lib/librte_eal/linuxapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdio.h>
  35 #include <stdlib.h>
  36 #include <stdint.h>
  37 #include <string.h>
  38 #include <stdarg.h>
  39 #include <unistd.h>
  40 #include <pthread.h>
  41 #include <syslog.h>
  42 #include <getopt.h>
  43 #include <sys/file.h>
  44 #include <stddef.h>
  45 #include <errno.h>
  46 #include <limits.h>
  47 #include <errno.h>
  48 #include <sys/mman.h>
  49 #include <sys/queue.h>
  50 #include <sys/user.h>
  51 #include <linux/binfmts.h>
  52
  53 #include <rte_common.h>
  54 #include <rte_debug.h>
  55 #include <rte_memory.h>
  56 #include <rte_memzone.h>
  57 #include <rte_launch.h>
  58 #include <rte_tailq.h>
  59 #include <rte_eal.h>
  60 #include <rte_eal_memconfig.h>
  61 #include <rte_per_lcore.h>
  62 #include <rte_lcore.h>
  63 #include <rte_log.h>
  64 #include <rte_random.h>
  65 #include <rte_cycles.h>
  66 #include <rte_string_fns.h>
  67 #include <rte_cpuflags.h>
  68 #include <rte_interrupts.h>
  69 #include <rte_pci.h>
  70 #include <rte_common.h>
  71 #include <rte_version.h>
  72 #include <rte_atomic.h>
  73 #include <malloc_heap.h>
  74
  75 #include "eal_private.h"
  76 #include "eal_thread.h"
  77 #include "eal_internal_cfg.h"
  78 #include "eal_filesystem.h"
  79 #include "eal_hugepages.h"
  80
  81 #define OPT_HUGE_DIR    "huge-dir"
  82 #define OPT_PROC_TYPE   "proc-type"
  83 #define OPT_NO_SHCONF   "no-shconf"
  84 #define OPT_NO_HPET     "no-hpet"
  85 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  86 #define OPT_NO_PCI      "no-pci"
  87 #define OPT_NO_HUGE     "no-huge"
  88 #define OPT_FILE_PREFIX "file-prefix"
  89 #define OPT_SOCKET_MEM  "socket-mem"
  90 #define OPT_SYSLOG      "syslog"
  91
  92 #define RTE_EAL_BLACKLIST_SIZE  0x100
  93
  94 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  95
  96 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
  97
  98 #define BITS_PER_HEX 4
  99
 100 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 101 {                                                               \
 102         unsigned long val;                                      \
 103         char *end;                                              \
 104         errno = 0;                                              \
 105         val = strtoul((in), &end, 16);                          \
 106         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 107                 return (-EINVAL);                               \
 108         (fd) = (typeof (fd))val;                                \
 109         (in) = end + 1;                                         \
 110 }
 111
 112 /* Allow the application to print its usage message too if set */
 113 static rte_usage_hook_t rte_application_usage_hook = NULL;
 114 /* early configuration structure, when memory config is not mmapped */
 115 static struct rte_mem_config early_mem_config;
 116
 117 /* define fd variable here, because file needs to be kept open for the
 118  * duration of the program, as we hold a write lock on it in the primary proc */
 119 static int mem_cfg_fd = -1;
 120
 121 static struct flock wr_lock = {
 122                 .l_type = F_WRLCK,
 123                 .l_whence = SEEK_SET,
 124                 .l_start = offsetof(struct rte_mem_config, memseg),
 125                 .l_len = sizeof(early_mem_config.memseg),
 126 };
 127
 128 /* Address of global and public configuration */
 129 static struct rte_config rte_config = {
 130                 .mem_config = &early_mem_config,
 131 };
 132
 133 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 134
 135 /* internal configuration (per-core) */
 136 struct lcore_config lcore_config[RTE_MAX_LCORE];
 137
 138 /* internal configuration */
 139 struct internal_config internal_config;
 140
 141 /* used by rte_rdtsc() */
 142 int rte_cycles_vmware_tsc_map;
 143
 144 /* Return a pointer to the configuration structure */
 145 struct rte_config *
 146 rte_eal_get_configuration(void)
 147 {
 148         return &rte_config;
 149 }
 150
 151 /* parse a sysfs (or other) file containing one integer value */
 152 int
 153 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 154 {
 155         FILE *f;
 156         char buf[BUFSIZ];
 157         char *end = NULL;
 158
 159         if ((f = fopen(filename, "r")) == NULL) {
 160                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 161                         __func__, filename);
 162                 return -1;
 163         }
 164
 165         if (fgets(buf, sizeof(buf), f) == NULL) {
 166                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 167                         __func__, filename);
 168                 fclose(f);
 169                 return -1;
 170         }
 171         *val = strtoul(buf, &end, 0);
 172         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 173                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 174                                 __func__, filename);
 175                 fclose(f);
 176                 return -1;
 177         }
 178         fclose(f);
 179         return 0;
 180 }
 181
 182
 183 /* create memory configuration in shared/mmap memory. Take out
 184  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 185  * This means we never close the file while running (auto-close on exit).
