1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Intel Corporation
8 #include <rte_cycles.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>
19 #define RTE_RWTEST_FAIL 0
21 #define TOTAL_ENTRY (16*1024*1024)
22 #define TOTAL_INSERT (15*1024*1024)
25 unsigned int core_cnt
[NUM_TEST
] = {2, 4, 8};
29 uint32_t single_write
;
30 uint32_t read_only
[NUM_TEST
];
31 uint32_t write_only
[NUM_TEST
];
32 uint32_t read_write_r
[NUM_TEST
];
33 uint32_t read_write_w
[NUM_TEST
];
36 static struct perf htm_results
, non_htm_results
;
42 uint32_t rounded_tot_insert
;
46 static rte_atomic64_t gcycles
;
47 static rte_atomic64_t ginsertions
;
49 static rte_atomic64_t gread_cycles
;
50 static rte_atomic64_t gwrite_cycles
;
52 static rte_atomic64_t greads
;
53 static rte_atomic64_t gwrites
;
56 test_hash_readwrite_worker(__attribute__((unused
)) void *arg
)
59 uint32_t lcore_id
= rte_lcore_id();
60 uint64_t begin
, cycles
;
63 offset
= (lcore_id
- rte_get_master_lcore())
64 * tbl_rw_test_param
.num_insert
;
66 printf("Core #%d inserting and reading %d: %'"PRId64
" - %'"PRId64
"\n",
67 lcore_id
, tbl_rw_test_param
.num_insert
,
68 offset
, offset
+ tbl_rw_test_param
.num_insert
);
70 begin
= rte_rdtsc_precise();
72 for (i
= offset
; i
< offset
+ tbl_rw_test_param
.num_insert
; i
++) {
74 if (rte_hash_lookup(tbl_rw_test_param
.h
,
75 tbl_rw_test_param
.keys
+ i
) > 0)
78 ret
= rte_hash_add_key(tbl_rw_test_param
.h
,
79 tbl_rw_test_param
.keys
+ i
);
83 if (rte_hash_lookup(tbl_rw_test_param
.h
,
84 tbl_rw_test_param
.keys
+ i
) != ret
)
88 cycles
= rte_rdtsc_precise() - begin
;
89 rte_atomic64_add(&gcycles
, cycles
);
90 rte_atomic64_add(&ginsertions
, i
- offset
);
92 for (; i
< offset
+ tbl_rw_test_param
.num_insert
; i
++)
93 tbl_rw_test_param
.keys
[i
] = RTE_RWTEST_FAIL
;
99 init_params(int use_htm
, int use_jhash
)
103 uint32_t *keys
= NULL
;
104 uint32_t *found
= NULL
;
105 struct rte_hash
*handle
;
107 struct rte_hash_parameters hash_params
= {
108 .entries
= TOTAL_ENTRY
,
109 .key_len
= sizeof(uint32_t),
110 .hash_func_init_val
= 0,
111 .socket_id
= rte_socket_id(),
114 hash_params
.hash_func
= rte_jhash
;
116 hash_params
.hash_func
= rte_hash_crc
;
119 hash_params
.extra_flag
=
120 RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT
|
121 RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY
;
123 hash_params
.extra_flag
=
124 RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY
;
126 hash_params
.name
= "tests";
128 handle
= rte_hash_create(&hash_params
);
129 if (handle
== NULL
) {
130 printf("hash creation failed");
134 tbl_rw_test_param
.h
= handle
;
135 keys
= rte_malloc(NULL
, sizeof(uint32_t) * TOTAL_ENTRY
, 0);
138 printf("RTE_MALLOC failed\n");
142 found
= rte_zmalloc(NULL
, sizeof(uint32_t) * TOTAL_ENTRY
, 0);
144 printf("RTE_ZMALLOC failed\n");
148 tbl_rw_test_param
.keys
= keys
;
149 tbl_rw_test_param
.found
= found
;
151 for (i
= 0; i
< TOTAL_ENTRY
; i
++)
158 rte_hash_free(handle
);
164 test_hash_readwrite_functional(int use_htm
)
167 const void *next_key
;
171 uint32_t duplicated_keys
= 0;
172 uint32_t lost_keys
= 0;
175 rte_atomic64_init(&gcycles
);
176 rte_atomic64_clear(&gcycles
);
178 rte_atomic64_init(&ginsertions
);
179 rte_atomic64_clear(&ginsertions
);
181 if (init_params(use_htm
, use_jhash
) != 0)
184 tbl_rw_test_param
.num_insert
=
185 TOTAL_INSERT
/ rte_lcore_count();
187 tbl_rw_test_param
