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 #include <rte_eth_ring.h>
  75
  76 #include "eal_private.h"
  77 #include "eal_thread.h"
  78 #include "eal_internal_cfg.h"
  79 #include "eal_filesystem.h"
  80 #include "eal_hugepages.h"
  81
  82 #define OPT_HUGE_DIR    "huge-dir"
  83 #define OPT_PROC_TYPE   "proc-type"
  84 #define OPT_NO_SHCONF   "no-shconf"
  85 #define OPT_NO_HPET     "no-hpet"
  86 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  87 #define OPT_NO_PCI      "no-pci"
  88 #define OPT_NO_HUGE     "no-huge"
  89 #define OPT_FILE_PREFIX "file-prefix"
  90 #define OPT_SOCKET_MEM  "socket-mem"
  91 #define OPT_USE_DEVICE  "use-device"
  92 #define OPT_SYSLOG      "syslog"
  93
  94 #define RTE_EAL_BLACKLIST_SIZE  0x100
  95
  96 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  97
  98 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
  99
 100 #define BITS_PER_HEX 4
 101
 102 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 103 {                                                               \
 104         unsigned long val;                                      \
 105         char *end;                                              \
 106         errno = 0;                                              \
 107         val = strtoul((in), &end, 16);                          \
 108         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 109                 return (-EINVAL);                               \
 110         (fd) = (typeof (fd))val;                                \
 111         (in) = end + 1;                                         \
 112 }
 113
 114 /* Allow the application to print its usage message too if set */
 115 static rte_usage_hook_t rte_application_usage_hook = NULL;
 116 /* early configuration structure, when memory config is not mmapped */
 117 static struct rte_mem_config early_mem_config;
 118
 119 /* define fd variable here, because file needs to be kept open for the
 120  * duration of the program, as we hold a write lock on it in the primary proc */
 121 static int mem_cfg_fd = -1;
 122
 123 static struct flock wr_lock = {
 124                 .l_type = F_WRLCK,
 125                 .l_whence = SEEK_SET,
 126                 .l_start = offsetof(struct rte_mem_config, memseg),
 127                 .l_len = sizeof(early_mem_config.memseg),
 128 };
 129
 130 /* Address of global and public configuration */
 131 static struct rte_config rte_config = {
 132                 .mem_config = &early_mem_config,
 133 };
 134
 135 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 136
 137 /* internal configuration (per-core) */
 138 struct lcore_config lcore_config[RTE_MAX_LCORE];
 139
 140 /* internal configuration */
 141 struct internal_config internal_config;
 142
 143 /* used by rte_rdtsc() */
 144 int rte_cycles_vmware_tsc_map;
 145
 146 /* Return a pointer to the configuration structure */
 147 struct rte_config *
 148 rte_eal_get_configuration(void)
 149 {
 150         return &rte_config;
 151 }
 152
 153 /* parse a sysfs (or other) file containing one integer value */
 154 int
 155 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 156 {
 157         FILE *f;
 158         char buf[BUFSIZ];
 159         char *end = NULL;
 160
 161         if ((f = fopen(filename, "r")) == NULL) {
 162                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 163                         __func__, filename);
 164                 return -1;
 165         }
 166
 167         if (fgets(buf, sizeof(buf), f) == NULL) {
 168                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 169                         __func__, filename);
 170                 fclose(f);
 171                 return -1;
 172         }
 173         *val = strtoul(buf, &end, 0);
 174         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 175                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 176                                 __func__, filename);
 177                 fclose(f);
 178                 return -1;
 179         }
 180         fclose(f);
 181         return 0;
 182 }
 183
 184
 185 /* create memory configuration in shared/mmap memory. Take out
 186  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 187  * This means we never close the file while running (auto-close on exit).
