xref: /dpdk/app/test/test_hash_readwrite.c (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2018 Intel Corporation
3  */
4 
5 #include <inttypes.h>
6 #include <locale.h>
7 
8 #include <rte_cycles.h>
9 #include <rte_hash.h>
10 #include <rte_hash_crc.h>
11 #include <rte_jhash.h>
12 #include <rte_launch.h>
13 #include <rte_malloc.h>
14 #include <rte_random.h>
15 #include <rte_spinlock.h>
16 
17 #include "test.h"
18 
19 #define RTE_RWTEST_FAIL 0
20 
21 #define TOTAL_ENTRY (5*1024*1024)
22 #define TOTAL_INSERT (4.5*1024*1024)
23 #define TOTAL_INSERT_EXT (5*1024*1024)
24 
25 #define NUM_TEST 3
26 unsigned int core_cnt[NUM_TEST] = {2, 4, 8};
27 
28 unsigned int worker_core_ids[RTE_MAX_LCORE];
29 struct perf {
30 	uint32_t single_read;
31 	uint32_t single_write;
32 	uint32_t read_only[NUM_TEST];
33 	uint32_t write_only[NUM_TEST];
34 	uint32_t read_write_r[NUM_TEST];
35 	uint32_t read_write_w[NUM_TEST];
36 };
37 
38 static struct perf htm_results, non_htm_results;
39 
40 struct {
41 	uint32_t *keys;
42 	uint8_t *found;
43 	uint32_t num_insert;
44 	uint32_t rounded_tot_insert;
45 	struct rte_hash *h;
46 } tbl_rw_test_param;
47 
48 static uint64_t gcycles;
49 static uint64_t ginsertions;
50 
51 static uint64_t gread_cycles;
52 static uint64_t gwrite_cycles;
53 
54 static uint64_t greads;
55 static uint64_t gwrites;
56 
57 static int
58 test_hash_readwrite_worker(__rte_unused void *arg)
59 {
60 	uint64_t i, offset;
61 	uint32_t lcore_id = rte_lcore_id();
62 	uint64_t begin, cycles;
63 	int *ret;
64 
65 	ret = rte_malloc(NULL, sizeof(int) *
66 				tbl_rw_test_param.num_insert, 0);
67 	for (i = 0; i < rte_lcore_count(); i++) {
68 		if (worker_core_ids[i] == lcore_id)
69 			break;
70 	}
71 	offset = tbl_rw_test_param.num_insert * i;
72 
73 	printf("Core #%d inserting and reading %d: %'"PRId64" - %'"PRId64"\n",
74 	       lcore_id, tbl_rw_test_param.num_insert,
75 	       offset, offset + tbl_rw_test_param.num_insert - 1);
76 
77 	begin = rte_rdtsc_precise();
78 
79 	for (i = offset; i < offset + tbl_rw_test_param.num_insert; i++) {
80 
81 		if (rte_hash_lookup(tbl_rw_test_param.h,
82 				tbl_rw_test_param.keys + i) > 0)
83 			break;
84 
85 		ret[i - offset] = rte_hash_add_key(tbl_rw_test_param.h,
86 				     tbl_rw_test_param.keys + i);
87 		if (ret[i - offset] < 0)
88 			break;
89 
90 		/* lookup a random key */
91 		uint32_t rand = rte_rand() % (i + 1 - offset);
92 
93 		if (rte_hash_lookup(tbl_rw_test_param.h,
94 				tbl_rw_test_param.keys + rand) != ret[rand])
95 			break;
96 
97 
98 		if (rte_hash_del_key(tbl_rw_test_param.h,
99 				tbl_rw_test_param.keys + rand) != ret[rand])
100 			break;
101 
102 		ret[rand] = rte_hash_add_key(tbl_rw_test_param.h,
103 					tbl_rw_test_param.keys + rand);
104 		if (ret[rand] < 0)
105 			break;
106 
107 		if (rte_hash_lookup(tbl_rw_test_param.h,
108 			tbl_rw_test_param.keys + rand) != ret[rand])
109 			break;
110 	}
111 
112 	cycles = rte_rdtsc_precise() - begin;
113 	__atomic_fetch_add(&gcycles, cycles, __ATOMIC_RELAXED);
114 	__atomic_fetch_add(&ginsertions, i - offset, __ATOMIC_RELAXED);
115 
116 	for (; i < offset + tbl_rw_test_param.num_insert; i++)
117 		tbl_rw_test_param.keys[i] = RTE_RWTEST_FAIL;
118 
