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 #define OPT_CREATE_UIO_DEV "create-uio-dev"
  98
  99 #define RTE_EAL_BLACKLIST_SIZE  0x100
 100
 101 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 102
 103 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 104
 105 #define HIGHEST_RPL 3
 106
 107 #define BITS_PER_HEX 4
 108
 109 /* Allow the application to print its usage message too if set */
 110 static rte_usage_hook_t rte_application_usage_hook = NULL;
 111
 112 TAILQ_HEAD(shared_driver_list, shared_driver);
 113
 114 /* Definition for shared object drivers. */
 115 struct shared_driver {
 116         TAILQ_ENTRY(shared_driver) next;
 117
 118         char    name[PATH_MAX];
 119         void*   lib_handle;
 120 };
 121
 122 /* List of external loadable drivers */
 123 static struct shared_driver_list solib_list =
 124 TAILQ_HEAD_INITIALIZER(solib_list);
 125
 126 /* early configuration structure, when memory config is not mmapped */
 127 static struct rte_mem_config early_mem_config;
 128
 129 /* define fd variable here, because file needs to be kept open for the
 130  * duration of the program, as we hold a write lock on it in the primary proc */
 131 static int mem_cfg_fd = -1;
 132
 133 static struct flock wr_lock = {
 134                 .l_type = F_WRLCK,
 135                 .l_whence = SEEK_SET,
 136                 .l_start = offsetof(struct rte_mem_config, memseg),
 137                 .l_len = sizeof(early_mem_config.memseg),
 138 };
 139
 140 /* Address of global and public configuration */
 141 static struct rte_config rte_config = {
 142                 .mem_config = &early_mem_config,
 143 };
 144
 145 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 146
 147 /* internal configuration (per-core) */
 148 struct lcore_config lcore_config[RTE_MAX_LCORE];
 149
 150 /* internal configuration */
 151 struct internal_config internal_config;
 152
 153 /* used by rte_rdtsc() */
 154 int rte_cycles_vmware_tsc_map;
 155
 156 /* Return a pointer to the configuration structure */
 157 struct rte_config *
 158 rte_eal_get_configuration(void)
 159 {
 160         return &rte_config;
 161 }
 162
 163 /* parse a sysfs (or other) file containing one integer value */
 164 int
 165 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 166 {
 167         FILE *f;
 168         char buf[BUFSIZ];
 169         char *end = NULL;
 170
 171         if ((f = fopen(filename, "r")) == NULL) {
 172                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 173                         __func__, filename);
 174                 return -1;
 175         }
 176
 177         if (fgets(buf, sizeof(buf), f) == NULL) {
 178                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 179                         __func__, filename);
 180                 fclose(f);
 181                 return -1;
 182         }
 183         *val = strtoul(buf, &end, 0);
 184         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 185                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 186                                 __func__, filename);
 187                 fclose(f);
 188                 return -1;
 189         }
 190         fclose(f);
 191         return 0;
 192 }
 193
 194
 195 /* create memory configuration in shared/mmap memory. Take out
 196  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 197  * This means we never close the file while running (auto-close on exit).
