lib/librte_eal/bsdapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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 #include <rte_eth_ring.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_USE_DEVICE  "use-device"
  90 #define OPT_SYSLOG      "syslog"
  91
  92 #define RTE_EAL_BLACKLIST_SIZE  0x100
  93
  94 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  95
  96 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
  97
  98 #define HIGHEST_RPL 3
  99
 100 #define BITS_PER_HEX 4
 101
 102 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 103 {                                                               \
 104         unsigned long val;                                      \
 105         char *end;                                              \
 106         errno = 0;                                              \
 107         val = strtoul((in), &end, 16);                          \
 108         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 109                 return (-EINVAL);                               \
 110         (fd) = (typeof (fd))val;                                \
 111         (in) = end + 1;                                         \
 112 }
 113
 114 /* Allow the application to print its usage message too if set */
 115 static rte_usage_hook_t rte_application_usage_hook = NULL;
 116 /* early configuration structure, when memory config is not mmapped */
 117 static struct rte_mem_config early_mem_config;
 118
 119 /* define fd variable here, because file needs to be kept open for the
 120  * duration of the program, as we hold a write lock on it in the primary proc */
 121 static int mem_cfg_fd = -1;
 122
 123 static struct flock wr_lock = {
 124                 .l_type = F_WRLCK,
 125                 .l_whence = SEEK_SET,
 126                 .l_start = offsetof(struct rte_mem_config, memseg),
 127                 .l_len = sizeof(early_mem_config.memseg),
 128 };
 129
 130 /* Address of global and public configuration */
 131 static struct rte_config rte_config = {
 132                 .mem_config = &early_mem_config,
 133 };
 134
 135 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 136
 137 /* internal configuration (per-core) */
 138 struct lcore_config lcore_config[RTE_MAX_LCORE];
 139
 140 /* internal configuration */
 141 struct internal_config internal_config;
 142
 143 /* used by rte_rdtsc() */
 144 int rte_cycles_vmware_tsc_map;
 145
 146 /* Return a pointer to the configuration structure */
 147 struct rte_config *
 148 rte_eal_get_configuration(void)
 149 {
 150         return &rte_config;
 151 }
 152
 153 /* parse a sysfs (or other) file containing one integer value */
 154 int
 155 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 156 {
 157         FILE *f;
 158         char buf[BUFSIZ];
 159         char *end = NULL;
 160
 161         if ((f = fopen(filename, "r")) == NULL) {
 162                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 163                         __func__, filename);
 164                 return -1;
 165         }
 166
 167         if (fgets(buf, sizeof(buf), f) == NULL) {
 168                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 169                         __func__, filename);
 170                 fclose(f);
 171                 return -1;
 172         }
 173         *val = strtoul(buf, &end, 0);
 174         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 175                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 176                                 __func__, filename);
 177                 fclose(f);
 178                 return -1;
 179         }
 180         fclose(f);
 181         return 0;
 182 }
 183
 184
 185 /* create memory configuration in shared/mmap memory. Take out
 186  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 187  * This means we never close the file while running (auto-close on exit).
