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         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 284                         eal_proc_type_detect() : /* for auto, detect the type */
 285                         internal_config.process_type; /* otherwise use what's already set */
 286
 287         switch (rte_config.process_type){
 288         case RTE_PROC_PRIMARY:
 289                 rte_eal_config_create();
 290                 break;
 291         case RTE_PROC_SECONDARY:
 292                 rte_eal_config_attach();
 293                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 294                 break;
 295         case RTE_PROC_AUTO:
 296         case RTE_PROC_INVALID:
 297                 rte_panic("Invalid process type\n");
 298         }
 299 }
 300
 301 /* display usage */
 302 static void
 303 eal_usage(const char *prgname)
 304 {
 305         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 306                "[--proc-type primary|secondary|auto] \n\n"
 307                "EAL options:\n"
 308                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 309                "  -n NUM       : Number of memory channels\n"
 310                "  -v           : Display version information on startup\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_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 315                "               Prevent EAL from using this PCI device. The argument\n"
 316                "               format is <domain:bus:devid.func>.\n"
 317                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 318                "               Only use the specified PCI devices. The argument format\n"
 319                "               is <[domain:]bus:devid.func>. This option can be present\n"
 320                "               several times (once per device).\n"
 321                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 322                "  --"OPT_VDEV": add a virtual device.\n"
 323                "               The argument format is <driver><id>[,key=val,...]\n"
 324                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 325                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 326                "\nEAL options for DEBUG use only:\n"
 327                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 328                "  --"OPT_NO_PCI"   : disable pci\n"
 329                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 330                "\n",
 331                prgname);
 332         /* Allow the application to print its usage message too if hook is set */
 333         if ( rte_application_usage_hook ) {
 334                 printf("===== Application Usage =====\n\n");
 335                 rte_application_usage_hook(prgname);
 336         }
 337 }
 338
 339 /* Set a per-application usage message */
 340 rte_usage_hook_t
 341 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 342 {
 343         rte_usage_hook_t        old_func;
 344
 345         /* Will be NULL on the first call to denote the last usage routine. */
 346         old_func                                        = rte_application_usage_hook;
 347         rte_application_usage_hook      = usage_func;
 348
 349         return old_func;
 350 }
 351
 352 /*
 353  * Parse the coremask given as argument (hexadecimal string) and fill
 354  * the global configuration (core role and core count) with the parsed
 355  * value.
 356  */
 357 static int xdigit2val(unsigned char c)
 358 {
 359         int val;
 360         if(isdigit(c))
 361                 val = c - '0';
 362         else if(isupper(c))
 363                 val = c - 'A' + 10;
 364         else
 365                 val = c - 'a' + 10;
 366         return val;
 367 }
 368 static int
 369 eal_parse_coremask(const char *coremask)
 370 {
 371         struct rte_config *cfg = rte_eal_get_configuration();
 372         int i, j, idx = 0 ;
 373         unsigned count = 0;
 374         char c;
 375         int val;
 376
 377         if (coremask == NULL)
 378                 return -1;
 379         /* Remove all blank characters ahead and after .
 380          * Remove 0x/0X if exists.
