lib/librte_eal/linuxapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2012-2013 6WIND S.A.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <string.h>
  39 #include <stdarg.h>
  40 #include <unistd.h>
  41 #include <pthread.h>
  42 #include <syslog.h>
  43 #include <getopt.h>
  44 #include <sys/file.h>
  45 #include <fcntl.h>
  46 #include <dlfcn.h>
  47 #include <stddef.h>
  48 #include <errno.h>
  49 #include <limits.h>
  50 #include <errno.h>
  51 #include <sys/mman.h>
  52 #include <sys/queue.h>
  53 #include <sys/io.h>
  54
  55 #include <rte_common.h>
  56 #include <rte_debug.h>
  57 #include <rte_memory.h>
  58 #include <rte_memzone.h>
  59 #include <rte_launch.h>
  60 #include <rte_tailq.h>
  61 #include <rte_eal.h>
  62 #include <rte_eal_memconfig.h>
  63 #include <rte_per_lcore.h>
  64 #include <rte_lcore.h>
  65 #include <rte_log.h>
  66 #include <rte_random.h>
  67 #include <rte_cycles.h>
  68 #include <rte_string_fns.h>
  69 #include <rte_cpuflags.h>
  70 #include <rte_interrupts.h>
  71 #include <rte_pci.h>
  72 #include <rte_common.h>
  73 #include <rte_version.h>
  74 #include <rte_atomic.h>
  75 #include <malloc_heap.h>
  76 #include <rte_eth_ring.h>
  77
  78 #include "eal_private.h"
  79 #include "eal_thread.h"
  80 #include "eal_internal_cfg.h"
  81 #include "eal_filesystem.h"
  82 #include "eal_hugepages.h"
  83
  84 #define OPT_HUGE_DIR    "huge-dir"
  85 #define OPT_PROC_TYPE   "proc-type"
  86 #define OPT_NO_SHCONF   "no-shconf"
  87 #define OPT_NO_HPET     "no-hpet"
  88 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  89 #define OPT_NO_PCI      "no-pci"
  90 #define OPT_NO_HUGE     "no-huge"
  91 #define OPT_FILE_PREFIX "file-prefix"
  92 #define OPT_SOCKET_MEM  "socket-mem"
  93 #define OPT_USE_DEVICE  "use-device"
  94 #define OPT_SYSLOG      "syslog"
  95 #define OPT_BASE_VIRTADDR   "base-virtaddr"
  96 #define OPT_XEN_DOM0    "xen-dom0"
  97
  98 #define RTE_EAL_BLACKLIST_SIZE  0x100
  99
 100 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 101
 102 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 103
 104 #define HIGHEST_RPL 3
 105
 106 #define BITS_PER_HEX 4
 107
 108 #define GET_BLACKLIST_FIELD(in, fd, lim, dlm)                   \
 109 {                                                               \
 110         unsigned long val;                                      \
 111         char *end;                                              \
 112         errno = 0;                                              \
 113         val = strtoul((in), &end, 16);                          \
 114         if (errno != 0 || end[0] != (dlm) || val > (lim))       \
 115                 return (-EINVAL);                               \
 116         (fd) = (typeof (fd))val;                                \
 117         (in) = end + 1;                                         \
 118 }
 119
 120 /* Allow the application to print its usage message too if set */
 121 static rte_usage_hook_t rte_application_usage_hook = NULL;
 122
 123 TAILQ_HEAD(shared_driver_list, shared_driver);
 124
 125 /* Definition for shared object drivers. */
 126 struct shared_driver {
 127         TAILQ_ENTRY(shared_driver) next;
 128
 129         char    name[PATH_MAX];
 130         void*   lib_handle;
 131 };
 132
 133 /* List of external loadable drivers */
 134 static struct shared_driver_list solib_list =
 135 TAILQ_HEAD_INITIALIZER(solib_list);
 136
 137 /* early configuration structure, when memory config is not mmapped */
 138 static struct rte_mem_config early_mem_config;
 139
 140 /* define fd variable here, because file needs to be kept open for the
 141  * duration of the program, as we hold a write lock on it in the primary proc */
 142 static int mem_cfg_fd = -1;
 143
 144 static struct flock wr_lock = {
 145                 .l_type = F_WRLCK,
 146                 .l_whence = SEEK_SET,
 147                 .l_start = offsetof(struct rte_mem_config, memseg),
 148                 .l_len = sizeof(early_mem_config.memseg),
 149 };
 150
 151 /* Address of global and public configuration */
 152 static struct rte_config rte_config = {
 153                 .mem_config = &early_mem_config,
 154 };
 155
 156 static struct rte_pci_addr eal_dev_blacklist[RTE_EAL_BLACKLIST_SIZE];
 157
 158 /* internal configuration (per-core) */
 159 struct lcore_config lcore_config[RTE_MAX_LCORE];
 160
 161 /* internal configuration */
 162 struct internal_config internal_config;
 163
 164 /* used by rte_rdtsc() */
 165 int rte_cycles_vmware_tsc_map;
 166
 167 /* Return a pointer to the configuration structure */
 168 struct rte_config *
 169 rte_eal_get_configuration(void)
 170 {
 171         return &rte_config;
 172 }
 173
 174 /* parse a sysfs (or other) file containing one integer value */
 175 int
 176 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 177 {
 178         FILE *f;
 179         char buf[BUFSIZ];
 180         char *end = NULL;
 181
 182         if ((f = fopen(filename, "r")) == NULL) {
