lib/librte_eal/linuxapp/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 #include <sys/io.h>
  51 #include <sys/user.h>
  52 #include <linux/binfmts.h>
  53
  54 #include <rte_common.h>
  55 #include <rte_debug.h>
  56 #include <rte_memory.h>
  57 #include <rte_memzone.h>
  58 #include <rte_launch.h>
  59 #include <rte_tailq.h>
  60 #include <rte_eal.h>
  61 #include <rte_eal_memconfig.h>
  62 #include <rte_per_lcore.h>
  63 #include <rte_lcore.h>
  64 #include <rte_log.h>
  65 #include <rte_random.h>
  66 #include <rte_cycles.h>
  67 #include <rte_string_fns.h>
  68 #include <rte_cpuflags.h>
  69 #include <rte_interrupts.h>
  70 #include <rte_pci.h>
  71 #include <rte_common.h>
  72 #include <rte_version.h>
  73 #include <rte_atomic.h>
  74 #include <malloc_heap.h>
  75 #include <rte_eth_ring.h>
  76
  77 #include "eal_private.h"
  78 #include "eal_thread.h"
  79 #include "eal_internal_cfg.h"
  80 #include "eal_filesystem.h"
  81 #include "eal_hugepages.h"
  82
  83 #define OPT_HUGE_DIR    "huge-dir"
  84 #define OPT_PROC_TYPE   "proc-type"
  85 #define OPT_NO_SHCONF   "no-shconf"
  86 #define OPT_NO_HPET     "no-hpet"
  87 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  88 #define OPT_NO_PCI      "no-pci"
  89 #define OPT_NO_HUGE     "no-huge"
  90 #define OPT_FILE_PREFIX "file-prefix"
  91 #define OPT_SOCKET_MEM  "socket-mem"
  92 #define OPT_USE_DEVICE  "use-device"
  93 #define OPT_SYSLOG      "syslog"
  94 #define OPT_BASE_VIRTADDR   "base-virtaddr"
  95
  96 #define RTE_EAL_BLACKLIST_SIZE  0x100
  97
  98 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  99
 100 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 101
 102 #define HIGHEST_RPL 3
 103
 104 #define BITS_PER_HEX 4
 105
 106 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 107 {                                                               \
 108         unsigned long val;                                      \
 109         char *end;                                              \
 110         errno = 0;                                              \
 111         val = strtoul((in), &end, 16);                          \
 112         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 113                 return (-EINVAL);                               \
 114         (fd) = (typeof (fd))val;                                \
 115         (in) = end + 1;                                         \
 116 }
 117
 118 /* Allow the application to print its usage message too if set */
 119 static rte_usage_hook_t rte_application_usage_hook = NULL;
 120 /* early configuration structure, when memory config is not mmapped */
 121 static struct rte_mem_config early_mem_config;
 122
 123 /* define fd variable here, because file needs to be kept open for the
 124  * duration of the program, as we hold a write lock on it in the primary proc */
 125 static int mem_cfg_fd = -1;
 126
 127 static struct flock wr_lock = {
 128                 .l_type = F_WRLCK,
 129                 .l_whence = SEEK_SET,
 130                 .l_start = offsetof(struct rte_mem_config, memseg),
 131                 .l_len = sizeof(early_mem_config.memseg),
 132 };
 133
 134 /* Address of global and public configuration */
 135 static struct rte_config rte_config = {
 136                 .mem_config = &early_mem_config,
 137 };
 138
 139 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 140
 141 /* internal configuration (per-core) */
 142 struct lcore_config lcore_config[RTE_MAX_LCORE];
 143
 144 /* internal configuration */
 145 struct internal_config internal_config;
 146
 147 /* used by rte_rdtsc() */
 148 int rte_cycles_vmware_tsc_map;
 149
 150 /* Return a pointer to the configuration structure */
 151 struct rte_config *
 152 rte_eal_get_configuration(void)
 153 {
 154         return &rte_config;
 155 }
 156
 157 /* parse a sysfs (or other) file containing one integer value */
 158 int
 159 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 160 {
 161         FILE *f;
 162         char buf[BUFSIZ];
 163         char *end = NULL;
 164
 165         if ((f = fopen(filename, "r")) == NULL) {
 166                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 167                         __func__, filename);
 168                 return -1;
 169         }
 170
 171         if (fgets(buf, sizeof(buf), f) == NULL) {
 172                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 173                         __func__, filename);
 174                 fclose(f);
 175                 return -1;
 176         }
 177         *val = strtoul(buf, &end, 0);
 178         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 179                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 180                                 __func__, filename);
 181                 fclose(f);
 182                 return -1;
 183         }
 184         fclose(f);
 185         return 0;
 186 }
 187
 188
 189 /* create memory configuration in shared/mmap memory. Take out
 190  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 191  * This means we never close the file while running (auto-close on exit).