 186  * We also don't lock the whole file, so that in future we can use read-locks
 187  * on other parts, e.g. memzones, to detect if there are running secondary
 188  * processes. */
 189 static void
 190 rte_eal_config_create(void)
 191 {
 192         void *rte_mem_cfg_addr;
 193         int retval;
 194
 195         const char *pathname = eal_runtime_config_path();
 196
 197         if (internal_config.no_shconf)
 198                 return;
 199
 200         if (mem_cfg_fd < 0){
 201                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 202                 if (mem_cfg_fd < 0)
 203                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 204         }
 205
 206         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 207         if (retval < 0){
 208                 close(mem_cfg_fd);
 209                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 210         }
 211
 212         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 213         if (retval < 0){
 214                 close(mem_cfg_fd);
 215                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 216                                 "process running?\n", pathname);
 217         }
 218
 219         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 220                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 221
 222         if (rte_mem_cfg_addr == MAP_FAILED){
 223                 rte_panic("Cannot mmap memory for rte_config\n");
 224         }
 225         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 226         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 227 }
 228
 229 /* attach to an existing shared memory config */
 230 static void
 231 rte_eal_config_attach(void)
 232 {
 233         void *rte_mem_cfg_addr;
 234         const char *pathname = eal_runtime_config_path();
 235
 236         if (internal_config.no_shconf)
 237                 return;
 238
 239         if (mem_cfg_fd < 0){
 240                 mem_cfg_fd = open(pathname, O_RDWR);
 241                 if (mem_cfg_fd < 0)
 242                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 243         }
 244
 245         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 246                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 247         close(mem_cfg_fd);
 248         if (rte_mem_cfg_addr == MAP_FAILED)
 249                 rte_panic("Cannot mmap memory for rte_config\n");
 250
 251         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 252 }
 253
 254 /* Detect if we are a primary or a secondary process */
 255 static enum rte_proc_type_t
 256 eal_proc_type_detect(void)
 257 {
 258         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 259         const char *pathname = eal_runtime_config_path();
 260
 261         /* if we can open the file but not get a write-lock we are a secondary
 262          * process. NOTE: if we get a file handle back, we keep that open
 263          * and don't close it to prevent a race condition between multiple opens */
 264         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 265                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 266                 ptype = RTE_PROC_SECONDARY;
 267
 268         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 269                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 270
 271         return ptype;
 272 }
 273
 274 /* Sets up rte_config structure with the pointer to shared memory config.*/
 275 static void
 276 rte_config_init(void)
 277 {
 278         /* set the magic in configuration structure */
 279         rte_config.magic = RTE_MAGIC;
 280         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 281                         eal_proc_type_detect() : /* for auto, detect the type */
 282                         internal_config.process_type; /* otherwise use what's already set */
 283
 284         switch (rte_config.process_type){
 285         case RTE_PROC_PRIMARY:
 286                 rte_eal_config_create();
 287                 break;
 288         case RTE_PROC_SECONDARY:
 289                 rte_eal_config_attach();
 290                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 291                 break;