 188  * We also don't lock the whole file, so that in future we can use read-locks
 189  * on other parts, e.g. memzones, to detect if there are running secondary
 190  * processes. */
 191 static void
 192 rte_eal_config_create(void)
 193 {
 194         void *rte_mem_cfg_addr;
 195         int retval;
 196
 197         const char *pathname = eal_runtime_config_path();
 198
 199         if (internal_config.no_shconf)
 200                 return;
 201
 202         if (mem_cfg_fd < 0){
 203                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 204                 if (mem_cfg_fd < 0)
 205                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 206         }
 207
 208         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 209         if (retval < 0){
 210                 close(mem_cfg_fd);
 211                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 212         }
 213
 214         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 215         if (retval < 0){
 216                 close(mem_cfg_fd);
 217                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 218                                 "process running?\n", pathname);
 219         }
 220
 221         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 222                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 223
 224         if (rte_mem_cfg_addr == MAP_FAILED){
 225                 rte_panic("Cannot mmap memory for rte_config\n");
 226         }
 227         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 228         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 229 }
 230
 231 /* attach to an existing shared memory config */
 232 static void
 233 rte_eal_config_attach(void)
 234 {
 235         void *rte_mem_cfg_addr;
 236         const char *pathname = eal_runtime_config_path();
 237
 238         if (internal_config.no_shconf)
 239                 return;
 240
 241         if (mem_cfg_fd < 0){
 242                 mem_cfg_fd = open(pathname, O_RDWR);
 243                 if (mem_cfg_fd < 0)
 244                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 245         }
 246
 247         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 248                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 249         close(mem_cfg_fd);
 250         if (rte_mem_cfg_addr == MAP_FAILED)
 251                 rte_panic("Cannot mmap memory for rte_config\n");
 252
 253         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 254 }
 255
 256 /* Detect if we are a primary or a secondary process */
 257 static enum rte_proc_type_t
 258 eal_proc_type_detect(void)
 259 {
 260         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 261         const char *pathname = eal_runtime_config_path();
 262
 263         /* if we can open the file but not get a write-lock we are a secondary
 264          * process. NOTE: if we get a file handle back, we keep that open
 265          * and don't close it to prevent a race condition between multiple opens */
 266         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 267                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 268                 ptype = RTE_PROC_SECONDARY;
 269
 270         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 271                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 272
 273         return ptype;
 274 }
 275
 276 /* Sets up rte_config structure with the pointer to shared memory config.*/
 277 static void
 278 rte_config_init(void)
 279 {
 280         /* set the magic in configuration structure */
 281         rte_config.magic = RTE_MAGIC;
 282         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 283                         eal_proc_type_detect() : /* for auto, detect the type */
 284                         internal_config.process_type; /* otherwise use what's already set */
 285
 286         switch (rte_config.process_type){
 287         case RTE_PROC_PRIMARY:
 288                 rte_eal_config_create();
 289                 break;
 290         case RTE_PROC_SECONDARY:
 291                 rte_eal_config_attach();
 292                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 293                 break;
 294         case RTE_PROC_AUTO:
 295         case RTE_PROC_INVALID:
 296                 rte_panic("Invalid process type\n");
 297         }
 298 }
 299
 300 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 301 static void
 302 eal_hugedirs_unlock(void)
 303 {
 304         int i;
 305
 306         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 307         {
 308                 /* skip uninitialized */
 309                 if (internal_config.hugepage_info[i].lock_descriptor == 0)
 310                         continue;
 311                 /* unlock hugepage file */
 312                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 313                 close(internal_config.hugepage_info[i].lock_descriptor);
 314                 /* reset the field */
 315                 internal_config.hugepage_info[i].lock_descriptor = 0;
 316         }
 317 }
 318
 319 /* display usage */
 320 static void
 321 eal_usage(const char *prgname)
 322 {
 323         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 324                "[--proc-type primary|secondary|auto] \n\n"
 325                "EAL options:\n"
 326                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 327                "  -n NUM       : Number of memory channels\n"
 328                    "  -v           : Display version information on startup\n"
 329                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 330            "PCI device\n"
 331                "                 (multiple -b options are allowed)\n"
 332                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 333                "  -r NUM       : force number of memory ranks (don't detect)\n"
 334                "  --"OPT_SYSLOG"     : set syslog facility\n"
 335                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 336                    "                 sockets (use comma separated values)\n"
 337                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 338                "  --"OPT_PROC_TYPE"  : type of this process\n"
 339                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 340                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only."
