app/test-compress-perf/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <signal.h>
   6 #include <sys/types.h>
   7 #include <unistd.h>
   8
   9 #include <rte_malloc.h>
  10 #include <rte_eal.h>
  11 #include <rte_log.h>
  12 #include <rte_compressdev.h>
  13
  14 #include "comp_perf.h"
  15 #include "comp_perf_options.h"
  16 #include "comp_perf_test_common.h"
  17 #include "comp_perf_test_cyclecount.h"
  18 #include "comp_perf_test_throughput.h"
  19 #include "comp_perf_test_verify.h"
  20
  21 #define NUM_MAX_XFORMS 16
  22 #define NUM_MAX_INFLIGHT_OPS 512
  23
  24 __extension__
  25 const char *comp_perf_test_type_strs[] = {
  26         [CPERF_TEST_TYPE_THROUGHPUT] = "throughput",
  27         [CPERF_TEST_TYPE_VERIFY] = "verify",
  28         [CPERF_TEST_TYPE_PMDCC] = "pmd-cyclecount"
  29 };
  30
  31 __extension__
  32 static const struct cperf_test cperf_testmap[] = {
  33         [CPERF_TEST_TYPE_THROUGHPUT] = {
  34                         cperf_throughput_test_constructor,
  35                         cperf_throughput_test_runner,
  36                         cperf_throughput_test_destructor
  37
  38         },
  39         [CPERF_TEST_TYPE_VERIFY] = {
  40                         cperf_verify_test_constructor,
  41                         cperf_verify_test_runner,
  42                         cperf_verify_test_destructor
  43         },
  44
  45         [CPERF_TEST_TYPE_PMDCC] = {
  46                         cperf_cyclecount_test_constructor,
  47                         cperf_cyclecount_test_runner,
  48                         cperf_cyclecount_test_destructor
  49         }
  50 };
  51
  52 static struct comp_test_data *test_data;
  53
  54 static int
  55 comp_perf_check_capabilities(struct comp_test_data *test_data, uint8_t cdev_id)
  56 {
  57         const struct rte_compressdev_capabilities *cap;
  58
  59         cap = rte_compressdev_capability_get(cdev_id,
  60                                              RTE_COMP_ALGO_DEFLATE);
  61
  62         if (cap == NULL) {
  63                 RTE_LOG(ERR, USER1,
  64                         "Compress device does not support DEFLATE\n");
  65                 return -1;
  66         }
  67
  68         uint64_t comp_flags = cap->comp_feature_flags;
  69
  70         /* Huffman enconding */
  71         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_FIXED &&
  72                         (comp_flags & RTE_COMP_FF_HUFFMAN_FIXED) == 0) {
  73                 RTE_LOG(ERR, USER1,
  74                         "Compress device does not supported Fixed Huffman\n");
  75                 return -1;
  76         }
  77
  78         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_DYNAMIC &&
  79                         (comp_flags & RTE_COMP_FF_HUFFMAN_DYNAMIC) == 0) {
  80                 RTE_LOG(ERR, USER1,
  81                         "Compress device does not supported Dynamic Huffman\n");
  82                 return -1;
  83         }
  84
  85         /* Window size */
  86         if (test_data->window_sz != -1) {
  87                 if (param_range_check(test_data->window_sz, &cap->window_size)
  88                                 < 0) {
  89                         RTE_LOG(ERR, USER1,
  90                                 "Compress device does not support "
  91                                 "this window size\n");
  92                         return -1;
  93                 }
  94         } else
  95                 /* Set window size to PMD maximum if none was specified */
  96                 test_data->window_sz = cap->window_size.max;
  97
  98         /* Check if chained mbufs is supported */
  99         if (test_data->max_sgl_segs > 1  &&
 100                         (comp_flags & RTE_COMP_FF_OOP_SGL_IN_SGL_OUT) == 0) {
 101                 RTE_LOG(INFO, USER1, "Compress device does not support "
 102                                 "chained mbufs. Max SGL segments set to 1\n");
 103                 test_data->max_sgl_segs = 1;
 104         }
 105
 106         /* Level 0 support */
 107         if (test_data->level_lst.min == 0 &&
 108                         (comp_flags & RTE_COMP_FF_NONCOMPRESSED_BLOCKS) == 0) {
 109                 RTE_LOG(ERR, USER1, "Compress device does not support "
 110                                 "level 0 (no compression)\n");
