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) 2012-2014 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_devargs.h>
  73 #include <rte_common.h>
  74 #include <rte_version.h>
  75 #include <rte_atomic.h>
  76 #include <malloc_heap.h>
  77 #include <rte_eth_ring.h>
  78
  79 #include "eal_private.h"
  80 #include "eal_thread.h"
  81 #include "eal_internal_cfg.h"
  82 #include "eal_filesystem.h"
  83 #include "eal_hugepages.h"
  84
  85 #define OPT_HUGE_DIR    "huge-dir"
  86 #define OPT_PROC_TYPE   "proc-type"
  87 #define OPT_NO_SHCONF   "no-shconf"
  88 #define OPT_NO_HPET     "no-hpet"
  89 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  90 #define OPT_NO_PCI      "no-pci"
  91 #define OPT_NO_HUGE     "no-huge"
  92 #define OPT_FILE_PREFIX "file-prefix"
  93 #define OPT_SOCKET_MEM  "socket-mem"
  94 #define OPT_USE_DEVICE  "use-device"
  95 #define OPT_SYSLOG      "syslog"
  96 #define OPT_BASE_VIRTADDR   "base-virtaddr"
  97 #define OPT_XEN_DOM0    "xen-dom0"
  98 #define OPT_CREATE_UIO_DEV "create-uio-dev"
  99
 100 #define RTE_EAL_BLACKLIST_SIZE  0x100
 101
 102 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 103
 104 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 105
 106 #define HIGHEST_RPL 3
 107
 108 #define BITS_PER_HEX 4
 109
 110 /* Allow the application to print its usage message too if set */
 111 static rte_usage_hook_t rte_application_usage_hook = NULL;
 112
 113 TAILQ_HEAD(shared_driver_list, shared_driver);
 114
 115 /* Definition for shared object drivers. */
 116 struct shared_driver {
 117         TAILQ_ENTRY(shared_driver) next;
 118
 119         char    name[PATH_MAX];
 120         void*   lib_handle;
 121 };
 122
 123 /* List of external loadable drivers */
 124 static struct shared_driver_list solib_list =
 125 TAILQ_HEAD_INITIALIZER(solib_list);
 126
 127 /* early configuration structure, when memory config is not mmapped */
 128 static struct rte_mem_config early_mem_config;
 129
 130 /* define fd variable here, because file needs to be kept open for the
 131  * duration of the program, as we hold a write lock on it in the primary proc */
 132 static int mem_cfg_fd = -1;
 133
 134 static struct flock wr_lock = {
 135                 .l_type = F_WRLCK,
 136                 .l_whence = SEEK_SET,
 137                 .l_start = offsetof(struct rte_mem_config, memseg),
 138                 .l_len = sizeof(early_mem_config.memseg),
 139 };
 140
 141 /* Address of global and public configuration */
 142 static struct rte_config rte_config = {
 143                 .mem_config = &early_mem_config,
 144 };
 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.\n"
 353                "               The argument format is <[domain:]bus:devid.func> to add\n"
 354                "               a PCI device to the white list or <driver><id>[;key=val;...]\n"
 355                "               to add a virtual device.\n"
 356                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 357                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 358                            "native RDTSC\n"
 359                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 360                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 361                "\nEAL options for DEBUG use only:\n"
 362                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 363                "  --"OPT_NO_PCI"   : disable pci\n"
 364                "  --"OPT_NO_HPET"  : disable hpet\n"
 365                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 366                "\n",
 367                prgname);
 368         /* Allow the application to print its usage message too if hook is set */
 369         if ( rte_application_usage_hook ) {
 370                 printf("===== Application Usage =====\n\n");
 371                 rte_application_usage_hook(prgname);
 372         }
 373 }
 374
 375 /* Set a per-application usage message */
 376 rte_usage_hook_t
 377 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 378 {
 379         rte_usage_hook_t        old_func;
 380
 381         /* Will be NULL on the first call to denote the last usage routine. */
 382         old_func                                        = rte_application_usage_hook;
 383         rte_application_usage_hook      = usage_func;
 384
 385         return old_func;
 386 }
 387
 388 /*
 389  * Parse the coremask given as argument (hexadecimal string) and fill
 390  * the global configuration (core role and core count) with the parsed
 391  * value.