 341                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 342                "               [NOTE: Cannot be used with -b option]\n"
 343                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 344                            "native RDTSC\n"
 345                "\nEAL options for DEBUG use only:\n"
 346                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 347                "  --"OPT_NO_PCI"   : disable pci\n"
 348                "  --"OPT_NO_HPET"  : disable hpet\n"
 349                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 350                "\n",
 351                prgname);
 352         /* Allow the application to print its usage message too if hook is set */
 353         if ( rte_application_usage_hook ) {
 354                 printf("===== Application Usage =====\n\n");
 355                 rte_application_usage_hook(prgname);
 356         }
 357 }
 358
 359 /* Set a per-application usage message */
 360 rte_usage_hook_t
 361 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 362 {
 363         rte_usage_hook_t        old_func;
 364
 365         /* Will be NULL on the first call to denote the last usage routine. */
 366         old_func                                        = rte_application_usage_hook;
 367         rte_application_usage_hook      = usage_func;
 368
 369         return old_func;
 370 }
 371
 372 /*
 373  * Parse the coremask given as argument (hexadecimal string) and fill
 374  * the global configuration (core role and core count) with the parsed
 375  * value.
 376  */
 377 static int xdigit2val(unsigned char c)
 378 {
 379         int val;
 380         if(isdigit(c))
 381                 val = c - '0';
 382         else if(isupper(c))
 383                 val = c - 'A' + 10;
 384         else
 385                 val = c - 'a' + 10;
 386         return val;
 387 }
 388 static int
 389 eal_parse_coremask(const char *coremask)
 390 {
 391         struct rte_config *cfg = rte_eal_get_configuration();
 392         int i, j, idx = 0 ;
 393         unsigned count = 0;
 394         char c;
 395         int val;
 396
 397         if (coremask == NULL)
 398                 return -1;
 399         /* Remove all blank characters ahead and after .
 400          * Remove 0x/0X if exists.
 401          */
 402         while (isblank(*coremask))
 403                 coremask++;
 404         if (coremask[0] == '0' && ((coremask[1] == 'x')
 405                 ||  (coremask[1] == 'X')) )
 406                 coremask += 2;
 407         i = strnlen(coremask, MAX_ARG_STRLEN);
 408         while ((i > 0) && isblank(coremask[i - 1]))
 409                 i--;
 410         if (i == 0)
 411                 return -1;
 412
 413         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 414                 c = coremask[i];
 415                 if (isxdigit(c) == 0) {
 416                         /* invalid characters */
 417                         return (-1);
 418                 }
 419                 val = xdigit2val(c);
 420                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 421                         if((1 << j) & val) {
 422                                 cfg->lcore_role[idx] = ROLE_RTE;
 423                                 if(count == 0)
 424                                         cfg->master_lcore = idx;
 425                                 count++;
 426                         } else  {
 427                                 cfg->lcore_role[idx] = ROLE_OFF;
 428                         }
 429                 }
 430         }
 431         for(; i >= 0; i--)
 432                 if(coremask[i] != '0')
 433                         return -1;
 434         for(; idx < RTE_MAX_LCORE; idx++)
 435                 cfg->lcore_role[idx] = ROLE_OFF;
 436         if(count == 0)
 437                 return -1;
 438         return 0;
 439 }
 440
 441 static int
 442 eal_parse_syslog(const char *facility)
 443 {
 444         int i;
 445         static struct {
 446                 const char *name;
 447                 int value;
 448         } map[] = {
 449                 { "auth", LOG_AUTH },
 450                 { "cron", LOG_CRON },
 451                 { "daemon", LOG_DAEMON },
 452                 { "ftp", LOG_FTP },
 453                 { "kern", LOG_KERN },
 454                 { "lpr", LOG_LPR },
 455                 { "mail", LOG_MAIL },
 456                 { "news", LOG_NEWS },
 457                 { "syslog", LOG_SYSLOG },
 458                 { "user", LOG_USER },
 459                 { "uucp", LOG_UUCP },
 460                 { "local0", LOG_LOCAL0 },
 461                 { "local1", LOG_LOCAL1 },
 462                 { "local2", LOG_LOCAL2 },
 463                 { "local3", LOG_LOCAL3 },
 464                 { "local4", LOG_LOCAL4 },
 465                 { "local5", LOG_LOCAL5 },
 466                 { "local6", LOG_LOCAL6 },
 467                 { "local7", LOG_LOCAL7 },
 468                 { NULL, 0 }
 469         };
 470
 471         for (i = 0; map[i].name; i++) {