 198  * We also don't lock the whole file, so that in future we can use read-locks
 199  * on other parts, e.g. memzones, to detect if there are running secondary
 200  * processes. */
 201 static void
 202 rte_eal_config_create(void)
 203 {
 204         void *rte_mem_cfg_addr;
 205         int retval;
 206
 207         const char *pathname = eal_runtime_config_path();
 208
 209         if (internal_config.no_shconf)
 210                 return;
 211
 212         if (mem_cfg_fd < 0){
 213                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 214                 if (mem_cfg_fd < 0)
 215                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 216         }
 217
 218         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 219         if (retval < 0){
 220                 close(mem_cfg_fd);
 221                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 222         }
 223
 224         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 225         if (retval < 0){
 226                 close(mem_cfg_fd);
 227                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 228                                 "process running?\n", pathname);
 229         }
 230
 231         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 232                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 233
 234         if (rte_mem_cfg_addr == MAP_FAILED){
 235                 rte_panic("Cannot mmap memory for rte_config\n");
 236         }
 237         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 238         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 239 }
 240
 241 /* attach to an existing shared memory config */
 242 static void
 243 rte_eal_config_attach(void)
 244 {
 245         void *rte_mem_cfg_addr;
 246         const char *pathname = eal_runtime_config_path();
 247
 248         if (internal_config.no_shconf)
 249                 return;
 250
 251         if (mem_cfg_fd < 0){
 252                 mem_cfg_fd = open(pathname, O_RDWR);
 253                 if (mem_cfg_fd < 0)
 254                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 255         }
 256
 257         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 258                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 259         close(mem_cfg_fd);
 260         if (rte_mem_cfg_addr == MAP_FAILED)
 261                 rte_panic("Cannot mmap memory for rte_config\n");
 262
 263         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 264 }
 265
 266 /* Detect if we are a primary or a secondary process */
 267 static enum rte_proc_type_t
 268 eal_proc_type_detect(void)
 269 {
 270         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 271         const char *pathname = eal_runtime_config_path();
 272
 273         /* if we can open the file but not get a write-lock we are a secondary
 274          * process. NOTE: if we get a file handle back, we keep that open
 275          * and don't close it to prevent a race condition between multiple opens */
 276         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 277                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 278                 ptype = RTE_PROC_SECONDARY;
 279
 280         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 281                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 282
 283         return ptype;
 284 }
 285
 286 /* Sets up rte_config structure with the pointer to shared memory config.*/
 287 static void
 288 rte_config_init(void)
 289 {
 290         /* set the magic in configuration structure */
 291         rte_config.magic = RTE_MAGIC;
 292         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 293                         eal_proc_type_detect() : /* for auto, detect the type */
 294                         internal_config.process_type; /* otherwise use what's already set */
 295
 296         switch (rte_config.process_type){
 297         case RTE_PROC_PRIMARY:
 298                 rte_eal_config_create();
 299                 break;
 300         case RTE_PROC_SECONDARY:
 301                 rte_eal_config_attach();
 302                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 303                 break;
 304         case RTE_PROC_AUTO:
 305         case RTE_PROC_INVALID:
 306                 rte_panic("Invalid process type\n");
 307         }
 308 }
 309
 310 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 311 static void
 312 eal_hugedirs_unlock(void)
 313 {
 314         int i;
 315
 316         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 317         {
 318                 /* skip uninitialized */
 319                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 320                         continue;
 321                 /* unlock hugepage file */
 322                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 323                 close(internal_config.hugepage_info[i].lock_descriptor);
 324                 /* reset the field */
 325                 internal_config.hugepage_info[i].lock_descriptor = -1;
 326         }
 327 }
 328
 329 /* display usage */
 330 static void
 331 eal_usage(const char *prgname)
 332 {
 333         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 334                "[--proc-type primary|secondary|auto] \n\n"
 335                "EAL options:\n"
 336                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 337                "  -n NUM       : Number of memory channels\n"
 338                    "  -v           : Display version information on startup\n"
 339                "  -d LIB.so    : add driver (can be used multiple times)\n"
 340                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 341            "PCI device\n"
 342                "                 (multiple -b options are allowed)\n"
 343                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 344                "  -r NUM       : force number of memory ranks (don't detect)\n"
 345                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 346                            "without hugetlbfs\n"
 347                "  --"OPT_SYSLOG"     : set syslog facility\n"
 348                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 349                    "                 sockets (use comma separated values)\n"
 350                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 351                "  --"OPT_PROC_TYPE"  : type of this process\n"
 352                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 353                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 354                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 355                "               [NOTE: Cannot be used with -b option]\n"
 356                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 357                            "native RDTSC\n"
 358                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 359                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 360                "\nEAL options for DEBUG use only:\n"
 361                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 362                "  --"OPT_NO_PCI"   : disable pci\n"
 363                "  --"OPT_NO_HPET"  : disable hpet\n"
 364                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 365                "\n",
 366                prgname);
 367         /* Allow the application to print its usage message too if hook is set */
 368         if ( rte_application_usage_hook ) {
 369                 printf("===== Application Usage =====\n\n");
 370                 rte_application_usage_hook(prgname);
 371         }
 372 }
 373
 374 /* Set a per-application usage message */
 375 rte_usage_hook_t
 376 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 377 {
 378         rte_usage_hook_t        old_func;
 379
 380         /* Will be NULL on the first call to denote the last usage routine. */
 381         old_func                                        = rte_application_usage_hook;
 382         rte_application_usage_hook      = usage_func;
 383
 384         return old_func;
 385 }
 386
 387 /*
 388  * Parse the coremask given as argument (hexadecimal string) and fill
 389  * the global configuration (core role and core count) with the parsed
 390  * value.