 392  */
 393 static int xdigit2val(unsigned char c)
 394 {
 395         int val;
 396         if(isdigit(c))
 397                 val = c - '0';
 398         else if(isupper(c))
 399                 val = c - 'A' + 10;
 400         else
 401                 val = c - 'a' + 10;
 402         return val;
 403 }
 404 static int
 405 eal_parse_coremask(const char *coremask)
 406 {
 407         struct rte_config *cfg = rte_eal_get_configuration();
 408         int i, j, idx = 0 ;
 409         unsigned count = 0;
 410         char c;
 411         int val;
 412
 413         if (coremask == NULL)
 414                 return -1;
 415         /* Remove all blank characters ahead and after .
 416          * Remove 0x/0X if exists.
 417          */
 418         while (isblank(*coremask))
 419                 coremask++;
 420         if (coremask[0] == '0' && ((coremask[1] == 'x')
 421                 ||  (coremask[1] == 'X')) )
 422                 coremask += 2;
 423         i = strnlen(coremask, PATH_MAX);
 424         while ((i > 0) && isblank(coremask[i - 1]))
 425                 i--;
 426         if (i == 0)
 427                 return -1;
 428
 429         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 430                 c = coremask[i];
 431                 if (isxdigit(c) == 0) {
 432                         /* invalid characters */
 433                         return (-1);
 434                 }
 435                 val = xdigit2val(c);
 436                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 437                         if((1 << j) & val) {
 438                                 cfg->lcore_role[idx] = ROLE_RTE;
 439                                 if(count == 0)
 440                                         cfg->master_lcore = idx;
 441                                 count++;
 442                         } else  {
 443                                 cfg->lcore_role[idx] = ROLE_OFF;
 444                         }
 445                 }
 446         }
 447         for(; i >= 0; i--)
 448                 if(coremask[i] != '0')
 449                         return -1;
 450         for(; idx < RTE_MAX_LCORE; idx++)
 451                 cfg->lcore_role[idx] = ROLE_OFF;
 452         if(count == 0)
 453                 return -1;
 454         /* Update the count of enabled logical cores of the EAL configuration */
 455         cfg->lcore_count = count;
 456         return 0;
 457 }
 458
 459 static int
 460 eal_parse_syslog(const char *facility)
 461 {
 462         int i;
 463         static struct {
 464                 const char *name;
 465                 int value;
 466         } map[] = {
 467                 { "auth", LOG_AUTH },
 468                 { "cron", LOG_CRON },
 469                 { "daemon", LOG_DAEMON },
 470                 { "ftp", LOG_FTP },
 471                 { "kern", LOG_KERN },
 472                 { "lpr", LOG_LPR },
 473                 { "mail", LOG_MAIL },
 474                 { "news", LOG_NEWS },
 475                 { "syslog", LOG_SYSLOG },
 476                 { "user", LOG_USER },
 477                 { "uucp", LOG_UUCP },
 478                 { "local0", LOG_LOCAL0 },
 479                 { "local1", LOG_LOCAL1 },
 480                 { "local2", LOG_LOCAL2 },
 481                 { "local3", LOG_LOCAL3 },
 482                 { "local4", LOG_LOCAL4 },
 483                 { "local5", LOG_LOCAL5 },
 484                 { "local6", LOG_LOCAL6 },
 485                 { "local7", LOG_LOCAL7 },
 486                 { NULL, 0 }
 487         };
 488
 489         for (i = 0; map[i].name; i++) {
 490                 if (!strcmp(facility, map[i].name)) {
 491                         internal_config.syslog_facility = map[i].value;
 492                         return 0;
 493                 }
 494         }
 495         return -1;
 496 }
 497
 498 static int
 499 eal_parse_socket_mem(char *socket_mem)
 500 {
 501         char * arg[RTE_MAX_NUMA_NODES];
 502         char *end;
 503         int arg_num, i, len;
 504         uint64_t total_mem = 0;
 505
 506         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 507         if (len == SOCKET_MEM_STRLEN) {
 508                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 509                 return -1;
 510         }
 511
 512         /* all other error cases will be caught later */
 513         if (!isdigit(socket_mem[len-1]))
 514                 return -1;
 515
 516         /* split the optarg into separate socket values */
 517         arg_num = rte_strsplit(socket_mem, len,
 518                         arg, RTE_MAX_NUMA_NODES, ',');
 519
 520         /* if split failed, or 0 arguments */
 521         if (arg_num <= 0)
 522                 return -1;
 523
 524         internal_config.force_sockets = 1;
 525
 526         /* parse each defined socket option */
 527         errno = 0;
 528         for (i = 0; i < arg_num; i++) {
 529                 end = NULL;
 530                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 531
 532                 /* check for invalid input */
 533                 if ((errno != 0)  ||
 534                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 535                         return -1;
 536                 internal_config.socket_mem[i] *= 1024ULL;
 537                 internal_config.socket_mem[i] *= 1024ULL;
 538                 total_mem += internal_config.socket_mem[i];
 539         }
 540
 541         /* check if we have a positive amount of total memory */
 542         if (total_mem == 0)
 543                 return -1;
 544
 545         return 0;
 546 }
 547
 548 static int
 549 eal_parse_base_virtaddr(const char *arg)
 550 {
 551         char *end;
 552         uint64_t addr;
 553
 554         addr = strtoull(arg, &end, 16);
 555