 111                 return -1;
 112         }
 113
 114         return 0;
 115 }
 116
 117 static int
 118 comp_perf_initialize_compressdev(struct comp_test_data *test_data,
 119                                  uint8_t *enabled_cdevs)
 120 {
 121         uint8_t enabled_cdev_count, nb_lcores, cdev_id;
 122         unsigned int i, j;
 123         int ret;
 124
 125         enabled_cdev_count = rte_compressdev_devices_get(test_data->driver_name,
 126                         enabled_cdevs, RTE_COMPRESS_MAX_DEVS);
 127         if (enabled_cdev_count == 0) {
 128                 RTE_LOG(ERR, USER1, "No compress devices type %s available,"
 129                                     " please check the list of specified devices in EAL section\n",
 130                                 test_data->driver_name);
 131                 return -EINVAL;
 132         }
 133
 134         nb_lcores = rte_lcore_count() - 1;
 135         /*
 136          * Use fewer devices,
 137          * if there are more available than cores.
 138          */
 139         if (enabled_cdev_count > nb_lcores) {
 140                 if (nb_lcores == 0) {
 141                         RTE_LOG(ERR, USER1, "Cannot run with 0 cores! Increase the number of cores\n");
 142                         return -EINVAL;
 143                 }
 144                 enabled_cdev_count = nb_lcores;
 145                 RTE_LOG(INFO, USER1,
 146                         "There's more available devices than cores!"
 147                         " The number of devices has been aligned to %d cores\n",
 148                         nb_lcores);
 149         }
 150
 151         /*
 152          * Calculate number of needed queue pairs, based on the amount
 153          * of available number of logical cores and compression devices.
 154          * For instance, if there are 4 cores and 2 compression devices,
 155          * 2 queue pairs will be set up per device.
 156          * One queue pair per one core.
 157          * if e.g.: there're 3 cores and 2 compression devices,
 158          * 2 queue pairs will be set up per device but one queue pair
 159          * will left unused in the last one device
 160          */
 161         test_data->nb_qps = (nb_lcores % enabled_cdev_count) ?
 162                                 (nb_lcores / enabled_cdev_count) + 1 :
 163                                 nb_lcores / enabled_cdev_count;
 164
 165         for (i = 0; i < enabled_cdev_count &&
 166                         i < RTE_COMPRESS_MAX_DEVS; i++,
 167                                         nb_lcores -= test_data->nb_qps) {
 168                 cdev_id = enabled_cdevs[i];
 169
 170                 struct rte_compressdev_info cdev_info;
 171                 uint8_t socket_id = rte_compressdev_socket_id(cdev_id);
 172
 173                 rte_compressdev_info_get(cdev_id, &cdev_info);
 174                 if (cdev_info.max_nb_queue_pairs &&
 175                         test_data->nb_qps > cdev_info.max_nb_queue_pairs) {
 176                         RTE_LOG(ERR, USER1,
 177                                 "Number of needed queue pairs is higher "
 178                                 "than the maximum number of queue pairs "
 179                                 "per device.\n");
 180                         RTE_LOG(ERR, USER1,
 181                                 "Lower the number of cores or increase "
 182                                 "the number of crypto devices\n");
 183                         return -EINVAL;
 184                 }
 185
 186                 if (comp_perf_check_capabilities(test_data, cdev_id) < 0)
 187                         return -EINVAL;
 188
 189                 /* Configure compressdev */
 190                 struct rte_compressdev_config config = {
 191                         .socket_id = socket_id,
 192                         .nb_queue_pairs = nb_lcores > test_data->nb_qps
 193                                         ? test_data->nb_qps : nb_lcores,
 194                         .max_nb_priv_xforms = NUM_MAX_XFORMS,
 195                         .max_nb_streams = 0
 196                 };
 197
 198                 if (rte_compressdev_configure(cdev_id, &config) < 0) {
 199                         RTE_LOG(ERR, USER1, "Device configuration failed\n");