 391  */
 392 static int xdigit2val(unsigned char c)
 393 {
 394         int val;
 395         if(isdigit(c))
 396                 val = c - '0';
 397         else if(isupper(c))
 398                 val = c - 'A' + 10;
 399         else
 400                 val = c - 'a' + 10;
 401         return val;
 402 }
 403 static int
 404 eal_parse_coremask(const char *coremask)
 405 {
 406         struct rte_config *cfg = rte_eal_get_configuration();
 407         int i, j, idx = 0 ;
 408         unsigned count = 0;
 409         char c;
 410         int val;
 411
 412         if (coremask == NULL)
 413                 return -1;
 414         /* Remove all blank characters ahead and after .
 415          * Remove 0x/0X if exists.
 416          */
 417         while (isblank(*coremask))
 418                 coremask++;
 419         if (coremask[0] == '0' && ((coremask[1] == 'x')
 420                 ||  (coremask[1] == 'X')) )
 421                 coremask += 2;
 422         i = strnlen(coremask, PATH_MAX);
 423         while ((i > 0) && isblank(coremask[i - 1]))
 424                 i--;
 425         if (i == 0)
 426                 return -1;
 427
 428         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 429                 c = coremask[i];
 430                 if (isxdigit(c) == 0) {
 431                         /* invalid characters */
 432                         return (-1);
 433                 }
 434                 val = xdigit2val(c);
 435                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 436                         if((1 << j) & val) {
 437                                 cfg->lcore_role[idx] = ROLE_RTE;
 438                                 if(count == 0)
 439                                         cfg->master_lcore = idx;
 440                                 count++;
 441                         } else  {
 442                                 cfg->lcore_role[idx] = ROLE_OFF;
 443                         }
 444                 }
 445         }
 446         for(; i >= 0; i--)
 447                 if(coremask[i] != '0')
 448                         return -1;
 449         for(; idx < RTE_MAX_LCORE; idx++)
 450                 cfg->lcore_role[idx] = ROLE_OFF;
 451         if(count == 0)
 452                 return -1;
 453         /* Update the count of enabled logical cores of the EAL configuration */
 454         cfg->lcore_count = count;
 455         return 0;
 456 }
 457
 458 static int
 459 eal_parse_syslog(const char *facility)
 460 {
 461         int i;
 462         static struct {
 463                 const char *name;
 464                 int value;
 465         } map[] = {
 466                 { "auth", LOG_AUTH },
 467                 { "cron", LOG_CRON },
 468                 { "daemon", LOG_DAEMON },
 469                 { "ftp", LOG_FTP },
 470                 { "kern", LOG_KERN },
 471                 { "lpr", LOG_LPR },
 472                 { "mail", LOG_MAIL },
 473                 { "news", LOG_NEWS },
 474                 { "syslog", LOG_SYSLOG },
 475                 { "user", LOG_USER },
 476                 { "uucp", LOG_UUCP },
 477                 { "local0", LOG_LOCAL0 },
 478                 { "local1", LOG_LOCAL1 },
 479                 { "local2", LOG_LOCAL2 },
 480                 { "local3", LOG_LOCAL3 },
 481                 { "local4", LOG_LOCAL4 },
 482                 { "local5", LOG_LOCAL5 },
 483                 { "local6", LOG_LOCAL6 },
 484                 { "local7", LOG_LOCAL7 },
 485                 { NULL, 0 }
 486         };
 487
 488         for (i = 0; map[i].name; i++) {
 489                 if (!strcmp(facility, map[i].name)) {
 490                         internal_config.syslog_facility = map[i].value;
 491                         return 0;
 492                 }
 493         }
 494         return -1;
 495 }
 496
 497 static int
 498 eal_parse_socket_mem(char *socket_mem)
 499 {
 500         char * arg[RTE_MAX_NUMA_NODES];
 501         char *end;
 502         int arg_num, i, len;
 503         uint64_t total_mem = 0;
 504
 505         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 506         if (len == SOCKET_MEM_STRLEN) {
 507                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 508                 return -1;
 509         }
 510
 511         /* all other error cases will be caught later */
 512         if (!isdigit(socket_mem[len-1]))
 513                 return -1;
 514
 515         /* split the optarg into separate socket values */
 516         arg_num = rte_strsplit(socket_mem, len,
 517                         arg, RTE_MAX_NUMA_NODES, ',');
 518
 519         /* if split failed, or 0 arguments */
 520         if (arg_num <= 0)
 521                 return -1;
 522
 523         internal_config.force_sockets = 1;
 524
