lib/librte_eal/bsdapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2014 6WIND S.A.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <string.h>
  39 #include <stdarg.h>
  40 #include <unistd.h>
  41 #include <pthread.h>
  42 #include <syslog.h>
  43 #include <getopt.h>
  44 #include <sys/file.h>
  45 #include <stddef.h>
  46 #include <errno.h>
  47 #include <limits.h>
  48 #include <errno.h>
  49 #include <sys/mman.h>
  50 #include <sys/queue.h>
  51
  52 #include <rte_common.h>
  53 #include <rte_debug.h>
  54 #include <rte_memory.h>
  55 #include <rte_memzone.h>
  56 #include <rte_launch.h>
  57 #include <rte_tailq.h>
  58 #include <rte_eal.h>
  59 #include <rte_eal_memconfig.h>
  60 #include <rte_per_lcore.h>
  61 #include <rte_lcore.h>
  62 #include <rte_log.h>
  63 #include <rte_random.h>
  64 #include <rte_cycles.h>
  65 #include <rte_string_fns.h>
  66 #include <rte_cpuflags.h>
  67 #include <rte_interrupts.h>
  68 #include <rte_pci.h>
  69 #include <rte_dev.h>
  70 #include <rte_devargs.h>
  71 #include <rte_common.h>
  72 #include <rte_version.h>
  73 #include <rte_atomic.h>
  74 #include <malloc_heap.h>
  75 #include <rte_eth_ring.h>
  76
  77 #include "eal_private.h"
  78 #include "eal_thread.h"
  79 #include "eal_internal_cfg.h"
  80 #include "eal_filesystem.h"
  81 #include "eal_hugepages.h"
  82
  83 #define OPT_HUGE_DIR    "huge-dir"
  84 #define OPT_PROC_TYPE   "proc-type"
  85 #define OPT_NO_SHCONF   "no-shconf"
  86 #define OPT_NO_HPET     "no-hpet"
  87 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  88 #define OPT_NO_PCI      "no-pci"
  89 #define OPT_NO_HUGE     "no-huge"
  90 #define OPT_FILE_PREFIX "file-prefix"
  91 #define OPT_SOCKET_MEM  "socket-mem"
  92 #define OPT_USE_DEVICE "use-device"
  93 #define OPT_PCI_WHITELIST "pci-whitelist"
  94 #define OPT_PCI_BLACKLIST "pci-blacklist"
  95 #define OPT_VDEV        "vdev"
  96 #define OPT_SYSLOG      "syslog"
  97
  98 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  99
 100 #define BITS_PER_HEX 4
 101
 102 /* Allow the application to print its usage message too if set */
 103 static rte_usage_hook_t rte_application_usage_hook = NULL;
 104 /* early configuration structure, when memory config is not mmapped */
 105 static struct rte_mem_config early_mem_config;
 106
 107 /* define fd variable here, because file needs to be kept open for the
 108  * duration of the program, as we hold a write lock on it in the primary proc */
 109 static int mem_cfg_fd = -1;
 110
 111 static struct flock wr_lock = {
 112                 .l_type = F_WRLCK,
 113                 .l_whence = SEEK_SET,
 114                 .l_start = offsetof(struct rte_mem_config, memseg),
 115                 .l_len = sizeof(early_mem_config.memseg),
 116 };
 117
 118 /* Address of global and public configuration */
 119 static struct rte_config rte_config = {
 120                 .mem_config = &early_mem_config,
 121 };
 122
 123 /* internal configuration (per-core) */
 124 struct lcore_config lcore_config[RTE_MAX_LCORE];
 125
 126 /* internal configuration */
 127 struct internal_config internal_config;
 128
 129 /* used by rte_rdtsc() */
 130 int rte_cycles_vmware_tsc_map;
 131
 132 /* Return a pointer to the configuration structure */
 133 struct rte_config *
 134 rte_eal_get_configuration(void)
 135 {
 136         return &rte_config;
 137 }
 138
 139 /* parse a sysfs (or other) file containing one integer value */
 140 int
 141 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 142 {
 143         FILE *f;
 144         char buf[BUFSIZ];
 145         char *end = NULL;
 146
 147         if ((f = fopen(filename, "r")) == NULL) {
 148                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 149                         __func__, filename);
 150                 return -1;
 151         }
 152
 153         if (fgets(buf, sizeof(buf), f) == NULL) {
 154                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 155                         __func__, filename);
 156                 fclose(f);
 157                 return -1;
 158         }
 159         *val = strtoul(buf, &end, 0);
 160         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 161                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 162                                 __func__, filename);
 163                 fclose(f);
 164                 return -1;
 165         }
 166         fclose(f);
 167         return 0;
 168 }
 169
 170
 171 /* create memory configuration in shared/mmap memory. Take out
 172  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 173  * This means we never close the file while running (auto-close on exit).