 188  * We also don't lock the whole file, so that in future we can use read-locks
 189  * on other parts, e.g. memzones, to detect if there are running secondary
 190  * processes. */
 191 static void
 192 rte_eal_config_create(void)
 193 {
 194         void *rte_mem_cfg_addr;
 195         int retval;
 196
 197         const char *pathname = eal_runtime_config_path();
 198
 199         if (internal_config.no_shconf)
 200                 return;
 201
 202         if (mem_cfg_fd < 0){
 203                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 204                 if (mem_cfg_fd < 0)
 205                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 206         }
 207
 208         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 209         if (retval < 0){
 210                 close(mem_cfg_fd);
 211                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 212         }
 213
 214         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 215         if (retval < 0){
 216                 close(mem_cfg_fd);
 217                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 218                                 "process running?\n", pathname);
 219         }
 220
 221         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 222                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 223
 224         if (rte_mem_cfg_addr == MAP_FAILED){
 225                 rte_panic("Cannot mmap memory for rte_config\n");
 226         }
 227         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 228         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 229 }
 230
 231 /* attach to an existing shared memory config */
 232 static void
 233 rte_eal_config_attach(void)
 234 {
 235         void *rte_mem_cfg_addr;
 236         const char *pathname = eal_runtime_config_path();
 237
 238         if (internal_config.no_shconf)
 239                 return;
 240
 241         if (mem_cfg_fd < 0){
 242                 mem_cfg_fd = open(pathname, O_RDWR);
 243                 if (mem_cfg_fd < 0)
 244                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 245         }
 246
 247         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 248                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 249         close(mem_cfg_fd);
 250         if (rte_mem_cfg_addr == MAP_FAILED)
 251                 rte_panic("Cannot mmap memory for rte_config\n");
 252
 253         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 254 }
 255
 256 /* Detect if we are a primary or a secondary process */
 257 static enum rte_proc_type_t
 258 eal_proc_type_detect(void)
 259 {
 260         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 261         const char *pathname = eal_runtime_config_path();
 262
 263         /* if we can open the file but not get a write-lock we are a secondary
 264          * process. NOTE: if we get a file handle back, we keep that open
 265          * and don't close it to prevent a race condition between multiple opens */
 266         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 267                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 268                 ptype = RTE_PROC_SECONDARY;
 269
 270         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 271                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 272
 273         return ptype;
 274 }
 275
 276 /* Sets up rte_config structure with the pointer to shared memory config.*/
 277 static void
 278 rte_config_init(void)
 279 {
 280         /* set the magic in configuration structure */
 281         rte_config.magic = RTE_MAGIC;
 282         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 283                         eal_proc_type_detect() : /* for auto, detect the type */
 284                         internal_config.process_type; /* otherwise use what's already set */
 285
 286         switch (rte_config.process_type){
 287         case RTE_PROC_PRIMARY:
 288                 rte_eal_config_create();
 289                 break;
 290         case RTE_PROC_SECONDARY:
 291                 rte_eal_config_attach();
 292                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 293                 break;
 294         case RTE_PROC_AUTO:
 295         case RTE_PROC_INVALID:
 296                 rte_panic("Invalid process type\n");
 297         }
 298 }
 299
 300 /* display usage */
 301 static void
 302 eal_usage(const char *prgname)
 303 {
 304         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 305                "[--proc-type primary|secondary|auto] \n\n"
 306                "EAL options:\n"
 307                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 308                "  -n NUM       : Number of memory channels\n"
 309                    "  -v           : Display version information on startup\n"
 310                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 311            "PCI device\n"
 312                "                 (multiple -b options are allowed)\n"
 313                "  -m MB        : memory to allocate\n"
 314                "  -r NUM       : force number of memory ranks (don't detect)\n"
 315                "  --"OPT_PROC_TYPE"  : type of this process\n"
 316                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 317                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 318                "               [NOTE: Cannot be used with -b option]\n"
 319                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 320                            "native RDTSC\n"
 321                "\nEAL options for DEBUG use only:\n"
 322                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 323                "  --"OPT_NO_PCI"   : disable pci\n"
 324                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 325                "\n",
 326                prgname);
 327         /* Allow the application to print its usage message too if hook is set */
 328         if ( rte_application_usage_hook ) {
 329                 printf("===== Application Usage =====\n\n");
 330                 rte_application_usage_hook(prgname);
 331         }
 332 }
 333
 334 /* Set a per-application usage message */
 335 rte_usage_hook_t
 336 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 337 {
 338         rte_usage_hook_t        old_func;
 339
 340         /* Will be NULL on the first call to denote the last usage routine. */
 341         old_func                                        = rte_application_usage_hook;
 342         rte_application_usage_hook      = usage_func;
 343
 344         return old_func;
 345 }
 346
 347 /*
 348  * Parse the coremask given as argument (hexadecimal string) and fill
 349  * the global configuration (core role and core count) with the parsed
 350  * value.
 351  */
 352 static int xdigit2val(unsigned char c)
 353 {
 354         int val;
 355         if(isdigit(c))
 356                 val = c - '0';
 357         else if(isupper(c))
 358                 val = c - 'A' + 10;
 359         else
 360                 val = c - 'a' + 10;
 361         return val;
 362 }
 363 static int
 364 eal_parse_coremask(const char *coremask)
 365 {
 366         struct rte_config *cfg = rte_eal_get_configuration();
 367         int i, j, idx = 0 ;
 368         unsigned count = 0;
 369         char c;
 370         int val;
 371
 372         if (coremask == NULL)
 373                 return -1;
 374         /* Remove all blank characters ahead and after .