 200                         return -EINVAL;
 201                 }
 202
 203                 for (j = 0; j < test_data->nb_qps; j++) {
 204                         ret = rte_compressdev_queue_pair_setup(cdev_id, j,
 205                                         NUM_MAX_INFLIGHT_OPS, socket_id);
 206                         if (ret < 0) {
 207                                 RTE_LOG(ERR, USER1,
 208                               "Failed to setup queue pair %u on compressdev %u",
 209                                         j, cdev_id);
 210                                 return -EINVAL;
 211                         }
 212                 }
 213
 214                 ret = rte_compressdev_start(cdev_id);
 215                 if (ret < 0) {
 216                         RTE_LOG(ERR, USER1,
 217                                 "Failed to start device %u: error %d\n",
 218                                 cdev_id, ret);
 219                         return -EPERM;
 220                 }
 221         }
 222
 223         return enabled_cdev_count;
 224 }
 225
 226 static int
 227 comp_perf_dump_input_data(struct comp_test_data *test_data)
 228 {
 229         FILE *f = fopen(test_data->input_file, "r");
 230         int ret = -1;
 231
 232         if (f == NULL) {
 233                 RTE_LOG(ERR, USER1, "Input file could not be opened\n");
 234                 return -1;
 235         }
 236
 237         if (fseek(f, 0, SEEK_END) != 0) {
 238                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 239                 goto end;
 240         }
 241         size_t actual_file_sz = ftell(f);
 242         /* If extended input data size has not been set,
 243          * input data size = file size
 244          */
 245
 246         if (test_data->input_data_sz == 0)
 247                 test_data->input_data_sz = actual_file_sz;
 248
 249         if (test_data->input_data_sz <= 0 || actual_file_sz <= 0 ||
 250                         fseek(f, 0, SEEK_SET) != 0) {
 251                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 252                 goto end;
 253         }
 254
 255         test_data->input_data = rte_zmalloc_socket(NULL,
 256                                 test_data->input_data_sz, 0, rte_socket_id());
 257
 258         if (test_data->input_data == NULL) {
 259                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 260                                 "file could not be allocated\n");
 261                 goto end;
 262         }
 263
 264         size_t remaining_data = test_data->input_data_sz;
 265         uint8_t *data = test_data->input_data;
 266
 267         while (remaining_data > 0) {
 268                 size_t data_to_read = RTE_MIN(remaining_data, actual_file_sz);
 269
 270                 if (fread(data, data_to_read, 1, f) != 1) {
 271                         RTE_LOG(ERR, USER1, "Input file could not be read\n");
 272                         goto end;
 273                 }
 274                 if (fseek(f, 0, SEEK_SET) != 0) {
 275                         RTE_LOG(ERR, USER1,
 276                                 "Size of input could not be calculated\n");
 277                         goto end;
 278                 }
 279                 remaining_data -= data_to_read;
 280                 data += data_to_read;
 281         }
 282
 283         printf("\n");
 284         if (test_data->input_data_sz > actual_file_sz)
 285                 RTE_LOG(INFO, USER1,
 286                   "%zu bytes read from file %s, extending the file %.2f times\n",
 287                         test_data->input_data_sz, test_data->input_file,
 288                         (double)test_data->input_data_sz/actual_file_sz);
 289         else
 290                 RTE_LOG(INFO, USER1,
 291                         "%zu bytes read from file %s\n",
 292                         test_data->input_data_sz, test_data->input_file);
 293
 294         ret = 0;
 295
 296 end:
 297         fclose(f);
 298         return ret;
 299 }
 300
 301 static void
 302 comp_perf_cleanup_on_signal(int signalNumber __rte_unused)
 303 {
 304         test_data->perf_comp_force_stop = 1;
 305 }
 306
 307 static void
 308 comp_perf_register_cleanup_on_signal(void)
 309 {
 310         signal(SIGTERM, comp_perf_cleanup_on_signal);