119 	rte_free(ret);
120 	return 0;
121 }
122 
123 static int
124 init_params(int use_ext, int use_htm, int rw_lf, int use_jhash)
125 {
126 	unsigned int i;
127 
128 	uint32_t *keys = NULL;
129 	uint8_t *found = NULL;
130 	struct rte_hash *handle;
131 
132 	struct rte_hash_parameters hash_params = {
133 		.entries = TOTAL_ENTRY,
134 		.key_len = sizeof(uint32_t),
135 		.hash_func_init_val = 0,
136 		.socket_id = rte_socket_id(),
137 	};
138 	if (use_jhash)
139 		hash_params.hash_func = rte_jhash;
140 	else
141 		hash_params.hash_func = rte_hash_crc;
142 
143 	hash_params.extra_flag = RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD;
144 	if (use_htm)
145 		hash_params.extra_flag |=
146 			RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT;
147 	if (rw_lf)
148 		hash_params.extra_flag |=
149 			RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF;
150 	else
151 		hash_params.extra_flag |=
152 			RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY;
153 
154 	if (use_ext)
155 		hash_params.extra_flag |=
156 			RTE_HASH_EXTRA_FLAGS_EXT_TABLE;
157 	else
158 		hash_params.extra_flag &=
159 		       ~RTE_HASH_EXTRA_FLAGS_EXT_TABLE;
160 
161 	hash_params.name = "tests";
162 
163 	handle = rte_hash_create(&hash_params);
164 	if (handle == NULL) {
165 		printf("hash creation failed");
166 		return -1;
167 	}
168 
169 	tbl_rw_test_param.h = handle;
170 	keys = rte_malloc(NULL, sizeof(uint32_t) * TOTAL_ENTRY, 0);
171 
172 	if (keys == NULL) {
173 		printf("RTE_MALLOC failed\n");
174 		goto err;
175 	}
176 
177 	found = rte_zmalloc(NULL, sizeof(uint8_t) * TOTAL_ENTRY, 0);
178 	if (found == NULL) {
179 		printf("RTE_ZMALLOC failed\n");
180 		goto err;
181 	}
182 
183 	tbl_rw_test_param.keys = keys;
184 	tbl_rw_test_param.found = found;
185 
186 	for (i = 0; i < TOTAL_ENTRY; i++)
187 		keys[i] = i;
188 
189 	return 0;
190 
191 err:
192 	rte_free(keys);
193 	rte_hash_free(handle);
194 
195 	return -1;
196 }
197 
198 static int
199 test_hash_readwrite_functional(int use_htm, int use_rw_lf, int use_ext)
200 {
201 	unsigned int i;
202 	const void *next_key;
203 	void *next_data;
204 	uint32_t iter = 0;
205 
206 	uint32_t duplicated_keys = 0;
207 	uint32_t lost_keys = 0;
208 	int use_jhash = 1;
209 	int worker_cnt = rte_lcore_count() - 1;
210 	uint32_t tot_insert = 0;
211 
212 	__atomic_store_n(&gcycles, 0, __ATOMIC_RELAXED);
213 	__atomic_store_n(&ginsertions, 0, __ATOMIC_RELAXED);
214 
215 	if (init_params(use_ext, use_htm, use_rw_lf, use_jhash) != 0)
216 		goto err;
217 
218 	if (use_ext)
219 		tot_insert = TOTAL_INSERT_EXT;
220 	else
221 		tot_insert = TOTAL_INSERT;
222 
223 	tbl_rw_test_param.num_insert =
224 		tot_insert / worker_cnt;
225 
226 	tbl_rw_test_param.rounded_tot_insert =
227 		tbl_rw_test_param.num_insert * worker_cnt;
228 
229 	printf("\nHTM = %d, RW-LF = %d, EXT-Table = %d\n",
230 		use_htm, use_rw_lf, use_ext);
231 	printf("++++++++Start function tests:+++++++++\n");
232 
233 	/* Fire all threads. */
234 	rte_eal_mp_remote_launch(test_hash_readwrite_worker,
235 				 NULL, SKIP_MAIN);
236 	rte_eal_mp_wait_lcore();
237 
238 	while (rte_hash_iterate(tbl_rw_test_param.h, &next_key,
239 			&next_data, &iter) >= 0) {