 375          * Remove 0x/0X if exists.
 376          */
 377         while (isblank(*coremask))
 378                 coremask++;
 379         if (coremask[0] == '0' && ((coremask[1] == 'x')
 380                 ||  (coremask[1] == 'X')) )
 381                 coremask += 2;
 382         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 383         while ((i > 0) && isblank(coremask[i - 1]))
 384                 i--;
 385         if (i == 0)
 386                 return -1;
 387
 388         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 389                 c = coremask[i];
 390                 if (isxdigit(c) == 0) {
 391                         /* invalid characters */
 392                         return (-1);
 393                 }
 394                 val = xdigit2val(c);
 395                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 396                         if((1 << j) & val) {
 397                                 cfg->lcore_role[idx] = ROLE_RTE;
 398                                 if(count == 0)
 399                                         cfg->master_lcore = idx;
 400                                 count++;
 401                         } else  {
 402                                 cfg->lcore_role[idx] = ROLE_OFF;
 403                         }
 404                 }
 405         }
 406         for(; i >= 0; i--)
 407                 if(coremask[i] != '0')
 408                         return -1;
 409         for(; idx < RTE_MAX_LCORE; idx++)
 410                 cfg->lcore_role[idx] = ROLE_OFF;
 411         if(count == 0)
 412                 return -1;
 413         return 0;
 414 }
 415
 416 static int
 417 eal_parse_syslog(const char *facility)
 418 {
 419         int i;
 420         static struct {
 421                 const char *name;
 422                 int value;
 423         } map[] = {
 424                 { "auth", LOG_AUTH },
 425                 { "cron", LOG_CRON },
 426                 { "daemon", LOG_DAEMON },
 427                 { "ftp", LOG_FTP },
 428                 { "kern", LOG_KERN },
 429                 { "lpr", LOG_LPR },
 430                 { "mail", LOG_MAIL },
 431                 { "news", LOG_NEWS },
 432                 { "syslog", LOG_SYSLOG },
 433                 { "user", LOG_USER },
 434                 { "uucp", LOG_UUCP },
 435                 { "local0", LOG_LOCAL0 },
 436                 { "local1", LOG_LOCAL1 },
 437                 { "local2", LOG_LOCAL2 },
 438                 { "local3", LOG_LOCAL3 },
 439                 { "local4", LOG_LOCAL4 },
 440                 { "local5", LOG_LOCAL5 },
 441                 { "local6", LOG_LOCAL6 },
 442                 { "local7", LOG_LOCAL7 },
 443                 { NULL, 0 }
 444         };
 445
 446         for (i = 0; map[i].name; i++) {
 447                 if (!strcmp(facility, map[i].name)) {
 448                         internal_config.syslog_facility = map[i].value;
 449                         return 0;
 450                 }
 451         }
 452         return -1;
 453 }
 454
 455 static inline size_t
 456 eal_get_hugepage_mem_size(void)
 457 {
 458         uint64_t size = 0;
 459         unsigned i, j;
 460
 461         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 462                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 463                 if (hpi->hugedir != NULL) {
 464                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 465                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 466                         }
 467                 }
 468         }
 469
 470         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 471 }
 472
 473 static enum rte_proc_type_t
 474 eal_parse_proc_type(const char *arg)
 475 {
 476         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 477                 return RTE_PROC_PRIMARY;
 478         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 479                 return RTE_PROC_SECONDARY;
 480         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 481                 return RTE_PROC_AUTO;
 482
 483         return RTE_PROC_INVALID;
 484 }
 485
 486 static ssize_t
 487 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 488 {
 489         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 490                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 491                 return (-EINVAL);
 492         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 493                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 494                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 495                 return (-EINVAL);
 496         }
 497
 498         idx += 1;
 499         return (idx);
 500 }
 501
 502 /* Parse the argument given in the command line of the application */
 503 static int
 504 eal_parse_args(int argc, char **argv)
 505 {
 506         int opt, ret, i;
 507         char **argvopt;
 508         int option_index;
 509         int coremask_ok = 0;
 510         ssize_t blacklist_index = 0;
 511         char *prgname = argv[0];
 512         static struct option lgopts[] = {
 513                 {OPT_NO_HUGE, 0, 0, 0},
 514                 {OPT_NO_PCI, 0, 0, 0},
 515                 {OPT_NO_HPET, 0, 0, 0},
 516                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 517                 {OPT_HUGE_DIR, 1, 0, 0},
 518                 {OPT_NO_SHCONF, 0, 0, 0},
