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_PCI_WHITELIST "pci-whitelist"
  93 #define OPT_PCI_BLACKLIST "pci-blacklist"
  94 #define OPT_VDEV        "vdev"
  95 #define OPT_SYSLOG      "syslog"
  96
  97 #define RTE_EAL_BLACKLIST_SIZE  0x100
  98
  99 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 100
 101 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 102
 103 #define HIGHEST_RPL 3
 104
 105 #define BITS_PER_HEX 4
 106
 107 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 108 {                                                               \
 109         unsigned long val;                                      \
 110         char *end;                                              \
 111         errno = 0;                                              \
 112         val = strtoul((in), &end, 16);                          \
 113         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 114                 return (-EINVAL);                               \
 115         (fd) = (typeof (fd))val;                                \
 116         (in) = end + 1;                                         \
 117 }
 118
 119 /* Allow the application to print its usage message too if set */
 120 static rte_usage_hook_t rte_application_usage_hook = NULL;
 121 /* early configuration structure, when memory config is not mmapped */
 122 static struct rte_mem_config early_mem_config;
 123
 124 /* define fd variable here, because file needs to be kept open for the
 125  * duration of the program, as we hold a write lock on it in the primary proc */
 126 static int mem_cfg_fd = -1;
 127
 128 static struct flock wr_lock = {
 129                 .l_type = F_WRLCK,
 130                 .l_whence = SEEK_SET,
 131                 .l_start = offsetof(struct rte_mem_config, memseg),
 132                 .l_len = sizeof(early_mem_config.memseg),
 133 };
 134
 135 /* Address of global and public configuration */
 136 static struct rte_config rte_config = {
 137                 .mem_config = &early_mem_config,
 138 };
 139
 140 /* internal configuration (per-core) */
 141 struct lcore_config lcore_config[RTE_MAX_LCORE];
 142
 143 /* internal configuration */
 144 struct internal_config internal_config;
 145
 146 /* used by rte_rdtsc() */
 147 int rte_cycles_vmware_tsc_map;
 148
 149 /* Return a pointer to the configuration structure */
 150 struct rte_config *
 151 rte_eal_get_configuration(void)
 152 {
 153         return &rte_config;
 154 }
 155
 156 /* parse a sysfs (or other) file containing one integer value */
 157 int
 158 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 159 {
 160         FILE *f;
 161         char buf[BUFSIZ];
 162         char *end = NULL;
 163
 164         if ((f = fopen(filename, "r")) == NULL) {
 165                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 166                         __func__, filename);
 167                 return -1;
 168         }
 169
 170         if (fgets(buf, sizeof(buf), f) == NULL) {
 171                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 172                         __func__, filename);
 173                 fclose(f);
 174                 return -1;
 175         }
 176         *val = strtoul(buf, &end, 0);
 177         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 178                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 179                                 __func__, filename);
 180                 fclose(f);
 181                 return -1;
 182         }
 183         fclose(f);
 184         return 0;
 185 }
 186
 187
 188 /* create memory configuration in shared/mmap memory. Take out
 189  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 190  * This means we never close the file while running (auto-close on exit).