 381          */
 382         while (isblank(*coremask))
 383                 coremask++;
 384         if (coremask[0] == '0' && ((coremask[1] == 'x')
 385                 ||  (coremask[1] == 'X')) )
 386                 coremask += 2;
 387         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 388         while ((i > 0) && isblank(coremask[i - 1]))
 389                 i--;
 390         if (i == 0)
 391                 return -1;
 392
 393         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 394                 c = coremask[i];
 395                 if (isxdigit(c) == 0) {
 396                         /* invalid characters */
 397                         return (-1);
 398                 }
 399                 val = xdigit2val(c);
 400                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 401                         if((1 << j) & val) {
 402                                 cfg->lcore_role[idx] = ROLE_RTE;
 403                                 if(count == 0)
 404                                         cfg->master_lcore = idx;
 405                                 count++;
 406                         } else  {
 407                                 cfg->lcore_role[idx] = ROLE_OFF;
 408                         }
 409                 }
 410         }
 411         for(; i >= 0; i--)
 412                 if(coremask[i] != '0')
 413                         return -1;
 414         for(; idx < RTE_MAX_LCORE; idx++)
 415                 cfg->lcore_role[idx] = ROLE_OFF;
 416         if(count == 0)
 417                 return -1;
 418         return 0;
 419 }
 420
 421 static int
 422 eal_parse_syslog(const char *facility)
 423 {
 424         int i;
 425         static struct {
 426                 const char *name;
 427                 int value;
 428         } map[] = {
 429                 { "auth", LOG_AUTH },
 430                 { "cron", LOG_CRON },
 431                 { "daemon", LOG_DAEMON },
 432                 { "ftp", LOG_FTP },
 433                 { "kern", LOG_KERN },
 434                 { "lpr", LOG_LPR },
 435                 { "mail", LOG_MAIL },
 436                 { "news", LOG_NEWS },
 437                 { "syslog", LOG_SYSLOG },
 438                 { "user", LOG_USER },
 439                 { "uucp", LOG_UUCP },
 440                 { "local0", LOG_LOCAL0 },
 441                 { "local1", LOG_LOCAL1 },
 442                 { "local2", LOG_LOCAL2 },
 443                 { "local3", LOG_LOCAL3 },
 444                 { "local4", LOG_LOCAL4 },
 445                 { "local5", LOG_LOCAL5 },
 446                 { "local6", LOG_LOCAL6 },
 447                 { "local7", LOG_LOCAL7 },
 448                 { NULL, 0 }
 449         };
 450
 451         for (i = 0; map[i].name; i++) {
 452                 if (!strcmp(facility, map[i].name)) {
 453                         internal_config.syslog_facility = map[i].value;
 454                         return 0;
 455                 }
 456         }
 457         return -1;
 458 }
 459
 460 static inline size_t
 461 eal_get_hugepage_mem_size(void)
 462 {
 463         uint64_t size = 0;
 464         unsigned i, j;
 465
 466         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 467                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 468                 if (hpi->hugedir != NULL) {
 469                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 470                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 471                         }
 472                 }
 473         }
 474
 475         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 476 }
 477
 478 static enum rte_proc_type_t
 479 eal_parse_proc_type(const char *arg)
 480 {
 481         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 482                 return RTE_PROC_PRIMARY;
 483         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 484                 return RTE_PROC_SECONDARY;
 485         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 486                 return RTE_PROC_AUTO;
 487
 488         return RTE_PROC_INVALID;
 489 }
 490
 491 /* Parse the argument given in the command line of the application */
 492 static int
 493 eal_parse_args(int argc, char **argv)
 494 {
 495         int opt, ret, i;
 496         char **argvopt;
 497         int option_index;
 498         int coremask_ok = 0;
 499         char *prgname = argv[0];
 500         static struct option lgopts[] = {
 501                 {OPT_NO_HUGE, 0, 0, 0},
 502                 {OPT_NO_PCI, 0, 0, 0},
 503                 {OPT_NO_HPET, 0, 0, 0},
 504                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 505                 {OPT_HUGE_DIR, 1, 0, 0},
 506                 {OPT_NO_SHCONF, 0, 0, 0},
 507                 {OPT_PROC_TYPE, 1, 0, 0},
 508                 {OPT_FILE_PREFIX, 1, 0, 0},
 509                 {OPT_SOCKET_MEM, 1, 0, 0},
 510                 {OPT_PCI_WHITELIST, 1, 0, 0},
 511                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 512                 {OPT_VDEV, 1, 0, 0},
 513                 {OPT_SYSLOG, 1, NULL, 0},
 514                 {0, 0, 0, 0}
 515         };
 516
 517         argvopt = argv;
 518
 519         internal_config.memory = 0;
 520         internal_config.force_nrank = 0;
 521         internal_config.force_nchannel = 0;
 522         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 523         internal_config.hugepage_dir = NULL;
 524         internal_config.force_sockets = 0;
 525         internal_config.syslog_facility = LOG_DAEMON;
 526 #ifdef RTE_LIBEAL_USE_HPET
 527         internal_config.no_hpet = 0;
 528 #else