 472                 if (!strcmp(facility, map[i].name)) {
 473                         internal_config.syslog_facility = map[i].value;
 474                         return 0;
 475                 }
 476         }
 477         return -1;
 478 }
 479
 480 static int
 481 eal_parse_socket_mem(char *socket_mem)
 482 {
 483         char * arg[RTE_MAX_NUMA_NODES];
 484         char *end;
 485         int arg_num, i, len;
 486         uint64_t total_mem = 0;
 487
 488         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 489         if (len == SOCKET_MEM_STRLEN) {
 490                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 491                 return -1;
 492         }
 493
 494         /* all other error cases will be caught later */
 495         if (!isdigit(socket_mem[len-1]))
 496                 return -1;
 497
 498         /* split the optarg into separate socket values */
 499         arg_num = rte_strsplit(socket_mem, len,
 500                         arg, RTE_MAX_NUMA_NODES, ',');
 501
 502         /* if split failed, or 0 arguments */
 503         if (arg_num <= 0)
 504                 return -1;
 505
 506         internal_config.force_sockets = 1;
 507
 508         /* parse each defined socket option */
 509         errno = 0;
 510         for (i = 0; i < arg_num; i++) {
 511                 end = NULL;
 512                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 513
 514                 /* check for invalid input */
 515                 if ((errno != 0)  ||
 516                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 517                         return -1;
 518                 internal_config.socket_mem[i] *= 1024ULL;
 519                 internal_config.socket_mem[i] *= 1024ULL;
 520                 total_mem += internal_config.socket_mem[i];
 521         }
 522
 523         /* check if we have a positive amount of total memory */
 524         if (total_mem == 0)
 525                 return -1;
 526
 527         return 0;
 528 }
 529
 530 static inline size_t
 531 eal_get_hugepage_mem_size(void)
 532 {
 533         uint64_t size = 0;
 534         unsigned i, j;
 535
 536         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 537                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 538                 if (hpi->hugedir != NULL) {
 539                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 540                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 541                         }
 542                 }
 543         }
 544
 545         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 546 }
 547
 548 static enum rte_proc_type_t
 549 eal_parse_proc_type(const char *arg)
 550 {
 551         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 552                 return RTE_PROC_PRIMARY;
 553         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 554                 return RTE_PROC_SECONDARY;
 555         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 556                 return RTE_PROC_AUTO;
 557
 558         return RTE_PROC_INVALID;
 559 }
 560
 561 static ssize_t
 562 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 563 {
 564         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 565                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 566                 return (-EINVAL);
 567         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 568                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 569                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 570                 return (-EINVAL);
 571         }
 572
 573         idx += 1;
 574         return (idx);
 575 }
 576
 577 /* Parse the argument given in the command line of the application */
 578 static int
 579 eal_parse_args(int argc, char **argv)
 580 {
 581         int opt, ret, i;
 582         char **argvopt;
 583         int option_index;
 584         int coremask_ok = 0;
 585         ssize_t blacklist_index = 0;
 586         char *prgname = argv[0];
 587         static struct option lgopts[] = {
 588                 {OPT_NO_HUGE, 0, 0, 0},
 589                 {OPT_NO_PCI, 0, 0, 0},
 590                 {OPT_NO_HPET, 0, 0, 0},
 591                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 592                 {OPT_HUGE_DIR, 1, 0, 0},
 593                 {OPT_NO_SHCONF, 0, 0, 0},
 594                 {OPT_PROC_TYPE, 1, 0, 0},
 595                 {OPT_FILE_PREFIX, 1, 0, 0},
 596                 {OPT_SOCKET_MEM, 1, 0, 0},
 597                 {OPT_USE_DEVICE, 1, 0, 0},
 598                 {OPT_SYSLOG, 1, NULL, 0},
 599                 {0, 0, 0, 0}
 600         };
 601
 602         argvopt = argv;
 603
 604         internal_config.memory = 0;
 605         internal_config.force_nrank = 0;
 606         internal_config.force_nchannel = 0;
 607         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 608         internal_config.hugepage_dir = NULL;
 609         internal_config.force_sockets = 0;
 610         internal_config.syslog_facility = LOG_DAEMON;
