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_PCI_WHITELIST "pci-whitelist"
  96 #define OPT_PCI_BLACKLIST "pci-blacklist"
  97 #define OPT_VDEV        "vdev"
  98 #define OPT_SYSLOG      "syslog"
  99 #define OPT_BASE_VIRTADDR   "base-virtaddr"
 100 #define OPT_XEN_DOM0    "xen-dom0"
 101 #define OPT_CREATE_UIO_DEV "create-uio-dev"
 102
 103 #define RTE_EAL_BLACKLIST_SIZE  0x100
 104
 105 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 106
 107 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 108
 109 #define HIGHEST_RPL 3
 110
 111 #define BITS_PER_HEX 4
 112
 113 /* Allow the application to print its usage message too if set */
 114 static rte_usage_hook_t rte_application_usage_hook = NULL;
 115
 116 TAILQ_HEAD(shared_driver_list, shared_driver);
 117
 118 /* Definition for shared object drivers. */
 119 struct shared_driver {
 120         TAILQ_ENTRY(shared_driver) next;
 121
 122         char    name[PATH_MAX];
 123         void*   lib_handle;
 124 };
 125
 126 /* List of external loadable drivers */
 127 static struct shared_driver_list solib_list =
 128 TAILQ_HEAD_INITIALIZER(solib_list);
 129
 130 /* early configuration structure, when memory config is not mmapped */
 131 static struct rte_mem_config early_mem_config;
 132
 133 /* define fd variable here, because file needs to be kept open for the
 134  * duration of the program, as we hold a write lock on it in the primary proc */
 135 static int mem_cfg_fd = -1;
 136
 137 static struct flock wr_lock = {
 138                 .l_type = F_WRLCK,
 139                 .l_whence = SEEK_SET,
 140                 .l_start = offsetof(struct rte_mem_config, memseg),
 141                 .l_len = sizeof(early_mem_config.memseg),
 142 };
 143
 144 /* Address of global and public configuration */
 145 static struct rte_config rte_config = {
 146                 .mem_config = &early_mem_config,
 147 };
 148
 149 /* internal configuration (per-core) */
 150 struct lcore_config lcore_config[RTE_MAX_LCORE];
 151
 152 /* internal configuration */
 153 struct internal_config internal_config;
 154
 155 /* used by rte_rdtsc() */
 156 int rte_cycles_vmware_tsc_map;
 157
 158 /* Return a pointer to the configuration structure */
 159 struct rte_config *
 160 rte_eal_get_configuration(void)
 161 {
 162         return &rte_config;
 163 }
 164
 165 /* parse a sysfs (or other) file containing one integer value */
 166 int
 167 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 168 {
 169         FILE *f;
 170         char buf[BUFSIZ];
 171         char *end = NULL;
 172
 173         if ((f = fopen(filename, "r")) == NULL) {
 174                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 175                         __func__, filename);
 176                 return -1;
 177         }
 178
 179         if (fgets(buf, sizeof(buf), f) == NULL) {
 180                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 181                         __func__, filename);
 182                 fclose(f);
 183                 return -1;
 184         }
 185         *val = strtoul(buf, &end, 0);
 186         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 187                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 188                                 __func__, filename);
 189                 fclose(f);
 190                 return -1;
 191         }
 192         fclose(f);
 193         return 0;
 194 }
 195
 196
 197 /* create memory configuration in shared/mmap memory. Take out
 198  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 199  * This means we never close the file while running (auto-close on exit).
