * Reader thread using rte_lpm data structure without RCU.
372  */
373 static int
374 test_lpm_reader(void *arg)
375 {
376 	int i;
377 	uint32_t ip_batch[QSBR_REPORTING_INTERVAL];
378 	uint32_t next_hop_return = 0;
379 
380 	RTE_SET_USED(arg);
381 	do {
382 		for (i = 0; i < QSBR_REPORTING_INTERVAL; i++)
383 			ip_batch[i] = rte_rand();
384 
385 		for (i = 0; i < QSBR_REPORTING_INTERVAL; i++)
386 			rte_lpm_lookup(lpm, ip_batch[i], &next_hop_return);
387 
388 	} while (!writer_done);
389 
390 	return 0;
391 }
392 
393 /*
394  * Reader thread using rte_lpm data structure with RCU.
395  */
396 static int
397 test_lpm_rcu_qsbr_reader(void *arg)
398 {
399 	int i;
400 	uint32_t thread_id = alloc_thread_id();
401 	uint32_t ip_batch[QSBR_REPORTING_INTERVAL];
402 	uint32_t next_hop_return = 0;
403 
404 	RTE_SET_USED(arg);
405 	/* Register this thread to report quiescent state */
406 	rte_rcu_qsbr_thread_register(rv, thread_id);
407 	rte_rcu_qsbr_thread_online(rv, thread_id);
408 
409 	do {
410 		for (i = 0; i < QSBR_REPORTING_INTERVAL; i++)
411 			ip_batch[i] = rte_rand();
412 
413 		for (i = 0; i < QSBR_REPORTING_INTERVAL; i++)
414 			rte_lpm_lookup(lpm, ip_batch[i], &next_hop_return);
415 
416 		/* Update quiescent state */
417 		rte_rcu_qsbr_quiescent(rv, thread_id);
418 	} while (!writer_done);
419 
420 	rte_rcu_qsbr_thread_offline(rv, thread_id);
421 	rte_rcu_qsbr_thread_unregister(rv, thread_id);
422 
423 	return 0;
424 }
425 
426 /*
427  * Writer thread using rte_lpm data structure with RCU.
428  */
429 static int
430 test_lpm_rcu_qsbr_writer(void *arg)
431 {
432 	unsigned int i, j, si, ei;
433 	uint64_t begin, total_cycles;
434 	uint32_t next_hop_add = 0xAA;
435 	uint8_t pos_core = (uint8_t)((uintptr_t)arg);
436 
437 	si = (pos_core * NUM_LDEPTH_ROUTE_ENTRIES) / num_writers;
438 	ei = ((pos_core + 1) * NUM_LDEPTH_ROUTE_ENTRIES) / num_writers;
439 
440 	/* Measure add/delete. */
441 	begin = rte_rdtsc_precise();
442 	for (i = 0; i < RCU_ITERATIONS; i++) {
443 		/* Add all the entries */
444 		for (j = si; j < ei; j++) {
445 			if (num_writers > 1)
446 				pthread_mutex_lock(&lpm_mutex);
447 			if (rte_lpm_add(lpm, large_ldepth_route_table[j].ip,
448 					large_ldepth_route_table[j].depth,
449 					next_hop_add) != 0) {
450 				printf("Failed to add iteration %d, route# %d\n",
451 					i, j);
452 				goto error;
453 			}
454 			if (num_writers > 1)
455 				pthread_mutex_unlock(&lpm_mutex);
456 		}
457 
458 		/* Delete all the entries */
459 		for (j = si; j < ei; j++) {
460 			if (num_writers > 1)
461 				pthread_mutex_lock(&lpm_mutex);
462 			if (rte_lpm_delete(lpm, large_ldepth_route_table[j].ip,
463 				large_ldepth_route_table[j].depth) != 0) {
464 				printf("Failed to delete iteration %d, route# %d\n",
465 					i, j);
466 				goto error;
467 			}
468 			if (num_writers > 1)