 183                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 184                         __func__, filename);
 185                 return -1;
 186         }
 187
 188         if (fgets(buf, sizeof(buf), f) == NULL) {
 189                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 190                         __func__, filename);
 191                 fclose(f);
 192                 return -1;
 193         }
 194         *val = strtoul(buf, &end, 0);
 195         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 196                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 197                                 __func__, filename);
 198                 fclose(f);
 199                 return -1;
 200         }
 201         fclose(f);
 202         return 0;
 203 }
 204
 205
 206 /* create memory configuration in shared/mmap memory. Take out
 207  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 208  * This means we never close the file while running (auto-close on exit).
 209  * We also don't lock the whole file, so that in future we can use read-locks
 210  * on other parts, e.g. memzones, to detect if there are running secondary
 211  * processes. */
 212 static void
 213 rte_eal_config_create(void)
 214 {
 215         void *rte_mem_cfg_addr;
 216         int retval;
 217
 218         const char *pathname = eal_runtime_config_path();
 219
 220         if (internal_config.no_shconf)
 221                 return;
 222
 223         if (mem_cfg_fd < 0){
 224                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 225                 if (mem_cfg_fd < 0)
 226                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 227         }
 228
 229         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 230         if (retval < 0){
 231                 close(mem_cfg_fd);
 232                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 233         }
 234
 235         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 236         if (retval < 0){
 237                 close(mem_cfg_fd);
 238                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 239                                 "process running?\n", pathname);
 240         }
 241
 242         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 243                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 244
 245         if (rte_mem_cfg_addr == MAP_FAILED){
 246                 rte_panic("Cannot mmap memory for rte_config\n");
 247         }
 248         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 249         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 250 }
 251
 252 /* attach to an existing shared memory config */
 253 static void
 254 rte_eal_config_attach(void)
 255 {
 256         void *rte_mem_cfg_addr;
 257         const char *pathname = eal_runtime_config_path();
 258
 259         if (internal_config.no_shconf)
 260                 return;
 261
 262         if (mem_cfg_fd < 0){
 263                 mem_cfg_fd = open(pathname, O_RDWR);
 264                 if (mem_cfg_fd < 0)
 265                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 266         }
 267
 268         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 269                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 270         close(mem_cfg_fd);
 271         if (rte_mem_cfg_addr == MAP_FAILED)
 272                 rte_panic("Cannot mmap memory for rte_config\n");
 273
 274         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 275 }
 276
 277 /* Detect if we are a primary or a secondary process */
 278 static enum rte_proc_type_t
 279 eal_proc_type_detect(void)
 280 {
 281         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 282         const char *pathname = eal_runtime_config_path();
 283
 284         /* if we can open the file but not get a write-lock we are a secondary
 285          * process. NOTE: if we get a file handle back, we keep that open
 286          * and don't close it to prevent a race condition between multiple opens */
 287         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 288                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 289                 ptype = RTE_PROC_SECONDARY;
 290
 291         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 292                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 293
 294         return ptype;
 295 }
 296
 297 /* Sets up rte_config structure with the pointer to shared memory config.*/
 298 static void
 299 rte_config_init(void)
 300 {
 301         /* set the magic in configuration structure */
 302         rte_config.magic = RTE_MAGIC;
 303         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 304                         eal_proc_type_detect() : /* for auto, detect the type */
 305                         internal_config.process_type; /* otherwise use what's already set */
 306
 307         switch (rte_config.process_type){
 308         case RTE_PROC_PRIMARY:
 309                 rte_eal_config_create();