 525         /* parse each defined socket option */
 526         errno = 0;
 527         for (i = 0; i < arg_num; i++) {
 528                 end = NULL;
 529                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 530
 531                 /* check for invalid input */
 532                 if ((errno != 0)  ||
 533                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 534                         return -1;
 535                 internal_config.socket_mem[i] *= 1024ULL;
 536                 internal_config.socket_mem[i] *= 1024ULL;
 537                 total_mem += internal_config.socket_mem[i];
 538         }
 539
 540         /* check if we have a positive amount of total memory */
 541         if (total_mem == 0)
 542                 return -1;
 543
 544         return 0;
 545 }
 546
 547 static int
 548 eal_parse_base_virtaddr(const char *arg)
 549 {
 550         char *end;
 551         uint64_t addr;
 552
 553         addr = strtoull(arg, &end, 16);
 554
 555         /* check for errors */
 556         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 557                 return -1;
 558
 559         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 560 #ifndef RTE_ARCH_X86_64
 561         if (addr >= UINTPTR_MAX)
 562                 return -1;
 563 #endif
 564
 565         /* align the addr on 2M boundary */
 566         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 567                                                            RTE_PGSIZE_2M);
 568
 569         return 0;
 570 }
 571
 572 static inline size_t
 573 eal_get_hugepage_mem_size(void)
 574 {
 575         uint64_t size = 0;
 576         unsigned i, j;
 577
 578         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 579                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 580                 if (hpi->hugedir != NULL) {
 581                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 582                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 583                         }
 584                 }
 585         }
 586
 587         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 588 }
 589
 590 static enum rte_proc_type_t
 591 eal_parse_proc_type(const char *arg)
 592 {
 593         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 594                 return RTE_PROC_PRIMARY;
 595         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 596                 return RTE_PROC_SECONDARY;
 597         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 598                 return RTE_PROC_AUTO;
 599
 600         return RTE_PROC_INVALID;
 601 }
 602
 603 static ssize_t
 604 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 605 {
 606         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 607                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 608                 return (-EINVAL);
 609         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 610                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 611                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 612                 return (-EINVAL);
 613         }
 614
 615         idx += 1;
 616         return (idx);
 617 }
 618
 619 /* Parse the argument given in the command line of the application */
 620 static int
 621 eal_parse_args(int argc, char **argv)
 622 {
 623         int opt, ret, i;
 624         char **argvopt;
 625         int option_index;
 626         int coremask_ok = 0;
 627         ssize_t blacklist_index = 0;
 628         char *prgname = argv[0];
 629         static struct option lgopts[] = {
 630                 {OPT_NO_HUGE, 0, 0, 0},
 631                 {OPT_NO_PCI, 0, 0, 0},
 632                 {OPT_NO_HPET, 0, 0, 0},
 633                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 634                 {OPT_HUGE_DIR, 1, 0, 0},
 635                 {OPT_NO_SHCONF, 0, 0, 0},
 636                 {OPT_PROC_TYPE, 1, 0, 0},
 637                 {OPT_FILE_PREFIX, 1, 0, 0},
 638                 {OPT_SOCKET_MEM, 1, 0, 0},
 639                 {OPT_USE_DEVICE, 1, 0, 0},
 640                 {OPT_SYSLOG, 1, NULL, 0},
 641                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 642                 {OPT_XEN_DOM0, 0, 0, 0},
 643                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 644                 {0, 0, 0, 0}
 645         };
 646         struct shared_driver *solib;
 647
 648         argvopt = argv;
 649
 650         internal_config.memory = 0;
 651         internal_config.force_nrank = 0;
 652         internal_config.force_nchannel = 0;
 653         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 654         internal_config.hugepage_dir = NULL;
 655         internal_config.force_sockets = 0;
