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                                 cfg->lcore_role[idx] = ROLE_RTE;
 445                                 if(count == 0)
 446                                         cfg->master_lcore = idx;
 447                                 count++;
 448                         } else  {
 449                                 cfg->lcore_role[idx] = ROLE_OFF;
 450                         }
 451                 }
 452         }
 453         for(; i >= 0; i--)
 454                 if(coremask[i] != '0')
 455                         return -1;
 456         for(; idx < RTE_MAX_LCORE; idx++)
 457                 cfg->lcore_role[idx] = ROLE_OFF;
 458         if(count == 0)
 459                 return -1;
 460         /* Update the count of enabled logical cores of the EAL configuration */
 461         cfg->lcore_count = count;
 462         return 0;
 463 }
 464
 465 static int
 466 eal_parse_syslog(const char *facility)
 467 {
 468         int i;
 469         static struct {
 470                 const char *name;
 471                 int value;
 472         } map[] = {
 473                 { "auth", LOG_AUTH },
 474                 { "cron", LOG_CRON },
 475                 { "daemon", LOG_DAEMON },
 476                 { "ftp", LOG_FTP },
 477                 { "kern", LOG_KERN },
 478                 { "lpr", LOG_LPR },
 479                 { "mail", LOG_MAIL },
 480                 { "news", LOG_NEWS },
 481                 { "syslog", LOG_SYSLOG },
 482                 { "user", LOG_USER },
 483                 { "uucp", LOG_UUCP },
 484                 { "local0", LOG_LOCAL0 },
 485                 { "local1", LOG_LOCAL1 },
 486                 { "local2", LOG_LOCAL2 },
 487                 { "local3", LOG_LOCAL3 },
 488                 { "local4", LOG_LOCAL4 },
 489                 { "local5", LOG_LOCAL5 },
 490                 { "local6", LOG_LOCAL6 },
 491                 { "local7", LOG_LOCAL7 },
 492                 { NULL, 0 }
 493         };
 494
 495         for (i = 0; map[i].name; i++) {
 496                 if (!strcmp(facility, map[i].name)) {
 497                         internal_config.syslog_facility = map[i].value;
 498                         return 0;
 499                 }
 500         }
 501         return -1;
 502 }
 503
 504 static int
 505 eal_parse_socket_mem(char *socket_mem)
 506 {
 507         char * arg[RTE_MAX_NUMA_NODES];
 508         char *end;
 509         int arg_num, i, len;
 510         uint64_t total_mem = 0;
 511
 512         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 513         if (len == SOCKET_MEM_STRLEN) {
 514                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 515                 return -1;
 516         }
 517
 518         /* all other error cases will be caught later */
 519         if (!isdigit(socket_mem[len-1]))
 520                 return -1;
 521
 522         /* split the optarg into separate socket values */
 523         arg_num = rte_strsplit(socket_mem, len,
 524                         arg, RTE_MAX_NUMA_NODES, ',');
 525
 526         /* if split failed, or 0 arguments */
 527         if (arg_num <= 0)
 528                 return -1;
 529
 530         internal_config.force_sockets = 1;
 531
 532         /* parse each defined socket option */
 533         errno = 0;
 534         for (i = 0; i < arg_num; i++) {
 535                 end = NULL;
 536                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 537
 538                 /* check for invalid input */
 539                 if ((errno != 0)  ||
 540                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 541                         return -1;
 542                 internal_config.socket_mem[i] *= 1024ULL;
 543                 internal_config.socket_mem[i] *= 1024ULL;
 544                 total_mem += internal_config.socket_mem[i];
 545         }
 546
 547         /* check if we have a positive amount of total memory */
 548         if (total_mem == 0)
 549                 return -1;
 550
 551         return 0;
 552 }
 553
 554 static int
 555 eal_parse_base_virtaddr(const char *arg)
 556 {
 557         char *end;
 558         uint64_t addr;
 559
 560         addr = strtoull(arg, &end, 16);
 561
 562         /* check for errors */
 563         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 564                 return -1;
 565
 566         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 567 #ifndef RTE_ARCH_X86_64
 568         if (addr >= UINTPTR_MAX)
 569                 return -1;
 570 #endif
 571
 572         /* align the addr on 2M boundary */
 573         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 574                                                            RTE_PGSIZE_2M);
 575
 576         return 0;
 577 }
 578
 579 static inline size_t
 580 eal_get_hugepage_mem_size(void)
 581 {
 582         uint64_t size = 0;
 583         unsigned i, j;
 584
 585         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 586                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 587                 if (hpi->hugedir != NULL) {
 588                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 589                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 590                         }
 591                 }
 592         }
 593
 594         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 595 }
 596
 597 static enum rte_proc_type_t
 598 eal_parse_proc_type(const char *arg)
 599 {
 600         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 601                 return RTE_PROC_PRIMARY;