 519                 {OPT_PROC_TYPE, 1, 0, 0},
 520                 {OPT_FILE_PREFIX, 1, 0, 0},
 521                 {OPT_SOCKET_MEM, 1, 0, 0},
 522                 {OPT_USE_DEVICE, 1, 0, 0},
 523                 {OPT_SYSLOG, 1, NULL, 0},
 524                 {0, 0, 0, 0}
 525         };
 526
 527         argvopt = argv;
 528
 529         internal_config.memory = 0;
 530         internal_config.force_nrank = 0;
 531         internal_config.force_nchannel = 0;
 532         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 533         internal_config.hugepage_dir = NULL;
 534         internal_config.force_sockets = 0;
 535         internal_config.syslog_facility = LOG_DAEMON;
 536 #ifdef RTE_LIBEAL_USE_HPET
 537         internal_config.no_hpet = 0;
 538 #else
 539         internal_config.no_hpet = 1;
 540 #endif
 541         /* zero out the NUMA config */
 542         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 543                 internal_config.socket_mem[i] = 0;
 544
 545         /* zero out hugedir descriptors */
 546         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 547                 internal_config.hugepage_info[i].lock_descriptor = 0;
 548
 549         internal_config.vmware_tsc_map = 0;
 550
 551         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 552                                   lgopts, &option_index)) != EOF) {
 553
 554                 switch (opt) {
 555                 /* blacklist */
 556                 case 'b':
 557                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 558                             blacklist_index)) < 0) {
 559                                 eal_usage(prgname);
 560                                 return (-1);
 561                         }
 562                         break;
 563                 /* coremask */
 564                 case 'c':
 565                         if (eal_parse_coremask(optarg) < 0) {
 566                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 567                                 eal_usage(prgname);
 568                                 return -1;
 569                         }
 570                         coremask_ok = 1;
 571                         break;
 572                 /* size of memory */
 573                 case 'm':
 574                         internal_config.memory = atoi(optarg);
 575                         internal_config.memory *= 1024ULL;
 576                         internal_config.memory *= 1024ULL;
 577                         break;
 578                 /* force number of channels */
 579                 case 'n':
 580                         internal_config.force_nchannel = atoi(optarg);
 581                         if (internal_config.force_nchannel == 0 ||
 582                             internal_config.force_nchannel > 4) {
 583                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 584                                 eal_usage(prgname);
 585                                 return -1;
 586                         }
 587                         break;
 588                 /* force number of ranks */
 589                 case 'r':
 590                         internal_config.force_nrank = atoi(optarg);
 591                         if (internal_config.force_nrank == 0 ||
 592                             internal_config.force_nrank > 16) {
 593                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 594                                 eal_usage(prgname);
 595                                 return -1;
 596                         }
 597                         break;
 598                 case 'v':
 599                         /* since message is explicitly requested by user, we
 600                          * write message at highest log level so it can always be seen
 601                          * even if info or warning messages are disabled */
 602                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 603                         break;
 604
 605                 /* long options */
 606                 case 0:
 607                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 608                                 internal_config.no_hugetlbfs = 1;
 609                         }
 610                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 611                                 internal_config.no_pci = 1;
 612                         }
 613                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 614                                 internal_config.no_hpet = 1;
 615                         }
 616                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 617                                 internal_config.vmware_tsc_map = 1;
 618                         }
 619                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 620                                 internal_config.no_shconf = 1;
 621                         }
 622                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 623                                 RTE_LOG(ERR, EAL, "Option "OPT_HUGE_DIR" is not supported on"
 624                                                 "FreeBSD\n");
 625                                 return -1;
 626                         }
 627                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 628                                 internal_config.process_type = eal_parse_proc_type(optarg);