 656         internal_config.syslog_facility = LOG_DAEMON;
 657         internal_config.xen_dom0_support = 0;
 658 #ifdef RTE_LIBEAL_USE_HPET
 659         internal_config.no_hpet = 0;
 660 #else
 661         internal_config.no_hpet = 1;
 662 #endif
 663         /* zero out the NUMA config */
 664         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 665                 internal_config.socket_mem[i] = 0;
 666
 667         /* zero out hugedir descriptors */
 668         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 669                 internal_config.hugepage_info[i].lock_descriptor = -1;
 670
 671         internal_config.vmware_tsc_map = 0;
 672         internal_config.base_virtaddr = 0;
 673
 674         while ((opt = getopt_long(argc, argvopt, "b:c:d:m:n:r:v",
 675                                   lgopts, &option_index)) != EOF) {
 676
 677                 switch (opt) {
 678                 /* blacklist */
 679                 case 'b':
 680                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 681                             blacklist_index)) < 0) {
 682                                 eal_usage(prgname);
 683                                 return (-1);
 684                         }
 685                         break;
 686                 /* coremask */
 687                 case 'c':
 688                         if (eal_parse_coremask(optarg) < 0) {
 689                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 690                                 eal_usage(prgname);
 691                                 return -1;
 692                         }
 693                         coremask_ok = 1;
 694                         break;
 695                 /* force loading of external driver */
 696                 case 'd':
 697                         solib = malloc(sizeof(*solib));
 698                         if (solib == NULL) {
 699                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 700                                 return -1;
 701                         }
 702                         memset(solib, 0, sizeof(*solib));
 703                         strncpy(solib->name, optarg, PATH_MAX-1);
 704                         solib->name[PATH_MAX-1] = 0;
 705                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 706                         break;
 707                 /* size of memory */
 708                 case 'm':
 709                         internal_config.memory = atoi(optarg);
 710                         internal_config.memory *= 1024ULL;
 711                         internal_config.memory *= 1024ULL;
 712                         break;
 713                 /* force number of channels */
 714                 case 'n':
 715                         internal_config.force_nchannel = atoi(optarg);
 716                         if (internal_config.force_nchannel == 0 ||
 717                             internal_config.force_nchannel > 4) {
 718                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 719                                 eal_usage(prgname);
 720                                 return -1;
 721                         }
 722                         break;
 723                 /* force number of ranks */
 724                 case 'r':
 725                         internal_config.force_nrank = atoi(optarg);
 726                         if (internal_config.force_nrank == 0 ||
 727                             internal_config.force_nrank > 16) {
 728                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 729                                 eal_usage(prgname);
 730                                 return -1;
 731                         }
 732                         break;
 733                 case 'v':
 734                         /* since message is explicitly requested by user, we
 735                          * write message at highest log level so it can always be seen
 736                          * even if info or warning messages are disabled */
 737                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 738                         break;
 739
 740                 /* long options */
 741                 case 0:
 742                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 743                                 internal_config.no_hugetlbfs = 1;
 744                         }
 745                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 746                 #ifdef RTE_LIBRTE_XEN_DOM0
 747                                 internal_config.xen_dom0_support = 1;
 748                 #else
 749                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 750                                         "running on Dom0, please configure"
 751                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 752                                 return -1;