240 		/* Search for the key in the list of keys added .*/
241 		i = *(const uint32_t *)next_key;
242 		tbl_rw_test_param.found[i]++;
243 	}
244 
245 	for (i = 0; i < tbl_rw_test_param.rounded_tot_insert; i++) {
246 		if (tbl_rw_test_param.keys[i] != RTE_RWTEST_FAIL) {
247 			if (tbl_rw_test_param.found[i] > 1) {
248 				duplicated_keys++;
249 				break;
250 			}
251 			if (tbl_rw_test_param.found[i] == 0) {
252 				lost_keys++;
253 				printf("key %d is lost\n", i);
254 				break;
255 			}
256 		}
257 	}
258 
259 	if (duplicated_keys > 0) {
260 		printf("%d key duplicated\n", duplicated_keys);
261 		goto err_free;
262 	}
263 
264 	if (lost_keys > 0) {
265 		printf("%d key lost\n", lost_keys);
266 		goto err_free;
267 	}
268 
269 	printf("No key corrupted during read-write test.\n");
270 
271 	unsigned long long int cycles_per_insertion =
272 		__atomic_load_n(&gcycles, __ATOMIC_RELAXED) /
273 		__atomic_load_n(&ginsertions, __ATOMIC_RELAXED);
274 
275 	printf("cycles per insertion and lookup: %llu\n", cycles_per_insertion);
276 
277 	rte_free(tbl_rw_test_param.found);
278 	rte_free(tbl_rw_test_param.keys);
279 	rte_hash_free(tbl_rw_test_param.h);
280 	printf("+++++++++Complete function tests+++++++++\n");
281 	return 0;
282 
283 err_free:
284 	rte_free(tbl_rw_test_param.found);
285 	rte_free(tbl_rw_test_param.keys);
286 	rte_hash_free(tbl_rw_test_param.h);
287 err:
288 	return -1;
289 }
290 
291 static int
292 test_rw_reader(void *arg)
293 {
294 	uint64_t i;
295 	uint64_t begin, cycles;
296 	uint64_t read_cnt = (uint64_t)((uintptr_t)arg);
297 
298 	begin = rte_rdtsc_precise();
299 	for (i = 0; i < read_cnt; i++) {
300 		void *data = arg;
301 		rte_hash_lookup_data(tbl_rw_test_param.h,
302 				tbl_rw_test_param.keys + i,
303 				&data);
304 		if (i != (uint64_t)(uintptr_t)data) {
305 			printf("lookup find wrong value %"PRIu64","
306 				"%"PRIu64"\n", i,
307 				(uint64_t)(uintptr_t)data);
308 			break;
309 		}
310 	}
311 
312 	cycles = rte_rdtsc_precise() - begin;
313 	__atomic_fetch_add(&gread_cycles, cycles, __ATOMIC_RELAXED);
314 	__atomic_fetch_add(&greads, i, __ATOMIC_RELAXED);
315 	return 0;
316 }
317 
318 static int
319 test_rw_writer(void *arg)
320 {
321 	uint64_t i;
322 	uint32_t lcore_id = rte_lcore_id();
323 	uint64_t begin, cycles;
324 	int ret;
325 	uint64_t start_coreid = (uint64_t)(uintptr_t)arg;
326 	uint64_t offset;
327 
328 	for (i = 0; i < rte_lcore_count(); i++) {
329 		if (worker_core_ids[i] == lcore_id)
330 			break;
331 	}
332 
333 	offset = TOTAL_INSERT / 2 + (i - (start_coreid)) *
334 				tbl_rw_test_param.num_insert;
335 	begin = rte_rdtsc_precise();
336 	for (i = offset; i < offset + tbl_rw_test_param.num_insert; i++) {
337 		ret = rte_hash_add_key_data(tbl_rw_test_param.h,
338 				tbl_rw_test_param.keys + i,
339 				(void *)((uintptr_t)i));
340 		if (ret < 0) {
341 			printf("writer failed %"PRIu64"\n", i);
342 			break;
343 		}
344 	}
345 
346 	cycles = rte_rdtsc_precise() - begin;
347 	__atomic_fetch_add(&gwrite_cycles, cycles, __ATOMIC_RELAXED);
348 	__atomic_fetch_add(&gwrites, tbl_rw_test_param.num_insert,
349 							__ATOMIC_RELAXED);
350 	return 0;
351 }
352 
353 static int