 292         case RTE_PROC_AUTO:
 293         case RTE_PROC_INVALID:
 294                 rte_panic("Invalid process type\n");
 295         }
 296 }
 297
 298 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 299 static void
 300 eal_hugedirs_unlock(void)
 301 {
 302         int i;
 303
 304         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 305         {
 306                 /* skip uninitialized */
 307                 if (internal_config.hugepage_info[i].lock_descriptor == 0)
 308                         continue;
 309                 /* unlock hugepage file */
 310                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 311                 close(internal_config.hugepage_info[i].lock_descriptor);
 312                 /* reset the field */
 313                 internal_config.hugepage_info[i].lock_descriptor = 0;
 314         }
 315 }
 316
 317 /* display usage */
 318 static void
 319 eal_usage(const char *prgname)
 320 {
 321         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 322                "[--proc-type primary|secondary|auto] \n\n"
 323                "EAL options:\n"
 324                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 325                "  -n NUM       : Number of memory channels\n"
 326                    "  -v           : Display version information on startup\n"
 327                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 328            "PCI device\n"
 329                "                 (multiple -b options are allowed)\n"
 330                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 331                "  -r NUM       : force number of memory ranks (don't detect)\n"
 332                "  --"OPT_SYSLOG"     : set syslog facility\n"
 333                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 334                    "                 sockets (use comma separated values)\n"
 335                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 336                "  --"OPT_PROC_TYPE"  : type of this process\n"
 337                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 338                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 339                            "native RDTSC\n"
 340                "\nEAL options for DEBUG use only:\n"
 341                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 342                "  --"OPT_NO_PCI"   : disable pci\n"
 343                "  --"OPT_NO_HPET"  : disable hpet\n"
 344                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 345                "\n",
 346                prgname);
 347         /* Allow the application to print its usage message too if hook is set */
 348         if ( rte_application_usage_hook ) {
 349                 printf("===== Application Usage =====\n\n");
 350                 rte_application_usage_hook(prgname);
 351         }
 352 }
 353
 354 /* Set a per-application usage message */
 355 rte_usage_hook_t
 356 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 357 {
 358         rte_usage_hook_t        old_func;
 359
 360         /* Will be NULL on the first call to denote the last usage routine. */
 361         old_func                                        = rte_application_usage_hook;
 362         rte_application_usage_hook      = usage_func;
 363
 364         return old_func;
 365 }
 366
 367 /*
 368  * Parse the coremask given as argument (hexadecimal string) and fill
 369  * the global configuration (core role and core count) with the parsed
 370  * value.
 371  */
 372 static int xdigit2val(unsigned char c)
 373 {
 374         int val;
 375         if(isdigit(c))
 376                 val = c - '0';
 377         else if(isupper(c))
 378                 val = c - 'A' + 10;
 379         else
 380                 val = c - 'a' + 10;
 381         return val;
 382 }
 383 static int
 384 eal_parse_coremask(const char *coremask)
 385 {
 386         struct rte_config *cfg = rte_eal_get_configuration();
 387         int i, j, idx = 0 ;
 388         unsigned count = 0;
 389         char c;
 390         int val;
 391
 392         if (coremask == NULL)
 393                 return -1;
 394         /* Remove all blank characters ahead and after .
 395          * Remove 0x/0X if exists.
