lib/librte_eal/linuxapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2012-2013 6WIND S.A.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <string.h>
  39 #include <stdarg.h>
  40 #include <unistd.h>
  41 #include <pthread.h>
  42 #include <syslog.h>
  43 #include <getopt.h>
  44 #include <sys/file.h>
  45 #include <fcntl.h>
  46 #include <dlfcn.h>
  47 #include <stddef.h>
  48 #include <errno.h>
  49 #include <limits.h>
  50 #include <errno.h>
  51 #include <sys/mman.h>
  52 #include <sys/queue.h>
  53 #include <sys/io.h>
  54
  55 #include <rte_common.h>
  56 #include <rte_debug.h>
  57 #include <rte_memory.h>
  58 #include <rte_memzone.h>
  59 #include <rte_launch.h>
  60 #include <rte_tailq.h>
  61 #include <rte_eal.h>
  62 #include <rte_eal_memconfig.h>
  63 #include <rte_per_lcore.h>
  64 #include <rte_lcore.h>
  65 #include <rte_log.h>
  66 #include <rte_random.h>
  67 #include <rte_cycles.h>
  68 #include <rte_string_fns.h>
  69 #include <rte_cpuflags.h>
  70 #include <rte_interrupts.h>
  71 #include <rte_pci.h>
  72 #include <rte_common.h>
  73 #include <rte_version.h>
  74 #include <rte_atomic.h>
  75 #include <malloc_heap.h>
  76 #include <rte_eth_ring.h>
  77
  78 #include "eal_private.h"
  79 #include "eal_thread.h"
  80 #include "eal_internal_cfg.h"
  81 #include "eal_filesystem.h"
  82 #include "eal_hugepages.h"
  83
  84 #define OPT_HUGE_DIR    "huge-dir"
  85 #define OPT_PROC_TYPE   "proc-type"
  86 #define OPT_NO_SHCONF   "no-shconf"
  87 #define OPT_NO_HPET     "no-hpet"
  88 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  89 #define OPT_NO_PCI      "no-pci"
  90 #define OPT_NO_HUGE     "no-huge"
  91 #define OPT_FILE_PREFIX "file-prefix"
  92 #define OPT_SOCKET_MEM  "socket-mem"
  93 #define OPT_USE_DEVICE  "use-device"
  94 #define OPT_SYSLOG      "syslog"
  95 #define OPT_BASE_VIRTADDR   "base-virtaddr"
  96 #define OPT_XEN_DOM0    "xen-dom0"
  97
  98 #define RTE_EAL_BLACKLIST_SIZE  0x100
  99
 100 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 101
 102 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 103
 104 #define HIGHEST_RPL 3
 105
 106 #define BITS_PER_HEX 4
 107
 108 /* Allow the application to print its usage message too if set */
 109 static rte_usage_hook_t rte_application_usage_hook = NULL;
 110
 111 TAILQ_HEAD(shared_driver_list, shared_driver);
 112
 113 /* Definition for shared object drivers. */
 114 struct shared_driver {
 115         TAILQ_ENTRY(shared_driver) next;
 116
 117         char    name[PATH_MAX];
 118         void*   lib_handle;
 119 };
 120
 121 /* List of external loadable drivers */
 122 static struct shared_driver_list solib_list =
 123 TAILQ_HEAD_INITIALIZER(solib_list);
 124
 125 /* early configuration structure, when memory config is not mmapped */
 126 static struct rte_mem_config early_mem_config;
 127
 128 /* define fd variable here, because file needs to be kept open for the
 129  * duration of the program, as we hold a write lock on it in the primary proc */
 130 static int mem_cfg_fd = -1;
 131
 132 static struct flock wr_lock = {
 133                 .l_type = F_WRLCK,
 134                 .l_whence = SEEK_SET,
 135                 .l_start = offsetof(struct rte_mem_config, memseg),
 136                 .l_len = sizeof(early_mem_config.memseg),
 137 };
 138
 139 /* Address of global and public configuration */
 140 static struct rte_config rte_config = {
 141                 .mem_config = &early_mem_config,
 142 };
 143
 144 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 145
 146 /* internal configuration (per-core) */
 147 struct lcore_config lcore_config[RTE_MAX_LCORE];
 148
 149 /* internal configuration */
 150 struct internal_config internal_config;
 151
 152 /* used by rte_rdtsc() */
 153 int rte_cycles_vmware_tsc_map;
 154
 155 /* Return a pointer to the configuration structure */
 156 struct rte_config *
 157 rte_eal_get_configuration(void)
 158 {
 159         return &rte_config;
 160 }
 161
 162 /* parse a sysfs (or other) file containing one integer value */
 163 int
 164 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 165 {
 166         FILE *f;
 167         char buf[BUFSIZ];
 168         char *end = NULL;
 169
 170         if ((f = fopen(filename, "r")) == NULL) {
 171                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 172                         __func__, filename);
 173                 return -1;
 174         }
 175
 176         if (fgets(buf, sizeof(buf), f) == NULL) {
 177                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 178                         __func__, filename);
 179                 fclose(f);
 180                 return -1;
 181         }
 182         *val = strtoul(buf, &end, 0);
