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_devargs.h>
  70 #include <rte_common.h>
  71 #include <rte_version.h>
  72 #include <rte_atomic.h>
  73 #include <malloc_heap.h>
  74 #include <rte_eth_ring.h>
  75
  76 #include "eal_private.h"
  77 #include "eal_thread.h"
  78 #include "eal_internal_cfg.h"
  79 #include "eal_filesystem.h"
  80 #include "eal_hugepages.h"
  81
  82 #define OPT_HUGE_DIR    "huge-dir"
  83 #define OPT_PROC_TYPE   "proc-type"
  84 #define OPT_NO_SHCONF   "no-shconf"
  85 #define OPT_NO_HPET     "no-hpet"
  86 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  87 #define OPT_NO_PCI      "no-pci"
  88 #define OPT_NO_HUGE     "no-huge"
  89 #define OPT_FILE_PREFIX "file-prefix"
  90 #define OPT_SOCKET_MEM  "socket-mem"
  91 #define OPT_USE_DEVICE  "use-device"
  92 #define OPT_SYSLOG      "syslog"
  93
  94 #define RTE_EAL_BLACKLIST_SIZE  0x100
  95
  96 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  97
  98 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
  99
 100 #define HIGHEST_RPL 3
 101
 102 #define BITS_PER_HEX 4
 103
 104 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 105 {                                                               \
 106         unsigned long val;                                      \
 107         char *end;                                              \
 108         errno = 0;                                              \
 109         val = strtoul((in), &end, 16);                          \
 110         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 111                 return (-EINVAL);                               \
 112         (fd) = (typeof (fd))val;                                \
 113         (in) = end + 1;                                         \
 114 }
 115
 116 /* Allow the application to print its usage message too if set */
 117 static rte_usage_hook_t rte_application_usage_hook = NULL;
 118 /* early configuration structure, when memory config is not mmapped */
 119 static struct rte_mem_config early_mem_config;
 120
 121 /* define fd variable here, because file needs to be kept open for the
 122  * duration of the program, as we hold a write lock on it in the primary proc */
 123 static int mem_cfg_fd = -1;
 124
 125 static struct flock wr_lock = {
 126                 .l_type = F_WRLCK,
 127                 .l_whence = SEEK_SET,
 128                 .l_start = offsetof(struct rte_mem_config, memseg),
 129                 .l_len = sizeof(early_mem_config.memseg),
 130 };
 131
 132 /* Address of global and public configuration */
 133 static struct rte_config rte_config = {
 134                 .mem_config = &early_mem_config,
 135 };
 136
 137 /* internal configuration (per-core) */
 138 struct lcore_config lcore_config[RTE_MAX_LCORE];
 139
 140 /* internal configuration */
 141 struct internal_config internal_config;
 142
 143 /* used by rte_rdtsc() */
 144 int rte_cycles_vmware_tsc_map;
 145
 146 /* Return a pointer to the configuration structure */
 147 struct rte_config *
 148 rte_eal_get_configuration(void)
 149 {
 150         return &rte_config;
 151 }
 152
 153 /* parse a sysfs (or other) file containing one integer value */
 154 int
 155 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 156 {
 157         FILE *f;
 158         char buf[BUFSIZ];
 159         char *end = NULL;
 160
 161         if ((f = fopen(filename, "r")) == NULL) {
 162                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 163                         __func__, filename);
 164                 return -1;
 165         }
 166
 167         if (fgets(buf, sizeof(buf), f) == NULL) {
 168                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 169                         __func__, filename);
 170                 fclose(f);
 171                 return -1;
 172         }
 173         *val = strtoul(buf, &end, 0);
 174         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 175                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 176                                 __func__, filename);
 177                 fclose(f);
 178                 return -1;
 179         }
 180         fclose(f);
 181         return 0;
 182 }
 183
 184
 185 /* create memory configuration in shared/mmap memory. Take out
 186  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 187  * This means we never close the file while running (auto-close on exit).