 191  * We also don't lock the whole file, so that in future we can use read-locks
 192  * on other parts, e.g. memzones, to detect if there are running secondary
 193  * processes. */
 194 static void
 195 rte_eal_config_create(void)
 196 {
 197         void *rte_mem_cfg_addr;
 198         int retval;
 199
 200         const char *pathname = eal_runtime_config_path();
 201
 202         if (internal_config.no_shconf)
 203                 return;
 204
 205         if (mem_cfg_fd < 0){
 206                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 207                 if (mem_cfg_fd < 0)
 208                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 209         }
 210
 211         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 212         if (retval < 0){
 213                 close(mem_cfg_fd);
 214                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 215         }
 216
 217         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 218         if (retval < 0){
 219                 close(mem_cfg_fd);
 220                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 221                                 "process running?\n", pathname);
 222         }
 223
 224         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 225                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 226
 227         if (rte_mem_cfg_addr == MAP_FAILED){
 228                 rte_panic("Cannot mmap memory for rte_config\n");
 229         }
 230         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 231         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 232 }
 233
 234 /* attach to an existing shared memory config */
 235 static void
 236 rte_eal_config_attach(void)
 237 {
 238         void *rte_mem_cfg_addr;
 239         const char *pathname = eal_runtime_config_path();
 240
 241         if (internal_config.no_shconf)
 242                 return;
 243
 244         if (mem_cfg_fd < 0){
 245                 mem_cfg_fd = open(pathname, O_RDWR);
 246                 if (mem_cfg_fd < 0)
 247                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 248         }
 249
 250         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 251                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 252         close(mem_cfg_fd);
 253         if (rte_mem_cfg_addr == MAP_FAILED)
 254                 rte_panic("Cannot mmap memory for rte_config\n");
 255
 256         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 257 }
 258
 259 /* Detect if we are a primary or a secondary process */
 260 static enum rte_proc_type_t
 261 eal_proc_type_detect(void)
 262 {
 263         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 264         const char *pathname = eal_runtime_config_path();
 265
 266         /* if we can open the file but not get a write-lock we are a secondary
 267          * process. NOTE: if we get a file handle back, we keep that open
 268          * and don't close it to prevent a race condition between multiple opens */
 269         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 270                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 271                 ptype = RTE_PROC_SECONDARY;
 272
 273         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 274                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 275
 276         return ptype;
 277 }
 278
 279 /* Sets up rte_config structure with the pointer to shared memory config.*/
 280 static void
 281 rte_config_init(void)
 282 {
 283         /* set the magic in configuration structure */
 284         rte_config.magic = RTE_MAGIC;
 285         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 286                         eal_proc_type_detect() : /* for auto, detect the type */
 287                         internal_config.process_type; /* otherwise use what's already set */
 288
 289         switch (rte_config.process_type){
 290         case RTE_PROC_PRIMARY:
 291                 rte_eal_config_create();
 292                 break;
 293         case RTE_PROC_SECONDARY:
 294                 rte_eal_config_attach();
 295                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 296                 break;
 297         case RTE_PROC_AUTO:
 298         case RTE_PROC_INVALID:
 299                 rte_panic("Invalid process type\n");
 300         }
 301 }
 302
 303 /* display usage */
 304 static void
 305 eal_usage(const char *prgname)
 306 {
 307         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 308                "[--proc-type primary|secondary|auto] \n\n"
 309                "EAL options:\n"
 310                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 311                "  -n NUM       : Number of memory channels\n"
 312                "  -v           : Display version information on startup\n"
 313                "  -m MB        : memory to allocate\n"
 314                "  -r NUM       : force number of memory ranks (don't detect)\n"
 315                "  --"OPT_PROC_TYPE"  : type of this process\n"
 316                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 317                "               Prevent EAL from using this PCI device. The argument\n"
 318                "               format is <domain:bus:devid.func>.\n"
 319                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 320                "               Only use the specified PCI devices. The argument format\n"
 321                "               is <[domain:]bus:devid.func>. This option can be present\n"
 322                "               several times (once per device).\n"
 323                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 324                "  --"OPT_VDEV": add a virtual device.\n"
 325                "               The argument format is <driver><id>[,key=val,...]\n"
 326                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 327                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 328                            "native RDTSC\n"
 329                "\nEAL options for DEBUG use only:\n"
 330                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 331                "  --"OPT_NO_PCI"   : disable pci\n"
 332                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 333                "\n",
 334                prgname);
 335         /* Allow the application to print its usage message too if hook is set */
 336         if ( rte_application_usage_hook ) {
 337                 printf("===== Application Usage =====\n\n");
 338                 rte_application_usage_hook(prgname);
 339         }
 340 }
 341
 342 /* Set a per-application usage message */
 343 rte_usage_hook_t
 344 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 345 {
 346         rte_usage_hook_t        old_func;
 347
 348         /* Will be NULL on the first call to denote the last usage routine. */
 349         old_func                                        = rte_application_usage_hook;
 350         rte_application_usage_hook      = usage_func;
 351
 352         return old_func;
 353 }
 354
 355 /*
 356  * Parse the coremask given as argument (hexadecimal string) and fill
 357  * the global configuration (core role and core count) with the parsed
 358  * value.
 359  */
 360 static int xdigit2val(unsigned char c)
 361 {
 362         int val;
 363         if(isdigit(c))
 364                 val = c - '0';
 365         else if(isupper(c))
 366                 val = c - 'A' + 10;
 367         else
 368                 val = c - 'a' + 10;
 369         return val;
 370 }
 371 static int
 372 eal_parse_coremask(const char *coremask)
 373 {
 374         struct rte_config *cfg = rte_eal_get_configuration();
 375         int i, j, idx = 0 ;
 376         unsigned count = 0;
 377         char c;
 378         int val;
 379
 380         if (coremask == NULL)
 381                 return -1;
 382         /* Remove all blank characters ahead and after .