 529         internal_config.no_hpet = 1;
 530 #endif
 531         /* zero out the NUMA config */
 532         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 533                 internal_config.socket_mem[i] = 0;
 534
 535         /* zero out hugedir descriptors */
 536         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 537                 internal_config.hugepage_info[i].lock_descriptor = 0;
 538
 539         internal_config.vmware_tsc_map = 0;
 540
 541         while ((opt = getopt_long(argc, argvopt, "b:w:c:m:n:r:v",
 542                                   lgopts, &option_index)) != EOF) {
 543
 544                 switch (opt) {
 545                 /* blacklist */
 546                 case 'b':
 547                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 548                                         optarg) < 0) {
 549                                 eal_usage(prgname);
 550                                 return (-1);
 551                         }
 552                         break;
 553                 /* whitelist */
 554                 case 'w':
 555                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 556                                         optarg) < 0) {
 557                                 eal_usage(prgname);
 558                                 return -1;
 559                         }
 560                         break;
 561                 /* coremask */
 562                 case 'c':
 563                         if (eal_parse_coremask(optarg) < 0) {
 564                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 565                                 eal_usage(prgname);
 566                                 return -1;
 567                         }
 568                         coremask_ok = 1;
 569                         break;
 570                 /* size of memory */
 571                 case 'm':
 572                         internal_config.memory = atoi(optarg);
 573                         internal_config.memory *= 1024ULL;
 574                         internal_config.memory *= 1024ULL;
 575                         break;
 576                 /* force number of channels */
 577                 case 'n':
 578                         internal_config.force_nchannel = atoi(optarg);
 579                         if (internal_config.force_nchannel == 0 ||
 580                             internal_config.force_nchannel > 4) {
 581                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 582                                 eal_usage(prgname);
 583                                 return -1;
 584                         }
 585                         break;
 586                 /* force number of ranks */
 587                 case 'r':
 588                         internal_config.force_nrank = atoi(optarg);
 589                         if (internal_config.force_nrank == 0 ||
 590                             internal_config.force_nrank > 16) {
 591                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 592                                 eal_usage(prgname);
 593                                 return -1;
 594                         }
 595                         break;
 596                 case 'v':
 597                         /* since message is explicitly requested by user, we
 598                          * write message at highest log level so it can always be seen
 599                          * even if info or warning messages are disabled */
 600                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 601                         break;
 602
 603                 /* long options */
 604                 case 0:
 605                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 606                                 internal_config.no_hugetlbfs = 1;
 607                         }
 608                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 609                                 internal_config.no_pci = 1;
 610                         }
 611                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 612                                 internal_config.no_hpet = 1;
 613                         }
 614                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 615                                 internal_config.vmware_tsc_map = 1;
 616                         }
 617                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 618                                 internal_config.no_shconf = 1;
 619                         }
 620                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 621                                 RTE_LOG(ERR, EAL, "Option "OPT_HUGE_DIR" is not supported on"
 622                                                 "FreeBSD\n");
 623                                 return -1;
 624                         }
 625                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 626                                 internal_config.process_type = eal_parse_proc_type(optarg);
 627                         }
 628                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 629                                 RTE_LOG(ERR, EAL, "Option "OPT_FILE_PREFIX" is not supported on"
 630                                                 "FreeBSD\n");
 631                                 return -1;
 632                         }
 633                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 634                                 RTE_LOG(ERR, EAL, "Option "OPT_SOCKET_MEM" is not supported on"
 635                                                 "FreeBSD\n");
 636                                 return -1;
