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