 183         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 184                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 185                                 __func__, filename);
 186                 fclose(f);
 187                 return -1;
 188         }
 189         fclose(f);
 190         return 0;
 191 }
 192
 193
 194 /* create memory configuration in shared/mmap memory. Take out
 195  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 196  * This means we never close the file while running (auto-close on exit).
 197  * We also don't lock the whole file, so that in future we can use read-locks
 198  * on other parts, e.g. memzones, to detect if there are running secondary
 199  * processes. */
 200 static void
 201 rte_eal_config_create(void)
 202 {
 203         void *rte_mem_cfg_addr;
 204         int retval;
 205
 206         const char *pathname = eal_runtime_config_path();
 207
 208         if (internal_config.no_shconf)
 209                 return;
 210
 211         if (mem_cfg_fd < 0){
 212                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 213                 if (mem_cfg_fd < 0)
 214                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 215         }
 216
 217         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 218         if (retval < 0){
 219                 close(mem_cfg_fd);
 220                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 221         }
 222
 223         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 224         if (retval < 0){
 225                 close(mem_cfg_fd);
 226                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 227                                 "process running?\n", pathname);
 228         }
 229
 230         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 231                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 232
 233         if (rte_mem_cfg_addr == MAP_FAILED){
 234                 rte_panic("Cannot mmap memory for rte_config\n");
 235         }
 236         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 237         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 238 }
 239
 240 /* attach to an existing shared memory config */
 241 static void
 242 rte_eal_config_attach(void)
 243 {
 244         void *rte_mem_cfg_addr;
 245         const char *pathname = eal_runtime_config_path();
 246
 247         if (internal_config.no_shconf)
 248                 return;
 249
 250         if (mem_cfg_fd < 0){
 251                 mem_cfg_fd = open(pathname, O_RDWR);
 252                 if (mem_cfg_fd < 0)
 253                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 254         }
 255
 256         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 257                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 258         close(mem_cfg_fd);
 259         if (rte_mem_cfg_addr == MAP_FAILED)
 260                 rte_panic("Cannot mmap memory for rte_config\n");
 261
 262         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 263 }
 264
 265 /* Detect if we are a primary or a secondary process */
 266 static enum rte_proc_type_t
 267 eal_proc_type_detect(void)
 268 {
 269         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 270         const char *pathname = eal_runtime_config_path();
 271
 272         /* if we can open the file but not get a write-lock we are a secondary
 273          * process. NOTE: if we get a file handle back, we keep that open
 274          * and don't close it to prevent a race condition between multiple opens */
 275         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 276                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 277                 ptype = RTE_PROC_SECONDARY;
 278
 279         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 280                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 281
 282         return ptype;
 283 }
 284
 285 /* Sets up rte_config structure with the pointer to shared memory config.*/
 286 static void
 287 rte_config_init(void)
 288 {
 289         /* set the magic in configuration structure */
 290         rte_config.magic = RTE_MAGIC;
 291         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 292                         eal_proc_type_detect() : /* for auto, detect the type */
 293                         internal_config.process_type; /* otherwise use what's already set */
 294
 295         switch (rte_config.process_type){
 296         case RTE_PROC_PRIMARY:
 297                 rte_eal_config_create();
 298                 break;
 299         case RTE_PROC_SECONDARY:
 300                 rte_eal_config_attach();
 301                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 302                 break;