 200  * We also don't lock the whole file, so that in future we can use read-locks
 201  * on other parts, e.g. memzones, to detect if there are running secondary
 202  * processes. */
 203 static void
 204 rte_eal_config_create(void)
 205 {
 206         void *rte_mem_cfg_addr;
 207         int retval;
 208
 209         const char *pathname = eal_runtime_config_path();
 210
 211         if (internal_config.no_shconf)
 212                 return;
 213
 214         if (mem_cfg_fd < 0){
 215                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 216                 if (mem_cfg_fd < 0)
 217                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 218         }
 219
 220         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 221         if (retval < 0){
 222                 close(mem_cfg_fd);
 223                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 224         }
 225
 226         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 227         if (retval < 0){
 228                 close(mem_cfg_fd);
 229                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 230                                 "process running?\n", pathname);
 231         }
 232
 233         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 234                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 235
 236         if (rte_mem_cfg_addr == MAP_FAILED){
 237                 rte_panic("Cannot mmap memory for rte_config\n");
 238         }
 239         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 240         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 241 }
 242
 243 /* attach to an existing shared memory config */
 244 static void
 245 rte_eal_config_attach(void)
 246 {
 247         void *rte_mem_cfg_addr;
 248         const char *pathname = eal_runtime_config_path();
 249
 250         if (internal_config.no_shconf)
 251                 return;
 252
 253         if (mem_cfg_fd < 0){
 254                 mem_cfg_fd = open(pathname, O_RDWR);
 255                 if (mem_cfg_fd < 0)
 256                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 257         }
 258
 259         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 260                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 261         close(mem_cfg_fd);
 262         if (rte_mem_cfg_addr == MAP_FAILED)
 263                 rte_panic("Cannot mmap memory for rte_config\n");
 264
 265         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 266 }
 267
 268 /* Detect if we are a primary or a secondary process */
 269 static enum rte_proc_type_t
 270 eal_proc_type_detect(void)
 271 {
 272         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 273         const char *pathname = eal_runtime_config_path();
 274
 275         /* if we can open the file but not get a write-lock we are a secondary
 276          * process. NOTE: if we get a file handle back, we keep that open
 277          * and don't close it to prevent a race condition between multiple opens */
 278         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 279                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 280                 ptype = RTE_PROC_SECONDARY;
 281
 282         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 283                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 284
 285         return ptype;
 286 }
 287
 288 /* Sets up rte_config structure with the pointer to shared memory config.*/
 289 static void
 290 rte_config_init(void)
 291 {
 292         /* set the magic in configuration structure */
 293         rte_config.magic = RTE_MAGIC;
 294         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 295                         eal_proc_type_detect() : /* for auto, detect the type */
 296                         internal_config.process_type; /* otherwise use what's already set */
 297
 298         switch (rte_config.process_type){
 299         case RTE_PROC_PRIMARY:
 300                 rte_eal_config_create();
 301                 break;
 302         case RTE_PROC_SECONDARY:
 303                 rte_eal_config_attach();
 304                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 305                 break;
 306         case RTE_PROC_AUTO:
 307         case RTE_PROC_INVALID:
 308                 rte_panic("Invalid process type\n");
 309         }
 310 }
 311
 312 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 313 static void
 314 eal_hugedirs_unlock(void)
 315 {
 316         int i;
 317
 318         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 319         {
 320                 /* skip uninitialized */
 321                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 322                         continue;
 323                 /* unlock hugepage file */
 324                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 325                 close(internal_config.hugepage_info[i].lock_descriptor);
 326                 /* reset the field */
 327                 internal_config.hugepage_info[i].lock_descriptor = -1;
 328         }
 329 }
 330
 331 /* display usage */
 332 static void
 333 eal_usage(const char *prgname)
 334 {
 335         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 336                "[--proc-type primary|secondary|auto] \n\n"
 337                "EAL options:\n"
 338                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 339                "  -n NUM       : Number of memory channels\n"
 340                    "  -v           : Display version information on startup\n"
 341                "  -d LIB.so    : add driver (can be used multiple times)\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_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 353                "               Prevent EAL from using this PCI device. The argument\n"
 354                "               format is <domain:bus:devid.func>.\n"
 355                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 356                "               Only use the specified PCI devices. The argument format\n"
 357                "               is <[domain:]bus:devid.func>. This option can be present\n"
 358                "               several times (once per device).\n"
 359                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 360                "  --"OPT_VDEV": add a virtual device.\n"
 361                "               The argument format is <driver><id>[,key=val,...]\n"
 362                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 363                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 364                            "native RDTSC\n"
 365                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 366                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 367                "\nEAL options for DEBUG use only:\n"
 368                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 369                "  --"OPT_NO_PCI"   : disable pci\n"
 370                "  --"OPT_NO_HPET"  : disable hpet\n"
 371                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 372                "\n",
 373                prgname);
 374         /* Allow the application to print its usage message too if hook is set */
 375         if ( rte_application_usage_hook ) {
 376                 printf("===== Application Usage =====\n\n");
 377                 rte_application_usage_hook(prgname);
 378         }
 379 }
 380
 381 /* Set a per-application usage message */
 382 rte_usage_hook_t
 383 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 384 {
 385         rte_usage_hook_t        old_func;
 386
 387         /* Will be NULL on the first call to denote the last usage routine. */
 388         old_func                                        = rte_application_usage_hook;
 389         rte_application_usage_hook      = usage_func;
 390
 391         return old_func;
 392 }
 393
 394 /*
 395  * Parse the coremask given as argument (hexadecimal string) and fill
 396  * the global configuration (core role and core count) with the parsed
 397  * value.
