ceph/src/spdk/dpdk/app/test-compress-perf/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <rte_malloc.h>
   6 #include <rte_eal.h>
   7 #include <rte_log.h>
   8 #include <rte_compressdev.h>
   9
  10 #include "comp_perf_options.h"
  11 #include "comp_perf_test_verify.h"
  12 #include "comp_perf_test_benchmark.h"
  13
  14 #define NUM_MAX_XFORMS 16
  15 #define NUM_MAX_INFLIGHT_OPS 512
  16
  17 #define DIV_CEIL(a, b)  ((a) / (b) + ((a) % (b) != 0))
  18
  19 /* Cleanup state machine */
  20 static enum cleanup_st {
  21         ST_CLEAR = 0,
  22         ST_TEST_DATA,
  23         ST_COMPDEV,
  24         ST_INPUT_DATA,
  25         ST_MEMORY_ALLOC,
  26         ST_PREPARE_BUF,
  27         ST_DURING_TEST
  28 } cleanup = ST_CLEAR;
  29
  30 static int
  31 param_range_check(uint16_t size, const struct rte_param_log2_range *range)
  32 {
  33         unsigned int next_size;
  34
  35         /* Check lower/upper bounds */
  36         if (size < range->min)
  37                 return -1;
  38
  39         if (size > range->max)
  40                 return -1;
  41
  42         /* If range is actually only one value, size is correct */
  43         if (range->increment == 0)
  44                 return 0;
  45
  46         /* Check if value is one of the supported sizes */
  47         for (next_size = range->min; next_size <= range->max;
  48                         next_size += range->increment)
  49                 if (size == next_size)
  50                         return 0;
  51
  52         return -1;
  53 }
  54
  55 static int
  56 comp_perf_check_capabilities(struct comp_test_data *test_data)
  57 {
  58         const struct rte_compressdev_capabilities *cap;
  59
  60         cap = rte_compressdev_capability_get(test_data->cdev_id,
  61                                              RTE_COMP_ALGO_DEFLATE);
  62
  63         if (cap == NULL) {
  64                 RTE_LOG(ERR, USER1,
  65                         "Compress device does not support DEFLATE\n");
  66                 return -1;
  67         }
  68
  69         uint64_t comp_flags = cap->comp_feature_flags;
  70
  71         /* Huffman enconding */
  72         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_FIXED &&
  73                         (comp_flags & RTE_COMP_FF_HUFFMAN_FIXED) == 0) {
  74                 RTE_LOG(ERR, USER1,
  75                         "Compress device does not supported Fixed Huffman\n");
  76                 return -1;
  77         }
  78
  79         if (test_data->huffman_enc == RTE_COMP_HUFFMAN_DYNAMIC &&
  80                         (comp_flags & RTE_COMP_FF_HUFFMAN_DYNAMIC) == 0) {
  81                 RTE_LOG(ERR, USER1,
  82                         "Compress device does not supported Dynamic Huffman\n");
  83                 return -1;
  84         }
  85
  86         /* Window size */
  87         if (test_data->window_sz != -1) {
  88                 if (param_range_check(test_data->window_sz, &cap->window_size)
  89                                 < 0) {
  90                         RTE_LOG(ERR, USER1,
  91                                 "Compress device does not support "
  92                                 "this window size\n");
  93                         return -1;
  94                 }
  95         } else
  96                 /* Set window size to PMD maximum if none was specified */
  97                 test_data->window_sz = cap->window_size.max;
  98
  99         /* Check if chained mbufs is supported */
 100         if (test_data->max_sgl_segs > 1  &&
 101                         (comp_flags & RTE_COMP_FF_OOP_SGL_IN_SGL_OUT) == 0) {
 102                 RTE_LOG(INFO, USER1, "Compress device does not support "
 103                                 "chained mbufs. Max SGL segments set to 1\n");
 104                 test_data->max_sgl_segs = 1;
 105         }
 106
 107         /* Level 0 support */
 108         if (test_data->level.min == 0 &&
 109                         (comp_flags & RTE_COMP_FF_NONCOMPRESSED_BLOCKS) == 0) {
 110                 RTE_LOG(ERR, USER1, "Compress device does not support "
 111                                 "level 0 (no compression)\n");