 310                 break;
 311         case RTE_PROC_SECONDARY:
 312                 rte_eal_config_attach();
 313                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 314                 break;
 315         case RTE_PROC_AUTO:
 316         case RTE_PROC_INVALID:
 317                 rte_panic("Invalid process type\n");
 318         }
 319 }
 320
 321 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 322 static void
 323 eal_hugedirs_unlock(void)
 324 {
 325         int i;
 326
 327         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 328         {
 329                 /* skip uninitialized */
 330                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 331                         continue;
 332                 /* unlock hugepage file */
 333                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 334                 close(internal_config.hugepage_info[i].lock_descriptor);
 335                 /* reset the field */
 336                 internal_config.hugepage_info[i].lock_descriptor = -1;
 337         }
 338 }
 339
 340 /* display usage */
 341 static void
 342 eal_usage(const char *prgname)
 343 {
 344         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 345                "[--proc-type primary|secondary|auto] \n\n"
 346                "EAL options:\n"
 347                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 348                "  -n NUM       : Number of memory channels\n"
 349                    "  -v           : Display version information on startup\n"
 350                "  -d LIB.so    : add driver (can be used multiple times)\n"
 351                "  -b <domain:bus:devid.func>: to prevent EAL from using specified "
 352            "PCI device\n"
 353                "                 (multiple -b options are allowed)\n"
 354                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 355                "  -r NUM       : force number of memory ranks (don't detect)\n"
 356                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 357                            "without hugetlbfs\n"
 358                "  --"OPT_SYSLOG"     : set syslog facility\n"
 359                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 360                    "                 sockets (use comma separated values)\n"
 361                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 362                "  --"OPT_PROC_TYPE"  : type of this process\n"
 363                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 364                "  --"OPT_USE_DEVICE": use the specified ethernet device(s) only. "
 365                            "Use comma-separate <[domain:]bus:devid.func> values.\n"
 366                "               [NOTE: Cannot be used with -b option]\n"
 367                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 368                            "native RDTSC\n"
 369                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 370                "\nEAL options for DEBUG use only:\n"
 371                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 372                "  --"OPT_NO_PCI"   : disable pci\n"
 373                "  --"OPT_NO_HPET"  : disable hpet\n"
 374                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 375                "\n",
 376                prgname);
 377         /* Allow the application to print its usage message too if hook is set */
 378         if ( rte_application_usage_hook ) {
 379                 printf("===== Application Usage =====\n\n");
 380                 rte_application_usage_hook(prgname);
 381         }
 382 }
 383
 384 /* Set a per-application usage message */
 385 rte_usage_hook_t
 386 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 387 {
 388         rte_usage_hook_t        old_func;
 389
 390         /* Will be NULL on the first call to denote the last usage routine. */
 391         old_func                                        = rte_application_usage_hook;
 392         rte_application_usage_hook      = usage_func;
 393
 394         return old_func;
 395 }
 396
 397 /*
 398  * Parse the coremask given as argument (hexadecimal string) and fill
 399  * the global configuration (core role and core count) with the parsed
 400  * value.
 401  */
 402 static int xdigit2val(unsigned char c)
 403 {
 404         int val;
 405         if(isdigit(c))
 406                 val = c - '0';
 407         else if(isupper(c))
 408                 val = c - 'A' + 10;
 409         else
 410                 val = c - 'a' + 10;
 411         return val;
 412 }
 413 static int
 414 eal_parse_coremask(const char *coremask)
 415 {
 416         struct rte_config *cfg = rte_eal_get_configuration();
 417         int i, j, idx = 0 ;
 418         unsigned count = 0;
 419         char c;
 420         int val;
 421
 422         if (coremask == NULL)
 423                 return -1;
 424         /* Remove all blank characters ahead and after .
 425          * Remove 0x/0X if exists.