 753                 #endif
 754                         }
 755                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 756                                 internal_config.no_pci = 1;
 757                         }
 758                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 759                                 internal_config.no_hpet = 1;
 760                         }
 761                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 762                                 internal_config.vmware_tsc_map = 1;
 763                         }
 764                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 765                                 internal_config.no_shconf = 1;
 766                         }
 767                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 768                                 internal_config.hugepage_dir = optarg;
 769                         }
 770                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 771                                 internal_config.process_type = eal_parse_proc_type(optarg);
 772                         }
 773                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 774                                 internal_config.hugefile_prefix = optarg;
 775                         }
 776                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 777                                 if (eal_parse_socket_mem(optarg) < 0) {
 778                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 779                                                         OPT_SOCKET_MEM "\n");
 780                                         eal_usage(prgname);
 781                                         return -1;
 782                                 }
 783                         }
 784                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 785                                 eal_dev_whitelist_add_entry(optarg);
 786                         }
 787                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 788                                 if (eal_parse_syslog(optarg) < 0) {
 789                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 790                                                         OPT_SYSLOG "\n");
 791                                         eal_usage(prgname);
 792                                         return -1;
 793                                 }
 794                         }
 795                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 796                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 797                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 798                                                         OPT_BASE_VIRTADDR "\n");
 799                                         eal_usage(prgname);
 800                                         return -1;
 801                                 }
 802                         }
 803                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 804                                 internal_config.create_uio_dev = 1;
 805                         }
 806                         break;
 807
 808                 default:
 809                         eal_usage(prgname);
 810                         return -1;
 811                 }
 812         }
 813
 814         /* sanity checks */
 815         if (!coremask_ok) {
 816                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 817                 eal_usage(prgname);
 818                 return -1;
 819         }
 820         if (internal_config.process_type == RTE_PROC_AUTO){
 821                 internal_config.process_type = eal_proc_type_detect();
 822         }
 823         if (internal_config.process_type == RTE_PROC_INVALID){
 824                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 825                 eal_usage(prgname);
 826                 return -1;
 827         }
 828         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 829                         internal_config.force_nchannel == 0) {
 830                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 831                 eal_usage(prgname);
 832                 return -1;
 833         }
 834         if (index(internal_config.hugefile_prefix,'%') != NULL){
 835                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 836                 eal_usage(prgname);
 837                 return -1;
 838         }
 839         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 840                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 841                                 "at the same time\n");
 842                 eal_usage(prgname);
 843                 return -1;