 383          * Remove 0x/0X if exists.
 384          */
 385         while (isblank(*coremask))
 386                 coremask++;
 387         if (coremask[0] == '0' && ((coremask[1] == 'x')
 388                 ||  (coremask[1] == 'X')) )
 389                 coremask += 2;
 390         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 391         while ((i > 0) && isblank(coremask[i - 1]))
 392                 i--;
 393         if (i == 0)
 394                 return -1;
 395
 396         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 397                 c = coremask[i];
 398                 if (isxdigit(c) == 0) {
 399                         /* invalid characters */
 400                         return (-1);
 401                 }
 402                 val = xdigit2val(c);
 403                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 404                         if((1 << j) & val) {
 405                                 cfg->lcore_role[idx] = ROLE_RTE;
 406                                 if(count == 0)
 407                                         cfg->master_lcore = idx;
 408                                 count++;
 409                         } else  {
 410                                 cfg->lcore_role[idx] = ROLE_OFF;
 411                         }
 412                 }
 413         }
 414         for(; i >= 0; i--)
 415                 if(coremask[i] != '0')
 416                         return -1;
 417         for(; idx < RTE_MAX_LCORE; idx++)
 418                 cfg->lcore_role[idx] = ROLE_OFF;
 419         if(count == 0)
 420                 return -1;
 421         return 0;
 422 }
 423
 424 static int
 425 eal_parse_syslog(const char *facility)
 426 {
 427         int i;
 428         static struct {
 429                 const char *name;
 430                 int value;
 431         } map[] = {
 432                 { "auth", LOG_AUTH },
 433                 { "cron", LOG_CRON },
 434                 { "daemon", LOG_DAEMON },
 435                 { "ftp", LOG_FTP },
 436                 { "kern", LOG_KERN },
 437                 { "lpr", LOG_LPR },
 438                 { "mail", LOG_MAIL },
 439                 { "news", LOG_NEWS },
 440                 { "syslog", LOG_SYSLOG },
 441                 { "user", LOG_USER },
 442                 { "uucp", LOG_UUCP },
 443                 { "local0", LOG_LOCAL0 },
 444                 { "local1", LOG_LOCAL1 },
 445                 { "local2", LOG_LOCAL2 },
 446                 { "local3", LOG_LOCAL3 },
 447                 { "local4", LOG_LOCAL4 },
 448                 { "local5", LOG_LOCAL5 },
 449                 { "local6", LOG_LOCAL6 },
 450                 { "local7", LOG_LOCAL7 },
 451                 { NULL, 0 }
 452         };
 453
 454         for (i = 0; map[i].name; i++) {
 455                 if (!strcmp(facility, map[i].name)) {
 456                         internal_config.syslog_facility = map[i].value;
 457                         return 0;
 458                 }
 459         }
 460         return -1;
 461 }
 462
 463 static inline size_t
 464 eal_get_hugepage_mem_size(void)
 465 {
 466         uint64_t size = 0;
 467         unsigned i, j;
 468
 469         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 470                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 471                 if (hpi->hugedir != NULL) {
 472                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 473                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 474                         }
 475                 }
 476         }
 477
 478         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 479 }
 480
 481 static enum rte_proc_type_t
 482 eal_parse_proc_type(const char *arg)
 483 {
 484         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 485                 return RTE_PROC_PRIMARY;
 486         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 487                 return RTE_PROC_SECONDARY;
 488         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 489                 return RTE_PROC_AUTO;
 490
 491         return RTE_PROC_INVALID;
 492 }
 493
 494 /* Parse the argument given in the command line of the application */
 495 static int
 496 eal_parse_args(int argc, char **argv)
 497 {
 498         int opt, ret, i;
 499         char **argvopt;
 500         int option_index;
 501         int coremask_ok = 0;
 502         char *prgname = argv[0];
 503         static struct option lgopts[] = {
 504                 {OPT_NO_HUGE, 0, 0, 0},
 505                 {OPT_NO_PCI, 0, 0, 0},
 506                 {OPT_NO_HPET, 0, 0, 0},
 507                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 508                 {OPT_HUGE_DIR, 1, 0, 0},
 509                 {OPT_NO_SHCONF, 0, 0, 0},
 510                 {OPT_PROC_TYPE, 1, 0, 0},
 511                 {OPT_FILE_PREFIX, 1, 0, 0},
 512                 {OPT_SOCKET_MEM, 1, 0, 0},
 513                 {OPT_PCI_WHITELIST, 1, 0, 0},
 514                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 515                 {OPT_VDEV, 1, 0, 0},
 516                 {OPT_SYSLOG, 1, NULL, 0},
 517                 {0, 0, 0, 0}
 518         };
 519
 520         argvopt = argv;
 521
 522         internal_config.memory = 0;
 523         internal_config.force_nrank = 0;
 524         internal_config.force_nchannel = 0;
 525         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 526         internal_config.hugepage_dir = NULL;
 527         internal_config.force_sockets = 0;
 528         internal_config.syslog_facility = LOG_DAEMON;
 529 #ifdef RTE_LIBEAL_USE_HPET
 530         internal_config.no_hpet = 0;