 174  * We also don't lock the whole file, so that in future we can use read-locks
 175  * on other parts, e.g. memzones, to detect if there are running secondary
 176  * processes. */
 177 static void
 178 rte_eal_config_create(void)
 179 {
 180         void *rte_mem_cfg_addr;
 181         int retval;
 182
 183         const char *pathname = eal_runtime_config_path();
 184
 185         if (internal_config.no_shconf)
 186                 return;
 187
 188         if (mem_cfg_fd < 0){
 189                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 190                 if (mem_cfg_fd < 0)
 191                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 192         }
 193
 194         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 195         if (retval < 0){
 196                 close(mem_cfg_fd);
 197                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 198         }
 199
 200         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 201         if (retval < 0){
 202                 close(mem_cfg_fd);
 203                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 204                                 "process running?\n", pathname);
 205         }
 206
 207         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 208                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 209
 210         if (rte_mem_cfg_addr == MAP_FAILED){
 211                 rte_panic("Cannot mmap memory for rte_config\n");
 212         }
 213         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 214         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 215 }
 216
 217 /* attach to an existing shared memory config */
 218 static void
 219 rte_eal_config_attach(void)
 220 {
 221         void *rte_mem_cfg_addr;
 222         const char *pathname = eal_runtime_config_path();
 223
 224         if (internal_config.no_shconf)
 225                 return;
 226
 227         if (mem_cfg_fd < 0){
 228                 mem_cfg_fd = open(pathname, O_RDWR);
 229                 if (mem_cfg_fd < 0)
 230                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 231         }
 232
 233         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 234                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 235         close(mem_cfg_fd);
 236         if (rte_mem_cfg_addr == MAP_FAILED)
 237                 rte_panic("Cannot mmap memory for rte_config\n");
 238
 239         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 240 }
 241
 242 /* Detect if we are a primary or a secondary process */
 243 static enum rte_proc_type_t
 244 eal_proc_type_detect(void)
 245 {
 246         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 247         const char *pathname = eal_runtime_config_path();
 248
 249         /* if we can open the file but not get a write-lock we are a secondary
 250          * process. NOTE: if we get a file handle back, we keep that open
 251          * and don't close it to prevent a race condition between multiple opens */
 252         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 253                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 254                 ptype = RTE_PROC_SECONDARY;
 255
 256         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 257                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 258
 259         return ptype;
 260 }
 261
 262 /* Sets up rte_config structure with the pointer to shared memory config.*/
 263 static void
 264 rte_config_init(void)
 265 {
 266         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 267                         eal_proc_type_detect() : /* for auto, detect the type */
 268                         internal_config.process_type; /* otherwise use what's already set */
 269
 270         switch (rte_config.process_type){
 271         case RTE_PROC_PRIMARY:
 272                 rte_eal_config_create();
 273                 break;
 274         case RTE_PROC_SECONDARY:
 275                 rte_eal_config_attach();
 276                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 277                 break;
 278         case RTE_PROC_AUTO:
 279         case RTE_PROC_INVALID:
 280                 rte_panic("Invalid process type\n");
 281         }
 282 }
 283
 284 /* display usage */
 285 static void
 286 eal_usage(const char *prgname)
 287 {
 288         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 289                "[--proc-type primary|secondary|auto] \n\n"
 290                "EAL options:\n"
 291                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 292                "  -n NUM       : Number of memory channels\n"
 293                "  -v           : Display version information on startup\n"
 294                "  -m MB        : memory to allocate\n"
 295                "  -r NUM       : force number of memory ranks (don't detect)\n"
 296                "  --"OPT_PROC_TYPE"  : type of this process\n"
 297                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 298                "               Prevent EAL from using this PCI device. The argument\n"