 396          */
 397         while (isblank(*coremask))
 398                 coremask++;
 399         if (coremask[0] == '0' && ((coremask[1] == 'x')
 400                 ||  (coremask[1] == 'X')) )
 401                 coremask += 2;
 402         i = strnlen(coremask, MAX_ARG_STRLEN);
 403         while ((i > 0) && isblank(coremask[i - 1]))
 404                 i--;
 405         if (i == 0)
 406                 return -1;
 407
 408         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 409                 c = coremask[i];
 410                 if (isxdigit(c) == 0) {
 411                         /* invalid characters */
 412                         return (-1);
 413                 }
 414                 val = xdigit2val(c);
 415                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 416                         if((1 << j) & val) {
 417                                 cfg->lcore_role[idx] = ROLE_RTE;
 418                                 if(count == 0)
 419                                         cfg->master_lcore = idx;
 420                                 count++;
 421                         } else  {
 422                                 cfg->lcore_role[idx] = ROLE_OFF;
 423                         }
 424                 }
 425         }
 426         for(; i >= 0; i--)
 427                 if(coremask[i] != '0')
 428                         return -1;
 429         for(; idx < RTE_MAX_LCORE; idx++)
 430                 cfg->lcore_role[idx] = ROLE_OFF;
 431         if(count == 0)
 432                 return -1;
 433         return 0;
 434 }
 435
 436 static int
 437 eal_parse_syslog(const char *facility)
 438 {
 439         int i;
 440         static struct {
 441                 const char *name;
 442                 int value;
 443         } map[] = {
 444                 { "auth", LOG_AUTH },
 445                 { "cron", LOG_CRON },
 446                 { "daemon", LOG_DAEMON },
 447                 { "ftp", LOG_FTP },
 448                 { "kern", LOG_KERN },
 449                 { "lpr", LOG_LPR },
 450                 { "mail", LOG_MAIL },
 451                 { "news", LOG_NEWS },
 452                 { "syslog", LOG_SYSLOG },
 453                 { "user", LOG_USER },
 454                 { "uucp", LOG_UUCP },
 455                 { "local0", LOG_LOCAL0 },
 456                 { "local1", LOG_LOCAL1 },
 457                 { "local2", LOG_LOCAL2 },
 458                 { "local3", LOG_LOCAL3 },
 459                 { "local4", LOG_LOCAL4 },
 460                 { "local5", LOG_LOCAL5 },
 461                 { "local6", LOG_LOCAL6 },
 462                 { "local7", LOG_LOCAL7 },
 463                 { NULL, 0 }
 464         };
 465
 466         for (i = 0; map[i].name; i++) {
 467                 if (!strcmp(facility, map[i].name)) {
 468                         internal_config.syslog_facility = map[i].value;
 469                         return 0;
 470                 }
 471         }
 472         return -1;
 473 }
 474
 475 static int
 476 eal_parse_socket_mem(char *socket_mem)
 477 {
 478         char * arg[RTE_MAX_NUMA_NODES];
 479         char *end;
 480         int arg_num, i, len;
 481         uint64_t total_mem = 0;
 482
 483         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 484         if (len == SOCKET_MEM_STRLEN) {
 485                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 486                 return -1;
 487         }
 488
 489         /* all other error cases will be caught later */
 490         if (!isdigit(socket_mem[len-1]))
 491                 return -1;
 492
 493         /* split the optarg into separate socket values */
 494         arg_num = rte_strsplit(socket_mem, len,
 495                         arg, RTE_MAX_NUMA_NODES, ',');
 496
 497         /* if split failed, or 0 arguments */
 498         if (arg_num <= 0)
 499                 return -1;
 500
 501         internal_config.force_sockets = 1;
 502
 503         /* parse each defined socket option */
 504         errno = 0;
 505         for (i = 0; i < arg_num; i++) {
 506                 end = NULL;
 507                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 508
 509                 /* check for invalid input */
 510                 if ((errno != 0)  ||
 511                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 512                         return -1;
 513                 internal_config.socket_mem[i] *= 1024ULL;
 514                 internal_config.socket_mem[i] *= 1024ULL;
 515                 total_mem += internal_config.socket_mem[i];
 516         }
 517
 518         /* check if we have a positive amount of total memory */
 519         if (total_mem == 0)
 520                 return -1;
 521
 522         return 0;
 523 }
 524
 525 static inline size_t
 526 eal_get_hugepage_mem_size(void)
 527 {
 528         uint64_t size = 0;
 529         unsigned i, j;
 530
 531         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 532                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 533                 if (hpi->hugedir != NULL) {
 534                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 535                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 536                         }
 537                 }
 538         }
 539
 540         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 541 }
 542
 543 static enum rte_proc_type_t
 544 eal_parse_proc_type(const char *arg)
 545 {
 546         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 547                 return RTE_PROC_PRIMARY;
