1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "env_internal.h" 37 38 #include "spdk/version.h" 39 #include "spdk/env_dpdk.h" 40 41 #include <rte_config.h> 42 #include <rte_eal.h> 43 #include <rte_errno.h> 44 45 #define SPDK_ENV_DPDK_DEFAULT_NAME "spdk" 46 #define SPDK_ENV_DPDK_DEFAULT_SHM_ID -1 47 #define SPDK_ENV_DPDK_DEFAULT_MEM_SIZE -1 48 #define SPDK_ENV_DPDK_DEFAULT_MASTER_CORE -1 49 #define SPDK_ENV_DPDK_DEFAULT_MEM_CHANNEL -1 50 #define SPDK_ENV_DPDK_DEFAULT_CORE_MASK "0x1" 51 52 static char **g_eal_cmdline; 53 static int g_eal_cmdline_argcount; 54 static bool g_external_init = true; 55 56 static char * 57 _sprintf_alloc(const char *format, ...) 58 { 59 va_list args; 60 va_list args_copy; 61 char *buf; 62 size_t bufsize; 63 int rc; 64 65 va_start(args, format); 66 67 /* Try with a small buffer first. */ 68 bufsize = 32; 69 70 /* Limit maximum buffer size to something reasonable so we don't loop forever. */ 71 while (bufsize <= 1024 * 1024) { 72 buf = malloc(bufsize); 73 if (buf == NULL) { 74 va_end(args); 75 return NULL; 76 } 77 78 va_copy(args_copy, args); 79 rc = vsnprintf(buf, bufsize, format, args_copy); 80 va_end(args_copy); 81 82 /* 83 * If vsnprintf() returned a count within our current buffer size, we are done. 84 * The count does not include the \0 terminator, so rc == bufsize is not OK. 85 */ 86 if (rc >= 0 && (size_t)rc < bufsize) { 87 va_end(args); 88 return buf; 89 } 90 91 /* 92 * vsnprintf() should return the required space, but some libc versions do not 93 * implement this correctly, so just double the buffer size and try again. 94 * 95 * We don't need the data in buf, so rather than realloc(), use free() and malloc() 96 * again to avoid a copy. 97 */ 98 free(buf); 99 bufsize *= 2; 100 } 101 102 va_end(args); 103 return NULL; 104 } 105 106 static void 107 spdk_env_unlink_shared_files(void) 108 { 109 /* Starting with DPDK 18.05, there are more files with unpredictable paths 110 * and filenames. The --no-shconf option prevents from creating them, but 111 * only for DPDK 18.08+. For DPDK 18.05 we just leave them be. 112 */ 113 #if RTE_VERSION < RTE_VERSION_NUM(18, 05, 0, 0) 114 char buffer[PATH_MAX]; 115 116 snprintf(buffer, PATH_MAX, "/var/run/.spdk_pid%d_hugepage_info", getpid()); 117 if (unlink(buffer)) { 118 fprintf(stderr, "Unable to unlink shared memory file: %s. Error code: %d\n", buffer, errno); 119 } 120 #endif 121 } 122 123 void 124 spdk_env_opts_init(struct spdk_env_opts *opts) 125 { 126 if (!opts) { 127 return; 128 } 129 130 memset(opts, 0, sizeof(*opts)); 131 132 opts->name = SPDK_ENV_DPDK_DEFAULT_NAME; 133 opts->core_mask = SPDK_ENV_DPDK_DEFAULT_CORE_MASK; 134 opts->shm_id = SPDK_ENV_DPDK_DEFAULT_SHM_ID; 135 opts->mem_size = SPDK_ENV_DPDK_DEFAULT_MEM_SIZE; 136 opts->master_core = SPDK_ENV_DPDK_DEFAULT_MASTER_CORE; 137 opts->mem_channel = SPDK_ENV_DPDK_DEFAULT_MEM_CHANNEL; 138 } 139 140 static void 141 spdk_free_args(char **args, int argcount) 142 { 143 int i; 144 145 for (i = 0; i < argcount; i++) { 146 free(args[i]); 147 } 148 149 if (argcount) { 150 free(args); 151 } 152 } 153 154 static char ** 155 spdk_push_arg(char *args[], int *argcount, char *arg) 156 { 157 char **tmp; 158 159 if (arg == NULL) { 160 fprintf(stderr, "%s: NULL arg supplied\n", __func__); 161 spdk_free_args(args, *argcount); 162 return NULL; 163 } 164 165 tmp = realloc(args, sizeof(char *) * (*argcount + 1)); 166 if (tmp == NULL) { 167 free(arg); 168 spdk_free_args(args, *argcount); 169 return NULL; 170 } 171 172 tmp[*argcount] = arg; 173 (*argcount)++; 174 175 return tmp; 176 } 177 178 #if defined(__linux__) && defined(__x86_64__) 179 180 /* TODO: Can likely get this value from rlimits in the future */ 181 #define SPDK_IOMMU_VA_REQUIRED_WIDTH 48 182 #define VTD_CAP_MGAW_SHIFT 16 183 #define VTD_CAP_MGAW_MASK (0x3F << VTD_CAP_MGAW_SHIFT) 184 185 static int 186 spdk_get_iommu_width(void) 187 { 188 DIR *dir; 189 FILE *file; 190 struct dirent *entry; 191 char mgaw_path[64]; 192 char buf[64]; 193 char *end; 194 long long int val; 195 int width, tmp; 196 197 dir = opendir("/sys/devices/virtual/iommu/"); 198 if (dir == NULL) { 199 return -EINVAL; 200 } 201 202 width = 0; 203 204 while ((entry = readdir(dir)) != NULL) { 205 /* Find directories named "dmar0", "dmar1", etc */ 206 if (strncmp(entry->d_name, "dmar", sizeof("dmar") - 1) != 0) { 207 continue; 208 } 209 210 tmp = snprintf(mgaw_path, sizeof(mgaw_path), "/sys/devices/virtual/iommu/%s/intel-iommu/cap", 211 entry->d_name); 212 if ((unsigned)tmp >= sizeof(mgaw_path)) { 213 continue; 214 } 215 216 file = fopen(mgaw_path, "r"); 217 if (file == NULL) { 218 continue; 219 } 220 221 if (fgets(buf, sizeof(buf), file) == NULL) { 222 fclose(file); 223 continue; 224 } 225 226 val = strtoll(buf, &end, 16); 227 if (val == LLONG_MIN || val == LLONG_MAX) { 228 fclose(file); 229 continue; 230 } 231 232 tmp = ((val & VTD_CAP_MGAW_MASK) >> VTD_CAP_MGAW_SHIFT) + 1; 233 if (width == 0 || tmp < width) { 234 width = tmp; 235 } 236 237 fclose(file); 238 } 239 240 closedir(dir); 241 242 return width; 243 } 244 245 #endif 246 247 static int 248 spdk_build_eal_cmdline(const struct spdk_env_opts *opts) 249 { 250 int argcount = 0; 251 char **args; 252 253 args = NULL; 254 255 /* set the program name */ 256 args = spdk_push_arg(args, &argcount, _sprintf_alloc("%s", opts->name)); 257 if (args == NULL) { 258 return -1; 259 } 260 261 /* disable shared configuration files when in single process mode. This allows for cleaner shutdown */ 262 if (opts->shm_id < 0) { 263 args = spdk_push_arg(args, &argcount, _sprintf_alloc("%s", "--no-shconf")); 264 if (args == NULL) { 265 return -1; 266 } 267 } 268 269 /* set the coremask */ 270 /* NOTE: If coremask starts with '[' and ends with ']' it is a core list 271 */ 272 if (opts->core_mask[0] == '[') { 273 char *l_arg = _sprintf_alloc("-l %s", opts->core_mask + 1); 274 int len = strlen(l_arg); 275 if (l_arg[len - 1] == ']') { 276 l_arg[len - 1] = '\0'; 277 } 278 args = spdk_push_arg(args, &argcount, l_arg); 279 } else { 280 args = spdk_push_arg(args, &argcount, _sprintf_alloc("-c %s", opts->core_mask)); 281 } 282 283 if (args == NULL) { 284 return -1; 285 } 286 287 /* set the memory channel number */ 288 if (opts->mem_channel > 0) { 289 args = spdk_push_arg(args, &argcount, _sprintf_alloc("-n %d", opts->mem_channel)); 290 if (args == NULL) { 291 return -1; 292 } 293 } 294 295 /* set the memory size */ 296 if (opts->mem_size >= 0) { 297 args = spdk_push_arg(args, &argcount, _sprintf_alloc("-m %d", opts->mem_size)); 298 if (args == NULL) { 299 return -1; 300 } 301 } 302 303 /* set the master core */ 304 if (opts->master_core > 0) { 305 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--master-lcore=%d", 306 opts->master_core)); 307 if (args == NULL) { 308 return -1; 309 } 310 } 311 312 /* set no pci if enabled */ 313 if (opts->no_pci) { 314 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--no-pci")); 315 if (args == NULL) { 316 return -1; 317 } 318 } 319 320 /* create just one hugetlbfs file */ 321 if (opts->hugepage_single_segments) { 322 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--single-file-segments")); 323 if (args == NULL) { 324 return -1; 325 } 326 } 327 328 /* unlink hugepages after initialization */ 329 if (opts->unlink_hugepage) { 330 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--huge-unlink")); 331 if (args == NULL) { 332 return -1; 333 } 334 } 335 336 /* use a specific hugetlbfs mount */ 337 if (opts->hugedir) { 338 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--huge-dir=%s", opts->hugedir)); 339 if (args == NULL) { 340 return -1; 341 } 342 } 343 344 #if RTE_VERSION >= RTE_VERSION_NUM(18, 05, 0, 0) && RTE_VERSION < RTE_VERSION_NUM(18, 5, 1, 0) 345 /* Dynamic memory management is buggy in DPDK 18.05.0. Don't use it. */ 346 if (!opts->env_context || strstr(opts->env_context, "--legacy-mem") == NULL) { 347 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--legacy-mem")); 348 if (args == NULL) { 349 return -1; 350 } 351 } 352 #endif 353 354 if (opts->num_pci_addr) { 355 size_t i; 356 char bdf[32]; 357 struct spdk_pci_addr *pci_addr = 358 opts->pci_blacklist ? opts->pci_blacklist : opts->pci_whitelist; 359 360 for (i = 0; i < opts->num_pci_addr; i++) { 361 spdk_pci_addr_fmt(bdf, 32, &pci_addr[i]); 362 args = spdk_push_arg(args, &argcount, _sprintf_alloc("%s=%s", 363 (opts->pci_blacklist ? "--pci-blacklist" : "--pci-whitelist"), 364 bdf)); 365 if (args == NULL) { 366 return -1; 367 } 368 } 369 } 370 371 /* The following log-level options are not understood by older DPDKs */ 372 #if RTE_VERSION >= RTE_VERSION_NUM(18, 05, 0, 0) 373 /* Lower default EAL loglevel to RTE_LOG_NOTICE - normal, but significant messages. 374 * This can be overridden by specifying the same option in opts->env_context 375 */ 376 args = spdk_push_arg(args, &argcount, strdup("--log-level=lib.eal:6")); 377 if (args == NULL) { 378 return -1; 379 } 380 381 /* Lower default CRYPTO loglevel to RTE_LOG_ERR to avoid a ton of init msgs. 382 * This can be overridden by specifying the same option in opts->env_context 383 */ 384 args = spdk_push_arg(args, &argcount, strdup("--log-level=lib.cryptodev:5")); 385 if (args == NULL) { 386 return -1; 387 } 388 389 /* `user1` log type is used by rte_vhost, which prints an INFO log for each received 390 * vhost user message. We don't want that. The same log type is also used by a couple 391 * of other DPDK libs, but none of which we make use right now. If necessary, this can 392 * be overridden via opts->env_context. 393 */ 394 args = spdk_push_arg(args, &argcount, strdup("--log-level=user1:6")); 395 if (args == NULL) { 396 return -1; 397 } 398 #endif 399 400 if (opts->env_context) { 401 args = spdk_push_arg(args, &argcount, strdup(opts->env_context)); 402 if (args == NULL) { 403 return -1; 404 } 405 } 406 407 #ifdef __linux__ 408 409 #if defined(__x86_64__) 410 /* DPDK by default guesses that it should be using iova-mode=va so that it can 411 * support running as an unprivileged user. However, some systems (especially 412 * virtual machines) don't have an IOMMU capable of handling the full virtual 413 * address space and DPDK doesn't currently catch that. Add a check in SPDK 414 * and force iova-mode=pa here. */ 415 if (spdk_get_iommu_width() < SPDK_IOMMU_VA_REQUIRED_WIDTH) { 416 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--iova-mode=pa")); 417 if (args == NULL) { 418 return -1; 419 } 420 } 421 #elif defined(__PPC64__) 422 /* On Linux + PowerPC, DPDK doesn't support VA mode at all. Unfortunately, it doesn't correctly 423 * auto-detect at the moment, so we'll just force it here. */ 424 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--iova-mode=pa")); 425 if (args == NULL) { 426 return -1; 427 } 428 #endif 429 430 431 /* Set the base virtual address - it must be an address that is not in the 432 * ASAN shadow region, otherwise ASAN-enabled builds will ignore the 433 * mmap hint. 434 * 435 * Ref: https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm 436 */ 437 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--base-virtaddr=0x200000000000")); 438 if (args == NULL) { 439 return -1; 440 } 441 442 /* --match-allocation prevents DPDK from merging or splitting system memory allocations under the hood. 443 * This is critical for RDMA when attempting to use an rte_mempool based buffer pool. If DPDK merges two 444 * physically or IOVA contiguous memory regions, then when we go to allocate a buffer pool, it can split 445 * the memory for a buffer over two allocations meaning the buffer will be split over a memory region. 