 426          */
 427         while (isblank(*coremask))
 428                 coremask++;
 429         if (coremask[0] == '0' && ((coremask[1] == 'x')
 430                 ||  (coremask[1] == 'X')) )
 431                 coremask += 2;
 432         i = strnlen(coremask, PATH_MAX);
 433         while ((i > 0) && isblank(coremask[i - 1]))
 434                 i--;
 435         if (i == 0)
 436                 return -1;
 437
 438         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 439                 c = coremask[i];
 440                 if (isxdigit(c) == 0) {
 441                         /* invalid characters */
 442                         return (-1);
 443                 }
 444                 val = xdigit2val(c);
 445                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 446                         if((1 << j) & val) {
 447                                 cfg->lcore_role[idx] = ROLE_RTE;
 448                                 if(count == 0)
 449                                         cfg->master_lcore = idx;
 450                                 count++;
 451                         } else  {
 452                                 cfg->lcore_role[idx] = ROLE_OFF;
 453                         }
 454                 }
 455         }
 456         for(; i >= 0; i--)
 457                 if(coremask[i] != '0')
 458                         return -1;
 459         for(; idx < RTE_MAX_LCORE; idx++)
 460                 cfg->lcore_role[idx] = ROLE_OFF;
 461         if(count == 0)
 462                 return -1;
 463         return 0;
 464 }
 465
 466 static int
 467 eal_parse_syslog(const char *facility)
 468 {
 469         int i;
 470         static struct {
 471                 const char *name;
 472                 int value;
 473         } map[] = {
 474                 { "auth", LOG_AUTH },
 475                 { "cron", LOG_CRON },
 476                 { "daemon", LOG_DAEMON },
 477                 { "ftp", LOG_FTP },
 478                 { "kern", LOG_KERN },
 479                 { "lpr", LOG_LPR },
 480                 { "mail", LOG_MAIL },
 481                 { "news", LOG_NEWS },
 482                 { "syslog", LOG_SYSLOG },
 483                 { "user", LOG_USER },
 484                 { "uucp", LOG_UUCP },
 485                 { "local0", LOG_LOCAL0 },
 486                 { "local1", LOG_LOCAL1 },
 487                 { "local2", LOG_LOCAL2 },
 488                 { "local3", LOG_LOCAL3 },
 489                 { "local4", LOG_LOCAL4 },
 490                 { "local5", LOG_LOCAL5 },
 491                 { "local6", LOG_LOCAL6 },
 492                 { "local7", LOG_LOCAL7 },
 493                 { NULL, 0 }
 494         };
 495
 496         for (i = 0; map[i].name; i++) {
 497                 if (!strcmp(facility, map[i].name)) {
 498                         internal_config.syslog_facility = map[i].value;
 499                         return 0;
 500                 }
 501         }
 502         return -1;
 503 }
 504
 505 static int
 506 eal_parse_socket_mem(char *socket_mem)
 507 {
 508         char * arg[RTE_MAX_NUMA_NODES];
 509         char *end;
 510         int arg_num, i, len;
 511         uint64_t total_mem = 0;
 512
 513         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 514         if (len == SOCKET_MEM_STRLEN) {
 515                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 516                 return -1;
 517         }
 518
 519         /* all other error cases will be caught later */
 520         if (!isdigit(socket_mem[len-1]))
 521                 return -1;
 522
 523         /* split the optarg into separate socket values */
 524         arg_num = rte_strsplit(socket_mem, len,
 525                         arg, RTE_MAX_NUMA_NODES, ',');
 526
 527         /* if split failed, or 0 arguments */
 528         if (arg_num <= 0)
 529                 return -1;
 530
 531         internal_config.force_sockets = 1;
 532
 533         /* parse each defined socket option */
 534         errno = 0;
 535         for (i = 0; i < arg_num; i++) {
 536                 end = NULL;
 537                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 538
 539                 /* check for invalid input */
 540                 if ((errno != 0)  ||
 541                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 542                         return -1;
 543                 internal_config.socket_mem[i] *= 1024ULL;
 544                 internal_config.socket_mem[i] *= 1024ULL;
 545                 total_mem += internal_config.socket_mem[i];
 546         }
 547
 548         /* check if we have a positive amount of total memory */
 549         if (total_mem == 0)
 550                 return -1;
 551
 552         return 0;
 553 }
 554
 555 static int
 556 eal_parse_base_virtaddr(const char *arg)
 557 {
 558         char *end;
 559         uint64_t addr;
 560
 561         addr = strtoull(arg, &end, 16);
 562
 563         /* check for errors */
 564         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 565                 return -1;
 566
 567         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 568 #ifndef RTE_ARCH_X86_64
 569         if (addr >= UINTPTR_MAX)
 570                 return -1;
 571 #endif
 572
 573         /* align the addr on 2M boundary */
 574         addr = RTE_PTR_ALIGN_CEIL(addr, RTE_PGSIZE_2M);
 575
 576         internal_config.base_virtaddr = (uintptr_t) addr;
 577         return 0;
 578 }
 579
 580 static inline size_t
 581 eal_get_hugepage_mem_size(void)