 548         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 549                 return RTE_PROC_SECONDARY;
 550         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 551                 return RTE_PROC_AUTO;
 552
 553         return RTE_PROC_INVALID;
 554 }
 555
 556 static int
 557 eal_parse_blacklist(const char *input,  struct rte_pci_addr *dev2bl)
 558 {
 559         GET_BLACKLIST_FIELD(input, dev2bl->domain, UINT16_MAX, ':');
 560         GET_BLACKLIST_FIELD(input, dev2bl->bus, UINT8_MAX, ':');
 561         GET_BLACKLIST_FIELD(input, dev2bl->devid, UINT8_MAX, '.');
 562         GET_BLACKLIST_FIELD(input, dev2bl->function, UINT8_MAX, 0);
 563         return (0);
 564 }
 565
 566 static ssize_t
 567 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 568 {
 569         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 570                 RTE_LOG(ERR, EAL,
 571                     "%s - too many devices to blacklist...\n",
 572                     optarg);
 573                 return (-EINVAL);
 574         } else if (eal_parse_blacklist(optarg, eal_dev_blacklist + idx) != 0) {
 575                 RTE_LOG(ERR, EAL,
 576                     "%s - invalid device to blacklist...\n",
 577                     optarg);
 578                 return (-EINVAL);
 579         }
 580
 581         idx += 1;
 582         return (idx);
 583 }
 584
 585 /* Parse the argument given in the command line of the application */
 586 static int
 587 eal_parse_args(int argc, char **argv)
 588 {
 589         int opt, ret, i;
 590         char **argvopt;
 591         int option_index;
 592         int coremask_ok = 0;
 593         ssize_t blacklist_index = 0;;
 594         char *prgname = argv[0];
 595         static struct option lgopts[] = {
 596                 {OPT_NO_HUGE, 0, 0, 0},
 597                 {OPT_NO_PCI, 0, 0, 0},
 598                 {OPT_NO_HPET, 0, 0, 0},
 599                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 600                 {OPT_HUGE_DIR, 1, 0, 0},
 601                 {OPT_NO_SHCONF, 0, 0, 0},
 602                 {OPT_PROC_TYPE, 1, 0, 0},
 603                 {OPT_FILE_PREFIX, 1, 0, 0},
 604                 {OPT_SOCKET_MEM, 1, 0, 0},
 605                 {OPT_SYSLOG, 1, NULL, 0},
 606                 {0, 0, 0, 0}
 607         };
 608
 609         argvopt = argv;
 610
 611         internal_config.memory = 0;
 612         internal_config.force_nrank = 0;
 613         internal_config.force_nchannel = 0;
 614         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 615         internal_config.hugepage_dir = NULL;
 616         internal_config.force_sockets = 0;
 617         internal_config.syslog_facility = LOG_DAEMON;
 618 #ifdef RTE_LIBEAL_USE_HPET
 619         internal_config.no_hpet = 0;
 620 #else
 621         internal_config.no_hpet = 1;
 622 #endif
 623         /* zero out the NUMA config */
 624         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 625                 internal_config.socket_mem[i] = 0;
 626
 627         /* zero out hugedir descriptors */
 628         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 629                 internal_config.hugepage_info[i].lock_descriptor = 0;
 630
 631         internal_config.vmware_tsc_map = 0;
 632
 633         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 634                                   lgopts, &option_index)) != EOF) {
 635
 636                 switch (opt) {
 637                 /* blacklist */
 638                 case 'b':
 639                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 640                             blacklist_index)) < 0) {
 641                                 eal_usage(prgname);
 642                                 return (-1);
 643                         }
 644                         break;
 645                 /* coremask */
 646                 case 'c':
 647                         if (eal_parse_coremask(optarg) < 0) {
 648                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 649                                 eal_usage(prgname);
 650                                 return -1;
 651                         }
 652                         coremask_ok = 1;
 653                         break;
 654                 /* size of memory */
 655                 case 'm':
 656                         internal_config.memory = atoi(optarg);
 657                         internal_config.memory *= 1024ULL;
 658                         internal_config.memory *= 1024ULL;
 659                         break;
 660                 /* force number of channels */
 661                 case 'n':
 662                         internal_config.force_nchannel = atoi(optarg);
 663                         if (internal_config.force_nchannel == 0 ||
 664                             internal_config.force_nchannel > 4) {
 665                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 666                                 eal_usage(prgname);
 667                                 return -1;
 668                         }
 669                         break;
 670                 /* force number of ranks */
 671                 case 'r':
 672                         internal_config.force_nrank = atoi(optarg);
 673                         if (internal_config.force_nrank == 0 ||
 674                             internal_config.force_nrank > 16) {
 675                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 676                                 eal_usage(prgname);
 677                                 return -1;
 678                         }
 679                         break;
 680                 case 'v':
