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