 303         case RTE_PROC_AUTO:
 304         case RTE_PROC_INVALID:
 305                 rte_panic("Invalid process type\n");
 306         }
 307 }
 308
 309 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 310 static void
 311 eal_hugedirs_unlock(void)
 312 {
 313         int i;
 314
 315         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 316         {
 317                 /* skip uninitialized */
 318                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 319                         continue;
 320                 /* unlock hugepage file */
 321                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 322                 close(internal_config.hugepage_info[i].lock_descriptor);
 323                 /* reset the field */
 324                 internal_config.hugepage_info[i].lock_descriptor = -1;
 325         }
 326 }
 327
 328 /* display usage */
 329 static void
 330 eal_usage(const char *prgname)
 331 {
 332         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 333                "[--proc-type primary|secondary|auto] \n\n"
 334                "EAL options:\n"
 335                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 336                "  -n NUM       : Number of memory channels\n"
 337                    "  -v           : Display version information on startup\n"
 338                "  -d LIB.so    : add driver (can be used multiple times)\n"
 339                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 340            "PCI device\n"
 341                "                 (multiple -b options are allowed)\n"
 342                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 343                "  -r NUM       : force number of memory ranks (don't detect)\n"
 344                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 345                            "without hugetlbfs\n"
 346                "  --"OPT_SYSLOG"     : set syslog facility\n"
 347                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 348                    "                 sockets (use comma separated values)\n"
 349                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 350                "  --"OPT_PROC_TYPE"  : type of this process\n"
 351                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 352                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 353                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 354                "               [NOTE: Cannot be used with -b option]\n"
 355                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 356                            "native RDTSC\n"
 357                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 358                "\nEAL options for DEBUG use only:\n"
 359                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 360                "  --"OPT_NO_PCI"   : disable pci\n"
 361                "  --"OPT_NO_HPET"  : disable hpet\n"
 362                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 363                "\n",
 364                prgname);
 365         /* Allow the application to print its usage message too if hook is set */
 366         if ( rte_application_usage_hook ) {
 367                 printf("===== Application Usage =====\n\n");
 368                 rte_application_usage_hook(prgname);
 369         }
 370 }
 371
 372 /* Set a per-application usage message */
 373 rte_usage_hook_t
 374 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 375 {
 376         rte_usage_hook_t        old_func;
 377
 378         /* Will be NULL on the first call to denote the last usage routine. */
 379         old_func                                        = rte_application_usage_hook;
 380         rte_application_usage_hook      = usage_func;
 381
 382         return old_func;
 383 }
 384
 385 /*
 386  * Parse the coremask given as argument (hexadecimal string) and fill
 387  * the global configuration (core role and core count) with the parsed
 388  * value.
 389  */
 390 static int xdigit2val(unsigned char c)
 391 {
 392         int val;
 393         if(isdigit(c))
 394                 val = c - '0';
 395         else if(isupper(c))
 396                 val = c - 'A' + 10;
 397         else
 398                 val = c - 'a' + 10;
 399         return val;
 400 }
 401 static int
 402 eal_parse_coremask(const char *coremask)
 403 {
 404         struct rte_config *cfg = rte_eal_get_configuration();
 405         int i, j, idx = 0 ;
 406         unsigned count = 0;
 407         char c;
 408         int val;
 409
 410         if (coremask == NULL)
 411                 return -1;
 412         /* Remove all blank characters ahead and after .
 413          * Remove 0x/0X if exists.