 629                         }
 630                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 631                                 RTE_LOG(ERR, EAL, "Option "OPT_FILE_PREFIX" is not supported on"
 632                                                 "FreeBSD\n");
 633                                 return -1;
 634                         }
 635                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 636                                 RTE_LOG(ERR, EAL, "Option "OPT_SOCKET_MEM" is not supported on"
 637                                                 "FreeBSD\n");
 638                                 return -1;
 639                         }
 640                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 641                                 eal_dev_whitelist_add_entry(optarg);
 642                         }
 643                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 644                                 if (eal_parse_syslog(optarg) < 0) {
 645                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 646                                                         OPT_SYSLOG "\n");
 647                                         eal_usage(prgname);
 648                                         return -1;
 649                                 }
 650                         }
 651                         break;
 652
 653                 default:
 654                         eal_usage(prgname);
 655                         return -1;
 656                 }
 657         }
 658
 659         /* sanity checks */
 660         if (!coremask_ok) {
 661                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 662                 eal_usage(prgname);
 663                 return -1;
 664         }
 665         if (internal_config.process_type == RTE_PROC_AUTO){
 666                 internal_config.process_type = eal_proc_type_detect();
 667         }
 668         if (internal_config.process_type == RTE_PROC_INVALID){
 669                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 670                 eal_usage(prgname);
 671                 return -1;
 672         }
 673         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 674                         internal_config.force_nchannel == 0) {
 675                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 676                 eal_usage(prgname);
 677                 return -1;
 678         }
 679         if (index(internal_config.hugefile_prefix,'%') != NULL){
 680                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 681                 eal_usage(prgname);
 682                 return -1;
 683         }
 684         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 685                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 686                                 "at the same time\n");
 687                 eal_usage(prgname);
 688                 return -1;
 689         }
 690         /* --no-huge doesn't make sense with either -m or --socket-mem */
 691         if (internal_config.no_hugetlbfs &&
 692                         (internal_config.memory > 0 ||
 693                                         internal_config.force_sockets == 1)) {
 694                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 695                                 "together with --no-huge!\n");
 696                 eal_usage(prgname);
 697                 return -1;
 698         }
 699
 700         /* if no blacklist, parse a whitelist */
 701         if (blacklist_index > 0) {
 702                 if (eal_dev_whitelist_exists()) {
 703                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 704                                         "[--use-device] options cannot be used at the same time\n");
 705                         eal_usage(prgname);
 706                         return -1;
 707                 }
 708                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 709         } else {
 710                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 711                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 712                         return -1;
 713                 }
 714         }
 715
 716         if (optind >= 0)
 717                 argv[optind-1] = prgname;
 718
 719         /* if no memory amounts were requested, this will result in 0 and
 720          * will be overriden later, right after eal_hugepage_info_init() */
 721         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 722                 internal_config.memory += internal_config.socket_mem[i];
 723
 724         ret = optind-1;
 725         optind = 0; /* reset getopt lib */
 726         return ret;
 727 }
 728
 729 static void
 730 eal_check_mem_on_local_socket(void)
 731 {
 732         const struct rte_memseg *ms;
 733         int i, socket_id;
 734
 735         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 736
 737         ms = rte_eal_get_physmem_layout();
 738
 739         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 740                 if (ms[i].socket_id == socket_id &&
 741                                 ms[i].len > 0)
 742                         return;
 743
 744         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 745                         "memory on local socket!\n");
 746 }
 747
 748 static int