 299                "               format is <domain:bus:devid.func>.\n"
 300                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 301                "               Only use the specified PCI devices. The argument format\n"
 302                "               is <[domain:]bus:devid.func>. This option can be present\n"
 303                "               several times (once per device).\n"
 304                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 305                "  --"OPT_VDEV": add a virtual device.\n"
 306                "               The argument format is <driver><id>[,key=val,...]\n"
 307                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 308                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 309                "\nEAL options for DEBUG use only:\n"
 310                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 311                "  --"OPT_NO_PCI"   : disable pci\n"
 312                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 313                "\n",
 314                prgname);
 315         /* Allow the application to print its usage message too if hook is set */
 316         if ( rte_application_usage_hook ) {
 317                 printf("===== Application Usage =====\n\n");
 318                 rte_application_usage_hook(prgname);
 319         }
 320 }
 321
 322 /* Set a per-application usage message */
 323 rte_usage_hook_t
 324 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 325 {
 326         rte_usage_hook_t        old_func;
 327
 328         /* Will be NULL on the first call to denote the last usage routine. */
 329         old_func                                        = rte_application_usage_hook;
 330         rte_application_usage_hook      = usage_func;
 331
 332         return old_func;
 333 }
 334
 335 /*
 336  * Parse the coremask given as argument (hexadecimal string) and fill
 337  * the global configuration (core role and core count) with the parsed
 338  * value.
 339  */
 340 static int xdigit2val(unsigned char c)
 341 {
 342         int val;
 343         if(isdigit(c))
 344                 val = c - '0';
 345         else if(isupper(c))
 346                 val = c - 'A' + 10;
 347         else
 348                 val = c - 'a' + 10;
 349         return val;
 350 }
 351 static int
 352 eal_parse_coremask(const char *coremask)
 353 {
 354         struct rte_config *cfg = rte_eal_get_configuration();
 355         int i, j, idx = 0 ;
 356         unsigned count = 0;
 357         char c;
 358         int val;
 359
 360         if (coremask == NULL)
 361                 return -1;
 362         /* Remove all blank characters ahead and after .
 363          * Remove 0x/0X if exists.
 364          */
 365         while (isblank(*coremask))
 366                 coremask++;
 367         if (coremask[0] == '0' && ((coremask[1] == 'x')
 368                 ||  (coremask[1] == 'X')) )
 369                 coremask += 2;
 370         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 371         while ((i > 0) && isblank(coremask[i - 1]))
 372                 i--;
 373         if (i == 0)
 374                 return -1;
 375
 376         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 377                 c = coremask[i];
 378                 if (isxdigit(c) == 0) {
 379                         /* invalid characters */
 380                         return (-1);
 381                 }
 382                 val = xdigit2val(c);
 383                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 384                         if((1 << j) & val) {
 385                                 cfg->lcore_role[idx] = ROLE_RTE;
 386                                 if(count == 0)
 387                                         cfg->master_lcore = idx;
 388                                 count++;
 389                         } else  {
 390                                 cfg->lcore_role[idx] = ROLE_OFF;
 391                         }
 392                 }
 393         }
 394         for(; i >= 0; i--)
 395                 if(coremask[i] != '0')
 396                         return -1;
 397         for(; idx < RTE_MAX_LCORE; idx++)
 398                 cfg->lcore_role[idx] = ROLE_OFF;
 399         if(count == 0)
 400                 return -1;
 401         return 0;
 402 }
 403
 404 static int
 405 eal_parse_syslog(const char *facility)
 406 {
 407         int i;
 408         static struct {
 409                 const char *name;
 410                 int value;
 411         } map[] = {
 412                 { "auth", LOG_AUTH },
 413                 { "cron", LOG_CRON },
 414                 { "daemon", LOG_DAEMON },
 415                 { "ftp", LOG_FTP },
 416                 { "kern", LOG_KERN },
 417                 { "lpr", LOG_LPR },
 418                 { "mail", LOG_MAIL },
 419                 { "news", LOG_NEWS },
 420                 { "syslog", LOG_SYSLOG },
 421                 { "user", LOG_USER },
 422                 { "uucp", LOG_UUCP },
 423                 { "local0", LOG_LOCAL0 },
 424                 { "local1", LOG_LOCAL1 },