 414          */
 415         while (isblank(*coremask))
 416                 coremask++;
 417         if (coremask[0] == '0' && ((coremask[1] == 'x')
 418                 ||  (coremask[1] == 'X')) )
 419                 coremask += 2;
 420         i = strnlen(coremask, PATH_MAX);
 421         while ((i > 0) && isblank(coremask[i - 1]))
 422                 i--;
 423         if (i == 0)
 424                 return -1;
 425
 426         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 427                 c = coremask[i];
 428                 if (isxdigit(c) == 0) {
 429                         /* invalid characters */
 430                         return (-1);
 431                 }
 432                 val = xdigit2val(c);
 433                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 434                         if((1 << j) & val) {
 435                                 cfg->lcore_role[idx] = ROLE_RTE;
 436                                 if(count == 0)
 437                                         cfg->master_lcore = idx;
 438                                 count++;
 439                         } else  {
 440                                 cfg->lcore_role[idx] = ROLE_OFF;
 441                         }
 442                 }
 443         }
 444         for(; i >= 0; i--)
 445                 if(coremask[i] != '0')
 446                         return -1;
 447         for(; idx < RTE_MAX_LCORE; idx++)
 448                 cfg->lcore_role[idx] = ROLE_OFF;
 449         if(count == 0)
 450                 return -1;
 451         /* Update the count of enabled logical cores of the EAL configuration */
 452         cfg->lcore_count = count;
 453         return 0;
 454 }
 455
 456 static int
 457 eal_parse_syslog(const char *facility)
 458 {
 459         int i;
 460         static struct {
 461                 const char *name;
 462                 int value;
 463         } map[] = {
 464                 { "auth", LOG_AUTH },
 465                 { "cron", LOG_CRON },
 466                 { "daemon", LOG_DAEMON },
 467                 { "ftp", LOG_FTP },
 468                 { "kern", LOG_KERN },
 469                 { "lpr", LOG_LPR },
 470                 { "mail", LOG_MAIL },
 471                 { "news", LOG_NEWS },
 472                 { "syslog", LOG_SYSLOG },
 473                 { "user", LOG_USER },
 474                 { "uucp", LOG_UUCP },
 475                 { "local0", LOG_LOCAL0 },
 476                 { "local1", LOG_LOCAL1 },
 477                 { "local2", LOG_LOCAL2 },
 478                 { "local3", LOG_LOCAL3 },
 479                 { "local4", LOG_LOCAL4 },
 480                 { "local5", LOG_LOCAL5 },
 481                 { "local6", LOG_LOCAL6 },
 482                 { "local7", LOG_LOCAL7 },
 483                 { NULL, 0 }
 484         };
 485
 486         for (i = 0; map[i].name; i++) {
 487                 if (!strcmp(facility, map[i].name)) {
 488                         internal_config.syslog_facility = map[i].value;
 489                         return 0;
 490                 }
 491         }
 492         return -1;
 493 }
 494
 495 static int
 496 eal_parse_socket_mem(char *socket_mem)
 497 {
 498         char * arg[RTE_MAX_NUMA_NODES];
 499         char *end;
 500         int arg_num, i, len;
 501         uint64_t total_mem = 0;
 502
 503         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 504         if (len == SOCKET_MEM_STRLEN) {
 505                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 506                 return -1;
 507         }
 508
 509         /* all other error cases will be caught later */
 510         if (!isdigit(socket_mem[len-1]))
 511                 return -1;
 512
 513         /* split the optarg into separate socket values */
 514         arg_num = rte_strsplit(socket_mem, len,
 515                         arg, RTE_MAX_NUMA_NODES, ',');
 516
 517         /* if split failed, or 0 arguments */
 518         if (arg_num <= 0)
 519                 return -1;
 520
 521         internal_config.force_sockets = 1;
 522
 523         /* parse each defined socket option */
 524         errno = 0;
 525         for (i = 0; i < arg_num; i++) {
 526                 end = NULL;
 527                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 528
 529                 /* check for invalid input */
 530                 if ((errno != 0)  ||
 531                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 532                         return -1;
 533                 internal_config.socket_mem[i] *= 1024ULL;
 534                 internal_config.socket_mem[i] *= 1024ULL;
 535                 total_mem += internal_config.socket_mem[i];
 536         }
 537
 538         /* check if we have a positive amount of total memory */
 539         if (total_mem == 0)
 540                 return -1;
 541
 542         return 0;
 543 }
 544
 545 static int
 546 eal_parse_base_virtaddr(const char *arg)
 547 {
 548         char *end;
 549         uint64_t addr;
 550
 551         addr = strtoull(arg, &end, 16);
 552
 553         /* check for errors */
 554         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 555                 return -1;
 556
 557         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 558 #ifndef RTE_ARCH_X86_64
 559         if (addr >= UINTPTR_MAX)
 560                 return -1;
 561 #endif
 562
 563         /* align the addr on 2M boundary */
 564         addr = RTE_PTR_ALIGN_CEIL(addr, RTE_PGSIZE_2M);
 565
 566         internal_config.base_virtaddr = (uintptr_t) addr;
