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