 398  */
 399 static int xdigit2val(unsigned char c)
 400 {
 401         int val;
 402         if(isdigit(c))
 403                 val = c - '0';
 404         else if(isupper(c))
 405                 val = c - 'A' + 10;
 406         else
 407                 val = c - 'a' + 10;
 408         return val;
 409 }
 410 static int
 411 eal_parse_coremask(const char *coremask)
 412 {
 413         struct rte_config *cfg = rte_eal_get_configuration();
 414         int i, j, idx = 0 ;
 415         unsigned count = 0;
 416         char c;
 417         int val;
 418
 419         if (coremask == NULL)
 420                 return -1;
 421         /* Remove all blank characters ahead and after .
 422          * Remove 0x/0X if exists.
 423          */
 424         while (isblank(*coremask))
 425                 coremask++;
 426         if (coremask[0] == '0' && ((coremask[1] == 'x')
 427                 ||  (coremask[1] == 'X')) )
 428                 coremask += 2;
 429         i = strnlen(coremask, PATH_MAX);
 430         while ((i > 0) && isblank(coremask[i - 1]))
 431                 i--;
 432         if (i == 0)
 433                 return -1;
 434
 435         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 436                 c = coremask[i];
 437                 if (isxdigit(c) == 0) {
 438                         /* invalid characters */
 439                         return (-1);
 440                 }
 441                 val = xdigit2val(c);
 442                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 443                         if((1 << j) & val) {
 444                                 if (!lcore_config[idx].detected) {
 445                                         RTE_LOG(ERR, EAL, "lcore %u "
 446                                                 "unavailable\n", idx);
 447                                         return -1;
 448                                 }
 449                                 cfg->lcore_role[idx] = ROLE_RTE;
 450                                 if(count == 0)
 451                                         cfg->master_lcore = idx;
 452                                 count++;
 453                         } else  {
 454                                 cfg->lcore_role[idx] = ROLE_OFF;
 455                         }
 456                 }
 457         }
 458         for(; i >= 0; i--)
 459                 if(coremask[i] != '0')
 460                         return -1;
 461         for(; idx < RTE_MAX_LCORE; idx++)
 462                 cfg->lcore_role[idx] = ROLE_OFF;
 463         if(count == 0)
 464                 return -1;
 465         /* Update the count of enabled logical cores of the EAL configuration */
 466         cfg->lcore_count = count;
 467         return 0;
 468 }
 469
 470 static int
 471 eal_parse_syslog(const char *facility)
 472 {
 473         int i;
 474         static struct {
 475                 const char *name;
 476                 int value;
 477         } map[] = {
 478                 { "auth", LOG_AUTH },
 479                 { "cron", LOG_CRON },
 480                 { "daemon", LOG_DAEMON },
 481                 { "ftp", LOG_FTP },
 482                 { "kern", LOG_KERN },
 483                 { "lpr", LOG_LPR },
 484                 { "mail", LOG_MAIL },
 485                 { "news", LOG_NEWS },
 486                 { "syslog", LOG_SYSLOG },
 487                 { "user", LOG_USER },
 488                 { "uucp", LOG_UUCP },
 489                 { "local0", LOG_LOCAL0 },
 490                 { "local1", LOG_LOCAL1 },
 491                 { "local2", LOG_LOCAL2 },
 492                 { "local3", LOG_LOCAL3 },
 493                 { "local4", LOG_LOCAL4 },
 494                 { "local5", LOG_LOCAL5 },
 495                 { "local6", LOG_LOCAL6 },
 496                 { "local7", LOG_LOCAL7 },
 497                 { NULL, 0 }
 498         };
 499
 500         for (i = 0; map[i].name; i++) {
 501                 if (!strcmp(facility, map[i].name)) {
 502                         internal_config.syslog_facility = map[i].value;
 503                         return 0;
 504                 }
 505         }
 506         return -1;
 507 }
 508
 509 static int
 510 eal_parse_socket_mem(char *socket_mem)
 511 {
 512         char * arg[RTE_MAX_NUMA_NODES];
 513         char *end;
 514         int arg_num, i, len;
 515         uint64_t total_mem = 0;
 516
 517         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 518         if (len == SOCKET_MEM_STRLEN) {
 519                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 520                 return -1;
 521         }
 522
 523         /* all other error cases will be caught later */
 524         if (!isdigit(socket_mem[len-1]))
 525                 return -1;
 526
 527         /* split the optarg into separate socket values */
 528         arg_num = rte_strsplit(socket_mem, len,
 529                         arg, RTE_MAX_NUMA_NODES, ',');
 530
 531         /* if split failed, or 0 arguments */
 532         if (arg_num <= 0)
 533                 return -1;
 534
 535         internal_config.force_sockets = 1;
 536
 537         /* parse each defined socket option */
 538         errno = 0;
 539         for (i = 0; i < arg_num; i++) {
 540                 end = NULL;
 541                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 542
 543                 /* check for invalid input */
 544                 if ((errno != 0)  ||
 545                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 546                         return -1;
 547                 internal_config.socket_mem[i] *= 1024ULL;
 548                 internal_config.socket_mem[i] *= 1024ULL;
 549                 total_mem += internal_config.socket_mem[i];