 112                 return -1;
 113         }
 114
 115         return 0;
 116 }
 117
 118 static uint32_t
 119 find_buf_size(uint32_t input_size)
 120 {
 121         uint32_t i;
 122
 123         /* From performance point of view the buffer size should be a
 124          * power of 2 but also should be enough to store incompressible data
 125          */
 126
 127         /* We're looking for nearest power of 2 buffer size, which is greather
 128          * than input_size
 129          */
 130         uint32_t size =
 131                 !input_size ? MIN_COMPRESSED_BUF_SIZE : (input_size << 1);
 132
 133         for (i = UINT16_MAX + 1; !(i & size); i >>= 1)
 134                 ;
 135
 136         return i > ((UINT16_MAX + 1) >> 1)
 137                         ? (uint32_t)((float)input_size * EXPANSE_RATIO)
 138                         : i;
 139 }
 140
 141 static int
 142 comp_perf_allocate_memory(struct comp_test_data *test_data)
 143 {
 144
 145         test_data->out_seg_sz = find_buf_size(test_data->seg_sz);
 146         /* Number of segments for input and output
 147          * (compression and decompression)
 148          */
 149         uint32_t total_segs = DIV_CEIL(test_data->input_data_sz,
 150                         test_data->seg_sz);
 151         test_data->comp_buf_pool = rte_pktmbuf_pool_create("comp_buf_pool",
 152                                 total_segs,
 153                                 0, 0,
 154                                 test_data->out_seg_sz + RTE_PKTMBUF_HEADROOM,
 155                                 rte_socket_id());
 156         if (test_data->comp_buf_pool == NULL) {
 157                 RTE_LOG(ERR, USER1, "Mbuf mempool could not be created\n");
 158                 return -1;
 159         }
 160
 161         cleanup = ST_MEMORY_ALLOC;
 162         test_data->decomp_buf_pool = rte_pktmbuf_pool_create("decomp_buf_pool",
 163                                 total_segs,
 164                                 0, 0, test_data->seg_sz + RTE_PKTMBUF_HEADROOM,
 165                                 rte_socket_id());
 166         if (test_data->decomp_buf_pool == NULL) {
 167                 RTE_LOG(ERR, USER1, "Mbuf mempool could not be created\n");
 168                 return -1;
 169         }
 170
 171         test_data->total_bufs = DIV_CEIL(total_segs, test_data->max_sgl_segs);
 172
 173         test_data->op_pool = rte_comp_op_pool_create("op_pool",
 174                                   test_data->total_bufs,
 175                                   0, 0, rte_socket_id());
 176         if (test_data->op_pool == NULL) {
 177                 RTE_LOG(ERR, USER1, "Comp op mempool could not be created\n");
 178                 return -1;
 179         }
 180
 181         /*
 182          * Compressed data might be a bit larger than input data,
 183          * if data cannot be compressed
 184          */
 185         test_data->compressed_data = rte_zmalloc_socket(NULL,
 186                                 test_data->input_data_sz * EXPANSE_RATIO
 187                                                 + MIN_COMPRESSED_BUF_SIZE, 0,
 188                                 rte_socket_id());
 189         if (test_data->compressed_data == NULL) {
 190                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 191                                 "file could not be allocated\n");
 192                 return -1;
 193         }
 194
 195         test_data->decompressed_data = rte_zmalloc_socket(NULL,
 196                                 test_data->input_data_sz, 0,
 197                                 rte_socket_id());
 198         if (test_data->decompressed_data == NULL) {
 199                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 200                                 "file could not be allocated\n");
 201                 return -1;
 202         }
 203
 204         test_data->comp_bufs = rte_zmalloc_socket(NULL,
 205                         test_data->total_bufs * sizeof(struct rte_mbuf *),
 206                         0, rte_socket_id());
 207         if (test_data->comp_bufs == NULL) {