 531 #else
 532         internal_config.no_hpet = 1;
 533 #endif
 534         /* zero out the NUMA config */
 535         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 536                 internal_config.socket_mem[i] = 0;
 537
 538         /* zero out hugedir descriptors */
 539         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 540                 internal_config.hugepage_info[i].lock_descriptor = 0;
 541
 542         internal_config.vmware_tsc_map = 0;
 543
 544         while ((opt = getopt_long(argc, argvopt, "b:w:c:m:n:r:v",
 545                                   lgopts, &option_index)) != EOF) {
 546
 547                 switch (opt) {
 548                 /* blacklist */
 549                 case 'b':
 550                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 551                                         optarg) < 0) {
 552                                 eal_usage(prgname);
 553                                 return (-1);
 554                         }
 555                         break;
 556                 /* whitelist */
 557                 case 'w':
 558                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 559                                         optarg) < 0) {
 560                                 eal_usage(prgname);
 561                                 return -1;
 562                         }
 563                         break;
 564                 /* coremask */
 565                 case 'c':
 566                         if (eal_parse_coremask(optarg) < 0) {
 567                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 568                                 eal_usage(prgname);
 569                                 return -1;
 570                         }
 571                         coremask_ok = 1;
 572                         break;
 573                 /* size of memory */
 574                 case 'm':
 575                         internal_config.memory = atoi(optarg);
 576                         internal_config.memory *= 1024ULL;
 577                         internal_config.memory *= 1024ULL;
 578                         break;
 579                 /* force number of channels */
 580                 case 'n':
 581                         internal_config.force_nchannel = atoi(optarg);
 582                         if (internal_config.force_nchannel == 0 ||
 583                             internal_config.force_nchannel > 4) {
 584                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 585                                 eal_usage(prgname);
 586                                 return -1;
 587                         }
 588                         break;
 589                 /* force number of ranks */
 590                 case 'r':
 591                         internal_config.force_nrank = atoi(optarg);
 592                         if (internal_config.force_nrank == 0 ||
 593                             internal_config.force_nrank > 16) {
 594                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 595                                 eal_usage(prgname);
 596                                 return -1;
 597                         }
 598                         break;
 599                 case 'v':
 600                         /* since message is explicitly requested by user, we
 601                          * write message at highest log level so it can always be seen
 602                          * even if info or warning messages are disabled */
 603                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 604                         break;
 605
 606                 /* long options */
 607                 case 0:
 608                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 609                                 internal_config.no_hugetlbfs = 1;
 610                         }
 611                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 612                                 internal_config.no_pci = 1;
 613                         }
 614                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 615                                 internal_config.no_hpet = 1;
 616                         }
 617                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 618                                 internal_config.vmware_tsc_map = 1;
 619                         }
 620                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 621                                 internal_config.no_shconf = 1;
 622                         }
 623                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 624                                 RTE_LOG(ERR, EAL, "Option "OPT_HUGE_DIR" is not supported on"
 625                                                 "FreeBSD\n");
 626                                 return -1;
 627                         }
 628                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 629                                 internal_config.process_type = eal_parse_proc_type(optarg);
 630                         }
 631                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 632                                 RTE_LOG(ERR, EAL, "Option "OPT_FILE_PREFIX" is not supported on"
 633                                                 "FreeBSD\n");
 634                                 return -1;
 635                         }
 636                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 637                                 RTE_LOG(ERR, EAL, "Option "OPT_SOCKET_MEM" is not supported on"
 638                                                 "FreeBSD\n");
 639                                 return -1;