446 */ 447 #if RTE_VERSION >= RTE_VERSION_NUM(19, 02, 0, 0) 448 if (!opts->env_context || strstr(opts->env_context, "--legacy-mem") == NULL) { 449 args = spdk_push_arg(args, &argcount, _sprintf_alloc("%s", "--match-allocations")); 450 if (args == NULL) { 451 return -1; 452 } 453 } 454 #endif 455 456 if (opts->shm_id < 0) { 457 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--file-prefix=spdk_pid%d", 458 getpid())); 459 if (args == NULL) { 460 return -1; 461 } 462 } else { 463 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--file-prefix=spdk%d", 464 opts->shm_id)); 465 if (args == NULL) { 466 return -1; 467 } 468 469 /* set the process type */ 470 args = spdk_push_arg(args, &argcount, _sprintf_alloc("--proc-type=auto")); 471 if (args == NULL) { 472 return -1; 473 } 474 } 475 #endif 476 477 g_eal_cmdline = args; 478 g_eal_cmdline_argcount = argcount; 479 return argcount; 480 } 481 482 int 483 spdk_env_dpdk_post_init(bool legacy_mem) 484 { 485 int rc; 486 487 spdk_pci_init(); 488 489 rc = spdk_mem_map_init(legacy_mem); 490 if (rc < 0) { 491 fprintf(stderr, "Failed to allocate mem_map\n"); 492 return rc; 493 } 494 495 rc = spdk_vtophys_init(); 496 if (rc < 0) { 497 fprintf(stderr, "Failed to initialize vtophys\n"); 498 return rc; 499 } 500 501 return 0; 502 } 503 504 void 505 spdk_env_dpdk_post_fini(void) 506 { 507 spdk_pci_fini(); 508 509 spdk_free_args(g_eal_cmdline, g_eal_cmdline_argcount); 510 } 511 512 int 513 spdk_env_init(const struct spdk_env_opts *opts) 514 { 515 char **dpdk_args = NULL; 516 int i, rc; 517 int orig_optind; 518 bool legacy_mem; 519 520 g_external_init = false; 521 522 rc = spdk_build_eal_cmdline(opts); 523 if (rc < 0) { 524 fprintf(stderr, "Invalid arguments to initialize DPDK\n"); 525 return -EINVAL; 526 } 527 528 printf("Starting %s / %s initialization...\n", SPDK_VERSION_STRING, rte_version()); 529 printf("[ DPDK EAL parameters: "); 530 for (i = 0; i < g_eal_cmdline_argcount; i++) { 531 printf("%s ", g_eal_cmdline[i]); 532 } 533 printf("]\n"); 534 535 /* DPDK rearranges the array we pass to it, so make a copy 536 * before passing so we can still free the individual strings 537 * correctly. 538 */ 539 dpdk_args = calloc(g_eal_cmdline_argcount, sizeof(char *)); 540 if (dpdk_args == NULL) { 541 fprintf(stderr, "Failed to allocate dpdk_args\n"); 542 return -ENOMEM; 543 } 544 memcpy(dpdk_args, g_eal_cmdline, sizeof(char *) * g_eal_cmdline_argcount); 545 546 fflush(stdout); 547 orig_optind = optind; 548 optind = 1; 549 rc = rte_eal_init(g_eal_cmdline_argcount, dpdk_args); 550 optind = orig_optind; 551 552 free(dpdk_args); 553 554 if (rc < 0) { 555 if (rte_errno == EALREADY) { 556 fprintf(stderr, "DPDK already initialized\n"); 557 } else { 558 fprintf(stderr, "Failed to initialize DPDK\n"); 559 } 560 return -rte_errno; 561 } 562 563 if (opts->shm_id < 0 && !opts->hugepage_single_segments) { 564 /* 565 * Unlink hugepage and config info files after init. This will ensure they get 566 * deleted on app exit, even if the app crashes and does not exit normally. 567 * Only do this when not in multi-process mode, since for multi-process other 568 * apps will need to open these files. These files are not created for 569 * "single file segments". 570 */ 571 spdk_env_unlink_shared_files(); 572 } 573 574 legacy_mem = false; 575 if (opts->env_context && strstr(opts->env_context, "--legacy-mem") != NULL) { 576 legacy_mem = true; 577 } 578 579 return spdk_env_dpdk_post_init(legacy_mem); 580 } 581 582 void 583 spdk_env_fini(void) 584 { 585 spdk_env_dpdk_post_fini(); 586 } 587 588 bool 589 spdk_env_dpdk_external_init(void) 590 { 591 return g_external_init; 592 } 593