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
  52 #include <rte_common.h>
  53 #include <rte_debug.h>
  54 #include <rte_memory.h>
  55 #include <rte_memzone.h>
  56 #include <rte_launch.h>
  57 #include <rte_tailq.h>
  58 #include <rte_eal.h>
  59 #include <rte_eal_memconfig.h>
  60 #include <rte_per_lcore.h>
  61 #include <rte_lcore.h>
  62 #include <rte_log.h>
  63 #include <rte_random.h>
  64 #include <rte_cycles.h>
  65 #include <rte_string_fns.h>
  66 #include <rte_cpuflags.h>
  67 #include <rte_interrupts.h>
  68 #include <rte_pci.h>
  69 #include <rte_common.h>
  70 #include <rte_version.h>
  71 #include <rte_atomic.h>
  72 #include <malloc_heap.h>
  73 #include <rte_eth_ring.h>
  74
  75 #include "eal_private.h"
  76 #include "eal_thread.h"
  77 #include "eal_internal_cfg.h"
  78 #include "eal_filesystem.h"
  79 #include "eal_hugepages.h"
  80
  81 #define OPT_HUGE_DIR    "huge-dir"
  82 #define OPT_PROC_TYPE   "proc-type"
  83 #define OPT_NO_SHCONF   "no-shconf"
  84 #define OPT_NO_HPET     "no-hpet"
  85 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  86 #define OPT_NO_PCI      "no-pci"
  87 #define OPT_NO_HUGE     "no-huge"
  88 #define OPT_FILE_PREFIX "file-prefix"
  89 #define OPT_SOCKET_MEM  "socket-mem"
  90 #define OPT_USE_DEVICE  "use-device"
  91 #define OPT_SYSLOG      "syslog"
  92 #define OPT_BASE_VIRTADDR   "base-virtaddr"
  93 #define OPT_XEN_DOM0    "xen-dom0"
  94
  95 #define RTE_EAL_BLACKLIST_SIZE  0x100
  96
  97 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  98
  99 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 100
 101 #define HIGHEST_RPL 3
 102
 103 #define BITS_PER_HEX 4
 104
 105 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 106 {                                                               \
 107         unsigned long val;                                      \
 108         char *end;                                              \
 109         errno = 0;                                              \
 110         val = strtoul((in), &end, 16);                          \
 111         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 112                 return (-EINVAL);                               \
 113         (fd) = (typeof (fd))val;                                \
 114         (in) = end + 1;                                         \
 115 }
 116
 117 /* Allow the application to print its usage message too if set */
 118 static rte_usage_hook_t rte_application_usage_hook = NULL;
 119 /* early configuration structure, when memory config is not mmapped */
 120 static struct rte_mem_config early_mem_config;
 121
 122 /* define fd variable here, because file needs to be kept open for the
 123  * duration of the program, as we hold a write lock on it in the primary proc */
 124 static int mem_cfg_fd = -1;
 125
 126 static struct flock wr_lock = {
 127                 .l_type = F_WRLCK,
 128                 .l_whence = SEEK_SET,
 129                 .l_start = offsetof(struct rte_mem_config, memseg),
 130                 .l_len = sizeof(early_mem_config.memseg),
 131 };
 132
 133 /* Address of global and public configuration */
 134 static struct rte_config rte_config = {
 135                 .mem_config = &early_mem_config,
 136 };
 137
 138 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 139
 140 /* internal configuration (per-core) */
 141 struct lcore_config lcore_config[RTE_MAX_LCORE];
 142
 143 /* internal configuration */
 144 struct internal_config internal_config;
 145
 146 /* used by rte_rdtsc() */
 147 int rte_cycles_vmware_tsc_map;
 148
 149 /* Return a pointer to the configuration structure */
 150 struct rte_config *
 151 rte_eal_get_configuration(void)
 152 {
 153         return &rte_config;
 154 }
 155
 156 /* parse a sysfs (or other) file containing one integer value */
 157 int
 158 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 159 {
 160         FILE *f;
 161         char buf[BUFSIZ];
 162         char *end = NULL;
 163
 164         if ((f = fopen(filename, "r")) == NULL) {
 165                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 166                         __func__, filename);
 167                 return -1;
 168         }
 169
 170         if (fgets(buf, sizeof(buf), f) == NULL) {
 171                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 172                         __func__, filename);
 173                 fclose(f);
 174                 return -1;
 175         }
 176         *val = strtoul(buf, &end, 0);
 177         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 178                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 179                                 __func__, filename);
 180                 fclose(f);
 181                 return -1;
 182         }
 183         fclose(f);
 184         return 0;
 185 }
 186
 187
 188 /* create memory configuration in shared/mmap memory. Take out
 189  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 190  * This means we never close the file while running (auto-close on exit).