 425                 { "local2", LOG_LOCAL2 },
 426                 { "local3", LOG_LOCAL3 },
 427                 { "local4", LOG_LOCAL4 },
 428                 { "local5", LOG_LOCAL5 },
 429                 { "local6", LOG_LOCAL6 },
 430                 { "local7", LOG_LOCAL7 },
 431                 { NULL, 0 }
 432         };
 433
 434         for (i = 0; map[i].name; i++) {
 435                 if (!strcmp(facility, map[i].name)) {
 436                         internal_config.syslog_facility = map[i].value;
 437                         return 0;
 438                 }
 439         }
 440         return -1;
 441 }
 442
 443 static inline size_t
 444 eal_get_hugepage_mem_size(void)
 445 {
 446         uint64_t size = 0;
 447         unsigned i, j;
 448
 449         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 450                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 451                 if (hpi->hugedir != NULL) {
 452                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 453                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 454                         }
 455                 }
 456         }
 457
 458         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 459 }
 460
 461 static enum rte_proc_type_t
 462 eal_parse_proc_type(const char *arg)
 463 {
 464         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 465                 return RTE_PROC_PRIMARY;
 466         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 467                 return RTE_PROC_SECONDARY;
 468         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 469                 return RTE_PROC_AUTO;
 470
 471         return RTE_PROC_INVALID;
 472 }
 473
 474 /* Parse the argument given in the command line of the application */
 475 static int
 476 eal_parse_args(int argc, char **argv)
 477 {
 478         int opt, ret, i;
 479         char **argvopt;
 480         int option_index;
 481         int coremask_ok = 0;
 482         char *prgname = argv[0];
 483         static struct option lgopts[] = {
 484                 {OPT_NO_HUGE, 0, 0, 0},
 485                 {OPT_NO_PCI, 0, 0, 0},
 486                 {OPT_NO_HPET, 0, 0, 0},
 487                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 488                 {OPT_HUGE_DIR, 1, 0, 0},
 489                 {OPT_NO_SHCONF, 0, 0, 0},
 490                 {OPT_PROC_TYPE, 1, 0, 0},
 491                 {OPT_FILE_PREFIX, 1, 0, 0},
 492                 {OPT_SOCKET_MEM, 1, 0, 0},
 493                 {OPT_PCI_WHITELIST, 1, 0, 0},
 494                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 495                 {OPT_VDEV, 1, 0, 0},
 496                 {OPT_SYSLOG, 1, NULL, 0},
 497                 {0, 0, 0, 0}
 498         };
 499
 500         argvopt = argv;
 501
 502         internal_config.memory = 0;
 503         internal_config.force_nrank = 0;
 504         internal_config.force_nchannel = 0;
 505         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 506         internal_config.hugepage_dir = NULL;
 507         internal_config.force_sockets = 0;
 508         internal_config.syslog_facility = LOG_DAEMON;
 509 #ifdef RTE_LIBEAL_USE_HPET
 510         internal_config.no_hpet = 0;
 511 #else
 512         internal_config.no_hpet = 1;
 513 #endif
 514         /* zero out the NUMA config */
 515         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 516                 internal_config.socket_mem[i] = 0;
 517
 518         /* zero out hugedir descriptors */
 519         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 520                 internal_config.hugepage_info[i].lock_descriptor = 0;
 521
 522         internal_config.vmware_tsc_map = 0;
 523
 524         while ((opt = getopt_long(argc, argvopt, "b:w:c:m:n:r:v",
 525                                   lgopts, &option_index)) != EOF) {
 526
 527                 switch (opt) {
 528                 /* blacklist */
 529                 case 'b':
 530                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 531                                         optarg) < 0) {
 532                                 eal_usage(prgname);
 533                                 return (-1);
 534                         }
 535                         break;
 536                 /* whitelist */
 537                 case 'w':
 538                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 539                                         optarg) < 0) {
 540                                 eal_usage(prgname);
 541                                 return -1;
 542                         }
 543                         break;
 544                 /* coremask */
 545                 case 'c':
 546                         if (eal_parse_coremask(optarg) < 0) {
 547                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 548                                 eal_usage(prgname);
 549                                 return -1;
 550                         }
 551                         coremask_ok = 1;
 552                         break;
 553                 /* size of memory */
 554                 case 'm':