 192  * We also don't lock the whole file, so that in future we can use read-locks
 193  * on other parts, e.g. memzones, to detect if there are running secondary
 194  * processes. */
 195 static void
 196 rte_eal_config_create(void)
 197 {
 198         void *rte_mem_cfg_addr;
 199         int retval;
 200
 201         const char *pathname = eal_runtime_config_path();
 202
 203         if (internal_config.no_shconf)
 204                 return;
 205
 206         if (mem_cfg_fd < 0){
 207                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 208                 if (mem_cfg_fd < 0)
 209                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 210         }
 211
 212         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 213         if (retval < 0){
 214                 close(mem_cfg_fd);
 215                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 216         }
 217
 218         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 219         if (retval < 0){
 220                 close(mem_cfg_fd);
 221                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 222                                 "process running?\n", pathname);
 223         }
 224
 225         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 226                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 227
 228         if (rte_mem_cfg_addr == MAP_FAILED){
 229                 rte_panic("Cannot mmap memory for rte_config\n");
 230         }
 231         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 232         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 233 }
 234
 235 /* attach to an existing shared memory config */
 236 static void
 237 rte_eal_config_attach(void)
 238 {
 239         void *rte_mem_cfg_addr;
 240         const char *pathname = eal_runtime_config_path();
 241
 242         if (internal_config.no_shconf)
 243                 return;
 244
 245         if (mem_cfg_fd < 0){
 246                 mem_cfg_fd = open(pathname, O_RDWR);
 247                 if (mem_cfg_fd < 0)
 248                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 249         }
 250
 251         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 252                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 253         close(mem_cfg_fd);
 254         if (rte_mem_cfg_addr == MAP_FAILED)
 255                 rte_panic("Cannot mmap memory for rte_config\n");
 256
 257         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 258 }
 259
 260 /* Detect if we are a primary or a secondary process */
 261 static enum rte_proc_type_t
 262 eal_proc_type_detect(void)
 263 {
 264         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 265         const char *pathname = eal_runtime_config_path();
 266
 267         /* if we can open the file but not get a write-lock we are a secondary
 268          * process. NOTE: if we get a file handle back, we keep that open
 269          * and don't close it to prevent a race condition between multiple opens */
 270         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 271                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 272                 ptype = RTE_PROC_SECONDARY;
 273
 274         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 275                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 276
 277         return ptype;
 278 }
 279
 280 /* Sets up rte_config structure with the pointer to shared memory config.*/
 281 static void
 282 rte_config_init(void)
 283 {
 284         /* set the magic in configuration structure */
 285         rte_config.magic = RTE_MAGIC;
 286         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 287                         eal_proc_type_detect() : /* for auto, detect the type */
 288                         internal_config.process_type; /* otherwise use what's already set */
 289
 290         switch (rte_config.process_type){
 291         case RTE_PROC_PRIMARY:
 292                 rte_eal_config_create();
 293                 break;
 294         case RTE_PROC_SECONDARY:
 295                 rte_eal_config_attach();
 296                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 297                 break;
 298         case RTE_PROC_AUTO:
 299         case RTE_PROC_INVALID:
 300                 rte_panic("Invalid process type\n");
 301         }
 302 }
 303
 304 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 305 static void
 306 eal_hugedirs_unlock(void)
 307 {
 308         int i;
 309
 310         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 311         {
 312                 /* skip uninitialized */
 313                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 314                         continue;
 315                 /* unlock hugepage file */
 316                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 317                 close(internal_config.hugepage_info[i].lock_descriptor);
 318                 /* reset the field */
 319                 internal_config.hugepage_info[i].lock_descriptor = -1;
 320         }
 321 }
 322
 323 /* display usage */
 324 static void
 325 eal_usage(const char *prgname)
 326 {
 327         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 328                "[--proc-type primary|secondary|auto] \n\n"
 329                "EAL options:\n"
 330                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 331                "  -n NUM       : Number of memory channels\n"
 332                    "  -v           : Display version information on startup\n"