 191  * We also don't lock the whole file, so that in future we can use read-locks
 192  * on other parts, e.g. memzones, to detect if there are running secondary
 193  * processes. */
 194 static void
 195 rte_eal_config_create(void)
 196 {
 197         void *rte_mem_cfg_addr;
 198         int retval;
 199
 200         const char *pathname = eal_runtime_config_path();
 201
 202         if (internal_config.no_shconf)
 203                 return;
 204
 205         if (mem_cfg_fd < 0){
 206                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 207                 if (mem_cfg_fd < 0)
 208                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 209         }
 210
 211         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 212         if (retval < 0){
 213                 close(mem_cfg_fd);
 214                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 215         }
 216
 217         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 218         if (retval < 0){
 219                 close(mem_cfg_fd);
 220                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 221                                 "process running?\n", pathname);
 222         }
 223
 224         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 225                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 226
 227         if (rte_mem_cfg_addr == MAP_FAILED){
 228                 rte_panic("Cannot mmap memory for rte_config\n");
 229         }
 230         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 231         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 232 }
 233
 234 /* attach to an existing shared memory config */
 235 static void
 236 rte_eal_config_attach(void)
 237 {
 238         void *rte_mem_cfg_addr;
 239         const char *pathname = eal_runtime_config_path();
 240
 241         if (internal_config.no_shconf)
 242                 return;
 243
 244         if (mem_cfg_fd < 0){
 245                 mem_cfg_fd = open(pathname, O_RDWR);
 246                 if (mem_cfg_fd < 0)
 247                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 248         }
 249
 250         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 251                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 252         close(mem_cfg_fd);
 253         if (rte_mem_cfg_addr == MAP_FAILED)
 254                 rte_panic("Cannot mmap memory for rte_config\n");
 255
 256         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 257 }
 258
 259 /* Detect if we are a primary or a secondary process */
 260 static enum rte_proc_type_t
 261 eal_proc_type_detect(void)
 262 {
 263         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 264         const char *pathname = eal_runtime_config_path();
 265
 266         /* if we can open the file but not get a write-lock we are a secondary
 267          * process. NOTE: if we get a file handle back, we keep that open
 268          * and don't close it to prevent a race condition between multiple opens */
 269         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 270                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 271                 ptype = RTE_PROC_SECONDARY;
 272
 273         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 274                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 275
 276         return ptype;
 277 }
 278
 279 /* Sets up rte_config structure with the pointer to shared memory config.*/
 280 static void
 281 rte_config_init(void)
 282 {
 283         /* set the magic in configuration structure */
 284         rte_config.magic = RTE_MAGIC;
 285         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 286                         eal_proc_type_detect() : /* for auto, detect the type */
 287                         internal_config.process_type; /* otherwise use what's already set */
 288
 289         switch (rte_config.process_type){
 290         case RTE_PROC_PRIMARY:
 291                 rte_eal_config_create();
 292                 break;
 293         case RTE_PROC_SECONDARY:
 294                 rte_eal_config_attach();
 295                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 296                 break;
 297         case RTE_PROC_AUTO:
 298         case RTE_PROC_INVALID:
 299                 rte_panic("Invalid process type\n");
 300         }
 301 }
 302
 303 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 304 static void