 188  * We also don't lock the whole file, so that in future we can use read-locks
 189  * on other parts, e.g. memzones, to detect if there are running secondary
 190  * processes. */
 191 static void
 192 rte_eal_config_create(void)
 193 {
 194         void *rte_mem_cfg_addr;
 195         int retval;
 196
 197         const char *pathname = eal_runtime_config_path();
 198
 199         if (internal_config.no_shconf)
 200                 return;
 201
 202         if (mem_cfg_fd < 0){
 203                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 204                 if (mem_cfg_fd < 0)
 205                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 206         }
 207
 208         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 209         if (retval < 0){
 210                 close(mem_cfg_fd);
 211                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 212         }
 213
 214         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 215         if (retval < 0){
 216                 close(mem_cfg_fd);
 217                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 218                                 "process running?\n", pathname);
 219         }
 220
 221         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 222                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 223
 224         if (rte_mem_cfg_addr == MAP_FAILED){
 225                 rte_panic("Cannot mmap memory for rte_config\n");
 226         }
 227         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 228         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 229 }
 230
 231 /* attach to an existing shared memory config */
 232 static void
 233 rte_eal_config_attach(void)
 234 {
 235         void *rte_mem_cfg_addr;
 236         const char *pathname = eal_runtime_config_path();
 237
 238         if (internal_config.no_shconf)
 239                 return;
 240
 241         if (mem_cfg_fd < 0){
 242                 mem_cfg_fd = open(pathname, O_RDWR);
 243                 if (mem_cfg_fd < 0)
 244                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 245         }
 246
 247         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 248                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 249         close(mem_cfg_fd);
 250         if (rte_mem_cfg_addr == MAP_FAILED)
 251                 rte_panic("Cannot mmap memory for rte_config\n");
 252
 253         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 254 }
 255
 256 /* Detect if we are a primary or a secondary process */
 257 static enum rte_proc_type_t
 258 eal_proc_type_detect(void)
 259 {
 260         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 261         const char *pathname = eal_runtime_config_path();
 262
 263         /* if we can open the file but not get a write-lock we are a secondary
 264          * process. NOTE: if we get a file handle back, we keep that open
 265          * and don't close it to prevent a race condition between multiple opens */
 266         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 267                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 268                 ptype = RTE_PROC_SECONDARY;
 269
 270         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 271                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 272
 273         return ptype;
 274 }
 275
 276 /* Sets up rte_config structure with the pointer to shared memory config.*/
 277 static void
 278 rte_config_init(void)
 279 {
 280         /* set the magic in configuration structure */
 281         rte_config.magic = RTE_MAGIC;
 282         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 283                         eal_proc_type_detect() : /* for auto, detect the type */
 284                         internal_config.process_type; /* otherwise use what's already set */
 285
 286         switch (rte_config.process_type){
 287         case RTE_PROC_PRIMARY:
 288                 rte_eal_config_create();
 289                 break;
 290         case RTE_PROC_SECONDARY:
 291                 rte_eal_config_attach();
 292                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 293                 break;
 294         case RTE_PROC_AUTO:
 295         case RTE_PROC_INVALID:
 296                 rte_panic("Invalid process type\n");
 297         }
 298 }
 299
 300 /* display usage */
 301 static void
 302 eal_usage(const char *prgname)
 303 {
 304         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 305                "[--proc-type primary|secondary|auto] \n\n"
 306                "EAL options:\n"
 307                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 308                "  -n NUM       : Number of memory channels\n"
 309                    "  -v           : Display version information on startup\n"
 310                "                 (multiple -b options are allowed)\n"
 311                "  -m MB        : memory to allocate\n"
 312                "  -r NUM       : force number of memory ranks (don't detect)\n"
 313                "  --"OPT_PROC_TYPE"  : type of this process\n"
 314                " --"OPT_USE_DEVICE": use the specified ethernet device(s) only.\n"
 315                " The argument format is <[domain:]bus:devid.func> to add\n"
 316                " a PCI device to the white list or <driver><id>[;key=val;...]\n"
 317                " to add a virtual device.\n"
 318                " [NOTE: PCI whitelist cannot be used with -b option]\n"
 319                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 320                            "native RDTSC\n"
 321                "\nEAL options for DEBUG use only:\n"
 322                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 323                "  --"OPT_NO_PCI"   : disable pci\n"
 324                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 325                "\n",
 326                prgname);
 327         /* Allow the application to print its usage message too if hook is set */
 328         if ( rte_application_usage_hook ) {
 329                 printf("===== Application Usage =====\n\n");
 330                 rte_application_usage_hook(prgname);
 331         }
 332 }
 333
 334 /* Set a per-application usage message */
 335 rte_usage_hook_t
 336 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 337 {
 338         rte_usage_hook_t        old_func;
 339
 340         /* Will be NULL on the first call to denote the last usage routine. */
 341         old_func                                        = rte_application_usage_hook;
 342         rte_application_usage_hook      = usage_func;
 343
 344         return old_func;
 345 }
 346
 347 /*
 348  * Parse the coremask given as argument (hexadecimal string) and fill
 349  * the global configuration (core role and core count) with the parsed
 350  * value.