 550         }
 551
 552         /* check if we have a positive amount of total memory */
 553         if (total_mem == 0)
 554                 return -1;
 555
 556         return 0;
 557 }
 558
 559 static int
 560 eal_parse_base_virtaddr(const char *arg)
 561 {
 562         char *end;
 563         uint64_t addr;
 564
 565         addr = strtoull(arg, &end, 16);
 566
 567         /* check for errors */
 568         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 569                 return -1;
 570
 571         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 572 #ifndef RTE_ARCH_X86_64
 573         if (addr >= UINTPTR_MAX)
 574                 return -1;
 575 #endif
 576
 577         /* align the addr on 2M boundary */
 578         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 579                                                            RTE_PGSIZE_2M);
 580
 581         return 0;
 582 }
 583
 584 static inline size_t
 585 eal_get_hugepage_mem_size(void)
 586 {
 587         uint64_t size = 0;
 588         unsigned i, j;
 589
 590         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 591                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 592                 if (hpi->hugedir != NULL) {
 593                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 594                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 595                         }
 596                 }
 597         }
 598
 599         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 600 }
 601
 602 static enum rte_proc_type_t
 603 eal_parse_proc_type(const char *arg)
 604 {
 605         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 606                 return RTE_PROC_PRIMARY;
 607         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 608                 return RTE_PROC_SECONDARY;
 609         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 610                 return RTE_PROC_AUTO;
 611
 612         return RTE_PROC_INVALID;
 613 }
 614
 615 /* Parse the argument given in the command line of the application */
 616 static int
 617 eal_parse_args(int argc, char **argv)
 618 {
 619         int opt, ret, i;
 620         char **argvopt;
 621         int option_index;
 622         int coremask_ok = 0;
 623         char *prgname = argv[0];
 624         static struct option lgopts[] = {
 625                 {OPT_NO_HUGE, 0, 0, 0},
 626                 {OPT_NO_PCI, 0, 0, 0},
 627                 {OPT_NO_HPET, 0, 0, 0},
 628                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 629                 {OPT_HUGE_DIR, 1, 0, 0},
 630                 {OPT_NO_SHCONF, 0, 0, 0},
 631                 {OPT_PROC_TYPE, 1, 0, 0},
 632                 {OPT_FILE_PREFIX, 1, 0, 0},
 633                 {OPT_SOCKET_MEM, 1, 0, 0},
 634                 {OPT_PCI_WHITELIST, 1, 0, 0},
 635                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 636                 {OPT_VDEV, 1, 0, 0},
 637                 {OPT_SYSLOG, 1, NULL, 0},
 638                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 639                 {OPT_XEN_DOM0, 0, 0, 0},
 640                 {OPT_CREATE_UIO_DEV, 1, NULL, 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:w:c:d:m:n:r:v",
 672                                   lgopts, &option_index)) != EOF) {
 673
 674                 switch (opt) {
 675                 /* blacklist */
 676                 case 'b':
 677                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 678                                         optarg) < 0) {
 679                                 eal_usage(prgname);
 680                                 return (-1);
 681                         }
 682                         break;
 683                 /* whitelist */
 684                 case 'w':
 685                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 686                                         optarg) < 0) {
 687                                 eal_usage(prgname);
 688                                 return -1;
 689                         }
 690                         break;
 691                 /* coremask */
 692                 case 'c':
 693                         if (eal_parse_coremask(optarg) < 0) {
 694                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 695                                 eal_usage(prgname);
 696                                 return -1;
 697                         }
 698                         coremask_ok = 1;
 699                         break;
 700                 /* force loading of external driver */
 701                 case 'd':
 702                         solib = malloc(sizeof(*solib));
 703                         if (solib == NULL) {
 704                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 705                                 return -1;
 706                         }
 707                         memset(solib, 0, sizeof(*solib));
 708                         strncpy(solib->name, optarg, PATH_MAX-1);
 709                         solib->name[PATH_MAX-1] = 0;
 710                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 711                         break;
 712                 /* size of memory */
 713                 case 'm':
 714                         internal_config.memory = atoi(optarg);
 715                         internal_config.memory *= 1024ULL;
 716                         internal_config.memory *= 1024ULL;
 717                         break;
 718                 /* force number of channels */
 719                 case 'n':
 720                         internal_config.force_nchannel = atoi(optarg);