 611 #ifdef RTE_LIBEAL_USE_HPET
 612         internal_config.no_hpet = 0;
 613 #else
 614         internal_config.no_hpet = 1;
 615 #endif
 616         /* zero out the NUMA config */
 617         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 618                 internal_config.socket_mem[i] = 0;
 619
 620         /* zero out hugedir descriptors */
 621         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 622                 internal_config.hugepage_info[i].lock_descriptor = 0;
 623
 624         internal_config.vmware_tsc_map = 0;
 625
 626         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 627                                   lgopts, &option_index)) != EOF) {
 628
 629                 switch (opt) {
 630                 /* blacklist */
 631                 case 'b':
 632                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 633                             blacklist_index)) < 0) {
 634                                 eal_usage(prgname);
 635                                 return (-1);
 636                         }
 637                         break;
 638                 /* coremask */
 639                 case 'c':
 640                         if (eal_parse_coremask(optarg) < 0) {
 641                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 642                                 eal_usage(prgname);
 643                                 return -1;
 644                         }
 645                         coremask_ok = 1;
 646                         break;
 647                 /* size of memory */
 648                 case 'm':
 649                         internal_config.memory = atoi(optarg);
 650                         internal_config.memory *= 1024ULL;
 651                         internal_config.memory *= 1024ULL;
 652                         break;
 653                 /* force number of channels */
 654                 case 'n':
 655                         internal_config.force_nchannel = atoi(optarg);
 656                         if (internal_config.force_nchannel == 0 ||
 657                             internal_config.force_nchannel > 4) {
 658                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 659                                 eal_usage(prgname);
 660                                 return -1;
 661                         }
 662                         break;
 663                 /* force number of ranks */
 664                 case 'r':
 665                         internal_config.force_nrank = atoi(optarg);
 666                         if (internal_config.force_nrank == 0 ||
 667                             internal_config.force_nrank > 16) {
 668                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 669                                 eal_usage(prgname);
 670                                 return -1;
 671                         }
 672                         break;
 673                 case 'v':
 674                         /* since message is explicitly requested by user, we
 675                          * write message at highest log level so it can always be seen
 676                          * even if info or warning messages are disabled */
 677                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 678                         break;
 679
 680                 /* long options */
 681                 case 0:
 682                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 683                                 internal_config.no_hugetlbfs = 1;
 684                         }
 685                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 686                                 internal_config.no_pci = 1;
 687                         }
 688                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 689                                 internal_config.no_hpet = 1;
 690                         }
 691                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 692                                 internal_config.vmware_tsc_map = 1;
 693                         }
 694                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 695                                 internal_config.no_shconf = 1;
 696                         }
 697                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 698                                 internal_config.hugepage_dir = optarg;
 699                         }
 700                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 701                                 internal_config.process_type = eal_parse_proc_type(optarg);
 702                         }
 703                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 704                                 internal_config.hugefile_prefix = optarg;
 705                         }
 706                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 707                                 if (eal_parse_socket_mem(optarg) < 0) {
 708                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 709                                                         OPT_SOCKET_MEM "\n");