 333                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 334            "PCI device\n"
 335                "                 (multiple -b options are allowed)\n"
 336                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 337                "  -r NUM       : force number of memory ranks (don't detect)\n"
 338                "  --"OPT_SYSLOG"     : set syslog facility\n"
 339                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 340                    "                 sockets (use comma separated values)\n"
 341                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 342                "  --"OPT_PROC_TYPE"  : type of this process\n"
 343                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 344                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 345                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 346                "               [NOTE: Cannot be used with -b option]\n"
 347                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 348                            "native RDTSC\n"
 349                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 350                "\nEAL options for DEBUG use only:\n"
 351                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 352                "  --"OPT_NO_PCI"   : disable pci\n"
 353                "  --"OPT_NO_HPET"  : disable hpet\n"
 354                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 355                "\n",
 356                prgname);
 357         /* Allow the application to print its usage message too if hook is set */
 358         if ( rte_application_usage_hook ) {
 359                 printf("===== Application Usage =====\n\n");
 360                 rte_application_usage_hook(prgname);
 361         }
 362 }
 363
 364 /* Set a per-application usage message */
 365 rte_usage_hook_t
 366 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 367 {
 368         rte_usage_hook_t        old_func;
 369
 370         /* Will be NULL on the first call to denote the last usage routine. */
 371         old_func                                        = rte_application_usage_hook;
 372         rte_application_usage_hook      = usage_func;
 373
 374         return old_func;
 375 }
 376
 377 /*
 378  * Parse the coremask given as argument (hexadecimal string) and fill
 379  * the global configuration (core role and core count) with the parsed
 380  * value.
 381  */
 382 static int xdigit2val(unsigned char c)
 383 {
 384         int val;
 385         if(isdigit(c))
 386                 val = c - '0';
 387         else if(isupper(c))
 388                 val = c - 'A' + 10;
 389         else
 390                 val = c - 'a' + 10;
 391         return val;
 392 }
 393 static int
 394 eal_parse_coremask(const char *coremask)
 395 {
 396         struct rte_config *cfg = rte_eal_get_configuration();
 397         int i, j, idx = 0 ;
 398         unsigned count = 0;
 399         char c;
 400         int val;
 401
 402         if (coremask == NULL)
 403                 return -1;
 404         /* Remove all blank characters ahead and after .
 405          * Remove 0x/0X if exists.
 406          */
 407         while (isblank(*coremask))
 408                 coremask++;
 409         if (coremask[0] == '0' && ((coremask[1] == 'x')
 410                 ||  (coremask[1] == 'X')) )
 411                 coremask += 2;
 412         i = strnlen(coremask, MAX_ARG_STRLEN);
 413         while ((i > 0) && isblank(coremask[i - 1]))
 414                 i--;
 415         if (i == 0)
 416                 return -1;
 417
 418         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 419                 c = coremask[i];
 420                 if (isxdigit(c) == 0) {
 421                         /* invalid characters */
 422                         return (-1);
 423                 }
 424                 val = xdigit2val(c);
 425                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 426                         if((1 << j) & val) {
 427                                 cfg->lcore_role[idx] = ROLE_RTE;
 428                                 if(count == 0)
 429                                         cfg->master_lcore = idx;
 430                                 count++;
 431                         } else  {
 432                                 cfg->lcore_role[idx] = ROLE_OFF;
 433                         }
 434                 }
 435         }
 436         for(; i >= 0; i--)
 437                 if(coremask[i] != '0')
 438                         return -1;
 439         for(; idx < RTE_MAX_LCORE; idx++)
 440                 cfg->lcore_role[idx] = ROLE_OFF;
 441         if(count == 0)
 442                 return -1;
 443         return 0;
 444 }
 445
 446 static int
 447 eal_parse_syslog(const char *facility)
 448 {
 449         int i;
 450         static struct {
 451                 const char *name;
 452                 int value;
 453         } map[] = {
 454                 { "auth", LOG_AUTH },
 455                 { "cron", LOG_CRON },
 456                 { "daemon", LOG_DAEMON },
 457                 { "ftp", LOG_FTP },
 458                 { "kern", LOG_KERN },
 459                 { "lpr", LOG_LPR },
 460                 { "mail", LOG_MAIL },
 461                 { "news", LOG_NEWS },
 462                 { "syslog", LOG_SYSLOG },
 463                 { "user", LOG_USER },
 464                 { "uucp", LOG_UUCP },
 465                 { "local0", LOG_LOCAL0 },