 305 eal_hugedirs_unlock(void)
 306 {
 307         int i;
 308
 309         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 310         {
 311                 /* skip uninitialized */
 312                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 313                         continue;
 314                 /* unlock hugepage file */
 315                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 316                 close(internal_config.hugepage_info[i].lock_descriptor);
 317                 /* reset the field */
 318                 internal_config.hugepage_info[i].lock_descriptor = -1;
 319         }
 320 }
 321
 322 /* display usage */
 323 static void
 324 eal_usage(const char *prgname)
 325 {
 326         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 327                "[--proc-type primary|secondary|auto] \n\n"
 328                "EAL options:\n"
 329                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 330                "  -n NUM       : Number of memory channels\n"
 331                    "  -v           : Display version information on startup\n"
 332                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 333            "PCI device\n"
 334                "                 (multiple -b options are allowed)\n"
 335                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 336                "  -r NUM       : force number of memory ranks (don't detect)\n"
 337                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 338                            "without hugetlbfs\n"
 339                "  --"OPT_SYSLOG"     : set syslog facility\n"
 340                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 341                    "                 sockets (use comma separated values)\n"
 342                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 343                "  --"OPT_PROC_TYPE"  : type of this process\n"
 344                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 345                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 346                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 347                "               [NOTE: Cannot be used with -b option]\n"
 348                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 349                            "native RDTSC\n"
 350                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 351                "\nEAL options for DEBUG use only:\n"
 352                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 353                "  --"OPT_NO_PCI"   : disable pci\n"
 354                "  --"OPT_NO_HPET"  : disable hpet\n"
 355                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 356                "\n",
 357                prgname);
 358         /* Allow the application to print its usage message too if hook is set */
 359         if ( rte_application_usage_hook ) {
 360                 printf("===== Application Usage =====\n\n");
 361                 rte_application_usage_hook(prgname);
 362         }
 363 }
 364
 365 /* Set a per-application usage message */
 366 rte_usage_hook_t
 367 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 368 {
 369         rte_usage_hook_t        old_func;
 370
 371         /* Will be NULL on the first call to denote the last usage routine. */
 372         old_func                                        = rte_application_usage_hook;
 373         rte_application_usage_hook      = usage_func;
 374
 375         return old_func;
 376 }
 377
 378 /*
 379  * Parse the coremask given as argument (hexadecimal string) and fill
 380  * the global configuration (core role and core count) with the parsed
 381  * value.
 382  */
 383 static int xdigit2val(unsigned char c)
 384 {
 385         int val;
 386         if(isdigit(c))
 387                 val = c - '0';
 388         else if(isupper(c))
 389                 val = c - 'A' + 10;
 390         else
 391                 val = c - 'a' + 10;
 392         return val;
 393 }
 394 static int
 395 eal_parse_coremask(const char *coremask)
 396 {
 397         struct rte_config *cfg = rte_eal_get_configuration();
 398         int i, j, idx = 0 ;
 399         unsigned count = 0;
 400         char c;
 401         int val;
 402
 403         if (coremask == NULL)
 404                 return -1;
 405         /* Remove all blank characters ahead and after .
 406          * Remove 0x/0X if exists.