 208                 RTE_LOG(ERR, USER1, "Memory to hold the compression mbufs"
 209                                 " could not be allocated\n");
 210                 return -1;
 211         }
 212
 213         test_data->decomp_bufs = rte_zmalloc_socket(NULL,
 214                         test_data->total_bufs * sizeof(struct rte_mbuf *),
 215                         0, rte_socket_id());
 216         if (test_data->decomp_bufs == NULL) {
 217                 RTE_LOG(ERR, USER1, "Memory to hold the decompression mbufs"
 218                                 " could not be allocated\n");
 219                 return -1;
 220         }
 221         return 0;
 222 }
 223
 224 static int
 225 comp_perf_dump_input_data(struct comp_test_data *test_data)
 226 {
 227         FILE *f = fopen(test_data->input_file, "r");
 228         int ret = -1;
 229
 230         if (f == NULL) {
 231                 RTE_LOG(ERR, USER1, "Input file could not be opened\n");
 232                 return -1;
 233         }
 234
 235         if (fseek(f, 0, SEEK_END) != 0) {
 236                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 237                 goto end;
 238         }
 239         size_t actual_file_sz = ftell(f);
 240         /* If extended input data size has not been set,
 241          * input data size = file size
 242          */
 243
 244         if (test_data->input_data_sz == 0)
 245                 test_data->input_data_sz = actual_file_sz;
 246
 247         if (fseek(f, 0, SEEK_SET) != 0) {
 248                 RTE_LOG(ERR, USER1, "Size of input could not be calculated\n");
 249                 goto end;
 250         }
 251
 252         test_data->input_data = rte_zmalloc_socket(NULL,
 253                                 test_data->input_data_sz, 0, rte_socket_id());
 254
 255         if (test_data->input_data == NULL) {
 256                 RTE_LOG(ERR, USER1, "Memory to hold the data from the input "
 257                                 "file could not be allocated\n");
 258                 goto end;
 259         }
 260
 261         size_t remaining_data = test_data->input_data_sz;
 262         uint8_t *data = test_data->input_data;
 263
 264         while (remaining_data > 0) {
 265                 size_t data_to_read = RTE_MIN(remaining_data, actual_file_sz);
 266
 267                 if (fread(data, data_to_read, 1, f) != 1) {
 268                         RTE_LOG(ERR, USER1, "Input file could not be read\n");
 269                         goto end;
 270                 }
 271                 if (fseek(f, 0, SEEK_SET) != 0) {
 272                         RTE_LOG(ERR, USER1,
 273                                 "Size of input could not be calculated\n");
 274                         goto end;
 275                 }
 276                 remaining_data -= data_to_read;
 277                 data += data_to_read;
 278         }
 279
 280         if (test_data->input_data_sz > actual_file_sz)
 281                 RTE_LOG(INFO, USER1,
 282                   "%zu bytes read from file %s, extending the file %.2f times\n",
 283                         test_data->input_data_sz, test_data->input_file,
 284                         (double)test_data->input_data_sz/actual_file_sz);
 285         else
 286                 RTE_LOG(INFO, USER1,
 287                         "%zu bytes read from file %s\n",
 288                         test_data->input_data_sz, test_data->input_file);
 289
 290         ret = 0;
 291
 292 end:
 293         fclose(f);
 294         return ret;
 295 }
 296
 297 static int
 298 comp_perf_initialize_compressdev(struct comp_test_data *test_data)
 299 {
 300         uint8_t enabled_cdev_count;
 301         uint8_t enabled_cdevs[RTE_COMPRESS_MAX_DEVS];
 302
 303         enabled_cdev_count = rte_compressdev_devices_get(test_data->driver_name,
 304                         enabled_cdevs, RTE_COMPRESS_MAX_DEVS);
 305         if (enabled_cdev_count == 0) {
 306                 RTE_LOG(ERR, USER1, "No compress devices type %s available\n",
 307                                 test_data->driver_name);
 308                 return -EINVAL;
 309         }
 310
 311         if (enabled_cdev_count > 1)