.rounded_tot_insert
=
188 tbl_rw_test_param
.num_insert
191 printf("++++++++Start function tests:+++++++++\n");
193 /* Fire all threads. */
194 rte_eal_mp_remote_launch(test_hash_readwrite_worker
,
196 rte_eal_mp_wait_lcore();
198 while (rte_hash_iterate(tbl_rw_test_param
.h
, &next_key
,
199 &next_data
, &iter
) >= 0) {
200 /* Search for the key in the list of keys added .*/
201 i
= *(const uint32_t *)next_key
;
202 tbl_rw_test_param
.found
[i
]++;
205 for (i
= 0; i
< tbl_rw_test_param
.rounded_tot_insert
; i
++) {
206 if (tbl_rw_test_param
.keys
[i
] != RTE_RWTEST_FAIL
) {
207 if (tbl_rw_test_param
.found
[i
] > 1) {
211 if (tbl_rw_test_param
.found
[i
] == 0) {
213 printf("key %d is lost\n", i
);
219 if (duplicated_keys
> 0) {
220 printf("%d key duplicated\n", duplicated_keys
);
225 printf("%d key lost\n", lost_keys
);
229 printf("No key corrupted during read-write test.\n");
231 unsigned long long int cycles_per_insertion
=
232 rte_atomic64_read(&gcycles
) /
233 rte_atomic64_read(&ginsertions
);
235 printf("cycles per insertion and lookup: %llu\n", cycles_per_insertion
);
237 rte_free(tbl_rw_test_param
.found
);
238 rte_free(tbl_rw_test_param
.keys
);
239 rte_hash_free(tbl_rw_test_param
.h
);
240 printf("+++++++++Complete function tests+++++++++\n");
244 rte_free(tbl_rw_test_param
.found
);
245 rte_free(tbl_rw_test_param
.keys
);
246 rte_hash_free(tbl_rw_test_param
.h
);
252 test_rw_reader(__attribute__((unused
)) void *arg
)
255 uint64_t begin
, cycles
;
256 uint64_t read_cnt
= (uint64_t)((uintptr_t)arg
);
258 begin
= rte_rdtsc_precise();
259 for (i
= 0; i
< read_cnt
; i
++) {
261 rte_hash_lookup_data(tbl_rw_test_param
.h
,
262 tbl_rw_test_param
.keys
+ i
,
264 if (i
!= (uint64_t)(uintptr_t)data
) {
265 printf("lookup find wrong value %"PRIu64
","
267 (uint64_t)(uintptr_t)data
);
272 cycles
= rte_rdtsc_precise() - begin
;
273 rte_atomic64_add(&gread_cycles
, cycles
);
274 rte_atomic64_add(&greads
, i
);
279 test_rw_writer(__attribute__((unused
)) void *arg
)
282 uint32_t lcore_id
= rte_lcore_id();
283 uint64_t begin
, cycles
;
285 uint64_t start_coreid
= (uint64_t)(uintptr_t)arg
;
288 offset
= TOTAL_INSERT
/ 2 + (lcore_id
- start_coreid
)
289 * tbl_rw_test_param
.num_insert
;
290 begin
= rte_rdtsc_precise();
291 for (i
= offset
; i
< offset
+ tbl_rw_test_param
.num_insert
; i
++) {
292 ret
= rte_hash_add_key_data(tbl_rw_test_param
.h
,
293 tbl_rw_test_param
.keys
+ i
,
294 (void *)((uintptr_t)i
));
296 printf("writer failed %"PRIu64
"\n", i
);
301 cycles
= rte_rdtsc_precise() - begin
;
302 rte_atomic64_add(&gwrite_cycles
, cycles
);
303 rte_atomic64_add(&gwrites
, tbl_rw_test_param
.num_insert
);
308 test_hash_readwrite_perf(struct perf
*perf_results
, int use_htm
,
314 uint64_t i
, read_cnt
;
316 const void *next_key
;
321 uint32_t duplicated_keys
= 0;
322 uint32_t lost_keys
= 0;
324 uint64_t start
= 0, end
= 0;
326 rte_atomic64_init(&greads
);
327 rte_atomic64_init(&gwrites
);
328 rte_atomic64_clear(&gwrites
);
329 rte_atomic64_clear(&greads
);
331 rte_atomic64_init(&gread_cycles
);
332 rte_atomic64_clear(&gread_cycles
);
333 rte_atomic64_init(&gwrite_cycles
);
334 rte_atomic64_clear(&gwrite_cycles
);
336 if (init_params(use_htm
, use_jhash
) != 0)
340 * Do a readers finish faster or writers finish faster test.
341 * When readers finish faster, we timing the readers, and when writers
342 * finish faster, we timing the writers.