 681                         /* since message is explicitly requested by user, we
 682                          * write message at highest log level so it can always be seen
 683                          * even if info or warning messages are disabled */
 684                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 685                         break;
 686
 687                 /* long options */
 688                 case 0:
 689                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 690                                 internal_config.no_hugetlbfs = 1;
 691                         }
 692                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 693                                 internal_config.no_pci = 1;
 694                         }
 695                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 696                                 internal_config.no_hpet = 1;
 697                         }
 698                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 699                                 internal_config.vmware_tsc_map = 1;
 700                         }
 701                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 702                                 internal_config.no_shconf = 1;
 703                         }
 704                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 705                                 internal_config.hugepage_dir = optarg;
 706                         }
 707                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 708                                 internal_config.process_type = eal_parse_proc_type(optarg);
 709                         }
 710                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 711                                 internal_config.hugefile_prefix = optarg;
 712                         }
 713                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 714                                 if (eal_parse_socket_mem(optarg) < 0) {
 715                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 716                                                         OPT_SOCKET_MEM "\n");
 717                                         eal_usage(prgname);
 718                                         return -1;
 719                                 }
 720                         }
 721                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 722                                 if (eal_parse_syslog(optarg) < 0) {
 723                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 724                                                         OPT_SYSLOG "\n");
 725                                         eal_usage(prgname);
 726                                         return -1;
 727                                 }
 728                         }
 729                         break;
 730
 731                 default:
 732                         eal_usage(prgname);
 733                         return -1;
 734                 }
 735         }
 736
 737         /* sanity checks */
 738         if (!coremask_ok) {
 739                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 740                 eal_usage(prgname);
 741                 return -1;
 742         }
 743         if (internal_config.process_type == RTE_PROC_AUTO){
 744                 internal_config.process_type = eal_proc_type_detect();
 745         }
 746         if (internal_config.process_type == RTE_PROC_INVALID){
 747                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 748                 eal_usage(prgname);
 749                 return -1;
 750         }
 751         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 752                         internal_config.force_nchannel == 0) {
 753                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 754                 eal_usage(prgname);
 755                 return -1;
 756         }
 757         if (index(internal_config.hugefile_prefix,'%') != NULL){
 758                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 759                 eal_usage(prgname);
 760                 return -1;
 761         }
 762         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 763                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 764                                 "at the same time\n");
 765                 eal_usage(prgname);
 766                 return -1;
 767         }
 768         /* --no-huge doesn't make sense with either -m or --socket-mem */
 769         if (internal_config.no_hugetlbfs &&
 770                         (internal_config.memory > 0 ||
 771                                         internal_config.force_sockets == 1)) {
 772                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 773                                 "together with --no-huge!\n");
 774                 eal_usage(prgname);
 775                 return -1;
 776         }
 777
 778         if (blacklist_index > 0)
 779                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 780
 781         if (optind >= 0)
 782                 argv[optind-1] = prgname;
 783
 784         /* if no memory amounts were requested, this will result in 0 and
 785          * will be overriden later, right after eal_hugepage_info_init() */
 786         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 787                 internal_config.memory += internal_config.socket_mem[i];
 788
 789         ret = optind-1;
 790         optind = 0; /* reset getopt lib */
 791         return ret;
 792 }
 793
 794 static void
 795 eal_check_mem_on_local_socket(void)