 640                         }
 641                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 642                                 printf("The --use-device option is deprecated, please use\n"
 643                                         "--whitelist or --vdev instead.\n");
 644                                 eal_usage(prgname);
 645                                 return -1;
 646                         }
 647                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 648                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 649                                                 optarg) < 0) {
 650                                         eal_usage(prgname);
 651                                         return -1;
 652                                 }
 653                         }
 654                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 655                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 656                                                 optarg) < 0) {
 657                                         eal_usage(prgname);
 658                                         return -1;
 659                                 }
 660                         }
 661                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 662                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 663                                                 optarg) < 0) {
 664                                         eal_usage(prgname);
 665                                         return -1;
 666                                 }
 667                         }
 668                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 669                                 if (eal_parse_syslog(optarg) < 0) {
 670                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 671                                                         OPT_SYSLOG "\n");
 672                                         eal_usage(prgname);
 673                                         return -1;
 674                                 }
 675                         }
 676                         break;
 677
 678                 default:
 679                         eal_usage(prgname);
 680                         return -1;
 681                 }
 682         }
 683
 684         /* sanity checks */
 685         if (!coremask_ok) {
 686                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 687                 eal_usage(prgname);
 688                 return -1;
 689         }
 690         if (internal_config.process_type == RTE_PROC_AUTO){
 691                 internal_config.process_type = eal_proc_type_detect();
 692         }
 693         if (internal_config.process_type == RTE_PROC_INVALID){
 694                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 695                 eal_usage(prgname);
 696                 return -1;
 697         }
 698         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 699                         internal_config.force_nchannel == 0) {
 700                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 701                 eal_usage(prgname);
 702                 return -1;
 703         }
 704         if (index(internal_config.hugefile_prefix,'%') != NULL){
 705                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 706                 eal_usage(prgname);
 707                 return -1;
 708         }
 709         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 710                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 711                                 "at the same time\n");
 712                 eal_usage(prgname);
 713                 return -1;
 714         }
 715         /* --no-huge doesn't make sense with either -m or --socket-mem */
 716         if (internal_config.no_hugetlbfs &&
 717                         (internal_config.memory > 0 ||
 718                                         internal_config.force_sockets == 1)) {
 719                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 720                                 "together with --no-huge!\n");
 721                 eal_usage(prgname);
 722                 return -1;
 723         }
 724
 725         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 726                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 727                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 728                         "[-w] options cannot be used at the same time\n");
 729                 eal_usage(prgname);
 730                 return -1;
 731         }
 732
 733         if (optind >= 0)
 734                 argv[optind-1] = prgname;
 735
 736         /* if no memory amounts were requested, this will result in 0 and
 737          * will be overriden later, right after eal_hugepage_info_init() */
 738         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 739                 internal_config.memory += internal_config.socket_mem[i];
 740
 741         ret = optind-1;
 742         optind = 0; /* reset getopt lib */
 743         return ret;
 744 }
 745
 746 static void
 747 eal_check_mem_on_local_socket(void)
 748 {
 749         const struct rte_memseg *ms;
 750         int i, socket_id;
 751
 752         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 753
 754         ms = rte_eal_get_physmem_layout();
 755
 756         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 757                 if (ms[i].socket_id == socket_id &&
 758                                 ms[i].len > 0)
 759                         return;
 760