343 * Divided by 10 or 2 is just experimental values to vary the workload
347 printf("++++++Start perf test: reader++++++++\n");
348 read_cnt
= TOTAL_INSERT
/ 10;
350 printf("++++++Start perf test: writer++++++++\n");
351 read_cnt
= TOTAL_INSERT
/ 2;
354 /* We first test single thread performance */
355 start
= rte_rdtsc_precise();
356 /* Insert half of the keys */
357 for (i
= 0; i
< TOTAL_INSERT
/ 2; i
++) {
358 ret
= rte_hash_add_key_data(tbl_rw_test_param
.h
,
359 tbl_rw_test_param
.keys
+ i
,
360 (void *)((uintptr_t)i
));
362 printf("Failed to insert half of keys\n");
366 end
= rte_rdtsc_precise() - start
;
367 perf_results
->single_write
= end
/ i
;
369 start
= rte_rdtsc_precise();
371 for (i
= 0; i
< read_cnt
; i
++) {
373 rte_hash_lookup_data(tbl_rw_test_param
.h
,
374 tbl_rw_test_param
.keys
+ i
,
376 if (i
!= (uint64_t)(uintptr_t)data
) {
377 printf("lookup find wrong value"
378 " %"PRIu64
",%"PRIu64
"\n", i
,
379 (uint64_t)(uintptr_t)data
);
383 end
= rte_rdtsc_precise() - start
;
384 perf_results
->single_read
= end
/ i
;
386 for (n
= 0; n
< NUM_TEST
; n
++) {
387 unsigned int tot_lcore
= rte_lcore_count();
388 if (tot_lcore
< core_cnt
[n
] * 2 + 1)
391 rte_atomic64_clear(&greads
);
392 rte_atomic64_clear(&gread_cycles
);
393 rte_atomic64_clear(&gwrites
);
394 rte_atomic64_clear(&gwrite_cycles
);
396 rte_hash_reset(tbl_rw_test_param
.h
);
398 tbl_rw_test_param
.num_insert
= TOTAL_INSERT
/ 2 / core_cnt
[n
];
399 tbl_rw_test_param
.rounded_tot_insert
= TOTAL_INSERT
/ 2 +
400 tbl_rw_test_param
.num_insert
*
403 for (i
= 0; i
< TOTAL_INSERT
/ 2; i
++) {
404 ret
= rte_hash_add_key_data(tbl_rw_test_param
.h
,
405 tbl_rw_test_param
.keys
+ i
,
406 (void *)((uintptr_t)i
));
408 printf("Failed to insert half of keys\n");
413 /* Then test multiple thread case but only all reads or
417 /* Test only reader cases */
418 for (i
= 1; i
<= core_cnt
[n
]; i
++)
419 rte_eal_remote_launch(test_rw_reader
,
420 (void *)(uintptr_t)read_cnt
, i
);
422 rte_eal_mp_wait_lcore();
425 /* Test only writer cases */
426 for (; i
<= core_cnt
[n
] * 2; i
++)
427 rte_eal_remote_launch(test_rw_writer
,
428 (void *)((uintptr_t)start_coreid
), i
);
430 rte_eal_mp_wait_lcore();
433 unsigned long long int cycles_per_insertion
=
434 rte_atomic64_read(&gread_cycles
) /
435 rte_atomic64_read(&greads
);
436 perf_results
->read_only
[n
] = cycles_per_insertion
;
437 printf("Reader only: cycles per lookup: %llu\n",
438 cycles_per_insertion
);
442 unsigned long long int cycles_per_insertion
=
443 rte_atomic64_read(&gwrite_cycles
) /
444 rte_atomic64_read(&gwrites
);
445 perf_results
->write_only
[n
] = cycles_per_insertion
;
446 printf("Writer only: cycles per writes: %llu\n",
447 cycles_per_insertion
);
450 rte_atomic64_clear(&greads
);
451 rte_atomic64_clear(&gread_cycles
);
452 rte_atomic64_clear(&gwrites
);
453 rte_atomic64_clear(&gwrite_cycles
);
455 rte_hash_reset(tbl_rw_test_param
.h
);
457 for (i
= 0; i
< TOTAL_INSERT
/ 2; i
++) {
458 ret
= rte_hash_add_key_data(tbl_rw_test_param
.h
,
459 tbl_rw_test_param
.keys
+ i
,
460 (void *)((uintptr_t)i
));
462 printf("Failed to insert half of keys\n");
467 start_coreid
= core_cnt
[n
] + 1;
470 for (i
= core_cnt
[n
] + 1; i
<= core_cnt
[n
] * 2; i
++)
471 rte_eal_remote_launch(test_rw_writer
,
472 (void *)((uintptr_t)start_coreid
), i
);
473 for (i
= 1; i
<= core_cnt
[n
]; i
++)
474 rte_eal_remote_launch(test_rw_reader
,
475 (void *)(uintptr_t)read_cnt
, i
);
477 for (i
= 1; i
<= core_cnt
[n
]; i
++)
478 rte_eal_remote_launch(test_rw_reader
,
479 (void *)(uintptr_t)read_cnt
, i
);
480 for (; i
<= core_cnt
[n
] * 2; i
++)