 407          */
 408         while (isblank(*coremask))
 409                 coremask++;
 410         if (coremask[0] == '0' && ((coremask[1] == 'x')
 411                 ||  (coremask[1] == 'X')) )
 412                 coremask += 2;
 413         i = strnlen(coremask, PATH_MAX);
 414         while ((i > 0) && isblank(coremask[i - 1]))
 415                 i--;
 416         if (i == 0)
 417                 return -1;
 418
 419         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 420                 c = coremask[i];
 421                 if (isxdigit(c) == 0) {
 422                         /* invalid characters */
 423                         return (-1);
 424                 }
 425                 val = xdigit2val(c);
 426                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 427                         if((1 << j) & val) {
 428                                 cfg->lcore_role[idx] = ROLE_RTE;
 429                                 if(count == 0)
 430                                         cfg->master_lcore = idx;
 431                                 count++;
 432                         } else  {
 433                                 cfg->lcore_role[idx] = ROLE_OFF;
 434                         }
 435                 }
 436         }
 437         for(; i >= 0; i--)
 438                 if(coremask[i] != '0')
 439                         return -1;
 440         for(; idx < RTE_MAX_LCORE; idx++)
 441                 cfg->lcore_role[idx] = ROLE_OFF;
 442         if(count == 0)
 443                 return -1;
 444         return 0;
 445 }
 446
 447 static int
 448 eal_parse_syslog(const char *facility)
 449 {
 450         int i;
 451         static struct {
 452                 const char *name;
 453                 int value;
 454         } map[] = {
 455                 { "auth", LOG_AUTH },
 456                 { "cron", LOG_CRON },
 457                 { "daemon", LOG_DAEMON },
 458                 { "ftp", LOG_FTP },
 459                 { "kern", LOG_KERN },
 460                 { "lpr", LOG_LPR },
 461                 { "mail", LOG_MAIL },
 462                 { "news", LOG_NEWS },
 463                 { "syslog", LOG_SYSLOG },
 464                 { "user", LOG_USER },
 465                 { "uucp", LOG_UUCP },
 466                 { "local0", LOG_LOCAL0 },
 467                 { "local1", LOG_LOCAL1 },
 468                 { "local2", LOG_LOCAL2 },
 469                 { "local3", LOG_LOCAL3 },
 470                 { "local4", LOG_LOCAL4 },
 471                 { "local5", LOG_LOCAL5 },
 472                 { "local6", LOG_LOCAL6 },
 473                 { "local7", LOG_LOCAL7 },
 474                 { NULL, 0 }
 475         };
 476
 477         for (i = 0; map[i].name; i++) {
 478                 if (!strcmp(facility, map[i].name)) {
 479                         internal_config.syslog_facility = map[i].value;
 480                         return 0;
 481                 }
 482         }
 483         return -1;
 484 }
 485
 486 static int
 487 eal_parse_socket_mem(char *socket_mem)
 488 {
 489         char * arg[RTE_MAX_NUMA_NODES];
 490         char *end;
 491         int arg_num, i, len;
 492         uint64_t total_mem = 0;
 493
 494         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 495         if (len == SOCKET_MEM_STRLEN) {
 496                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 497                 return -1;
 498         }
 499
 500         /* all other error cases will be caught later */
 501         if (!isdigit(socket_mem[len-1]))
 502                 return -1;
 503
 504         /* split the optarg into separate socket values */
 505         arg_num = rte_strsplit(socket_mem, len,
 506                         arg, RTE_MAX_NUMA_NODES, ',');
 507
 508         /* if split failed, or 0 arguments */
 509         if (arg_num <= 0)
 510                 return -1;
 511
 512         internal_config.force_sockets = 1;
 513
 514         /* parse each defined socket option */
 515         errno = 0;
 516         for (i = 0; i < arg_num; i++) {
 517                 end = NULL;
 518                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 519
 520                 /* check for invalid input */
 521                 if ((errno != 0)  ||
 522                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 523                         return -1;
 524                 internal_config.socket_mem[i] *= 1024ULL;
 525                 internal_config.socket_mem[i] *= 1024ULL;
 526                 total_mem += internal_config.socket_mem[i];
 527         }
 528
 529         /* check if we have a positive amount of total memory */
 530         if (total_mem == 0)
 531                 return -1;
 532
 533         return 0;
 534 }
 535
 536 static int
 537 eal_parse_base_virtaddr(const char *arg)
 538 {
 539         char *end;
 540         uint64_t addr;
 541
 542         addr = strtoull(arg, &end, 16);
 543
 544         /* check for errors */
 545         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 546                 return -1;
 547
 548         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 549 #ifndef RTE_ARCH_X86_64
 550         if (addr >= UINTPTR_MAX)