 567         return 0;
 568 }
 569
 570 static inline size_t
 571 eal_get_hugepage_mem_size(void)
 572 {
 573         uint64_t size = 0;
 574         unsigned i, j;
 575
 576         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 577                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 578                 if (hpi->hugedir != NULL) {
 579                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 580                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 581                         }
 582                 }
 583         }
 584
 585         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 586 }
 587
 588 static enum rte_proc_type_t
 589 eal_parse_proc_type(const char *arg)
 590 {
 591         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 592                 return RTE_PROC_PRIMARY;
 593         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 594                 return RTE_PROC_SECONDARY;
 595         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 596                 return RTE_PROC_AUTO;
 597
 598         return RTE_PROC_INVALID;
 599 }
 600
 601 static ssize_t
 602 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 603 {
 604         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 605                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 606                 return (-EINVAL);
 607         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 608                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 609                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 610                 return (-EINVAL);
 611         }
 612
 613         idx += 1;
 614         return (idx);
 615 }
 616
 617 /* Parse the argument given in the command line of the application */
 618 static int
 619 eal_parse_args(int argc, char **argv)
 620 {
 621         int opt, ret, i;
 622         char **argvopt;
 623         int option_index;
 624         int coremask_ok = 0;
 625         ssize_t blacklist_index = 0;
 626         char *prgname = argv[0];
 627         static struct option lgopts[] = {
 628                 {OPT_NO_HUGE, 0, 0, 0},
 629                 {OPT_NO_PCI, 0, 0, 0},
 630                 {OPT_NO_HPET, 0, 0, 0},
 631                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 632                 {OPT_HUGE_DIR, 1, 0, 0},
 633                 {OPT_NO_SHCONF, 0, 0, 0},
 634                 {OPT_PROC_TYPE, 1, 0, 0},
 635                 {OPT_FILE_PREFIX, 1, 0, 0},
 636                 {OPT_SOCKET_MEM, 1, 0, 0},
 637                 {OPT_USE_DEVICE, 1, 0, 0},
 638                 {OPT_SYSLOG, 1, NULL, 0},
 639                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 640                 {OPT_XEN_DOM0, 0, 0, 0},
 641                 {0, 0, 0, 0}
 642         };
 643         struct shared_driver *solib;
 644
 645         argvopt = argv;
 646
 647         internal_config.memory = 0;
 648         internal_config.force_nrank = 0;
 649         internal_config.force_nchannel = 0;
 650         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 651         internal_config.hugepage_dir = NULL;
 652         internal_config.force_sockets = 0;
 653         internal_config.syslog_facility = LOG_DAEMON;
 654         internal_config.xen_dom0_support = 0;
 655 #ifdef RTE_LIBEAL_USE_HPET
 656         internal_config.no_hpet = 0;
 657 #else
 658         internal_config.no_hpet = 1;
 659 #endif
 660         /* zero out the NUMA config */
 661         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 662                 internal_config.socket_mem[i] = 0;
 663
 664         /* zero out hugedir descriptors */
 665         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 666                 internal_config.hugepage_info[i].lock_descriptor = -1;
 667
 668         internal_config.vmware_tsc_map = 0;
 669         internal_config.base_virtaddr = 0;
 670
 671         while ((opt = getopt_long(argc, argvopt, "b:c:d:m:n:r:v",
 672                                   lgopts, &option_index)) != EOF) {
 673
 674                 switch (opt) {
 675                 /* blacklist */
 676                 case 'b':
 677                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 678                             blacklist_index)) < 0) {
 679                                 eal_usage(prgname);
 680                                 return (-1);
 681                         }
 682                         break;
 683                 /* coremask */
 684                 case 'c':
 685                         if (eal_parse_coremask(optarg) < 0) {
 686                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 687                                 eal_usage(prgname);
 688                                 return -1;
 689                         }
 690                         coremask_ok = 1;
 691                         break;
 692                 /* force loading of external driver */
 693                 case 'd':
 694                         solib = malloc(sizeof(*solib));
 695                         if (solib == NULL) {
 696                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 697                                 return -1;
 698                         }
 699                         memset(solib, 0, sizeof(*solib));
 700                         strncpy(solib->name, optarg, PATH_MAX-1);
 701                         solib->name[PATH_MAX-1] = 0;
 702                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 703                         break;