 796 {
 797         const struct rte_memseg *ms;
 798         int i, socket_id;
 799
 800         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 801
 802         ms = rte_eal_get_physmem_layout();
 803
 804         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 805                 if (ms[i].socket_id == socket_id &&
 806                                 ms[i].len > 0)
 807                         return;
 808
 809         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 810                         "memory on local socket!\n");
 811 }
 812
 813 static int
 814 sync_func(__attribute__((unused)) void *arg)
 815 {
 816         return 0;
 817 }
 818
 819 inline static void
 820 rte_eal_mcfg_complete(void)
 821 {
 822         /* ALL shared mem_config related INIT DONE */
 823         if (rte_config.process_type == RTE_PROC_PRIMARY)
 824                 rte_config.mem_config->magic = RTE_MAGIC;
 825 }
 826
 827 /* Launch threads, called at application init(). */
 828 int
 829 rte_eal_init(int argc, char **argv)
 830 {
 831         int i, fctret, ret;
 832         pthread_t thread_id;
 833         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 834
 835         if (!rte_atomic32_test_and_set(&run_once))
 836                 return -1;
 837
 838         thread_id = pthread_self();
 839
 840         if (rte_eal_log_early_init() < 0)
 841                 rte_panic("Cannot init early logs\n");
 842
 843         fctret = eal_parse_args(argc, argv);
 844         if (fctret < 0)
 845                 exit(1);
 846
 847         if (internal_config.no_hugetlbfs == 0 &&
 848                         internal_config.process_type != RTE_PROC_SECONDARY &&
 849                         eal_hugepage_info_init() < 0)
 850                 rte_panic("Cannot get hugepage information\n");
 851
 852         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 853                 if (internal_config.no_hugetlbfs)
 854                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 855                 else
 856                         internal_config.memory = eal_get_hugepage_mem_size();
 857         }
 858
 859         if (internal_config.vmware_tsc_map == 1) {
 860 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 861                 rte_cycles_vmware_tsc_map = 1;
 862                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 863                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 864 #else
 865                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 866                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 867 #endif
 868         }
 869
 870         rte_srand(rte_rdtsc());
 871
 872         rte_config_init();
 873
 874         if (rte_eal_cpu_init() < 0)
 875                 rte_panic("Cannot detect lcores\n");
 876
 877         if (rte_eal_memory_init() < 0)
 878                 rte_panic("Cannot init memory\n");
 879
 880         /* the directories are locked during eal_hugepage_info_init */
 881         eal_hugedirs_unlock();
 882
 883         if (rte_eal_memzone_init() < 0)
 884                 rte_panic("Cannot init memzone\n");
 885
 886         if (rte_eal_tailqs_init() < 0)
 887                 rte_panic("Cannot init tail queues for objects\n");
 888
 889         if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 890                 rte_panic("Cannot init logs\n");
 891
 892         if (rte_eal_alarm_init() < 0)
 893                 rte_panic("Cannot init interrupt-handling thread\n");
 894
 895         if (rte_eal_intr_init() < 0)
 896                 rte_panic("Cannot init interrupt-handling thread\n");
 897
 898         if (rte_eal_timer_init() < 0)
 899                 rte_panic("Cannot init HPET or TSC timers\n");
 900
 901         if (rte_eal_pci_init() < 0)
 902                 rte_panic("Cannot init PCI\n");
 903
 904         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
 905                 rte_config.master_lcore, (int)thread_id);
 906
 907         eal_check_mem_on_local_socket();
 908
 909         rte_eal_mcfg_complete();
 910
 911         RTE_LCORE_FOREACH_SLAVE(i) {
 912
 913                 /*
 914                  * create communication pipes between master thread
 915                  * and children
 916                  */
 917                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 918                         rte_panic("Cannot create pipe\n");
 919                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 920                         rte_panic("Cannot create pipe\n");
 921
 922                 lcore_config[i].state = WAIT;
 923
 924                 /* create a thread for each lcore */
 925                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 926                                      eal_thread_loop, NULL);
 927                 if (ret != 0)
 928                         rte_panic("Cannot create thread\n");
 929         }
 930
 931         eal_thread_init_master(rte_config.master_lcore);
 932
 933         /*
 934          * Launch a dummy function on all slave lcores, so that master lcore
 935          * knows they are all ready when this function returns.
 936          */
 937         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 938         rte_eal_mp_wait_lcore();
 939
 940         return fctret;
 941 }
 942
 943 /* get core role */
 944 enum rte_lcore_role_t
 945 rte_eal_lcore_role(unsigned lcore_id)
 946 {
 947         return (rte_config.lcore_role[lcore_id]);
 948 }
 949
 950 enum rte_proc_type_t
 951 rte_eal_process_type(void)
 952 {
 953         return (rte_config.process_type);
 954 }
 955