 551                 return -1;
 552 #endif
 553
 554         /* align the addr on 2M boundary */
 555         addr = RTE_PTR_ALIGN_CEIL(addr, RTE_PGSIZE_2M);
 556
 557         internal_config.base_virtaddr = (uintptr_t) addr;
 558         return 0;
 559 }
 560
 561 static inline size_t
 562 eal_get_hugepage_mem_size(void)
 563 {
 564         uint64_t size = 0;
 565         unsigned i, j;
 566
 567         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 568                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 569                 if (hpi->hugedir != NULL) {
 570                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 571                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 572                         }
 573                 }
 574         }
 575
 576         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 577 }
 578
 579 static enum rte_proc_type_t
 580 eal_parse_proc_type(const char *arg)
 581 {
 582         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 583                 return RTE_PROC_PRIMARY;
 584         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 585                 return RTE_PROC_SECONDARY;
 586         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 587                 return RTE_PROC_AUTO;
 588
 589         return RTE_PROC_INVALID;
 590 }
 591
 592 static ssize_t
 593 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 594 {
 595         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 596                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 597                 return (-EINVAL);
 598         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 599                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 600                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 601                 return (-EINVAL);
 602         }
 603
 604         idx += 1;
 605         return (idx);
 606 }
 607
 608 /* Parse the argument given in the command line of the application */
 609 static int
 610 eal_parse_args(int argc, char **argv)
 611 {
 612         int opt, ret, i;
 613         char **argvopt;
 614         int option_index;
 615         int coremask_ok = 0;
 616         ssize_t blacklist_index = 0;
 617         char *prgname = argv[0];
 618         static struct option lgopts[] = {
 619                 {OPT_NO_HUGE, 0, 0, 0},
 620                 {OPT_NO_PCI, 0, 0, 0},
 621                 {OPT_NO_HPET, 0, 0, 0},
 622                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 623                 {OPT_HUGE_DIR, 1, 0, 0},
 624                 {OPT_NO_SHCONF, 0, 0, 0},
 625                 {OPT_PROC_TYPE, 1, 0, 0},
 626                 {OPT_FILE_PREFIX, 1, 0, 0},
 627                 {OPT_SOCKET_MEM, 1, 0, 0},
 628                 {OPT_USE_DEVICE, 1, 0, 0},
 629                 {OPT_SYSLOG, 1, NULL, 0},
 630                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 631                 {OPT_XEN_DOM0, 0, 0, 0},
 632                 {0, 0, 0, 0}
 633         };
 634
 635         argvopt = argv;
 636
 637         internal_config.memory = 0;
 638         internal_config.force_nrank = 0;
 639         internal_config.force_nchannel = 0;
 640         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 641         internal_config.hugepage_dir = NULL;
 642         internal_config.force_sockets = 0;
 643         internal_config.syslog_facility = LOG_DAEMON;
 644         internal_config.xen_dom0_support = 0;
 645 #ifdef RTE_LIBEAL_USE_HPET
 646         internal_config.no_hpet = 0;
 647 #else
 648         internal_config.no_hpet = 1;
 649 #endif
 650         /* zero out the NUMA config */
 651         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 652                 internal_config.socket_mem[i] = 0;
 653
 654         /* zero out hugedir descriptors */
 655         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 656                 internal_config.hugepage_info[i].lock_descriptor = -1;
 657
 658         internal_config.vmware_tsc_map = 0;
 659         internal_config.base_virtaddr = 0;
 660
 661         while ((opt = getopt_long(argc, argvopt, "b:c:m:n:r:v",
 662                                   lgopts, &option_index)) != EOF) {
 663
 664                 switch (opt) {
 665                 /* blacklist */
 666                 case 'b':
 667                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 668                             blacklist_index)) < 0) {
 669                                 eal_usage(prgname);
 670                                 return (-1);
 671                         }
 672                         break;
 673                 /* coremask */
 674                 case 'c':
 675                         if (eal_parse_coremask(optarg) < 0) {
 676                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 677                                 eal_usage(prgname);
 678                                 return -1;
 679                         }
 680                         coremask_ok = 1;
 681                         break;
 682                 /* size of memory */
 683                 case 'm':
 684                         internal_config.memory = atoi(optarg);