469 				pthread_mutex_unlock(&lpm_mutex);
470 		}
471 	}
472 
473 	total_cycles = rte_rdtsc_precise() - begin;
474 
475 	__atomic_fetch_add(&gwrite_cycles, total_cycles, __ATOMIC_RELAXED);
476 
477 	return 0;
478 
479 error:
480 	if (num_writers > 1)
481 		pthread_mutex_unlock(&lpm_mutex);
482 	return -1;
483 }
484 
485 /*
486  * Functional test:
487  * 1/2 writers, rest are readers
488  */
489 static int
490 test_lpm_rcu_perf_multi_writer(uint8_t use_rcu)
491 {
492 	struct rte_lpm_config config;
493 	size_t sz;
494 	unsigned int i, j;
495 	uint16_t core_id;
496 	struct rte_lpm_rcu_config rcu_cfg = {0};
497 	int (*reader_f)(void *arg) = NULL;
498 
499 	if (rte_lcore_count() < 3) {
500 		printf("Not enough cores for lpm_rcu_perf_autotest, expecting at least 3\n");
501 		return TEST_SKIPPED;
502 	}
503 
504 	num_cores = 0;
505 	RTE_LCORE_FOREACH_WORKER(core_id) {
506 		enabled_core_ids[num_cores] = core_id;
507 		num_cores++;
508 	}
509 
510 	for (j = 1; j < 3; j++) {
511 		if (use_rcu)
512 			printf("\nPerf test: %d writer(s), %d reader(s),"
513 			       " RCU integration enabled\n", j, num_cores - j);
514 		else
515 			printf("\nPerf test: %d writer(s), %d reader(s),"
516 			       " RCU integration disabled\n", j, num_cores - j);
517 
518 		num_writers = j;
519 
520 		/* Create LPM table */
521 		config.max_rules = NUM_LDEPTH_ROUTE_ENTRIES;
522 		config.number_tbl8s = NUM_LDEPTH_ROUTE_ENTRIES;
523 		config.flags = 0;
524 		lpm = rte_lpm_create(__func__, SOCKET_ID_ANY, &config);
525 		TEST_LPM_ASSERT(lpm != NULL);
526 
527 		/* Init RCU variable */
528 		if (use_rcu) {
529 			sz = rte_rcu_qsbr_get_memsize(num_cores);
530 			rv = (struct rte_rcu_qsbr *)rte_zmalloc("rcu0", sz,
531 							RTE_CACHE_LINE_SIZE);
532 			rte_rcu_qsbr_init(rv, num_cores);
533 
534 			rcu_cfg.v = rv;
535 			/* Assign the RCU variable to LPM */
536 			if (rte_lpm_rcu_qsbr_add(lpm, &rcu_cfg) != 0) {
537 				printf("RCU variable assignment failed\n");
538 				goto error;
539 			}
540 
541 			reader_f = test_lpm_rcu_qsbr_reader;
542 		} else
543 			reader_f = test_lpm_reader;
544 
545 		writer_done = 0;
546 		__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
547 
548 		__atomic_store_n(&thr_id, 0, __ATOMIC_SEQ_CST);
549 
550 		/* Launch reader threads */
551 		for (i = j; i < num_cores; i++)
552 			rte_eal_remote_launch(reader_f, NULL,
553 						enabled_core_ids[i]);
554 
555 		/* Launch writer threads */
556 		for (i = 0; i < j; i++)
557 			rte_eal_remote_launch(test_lpm_rcu_qsbr_writer,
558 						(void *)(uintptr_t)i,
559 						enabled_core_ids[i]);
560 
561 		/* Wait for writer threads */
562 		for (i = 0; i < j; i++)
563 			if (rte_eal_wait_lcore(enabled_core_ids[i]) < 0)
564 				goto error;
565 
566 		printf("Total LPM Adds: %d\n", TOTAL_WRITES);
567 		printf("Total LPM Deletes: %d\n", TOTAL_WRITES);
