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         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 293                         eal_proc_type_detect() : /* for auto, detect the type */
 294                         internal_config.process_type; /* otherwise use what's already set */
 295
 296         switch (rte_config.process_type){
 297         case RTE_PROC_PRIMARY:
 298                 rte_eal_config_create();
 299                 break;
 300         case RTE_PROC_SECONDARY:
 301                 rte_eal_config_attach();
 302                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 303                 break;
 304         case RTE_PROC_AUTO:
 305         case RTE_PROC_INVALID:
 306                 rte_panic("Invalid process type\n");
 307         }
 308 }
 309
 310 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 311 static void
 312 eal_hugedirs_unlock(void)
 313 {
 314         int i;
 315
 316         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 317         {
 318                 /* skip uninitialized */
 319                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 320                         continue;
 321                 /* unlock hugepage file */
 322                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 323                 close(internal_config.hugepage_info[i].lock_descriptor);
 324                 /* reset the field */
 325                 internal_config.hugepage_info[i].lock_descriptor = -1;
 326         }
 327 }
 328
 329 /* display usage */
 330 static void
 331 eal_usage(const char *prgname)
 332 {
 333         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 334                "[--proc-type primary|secondary|auto] \n\n"
 335                "EAL options:\n"
 336                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 337                "  -n NUM       : Number of memory channels\n"
 338                    "  -v           : Display version information on startup\n"
 339                "  -d LIB.so    : add driver (can be used multiple times)\n"
 340                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 341                "  -r NUM       : force number of memory ranks (don't detect)\n"
 342                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 343                            "without hugetlbfs\n"
 344                "  --"OPT_SYSLOG"     : set syslog facility\n"
 345                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 346                    "                 sockets (use comma separated values)\n"
 347                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 348                "  --"OPT_PROC_TYPE"  : type of this process\n"
 349                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 350                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 351                "               Prevent EAL from using this PCI device. The argument\n"
 352                "               format is <domain:bus:devid.func>.\n"
 353                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 354                "               Only use the specified PCI devices. The argument format\n"
 355                "               is <[domain:]bus:devid.func>. This option can be present\n"
 356                "               several times (once per device).\n"
 357                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 358                "  --"OPT_VDEV": add a virtual device.\n"
 359                "               The argument format is <driver><id>[,key=val,...]\n"
 360                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 361                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of "
 362                            "native RDTSC\n"
 363                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 364                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 365                "\nEAL options for DEBUG use only:\n"
 366                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 367                "  --"OPT_NO_PCI"   : disable pci\n"
 368                "  --"OPT_NO_HPET"  : disable hpet\n"
 369                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 370                "\n",
 371                prgname);
 372         /* Allow the application to print its usage message too if hook is set */
 373         if ( rte_application_usage_hook ) {
 374                 printf("===== Application Usage =====\n\n");
 375                 rte_application_usage_hook(prgname);
 376         }
 377 }
 378
 379 /* Set a per-application usage message */
 380 rte_usage_hook_t
 381 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 382 {
 383         rte_usage_hook_t        old_func;
 384
 385         /* Will be NULL on the first call to denote the last usage routine. */
 386         old_func                                        = rte_application_usage_hook;
 387         rte_application_usage_hook      = usage_func;
 388
 389         return old_func;
 390 }
 391
 392 /*
 393  * Parse the coremask given as argument (hexadecimal string) and fill
 394  * the global configuration (core role and core count) with the parsed
 395  * value.
 396  */
 397 static int xdigit2val(unsigned char c)
 398 {
 399         int val;
 400         if(isdigit(c))
 401                 val = c - '0';
 402         else if(isupper(c))
 403                 val = c - 'A' + 10;
 404         else
 405                 val = c - 'a' + 10;
 406         return val;
 407 }
 408 static int
 409 eal_parse_coremask(const char *coremask)
 410 {
 411         struct rte_config *cfg = rte_eal_get_configuration();
 412         int i, j, idx = 0 ;
 413         unsigned count = 0;
 414         char c;
 415         int val;
 416
 417         if (coremask == NULL)
 418                 return -1;
 419         /* Remove all blank characters ahead and after .
 420          * Remove 0x/0X if exists.