 685                         internal_config.memory *= 1024ULL;
 686                         internal_config.memory *= 1024ULL;
 687                         break;
 688                 /* force number of channels */
 689                 case 'n':
 690                         internal_config.force_nchannel = atoi(optarg);
 691                         if (internal_config.force_nchannel == 0 ||
 692                             internal_config.force_nchannel > 4) {
 693                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 694                                 eal_usage(prgname);
 695                                 return -1;
 696                         }
 697                         break;
 698                 /* force number of ranks */
 699                 case 'r':
 700                         internal_config.force_nrank = atoi(optarg);
 701                         if (internal_config.force_nrank == 0 ||
 702                             internal_config.force_nrank > 16) {
 703                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 704                                 eal_usage(prgname);
 705                                 return -1;
 706                         }
 707                         break;
 708                 case 'v':
 709                         /* since message is explicitly requested by user, we
 710                          * write message at highest log level so it can always be seen
 711                          * even if info or warning messages are disabled */
 712                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 713                         break;
 714
 715                 /* long options */
 716                 case 0:
 717                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 718                                 internal_config.no_hugetlbfs = 1;
 719                         }
 720                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 721                 #ifdef RTE_LIBRTE_XEN_DOM0
 722                                 internal_config.xen_dom0_support = 1;
 723                 #else
 724                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 725                                         "running on Dom0, please configure"
 726                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 727                                 return -1;
 728                 #endif
 729                         }
 730                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 731                                 internal_config.no_pci = 1;
 732                         }
 733                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 734                                 internal_config.no_hpet = 1;
 735                         }
 736                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 737                                 internal_config.vmware_tsc_map = 1;
 738                         }
 739                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 740                                 internal_config.no_shconf = 1;
 741                         }
 742                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 743                                 internal_config.hugepage_dir = optarg;
 744                         }
 745                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 746                                 internal_config.process_type = eal_parse_proc_type(optarg);
 747                         }
 748                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 749                                 internal_config.hugefile_prefix = optarg;
 750                         }
 751                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 752                                 if (eal_parse_socket_mem(optarg) < 0) {
 753                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 754                                                         OPT_SOCKET_MEM "\n");
 755                                         eal_usage(prgname);
 756                                         return -1;
 757                                 }
 758                         }
 759                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 760                                 eal_dev_whitelist_add_entry(optarg);
 761                         }
 762                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 763                                 if (eal_parse_syslog(optarg) < 0) {
 764                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 765                                                         OPT_SYSLOG "\n");
 766                                         eal_usage(prgname);
 767                                         return -1;
 768                                 }
 769                         }
 770                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 771                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 772                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 773                                                         OPT_BASE_VIRTADDR "\n");
 774                                         eal_usage(prgname);