 312                 RTE_LOG(INFO, USER1,
 313                         "Only the first compress device will be used\n");
 314
 315         test_data->cdev_id = enabled_cdevs[0];
 316
 317         if (comp_perf_check_capabilities(test_data) < 0)
 318                 return -1;
 319
 320         /* Configure compressdev (one device, one queue pair) */
 321         struct rte_compressdev_config config = {
 322                 .socket_id = rte_socket_id(),
 323                 .nb_queue_pairs = 1,
 324                 .max_nb_priv_xforms = NUM_MAX_XFORMS,
 325                 .max_nb_streams = 0
 326         };
 327
 328         if (rte_compressdev_configure(test_data->cdev_id, &config) < 0) {
 329                 RTE_LOG(ERR, USER1, "Device configuration failed\n");
 330                 return -1;
 331         }
 332
 333         if (rte_compressdev_queue_pair_setup(test_data->cdev_id, 0,
 334                         NUM_MAX_INFLIGHT_OPS, rte_socket_id()) < 0) {
 335                 RTE_LOG(ERR, USER1, "Queue pair setup failed\n");
 336                 return -1;
 337         }
 338
 339         if (rte_compressdev_start(test_data->cdev_id) < 0) {
 340                 RTE_LOG(ERR, USER1, "Device could not be started\n");
 341                 return -1;
 342         }
 343
 344         return 0;
 345 }
 346
 347 static int
 348 prepare_bufs(struct comp_test_data *test_data)
 349 {
 350         uint32_t remaining_data = test_data->input_data_sz;
 351         uint8_t *input_data_ptr = test_data->input_data;
 352         size_t data_sz;
 353         uint8_t *data_addr;
 354         uint32_t i, j;
 355
 356         for (i = 0; i < test_data->total_bufs; i++) {
 357                 /* Allocate data in input mbuf and copy data from input file */
 358                 test_data->decomp_bufs[i] =
 359                         rte_pktmbuf_alloc(test_data->decomp_buf_pool);
 360                 if (test_data->decomp_bufs[i] == NULL) {
 361                         RTE_LOG(ERR, USER1, "Could not allocate mbuf\n");
 362                         return -1;
 363                 }
 364
 365                 cleanup = ST_PREPARE_BUF;
 366                 data_sz = RTE_MIN(remaining_data, test_data->seg_sz);
 367                 data_addr = (uint8_t *) rte_pktmbuf_append(
 368                                         test_data->decomp_bufs[i], data_sz);
 369                 if (data_addr == NULL) {
 370                         RTE_LOG(ERR, USER1, "Could not append data\n");
 371                         return -1;
 372                 }
 373                 rte_memcpy(data_addr, input_data_ptr, data_sz);
 374
 375                 input_data_ptr += data_sz;
 376                 remaining_data -= data_sz;
 377
 378                 /* Already one segment in the mbuf */
 379                 uint16_t segs_per_mbuf = 1;
 380
 381                 /* Chain mbufs if needed for input mbufs */
 382                 while (segs_per_mbuf < test_data->max_sgl_segs
 383                                 && remaining_data > 0) {
 384                         struct rte_mbuf *next_seg =
 385                                 rte_pktmbuf_alloc(test_data->decomp_buf_pool);
 386
 387                         if (next_seg == NULL) {
 388                                 RTE_LOG(ERR, USER1,
 389                                         "Could not allocate mbuf\n");
 390                                 return -1;
 391                         }
 392
 393                         data_sz = RTE_MIN(remaining_data, test_data->seg_sz);
 394                         data_addr = (uint8_t *)rte_pktmbuf_append(next_seg,
 395                                 data_sz);
 396
 397                         if (data_addr == NULL) {
 398                                 RTE_LOG(ERR, USER1, "Could not append data\n");
 399                                 return -1;
 400                         }
 401
 402                         rte_memcpy(data_addr, input_data_ptr, data_sz);
 403                         input_data_ptr += data_sz;
 404                         remaining_data -= data_sz;
 405
 406                         if (rte_pktmbuf_chain(test_data->decomp_bufs[i],