 844         }
 845         /* --no-huge doesn't make sense with either -m or --socket-mem */
 846         if (internal_config.no_hugetlbfs &&
 847                         (internal_config.memory > 0 ||
 848                                         internal_config.force_sockets == 1)) {
 849                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 850                                 "together with --no-huge!\n");
 851                 eal_usage(prgname);
 852                 return -1;
 853         }
 854         /* --xen-dom0 doesn't make sense with --socket-mem */
 855         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 856                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 857                                         "together with --xen_dom0!\n");
 858                 eal_usage(prgname);
 859                 return -1;
 860         }
 861         /* if no blacklist, parse a whitelist */
 862         if (blacklist_index > 0) {
 863                 if (eal_dev_whitelist_exists()) {
 864                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 865                                         "[--use-device] options cannot be used at the same time\n");
 866                         eal_usage(prgname);
 867                         return -1;
 868                 }
 869                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 870         } else {
 871                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 872                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 873                         return -1;
 874                 }
 875         }
 876
 877         if (optind >= 0)
 878                 argv[optind-1] = prgname;
 879
 880         /* if no memory amounts were requested, this will result in 0 and
 881          * will be overriden later, right after eal_hugepage_info_init() */
 882         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 883                 internal_config.memory += internal_config.socket_mem[i];
 884
 885         ret = optind-1;
 886         optind = 0; /* reset getopt lib */
 887         return ret;
 888 }
 889
 890 static void
 891 eal_check_mem_on_local_socket(void)
 892 {
 893         const struct rte_memseg *ms;
 894         int i, socket_id;
 895
 896         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 897
 898         ms = rte_eal_get_physmem_layout();
 899
 900         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 901                 if (ms[i].socket_id == socket_id &&
 902                                 ms[i].len > 0)
 903                         return;
 904
 905         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 906                         "memory on local socket!\n");
 907 }
 908
 909 static int
 910 sync_func(__attribute__((unused)) void *arg)
 911 {
 912         return 0;
 913 }
 914
 915 inline static void
 916 rte_eal_mcfg_complete(void)
 917 {
 918         /* ALL shared mem_config related INIT DONE */
 919         if (rte_config.process_type == RTE_PROC_PRIMARY)
 920                 rte_config.mem_config->magic = RTE_MAGIC;
 921 }
 922
 923 /*
 924  * Request iopl priviledge for all RPL, returns 0 on success
 925  */
 926 static int
 927 rte_eal_iopl_init(void)
 928 {
 929         return iopl(HIGHEST_RPL);
 930 }
 931
 932 /* Launch threads, called at application init(). */
 933 int
 934 rte_eal_init(int argc, char **argv)
 935 {
 936         int i, fctret, ret;
 937         pthread_t thread_id;
 938         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 939         struct shared_driver *solib = NULL;
 940         const char *logid;
 941
 942         if (!rte_atomic32_test_and_set(&run_once))
 943                 return -1;
 944
 945         logid = strrchr(argv[0], '/');
 946         logid = strdup(logid ? logid + 1: argv[0]);
 947
 948         thread_id = pthread_self();
 949
 950         if (rte_eal_log_early_init() < 0)
 951                 rte_panic("Cannot init early logs\n");
 952
 953         if (rte_eal_cpu_init() < 0)
 954                 rte_panic("Cannot detect lcores\n");
 955
 956         fctret = eal_parse_args(argc, argv);
 957         if (fctret < 0)
 958                 exit(1);
 959
 960         if (internal_config.no_hugetlbfs == 0 &&
 961                         internal_config.process_type != RTE_PROC_SECONDARY &&
 962                         internal_config.xen_dom0_support == 0 &&
 963                         eal_hugepage_info_init() < 0)
 964                 rte_panic("Cannot get hugepage information\n");
 965
 966         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 967                 if (internal_config.no_hugetlbfs)
 968                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 969                 else