 556         /* check for errors */
 557         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 558                 return -1;
 559
 560         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 561 #ifndef RTE_ARCH_X86_64
 562         if (addr >= UINTPTR_MAX)
 563                 return -1;
 564 #endif
 565
 566         /* align the addr on 2M boundary */
 567         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 568                                                            RTE_PGSIZE_2M);
 569
 570         return 0;
 571 }
 572
 573 static inline size_t
 574 eal_get_hugepage_mem_size(void)
 575 {
 576         uint64_t size = 0;
 577         unsigned i, j;
 578
 579         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 580                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 581                 if (hpi->hugedir != NULL) {
 582                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 583                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 584                         }
 585                 }
 586         }
 587
 588         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 589 }
 590
 591 static enum rte_proc_type_t
 592 eal_parse_proc_type(const char *arg)
 593 {
 594         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 595                 return RTE_PROC_PRIMARY;
 596         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 597                 return RTE_PROC_SECONDARY;
 598         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 599                 return RTE_PROC_AUTO;
 600
 601         return RTE_PROC_INVALID;
 602 }
 603
 604 static int
 605 eal_parse_use_device(const char *optarg)
 606 {
 607         struct rte_pci_addr addr;
 608         char *dup, *sep;
 609
 610         dup = strdup(optarg);
 611         if (dup == NULL)
 612                 return -1;
 613
 614         /* remove arguments in 'dup' string */
 615         sep = strchr(dup, ';');
 616         if (sep != NULL)
 617                 *sep = '\0';
 618
 619         /* if argument is a PCI address, it's a whitelisted device */
 620         if (eal_parse_pci_DomBDF(dup, &addr) == 0 ||
 621                 eal_parse_pci_BDF(dup, &addr) == 0) {
 622                 rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI, optarg);
 623         } else {
 624                 rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL, optarg);
 625         }
 626
 627         free(dup);
 628         return 0;
 629 }
 630
 631 /* Parse the argument given in the command line of the application */
 632 static int
 633 eal_parse_args(int argc, char **argv)
 634 {
 635         int opt, ret, i;
 636         char **argvopt;
 637         int option_index;
 638         int coremask_ok = 0;
 639         char *prgname = argv[0];
 640         static struct option lgopts[] = {
 641                 {OPT_NO_HUGE, 0, 0, 0},
 642                 {OPT_NO_PCI, 0, 0, 0},
 643                 {OPT_NO_HPET, 0, 0, 0},
 644                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 645                 {OPT_HUGE_DIR, 1, 0, 0},
 646                 {OPT_NO_SHCONF, 0, 0, 0},
 647                 {OPT_PROC_TYPE, 1, 0, 0},
 648                 {OPT_FILE_PREFIX, 1, 0, 0},
 649                 {OPT_SOCKET_MEM, 1, 0, 0},
 650                 {OPT_USE_DEVICE, 1, 0, 0},
 651                 {OPT_SYSLOG, 1, NULL, 0},
 652                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 653                 {OPT_XEN_DOM0, 0, 0, 0},
 654                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 655                 {0, 0, 0, 0}
 656         };
 657         struct shared_driver *solib;
 658
 659         argvopt = argv;
 660
 661         internal_config.memory = 0;
 662         internal_config.force_nrank = 0;
 663         internal_config.force_nchannel = 0;
 664         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 665         internal_config.hugepage_dir = NULL;
 666         internal_config.force_sockets = 0;
 667         internal_config.syslog_facility = LOG_DAEMON;
 668         internal_config.xen_dom0_support = 0;
 669 #ifdef RTE_LIBEAL_USE_HPET
 670         internal_config.no_hpet = 0;
 671 #else
 672         internal_config.no_hpet = 1;
 673 #endif
 674         /* zero out the NUMA config */
 675         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 676                 internal_config.socket_mem[i] = 0;
 677
 678         /* zero out hugedir descriptors */
 679         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 680                 internal_config.hugepage_info[i].lock_descriptor = -1;
 681
 682         internal_config.vmware_tsc_map = 0;
 683         internal_config.base_virtaddr = 0;
 684
 685         while ((opt = getopt_long(argc, argvopt, "b:c:d:m:n:r:v",
 686                                   lgopts, &option_index)) != EOF) {
 687
 688                 switch (opt) {
 689                 /* blacklist */
 690                 case 'b':
 691                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 692                                         optarg) < 0) {
 693                                 eal_usage(prgname);
 694                                 return (-1);
 695                         }
 696                         break;
 697                 /* coremask */
 698                 case 'c':
 699                         if (eal_parse_coremask(optarg) < 0) {
 700                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 701                                 eal_usage(prgname);
 702                                 return -1;
 703                         }
 704                         coremask_ok = 1;
 705                         break;
 706                 /* force loading of external driver */
 707                 case 'd':