 602         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 603                 return RTE_PROC_SECONDARY;
 604         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 605                 return RTE_PROC_AUTO;
 606
 607         return RTE_PROC_INVALID;
 608 }
 609
 610 /* Parse the argument given in the command line of the application */
 611 static int
 612 eal_parse_args(int argc, char **argv)
 613 {
 614         int opt, ret, i;
 615         char **argvopt;
 616         int option_index;
 617         int coremask_ok = 0;
 618         char *prgname = argv[0];
 619         static struct option lgopts[] = {
 620                 {OPT_NO_HUGE, 0, 0, 0},
 621                 {OPT_NO_PCI, 0, 0, 0},
 622                 {OPT_NO_HPET, 0, 0, 0},
 623                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 624                 {OPT_HUGE_DIR, 1, 0, 0},
 625                 {OPT_NO_SHCONF, 0, 0, 0},
 626                 {OPT_PROC_TYPE, 1, 0, 0},
 627                 {OPT_FILE_PREFIX, 1, 0, 0},
 628                 {OPT_SOCKET_MEM, 1, 0, 0},
 629                 {OPT_PCI_WHITELIST, 1, 0, 0},
 630                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 631                 {OPT_VDEV, 1, 0, 0},
 632                 {OPT_SYSLOG, 1, NULL, 0},
 633                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 634                 {OPT_XEN_DOM0, 0, 0, 0},
 635                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 636                 {0, 0, 0, 0}
 637         };
 638         struct shared_driver *solib;
 639
 640         argvopt = argv;
 641
 642         internal_config.memory = 0;
 643         internal_config.force_nrank = 0;
 644         internal_config.force_nchannel = 0;
 645         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 646         internal_config.hugepage_dir = NULL;
 647         internal_config.force_sockets = 0;
 648         internal_config.syslog_facility = LOG_DAEMON;
 649         internal_config.xen_dom0_support = 0;
 650 #ifdef RTE_LIBEAL_USE_HPET
 651         internal_config.no_hpet = 0;
 652 #else
 653         internal_config.no_hpet = 1;
 654 #endif
 655         /* zero out the NUMA config */
 656         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 657                 internal_config.socket_mem[i] = 0;
 658
 659         /* zero out hugedir descriptors */
 660         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 661                 internal_config.hugepage_info[i].lock_descriptor = -1;
 662
 663         internal_config.vmware_tsc_map = 0;
 664         internal_config.base_virtaddr = 0;
 665
 666         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
 667                                   lgopts, &option_index)) != EOF) {
 668
 669                 switch (opt) {
 670                 /* blacklist */
 671                 case 'b':
 672                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 673                                         optarg) < 0) {
 674                                 eal_usage(prgname);
 675                                 return (-1);
 676                         }
 677                         break;
 678                 /* whitelist */
 679                 case 'w':
 680                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 681                                         optarg) < 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                                 printf("The --use-device option is deprecated, please use\n"
 786                                         "--whitelist or --vdev instead.\n");
 787                                 eal_usage(prgname);
 788                                 return -1;
 789                         }
 790                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 791                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 792                                                 optarg) < 0) {
 793                                         eal_usage(prgname);
 794                                         return -1;
 795                                 }
 796                         }
 797                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 798                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 799                                                 optarg) < 0) {
 800                                         eal_usage(prgname);
 801                                         return -1;
 802                                 }
 803                         }
 804                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 805                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 806                                                 optarg) < 0) {
 807                                         eal_usage(prgname);
 808                                         return -1;
 809                                 }
 810                         }
 811                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 812                                 if (eal_parse_syslog(optarg) < 0) {
 813                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 814                                                         OPT_SYSLOG "\n");
 815                                         eal_usage(prgname);
 816                                         return -1;
 817                                 }
 818                         }
 819                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 820                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 821                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 822                                                         OPT_BASE_VIRTADDR "\n");