 582 {
 583         uint64_t size = 0;
 584         unsigned i, j;
 585
 586         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 587                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 588                 if (hpi->hugedir != NULL) {
 589                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 590                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 591                         }
 592                 }
 593         }
 594
 595         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 596 }
 597
 598 static enum rte_proc_type_t
 599 eal_parse_proc_type(const char *arg)
 600 {
 601         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 602                 return RTE_PROC_PRIMARY;
 603         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 604                 return RTE_PROC_SECONDARY;
 605         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 606                 return RTE_PROC_AUTO;
 607
 608         return RTE_PROC_INVALID;
 609 }
 610
 611 static ssize_t
 612 eal_parse_blacklist_opt(const char *optarg, size_t idx)
 613 {
 614         if (idx >= sizeof (eal_dev_blacklist) / sizeof (eal_dev_blacklist[0])) {
 615                 RTE_LOG(ERR, EAL, "%s - too many devices to blacklist...\n", optarg);
 616                 return (-EINVAL);
 617         } else if (eal_parse_pci_DomBDF(optarg, eal_dev_blacklist + idx) < 0 &&
 618                         eal_parse_pci_BDF(optarg, eal_dev_blacklist + idx) < 0) {
 619                 RTE_LOG(ERR, EAL, "%s - invalid device to blacklist...\n", optarg);
 620                 return (-EINVAL);
 621         }
 622
 623         idx += 1;
 624         return (idx);
 625 }
 626
 627 /* Parse the argument given in the command line of the application */
 628 static int
 629 eal_parse_args(int argc, char **argv)
 630 {
 631         int opt, ret, i;
 632         char **argvopt;
 633         int option_index;
 634         int coremask_ok = 0;
 635         ssize_t blacklist_index = 0;
 636         char *prgname = argv[0];
 637         static struct option lgopts[] = {
 638                 {OPT_NO_HUGE, 0, 0, 0},
 639                 {OPT_NO_PCI, 0, 0, 0},
 640                 {OPT_NO_HPET, 0, 0, 0},
 641                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 642                 {OPT_HUGE_DIR, 1, 0, 0},
 643                 {OPT_NO_SHCONF, 0, 0, 0},
 644                 {OPT_PROC_TYPE, 1, 0, 0},
 645                 {OPT_FILE_PREFIX, 1, 0, 0},
 646                 {OPT_SOCKET_MEM, 1, 0, 0},
 647                 {OPT_USE_DEVICE, 1, 0, 0},
 648                 {OPT_SYSLOG, 1, NULL, 0},
 649                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 650                 {OPT_XEN_DOM0, 0, 0, 0},
 651                 {0, 0, 0, 0}
 652         };
 653         struct shared_driver *solib;
 654
 655         argvopt = argv;
 656
 657         internal_config.memory = 0;
 658         internal_config.force_nrank = 0;
 659         internal_config.force_nchannel = 0;
 660         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 661         internal_config.hugepage_dir = NULL;
 662         internal_config.force_sockets = 0;
 663         internal_config.syslog_facility = LOG_DAEMON;
 664         internal_config.xen_dom0_support = 0;
 665 #ifdef RTE_LIBEAL_USE_HPET
 666         internal_config.no_hpet = 0;
 667 #else
 668         internal_config.no_hpet = 1;
 669 #endif
 670         /* zero out the NUMA config */
 671         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 672                 internal_config.socket_mem[i] = 0;
 673
 674         /* zero out hugedir descriptors */
 675         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 676                 internal_config.hugepage_info[i].lock_descriptor = -1;
 677
 678         internal_config.vmware_tsc_map = 0;
 679         internal_config.base_virtaddr = 0;
 680
 681         while ((opt = getopt_long(argc, argvopt, "b:c:d:m:n:r:v",
 682                                   lgopts, &option_index)) != EOF) {
 683
 684                 switch (opt) {
 685                 /* blacklist */
 686                 case 'b':
 687                         if ((blacklist_index = eal_parse_blacklist_opt(optarg,
 688                             blacklist_index)) < 0) {
 689                                 eal_usage(prgname);
 690                                 return (-1);
 691                         }
 692                         break;
 693                 /* coremask */
 694                 case 'c':
 695                         if (eal_parse_coremask(optarg) < 0) {
 696                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 697                                 eal_usage(prgname);
 698                                 return -1;
 699                         }
 700                         coremask_ok = 1;
 701                         break;
 702                 /* force loading of external driver */
 703                 case 'd':
 704                         solib = malloc(sizeof(*solib));
 705                         if (solib == NULL) {
 706                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 707                                 return -1;
 708                         }
 709                         memset(solib, 0, sizeof(*solib));
 710                         strncpy(solib->name, optarg, PATH_MAX-1);
 711                         solib->name[PATH_MAX-1] = 0;
 712                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 713                         break;