 708                         solib = malloc(sizeof(*solib));
 709                         if (solib == NULL) {
 710                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 711                                 return -1;
 712                         }
 713                         memset(solib, 0, sizeof(*solib));
 714                         strncpy(solib->name, optarg, PATH_MAX-1);
 715                         solib->name[PATH_MAX-1] = 0;
 716                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 717                         break;
 718                 /* size of memory */
 719                 case 'm':
 720                         internal_config.memory = atoi(optarg);
 721                         internal_config.memory *= 1024ULL;
 722                         internal_config.memory *= 1024ULL;
 723                         break;
 724                 /* force number of channels */
 725                 case 'n':
 726                         internal_config.force_nchannel = atoi(optarg);
 727                         if (internal_config.force_nchannel == 0 ||
 728                             internal_config.force_nchannel > 4) {
 729                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 730                                 eal_usage(prgname);
 731                                 return -1;
 732                         }
 733                         break;
 734                 /* force number of ranks */
 735                 case 'r':
 736                         internal_config.force_nrank = atoi(optarg);
 737                         if (internal_config.force_nrank == 0 ||
 738                             internal_config.force_nrank > 16) {
 739                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 740                                 eal_usage(prgname);
 741                                 return -1;
 742                         }
 743                         break;
 744                 case 'v':
 745                         /* since message is explicitly requested by user, we
 746                          * write message at highest log level so it can always be seen
 747                          * even if info or warning messages are disabled */
 748                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 749                         break;
 750
 751                 /* long options */
 752                 case 0:
 753                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 754                                 internal_config.no_hugetlbfs = 1;
 755                         }
 756                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 757                 #ifdef RTE_LIBRTE_XEN_DOM0
 758                                 internal_config.xen_dom0_support = 1;
 759                 #else
 760                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 761                                         "running on Dom0, please configure"
 762                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 763                                 return -1;
 764                 #endif
 765                         }
 766                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 767                                 internal_config.no_pci = 1;
 768                         }
 769                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 770                                 internal_config.no_hpet = 1;
 771                         }
 772                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 773                                 internal_config.vmware_tsc_map = 1;
 774                         }
 775                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 776                                 internal_config.no_shconf = 1;
 777                         }
 778                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 779                                 internal_config.hugepage_dir = optarg;
 780                         }
 781                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 782                                 internal_config.process_type = eal_parse_proc_type(optarg);
 783                         }
 784                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 785                                 internal_config.hugefile_prefix = optarg;
 786                         }
 787                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 788                                 if (eal_parse_socket_mem(optarg) < 0) {
 789                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 790                                                         OPT_SOCKET_MEM "\n");
 791                                         eal_usage(prgname);
 792                                         return -1;
 793                                 }
 794                         }
 795                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 796                                 if (eal_parse_use_device(optarg) < 0) {
 797                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 798                                                         OPT_USE_DEVICE "\n");
 799                                         eal_usage(prgname);
 800                                         return -1;
 801                                 }
 802                         }
 803                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 804                                 if (eal_parse_syslog(optarg) < 0) {
 805                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 806                                                         OPT_SYSLOG "\n");
 807                                         eal_usage(prgname);
 808                                         return -1;