 823                                         eal_usage(prgname);
 824                                         return -1;
 825                                 }
 826                         }
 827                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 828                                 internal_config.create_uio_dev = 1;
 829                         }
 830                         break;
 831
 832                 default:
 833                         eal_usage(prgname);
 834                         return -1;
 835                 }
 836         }
 837
 838         /* sanity checks */
 839         if (!coremask_ok) {
 840                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 841                 eal_usage(prgname);
 842                 return -1;
 843         }
 844         if (internal_config.process_type == RTE_PROC_AUTO){
 845                 internal_config.process_type = eal_proc_type_detect();
 846         }
 847         if (internal_config.process_type == RTE_PROC_INVALID){
 848                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 849                 eal_usage(prgname);
 850                 return -1;
 851         }
 852         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 853                         internal_config.force_nchannel == 0) {
 854                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 855                 eal_usage(prgname);
 856                 return -1;
 857         }
 858         if (index(internal_config.hugefile_prefix,'%') != NULL){
 859                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 860                 eal_usage(prgname);
 861                 return -1;
 862         }
 863         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 864                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 865                                 "at the same time\n");
 866                 eal_usage(prgname);
 867                 return -1;
 868         }
 869         /* --no-huge doesn't make sense with either -m or --socket-mem */
 870         if (internal_config.no_hugetlbfs &&
 871                         (internal_config.memory > 0 ||
 872                                         internal_config.force_sockets == 1)) {
 873                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 874                                 "together with --no-huge!\n");
 875                 eal_usage(prgname);
 876                 return -1;
 877         }
 878         /* --xen-dom0 doesn't make sense with --socket-mem */
 879         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 880                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 881                                         "together with --xen_dom0!\n");
 882                 eal_usage(prgname);
 883                 return -1;
 884         }
 885
 886         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 887                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 888                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 889                         "[-w] options cannot be used at the same time\n");
 890                 eal_usage(prgname);
 891                 return -1;
 892         }
 893
 894         if (optind >= 0)
 895                 argv[optind-1] = prgname;
 896
 897         /* if no memory amounts were requested, this will result in 0 and
 898          * will be overriden later, right after eal_hugepage_info_init() */
 899         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 900                 internal_config.memory += internal_config.socket_mem[i];
 901
 902         ret = optind-1;
 903         optind = 0; /* reset getopt lib */
 904         return ret;
 905 }
 906
 907 static void
 908 eal_check_mem_on_local_socket(void)
 909 {
 910         const struct rte_memseg *ms;
 911         int i, socket_id;
 912
 913         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 914
 915         ms = rte_eal_get_physmem_layout();
 916
 917         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 918                 if (ms[i].socket_id == socket_id &&
 919                                 ms[i].len > 0)
 920                         return;
 921
 922         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 923                         "memory on local socket!\n");
 924 }
 925
 926 static int
 927 sync_func(__attribute__((unused)) void *arg)
 928 {
 929         return 0;
 930 }
 931
 932 inline static void
 933 rte_eal_mcfg_complete(void)
 934 {
 935         /* ALL shared mem_config related INIT DONE */
 936         if (rte_config.process_type == RTE_PROC_PRIMARY)
 937                 rte_config.mem_config->magic = RTE_MAGIC;
 938 }
 939
 940 /*
 941  * Request iopl priviledge for all RPL, returns 0 on success
 942  */
 943 static int
 944 rte_eal_iopl_init(void)
 945 {
 946         return iopl(HIGHEST_RPL);
 947 }
 948
 949 /* Launch threads, called at application init(). */
 950 int
 951 rte_eal_init(int argc, char **argv)
 952 {
 953         int i, fctret, ret;
 954         pthread_t thread_id;
 955         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 956         struct shared_driver *solib = NULL;
 957         const char *logid;
 958
 959         if (!rte_atomic32_test_and_set(&run_once))
 960                 return -1;
 961
 962         logid = strrchr(argv[0], '/');
 963         logid = strdup(logid ? logid + 1: argv[0]);
 964
 965         thread_id = pthread_self();
 966
 967         if (rte_eal_log_early_init() < 0)
 968                 rte_panic("Cannot init early logs\n");