 714                 /* size of memory */
 715                 case 'm':
 716                         internal_config.memory = atoi(optarg);
 717                         internal_config.memory *= 1024ULL;
 718                         internal_config.memory *= 1024ULL;
 719                         break;
 720                 /* force number of channels */
 721                 case 'n':
 722                         internal_config.force_nchannel = atoi(optarg);
 723                         if (internal_config.force_nchannel == 0 ||
 724                             internal_config.force_nchannel > 4) {
 725                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 726                                 eal_usage(prgname);
 727                                 return -1;
 728                         }
 729                         break;
 730                 /* force number of ranks */
 731                 case 'r':
 732                         internal_config.force_nrank = atoi(optarg);
 733                         if (internal_config.force_nrank == 0 ||
 734                             internal_config.force_nrank > 16) {
 735                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 736                                 eal_usage(prgname);
 737                                 return -1;
 738                         }
 739                         break;
 740                 case 'v':
 741                         /* since message is explicitly requested by user, we
 742                          * write message at highest log level so it can always be seen
 743                          * even if info or warning messages are disabled */
 744                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 745                         break;
 746
 747                 /* long options */
 748                 case 0:
 749                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 750                                 internal_config.no_hugetlbfs = 1;
 751                         }
 752                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 753                 #ifdef RTE_LIBRTE_XEN_DOM0
 754                                 internal_config.xen_dom0_support = 1;
 755                 #else
 756                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 757                                         "running on Dom0, please configure"
 758                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 759                                 return -1;
 760                 #endif
 761                         }
 762                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 763                                 internal_config.no_pci = 1;
 764                         }
 765                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 766                                 internal_config.no_hpet = 1;
 767                         }
 768                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 769                                 internal_config.vmware_tsc_map = 1;
 770                         }
 771                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 772                                 internal_config.no_shconf = 1;
 773                         }
 774                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 775                                 internal_config.hugepage_dir = optarg;
 776                         }
 777                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 778                                 internal_config.process_type = eal_parse_proc_type(optarg);
 779                         }
 780                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 781                                 internal_config.hugefile_prefix = optarg;
 782                         }
 783                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 784                                 if (eal_parse_socket_mem(optarg) < 0) {
 785                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 786                                                         OPT_SOCKET_MEM "\n");
 787                                         eal_usage(prgname);
 788                                         return -1;
 789                                 }
 790                         }
 791                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 792                                 eal_dev_whitelist_add_entry(optarg);
 793                         }
 794                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 795                                 if (eal_parse_syslog(optarg) < 0) {
 796                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 797                                                         OPT_SYSLOG "\n");
 798                                         eal_usage(prgname);
 799                                         return -1;
 800                                 }
 801                         }
 802                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 803                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 804                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 805                                                         OPT_BASE_VIRTADDR "\n");
 806                                         eal_usage(prgname);
 807                                         return -1;
 808                                 }
 809                         }