 809                                 }
 810                         }
 811                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 812                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 813                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 814                                                         OPT_BASE_VIRTADDR "\n");
 815                                         eal_usage(prgname);
 816                                         return -1;
 817                                 }
 818                         }
 819                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 820                                 internal_config.create_uio_dev = 1;
 821                         }
 822                         break;
 823
 824                 default:
 825                         eal_usage(prgname);
 826                         return -1;
 827                 }
 828         }
 829
 830         /* sanity checks */
 831         if (!coremask_ok) {
 832                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 833                 eal_usage(prgname);
 834                 return -1;
 835         }
 836         if (internal_config.process_type == RTE_PROC_AUTO){
 837                 internal_config.process_type = eal_proc_type_detect();
 838         }
 839         if (internal_config.process_type == RTE_PROC_INVALID){
 840                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 841                 eal_usage(prgname);
 842                 return -1;
 843         }
 844         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 845                         internal_config.force_nchannel == 0) {
 846                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 847                 eal_usage(prgname);
 848                 return -1;
 849         }
 850         if (index(internal_config.hugefile_prefix,'%') != NULL){
 851                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 852                 eal_usage(prgname);
 853                 return -1;
 854         }
 855         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 856                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 857                                 "at the same time\n");
 858                 eal_usage(prgname);
 859                 return -1;
 860         }
 861         /* --no-huge doesn't make sense with either -m or --socket-mem */
 862         if (internal_config.no_hugetlbfs &&
 863                         (internal_config.memory > 0 ||
 864                                         internal_config.force_sockets == 1)) {
 865                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 866                                 "together with --no-huge!\n");
 867                 eal_usage(prgname);
 868                 return -1;
 869         }
 870         /* --xen-dom0 doesn't make sense with --socket-mem */
 871         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 872                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 873                                         "together with --xen_dom0!\n");
 874                 eal_usage(prgname);
 875                 return -1;
 876         }
 877
 878         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 879                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 880                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 881                         "[--use-device] options cannot be used at the same time\n");
 882                 eal_usage(prgname);
 883                 return -1;
 884         }
 885
 886         if (optind >= 0)
 887                 argv[optind-1] = prgname;
 888
 889         /* if no memory amounts were requested, this will result in 0 and
 890          * will be overriden later, right after eal_hugepage_info_init() */
 891         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 892                 internal_config.memory += internal_config.socket_mem[i];
 893
 894         ret = optind-1;
 895         optind = 0; /* reset getopt lib */
 896         return ret;
 897 }
 898
 899 static void
 900 eal_check_mem_on_local_socket(void)
 901 {
 902         const struct rte_memseg *ms;
 903         int i, socket_id;
 904
 905         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 906
 907         ms = rte_eal_get_physmem_layout();
 908
 909         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 910                 if (ms[i].socket_id == socket_id &&
 911                                 ms[i].len > 0)
 912                         return;
 913
 914         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 915                         "memory on local socket!\n");
 916 }
 917
 918 static int
 919 sync_func(__attribute__((unused)) void *arg)
 920 {
 921         return 0;
 922 }
 923
 924 inline static void
 925 rte_eal_mcfg_complete(void)
 926 {
 927         /* ALL shared mem_config related INIT DONE */
 928         if (rte_config.process_type == RTE_PROC_PRIMARY)
 929                 rte_config.mem_config->magic = RTE_MAGIC;
 930 }
 931
 932 /*
 933  * Request iopl priviledge for all RPL, returns 0 on success
 934  */
 935 static int
 936 rte_eal_iopl_init(void)
 937 {
 938         return iopl(HIGHEST_RPL);
 939 }
 940
 941 /* Launch threads, called at application init(). */
 942 int
 943 rte_eal_init(int argc, char **argv)
 944 {
 945         int i, fctret, ret;
 946         pthread_t thread_id;
 947         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 948         struct shared_driver *solib = NULL;
 949         const char *logid;
 950
 951         if (!rte_atomic32_test_and_set(&run_once))
 952                 return -1;
 953
 954         logid = strrchr(argv[0], '/');