 761         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 762                         "memory on local socket!\n");
 763 }
 764
 765 static int
 766 sync_func(__attribute__((unused)) void *arg)
 767 {
 768         return 0;
 769 }
 770
 771 inline static void
 772 rte_eal_mcfg_complete(void)
 773 {
 774         /* ALL shared mem_config related INIT DONE */
 775         if (rte_config.process_type == RTE_PROC_PRIMARY)
 776                 rte_config.mem_config->magic = RTE_MAGIC;
 777 }
 778
 779 /* return non-zero if hugepages are enabled. */
 780 int rte_eal_has_hugepages(void)
 781 {
 782         return !internal_config.no_hugetlbfs;
 783 }
 784
 785 /* Abstraction for port I/0 privilage */
 786 static int
 787 rte_eal_iopl_init(void)
 788 {
 789         int fd = -1;
 790         fd = open("/dev/io", O_RDWR);
 791         if (fd < 0)
 792                 return -1;
 793         return 0;
 794 }
 795
 796 /* Launch threads, called at application init(). */
 797 int
 798 rte_eal_init(int argc, char **argv)
 799 {
 800         int i, fctret, ret;
 801         pthread_t thread_id;
 802         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 803
 804         if (!rte_atomic32_test_and_set(&run_once))
 805                 return -1;
 806
 807         thread_id = pthread_self();
 808
 809         if (rte_eal_log_early_init() < 0)
 810                 rte_panic("Cannot init early logs\n");
 811
 812         fctret = eal_parse_args(argc, argv);
 813         if (fctret < 0)
 814                 exit(1);
 815
 816         if (internal_config.no_hugetlbfs == 0 &&
 817                         internal_config.process_type != RTE_PROC_SECONDARY &&
 818                         eal_hugepage_info_init() < 0)
 819                 rte_panic("Cannot get hugepage information\n");
 820
 821         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 822                 if (internal_config.no_hugetlbfs)
 823                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 824                 else
 825                         internal_config.memory = eal_get_hugepage_mem_size();
 826         }
 827
 828         if (internal_config.vmware_tsc_map == 1) {
 829 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 830                 rte_cycles_vmware_tsc_map = 1;
 831                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 832                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 833 #else
 834                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 835                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 836 #endif
 837         }
 838
 839         rte_srand(rte_rdtsc());
 840
 841         rte_config_init();
 842
 843         if (rte_eal_iopl_init() == 0)
 844                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 845
 846         if (rte_eal_cpu_init() < 0)
 847                 rte_panic("Cannot detect lcores\n");
 848
 849         if (rte_eal_memory_init() < 0)
 850                 rte_panic("Cannot init memory\n");
 851
 852         if (rte_eal_memzone_init() < 0)
 853                 rte_panic("Cannot init memzone\n");
 854
 855         if (rte_eal_tailqs_init() < 0)
 856                 rte_panic("Cannot init tail queues for objects\n");
 857
 858 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 859                 rte_panic("Cannot init logs\n");*/
 860
 861         if (rte_eal_alarm_init() < 0)
 862                 rte_panic("Cannot init interrupt-handling thread\n");
 863
 864         if (rte_eal_intr_init() < 0)
 865                 rte_panic("Cannot init interrupt-handling thread\n");
 866
 867         if (rte_eal_timer_init() < 0)
 868                 rte_panic("Cannot init HPET or TSC timers\n");
 869
 870         if (rte_eal_pci_init() < 0)
 871                 rte_panic("Cannot init PCI\n");
 872
 873         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 874                 rte_config.master_lcore, thread_id);
 875
 876         eal_check_mem_on_local_socket();
 877
 878         rte_eal_mcfg_complete();
 879
 880         if (rte_eal_vdev_init() < 0)
 881                 rte_panic("Cannot init virtual devices\n");
 882
 883         RTE_LCORE_FOREACH_SLAVE(i) {
 884
 885                 /*
 886                  * create communication pipes between master thread
 887                  * and children
 888                  */
 889                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 890                         rte_panic("Cannot create pipe\n");
 891                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 892                         rte_panic("Cannot create pipe\n");
 893
 894                 lcore_config[i].state = WAIT;
 895
 896                 /* create a thread for each lcore */
 897                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 898                                      eal_thread_loop, NULL);
 899                 if (ret != 0)
 900                         rte_panic("Cannot create thread\n");
 901         }
 902
 903         eal_thread_init_master(rte_config.master_lcore);
 904
 905         /*
 906          * Launch a dummy function on all slave lcores, so that master lcore
 907          * knows they are all ready when this function returns.
 908          */
 909         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 910         rte_eal_mp_wait_lcore();
 911
 912         return fctret;
 913 }
 914
 915 /* get core role */
 916 enum rte_lcore_role_t
 917 rte_eal_lcore_role(unsigned lcore_id)
 918 {
 919         return (rte_config.lcore_role[lcore_id]);
 920 }
 921
 922 enum rte_proc_type_t
 923 rte_eal_process_type(void)
 924 {
 925         return (rte_config.process_type);
 926 }
 927