 466                 { "local1", LOG_LOCAL1 },
 467                 { "local2", LOG_LOCAL2 },
 468                 { "local3", LOG_LOCAL3 },
 469                 { "local4", LOG_LOCAL4 },
 470                 { "local5", LOG_LOCAL5 },
 471                 { "local6", LOG_LOCAL6 },
 472                 { "local7", LOG_LOCAL7 },
 473                 { NULL, 0 }
 474         };
 475
 476         for (i = 0; map[i].name; i++) {
 477                 if (!strcmp(facility, map[i].name)) {
 478                         internal_config.syslog_facility = map[i].value;
 479                         return 0;
 480                 }
 481         }
 482         return -1;
 483 }
 484
 485 static int
 486 eal_parse_socket_mem(char *socket_mem)
 487 {
 488         char * arg[RTE_MAX_NUMA_NODES];
 489         char *end;
 490         int arg_num, i, len;
 491         uint64_t total_mem = 0;
 492
 493         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 494         if (len == SOCKET_MEM_STRLEN) {
 495                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 496                 return -1;
 497         }
 498
 499         /* all other error cases will be caught later */
 500         if (!isdigit(socket_mem[len-1]))
 501                 return -1;
 502
 503         /* split the optarg into separate socket values */
 504         arg_num = rte_strsplit(socket_mem, len,
 505                         arg, RTE_MAX_NUMA_NODES, ',');
 506
 507         /* if split failed, or 0 arguments */
 508         if (arg_num <= 0)
 509                 return -1;
 510
 511         internal_config.force_sockets = 1;
 512
 513         /* parse each defined socket option */
 514         errno = 0;
 515         for (i = 0; i < arg_num; i++) {
 516                 end = NULL;
 517                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 518
 519                 /* check for invalid input */
 520                 if ((errno != 0)  ||
 521                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 522                         return -1;
 523                 internal_config.socket_mem[i] *= 1024ULL;
 524                 internal_config.socket_mem[i] *= 1024ULL;
 525                 total_mem += internal_config.socket_mem[i];
 526         }
 527
 528         /* check if we have a positive amount of total memory */
 529         if (total_mem == 0)
 530                 return -1;
 531
 532         return 0;
 533 }
 534
 535 static int
 536 eal_parse_base_virtaddr(const char *arg)
 537 {
 538         char *end;
 539         uint64_t addr;
 540
 541         addr = strtoull(arg, &end, 16);
 542
 543         /* check for errors */
 544         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 545                 return -1;
 546
 547         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 548 #ifndef RTE_ARCH_X86_64
 549         if (addr >= UINTPTR_MAX)
 550                 return -1;
 551 #endif
 552
 553         /* align the addr on 2M boundary */
 554         addr = RTE_PTR_ALIGN_CEIL(addr, RTE_PGSIZE_2M);
 555
 556         internal_config.base_virtaddr = (uintptr_t) addr;
 557         return 0;
 558 }
 559
 560 static inline size_t
 561 eal_get_hugepage_mem_size(void)
 562 {
 563         uint64_t size = 0;
 564         unsigned i, j;
 565
 566         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 567                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 568                 if (hpi->hugedir != NULL) {
 569                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 570                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 571                         }
 572                 }
 573         }
 574
 575         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 576 }
 577
 578 static enum rte_proc_type_t
 579 eal_parse_proc_type(const char *arg)
 580 {
 581         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 582                 return RTE_PROC_PRIMARY;
 583         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 584                 return RTE_PROC_SECONDARY;
 585         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 586                 return RTE_PROC_AUTO;
 587
 588         return RTE_PROC_INVALID;
 589 }
 590
 591 static ssize_t
 592 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 593 {
 594         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 595                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 596                 return (-EINVAL);
 597         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 598                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 599                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 600                 return (-EINVAL);
 601         }
 602
 603         idx += 1;
 604         return (idx);
 605 }
 606
 607 /* Parse the argument given in the command line of the application */
 608 static int
 609 eal_parse_args(int argc, char **argv)
 610 {
 611         int opt, ret, i;
 612         char **argvopt;
 613         int option_index;
 614         int coremask_ok = 0;
 615         ssize_t blacklist_index = 0;
 616         char *prgname = argv[0];
 617         static struct option lgopts[] = {
 618                 {OPT_NO_HUGE, 0, 0, 0},
 619                 {OPT_NO_PCI, 0, 0, 0},
 620                 {OPT_NO_HPET, 0, 0, 0},
 621                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 622                 {OPT_HUGE_DIR, 1, 0, 0},
 623                 {OPT_NO_SHCONF, 0, 0, 0},
 624                 {OPT_PROC_TYPE, 1, 0, 0},