 555                         internal_config.memory = atoi(optarg);
 556                         internal_config.memory *= 1024ULL;
 557                         internal_config.memory *= 1024ULL;
 558                         break;
 559                 /* force number of channels */
 560                 case 'n':
 561                         internal_config.force_nchannel = atoi(optarg);
 562                         if (internal_config.force_nchannel == 0 ||
 563                             internal_config.force_nchannel > 4) {
 564                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 565                                 eal_usage(prgname);
 566                                 return -1;
 567                         }
 568                         break;
 569                 /* force number of ranks */
 570                 case 'r':
 571                         internal_config.force_nrank = atoi(optarg);
 572                         if (internal_config.force_nrank == 0 ||
 573                             internal_config.force_nrank > 16) {
 574                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 575                                 eal_usage(prgname);
 576                                 return -1;
 577                         }
 578                         break;
 579                 case 'v':
 580                         /* since message is explicitly requested by user, we
 581                          * write message at highest log level so it can always be seen
 582                          * even if info or warning messages are disabled */
 583                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 584                         break;
 585
 586                 /* long options */
 587                 case 0:
 588                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 589                                 internal_config.no_hugetlbfs = 1;
 590                         }
 591                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 592                                 internal_config.no_pci = 1;
 593                         }
 594                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 595                                 internal_config.no_hpet = 1;
 596                         }
 597                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 598                                 internal_config.vmware_tsc_map = 1;
 599                         }
 600                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 601                                 internal_config.no_shconf = 1;
 602                         }
 603                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 604                                 RTE_LOG(ERR, EAL, "Option "OPT_HUGE_DIR" is not supported on"
 605                                                 "FreeBSD\n");
 606                                 return -1;
 607                         }
 608                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 609                                 internal_config.process_type = eal_parse_proc_type(optarg);
 610                         }
 611                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 612                                 RTE_LOG(ERR, EAL, "Option "OPT_FILE_PREFIX" is not supported on"
 613                                                 "FreeBSD\n");
 614                                 return -1;
 615                         }
 616                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 617                                 RTE_LOG(ERR, EAL, "Option "OPT_SOCKET_MEM" is not supported on"
 618                                                 "FreeBSD\n");
 619                                 return -1;
 620                         }
 621                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 622                                 printf("The --use-device option is deprecated, please use\n"
 623                                         "--whitelist or --vdev instead.\n");
 624                                 eal_usage(prgname);
 625                                 return -1;
 626                         }
 627                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 628                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 629                                                 optarg) < 0) {
 630                                         eal_usage(prgname);
 631                                         return -1;
 632                                 }
 633                         }
 634                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 635                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 636                                                 optarg) < 0) {
 637                                         eal_usage(prgname);
 638                                         return -1;
 639                                 }
 640                         }
 641                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 642                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 643                                                 optarg) < 0) {