 710                                         eal_usage(prgname);
 711                                         return -1;
 712                                 }
 713                         }
 714                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 715                                 eal_dev_whitelist_add_entry(optarg);
 716                         }
 717                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 718                                 if (eal_parse_syslog(optarg) < 0) {
 719                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 720                                                         OPT_SYSLOG "\n");
 721                                         eal_usage(prgname);
 722                                         return -1;
 723                                 }
 724                         }
 725                         break;
 726
 727                 default:
 728                         eal_usage(prgname);
 729                         return -1;
 730                 }
 731         }
 732
 733         /* sanity checks */
 734         if (!coremask_ok) {
 735                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 736                 eal_usage(prgname);
 737                 return -1;
 738         }
 739         if (internal_config.process_type == RTE_PROC_AUTO){
 740                 internal_config.process_type = eal_proc_type_detect();
 741         }
 742         if (internal_config.process_type == RTE_PROC_INVALID){
 743                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 744                 eal_usage(prgname);
 745                 return -1;
 746         }
 747         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 748                         internal_config.force_nchannel == 0) {
 749                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 750                 eal_usage(prgname);
 751                 return -1;
 752         }
 753         if (index(internal_config.hugefile_prefix,'%') != NULL){
 754                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 755                 eal_usage(prgname);
 756                 return -1;
 757         }
 758         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 759                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 760                                 "at the same time\n");
 761                 eal_usage(prgname);
 762                 return -1;
 763         }
 764         /* --no-huge doesn't make sense with either -m or --socket-mem */
 765         if (internal_config.no_hugetlbfs &&
 766                         (internal_config.memory > 0 ||
 767                                         internal_config.force_sockets == 1)) {
 768                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 769                                 "together with --no-huge!\n");
 770                 eal_usage(prgname);
 771                 return -1;
 772         }
 773
 774         /* if no blacklist, parse a whitelist */
 775         if (blacklist_index > 0) {
 776                 if (eal_dev_whitelist_exists()) {
 777                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 778                                         "[--use-device] options cannot be used at the same time\n");
 779                         eal_usage(prgname);
 780                         return -1;
 781                 }
 782                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 783         } else {
 784                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 785                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 786                         return -1;
 787                 }
 788         }
 789
 790         if (optind >= 0)
 791                 argv[optind-1] = prgname;
 792
 793         /* if no memory amounts were requested, this will result in 0 and
 794          * will be overriden later, right after eal_hugepage_info_init() */
 795         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 796                 internal_config.memory += internal_config.socket_mem[i];
 797
 798         ret = optind-1;
 799         optind = 0; /* reset getopt lib */
 800         return ret;
 801 }
 802
 803 static void
 804 eal_check_mem_on_local_socket(void)
 805 {
 806         const struct rte_memseg *ms;
 807         int i, socket_id;
 808
 809         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 810
 811         ms = rte_eal_get_physmem_layout();
 812
 813         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 814                 if (ms[i].socket_id == socket_id &&
 815                                 ms[i].len > 0)
 816                         return;
 817
 818         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 819                         "memory on local socket!\n");
 820 }
 821
 822 static int
 823 sync_func(__attribute__((unused)) void *arg)
 824 {
 825         return 0;
 826 }
 827
 828 inline static void