481 rte_eal_remote_launch(test_rw_writer
,
482 (void *)((uintptr_t)start_coreid
), i
);
485 rte_eal_mp_wait_lcore();
487 while (rte_hash_iterate(tbl_rw_test_param
.h
,
488 &next_key
, &next_data
, &iter
) >= 0) {
489 /* Search for the key in the list of keys added .*/
490 i
= *(const uint32_t *)next_key
;
491 tbl_rw_test_param
.found
[i
]++;
494 for (i
= 0; i
< tbl_rw_test_param
.rounded_tot_insert
; i
++) {
495 if (tbl_rw_test_param
.keys
[i
] != RTE_RWTEST_FAIL
) {
496 if (tbl_rw_test_param
.found
[i
] > 1) {
500 if (tbl_rw_test_param
.found
[i
] == 0) {
502 printf("key %"PRIu64
" is lost\n", i
);
508 if (duplicated_keys
> 0) {
509 printf("%d key duplicated\n", duplicated_keys
);
514 printf("%d key lost\n", lost_keys
);
518 printf("No key corrupted during read-write test.\n");
521 unsigned long long int cycles_per_insertion
=
522 rte_atomic64_read(&gread_cycles
) /
523 rte_atomic64_read(&greads
);
524 perf_results
->read_write_r
[n
] = cycles_per_insertion
;
525 printf("Read-write cycles per lookup: %llu\n",
526 cycles_per_insertion
);
530 unsigned long long int cycles_per_insertion
=
531 rte_atomic64_read(&gwrite_cycles
) /
532 rte_atomic64_read(&gwrites
);
533 perf_results
->read_write_w
[n
] = cycles_per_insertion
;
534 printf("Read-write cycles per writes: %llu\n",
535 cycles_per_insertion
);
540 rte_free(tbl_rw_test_param
.found
);
541 rte_free(tbl_rw_test_param
.keys
);
542 rte_hash_free(tbl_rw_test_param
.h
);
546 rte_free(tbl_rw_test_param
.found
);
547 rte_free(tbl_rw_test_param
.keys
);
548 rte_hash_free(tbl_rw_test_param
.h
);
555 test_hash_readwrite_main(void)
558 * Variables used to choose different tests.
559 * use_htm indicates if hardware transactional memory should be used.
560 * reader_faster indicates if the reader threads should finish earlier
561 * than writer threads. This is to timing either reader threads or
562 * writer threads for performance numbers.
564 int use_htm
, reader_faster
;
566 if (rte_lcore_count() == 1) {
567 printf("More than one lcore is required "
568 "to do read write test\n");
572 setlocale(LC_NUMERIC
, "");
574 if (rte_tm_supported()) {
575 printf("Hardware transactional memory (lock elision) "
578 printf("Test read-write with Hardware transactional memory\n");
581 if (test_hash_readwrite_functional(use_htm
) < 0)
585 if (test_hash_readwrite_perf(&htm_results
, use_htm
,
590 if (test_hash_readwrite_perf(&htm_results
, use_htm
,
594 printf("Hardware transactional memory (lock elision) "
595 "is NOT supported\n");
598 printf("Test read-write without Hardware transactional memory\n");
600 if (test_hash_readwrite_functional(use_htm
) < 0)
603 if (test_hash_readwrite_perf(&non_htm_results
, use_htm
,
607 if (test_hash_readwrite_perf(&non_htm_results
, use_htm
,
611 printf("Results summary:\n");
615 printf("single read: %u\n", htm_results
.single_read
);
616 printf("single write: %u\n", htm_results
.single_write
);
617 for (i
= 0; i
< NUM_TEST
; i
++) {
618 printf("core_cnt: %u\n", core_cnt
[i
]);
620 printf("read only: %u\n", htm_results
.read_only
[i
]);
621 printf("write only: %u\n", htm_results
.write_only
[i
]);
622 printf("read-write read: %u\n", htm_results
.read_write_r
[i
]);
623 printf("read-write write: %u\n", htm_results
.read_write_w
[i
]);
625 printf("non HTM:\n");
626 printf("read only: %u\n", non_htm_results
.read_only
[i
]);
627 printf("write only: %u\n", non_htm_results
.write_only
[i
]);
628 printf("read-write read: %u\n",
629 non_htm_results
.read_write_r
[i
]);
630 printf("read-write write: %u\n",
631 non_htm_results
.read_write_w
[i
]);
637 REGISTER_TEST_COMMAND(hash_readwrite_autotest
, test_hash_readwrite_main
);