 351  */
 352 static int xdigit2val(unsigned char c)
 353 {
 354         int val;
 355         if(isdigit(c))
 356                 val = c - '0';
 357         else if(isupper(c))
 358                 val = c - 'A' + 10;
 359         else
 360                 val = c - 'a' + 10;
 361         return val;
 362 }
 363 static int
 364 eal_parse_coremask(const char *coremask)
 365 {
 366         struct rte_config *cfg = rte_eal_get_configuration();
 367         int i, j, idx = 0 ;
 368         unsigned count = 0;
 369         char c;
 370         int val;
 371
 372         if (coremask == NULL)
 373                 return -1;
 374         /* Remove all blank characters ahead and after .
 375          * Remove 0x/0X if exists.
 376          */
 377         while (isblank(*coremask))
 378                 coremask++;
 379         if (coremask[0] == '0' && ((coremask[1] == 'x')
 380                 ||  (coremask[1] == 'X')) )
 381                 coremask += 2;
 382         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 383         while ((i > 0) && isblank(coremask[i - 1]))
 384                 i--;
 385         if (i == 0)
 386                 return -1;
 387
 388         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 389                 c = coremask[i];
 390                 if (isxdigit(c) == 0) {
 391                         /* invalid characters */
 392                         return (-1);
 393                 }
 394                 val = xdigit2val(c);
 395                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 396                         if((1 << j) & val) {
 397                                 cfg->lcore_role[idx] = ROLE_RTE;
 398                                 if(count == 0)
 399                                         cfg->master_lcore = idx;
 400                                 count++;
 401                         } else  {
 402                                 cfg->lcore_role[idx] = ROLE_OFF;
 403                         }
 404                 }
 405         }
 406         for(; i >= 0; i--)
 407                 if(coremask[i] != '0')
 408                         return -1;
 409         for(; idx < RTE_MAX_LCORE; idx++)
 410                 cfg->lcore_role[idx] = ROLE_OFF;
 411         if(count == 0)
 412                 return -1;
 413         return 0;
 414 }
 415
 416 static int
 417 eal_parse_syslog(const char *facility)
 418 {
 419         int i;
 420         static struct {
 421                 const char *name;
 422                 int value;
 423         } map[] = {
 424                 { "auth", LOG_AUTH },
 425                 { "cron", LOG_CRON },
 426                 { "daemon", LOG_DAEMON },
 427                 { "ftp", LOG_FTP },
 428                 { "kern", LOG_KERN },
 429                 { "lpr", LOG_LPR },
 430                 { "mail", LOG_MAIL },
 431                 { "news", LOG_NEWS },
 432                 { "syslog", LOG_SYSLOG },
 433                 { "user", LOG_USER },
 434                 { "uucp", LOG_UUCP },
 435                 { "local0", LOG_LOCAL0 },
 436                 { "local1", LOG_LOCAL1 },
 437                 { "local2", LOG_LOCAL2 },
 438                 { "local3", LOG_LOCAL3 },
 439                 { "local4", LOG_LOCAL4 },
 440                 { "local5", LOG_LOCAL5 },
 441                 { "local6", LOG_LOCAL6 },
 442                 { "local7", LOG_LOCAL7 },
 443                 { NULL, 0 }
 444         };
 445
 446         for (i = 0; map[i].name; i++) {
 447                 if (!strcmp(facility, map[i].name)) {
 448                         internal_config.syslog_facility = map[i].value;
 449                         return 0;
 450                 }
 451         }
 452         return -1;
 453 }
 454
 455 static inline size_t
 456 eal_get_hugepage_mem_size(void)
 457 {
 458         uint64_t size = 0;
 459         unsigned i, j;
 460
 461         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 462                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 463                 if (hpi->hugedir != NULL) {
 464                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 465                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 466                         }
 467                 }
 468         }
 469
 470         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 471 }
 472
 473 static enum rte_proc_type_t
 474 eal_parse_proc_type(const char *arg)
 475 {
 476         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 477                 return RTE_PROC_PRIMARY;
 478         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 479                 return RTE_PROC_SECONDARY;
 480         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 481                 return RTE_PROC_AUTO;
 482
 483         return RTE_PROC_INVALID;
 484 }
 485
 486 static int
 487 eal_parse_use_device(const char *optarg)
 488 {
 489         struct rte_pci_addr addr;
 490         char *dup, *sep;
 491
 492         dup = strdup(optarg);
 493         if (dup == NULL)
 494                 return -1;
 495
 496         /* remove arguments in 'dup' string */
 497         sep = strchr(dup, ';');
 498         if (sep != NULL)
 499                 *sep = '\0';
 500
 501         /* if argument is a PCI address, it's a whitelisted device */
 502         if (eal_parse_pci_DomBDF(dup, &addr) == 0 ||
 503                 eal_parse_pci_BDF(dup, &addr) == 0) {
 504                 rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI, optarg);
 505         } else {
 506                 rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL, optarg);
 507         }
 508
 509         free(dup);