354 test_hash_readwrite_perf(struct perf *perf_results, int use_htm,
355 							int reader_faster)
356 {
357 	unsigned int n;
358 	int ret;
359 	int start_coreid;
360 	uint64_t i, read_cnt;
361 
362 	const void *next_key;
363 	void *next_data;
364 	uint32_t iter;
365 	int use_jhash = 0;
366 
367 	uint32_t duplicated_keys = 0;
368 	uint32_t lost_keys = 0;
369 
370 	uint64_t start = 0, end = 0;
371 
372 	__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
373 	__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
374 
375 	__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
376 	__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
377 
378 	if (init_params(0, use_htm, 0, use_jhash) != 0)
379 		goto err;
380 
381 	/*
382 	 * Do a readers finish faster or writers finish faster test.
383 	 * When readers finish faster, we timing the readers, and when writers
384 	 * finish faster, we timing the writers.
385 	 * Divided by 10 or 2 is just experimental values to vary the workload
386 	 * of readers.
387 	 */
388 	if (reader_faster) {
389 		printf("++++++Start perf test: reader++++++++\n");
390 		read_cnt = TOTAL_INSERT / 10;
391 	} else {
392 		printf("++++++Start perf test: writer++++++++\n");
393 		read_cnt = TOTAL_INSERT / 2;
394 	}
395 
396 	/* We first test single thread performance */
397 	start = rte_rdtsc_precise();
398 	/* Insert half of the keys */
399 	for (i = 0; i < TOTAL_INSERT / 2; i++) {
400 		ret = rte_hash_add_key_data(tbl_rw_test_param.h,
401 				     tbl_rw_test_param.keys + i,
402 					(void *)((uintptr_t)i));
403 		if (ret < 0) {
404 			printf("Failed to insert half of keys\n");
405 			goto err_free;
406 		}
407 	}
408 	end = rte_rdtsc_precise() - start;
409 	perf_results->single_write = end / i;
410 
411 	start = rte_rdtsc_precise();
412 
413 	for (i = 0; i < read_cnt; i++) {
414 		void *data;
415 		rte_hash_lookup_data(tbl_rw_test_param.h,
416 				tbl_rw_test_param.keys + i,
417 				&data);
418 		if (i != (uint64_t)(uintptr_t)data) {
419 			printf("lookup find wrong value"
420 					" %"PRIu64",%"PRIu64"\n", i,
421 					(uint64_t)(uintptr_t)data);
422 			break;
423 		}
424 	}
425 	end = rte_rdtsc_precise() - start;
426 	perf_results->single_read = end / i;
427 
428 	for (n = 0; n < NUM_TEST; n++) {
429 		unsigned int tot_worker_lcore = rte_lcore_count() - 1;
430 		if (tot_worker_lcore < core_cnt[n] * 2)
431 			goto finish;
432 
433 		__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
434 		__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
435 		__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
436 		__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
437 
438 		rte_hash_reset(tbl_rw_test_param.h);
439 
440 		tbl_rw_test_param.num_insert = TOTAL_INSERT / 2 / core_cnt[n];
441 		tbl_rw_test_param.rounded_tot_insert = TOTAL_INSERT / 2 +
442 						tbl_rw_test_param.num_insert *
443 						core_cnt[n];
444 
445 		for (i = 0; i < TOTAL_INSERT / 2; i++) {
446 			ret = rte_hash_add_key_data(tbl_rw_test_param.h,
447 					tbl_rw_test_param.keys + i,
448 					(void *)((uintptr_t)i));
449 			if (ret < 0) {
450 				printf("Failed to insert half of keys\n");
451 				goto err_free;
452 			}
453 		}
454 
455 		/* Then test multiple thread case but only all reads or