 704                 /* size of memory */
 705                 case 'm':
 706                         internal_config.memory = atoi(optarg);
 707                         internal_config.memory *= 1024ULL;
 708                         internal_config.memory *= 1024ULL;
 709                         break;
 710                 /* force number of channels */
 711                 case 'n':
 712                         internal_config.force_nchannel = atoi(optarg);
 713                         if (internal_config.force_nchannel == 0 ||
 714                             internal_config.force_nchannel > 4) {
 715                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 716                                 eal_usage(prgname);
 717                                 return -1;
 718                         }
 719                         break;
 720                 /* force number of ranks */
 721                 case 'r':
 722                         internal_config.force_nrank = atoi(optarg);
 723                         if (internal_config.force_nrank == 0 ||
 724                             internal_config.force_nrank > 16) {
 725                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 726                                 eal_usage(prgname);
 727                                 return -1;
 728                         }
 729                         break;
 730                 case 'v':
 731                         /* since message is explicitly requested by user, we
 732                          * write message at highest log level so it can always be seen
 733                          * even if info or warning messages are disabled */
 734                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 735                         break;
 736
 737                 /* long options */
 738                 case 0:
 739                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 740                                 internal_config.no_hugetlbfs = 1;
 741                         }
 742                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 743                 #ifdef RTE_LIBRTE_XEN_DOM0
 744                                 internal_config.xen_dom0_support = 1;
 745                 #else
 746                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 747                                         "running on Dom0, please configure"
 748                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 749                                 return -1;
 750                 #endif
 751                         }
 752                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 753                                 internal_config.no_pci = 1;
 754                         }
 755                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 756                                 internal_config.no_hpet = 1;
 757                         }
 758                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 759                                 internal_config.vmware_tsc_map = 1;
 760                         }
 761                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 762                                 internal_config.no_shconf = 1;
 763                         }
 764                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 765                                 internal_config.hugepage_dir = optarg;
 766                         }
 767                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 768                                 internal_config.process_type = eal_parse_proc_type(optarg);
 769                         }
 770                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 771                                 internal_config.hugefile_prefix = optarg;
 772                         }
 773                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 774                                 if (eal_parse_socket_mem(optarg) < 0) {
 775                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 776                                                         OPT_SOCKET_MEM "\n");
 777                                         eal_usage(prgname);
 778                                         return -1;
 779                                 }
 780                         }
 781                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 782                                 eal_dev_whitelist_add_entry(optarg);
 783                         }
 784                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 785                                 if (eal_parse_syslog(optarg) < 0) {
 786                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 787                                                         OPT_SYSLOG "\n");
 788                                         eal_usage(prgname);
 789                                         return -1;
 790                                 }
 791                         }
 792                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 793                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 794                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 795                                                         OPT_BASE_VIRTADDR "\n");
 796                                         eal_usage(prgname);
 797                                         return -1;
 798                                 }
 799                         }
 800                         break;
 801
 802                 default:
 803                         eal_usage(prgname);
 804                         return -1;
 805                 }
 806         }
 807
 808         /* sanity checks */
 809         if (!coremask_ok) {
 810                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 811                 eal_usage(prgname);
 812                 return -1;
 813         }
 814         if (internal_config.process_type == RTE_PROC_AUTO){
 815                 internal_config.process_type = eal_proc_type_detect();
 816         }
 817         if (internal_config.process_type == RTE_PROC_INVALID){
 818                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 819                 eal_usage(prgname);
 820                 return -1;
 821         }
 822         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 823                         internal_config.force_nchannel == 0) {
 824                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 825                 eal_usage(prgname);
 826                 return -1;
 827         }
 828         if (index(internal_config.hugefile_prefix,'%') != NULL){
 829                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 830                 eal_usage(prgname);
 831                 return -1;
 832         }
 833         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 834                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 835                                 "at the same time\n");
 836                 eal_usage(prgname);
 837                 return -1;
 838         }
 839         /* --no-huge doesn't make sense with either -m or --socket-mem */
 840         if (internal_config.no_hugetlbfs &&
 841                         (internal_config.memory > 0 ||
 842                                         internal_config.force_sockets == 1)) {
 843                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 844                                 "together with --no-huge!\n");
 845                 eal_usage(prgname);
 846                 return -1;
 847         }
 848         /* --xen-dom0 doesn't make sense with --socket-mem */
 849         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 850                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 851                                         "together with --xen_dom0!\n");
 852                 eal_usage(prgname);
 853                 return -1;
 854         }
 855         /* if no blacklist, parse a whitelist */
 856         if (blacklist_index > 0) {
 857                 if (eal_dev_whitelist_exists()) {
 858                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 859                                         "[--use-device] options cannot be used at the same time\n");
 860                         eal_usage(prgname);
 861                         return -1;
 862                 }
 863                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 864         } else {
 865                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 866                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 867                         return -1;
 868                 }
 869         }
 870
 871         if (optind >= 0)
 872                 argv[optind-1] = prgname;
 873
 874         /* if no memory amounts were requested, this will result in 0 and
 875          * will be overriden later, right after eal_hugepage_info_init() */
 876         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 877                 internal_config.memory += internal_config.socket_mem[i];
 878
 879         ret = optind-1;
 880         optind = 0; /* reset getopt lib */
 881         return ret;
 882 }
 883
 884 static void
 885 eal_check_mem_on_local_socket(void)
 886 {
 887         const struct rte_memseg *ms;
 888         int i, socket_id;
 889
 890         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 891
 892         ms = rte_eal_get_physmem_layout();
 893
 894         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 895                 if (ms[i].socket_id == socket_id &&
 896                                 ms[i].len > 0)
 897                         return;
 898
 899         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 900                         "memory on local socket!\n");
 901 }
 902
 903 static int
 904 sync_func(__attribute__((unused)) void *arg)
 905 {
 906         return 0;
 907 }
 908
 909 inline static void
 910 rte_eal_mcfg_complete(void)
 911 {
 912         /* ALL shared mem_config related INIT DONE */
 913         if (rte_config.process_type == RTE_PROC_PRIMARY)
 914                 rte_config.mem_config->magic = RTE_MAGIC;
 915 }
 916
 917 /*
 918  * Request iopl priviledge for all RPL, returns 0 on success
 919  */
 920 static int
 921 rte_eal_iopl_init(void)
 922 {
 923         return iopl(HIGHEST_RPL);
 924 }
 925
 926 /* Launch threads, called at application init(). */
 927 int
 928 rte_eal_init(int argc, char **argv)
 929 {
 930         int i, fctret, ret;
 931         pthread_t thread_id;
 932         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 933         struct shared_driver *solib = NULL;
 934         const char *logid;
 935
 936         if (!rte_atomic32_test_and_set(&run_once))
 937                 return -1;
 938
 939         logid = strrchr(argv[0], '/');
 940         logid = strdup(logid ? logid + 1: argv[0]);
 941
 942         thread_id = pthread_self();
 943
 944         if (rte_eal_log_early_init() < 0)
 945                 rte_panic("Cannot init early logs\n");
 946
 947         if (rte_eal_cpu_init() < 0)