 510         return 0;
 511 }
 512
 513 /* Parse the argument given in the command line of the application */
 514 static int
 515 eal_parse_args(int argc, char **argv)
 516 {
 517         int opt, ret, i;
 518         char **argvopt;
 519         int option_index;
 520         int coremask_ok = 0;
 521         char *prgname = argv[0];
 522         static struct option lgopts[] = {
 523                 {OPT_NO_HUGE, 0, 0, 0},
 524                 {OPT_NO_PCI, 0, 0, 0},
 525                 {OPT_NO_HPET, 0, 0, 0},
 526                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 527                 {OPT_HUGE_DIR, 1, 0, 0},
 528                 {OPT_NO_SHCONF, 0, 0, 0},
 529                 {OPT_PROC_TYPE, 1, 0, 0},
 530                 {OPT_FILE_PREFIX, 1, 0, 0},
 531                 {OPT_SOCKET_MEM, 1, 0, 0},
 532                 {OPT_USE_DEVICE, 1, 0, 0},
 533                 {OPT_SYSLOG, 1, NULL, 0},
 534                 {0, 0, 0, 0}
 535         };
 536
 537         argvopt = argv;
 538
 539         internal_config.memory = 0;
 540         internal_config.force_nrank = 0;
 541         internal_config.force_nchannel = 0;
 542         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 543         internal_config.hugepage_dir = NULL;
 544         internal_config.force_sockets = 0;
 545         internal_config.syslog_facility = LOG_DAEMON;
 546 #ifdef RTE_LIBEAL_USE_HPET
 547         internal_config.no_hpet = 0;
 548 #else
 549         internal_config.no_hpet = 1;
 550 #endif
 551         /* zero out the NUMA config */
 552         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 553                 internal_config.socket_mem[i] = 0;
 554
 555         /* zero out hugedir descriptors */
 556         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 557                 internal_config.hugepage_info[i].lock_descriptor = 0;
 558
 559         internal_config.vmware_tsc_map = 0;
 560
 561         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 562                                   lgopts, &option_index)) != EOF) {
 563
 564                 switch (opt) {
 565                 /* blacklist */
 566                 case 'b':
 567                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 568                                         optarg) < 0) {
 569                                 eal_usage(prgname);
 570                                 return (-1);
 571                         }
 572                         break;
 573                 /* coremask */
 574                 case 'c':
 575                         if (eal_parse_coremask(optarg) < 0) {
 576                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 577                                 eal_usage(prgname);
 578                                 return -1;
 579                         }
 580                         coremask_ok = 1;
 581                         break;
 582                 /* size of memory */
 583                 case 'm':
 584                         internal_config.memory = atoi(optarg);
 585                         internal_config.memory *= 1024ULL;
 586                         internal_config.memory *= 1024ULL;
 587                         break;
 588                 /* force number of channels */
 589                 case 'n':
 590                         internal_config.force_nchannel = atoi(optarg);
 591                         if (internal_config.force_nchannel == 0 ||
 592                             internal_config.force_nchannel > 4) {
 593                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 594                                 eal_usage(prgname);
 595                                 return -1;
 596                         }
 597                         break;
 598                 /* force number of ranks */
 599                 case 'r':
 600                         internal_config.force_nrank = atoi(optarg);
 601                         if (internal_config.force_nrank == 0 ||
 602                             internal_config.force_nrank > 16) {
 603                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 604                                 eal_usage(prgname);
 605                                 return -1;
 606                         }
 607                         break;
 608                 case 'v':
 609                         /* since message is explicitly requested by user, we
 610                          * write message at highest log level so it can always be seen
 611                          * even if info or warning messages are disabled */
 612                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 613                         break;
 614
 615                 /* long options */
 616                 case 0:
 617                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 618                                 internal_config.no_hugetlbfs = 1;
 619                         }
 620                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 621                                 internal_config.no_pci = 1;
 622                         }
 623                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 624                                 internal_config.no_hpet = 1;
 625                         }
 626                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 627                                 internal_config.vmware_tsc_map = 1;
 628                         }
 629                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 630                                 internal_config.no_shconf = 1;
 631                         }