 407                                         next_seg) < 0) {
 408                                 RTE_LOG(ERR, USER1, "Could not chain mbufs\n");
 409                                 return -1;
 410                         }
 411                         segs_per_mbuf++;
 412                 }
 413
 414                 /* Allocate data in output mbuf */
 415                 test_data->comp_bufs[i] =
 416                         rte_pktmbuf_alloc(test_data->comp_buf_pool);
 417                 if (test_data->comp_bufs[i] == NULL) {
 418                         RTE_LOG(ERR, USER1, "Could not allocate mbuf\n");
 419                         return -1;
 420                 }
 421                 data_addr = (uint8_t *) rte_pktmbuf_append(
 422                                         test_data->comp_bufs[i],
 423                                         test_data->out_seg_sz);
 424                 if (data_addr == NULL) {
 425                         RTE_LOG(ERR, USER1, "Could not append data\n");
 426                         return -1;
 427                 }
 428
 429                 /* Chain mbufs if needed for output mbufs */
 430                 for (j = 1; j < segs_per_mbuf; j++) {
 431                         struct rte_mbuf *next_seg =
 432                                 rte_pktmbuf_alloc(test_data->comp_buf_pool);
 433
 434                         if (next_seg == NULL) {
 435                                 RTE_LOG(ERR, USER1,
 436                                         "Could not allocate mbuf\n");
 437                                 return -1;
 438                         }
 439
 440                         data_addr = (uint8_t *)rte_pktmbuf_append(next_seg,
 441                                 test_data->out_seg_sz);
 442
 443                         if (data_addr == NULL) {
 444                                 RTE_LOG(ERR, USER1, "Could not append data\n");
 445                                 return -1;
 446                         }
 447
 448                         if (rte_pktmbuf_chain(test_data->comp_bufs[i],
 449                                         next_seg) < 0) {
 450                                 RTE_LOG(ERR, USER1, "Could not chain mbufs\n");
 451                                 return -1;
 452                         }
 453                 }
 454         }
 455
 456         return 0;
 457 }
 458
 459 static void
 460 free_bufs(struct comp_test_data *test_data)
 461 {
 462         uint32_t i;
 463
 464         for (i = 0; i < test_data->total_bufs; i++) {
 465                 rte_pktmbuf_free(test_data->comp_bufs[i]);
 466                 rte_pktmbuf_free(test_data->decomp_bufs[i]);
 467         }
 468 }
 469
 470
 471
 472 int
 473 main(int argc, char **argv)
 474 {
 475         uint8_t level, level_idx = 0;
 476         int ret, i;
 477         struct comp_test_data *test_data;
 478
 479         /* Initialise DPDK EAL */
 480         ret = rte_eal_init(argc, argv);
 481         if (ret < 0)
 482                 rte_exit(EXIT_FAILURE, "Invalid EAL arguments!\n");
 483         argc -= ret;
 484         argv += ret;
 485
 486         test_data = rte_zmalloc_socket(NULL, sizeof(struct comp_test_data),
 487                                         0, rte_socket_id());
 488
 489         if (test_data == NULL)
 490                 rte_exit(EXIT_FAILURE, "Cannot reserve memory in socket %d\n",
 491                                 rte_socket_id());
 492
 493         ret = EXIT_SUCCESS;
 494         cleanup = ST_TEST_DATA;
 495         comp_perf_options_default(test_data);
 496
 497         if (comp_perf_options_parse(test_data, argc, argv) < 0) {
 498                 RTE_LOG(ERR, USER1,
 499                         "Parsing one or more user options failed\n");
 500                 ret = EXIT_FAILURE;
 501                 goto end;
 502         }
 503
 504         if (comp_perf_options_check(test_data) < 0) {
 505                 ret = EXIT_FAILURE;
 506                 goto end;
 507         }
 508
 509         if (comp_perf_initialize_compressdev(test_data) < 0) {
 510                 ret = EXIT_FAILURE;
 511                 goto end;