 625                 {OPT_FILE_PREFIX, 1, 0, 0},
 626                 {OPT_SOCKET_MEM, 1, 0, 0},
 627                 {OPT_USE_DEVICE, 1, 0, 0},
 628                 {OPT_SYSLOG, 1, NULL, 0},
 629                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 630                 {0, 0, 0, 0}
 631         };
 632
 633         argvopt = argv;
 634
 635         internal_config.memory = 0;
 636         internal_config.force_nrank = 0;
 637         internal_config.force_nchannel = 0;
 638         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 639         internal_config.hugepage_dir = NULL;
 640         internal_config.force_sockets = 0;
 641         internal_config.syslog_facility = LOG_DAEMON;
 642 #ifdef RTE_LIBEAL_USE_HPET
 643         internal_config.no_hpet = 0;
 644 #else
 645         internal_config.no_hpet = 1;
 646 #endif
 647         /* zero out the NUMA config */
 648         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 649                 internal_config.socket_mem[i] = 0;
 650
 651         /* zero out hugedir descriptors */
 652         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 653                 internal_config.hugepage_info[i].lock_descriptor = -1;
 654
 655         internal_config.vmware_tsc_map = 0;
 656         internal_config.base_virtaddr = 0;
 657
 658         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 659                                   lgopts, &option_index)) != EOF) {
 660
 661                 switch (opt) {
 662                 /* blacklist */
 663                 case 'b':
 664                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 665                             blacklist_index)) < 0) {
 666                                 eal_usage(prgname);
 667                                 return (-1);
 668                         }
 669                         break;
 670                 /* coremask */
 671                 case 'c':
 672                         if (eal_parse_coremask(optarg) < 0) {
 673                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 674                                 eal_usage(prgname);
 675                                 return -1;
 676                         }
 677                         coremask_ok = 1;
 678                         break;
 679                 /* size of memory */
 680                 case 'm':
 681                         internal_config.memory = atoi(optarg);
 682                         internal_config.memory *= 1024ULL;
 683                         internal_config.memory *= 1024ULL;
 684                         break;
 685                 /* force number of channels */
 686                 case 'n':
 687                         internal_config.force_nchannel = atoi(optarg);
 688                         if (internal_config.force_nchannel == 0 ||
 689                             internal_config.force_nchannel > 4) {
 690                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 691                                 eal_usage(prgname);
 692                                 return -1;
 693                         }
 694                         break;
 695                 /* force number of ranks */
 696                 case 'r':
 697                         internal_config.force_nrank = atoi(optarg);
 698                         if (internal_config.force_nrank == 0 ||
 699                             internal_config.force_nrank > 16) {
 700                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 701                                 eal_usage(prgname);
 702                                 return -1;
 703                         }
 704                         break;
 705                 case 'v':
 706                         /* since message is explicitly requested by user, we
 707                          * write message at highest log level so it can always be seen
 708                          * even if info or warning messages are disabled */
 709                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 710                         break;
 711
 712                 /* long options */
 713                 case 0:
 714                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 715                                 internal_config.no_hugetlbfs = 1;
 716                         }
 717                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 718                                 internal_config.no_pci = 1;
 719                         }
 720                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 721                                 internal_config.no_hpet = 1;
 722                         }
 723                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 724                                 internal_config.vmware_tsc_map = 1;
 725                         }
 726                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 727                                 internal_config.no_shconf = 1;
 728                         }
 729                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 730                                 internal_config.hugepage_dir = optarg;
 731                         }
 732                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 733                                 internal_config.process_type = eal_parse_proc_type(optarg);
 734                         }
 735                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 736                                 internal_config.hugefile_prefix = optarg;
 737                         }