 632                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 633                                 RTE_LOG(ERR, EAL, "Option "OPT_HUGE_DIR" is not supported on"
 634                                                 "FreeBSD\n");
 635                                 return -1;
 636                         }
 637                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 638                                 internal_config.process_type = eal_parse_proc_type(optarg);
 639                         }
 640                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 641                                 RTE_LOG(ERR, EAL, "Option "OPT_FILE_PREFIX" is not supported on"
 642                                                 "FreeBSD\n");
 643                                 return -1;
 644                         }
 645                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 646                                 RTE_LOG(ERR, EAL, "Option "OPT_SOCKET_MEM" is not supported on"
 647                                                 "FreeBSD\n");
 648                                 return -1;
 649                         }
 650                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 651                                 if (eal_parse_use_device(optarg) < 0) {
 652                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 653                                                 OPT_USE_DEVICE "\n");
 654                                         eal_usage(prgname);
 655                                         return -1;
 656                                 }
 657                         }
 658                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 659                                 if (eal_parse_syslog(optarg) < 0) {
 660                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 661                                                         OPT_SYSLOG "\n");
 662                                         eal_usage(prgname);
 663                                         return -1;
 664                                 }
 665                         }
 666                         break;
 667
 668                 default:
 669                         eal_usage(prgname);
 670                         return -1;
 671                 }
 672         }
 673
 674         /* sanity checks */
 675         if (!coremask_ok) {
 676                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 677                 eal_usage(prgname);
 678                 return -1;
 679         }
 680         if (internal_config.process_type == RTE_PROC_AUTO){
 681                 internal_config.process_type = eal_proc_type_detect();
 682         }
 683         if (internal_config.process_type == RTE_PROC_INVALID){
 684                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 685                 eal_usage(prgname);
 686                 return -1;
 687         }
 688         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 689                         internal_config.force_nchannel == 0) {
 690                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 691                 eal_usage(prgname);
 692                 return -1;
 693         }
 694         if (index(internal_config.hugefile_prefix,'%') != NULL){
 695                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 696                 eal_usage(prgname);
 697                 return -1;
 698         }
 699         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 700                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 701                                 "at the same time\n");
 702                 eal_usage(prgname);
 703                 return -1;
 704         }
 705         /* --no-huge doesn't make sense with either -m or --socket-mem */
 706         if (internal_config.no_hugetlbfs &&
 707                         (internal_config.memory > 0 ||
 708                                         internal_config.force_sockets == 1)) {
 709                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 710                                 "together with --no-huge!\n");
 711                 eal_usage(prgname);
 712                 return -1;
 713         }
 714
 715         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 716                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 717                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 718                         "[--use-device] options cannot be used at the same time\n");
 719                 eal_usage(prgname);
 720                 return -1;
 721         }
 722
 723         if (optind >= 0)
 724                 argv[optind-1] = prgname;
 725
 726         /* if no memory amounts were requested, this will result in 0 and
 727          * will be overriden later, right after eal_hugepage_info_init() */
 728         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 729                 internal_config.memory += internal_config.socket_mem[i];
 730
 731         ret = optind-1;
 732         optind = 0; /* reset getopt lib */
 733         return ret;
 734 }
 735
 736 static void
 737 eal_check_mem_on_local_socket(void)
 738 {
 739         const struct rte_memseg *ms;
 740         int i, socket_id;
 741
 742         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 743
 744         ms = rte_eal_get_physmem_layout();
 745
 746         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 747                 if (ms[i].socket_id == socket_id &&
 748                                 ms[i].len > 0)
 749                         return;
 750