 311         signal(SIGINT, comp_perf_cleanup_on_signal);
 312 }
 313
 314 int
 315 main(int argc, char **argv)
 316 {
 317         uint8_t level_idx = 0;
 318         int ret, i;
 319         void *ctx[RTE_MAX_LCORE] = {};
 320         uint8_t enabled_cdevs[RTE_COMPRESS_MAX_DEVS];
 321         int nb_compressdevs = 0;
 322         uint16_t total_nb_qps = 0;
 323         uint8_t cdev_id;
 324         uint32_t lcore_id;
 325
 326         /* Initialise DPDK EAL */
 327         ret = rte_eal_init(argc, argv);
 328         if (ret < 0)
 329                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments!\n");
 330         argc -= ret;
 331         argv += ret;
 332
 333         test_data = rte_zmalloc_socket(NULL, sizeof(struct comp_test_data),
 334                                         0, rte_socket_id());
 335
 336         if (test_data == NULL)
 337                 rte_exit(EXIT_FAILURE, "Cannot reserve memory in socket %d\n",
 338                                 rte_socket_id());
 339
 340         comp_perf_register_cleanup_on_signal();
 341
 342         ret = EXIT_SUCCESS;
 343         test_data->cleanup = ST_TEST_DATA;
 344         comp_perf_options_default(test_data);
 345
 346         if (comp_perf_options_parse(test_data, argc, argv) < 0) {
 347                 RTE_LOG(ERR, USER1,
 348                         "Parsing one or more user options failed\n");
 349                 ret = EXIT_FAILURE;
 350                 goto end;
 351         }
 352
 353         if (comp_perf_options_check(test_data) < 0) {
 354                 ret = EXIT_FAILURE;
 355                 goto end;
 356         }
 357
 358         nb_compressdevs =
 359                 comp_perf_initialize_compressdev(test_data, enabled_cdevs);
 360
 361         if (nb_compressdevs < 1) {
 362                 ret = EXIT_FAILURE;
 363                 goto end;
 364         }
 365
 366         test_data->cleanup = ST_COMPDEV;
 367         if (comp_perf_dump_input_data(test_data) < 0) {
 368                 ret = EXIT_FAILURE;
 369                 goto end;
 370         }
 371
 372         test_data->cleanup = ST_INPUT_DATA;
 373
 374         if (test_data->level_lst.inc != 0)
 375                 test_data->level = test_data->level_lst.min;
 376         else
 377                 test_data->level = test_data->level_lst.list[0];
 378
 379         printf("\nApp uses socket: %u\n", rte_socket_id());
 380         printf("Burst size = %u\n", test_data->burst_sz);
 381         printf("Input data size = %zu\n", test_data->input_data_sz);
 382         if (test_data->test == CPERF_TEST_TYPE_PMDCC)
 383                 printf("Cycle-count delay = %u [us]\n",
 384                        test_data->cyclecount_delay);
 385
 386         test_data->cleanup = ST_DURING_TEST;
 387         total_nb_qps = nb_compressdevs * test_data->nb_qps;
 388
 389         i = 0;
 390         uint8_t qp_id = 0, cdev_index = 0;
 391
 392         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 393
 394                 if (i == total_nb_qps)
 395                         break;
 396
 397                 cdev_id = enabled_cdevs[cdev_index];
 398                 ctx[i] = cperf_testmap[test_data->test].constructor(
 399                                                         cdev_id, qp_id,
 400                                                         test_data);
 401                 if (ctx[i] == NULL) {
 402                         RTE_LOG(ERR, USER1, "Test run constructor failed\n");
 403                         goto end;
 404                 }
 405                 qp_id = (qp_id + 1) % test_data->nb_qps;
 406                 if (qp_id == 0)
 407                         cdev_index++;
 408                 i++;
 409         }
 410
 411         print_test_dynamics(test_data);
 412
 413         while (test_data->level <= test_data->level_lst.max) {
 414
 415                 i = 0;
 416                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 417
 418                         if (i == total_nb_qps)
 419                                 break;
 420
 421                         rte_eal_remote_launch(
 422                                         cperf_testmap[test_data->test].runner,
 423                                         ctx[i], lcore_id);
 424                         i++;