 738                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 739                                 if (eal_parse_socket_mem(optarg) < 0) {
 740                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 741                                                         OPT_SOCKET_MEM "\n");
 742                                         eal_usage(prgname);
 743                                         return -1;
 744                                 }
 745                         }
 746                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 747                                 eal_dev_whitelist_add_entry(optarg);
 748                         }
 749                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 750                                 if (eal_parse_syslog(optarg) < 0) {
 751                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 752                                                         OPT_SYSLOG "\n");
 753                                         eal_usage(prgname);
 754                                         return -1;
 755                                 }
 756                         }
 757                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 758                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 759                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 760                                                         OPT_BASE_VIRTADDR "\n");
 761                                         eal_usage(prgname);
 762                                         return -1;
 763                                 }
 764                         }
 765                         break;
 766
 767                 default:
 768                         eal_usage(prgname);
 769                         return -1;
 770                 }
 771         }
 772
 773         /* sanity checks */
 774         if (!coremask_ok) {
 775                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 776                 eal_usage(prgname);
 777                 return -1;
 778         }
 779         if (internal_config.process_type == RTE_PROC_AUTO){
 780                 internal_config.process_type = eal_proc_type_detect();
 781         }
 782         if (internal_config.process_type == RTE_PROC_INVALID){
 783                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 784                 eal_usage(prgname);
 785                 return -1;
 786         }
 787         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 788                         internal_config.force_nchannel == 0) {
 789                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 790                 eal_usage(prgname);
 791                 return -1;
 792         }
 793         if (index(internal_config.hugefile_prefix,'%') != NULL){
 794                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 795                 eal_usage(prgname);
 796                 return -1;
 797         }
 798         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 799                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 800                                 "at the same time\n");
 801                 eal_usage(prgname);
 802                 return -1;
 803         }
 804         /* --no-huge doesn't make sense with either -m or --socket-mem */
 805         if (internal_config.no_hugetlbfs &&
 806                         (internal_config.memory > 0 ||
 807                                         internal_config.force_sockets == 1)) {
 808                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 809                                 "together with --no-huge!\n");
 810                 eal_usage(prgname);
 811                 return -1;
 812         }
 813
 814         /* if no blacklist, parse a whitelist */
 815         if (blacklist_index > 0) {
 816                 if (eal_dev_whitelist_exists()) {
 817                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 818                                         "[--use-device] options cannot be used at the same time\n");
 819                         eal_usage(prgname);
 820                         return -1;
 821                 }
 822                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 823         } else {
 824                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 825                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 826                         return -1;
 827                 }
 828         }
 829
 830         if (optind >= 0)
 831                 argv[optind-1] = prgname;
 832
 833         /* if no memory amounts were requested, this will result in 0 and
 834          * will be overriden later, right after eal_hugepage_info_init() */
 835         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 836                 internal_config.memory += internal_config.socket_mem[i];
 837
 838         ret = optind-1;
 839         optind = 0; /* reset getopt lib */
 840         return ret;
 841 }
 842
 843 static void
 844 eal_check_mem_on_local_socket(void)
 845 {
 846         const struct rte_memseg *ms;
 847         int i, socket_id;
 848
 849         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 850
 851         ms = rte_eal_get_physmem_layout();
 852
 853         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 854                 if (ms[i].socket_id == socket_id &&
 855                                 ms[i].len > 0)
 856                         return;
 857