456 		 * all writes
457 		 */
458 
459 		/* Test only reader cases */
460 		for (i = 0; i < core_cnt[n]; i++)
461 			rte_eal_remote_launch(test_rw_reader,
462 					(void *)(uintptr_t)read_cnt,
463 					worker_core_ids[i]);
464 
465 		rte_eal_mp_wait_lcore();
466 
467 		start_coreid = i;
468 		/* Test only writer cases */
469 		for (; i < core_cnt[n] * 2; i++)
470 			rte_eal_remote_launch(test_rw_writer,
471 					(void *)((uintptr_t)start_coreid),
472 					worker_core_ids[i]);
473 
474 		rte_eal_mp_wait_lcore();
475 
476 		if (reader_faster) {
477 			unsigned long long int cycles_per_insertion =
478 				__atomic_load_n(&gread_cycles, __ATOMIC_RELAXED) /
479 				__atomic_load_n(&greads, __ATOMIC_RELAXED);
480 			perf_results->read_only[n] = cycles_per_insertion;
481 			printf("Reader only: cycles per lookup: %llu\n",
482 							cycles_per_insertion);
483 		}
484 
485 		else {
486 			unsigned long long int cycles_per_insertion =
487 				__atomic_load_n(&gwrite_cycles, __ATOMIC_RELAXED) /
488 				__atomic_load_n(&gwrites, __ATOMIC_RELAXED);
489 			perf_results->write_only[n] = cycles_per_insertion;
490 			printf("Writer only: cycles per writes: %llu\n",
491 							cycles_per_insertion);
492 		}
493 
494 		__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
495 		__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
496 		__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
497 		__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
498 
499 		rte_hash_reset(tbl_rw_test_param.h);
500 
501 		for (i = 0; i < TOTAL_INSERT / 2; i++) {
502 			ret = rte_hash_add_key_data(tbl_rw_test_param.h,
503 					tbl_rw_test_param.keys + i,
504 					(void *)((uintptr_t)i));
505 			if (ret < 0) {
506 				printf("Failed to insert half of keys\n");
507 				goto err_free;
508 			}
509 		}
510 
511 		start_coreid = core_cnt[n];
512 
513 		if (reader_faster) {
514 			for (i = core_cnt[n]; i < core_cnt[n] * 2; i++)
515 				rte_eal_remote_launch(test_rw_writer,
516 					(void *)((uintptr_t)start_coreid),
517 					worker_core_ids[i]);
518 			for (i = 0; i < core_cnt[n]; i++)
519 				rte_eal_remote_launch(test_rw_reader,
520 					(void *)(uintptr_t)read_cnt,
521 					worker_core_ids[i]);
522 		} else {
523 			for (i = 0; i < core_cnt[n]; i++)
524 				rte_eal_remote_launch(test_rw_reader,
525 					(void *)(uintptr_t)read_cnt,
526 					worker_core_ids[i]);
527 			for (; i < core_cnt[n] * 2; i++)
528 				rte_eal_remote_launch(test_rw_writer,
529 					(void *)((uintptr_t)start_coreid),
530 					worker_core_ids[i]);
531 		}
532 
533 		rte_eal_mp_wait_lcore();
534 
535 		iter = 0;
536 		memset(tbl_rw_test_param.found, 0, TOTAL_ENTRY);
537 		while (rte_hash_iterate(tbl_rw_test_param.h,
538 				&next_key, &next_data, &iter) >= 0) {
539 			/* Search for the key in the list of keys added .*/
540 			i = *(const uint32_t *)next_key;
541 			tbl_rw_test_param.found[i]++;
542 		}
543 
544 		for (i = 0; i < tbl_rw_test_param.rounded_tot_insert; i++) {
545 			if (tbl_rw_test_param.keys[i] != RTE_RWTEST_FAIL) {
546 				if (tbl_rw_test_param.found[i] > 1) {
547 					duplicated_keys++;
548 					break;
549 				}
550 				if (tbl_rw_test_param.found[i] == 0) {