 955         logid = strdup(logid ? logid + 1: argv[0]);
 956
 957         thread_id = pthread_self();
 958
 959         if (rte_eal_log_early_init() < 0)
 960                 rte_panic("Cannot init early logs\n");
 961
 962         if (rte_eal_cpu_init() < 0)
 963                 rte_panic("Cannot detect lcores\n");
 964
 965         fctret = eal_parse_args(argc, argv);
 966         if (fctret < 0)
 967                 exit(1);
 968
 969         if (internal_config.no_hugetlbfs == 0 &&
 970                         internal_config.process_type != RTE_PROC_SECONDARY &&
 971                         internal_config.xen_dom0_support == 0 &&
 972                         eal_hugepage_info_init() < 0)
 973                 rte_panic("Cannot get hugepage information\n");
 974
 975         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 976                 if (internal_config.no_hugetlbfs)
 977                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 978                 else
 979                         internal_config.memory = eal_get_hugepage_mem_size();
 980         }
 981
 982         if (internal_config.vmware_tsc_map == 1) {
 983 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 984                 rte_cycles_vmware_tsc_map = 1;
 985                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 986                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 987 #else
 988                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 989                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 990 #endif
 991         }
 992
 993         rte_srand(rte_rdtsc());
 994
 995         rte_config_init();
 996
 997         if (rte_eal_iopl_init() == 0)
 998                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 999
1000         if (rte_eal_pci_init() < 0)
1001                 rte_panic("Cannot init PCI\n");
1002
1003 #ifdef RTE_LIBRTE_IVSHMEM
1004         if (rte_eal_ivshmem_init() < 0)
1005                 rte_panic("Cannot init IVSHMEM\n");
1006 #endif
1007
1008         if (rte_eal_memory_init() < 0)
1009                 rte_panic("Cannot init memory\n");
1010
1011         /* the directories are locked during eal_hugepage_info_init */
1012         eal_hugedirs_unlock();
1013
1014         if (rte_eal_memzone_init() < 0)
1015                 rte_panic("Cannot init memzone\n");
1016
1017         if (rte_eal_tailqs_init() < 0)
1018                 rte_panic("Cannot init tail queues for objects\n");
1019
1020 #ifdef RTE_LIBRTE_IVSHMEM
1021         if (rte_eal_ivshmem_obj_init() < 0)
1022                 rte_panic("Cannot init IVSHMEM objects\n");
1023 #endif
1024
1025         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1026                 rte_panic("Cannot init logs\n");
1027
1028         if (rte_eal_alarm_init() < 0)
1029                 rte_panic("Cannot init interrupt-handling thread\n");
1030
1031         if (rte_eal_intr_init() < 0)
1032                 rte_panic("Cannot init interrupt-handling thread\n");
1033
1034         if (rte_eal_timer_init() < 0)
1035                 rte_panic("Cannot init HPET or TSC timers\n");
1036
1037         eal_check_mem_on_local_socket();
1038
1039         rte_eal_mcfg_complete();
1040
1041         if (rte_eal_non_pci_ethdev_init() < 0)
1042                 rte_panic("Cannot init non-PCI eth_devs\n");
1043
1044         TAILQ_FOREACH(solib, &solib_list, next) {
1045                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1046                 if ((solib->lib_handle == NULL) && (solib->name[0] != '/')) {
1047                         /* relative path: try again with "./" prefix */
1048                         char sopath[PATH_MAX];
1049                         snprintf(sopath, sizeof(sopath), "./%s", solib->name);
1050                         solib->lib_handle = dlopen(sopath, RTLD_NOW);
1051                 }
1052                 if (solib->lib_handle == NULL)
1053                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1054         }
1055
1056         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1057                 rte_config.master_lcore, (int)thread_id);
1058
1059         RTE_LCORE_FOREACH_SLAVE(i) {
1060
1061                 /*
1062                  * create communication pipes between master thread
1063                  * and children
1064                  */
1065                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1066                         rte_panic("Cannot create pipe\n");
1067                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1068                         rte_panic("Cannot create pipe\n");
1069
1070                 lcore_config[i].state = WAIT;
1071
1072                 /* create a thread for each lcore */
1073                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1074                                      eal_thread_loop, NULL);
1075                 if (ret != 0)
1076                         rte_panic("Cannot create thread\n");
1077         }
1078
1079         eal_thread_init_master(rte_config.master_lcore);
1080
1081         /*
1082          * Launch a dummy function on all slave lcores, so that master lcore
1083          * knows they are all ready when this function returns.
1084          */
1085         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1086         rte_eal_mp_wait_lcore();
1087
1088         return fctret;
1089 }
1090
1091 /* get core role */
1092 enum rte_lcore_role_t
1093 rte_eal_lcore_role(unsigned lcore_id)
1094 {
1095         return (rte_config.lcore_role[lcore_id]);
1096 }
1097
1098 enum rte_proc_type_t
1099 rte_eal_process_type(void)
1100 {
1101         return (rte_config.process_type);
1102 }
1103
1104 int rte_eal_has_hugepages(void)
1105 {
1106         return ! internal_config.no_hugetlbfs;
1107 }