 512         }
 513
 514         cleanup = ST_COMPDEV;
 515         if (comp_perf_dump_input_data(test_data) < 0) {
 516                 ret = EXIT_FAILURE;
 517                 goto end;
 518         }
 519
 520         cleanup = ST_INPUT_DATA;
 521         if (comp_perf_allocate_memory(test_data) < 0) {
 522                 ret = EXIT_FAILURE;
 523                 goto end;
 524         }
 525
 526         if (prepare_bufs(test_data) < 0) {
 527                 ret = EXIT_FAILURE;
 528                 goto end;
 529         }
 530
 531         if (test_data->level.inc != 0)
 532                 level = test_data->level.min;
 533         else
 534                 level = test_data->level.list[0];
 535
 536         printf("Burst size = %u\n", test_data->burst_sz);
 537         printf("File size = %zu\n", test_data->input_data_sz);
 538
 539         printf("%6s%12s%17s%19s%21s%15s%21s%23s%16s\n",
 540                 "Level", "Comp size", "Comp ratio [%]",
 541                 "Comp [Cycles/it]", "Comp [Cycles/Byte]", "Comp [Gbps]",
 542                 "Decomp [Cycles/it]", "Decomp [Cycles/Byte]", "Decomp [Gbps]");
 543
 544         cleanup = ST_DURING_TEST;
 545         while (level <= test_data->level.max) {
 546
 547                 /*
 548                  * Run a first iteration, to verify compression and
 549                  * get the compression ratio for the level
 550                  */
 551                 if (cperf_verification(test_data, level) != EXIT_SUCCESS)
 552                         break;
 553
 554                 /*
 555                  * Run benchmarking test
 556                  */
 557                 if (cperf_benchmark(test_data, level) != EXIT_SUCCESS)
 558                         break;
 559
 560                 printf("%6u%12zu%17.2f%19"PRIu64"%21.2f"
 561                                         "%15.2f%21"PRIu64"%23.2f%16.2f\n",
 562                        level, test_data->comp_data_sz, test_data->ratio,
 563                        test_data->comp_tsc_duration[level],
 564                        test_data->comp_tsc_byte, test_data->comp_gbps,
 565                        test_data->decomp_tsc_duration[level],
 566                        test_data->decomp_tsc_byte, test_data->decomp_gbps);
 567
 568                 if (test_data->level.inc != 0)
 569                         level += test_data->level.inc;
 570                 else {
 571                         if (++level_idx == test_data->level.count)
 572                                 break;
 573                         level = test_data->level.list[level_idx];
 574                 }
 575         }
 576
 577 end:
 578         switch (cleanup) {
 579
 580         case ST_DURING_TEST:
 581         case ST_PREPARE_BUF:
 582                 free_bufs(test_data);
 583                 /* fallthrough */
 584         case ST_MEMORY_ALLOC:
 585                 rte_free(test_data->decomp_bufs);
 586                 rte_free(test_data->comp_bufs);
 587                 rte_free(test_data->decompressed_data);
 588                 rte_free(test_data->compressed_data);
 589                 rte_mempool_free(test_data->op_pool);
 590                 rte_mempool_free(test_data->decomp_buf_pool);
 591                 rte_mempool_free(test_data->comp_buf_pool);
 592                 /* fallthrough */
 593         case ST_INPUT_DATA:
 594                 rte_free(test_data->input_data);
 595                 /* fallthrough */
 596         case ST_COMPDEV:
 597                 if (test_data->cdev_id != -1)
 598                         rte_compressdev_stop(test_data->cdev_id);
 599                 /* fallthrough */
 600         case ST_TEST_DATA:
 601                 rte_free(test_data);
 602                 /* fallthrough */
 603         case ST_CLEAR:
 604         default:
 605                 i = rte_eal_cleanup();
 606                 if (i) {
 607                         RTE_LOG(ERR, USER1,
 608                                 "Error from rte_eal_cleanup(), %d\n", i);
 609                         ret = i;
 610                 }
 611                 break;
 612         }
 613         return ret;
 614 }