 775                                         return -1;
 776                                 }
 777                         }
 778                         break;
 779
 780                 default:
 781                         eal_usage(prgname);
 782                         return -1;
 783                 }
 784         }
 785
 786         /* sanity checks */
 787         if (!coremask_ok) {
 788                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 789                 eal_usage(prgname);
 790                 return -1;
 791         }
 792         if (internal_config.process_type == RTE_PROC_AUTO){
 793                 internal_config.process_type = eal_proc_type_detect();
 794         }
 795         if (internal_config.process_type == RTE_PROC_INVALID){
 796                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 797                 eal_usage(prgname);
 798                 return -1;
 799         }
 800         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 801                         internal_config.force_nchannel == 0) {
 802                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 803                 eal_usage(prgname);
 804                 return -1;
 805         }
 806         if (index(internal_config.hugefile_prefix,'%') != NULL){
 807                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 808                 eal_usage(prgname);
 809                 return -1;
 810         }
 811         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 812                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 813                                 "at the same time\n");
 814                 eal_usage(prgname);
 815                 return -1;
 816         }
 817         /* --no-huge doesn't make sense with either -m or --socket-mem */
 818         if (internal_config.no_hugetlbfs &&
 819                         (internal_config.memory > 0 ||
 820                                         internal_config.force_sockets == 1)) {
 821                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 822                                 "together with --no-huge!\n");
 823                 eal_usage(prgname);
 824                 return -1;
 825         }
 826         /* --xen-dom0 doesn't make sense with --socket-mem */
 827         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 828                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 829                                         "together with --xen_dom0!\n");
 830                 eal_usage(prgname);
 831                 return -1;
 832         }
 833         /* if no blacklist, parse a whitelist */
 834         if (blacklist_index > 0) {
 835                 if (eal_dev_whitelist_exists()) {
 836                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 837                                         "[--use-device] options cannot be used at the same time\n");
 838                         eal_usage(prgname);
 839                         return -1;
 840                 }
 841                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 842         } else {
 843                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 844                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 845                         return -1;
 846                 }
 847         }
 848
 849         if (optind >= 0)
 850                 argv[optind-1] = prgname;
 851
 852         /* if no memory amounts were requested, this will result in 0 and
 853          * will be overriden later, right after eal_hugepage_info_init() */
 854         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 855                 internal_config.memory += internal_config.socket_mem[i];
 856
 857         ret = optind-1;
 858         optind = 0; /* reset getopt lib */
 859         return ret;
 860 }
 861
 862 static void
 863 eal_check_mem_on_local_socket(void)
 864 {
 865         const struct rte_memseg *ms;
 866         int i, socket_id;
 867
 868         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 869
 870         ms = rte_eal_get_physmem_layout();
 871
 872         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 873                 if (ms[i].socket_id == socket_id &&
 874                                 ms[i].len > 0)
 875                         return;
 876
 877         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 878                         "memory on local socket!\n");
 879 }
 880
 881 static int
 882 sync_func(__attribute__((unused)) void *arg)
 883 {
 884         return 0;
 885 }
 886
 887 inline static void
 888 rte_eal_mcfg_complete(void)
 889 {
 890         /* ALL shared mem_config related INIT DONE */
 891         if (rte_config.process_type == RTE_PROC_PRIMARY)
 892                 rte_config.mem_config->magic = RTE_MAGIC;
 893 }
 894
 895 /*
 896  * Request iopl priviledge for all RPL, returns 0 on success
 897  */
 898 static int
 899 rte_eal_iopl_init(void)
 900 {
 901         return iopl(HIGHEST_RPL);
 902 }
 903
 904 /* Launch threads, called at application init(). */
 905 int
 906 rte_eal_init(int argc, char **argv)
 907 {