 644                                         eal_usage(prgname);
 645                                         return -1;
 646                                 }
 647                         }
 648                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 649                                 if (eal_parse_syslog(optarg) < 0) {
 650                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 651                                                         OPT_SYSLOG "\n");
 652                                         eal_usage(prgname);
 653                                         return -1;
 654                                 }
 655                         }
 656                         break;
 657
 658                 default:
 659                         eal_usage(prgname);
 660                         return -1;
 661                 }
 662         }
 663
 664         /* sanity checks */
 665         if (!coremask_ok) {
 666                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 667                 eal_usage(prgname);
 668                 return -1;
 669         }
 670         if (internal_config.process_type == RTE_PROC_AUTO){
 671                 internal_config.process_type = eal_proc_type_detect();
 672         }
 673         if (internal_config.process_type == RTE_PROC_INVALID){
 674                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 675                 eal_usage(prgname);
 676                 return -1;
 677         }
 678         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 679                         internal_config.force_nchannel == 0) {
 680                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 681                 eal_usage(prgname);
 682                 return -1;
 683         }
 684         if (index(internal_config.hugefile_prefix,'%') != NULL){
 685                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 686                 eal_usage(prgname);
 687                 return -1;
 688         }
 689         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 690                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 691                                 "at the same time\n");
 692                 eal_usage(prgname);
 693                 return -1;
 694         }
 695         /* --no-huge doesn't make sense with either -m or --socket-mem */
 696         if (internal_config.no_hugetlbfs &&
 697                         (internal_config.memory > 0 ||
 698                                         internal_config.force_sockets == 1)) {
 699                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 700                                 "together with --no-huge!\n");
 701                 eal_usage(prgname);
 702                 return -1;
 703         }
 704
 705         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 706                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 707                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 708                         "[-w] options cannot be used at the same time\n");
 709                 eal_usage(prgname);
 710                 return -1;
 711         }
 712
 713         if (optind >= 0)
 714                 argv[optind-1] = prgname;
 715
 716         /* if no memory amounts were requested, this will result in 0 and
 717          * will be overriden later, right after eal_hugepage_info_init() */
 718         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 719                 internal_config.memory += internal_config.socket_mem[i];
 720
 721         ret = optind-1;
 722         optind = 0; /* reset getopt lib */
 723         return ret;
 724 }
 725
 726 static void
 727 eal_check_mem_on_local_socket(void)
 728 {
 729         const struct rte_memseg *ms;
 730         int i, socket_id;
 731
 732         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 733
 734         ms = rte_eal_get_physmem_layout();
 735
 736         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 737                 if (ms[i].socket_id == socket_id &&
 738                                 ms[i].len > 0)
 739                         return;
 740
 741         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 742                         "memory on local socket!\n");
 743 }
 744
 745 static int
 746 sync_func(__attribute__((unused)) void *arg)
 747 {
 748         return 0;
 749 }
 750
 751 inline static void
 752 rte_eal_mcfg_complete(void)
 753 {
 754         /* ALL shared mem_config related INIT DONE */
 755         if (rte_config.process_type == RTE_PROC_PRIMARY)
 756                 rte_config.mem_config->magic = RTE_MAGIC;
 757 }
 758
 759 /* return non-zero if hugepages are enabled. */
 760 int rte_eal_has_hugepages(void)
 761 {
 762         return !internal_config.no_hugetlbfs;
 763 }
 764
 765 /* Abstraction for port I/0 privilege */
 766 static int
 767 rte_eal_iopl_init(void)
 768 {
 769         int fd = -1;
 770         fd = open("/dev/io", O_RDWR);
 771         if (fd < 0)
 772                 return -1;
 773         return 0;
 774 }
 775
