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 encoding */
  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                 int 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                 test_data->nb_qps = config.nb_queue_pairs;
 198
 199                 if (rte_compressdev_configure(cdev_id, &config) < 0) {
 200                         RTE_LOG(ERR, USER1, "Device configuration failed\n");
 201                         return -EINVAL;
 202                 }
 203
 204                 for (j = 0; j < test_data->nb_qps; j++) {
 205                         ret = rte_compressdev_queue_pair_setup(cdev_id, j,
 206                                         NUM_MAX_INFLIGHT_OPS, socket_id);
 207                         if (ret < 0) {
 208                                 RTE_LOG(ERR, USER1,
 209                               "Failed to setup queue pair %u on compressdev %u",
 210                                         j, cdev_id);
 211                                 return -EINVAL;
 212                         }
 213                 }
 214
 215                 ret = rte_compressdev_start(cdev_id);
 216                 if (ret < 0) {
 217                         RTE_LOG(ERR, USER1,
 218                                 "Failed to start device %u: error %d\n",
 219                                 cdev_id, ret);
 220                         return -EPERM;
 221                 }
 222         }
 223
 224         return enabled_cdev_count;
 225 }
 226
 227 static int
 228 comp_perf_dump_input_data(struct comp_test_data *test_data)
 229 {
 230         FILE *f = fopen(test_data->input_file, "r");
 231         int ret = -1;
 232
 233         if (f == NULL) {
 234                 RTE_LOG(ERR, USER1, "Input file could not be opened\n");
 235                 return -1;
 236         }
 237
 238         if (fseek(f, 0, SEEK_END) != 0) {
 239                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 240                 goto end;
 241         }
 242         size_t actual_file_sz = ftell(f);
 243         /* If extended input data size has not been set,
 244          * input data size = file size
 245          */
 246
 247         if (test_data->input_data_sz == 0)
 248                 test_data->input_data_sz = actual_file_sz;
 249
 250         if (test_data->input_data_sz <= 0 || actual_file_sz <= 0 ||
 251                         fseek(f, 0, SEEK_SET) != 0) {
 252                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 253                 goto end;
 254         }
 255
 256         test_data->input_data = rte_zmalloc_socket(NULL,
 257                                 test_data->input_data_sz, 0, rte_socket_id());
 258
 259         if (test_data->input_data == NULL) {
 260                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 261                                 "file could not be allocated\n");
 262                 goto end;
 263         }
 264
 265         size_t remaining_data = test_data->input_data_sz;
 266         uint8_t *data = test_data->input_data;
 267
 268         while (remaining_data > 0) {
 269                 size_t data_to_read = RTE_MIN(remaining_data, actual_file_sz);
 270
 271                 if (fread(data, data_to_read, 1, f) != 1) {
 272                         RTE_LOG(ERR, USER1, "Input file could not be read\n");
 273                         goto end;
 274                 }
 275                 if (fseek(f, 0, SEEK_SET) != 0) {
 276                         RTE_LOG(ERR, USER1,
 277                                 "Size of input could not be calculated\n");
 278                         goto end;
 279                 }
 280                 remaining_data -= data_to_read;
 281                 data += data_to_read;
 282         }
 283
 284         printf("\n");
 285         if (test_data->input_data_sz > actual_file_sz)
 286                 RTE_LOG(INFO, USER1,
 287                   "%zu bytes read from file %s, extending the file %.2f times\n",
 288                         test_data->input_data_sz, test_data->input_file,
 289                         (double)test_data->input_data_sz/actual_file_sz);
 290         else
 291                 RTE_LOG(INFO, USER1,
 292                         "%zu bytes read from file %s\n",
 293                         test_data->input_data_sz, test_data->input_file);
 294
 295         ret = 0;
 296
 297 end:
 298         fclose(f);
 299         return ret;