 810                         break;
 811
 812                 default:
 813                         eal_usage(prgname);
 814                         return -1;
 815                 }
 816         }
 817
 818         /* sanity checks */
 819         if (!coremask_ok) {
 820                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 821                 eal_usage(prgname);
 822                 return -1;
 823         }
 824         if (internal_config.process_type == RTE_PROC_AUTO){
 825                 internal_config.process_type = eal_proc_type_detect();
 826         }
 827         if (internal_config.process_type == RTE_PROC_INVALID){
 828                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 829                 eal_usage(prgname);
 830                 return -1;
 831         }
 832         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 833                         internal_config.force_nchannel == 0) {
 834                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 835                 eal_usage(prgname);
 836                 return -1;
 837         }
 838         if (index(internal_config.hugefile_prefix,'%') != NULL){
 839                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 840                 eal_usage(prgname);
 841                 return -1;
 842         }
 843         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 844                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 845                                 "at the same time\n");
 846                 eal_usage(prgname);
 847                 return -1;
 848         }
 849         /* --no-huge doesn't make sense with either -m or --socket-mem */
 850         if (internal_config.no_hugetlbfs &&
 851                         (internal_config.memory > 0 ||
 852                                         internal_config.force_sockets == 1)) {
 853                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 854                                 "together with --no-huge!\n");
 855                 eal_usage(prgname);
 856                 return -1;
 857         }
 858         /* --xen-dom0 doesn't make sense with --socket-mem */
 859         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 860                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 861                                         "together with --xen_dom0!\n");
 862                 eal_usage(prgname);
 863                 return -1;
 864         }
 865         /* if no blacklist, parse a whitelist */
 866         if (blacklist_index > 0) {
 867                 if (eal_dev_whitelist_exists()) {
 868                         RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 869                                         "[--use-device] options cannot be used at the same time\n");
 870                         eal_usage(prgname);
 871                         return -1;
 872                 }
 873                 rte_eal_pci_set_blacklist(eal_dev_blacklist, blacklist_index);
 874         } else {
 875                 if (eal_dev_whitelist_exists() && eal_dev_whitelist_parse() < 0) {
 876                         RTE_LOG(ERR,EAL, "Error parsing whitelist[--use-device] options\n");
 877                         return -1;
 878                 }
 879         }
 880
 881         if (optind >= 0)
 882                 argv[optind-1] = prgname;
 883
 884         /* if no memory amounts were requested, this will result in 0 and
 885          * will be overriden later, right after eal_hugepage_info_init() */
 886         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 887                 internal_config.memory += internal_config.socket_mem[i];
 888
 889         ret = optind-1;
 890         optind = 0; /* reset getopt lib */
 891         return ret;
 892 }
 893
 894 static void
 895 eal_check_mem_on_local_socket(void)
 896 {
 897         const struct rte_memseg *ms;
 898         int i, socket_id;
 899
 900         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 901
 902         ms = rte_eal_get_physmem_layout();
 903
 904         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 905                 if (ms[i].socket_id == socket_id &&
 906                                 ms[i].len > 0)
 907                         return;
 908
 909         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 910                         "memory on local socket!\n");
 911 }
 912
 913 static int
 914 sync_func(__attribute__((unused)) void *arg)
 915 {
 916         return 0;
 917 }
 918
 919 inline static void
 920 rte_eal_mcfg_complete(void)
 921 {
 922         /* ALL shared mem_config related INIT DONE */
 923         if (rte_config.process_type == RTE_PROC_PRIMARY)
 924                 rte_config.mem_config->magic = RTE_MAGIC;
 925 }
 926
 927 /*
 928  * Request iopl priviledge for all RPL, returns 0 on success
 929  */
 930 static int
 931 rte_eal_iopl_init(void)
 932 {
 933         return iopl(HIGHEST_RPL);
 934 }
 935
 936 /* Launch threads, called at application init(). */
 937 int
 938 rte_eal_init(int argc, char **argv)
 939 {
 940         int i, fctret, ret;
 941         pthread_t thread_id;
 942         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 943         struct shared_driver *solib = NULL;
 944
 945         if (!rte_atomic32_test_and_set(&run_once))
 946                 return -1;
 947
 948         thread_id = pthread_self();
 949
 950         if (rte_eal_log_early_init() < 0)
 951                 rte_panic("Cannot init early logs\n");