 721                         if (internal_config.force_nchannel == 0 ||
 722                             internal_config.force_nchannel > 4) {
 723                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 724                                 eal_usage(prgname);
 725                                 return -1;
 726                         }
 727                         break;
 728                 /* force number of ranks */
 729                 case 'r':
 730                         internal_config.force_nrank = atoi(optarg);
 731                         if (internal_config.force_nrank == 0 ||
 732                             internal_config.force_nrank > 16) {
 733                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 734                                 eal_usage(prgname);
 735                                 return -1;
 736                         }
 737                         break;
 738                 case 'v':
 739                         /* since message is explicitly requested by user, we
 740                          * write message at highest log level so it can always be seen
 741                          * even if info or warning messages are disabled */
 742                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 743                         break;
 744
 745                 /* long options */
 746                 case 0:
 747                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 748                                 internal_config.no_hugetlbfs = 1;
 749                         }
 750                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 751                 #ifdef RTE_LIBRTE_XEN_DOM0
 752                                 internal_config.xen_dom0_support = 1;
 753                 #else
 754                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 755                                         "running on Dom0, please configure"
 756                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 757                                 return -1;
 758                 #endif
 759                         }
 760                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 761                                 internal_config.no_pci = 1;
 762                         }
 763                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 764                                 internal_config.no_hpet = 1;
 765                         }
 766                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 767                                 internal_config.vmware_tsc_map = 1;
 768                         }
 769                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 770                                 internal_config.no_shconf = 1;
 771                         }
 772                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 773                                 internal_config.hugepage_dir = optarg;
 774                         }
 775                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 776                                 internal_config.process_type = eal_parse_proc_type(optarg);
 777                         }
 778                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 779                                 internal_config.hugefile_prefix = optarg;
 780                         }
 781                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 782                                 if (eal_parse_socket_mem(optarg) < 0) {
 783                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 784                                                         OPT_SOCKET_MEM "\n");
 785                                         eal_usage(prgname);
 786                                         return -1;
 787                                 }
 788                         }
 789                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 790                                 printf("The --use-device option is deprecated, please use\n"
 791                                         "--whitelist or --vdev instead.\n");
 792                                 eal_usage(prgname);
 793                                 return -1;
 794                         }
 795                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 796                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 797                                                 optarg) < 0) {
 798                                         eal_usage(prgname);
 799                                         return -1;
 800                                 }
 801                         }
 802                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 803                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 804                                                 optarg) < 0) {
 805                                         eal_usage(prgname);
 806                                         return -1;
 807                                 }
 808                         }
 809                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 810                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 811                                                 optarg) < 0) {
 812                                         eal_usage(prgname);
 813                                         return -1;
 814                                 }
 815                         }
 816                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 817                                 if (eal_parse_syslog(optarg) < 0) {
 818                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 819                                                         OPT_SYSLOG "\n");