551 					lost_keys++;
552 					printf("key %"PRIu64" is lost\n", i);
553 					break;
554 				}
555 			}
556 		}
557 
558 		if (duplicated_keys > 0) {
559 			printf("%d key duplicated\n", duplicated_keys);
560 			goto err_free;
561 		}
562 
563 		if (lost_keys > 0) {
564 			printf("%d key lost\n", lost_keys);
565 			goto err_free;
566 		}
567 
568 		printf("No key corrupted during read-write test.\n");
569 
570 		if (reader_faster) {
571 			unsigned long long int cycles_per_insertion =
572 				__atomic_load_n(&gread_cycles, __ATOMIC_RELAXED) /
573 				__atomic_load_n(&greads, __ATOMIC_RELAXED);
574 			perf_results->read_write_r[n] = cycles_per_insertion;
575 			printf("Read-write cycles per lookup: %llu\n",
576 							cycles_per_insertion);
577 		}
578 
579 		else {
580 			unsigned long long int cycles_per_insertion =
581 				__atomic_load_n(&gwrite_cycles, __ATOMIC_RELAXED) /
582 				__atomic_load_n(&gwrites, __ATOMIC_RELAXED);
583 			perf_results->read_write_w[n] = cycles_per_insertion;
584 			printf("Read-write cycles per writes: %llu\n",
585 							cycles_per_insertion);
586 		}
587 	}
588 
589 finish:
590 	rte_free(tbl_rw_test_param.found);
591 	rte_free(tbl_rw_test_param.keys);
592 	rte_hash_free(tbl_rw_test_param.h);
593 	return 0;
594 
595 err_free:
596 	rte_free(tbl_rw_test_param.found);
597 	rte_free(tbl_rw_test_param.keys);
598 	rte_hash_free(tbl_rw_test_param.h);
599 
600 err:
601 	return -1;
602 }
603 
604 static int
605 test_hash_rw_perf_main(void)
606 {
607 	/*
608 	 * Variables used to choose different tests.
609 	 * use_htm indicates if hardware transactional memory should be used.
610 	 * reader_faster indicates if the reader threads should finish earlier
611 	 * than writer threads. This is to timing either reader threads or
612 	 * writer threads for performance numbers.
613 	 */
614 	int use_htm, reader_faster;
615 	unsigned int i = 0, core_id = 0;
616 
617 	if (rte_lcore_count() < 3) {
618 		printf("Not enough cores for hash_readwrite_autotest, expecting at least 3\n");
619 		return TEST_SKIPPED;
620 	}
621 
622 	RTE_LCORE_FOREACH_WORKER(core_id) {
623 		worker_core_ids[i] = core_id;
624 		i++;
625 	}
626 
627 	setlocale(LC_NUMERIC, "");
628 
629 	if (rte_tm_supported()) {
630 		printf("Hardware transactional memory (lock elision) "
631 			"is supported\n");
632 
633 		printf("Test read-write with Hardware transactional memory\n");
634 
635 		use_htm = 1;
636 
637 		reader_faster = 1;
638 		if (test_hash_readwrite_perf(&htm_results, use_htm,
639 							reader_faster) < 0)
640 			return -1;
641 
642 		reader_faster = 0;
643 		if (test_hash_readwrite_perf(&htm_results, use_htm,
644 							reader_faster) < 0)
645 			return -1;
646 	} else {
647 		printf("Hardware transactional memory (lock elision) "
648 			"is NOT supported\n");
649 	}
650 
651 	printf("Test read-write without Hardware transactional memory\n");
652 	use_htm = 0;
653 
654 	reader_faster = 1;
655 	if (test_hash_readwrite_perf(&non_htm_results, use_htm,
656 							reader_faster) < 0)
657 		return -1;
658 	reader_faster = 0;
659 	if (test_hash_readwrite_perf(&non_htm_results, use_htm,
660 							reader_faster) < 0)
661 		return -1;
662 