 952
 953         fctret = eal_parse_args(argc, argv);
 954         if (fctret < 0)
 955                 exit(1);
 956
 957         if (internal_config.no_hugetlbfs == 0 &&
 958                         internal_config.process_type != RTE_PROC_SECONDARY &&
 959                         internal_config.xen_dom0_support == 0 &&
 960                         eal_hugepage_info_init() < 0)
 961                 rte_panic("Cannot get hugepage information\n");
 962
 963         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 964                 if (internal_config.no_hugetlbfs)
 965                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 966                 else
 967                         internal_config.memory = eal_get_hugepage_mem_size();
 968         }
 969
 970         if (internal_config.vmware_tsc_map == 1) {
 971 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 972                 rte_cycles_vmware_tsc_map = 1;
 973                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 974                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 975 #else
 976                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 977                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 978 #endif
 979         }
 980
 981         rte_srand(rte_rdtsc());
 982
 983         rte_config_init();
 984
 985         if (rte_eal_iopl_init() == 0)
 986                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 987
 988         if (rte_eal_cpu_init() < 0)
 989                 rte_panic("Cannot detect lcores\n");
 990
 991         if (rte_eal_pci_init() < 0)
 992                 rte_panic("Cannot init PCI\n");
 993
 994 #ifdef RTE_LIBRTE_IVSHMEM
 995         if (rte_eal_ivshmem_init() < 0)
 996                 rte_panic("Cannot init IVSHMEM\n");
 997 #endif
 998
 999         if (rte_eal_memory_init() < 0)
1000                 rte_panic("Cannot init memory\n");
1001
1002         /* the directories are locked during eal_hugepage_info_init */
1003         eal_hugedirs_unlock();
1004
1005         if (rte_eal_memzone_init() < 0)
1006                 rte_panic("Cannot init memzone\n");
1007
1008         if (rte_eal_tailqs_init() < 0)
1009                 rte_panic("Cannot init tail queues for objects\n");
1010
1011 #ifdef RTE_LIBRTE_IVSHMEM
1012         if (rte_eal_ivshmem_obj_init() < 0)
1013                 rte_panic("Cannot init IVSHMEM objects\n");
1014 #endif
1015
1016         if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
1017                 rte_panic("Cannot init logs\n");
1018
1019         if (rte_eal_alarm_init() < 0)
1020                 rte_panic("Cannot init interrupt-handling thread\n");
1021
1022         if (rte_eal_intr_init() < 0)
1023                 rte_panic("Cannot init interrupt-handling thread\n");
1024
1025         if (rte_eal_timer_init() < 0)
1026                 rte_panic("Cannot init HPET or TSC timers\n");
1027
1028         eal_check_mem_on_local_socket();
1029
1030         rte_eal_mcfg_complete();
1031
1032         if (rte_eal_non_pci_ethdev_init() < 0)
1033                 rte_panic("Cannot init non-PCI eth_devs\n");
1034
1035         TAILQ_FOREACH(solib, &solib_list, next) {
1036                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1037                 if ((solib->lib_handle == NULL) && (solib->name[0] != '/')) {
1038                         /* relative path: try again with "./" prefix */
1039                         char sopath[PATH_MAX];
1040                         snprintf(sopath, sizeof(sopath), "./%s", solib->name);
1041                         solib->lib_handle = dlopen(sopath, RTLD_NOW);
1042                 }
1043                 if (solib->lib_handle == NULL)
1044                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1045         }
1046
1047         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1048                 rte_config.master_lcore, (int)thread_id);
1049
1050         RTE_LCORE_FOREACH_SLAVE(i) {
1051
1052                 /*
1053                  * create communication pipes between master thread
1054                  * and children
1055                  */
1056                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1057                         rte_panic("Cannot create pipe\n");
1058                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1059                         rte_panic("Cannot create pipe\n");
1060
1061                 lcore_config[i].state = WAIT;
1062
1063                 /* create a thread for each lcore */
1064                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1065                                      eal_thread_loop, NULL);
1066                 if (ret != 0)
1067                         rte_panic("Cannot create thread\n");
1068         }
1069
1070         eal_thread_init_master(rte_config.master_lcore);
1071
1072         /*
1073          * Launch a dummy function on all slave lcores, so that master lcore
1074          * knows they are all ready when this function returns.
1075          */
1076         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1077         rte_eal_mp_wait_lcore();
1078
1079         return fctret;
1080 }
1081
1082 /* get core role */
1083 enum rte_lcore_role_t
1084 rte_eal_lcore_role(unsigned lcore_id)
1085 {
1086         return (rte_config.lcore_role[lcore_id]);
1087 }
1088
1089 enum rte_proc_type_t
1090 rte_eal_process_type(void)
1091 {
1092         return (rte_config.process_type);
1093 }
1094