 637                         }
 638                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 639                                 printf("The --use-device option is deprecated, please use\n"
 640                                         "--whitelist or --vdev instead.\n");
 641                                 eal_usage(prgname);
 642                                 return -1;
 643                         }
 644                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 645                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 646                                                 optarg) < 0) {
 647                                         eal_usage(prgname);
 648                                         return -1;
 649                                 }
 650                         }
 651                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 652                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 653                                                 optarg) < 0) {
 654                                         eal_usage(prgname);
 655                                         return -1;
 656                                 }
 657                         }
 658                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 659                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 660                                                 optarg) < 0) {
 661                                         eal_usage(prgname);
 662                                         return -1;
 663                                 }
 664                         }
 665                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 666                                 if (eal_parse_syslog(optarg) < 0) {
 667                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 668                                                         OPT_SYSLOG "\n");
 669                                         eal_usage(prgname);
 670                                         return -1;
 671                                 }
 672                         }
 673                         break;
 674
 675                 default:
 676                         eal_usage(prgname);
 677                         return -1;
 678                 }
 679         }
 680
 681         /* sanity checks */
 682         if (!coremask_ok) {
 683                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 684                 eal_usage(prgname);
 685                 return -1;
 686         }
 687         if (internal_config.process_type == RTE_PROC_AUTO){
 688                 internal_config.process_type = eal_proc_type_detect();
 689         }
 690         if (internal_config.process_type == RTE_PROC_INVALID){
 691                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 692                 eal_usage(prgname);
 693                 return -1;
 694         }
 695         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 696                         internal_config.force_nchannel == 0) {
 697                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 698                 eal_usage(prgname);
 699                 return -1;
 700         }
 701         if (index(internal_config.hugefile_prefix,'%') != NULL){
 702                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 703                 eal_usage(prgname);
 704                 return -1;
 705         }
 706         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 707                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 708                                 "at the same time\n");
 709                 eal_usage(prgname);
 710                 return -1;
 711         }
 712         /* --no-huge doesn't make sense with either -m or --socket-mem */
 713         if (internal_config.no_hugetlbfs &&
 714                         (internal_config.memory > 0 ||
 715                                         internal_config.force_sockets == 1)) {
 716                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 717                                 "together with --no-huge!\n");
 718                 eal_usage(prgname);
 719                 return -1;
 720         }
 721
 722         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 723                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 724                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 725                         "[-w] options cannot be used at the same time\n");
 726                 eal_usage(prgname);
 727                 return -1;
 728         }
 729
 730         if (optind >= 0)
 731                 argv[optind-1] = prgname;
 732
 733         /* if no memory amounts were requested, this will result in 0 and
 734          * will be overriden later, right after eal_hugepage_info_init() */
 735         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 736                 internal_config.memory += internal_config.socket_mem[i];
 737
 738         ret = optind-1;
 739         optind = 0; /* reset getopt lib */
 740         return ret;
 741 }
 742
 743 static void
 744 eal_check_mem_on_local_socket(void)
 745 {
 746         const struct rte_memseg *ms;
 747         int i, socket_id;
 748
 749         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 750
 751         ms = rte_eal_get_physmem_layout();
 752
 753         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 754                 if (ms[i].socket_id == socket_id &&
 755                                 ms[i].len > 0)
 756                         return;
 757