 820                                         eal_usage(prgname);
 821                                         return -1;
 822                                 }
 823                         }
 824                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 825                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 826                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 827                                                         OPT_BASE_VIRTADDR "\n");
 828                                         eal_usage(prgname);
 829                                         return -1;
 830                                 }
 831                         }
 832                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 833                                 internal_config.create_uio_dev = 1;
 834                         }
 835                         break;
 836
 837                 default:
 838                         eal_usage(prgname);
 839                         return -1;
 840                 }
 841         }
 842
 843         /* sanity checks */
 844         if (!coremask_ok) {
 845                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 846                 eal_usage(prgname);
 847                 return -1;
 848         }
 849         if (internal_config.process_type == RTE_PROC_AUTO){
 850                 internal_config.process_type = eal_proc_type_detect();
 851         }
 852         if (internal_config.process_type == RTE_PROC_INVALID){
 853                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 854                 eal_usage(prgname);
 855                 return -1;
 856         }
 857         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 858                         internal_config.force_nchannel == 0) {
 859                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 860                 eal_usage(prgname);
 861                 return -1;
 862         }
 863         if (index(internal_config.hugefile_prefix,'%') != NULL){
 864                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 865                 eal_usage(prgname);
 866                 return -1;
 867         }
 868         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 869                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 870                                 "at the same time\n");
 871                 eal_usage(prgname);
 872                 return -1;
 873         }
 874         /* --no-huge doesn't make sense with either -m or --socket-mem */
 875         if (internal_config.no_hugetlbfs &&
 876                         (internal_config.memory > 0 ||
 877                                         internal_config.force_sockets == 1)) {
 878                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 879                                 "together with --no-huge!\n");
 880                 eal_usage(prgname);
 881                 return -1;
 882         }
 883         /* --xen-dom0 doesn't make sense with --socket-mem */
 884         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 885                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 886                                         "together with --xen_dom0!\n");
 887                 eal_usage(prgname);
 888                 return -1;
 889         }
 890
 891         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 892                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 893                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 894                         "[-w] options cannot be used at the same time\n");
 895                 eal_usage(prgname);
 896                 return -1;
 897         }
 898
 899         if (optind >= 0)
 900                 argv[optind-1] = prgname;
 901
 902         /* if no memory amounts were requested, this will result in 0 and
 903          * will be overriden later, right after eal_hugepage_info_init() */
 904         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 905                 internal_config.memory += internal_config.socket_mem[i];
 906
 907         ret = optind-1;
 908         optind = 0; /* reset getopt lib */
 909         return ret;
 910 }
 911
 912 static void
 913 eal_check_mem_on_local_socket(void)
 914 {
 915         const struct rte_memseg *ms;
 916         int i, socket_id;
 917
 918         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 919
 920         ms = rte_eal_get_physmem_layout();
 921
 922         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 923                 if (ms[i].socket_id == socket_id &&
 924                                 ms[i].len > 0)
 925                         return;
 926
 927         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 928                         "memory on local socket!\n");
 929 }
 930
 931 static int
 932 sync_func(__attribute__((unused)) void *arg)
 933 {
 934         return 0;
 935 }
 936
 937 inline static void
 938 rte_eal_mcfg_complete(void)
 939 {
 940         /* ALL shared mem_config related INIT DONE */
 941         if (rte_config.process_type == RTE_PROC_PRIMARY)
 942                 rte_config.mem_config->magic = RTE_MAGIC;
 943 }
 944
 945 /*
 946  * Request iopl priviledge for all RPL, returns 0 on success
 947  */
 948 static int
 949 rte_eal_iopl_init(void)
 950 {
 951         return iopl(HIGHEST_RPL);
 952 }
 953
 954 /* Launch threads, called at application init(). */
 955 int
 956 rte_eal_init(int argc, char **argv)
 957 {
 958         int i, fctret, ret;
 959         pthread_t thread_id;
 960         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 961         struct shared_driver *solib = NULL;