 749 sync_func(__attribute__((unused)) void *arg)
 750 {
 751         return 0;
 752 }
 753
 754 inline static void
 755 rte_eal_mcfg_complete(void)
 756 {
 757         /* ALL shared mem_config related INIT DONE */
 758         if (rte_config.process_type == RTE_PROC_PRIMARY)
 759                 rte_config.mem_config->magic = RTE_MAGIC;
 760 }
 761
 762 /* Abstraction for port I/0 privilage */
 763 static int
 764 rte_eal_iopl_init(void)
 765 {
 766         int fd = -1;
 767         fd = open("/dev/io", O_RDWR);
 768         if (fd < 0)
 769                 return -1;
 770         return 0;
 771 }
 772
 773 /* Launch threads, called at application init(). */
 774 int
 775 rte_eal_init(int argc, char **argv)
 776 {
 777         int i, fctret, ret;
 778         pthread_t thread_id;
 779         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 780
 781         if (!rte_atomic32_test_and_set(&run_once))
 782                 return -1;
 783
 784         thread_id = pthread_self();
 785
 786         if (rte_eal_log_early_init() < 0)
 787                 rte_panic("Cannot init early logs\n");
 788
 789         fctret = eal_parse_args(argc, argv);
 790         if (fctret < 0)
 791                 exit(1);
 792
 793         if (internal_config.no_hugetlbfs == 0 &&
 794                         internal_config.process_type != RTE_PROC_SECONDARY &&
 795                         eal_hugepage_info_init() < 0)
 796                 rte_panic("Cannot get hugepage information\n");
 797
 798         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 799                 if (internal_config.no_hugetlbfs)
 800                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 801                 else
 802                         internal_config.memory = eal_get_hugepage_mem_size();
 803         }
 804
 805         if (internal_config.vmware_tsc_map == 1) {
 806 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 807                 rte_cycles_vmware_tsc_map = 1;
 808                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 809                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 810 #else
 811                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 812                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 813 #endif
 814         }
 815
 816         rte_srand(rte_rdtsc());
 817
 818         rte_config_init();
 819
 820         if (rte_eal_iopl_init() == 0)
 821                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 822
 823         if (rte_eal_cpu_init() < 0)
 824                 rte_panic("Cannot detect lcores\n");
 825
 826         if (rte_eal_memory_init() < 0)
 827                 rte_panic("Cannot init memory\n");
 828
 829         if (rte_eal_memzone_init() < 0)
 830                 rte_panic("Cannot init memzone\n");
 831
 832         if (rte_eal_tailqs_init() < 0)
 833                 rte_panic("Cannot init tail queues for objects\n");
 834
 835 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 836                 rte_panic("Cannot init logs\n");*/
 837
 838         if (rte_eal_alarm_init() < 0)
 839                 rte_panic("Cannot init interrupt-handling thread\n");
 840
 841         if (rte_eal_intr_init() < 0)
 842                 rte_panic("Cannot init interrupt-handling thread\n");
 843
 844         if (rte_eal_timer_init() < 0)
 845                 rte_panic("Cannot init HPET or TSC timers\n");
 846
 847         if (rte_eal_pci_init() < 0)
 848                 rte_panic("Cannot init PCI\n");
 849
 850         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 851                 rte_config.master_lcore, thread_id);
 852
 853         eal_check_mem_on_local_socket();
 854
 855         rte_eal_mcfg_complete();
 856
 857         if (rte_eal_non_pci_ethdev_init() < 0)
 858                 rte_panic("Cannot init non-PCI eth_devs\n");
 859
 860         RTE_LCORE_FOREACH_SLAVE(i) {
 861
 862                 /*
 863                  * create communication pipes between master thread
 864                  * and children
 865                  */
 866                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 867                         rte_panic("Cannot create pipe\n");
 868                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 869                         rte_panic("Cannot create pipe\n");
 870
 871                 lcore_config[i].state = WAIT;
 872
 873                 /* create a thread for each lcore */
 874                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 875                                      eal_thread_loop, NULL);
 876                 if (ret != 0)
 877                         rte_panic("Cannot create thread\n");
 878         }
 879
 880         eal_thread_init_master(rte_config.master_lcore);
 881
 882         /*
 883          * Launch a dummy function on all slave lcores, so that master lcore
 884          * knows they are all ready when this function returns.
 885          */
 886         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 887         rte_eal_mp_wait_lcore();
 888
 889         return fctret;
 890 }
 891
 892 /* get core role */
 893 enum rte_lcore_role_t
 894 rte_eal_lcore_role(unsigned lcore_id)
 895 {
 896         return (rte_config.lcore_role[lcore_id]);
 897 }
 898
 899 enum rte_proc_type_t
 900 rte_eal_process_type(void)
 901 {
 902         return (rte_config.process_type);
 903 }
 904