 300 }
 301
 302 static void
 303 comp_perf_cleanup_on_signal(int signalNumber __rte_unused)
 304 {
 305         test_data->perf_comp_force_stop = 1;
 306 }
 307
 308 static void
 309 comp_perf_register_cleanup_on_signal(void)
 310 {
 311         signal(SIGTERM, comp_perf_cleanup_on_signal);
 312         signal(SIGINT, comp_perf_cleanup_on_signal);
 313 }
 314
 315 int
 316 main(int argc, char **argv)
 317 {
 318         uint8_t level_idx = 0;
 319         int ret, i;
 320         void *ctx[RTE_MAX_LCORE] = {};
 321         uint8_t enabled_cdevs[RTE_COMPRESS_MAX_DEVS];
 322         int nb_compressdevs = 0;
 323         uint16_t total_nb_qps = 0;
 324         uint8_t cdev_id;
 325         uint32_t lcore_id;
 326
 327         /* Initialise DPDK EAL */
 328         ret = rte_eal_init(argc, argv);
 329         if (ret < 0)
 330                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments!\n");
 331         argc -= ret;
 332         argv += ret;
 333
 334         test_data = rte_zmalloc_socket(NULL, sizeof(struct comp_test_data),
 335                                         0, rte_socket_id());
 336
 337         if (test_data == NULL)
 338                 rte_exit(EXIT_FAILURE, "Cannot reserve memory in socket %d\n",
 339                                 rte_socket_id());
 340
 341         comp_perf_register_cleanup_on_signal();
 342
 343         ret = EXIT_SUCCESS;
 344         test_data->cleanup = ST_TEST_DATA;
 345         comp_perf_options_default(test_data);
 346
 347         if (comp_perf_options_parse(test_data, argc, argv) < 0) {
 348                 RTE_LOG(ERR, USER1,
 349                         "Parsing one or more user options failed\n");
 350                 ret = EXIT_FAILURE;
 351                 goto end;
 352         }
 353
 354         if (comp_perf_options_check(test_data) < 0) {
 355                 ret = EXIT_FAILURE;
 356                 goto end;
 357         }
 358
 359         nb_compressdevs =
 360                 comp_perf_initialize_compressdev(test_data, enabled_cdevs);
 361
 362         if (nb_compressdevs < 1) {
 363                 ret = EXIT_FAILURE;
 364                 goto end;
 365         }
 366
 367         test_data->cleanup = ST_COMPDEV;
 368         if (comp_perf_dump_input_data(test_data) < 0) {
 369                 ret = EXIT_FAILURE;
 370                 goto end;
 371         }
 372
 373         test_data->cleanup = ST_INPUT_DATA;
 374
 375         if (test_data->level_lst.inc != 0)
 376                 test_data->level = test_data->level_lst.min;
 377         else
 378                 test_data->level = test_data->level_lst.list[0];
 379
 380         printf("\nApp uses socket: %u\n", rte_socket_id());
 381         printf("Burst size = %u\n", test_data->burst_sz);
 382         printf("Input data size = %zu\n", test_data->input_data_sz);
 383         if (test_data->test == CPERF_TEST_TYPE_PMDCC)
 384                 printf("Cycle-count delay = %u [us]\n",
 385                        test_data->cyclecount_delay);
 386
 387         test_data->cleanup = ST_DURING_TEST;
 388         total_nb_qps = nb_compressdevs * test_data->nb_qps;
 389
 390         i = 0;
 391         uint8_t qp_id = 0, cdev_index = 0;
 392
 393         RTE_LCORE_FOREACH_WORKER(lcore_id) {
 394
 395                 if (i == total_nb_qps)
 396                         break;
 397
 398                 cdev_id = enabled_cdevs[cdev_index];
 399                 ctx[i] = cperf_testmap[test_data->test].constructor(
 400                                                         cdev_id, qp_id,
 401                                                         test_data);
 402                 if (ctx[i] == NULL) {
 403                         RTE_LOG(ERR, USER1, "Test run constructor failed\n");
 404                         goto end;
 405                 }
 406                 qp_id = (qp_id + 1) % test_data->nb_qps;
 407                 if (qp_id == 0)
 408                         cdev_index++;
 409                 i++;
 410         }
 411
 412         print_test_dynamics(test_data);
 413
 414         while (test_data->level <= test_data->level_lst.max) {
 415
 416                 i = 0;
 417                 RTE_LCORE_FOREACH_WORKER(lcore_id) {
 418
 419                         if (i == total_nb_qps)
 420                                 break;
 421
 422                         rte_eal_remote_launch(