 908         int i, fctret, ret;
 909         pthread_t thread_id;
 910         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 911
 912         if (!rte_atomic32_test_and_set(&run_once))
 913                 return -1;
 914
 915         thread_id = pthread_self();
 916
 917         if (rte_eal_log_early_init() < 0)
 918                 rte_panic("Cannot init early logs\n");
 919
 920         fctret = eal_parse_args(argc, argv);
 921         if (fctret < 0)
 922                 exit(1);
 923
 924         if (internal_config.no_hugetlbfs == 0 &&
 925                         internal_config.process_type != RTE_PROC_SECONDARY &&
 926                         internal_config.xen_dom0_support == 0 &&
 927                         eal_hugepage_info_init() < 0)
 928                 rte_panic("Cannot get hugepage information\n");
 929
 930         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 931                 if (internal_config.no_hugetlbfs)
 932                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 933                 else
 934                         internal_config.memory = eal_get_hugepage_mem_size();
 935         }
 936
 937         if (internal_config.vmware_tsc_map == 1) {
 938 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 939                 rte_cycles_vmware_tsc_map = 1;
 940                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 941                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 942 #else
 943                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 944                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 945 #endif
 946         }
 947
 948         rte_srand(rte_rdtsc());
 949
 950         rte_config_init();
 951
 952         if (rte_eal_iopl_init() == 0)
 953                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 954
 955         if (rte_eal_cpu_init() < 0)
 956                 rte_panic("Cannot detect lcores\n");
 957
 958         if (rte_eal_pci_init() < 0)
 959                 rte_panic("Cannot init PCI\n");
 960
 961 #ifdef RTE_LIBRTE_IVSHMEM
 962         if (rte_eal_ivshmem_init() < 0)
 963                 rte_panic("Cannot init IVSHMEM\n");
 964 #endif
 965
 966         if (rte_eal_memory_init() < 0)
 967                 rte_panic("Cannot init memory\n");
 968
 969         /* the directories are locked during eal_hugepage_info_init */
 970         eal_hugedirs_unlock();
 971
 972         if (rte_eal_memzone_init() < 0)
 973                 rte_panic("Cannot init memzone\n");
 974
 975         if (rte_eal_tailqs_init() < 0)
 976                 rte_panic("Cannot init tail queues for objects\n");
 977
 978 #ifdef RTE_LIBRTE_IVSHMEM
 979         if (rte_eal_ivshmem_obj_init() < 0)
 980                 rte_panic("Cannot init IVSHMEM objects\n");
 981 #endif
 982
 983         if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 984                 rte_panic("Cannot init logs\n");
 985
 986         if (rte_eal_alarm_init() < 0)
 987                 rte_panic("Cannot init interrupt-handling thread\n");
 988
 989         if (rte_eal_intr_init() < 0)
 990                 rte_panic("Cannot init interrupt-handling thread\n");
 991
 992         if (rte_eal_timer_init() < 0)
 993                 rte_panic("Cannot init HPET or TSC timers\n");
 994
 995         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
 996                 rte_config.master_lcore, (int)thread_id);
 997
 998         eal_check_mem_on_local_socket();
 999
1000         rte_eal_mcfg_complete();
1001
1002         if (rte_eal_non_pci_ethdev_init() < 0)
1003                 rte_panic("Cannot init non-PCI eth_devs\n");
1004
1005         RTE_LCORE_FOREACH_SLAVE(i) {
1006
1007                 /*
1008                  * create communication pipes between master thread
1009                  * and children
1010                  */
1011                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1012                         rte_panic("Cannot create pipe\n");
1013                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1014                         rte_panic("Cannot create pipe\n");
1015
1016                 lcore_config[i].state = WAIT;
1017
1018                 /* create a thread for each lcore */
1019                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1020                                      eal_thread_loop, NULL);
1021                 if (ret != 0)
1022                         rte_panic("Cannot create thread\n");
1023         }
1024
1025         eal_thread_init_master(rte_config.master_lcore);
1026
1027         /*
1028          * Launch a dummy function on all slave lcores, so that master lcore
1029          * knows they are all ready when this function returns.
1030          */
1031         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1032         rte_eal_mp_wait_lcore();
1033
1034         return fctret;
1035 }
1036
1037 /* get core role */
1038 enum rte_lcore_role_t
1039 rte_eal_lcore_role(unsigned lcore_id)
1040 {
1041         return (rte_config.lcore_role[lcore_id]);
1042 }
1043
1044 enum rte_proc_type_t
1045 rte_eal_process_type(void)
1046 {
1047         return (rte_config.process_type);
1048 }
1049