 969
 970         if (rte_eal_cpu_init() < 0)
 971                 rte_panic("Cannot detect lcores\n");
 972
 973         fctret = eal_parse_args(argc, argv);
 974         if (fctret < 0)
 975                 exit(1);
 976
 977         if (internal_config.no_hugetlbfs == 0 &&
 978                         internal_config.process_type != RTE_PROC_SECONDARY &&
 979                         internal_config.xen_dom0_support == 0 &&
 980                         eal_hugepage_info_init() < 0)
 981                 rte_panic("Cannot get hugepage information\n");
 982
 983         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 984                 if (internal_config.no_hugetlbfs)
 985                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 986                 else
 987                         internal_config.memory = eal_get_hugepage_mem_size();
 988         }
 989
 990         if (internal_config.vmware_tsc_map == 1) {
 991 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 992                 rte_cycles_vmware_tsc_map = 1;
 993                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 994                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 995 #else
 996                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 997                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 998 #endif
 999         }
1000
1001         rte_srand(rte_rdtsc());
1002
1003         rte_config_init();
1004
1005         if (rte_eal_iopl_init() == 0)
1006                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1007
1008         if (rte_eal_pci_init() < 0)
1009                 rte_panic("Cannot init PCI\n");
1010
1011 #ifdef RTE_LIBRTE_IVSHMEM
1012         if (rte_eal_ivshmem_init() < 0)
1013                 rte_panic("Cannot init IVSHMEM\n");
1014 #endif
1015
1016         if (rte_eal_memory_init() < 0)
1017                 rte_panic("Cannot init memory\n");
1018
1019         /* the directories are locked during eal_hugepage_info_init */
1020         eal_hugedirs_unlock();
1021
1022         if (rte_eal_memzone_init() < 0)
1023                 rte_panic("Cannot init memzone\n");
1024
1025         if (rte_eal_tailqs_init() < 0)
1026                 rte_panic("Cannot init tail queues for objects\n");
1027
1028 #ifdef RTE_LIBRTE_IVSHMEM
1029         if (rte_eal_ivshmem_obj_init() < 0)
1030                 rte_panic("Cannot init IVSHMEM objects\n");
1031 #endif
1032
1033         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1034                 rte_panic("Cannot init logs\n");
1035
1036         if (rte_eal_alarm_init() < 0)
1037                 rte_panic("Cannot init interrupt-handling thread\n");
1038
1039         if (rte_eal_intr_init() < 0)
1040                 rte_panic("Cannot init interrupt-handling thread\n");
1041
1042         if (rte_eal_timer_init() < 0)
1043                 rte_panic("Cannot init HPET or TSC timers\n");
1044
1045         eal_check_mem_on_local_socket();
1046
1047         rte_eal_mcfg_complete();
1048
1049         TAILQ_FOREACH(solib, &solib_list, next) {
1050                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1051                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1052                 if ((solib->lib_handle == NULL) && (solib->name[0] != '/')) {
1053                         /* relative path: try again with "./" prefix */
1054                         char sopath[PATH_MAX];
1055                         snprintf(sopath, sizeof(sopath), "./%s", solib->name);
1056                         solib->lib_handle = dlopen(sopath, RTLD_NOW);
1057                 }
1058                 if (solib->lib_handle == NULL)
1059                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1060         }
1061
1062         if (rte_eal_vdev_init() < 0)
1063                 rte_panic("Cannot init virtual devices\n");
1064
1065         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1066                 rte_config.master_lcore, (int)thread_id);
1067
1068         RTE_LCORE_FOREACH_SLAVE(i) {
1069
1070                 /*
1071                  * create communication pipes between master thread
1072                  * and children
1073                  */
1074                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1075                         rte_panic("Cannot create pipe\n");
1076                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1077                         rte_panic("Cannot create pipe\n");
1078
1079                 lcore_config[i].state = WAIT;
1080
1081                 /* create a thread for each lcore */
1082                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1083                                      eal_thread_loop, NULL);
1084                 if (ret != 0)
1085                         rte_panic("Cannot create thread\n");
1086         }
1087
1088         eal_thread_init_master(rte_config.master_lcore);
1089
1090         /*
1091          * Launch a dummy function on all slave lcores, so that master lcore
1092          * knows they are all ready when this function returns.
1093          */
1094         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1095         rte_eal_mp_wait_lcore();
1096
1097         return fctret;
1098 }
1099
1100 /* get core role */
1101 enum rte_lcore_role_t
1102 rte_eal_lcore_role(unsigned lcore_id)
1103 {
1104         return (rte_config.lcore_role[lcore_id]);
1105 }
1106
1107 enum rte_proc_type_t
1108 rte_eal_process_type(void)
1109 {
1110         return (rte_config.process_type);
1111 }
1112
1113 int rte_eal_has_hugepages(void)
1114 {
1115         return ! internal_config.no_hugetlbfs;
1116 }