 751         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 752                         "memory on local socket!\n");
 753 }
 754
 755 static int
 756 sync_func(__attribute__((unused)) void *arg)
 757 {
 758         return 0;
 759 }
 760
 761 inline static void
 762 rte_eal_mcfg_complete(void)
 763 {
 764         /* ALL shared mem_config related INIT DONE */
 765         if (rte_config.process_type == RTE_PROC_PRIMARY)
 766                 rte_config.mem_config->magic = RTE_MAGIC;
 767 }
 768
 769 /* return non-zero if hugepages are enabled. */
 770 int rte_eal_has_hugepages(void)
 771 {
 772         return !internal_config.no_hugetlbfs;
 773 }
 774
 775 /* Abstraction for port I/0 privilage */
 776 static int
 777 rte_eal_iopl_init(void)
 778 {
 779         int fd = -1;
 780         fd = open("/dev/io", O_RDWR);
 781         if (fd < 0)
 782                 return -1;
 783         return 0;
 784 }
 785
 786 /* Launch threads, called at application init(). */
 787 int
 788 rte_eal_init(int argc, char **argv)
 789 {
 790         int i, fctret, ret;
 791         pthread_t thread_id;
 792         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 793
 794         if (!rte_atomic32_test_and_set(&run_once))
 795                 return -1;
 796
 797         thread_id = pthread_self();
 798
 799         if (rte_eal_log_early_init() < 0)
 800                 rte_panic("Cannot init early logs\n");
 801
 802         fctret = eal_parse_args(argc, argv);
 803         if (fctret < 0)
 804                 exit(1);
 805
 806         if (internal_config.no_hugetlbfs == 0 &&
 807                         internal_config.process_type != RTE_PROC_SECONDARY &&
 808                         eal_hugepage_info_init() < 0)
 809                 rte_panic("Cannot get hugepage information\n");
 810
 811         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 812                 if (internal_config.no_hugetlbfs)
 813                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 814                 else
 815                         internal_config.memory = eal_get_hugepage_mem_size();
 816         }
 817
 818         if (internal_config.vmware_tsc_map == 1) {
 819 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 820                 rte_cycles_vmware_tsc_map = 1;
 821                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 822                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 823 #else
 824                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 825                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 826 #endif
 827         }
 828
 829         rte_srand(rte_rdtsc());
 830
 831         rte_config_init();
 832
 833         if (rte_eal_iopl_init() == 0)
 834                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 835
 836         if (rte_eal_cpu_init() < 0)
 837                 rte_panic("Cannot detect lcores\n");
 838
 839         if (rte_eal_memory_init() < 0)
 840                 rte_panic("Cannot init memory\n");
 841
 842         if (rte_eal_memzone_init() < 0)
 843                 rte_panic("Cannot init memzone\n");
 844
 845         if (rte_eal_tailqs_init() < 0)
 846                 rte_panic("Cannot init tail queues for objects\n");
 847
 848 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 849                 rte_panic("Cannot init logs\n");*/
 850
 851         if (rte_eal_alarm_init() < 0)
 852                 rte_panic("Cannot init interrupt-handling thread\n");
 853
 854         if (rte_eal_intr_init() < 0)
 855                 rte_panic("Cannot init interrupt-handling thread\n");
 856
 857         if (rte_eal_timer_init() < 0)
 858                 rte_panic("Cannot init HPET or TSC timers\n");
 859
 860         if (rte_eal_pci_init() < 0)
 861                 rte_panic("Cannot init PCI\n");
 862
 863         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 864                 rte_config.master_lcore, thread_id);
 865
 866         eal_check_mem_on_local_socket();
 867
 868         rte_eal_mcfg_complete();
 869
 870         if (rte_eal_vdev_init() < 0)
 871                 rte_panic("Cannot init virtual devices\n");
 872
 873         RTE_LCORE_FOREACH_SLAVE(i) {
 874
 875                 /*
 876                  * create communication pipes between master thread
 877                  * and children
 878                  */
 879                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 880                         rte_panic("Cannot create pipe\n");
 881                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 882                         rte_panic("Cannot create pipe\n");
 883
 884                 lcore_config[i].state = WAIT;
 885
 886                 /* create a thread for each lcore */
 887                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 888                                      eal_thread_loop, NULL);
 889                 if (ret != 0)
 890                         rte_panic("Cannot create thread\n");
 891         }
 892
 893         eal_thread_init_master(rte_config.master_lcore);
 894
 895         /*
 896          * Launch a dummy function on all slave lcores, so that master lcore
 897          * knows they are all ready when this function returns.
 898          */
 899         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 900         rte_eal_mp_wait_lcore();
 901
 902         return fctret;
 903 }
 904
 905 /* get core role */
 906 enum rte_lcore_role_t
 907 rte_eal_lcore_role(unsigned lcore_id)
 908 {
 909         return (rte_config.lcore_role[lcore_id]);
 910 }
 911
 912 enum rte_proc_type_t
 913 rte_eal_process_type(void)
 914 {
 915         return (rte_config.process_type);
 916 }
 917