 425                 }
 426                 i = 0;
 427                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 428
 429                         if (i == total_nb_qps)
 430                                 break;
 431                         ret |= rte_eal_wait_lcore(lcore_id);
 432                         i++;
 433                 }
 434
 435                 if (ret != EXIT_SUCCESS)
 436                         break;
 437
 438                 if (test_data->level_lst.inc != 0)
 439                         test_data->level += test_data->level_lst.inc;
 440                 else {
 441                         if (++level_idx == test_data->level_lst.count)
 442                                 break;
 443                         test_data->level = test_data->level_lst.list[level_idx];
 444                 }
 445         }
 446
 447 end:
 448         switch (test_data->cleanup) {
 449
 450         case ST_DURING_TEST:
 451                 i = 0;
 452                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 453                         if (i == total_nb_qps)
 454                                 break;
 455
 456                         if (ctx[i] && cperf_testmap[test_data->test].destructor)
 457                                 cperf_testmap[test_data->test].destructor(
 458                                                                         ctx[i]);
 459                         i++;
 460                 }
 461                 /* fallthrough */
 462         case ST_INPUT_DATA:
 463                 rte_free(test_data->input_data);
 464                 /* fallthrough */
 465         case ST_COMPDEV:
 466                 for (i = 0; i < nb_compressdevs &&
 467                      i < RTE_COMPRESS_MAX_DEVS; i++) {
 468                         rte_compressdev_stop(enabled_cdevs[i]);
 469                         rte_compressdev_close(enabled_cdevs[i]);
 470                 }
 471                 /* fallthrough */
 472         case ST_TEST_DATA:
 473                 rte_free(test_data);
 474                 /* fallthrough */
 475         case ST_CLEAR:
 476         default:
 477                 i = rte_eal_cleanup();
 478                 if (i) {
 479                         RTE_LOG(ERR, USER1,
 480                                 "Error from rte_eal_cleanup(), %d\n", i);
 481                         ret = i;
 482                 }
 483                 break;
 484         }
 485         return ret;
 486 }
 487
 488 __rte_weak void *
 489 cperf_cyclecount_test_constructor(uint8_t dev_id __rte_unused,
 490                                  uint16_t qp_id __rte_unused,
 491                                  struct comp_test_data *options __rte_unused)
 492 {
 493         RTE_LOG(INFO, USER1, "Cycle count test is not supported yet\n");
 494         return NULL;
 495 }
 496
 497 __rte_weak void
 498 cperf_cyclecount_test_destructor(void *arg __rte_unused)
 499 {
 500         RTE_LOG(INFO, USER1, "Something wrong happened!!!\n");
 501 }
 502
 503 __rte_weak int
 504 cperf_cyclecount_test_runner(void *test_ctx __rte_unused)
 505 {
 506         return 0;
 507 }
 508
 509 __rte_weak void *
 510 cperf_throughput_test_constructor(uint8_t dev_id __rte_unused,
 511                                  uint16_t qp_id __rte_unused,
 512                                  struct comp_test_data *options __rte_unused)
 513 {
 514         RTE_LOG(INFO, USER1, "Benchmark test is not supported yet\n");
 515         return NULL;
 516 }
 517
 518 __rte_weak void
 519 cperf_throughput_test_destructor(void *arg __rte_unused)
 520 {
 521
 522 }
 523
 524 __rte_weak int
 525 cperf_throughput_test_runner(void *test_ctx __rte_unused)
 526 {
 527         return 0;
 528 }
 529 __rte_weak void *
 530 cperf_verify_test_constructor(uint8_t dev_id __rte_unused,
 531                                  uint16_t qp_id __rte_unused,
 532                                  struct comp_test_data *options __rte_unused)
 533 {
 534         RTE_LOG(INFO, USER1, "Verify test is not supported yet\n");
 535         return NULL;
 536 }
 537
 538 __rte_weak void
 539 cperf_verify_test_destructor(void *arg __rte_unused)
 540 {
 541
 542 }
 543
 544 __rte_weak int
 545 cperf_verify_test_runner(void *test_ctx __rte_unused)
 546 {
 547         return 0;
 548 }