 858         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 859                         "memory on local socket!\n");
 860 }
 861
 862 static int
 863 sync_func(__attribute__((unused)) void *arg)
 864 {
 865         return 0;
 866 }
 867
 868 inline static void
 869 rte_eal_mcfg_complete(void)
 870 {
 871         /* ALL shared mem_config related INIT DONE */
 872         if (rte_config.process_type == RTE_PROC_PRIMARY)
 873                 rte_config.mem_config->magic = RTE_MAGIC;
 874 }
 875
 876 /*
 877  * Request iopl priviledge for all RPL, returns 0 on success
 878  */
 879 static int
 880 rte_eal_iopl_init(void)
 881 {
 882         return iopl(HIGHEST_RPL);
 883 }
 884
 885 /* Launch threads, called at application init(). */
 886 int
 887 rte_eal_init(int argc, char **argv)
 888 {
 889         int i, fctret, ret;
 890         pthread_t thread_id;
 891         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 892
 893         if (!rte_atomic32_test_and_set(&run_once))
 894                 return -1;
 895
 896         thread_id = pthread_self();
 897
 898         if (rte_eal_log_early_init() < 0)
 899                 rte_panic("Cannot init early logs\n");
 900
 901         fctret = eal_parse_args(argc, argv);
 902         if (fctret < 0)
 903                 exit(1);
 904
 905         if (internal_config.no_hugetlbfs == 0 &&
 906                         internal_config.process_type != RTE_PROC_SECONDARY &&
 907                         eal_hugepage_info_init() < 0)
 908                 rte_panic("Cannot get hugepage information\n");
 909
 910         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 911                 if (internal_config.no_hugetlbfs)
 912                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 913                 else
 914                         internal_config.memory = eal_get_hugepage_mem_size();
 915         }
 916
 917         if (internal_config.vmware_tsc_map == 1) {
 918 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 919                 rte_cycles_vmware_tsc_map = 1;
 920                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 921                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 922 #else
 923                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 924                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 925 #endif
 926         }
 927
 928         rte_srand(rte_rdtsc());
 929
 930         rte_config_init();
 931
 932         if (rte_eal_iopl_init() == 0)
 933                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 934
 935         if (rte_eal_cpu_init() < 0)
 936                 rte_panic("Cannot detect lcores\n");
 937
 938         if (rte_eal_memory_init() < 0)
 939                 rte_panic("Cannot init memory\n");
 940
 941         /* the directories are locked during eal_hugepage_info_init */
 942         eal_hugedirs_unlock();
 943
 944         if (rte_eal_memzone_init() < 0)
 945                 rte_panic("Cannot init memzone\n");
 946
 947         if (rte_eal_tailqs_init() < 0)
 948                 rte_panic("Cannot init tail queues for objects\n");
 949
 950         if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 951                 rte_panic("Cannot init logs\n");
 952
 953         if (rte_eal_alarm_init() < 0)
 954                 rte_panic("Cannot init interrupt-handling thread\n");
 955
 956         if (rte_eal_intr_init() < 0)
 957                 rte_panic("Cannot init interrupt-handling thread\n");
 958
 959         if (rte_eal_timer_init() < 0)
 960                 rte_panic("Cannot init HPET or TSC timers\n");
 961
 962         if (rte_eal_pci_init() < 0)
 963                 rte_panic("Cannot init PCI\n");
 964
 965         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
 966                 rte_config.master_lcore, (int)thread_id);
 967
 968         eal_check_mem_on_local_socket();
 969
 970         rte_eal_mcfg_complete();
 971
 972         if (rte_eal_non_pci_ethdev_init() < 0)
 973                 rte_panic("Cannot init non-PCI eth_devs\n");
 974
 975         RTE_LCORE_FOREACH_SLAVE(i) {
 976
 977                 /*
 978                  * create communication pipes between master thread
 979                  * and children
 980                  */
 981                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 982                         rte_panic("Cannot create pipe\n");
 983                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 984                         rte_panic("Cannot create pipe\n");
 985
 986                 lcore_config[i].state = WAIT;
 987
 988                 /* create a thread for each lcore */
 989                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 990                                      eal_thread_loop, NULL);
 991                 if (ret != 0)
 992                         rte_panic("Cannot create thread\n");
 993         }
 994
 995         eal_thread_init_master(rte_config.master_lcore);
 996
 997         /*
 998          * Launch a dummy function on all slave lcores, so that master lcore
 999          * knows they are all ready when this function returns.
1000          */
1001         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1002         rte_eal_mp_wait_lcore();
1003
1004         return fctret;
1005 }
1006
1007 /* get core role */
1008 enum rte_lcore_role_t
1009 rte_eal_lcore_role(unsigned lcore_id)
1010 {
1011         return (rte_config.lcore_role[lcore_id]);
1012 }
1013
1014 enum rte_proc_type_t
1015 rte_eal_process_type(void)
1016 {
1017         return (rte_config.process_type);
1018 }
1019