663 	printf("================\n");
664 	printf("Results summary:\n");
665 	printf("================\n");
666 
667 	printf("single read: %u\n", htm_results.single_read);
668 	printf("single write: %u\n", htm_results.single_write);
669 	for (i = 0; i < NUM_TEST; i++) {
670 		printf("+++ core_cnt: %u +++\n", core_cnt[i]);
671 		printf("HTM:\n");
672 		printf("  read only: %u\n", htm_results.read_only[i]);
673 		printf("  write only: %u\n", htm_results.write_only[i]);
674 		printf("  read-write read: %u\n", htm_results.read_write_r[i]);
675 		printf("  read-write write: %u\n", htm_results.read_write_w[i]);
676 
677 		printf("non HTM:\n");
678 		printf("  read only: %u\n", non_htm_results.read_only[i]);
679 		printf("  write only: %u\n", non_htm_results.write_only[i]);
680 		printf("  read-write read: %u\n",
681 			non_htm_results.read_write_r[i]);
682 		printf("  read-write write: %u\n",
683 			non_htm_results.read_write_w[i]);
684 	}
685 
686 	return 0;
687 }
688 
689 static int
690 test_hash_rw_func_main(void)
691 {
692 	/*
693 	 * Variables used to choose different tests.
694 	 * use_htm indicates if hardware transactional memory should be used.
695 	 * reader_faster indicates if the reader threads should finish earlier
696 	 * than writer threads. This is to timing either reader threads or
697 	 * writer threads for performance numbers.
698 	 */
699 	unsigned int i = 0, core_id = 0;
700 
701 	if (rte_lcore_count() < 3) {
702 		printf("Not enough cores for hash_readwrite_autotest, expecting at least 3\n");
703 		return TEST_SKIPPED;
704 	}
705 
706 	RTE_LCORE_FOREACH_WORKER(core_id) {
707 		worker_core_ids[i] = core_id;
708 		i++;
709 	}
710 
711 	setlocale(LC_NUMERIC, "");
712 
713 	if (rte_tm_supported()) {
714 		printf("Hardware transactional memory (lock elision) "
715 			"is supported\n");
716 
717 		printf("Test read-write with Hardware transactional memory\n");
718 
719 		/* htm = 1, rw_lf = 0, ext = 0 */
720 		if (test_hash_readwrite_functional(1, 0, 0) < 0)
721 			return -1;
722 
723 		/* htm = 1, rw_lf = 1, ext = 0 */
724 		if (test_hash_readwrite_functional(1, 1, 0) < 0)
725 			return -1;
726 
727 		/* htm = 1, rw_lf = 0, ext = 1 */
728 		if (test_hash_readwrite_functional(1, 0, 1) < 0)
729 			return -1;
730 
731 		/* htm = 1, rw_lf = 1, ext = 1 */
732 		if (test_hash_readwrite_functional(1, 1, 1) < 0)
733 			return -1;
734 	} else {
735 		printf("Hardware transactional memory (lock elision) "
736 			"is NOT supported\n");
737 	}
738 
739 	printf("Test read-write without Hardware transactional memory\n");
740 	/* htm = 0, rw_lf = 0, ext = 0 */
741 	if (test_hash_readwrite_functional(0, 0, 0) < 0)
742 		return -1;
743 
744 	/* htm = 0, rw_lf = 1, ext = 0 */
745 	if (test_hash_readwrite_functional(0, 1, 0) < 0)
746 		return -1;
747 
748 	/* htm = 0, rw_lf = 0, ext = 1 */
749 	if (test_hash_readwrite_functional(0, 0, 1) < 0)
750 		return -1;
751 
752 	/* htm = 0, rw_lf = 1, ext = 1 */
753 	if (test_hash_readwrite_functional(0, 1, 1) < 0)
754 		return -1;
755 
756 	return 0;
757 }
758 
759 REGISTER_TEST_COMMAND(hash_readwrite_func_autotest, test_hash_rw_func_main);
760 REGISTER_TEST_COMMAND(hash_readwrite_perf_autotest, test_hash_rw_perf_main);
761