 962         const char *logid;
 963
 964         if (!rte_atomic32_test_and_set(&run_once))
 965                 return -1;
 966
 967         logid = strrchr(argv[0], '/');
 968         logid = strdup(logid ? logid + 1: argv[0]);
 969
 970         thread_id = pthread_self();
 971
 972         if (rte_eal_log_early_init() < 0)
 973                 rte_panic("Cannot init early logs\n");
 974
 975         if (rte_eal_cpu_init() < 0)
 976                 rte_panic("Cannot detect lcores\n");
 977
 978         fctret = eal_parse_args(argc, argv);
 979         if (fctret < 0)
 980                 exit(1);
 981
 982         if (internal_config.no_hugetlbfs == 0 &&
 983                         internal_config.process_type != RTE_PROC_SECONDARY &&
 984                         internal_config.xen_dom0_support == 0 &&
 985                         eal_hugepage_info_init() < 0)
 986                 rte_panic("Cannot get hugepage information\n");
 987
 988         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 989                 if (internal_config.no_hugetlbfs)
 990                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 991                 else
 992                         internal_config.memory = eal_get_hugepage_mem_size();
 993         }
 994
 995         if (internal_config.vmware_tsc_map == 1) {
 996 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 997                 rte_cycles_vmware_tsc_map = 1;
 998                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 999                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
1000 #else
1001                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
1002                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1003 #endif
1004         }
1005
1006         rte_srand(rte_rdtsc());
1007
1008         rte_config_init();
1009
1010         if (rte_eal_iopl_init() == 0)
1011                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1012
1013         if (rte_eal_pci_init() < 0)
1014                 rte_panic("Cannot init PCI\n");
1015
1016 #ifdef RTE_LIBRTE_IVSHMEM
1017         if (rte_eal_ivshmem_init() < 0)
1018                 rte_panic("Cannot init IVSHMEM\n");
1019 #endif
1020
1021         if (rte_eal_memory_init() < 0)
1022                 rte_panic("Cannot init memory\n");
1023
1024         /* the directories are locked during eal_hugepage_info_init */
1025         eal_hugedirs_unlock();
1026
1027         if (rte_eal_memzone_init() < 0)
1028                 rte_panic("Cannot init memzone\n");
1029
1030         if (rte_eal_tailqs_init() < 0)
1031                 rte_panic("Cannot init tail queues for objects\n");
1032
1033 #ifdef RTE_LIBRTE_IVSHMEM
1034         if (rte_eal_ivshmem_obj_init() < 0)
1035                 rte_panic("Cannot init IVSHMEM objects\n");
1036 #endif
1037
1038         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1039                 rte_panic("Cannot init logs\n");
1040
1041         if (rte_eal_alarm_init() < 0)
1042                 rte_panic("Cannot init interrupt-handling thread\n");
1043
1044         if (rte_eal_intr_init() < 0)
1045                 rte_panic("Cannot init interrupt-handling thread\n");
1046
1047         if (rte_eal_timer_init() < 0)
1048                 rte_panic("Cannot init HPET or TSC timers\n");
1049
1050         eal_check_mem_on_local_socket();
1051
1052         rte_eal_mcfg_complete();
1053
1054         TAILQ_FOREACH(solib, &solib_list, next) {
1055                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1056                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1057                 if (solib->lib_handle == NULL)
1058                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1059         }
1060
1061         if (rte_eal_vdev_init() < 0)
1062                 rte_panic("Cannot init virtual devices\n");
1063
1064         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1065                 rte_config.master_lcore, (int)thread_id);
1066
1067         RTE_LCORE_FOREACH_SLAVE(i) {
1068
1069                 /*
1070                  * create communication pipes between master thread
1071                  * and children
1072                  */
1073                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1074                         rte_panic("Cannot create pipe\n");
1075                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1076                         rte_panic("Cannot create pipe\n");
1077
1078                 lcore_config[i].state = WAIT;
1079
1080                 /* create a thread for each lcore */
1081                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1082                                      eal_thread_loop, NULL);
1083                 if (ret != 0)
1084                         rte_panic("Cannot create thread\n");
1085         }
1086
1087         eal_thread_init_master(rte_config.master_lcore);
1088
1089         /*
1090          * Launch a dummy function on all slave lcores, so that master lcore
1091          * knows they are all ready when this function returns.
1092          */
1093         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1094         rte_eal_mp_wait_lcore();
1095
1096         return fctret;
1097 }
1098
1099 /* get core role */
1100 enum rte_lcore_role_t
1101 rte_eal_lcore_role(unsigned lcore_id)
1102 {
1103         return (rte_config.lcore_role[lcore_id]);
1104 }
1105
1106 enum rte_proc_type_t
1107 rte_eal_process_type(void)
1108 {
1109         return (rte_config.process_type);
1110 }
1111
1112 int rte_eal_has_hugepages(void)
1113 {
1114         return ! internal_config.no_hugetlbfs;
1115 }