 423                                         cperf_testmap[test_data->test].runner,
 424                                         ctx[i], lcore_id);
 425                         i++;
 426                 }
 427                 i = 0;
 428                 RTE_LCORE_FOREACH_WORKER(lcore_id) {
 429
 430                         if (i == total_nb_qps)
 431                                 break;
 432                         ret |= rte_eal_wait_lcore(lcore_id);
 433                         i++;
 434                 }
 435
 436                 if (ret != EXIT_SUCCESS)
 437                         break;
 438
 439                 if (test_data->level_lst.inc != 0)
 440                         test_data->level += test_data->level_lst.inc;
 441                 else {
 442                         if (++level_idx == test_data->level_lst.count)
 443                                 break;
 444                         test_data->level = test_data->level_lst.list[level_idx];
 445                 }
 446         }
 447
 448 end:
 449         switch (test_data->cleanup) {
 450
 451         case ST_DURING_TEST:
 452                 i = 0;
 453                 RTE_LCORE_FOREACH_WORKER(lcore_id) {
 454                         if (i == total_nb_qps)
 455                                 break;
 456
 457                         if (ctx[i] && cperf_testmap[test_data->test].destructor)
 458                                 cperf_testmap[test_data->test].destructor(
 459                                                                         ctx[i]);
 460                         i++;
 461                 }
 462                 /* fallthrough */
 463         case ST_INPUT_DATA:
 464                 rte_free(test_data->input_data);
 465                 /* fallthrough */
 466         case ST_COMPDEV:
 467                 for (i = 0; i < nb_compressdevs &&
 468                      i < RTE_COMPRESS_MAX_DEVS; i++) {
 469                         rte_compressdev_stop(enabled_cdevs[i]);
 470                         rte_compressdev_close(enabled_cdevs[i]);
 471                 }
 472                 /* fallthrough */
 473         case ST_TEST_DATA:
 474                 rte_free(test_data);
 475                 /* fallthrough */
 476         case ST_CLEAR:
 477         default:
 478                 i = rte_eal_cleanup();
 479                 if (i) {
 480                         RTE_LOG(ERR, USER1,
 481                                 "Error from rte_eal_cleanup(), %d\n", i);
 482                         ret = i;
 483                 }
 484                 break;
 485         }
 486         return ret;
 487 }
 488
 489 __rte_weak void *
 490 cperf_cyclecount_test_constructor(uint8_t dev_id __rte_unused,
 491                                  uint16_t qp_id __rte_unused,
 492                                  struct comp_test_data *options __rte_unused)
 493 {
 494         RTE_LOG(INFO, USER1, "Cycle count test is not supported yet\n");
 495         return NULL;
 496 }
 497
 498 __rte_weak void
 499 cperf_cyclecount_test_destructor(void *arg __rte_unused)
 500 {
 501         RTE_LOG(INFO, USER1, "Something wrong happened!!!\n");
 502 }
 503
 504 __rte_weak int
 505 cperf_cyclecount_test_runner(void *test_ctx __rte_unused)
 506 {
 507         return 0;
 508 }
 509
 510 __rte_weak void *
 511 cperf_throughput_test_constructor(uint8_t dev_id __rte_unused,
 512                                  uint16_t qp_id __rte_unused,
 513                                  struct comp_test_data *options __rte_unused)
 514 {
 515         RTE_LOG(INFO, USER1, "Benchmark test is not supported yet\n");
 516         return NULL;
 517 }
 518
 519 __rte_weak void
 520 cperf_throughput_test_destructor(void *arg __rte_unused)
 521 {
 522
 523 }
 524
 525 __rte_weak int
 526 cperf_throughput_test_runner(void *test_ctx __rte_unused)
 527 {
 528         return 0;
 529 }
 530 __rte_weak void *
 531 cperf_verify_test_constructor(uint8_t dev_id __rte_unused,
 532                                  uint16_t qp_id __rte_unused,
 533                                  struct comp_test_data *options __rte_unused)
 534 {
 535         RTE_LOG(INFO, USER1, "Verify test is not supported yet\n");
 536         return NULL;
 537 }
 538
 539 __rte_weak void
 540 cperf_verify_test_destructor(void *arg __rte_unused)
 541 {
 542
 543 }
 544
 545 __rte_weak int
 546 cperf_verify_test_runner(void *test_ctx __rte_unused)
 547 {
 548         return 0;
 549 }