568 		printf("Average LPM Add/Del: %"PRIu64" cycles\n",
569 			__atomic_load_n(&gwrite_cycles, __ATOMIC_RELAXED)
570 			/ TOTAL_WRITES);
571 
572 		writer_done = 1;
573 		/* Wait until all readers have exited */
574 		for (i = j; i < num_cores; i++)
575 			rte_eal_wait_lcore(enabled_core_ids[i]);
576 
577 		rte_lpm_free(lpm);
578 		rte_free(rv);
579 		lpm = NULL;
580 		rv = NULL;
581 	}
582 
583 	return 0;
584 
585 error:
586 	writer_done = 1;
587 	/* Wait until all readers have exited */
588 	rte_eal_mp_wait_lcore();
589 
590 	rte_lpm_free(lpm);
591 	rte_free(rv);
592 
593 	return -1;
594 }
595 
596 static int
597 test_lpm_perf(void)
598 {
599 	struct rte_lpm_config config;
600 
601 	config.max_rules = 2000000;
602 	config.number_tbl8s = 2048;
603 	config.flags = 0;
604 	uint64_t begin, total_time, lpm_used_entries = 0;
605 	unsigned i, j;
606 	uint32_t next_hop_add = 0xAA, next_hop_return = 0;
607 	int status = 0;
608 	uint64_t cache_line_counter = 0;
609 	int64_t count = 0;
610 
611 	rte_srand(rte_rdtsc());
612 
613 	generate_large_route_rule_table();
614 
615 	printf("No. routes = %u\n", (unsigned) NUM_ROUTE_ENTRIES);
616 
617 	print_route_distribution(large_route_table, (uint32_t) NUM_ROUTE_ENTRIES);
618 
619 	lpm = rte_lpm_create(__func__, SOCKET_ID_ANY, &config);
620 	TEST_LPM_ASSERT(lpm != NULL);
621 
622 	/* Measure add. */
623 	begin = rte_rdtsc();
624 
625 	for (i = 0; i < NUM_ROUTE_ENTRIES; i++) {
626 		if (rte_lpm_add(lpm, large_route_table[i].ip,
627 				large_route_table[i].depth, next_hop_add) == 0)
628 			status++;
629 	}
630 	/* End Timer. */
631 	total_time = rte_rdtsc() - begin;
632 
633 	printf("Unique added entries = %d\n", status);
634 	/* Obtain add statistics. */
635 	for (i = 0; i < RTE_LPM_TBL24_NUM_ENTRIES; i++) {
636 		if (lpm->tbl24[i].valid)
637 			lpm_used_entries++;
638 
639 		if (i % 32 == 0) {
640 			if ((uint64_t)count < lpm_used_entries) {
641 				cache_line_counter++;
642 				count = lpm_used_entries;
643 			}
644 		}
645 	}
646 
647 	printf("Used table 24 entries = %u (%g%%)\n",
648 			(unsigned) lpm_used_entries,
649 			(lpm_used_entries * 100.0) / RTE_LPM_TBL24_NUM_ENTRIES);
650 	printf("64 byte Cache entries used = %u (%u bytes)\n",
651 			(unsigned) cache_line_counter, (unsigned) cache_line_counter * 64);
652 
653 	printf("Average LPM Add: %g cycles\n",
654 			(double)total_time / NUM_ROUTE_ENTRIES);
655 
656 	/* Measure single Lookup */
657 	total_time = 0;
658 	count = 0;
659 
660 	for (i = 0; i < ITERATIONS; i++) {
661 		static uint32_t ip_batch[BATCH_SIZE];
662 
663 		for (j = 0; j < BATCH_SIZE; j++)
664 			ip_batch[j] = rte_rand();
665 
666 		/* Lookup per batch */
667 		begin = rte_rdtsc();
668 
669 		for (j = 0; j < BATCH_SIZE; j++) {
670 			if (rte_lpm_lookup(lpm, ip_batch[j], &next_hop_return) != 0)
671 				count++;
672 		}
673 