 829 rte_eal_mcfg_complete(void)
 830 {
 831         /* ALL shared mem_config related INIT DONE */
 832         if (rte_config.process_type == RTE_PROC_PRIMARY)
 833                 rte_config.mem_config->magic = RTE_MAGIC;
 834 }
 835
 836 /* Launch threads, called at application init(). */
 837 int
 838 rte_eal_init(int argc, char **argv)
 839 {
 840         int i, fctret, ret;
 841         pthread_t thread_id;
 842         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 843
 844         if (!rte_atomic32_test_and_set(&run_once))
 845                 return -1;
 846
 847         thread_id = pthread_self();
 848
 849         if (rte_eal_log_early_init() < 0)
 850                 rte_panic("Cannot init early logs\n");
 851
 852         fctret = eal_parse_args(argc, argv);
 853         if (fctret < 0)
 854                 exit(1);
 855
 856         if (internal_config.no_hugetlbfs == 0 &&
 857                         internal_config.process_type != RTE_PROC_SECONDARY &&
 858                         eal_hugepage_info_init() < 0)
 859                 rte_panic("Cannot get hugepage information\n");
 860
 861         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 862                 if (internal_config.no_hugetlbfs)
 863                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 864                 else
 865                         internal_config.memory = eal_get_hugepage_mem_size();
 866         }
 867
 868         if (internal_config.vmware_tsc_map == 1) {
 869 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 870                 rte_cycles_vmware_tsc_map = 1;
 871                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 872                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 873 #else
 874                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 875                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 876 #endif
 877         }
 878
 879         rte_srand(rte_rdtsc());
 880
 881         rte_config_init();
 882
 883         if (rte_eal_cpu_init() < 0)
 884                 rte_panic("Cannot detect lcores\n");
 885
 886         if (rte_eal_memory_init() < 0)
 887                 rte_panic("Cannot init memory\n");
 888
 889         /* the directories are locked during eal_hugepage_info_init */
 890         eal_hugedirs_unlock();
 891
 892         if (rte_eal_memzone_init() < 0)
 893                 rte_panic("Cannot init memzone\n");
 894
 895         if (rte_eal_tailqs_init() < 0)
 896                 rte_panic("Cannot init tail queues for objects\n");
 897
 898         if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 899                 rte_panic("Cannot init logs\n");
 900
 901         if (rte_eal_alarm_init() < 0)
 902                 rte_panic("Cannot init interrupt-handling thread\n");
 903
 904         if (rte_eal_intr_init() < 0)
 905                 rte_panic("Cannot init interrupt-handling thread\n");
 906
 907         if (rte_eal_timer_init() < 0)
 908                 rte_panic("Cannot init HPET or TSC timers\n");
 909
 910         if (rte_eal_pci_init() < 0)
 911                 rte_panic("Cannot init PCI\n");
 912
 913         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
 914                 rte_config.master_lcore, (int)thread_id);
 915
 916         eal_check_mem_on_local_socket();
 917
 918         rte_eal_mcfg_complete();
 919
 920         if (rte_eal_non_pci_ethdev_init() < 0)
 921                 rte_panic("Cannot init non-PCI eth_devs\n");
 922
 923         RTE_LCORE_FOREACH_SLAVE(i) {
 924
 925                 /*
 926                  * create communication pipes between master thread
 927                  * and children
 928                  */
 929                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 930                         rte_panic("Cannot create pipe\n");
 931                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 932                         rte_panic("Cannot create pipe\n");
 933
 934                 lcore_config[i].state = WAIT;
 935
 936                 /* create a thread for each lcore */
 937                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 938                                      eal_thread_loop, NULL);
 939                 if (ret != 0)
 940                         rte_panic("Cannot create thread\n");
 941         }
 942
 943         eal_thread_init_master(rte_config.master_lcore);
 944
 945         /*
 946          * Launch a dummy function on all slave lcores, so that master lcore
 947          * knows they are all ready when this function returns.
 948          */
 949         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 950         rte_eal_mp_wait_lcore();
 951
 952         return fctret;
 953 }
 954
 955 /* get core role */
 956 enum rte_lcore_role_t
 957 rte_eal_lcore_role(unsigned lcore_id)
 958 {
 959         return (rte_config.lcore_role[lcore_id]);
 960 }
 961
 962 enum rte_proc_type_t
 963 rte_eal_process_type(void)
 964 {
 965         return (rte_config.process_type);
 966 }
 967