 776 /* Launch threads, called at application init(). */
 777 int
 778 rte_eal_init(int argc, char **argv)
 779 {
 780         int i, fctret, ret;
 781         pthread_t thread_id;
 782         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 783
 784         if (!rte_atomic32_test_and_set(&run_once))
 785                 return -1;
 786
 787         thread_id = pthread_self();
 788
 789         if (rte_eal_log_early_init() < 0)
 790                 rte_panic("Cannot init early logs\n");
 791
 792         fctret = eal_parse_args(argc, argv);
 793         if (fctret < 0)
 794                 exit(1);
 795
 796         if (internal_config.no_hugetlbfs == 0 &&
 797                         internal_config.process_type != RTE_PROC_SECONDARY &&
 798                         eal_hugepage_info_init() < 0)
 799                 rte_panic("Cannot get hugepage information\n");
 800
 801         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 802                 if (internal_config.no_hugetlbfs)
 803                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 804                 else
 805                         internal_config.memory = eal_get_hugepage_mem_size();
 806         }
 807
 808         if (internal_config.vmware_tsc_map == 1) {
 809 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 810                 rte_cycles_vmware_tsc_map = 1;
 811                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 812                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 813 #else
 814                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 815                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 816 #endif
 817         }
 818
 819         rte_srand(rte_rdtsc());
 820
 821         rte_config_init();
 822
 823         if (rte_eal_iopl_init() == 0)
 824                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 825
 826         if (rte_eal_cpu_init() < 0)
 827                 rte_panic("Cannot detect lcores\n");
 828
 829         if (rte_eal_memory_init() < 0)
 830                 rte_panic("Cannot init memory\n");
 831
 832         if (rte_eal_memzone_init() < 0)
 833                 rte_panic("Cannot init memzone\n");
 834
 835         if (rte_eal_tailqs_init() < 0)
 836                 rte_panic("Cannot init tail queues for objects\n");
 837
 838 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 839                 rte_panic("Cannot init logs\n");*/
 840
 841         if (rte_eal_alarm_init() < 0)
 842                 rte_panic("Cannot init interrupt-handling thread\n");
 843
 844         if (rte_eal_intr_init() < 0)
 845                 rte_panic("Cannot init interrupt-handling thread\n");
 846
 847         if (rte_eal_timer_init() < 0)
 848                 rte_panic("Cannot init HPET or TSC timers\n");
 849
 850         if (rte_eal_pci_init() < 0)
 851                 rte_panic("Cannot init PCI\n");
 852
 853         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 854                 rte_config.master_lcore, thread_id);
 855
 856         eal_check_mem_on_local_socket();
 857
 858         rte_eal_mcfg_complete();
 859
 860         if (rte_eal_dev_init(PMD_INIT_PRE_PCI_PROBE) < 0)
 861                 rte_panic("Cannot init pmd devices\n");
 862
 863         RTE_LCORE_FOREACH_SLAVE(i) {
 864
 865                 /*
 866                  * create communication pipes between master thread
 867                  * and children
 868                  */
 869                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 870                         rte_panic("Cannot create pipe\n");
 871                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 872                         rte_panic("Cannot create pipe\n");
 873
 874                 lcore_config[i].state = WAIT;
 875
 876                 /* create a thread for each lcore */
 877                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 878                                      eal_thread_loop, NULL);
 879                 if (ret != 0)
 880                         rte_panic("Cannot create thread\n");
 881         }
 882
 883         eal_thread_init_master(rte_config.master_lcore);
 884
 885         /*
 886          * Launch a dummy function on all slave lcores, so that master lcore
 887          * knows they are all ready when this function returns.
 888          */
 889         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 890         rte_eal_mp_wait_lcore();
 891
 892         /* Probe & Initialize PCI devices */
 893         if (rte_eal_pci_probe())
 894                         rte_panic("Cannot probe PCI\n");
 895
 896         /* Initialize any outstanding devices */
 897         if (rte_eal_dev_init(PMD_INIT_POST_PCI_PROBE) < 0)
 898                 rte_panic("Cannot init pmd devices\n");
 899
 900         return fctret;
 901 }
 902
 903 /* get core role */
 904 enum rte_lcore_role_t
 905 rte_eal_lcore_role(unsigned lcore_id)
 906 {
 907         return (rte_config.lcore_role[lcore_id]);
 908 }
 909
 910 enum rte_proc_type_t
 911 rte_eal_process_type(void)
 912 {
 913         return (rte_config.process_type);
 914 }
 915