 948                 rte_panic("Cannot detect lcores\n");
 949
 950         fctret = eal_parse_args(argc, argv);
 951         if (fctret < 0)
 952                 exit(1);
 953
 954         if (internal_config.no_hugetlbfs == 0 &&
 955                         internal_config.process_type != RTE_PROC_SECONDARY &&
 956                         internal_config.xen_dom0_support == 0 &&
 957                         eal_hugepage_info_init() < 0)
 958                 rte_panic("Cannot get hugepage information\n");
 959
 960         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 961                 if (internal_config.no_hugetlbfs)
 962                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 963                 else
 964                         internal_config.memory = eal_get_hugepage_mem_size();
 965         }
 966
 967         if (internal_config.vmware_tsc_map == 1) {
 968 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 969                 rte_cycles_vmware_tsc_map = 1;
 970                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 971                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 972 #else
 973                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 974                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 975 #endif
 976         }
 977
 978         rte_srand(rte_rdtsc());
 979
 980         rte_config_init();
 981
 982         if (rte_eal_iopl_init() == 0)
 983                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 984
 985         if (rte_eal_pci_init() < 0)
 986                 rte_panic("Cannot init PCI\n");
 987
 988 #ifdef RTE_LIBRTE_IVSHMEM
 989         if (rte_eal_ivshmem_init() < 0)
 990                 rte_panic("Cannot init IVSHMEM\n");
 991 #endif
 992
 993         if (rte_eal_memory_init() < 0)
 994                 rte_panic("Cannot init memory\n");
 995
 996         /* the directories are locked during eal_hugepage_info_init */
 997         eal_hugedirs_unlock();
 998
 999         if (rte_eal_memzone_init() < 0)
1000                 rte_panic("Cannot init memzone\n");
1001
1002         if (rte_eal_tailqs_init() < 0)
1003                 rte_panic("Cannot init tail queues for objects\n");
1004
1005 #ifdef RTE_LIBRTE_IVSHMEM
1006         if (rte_eal_ivshmem_obj_init() < 0)
1007                 rte_panic("Cannot init IVSHMEM objects\n");
1008 #endif
1009
1010         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1011                 rte_panic("Cannot init logs\n");
1012
1013         if (rte_eal_alarm_init() < 0)
1014                 rte_panic("Cannot init interrupt-handling thread\n");
1015
1016         if (rte_eal_intr_init() < 0)
1017                 rte_panic("Cannot init interrupt-handling thread\n");
1018
1019         if (rte_eal_timer_init() < 0)
1020                 rte_panic("Cannot init HPET or TSC timers\n");
1021
1022         eal_check_mem_on_local_socket();
1023
1024         rte_eal_mcfg_complete();
1025
1026         if (rte_eal_non_pci_ethdev_init() < 0)
1027                 rte_panic("Cannot init non-PCI eth_devs\n");
1028
1029         TAILQ_FOREACH(solib, &solib_list, next) {
1030                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1031                 if ((solib->lib_handle == NULL) && (solib->name[0] != '/')) {
1032                         /* relative path: try again with "./" prefix */
1033                         char sopath[PATH_MAX];
1034                         snprintf(sopath, sizeof(sopath), "./%s", solib->name);
1035                         solib->lib_handle = dlopen(sopath, RTLD_NOW);
1036                 }
1037                 if (solib->lib_handle == NULL)
1038                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1039         }
1040
1041         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1042                 rte_config.master_lcore, (int)thread_id);
1043
1044         RTE_LCORE_FOREACH_SLAVE(i) {
1045
1046                 /*
1047                  * create communication pipes between master thread
1048                  * and children
1049                  */
1050                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1051                         rte_panic("Cannot create pipe\n");
1052                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1053                         rte_panic("Cannot create pipe\n");
1054
1055                 lcore_config[i].state = WAIT;
1056
1057                 /* create a thread for each lcore */
1058                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1059                                      eal_thread_loop, NULL);
1060                 if (ret != 0)
1061                         rte_panic("Cannot create thread\n");
1062         }
1063
1064         eal_thread_init_master(rte_config.master_lcore);
1065
1066         /*
1067          * Launch a dummy function on all slave lcores, so that master lcore
1068          * knows they are all ready when this function returns.
1069          */
1070         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1071         rte_eal_mp_wait_lcore();
1072
1073         return fctret;
1074 }
1075
1076 /* get core role */
1077 enum rte_lcore_role_t
1078 rte_eal_lcore_role(unsigned lcore_id)
1079 {
1080         return (rte_config.lcore_role[lcore_id]);
1081 }
1082
1083 enum rte_proc_type_t
1084 rte_eal_process_type(void)
1085 {
1086         return (rte_config.process_type);
1087 }
1088
1089 int rte_eal_has_hugepages(void)
1090 {
1091         return ! internal_config.no_hugetlbfs;
1092 }