674 		total_time += rte_rdtsc() - begin;
675 
676 	}
677 	printf("Average LPM Lookup: %.1f cycles (fails = %.1f%%)\n",
678 			(double)total_time / ((double)ITERATIONS * BATCH_SIZE),
679 			(count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
680 
681 	/* Measure bulk Lookup */
682 	total_time = 0;
683 	count = 0;
684 	for (i = 0; i < ITERATIONS; i++) {
685 		static uint32_t ip_batch[BATCH_SIZE];
686 		uint32_t next_hops[BULK_SIZE];
687 
688 		/* Create array of random IP addresses */
689 		for (j = 0; j < BATCH_SIZE; j++)
690 			ip_batch[j] = rte_rand();
691 
692 		/* Lookup per batch */
693 		begin = rte_rdtsc();
694 		for (j = 0; j < BATCH_SIZE; j += BULK_SIZE) {
695 			unsigned k;
696 			rte_lpm_lookup_bulk(lpm, &ip_batch[j], next_hops, BULK_SIZE);
697 			for (k = 0; k < BULK_SIZE; k++)
698 				if (unlikely(!(next_hops[k] & RTE_LPM_LOOKUP_SUCCESS)))
699 					count++;
700 		}
701 
702 		total_time += rte_rdtsc() - begin;
703 	}
704 	printf("BULK LPM Lookup: %.1f cycles (fails = %.1f%%)\n",
705 			(double)total_time / ((double)ITERATIONS * BATCH_SIZE),
706 			(count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
707 
708 	/* Measure LookupX4 */
709 	total_time = 0;
710 	count = 0;
711 	for (i = 0; i < ITERATIONS; i++) {
712 		static uint32_t ip_batch[BATCH_SIZE];
713 		uint32_t next_hops[4];
714 
715 		/* Create array of random IP addresses */
716 		for (j = 0; j < BATCH_SIZE; j++)
717 			ip_batch[j] = rte_rand();
718 
719 		/* Lookup per batch */
720 		begin = rte_rdtsc();
721 		for (j = 0; j < BATCH_SIZE; j += RTE_DIM(next_hops)) {
722 			unsigned k;
723 			xmm_t ipx4;
724 
725 			ipx4 = vect_loadu_sil128((xmm_t *)(ip_batch + j));
726 			ipx4 = *(xmm_t *)(ip_batch + j);
727 			rte_lpm_lookupx4(lpm, ipx4, next_hops, UINT32_MAX);
728 			for (k = 0; k < RTE_DIM(next_hops); k++)
729 				if (unlikely(next_hops[k] == UINT32_MAX))
730 					count++;
731 		}
732 
733 		total_time += rte_rdtsc() - begin;
734 	}
735 	printf("LPM LookupX4: %.1f cycles (fails = %.1f%%)\n",
736 			(double)total_time / ((double)ITERATIONS * BATCH_SIZE),
737 			(count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
738 
739 	/* Measure Delete */
740 	status = 0;
741 	begin = rte_rdtsc();
742 
743 	for (i = 0; i < NUM_ROUTE_ENTRIES; i++) {
744 		/* rte_lpm_delete(lpm, ip, depth) */
745 		status += rte_lpm_delete(lpm, large_route_table[i].ip,
746 				large_route_table[i].depth);
747 	}
748 
749 	total_time = rte_rdtsc() - begin;
750 
751 	printf("Average LPM Delete: %g cycles\n",
752 			(double)total_time / NUM_ROUTE_ENTRIES);
753 
754 	rte_lpm_delete_all(lpm);
755 	rte_lpm_free(lpm);
756 
757 	if (test_lpm_rcu_perf_multi_writer(0) < 0)
758 		return -1;
759 
760 	if (test_lpm_rcu_perf_multi_writer(1) < 0)
761 		return -1;
762 
763 	return 0;
764 }
765 
766 REGISTER_TEST_COMMAND(lpm_perf_autotest, test_lpm_perf);
767