 970                         internal_config.memory = eal_get_hugepage_mem_size();
 971         }
 972
 973         if (internal_config.vmware_tsc_map == 1) {
 974 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 975                 rte_cycles_vmware_tsc_map = 1;
 976                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 977                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 978 #else
 979                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 980                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 981 #endif
 982         }
 983
 984         rte_srand(rte_rdtsc());
 985
 986         rte_config_init();
 987
 988         if (rte_eal_iopl_init() == 0)
 989                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 990
 991         if (rte_eal_pci_init() < 0)
 992                 rte_panic("Cannot init PCI\n");
 993
 994 #ifdef RTE_LIBRTE_IVSHMEM
 995         if (rte_eal_ivshmem_init() < 0)
 996                 rte_panic("Cannot init IVSHMEM\n");
 997 #endif
 998
 999         if (rte_eal_memory_init() < 0)
1000                 rte_panic("Cannot init memory\n");
1001
1002         /* the directories are locked during eal_hugepage_info_init */
1003         eal_hugedirs_unlock();
1004
1005         if (rte_eal_memzone_init() < 0)
1006                 rte_panic("Cannot init memzone\n");
1007
1008         if (rte_eal_tailqs_init() < 0)
1009                 rte_panic("Cannot init tail queues for objects\n");
1010
1011 #ifdef RTE_LIBRTE_IVSHMEM
1012         if (rte_eal_ivshmem_obj_init() < 0)
1013                 rte_panic("Cannot init IVSHMEM objects\n");
1014 #endif
1015
1016         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1017                 rte_panic("Cannot init logs\n");
1018
1019         if (rte_eal_alarm_init() < 0)
1020                 rte_panic("Cannot init interrupt-handling thread\n");
1021
1022         if (rte_eal_intr_init() < 0)
1023                 rte_panic("Cannot init interrupt-handling thread\n");
1024
1025         if (rte_eal_timer_init() < 0)
1026                 rte_panic("Cannot init HPET or TSC timers\n");
1027
1028         eal_check_mem_on_local_socket();
1029
1030         rte_eal_mcfg_complete();
1031
1032         if (rte_eal_non_pci_ethdev_init() < 0)
1033                 rte_panic("Cannot init non-PCI eth_devs\n");
1034
1035         TAILQ_FOREACH(solib, &solib_list, next) {
1036                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1037                 if ((solib->lib_handle == NULL) && (solib->name[0] != '/')) {
1038                         /* relative path: try again with "./" prefix */
1039                         char sopath[PATH_MAX];
1040                         snprintf(sopath, sizeof(sopath), "./%s", solib->name);
1041                         solib->lib_handle = dlopen(sopath, RTLD_NOW);
1042                 }
1043                 if (solib->lib_handle == NULL)
1044                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1045         }
1046
1047         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1048                 rte_config.master_lcore, (int)thread_id);
1049
1050         RTE_LCORE_FOREACH_SLAVE(i) {
1051
1052                 /*
1053                  * create communication pipes between master thread
1054                  * and children
1055                  */
1056                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1057                         rte_panic("Cannot create pipe\n");
1058                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1059                         rte_panic("Cannot create pipe\n");
1060
1061                 lcore_config[i].state = WAIT;
1062
1063                 /* create a thread for each lcore */
1064                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1065                                      eal_thread_loop, NULL);
1066                 if (ret != 0)
1067                         rte_panic("Cannot create thread\n");
1068         }
1069
1070         eal_thread_init_master(rte_config.master_lcore);
1071
1072         /*
1073          * Launch a dummy function on all slave lcores, so that master lcore
1074          * knows they are all ready when this function returns.
1075          */
1076         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1077         rte_eal_mp_wait_lcore();
1078
1079         return fctret;
1080 }
1081
1082 /* get core role */
1083 enum rte_lcore_role_t
1084 rte_eal_lcore_role(unsigned lcore_id)
1085 {
1086         return (rte_config.lcore_role[lcore_id]);
1087 }
1088
1089 enum rte_proc_type_t
1090 rte_eal_process_type(void)
1091 {
1092         return (rte_config.process_type);
1093 }
1094
1095 int rte_eal_has_hugepages(void)
1096 {
1097         return ! internal_config.no_hugetlbfs;
1098 }