 758         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 759                         "memory on local socket!\n");
 760 }
 761
 762 static int
 763 sync_func(__attribute__((unused)) void *arg)
 764 {
 765         return 0;
 766 }
 767
 768 inline static void
 769 rte_eal_mcfg_complete(void)
 770 {
 771         /* ALL shared mem_config related INIT DONE */
 772         if (rte_config.process_type == RTE_PROC_PRIMARY)
 773                 rte_config.mem_config->magic = RTE_MAGIC;
 774 }
 775
 776 /* return non-zero if hugepages are enabled. */
 777 int rte_eal_has_hugepages(void)
 778 {
 779         return !internal_config.no_hugetlbfs;
 780 }
 781
 782 /* Abstraction for port I/0 privilege */
 783 static int
 784 rte_eal_iopl_init(void)
 785 {
 786         int fd = -1;
 787         fd = open("/dev/io", O_RDWR);
 788         if (fd < 0)
 789                 return -1;
 790         return 0;
 791 }
 792
 793 /* Launch threads, called at application init(). */
 794 int
 795 rte_eal_init(int argc, char **argv)
 796 {
 797         int i, fctret, ret;
 798         pthread_t thread_id;
 799         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 800
 801         if (!rte_atomic32_test_and_set(&run_once))
 802                 return -1;
 803
 804         thread_id = pthread_self();
 805
 806         if (rte_eal_log_early_init() < 0)
 807                 rte_panic("Cannot init early logs\n");
 808
 809         fctret = eal_parse_args(argc, argv);
 810         if (fctret < 0)
 811                 exit(1);
 812
 813         if (internal_config.no_hugetlbfs == 0 &&
 814                         internal_config.process_type != RTE_PROC_SECONDARY &&
 815                         eal_hugepage_info_init() < 0)
 816                 rte_panic("Cannot get hugepage information\n");
 817
 818         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 819                 if (internal_config.no_hugetlbfs)
 820                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 821                 else
 822                         internal_config.memory = eal_get_hugepage_mem_size();
 823         }
 824
 825         if (internal_config.vmware_tsc_map == 1) {
 826 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 827                 rte_cycles_vmware_tsc_map = 1;
 828                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 829                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 830 #else
 831                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 832                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 833 #endif
 834         }
 835
 836         rte_srand(rte_rdtsc());
 837
 838         rte_config_init();
 839
 840         if (rte_eal_iopl_init() == 0)
 841                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 842
 843         if (rte_eal_cpu_init() < 0)
 844                 rte_panic("Cannot detect lcores\n");
 845
 846         if (rte_eal_memory_init() < 0)
 847                 rte_panic("Cannot init memory\n");
 848
 849         if (rte_eal_memzone_init() < 0)
 850                 rte_panic("Cannot init memzone\n");
 851
 852         if (rte_eal_tailqs_init() < 0)
 853                 rte_panic("Cannot init tail queues for objects\n");
 854
 855 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 856                 rte_panic("Cannot init logs\n");*/
 857
 858         if (rte_eal_alarm_init() < 0)
 859                 rte_panic("Cannot init interrupt-handling thread\n");
 860
 861         if (rte_eal_intr_init() < 0)
 862                 rte_panic("Cannot init interrupt-handling thread\n");
 863
 864         if (rte_eal_timer_init() < 0)
 865                 rte_panic("Cannot init HPET or TSC timers\n");
 866
 867         if (rte_eal_pci_init() < 0)
 868                 rte_panic("Cannot init PCI\n");
 869
 870         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 871                 rte_config.master_lcore, thread_id);
 872
 873         eal_check_mem_on_local_socket();
 874
 875         rte_eal_mcfg_complete();
 876
 877         if (rte_eal_dev_init() < 0)
 878                 rte_panic("Cannot init pmd devices\n");
 879
 880         RTE_LCORE_FOREACH_SLAVE(i) {
 881
 882                 /*
 883                  * create communication pipes between master thread
 884                  * and children
 885                  */
 886                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 887                         rte_panic("Cannot create pipe\n");
 888                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 889                         rte_panic("Cannot create pipe\n");
 890
 891                 lcore_config[i].state = WAIT;
 892
 893                 /* create a thread for each lcore */
 894                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 895                                      eal_thread_loop, NULL);
 896                 if (ret != 0)
 897                         rte_panic("Cannot create thread\n");
 898         }
 899
 900         eal_thread_init_master(rte_config.master_lcore);
 901
 902         /*
 903          * Launch a dummy function on all slave lcores, so that master lcore
 904          * knows they are all ready when this function returns.
 905          */
 906         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 907         rte_eal_mp_wait_lcore();
 908
 909         return fctret;
 910 }
 911
 912 /* get core role */
 913 enum rte_lcore_role_t
 914 rte_eal_lcore_role(unsigned lcore_id)
 915 {
 916         return (rte_config.lcore_role[lcore_id]);
 917 }
 918
 919 enum rte_proc_type_t
 920 rte_eal_process_type(void)
 921 {
 922         return (rte_config.process_type);
 923 }
 924