 421          */
 422         while (isblank(*coremask))
 423                 coremask++;
 424         if (coremask[0] == '0' && ((coremask[1] == 'x')
 425                 ||  (coremask[1] == 'X')) )
 426                 coremask += 2;
 427         i = strnlen(coremask, PATH_MAX);
 428         while ((i > 0) && isblank(coremask[i - 1]))
 429                 i--;
 430         if (i == 0)
 431                 return -1;
 432
 433         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 434                 c = coremask[i];
 435                 if (isxdigit(c) == 0) {
 436                         /* invalid characters */
 437                         return (-1);
 438                 }
 439                 val = xdigit2val(c);
 440                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 441                         if((1 << j) & val) {
 442                                 if (!lcore_config[idx].detected) {
 443                                         RTE_LOG(ERR, EAL, "lcore %u "
 444                                                 "unavailable\n", idx);
 445                                         return -1;
 446                                 }
 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         /* Update the count of enabled logical cores of the EAL configuration */
 464         cfg->lcore_count = count;
 465         return 0;
 466 }
 467
 468 static int
 469 eal_parse_syslog(const char *facility)
 470 {
 471         int i;
 472         static struct {
 473                 const char *name;
 474                 int value;
 475         } map[] = {
 476                 { "auth", LOG_AUTH },
 477                 { "cron", LOG_CRON },
 478                 { "daemon", LOG_DAEMON },
 479                 { "ftp", LOG_FTP },
 480                 { "kern", LOG_KERN },
 481                 { "lpr", LOG_LPR },
 482                 { "mail", LOG_MAIL },
 483                 { "news", LOG_NEWS },
 484                 { "syslog", LOG_SYSLOG },
 485                 { "user", LOG_USER },
 486                 { "uucp", LOG_UUCP },
 487                 { "local0", LOG_LOCAL0 },
 488                 { "local1", LOG_LOCAL1 },
 489                 { "local2", LOG_LOCAL2 },
 490                 { "local3", LOG_LOCAL3 },
 491                 { "local4", LOG_LOCAL4 },
 492                 { "local5", LOG_LOCAL5 },
 493                 { "local6", LOG_LOCAL6 },
 494                 { "local7", LOG_LOCAL7 },
 495                 { NULL, 0 }
 496         };
 497
 498         for (i = 0; map[i].name; i++) {
 499                 if (!strcmp(facility, map[i].name)) {
 500                         internal_config.syslog_facility = map[i].value;
 501                         return 0;
 502                 }
 503         }
 504         return -1;
 505 }
 506
 507 static int
 508 eal_parse_socket_mem(char *socket_mem)
 509 {
 510         char * arg[RTE_MAX_NUMA_NODES];
 511         char *end;
 512         int arg_num, i, len;
 513         uint64_t total_mem = 0;
 514
 515         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 516         if (len == SOCKET_MEM_STRLEN) {
 517                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 518                 return -1;
 519         }
 520
 521         /* all other error cases will be caught later */
 522         if (!isdigit(socket_mem[len-1]))
 523                 return -1;
 524
 525         /* split the optarg into separate socket values */
 526         arg_num = rte_strsplit(socket_mem, len,
 527                         arg, RTE_MAX_NUMA_NODES, ',');
 528
 529         /* if split failed, or 0 arguments */
 530         if (arg_num <= 0)
 531                 return -1;
 532
 533         internal_config.force_sockets = 1;
 534
 535         /* parse each defined socket option */
 536         errno = 0;
 537         for (i = 0; i < arg_num; i++) {
 538                 end = NULL;
 539                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 540
 541                 /* check for invalid input */
 542                 if ((errno != 0)  ||
 543                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 544                         return -1;
 545                 internal_config.socket_mem[i] *= 1024ULL;
 546                 internal_config.socket_mem[i] *= 1024ULL;
 547                 total_mem += internal_config.socket_mem[i];
 548         }
 549
 550         /* check if we have a positive amount of total memory */
 551         if (total_mem == 0)
 552                 return -1;
 553
 554         return 0;
 555 }
 556
 557 static int
 558 eal_parse_base_virtaddr(const char *arg)
 559 {
 560         char *end;
 561         uint64_t addr;
 562
 563         addr = strtoull(arg, &end, 16);
 564
 565         /* check for errors */
 566         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 567                 return -1;
 568
 569         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 570 #ifndef RTE_ARCH_X86_64
 571         if (addr >= UINTPTR_MAX)
 572                 return -1;
 573 #endif
 574
 575         /* align the addr on 2M boundary */
 576         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 577                                                            RTE_PGSIZE_2M);
 578
 579         return 0;
 580 }
 581
 582 static inline size_t
 583 eal_get_hugepage_mem_size(void)
 584 {
 585         uint64_t size = 0;
 586         unsigned i, j;
 587
 588         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 589                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 590                 if (hpi->hugedir != NULL) {
 591                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 592                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 593                         }
 594                 }
 595         }
 596
 597         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 598 }
 599
 600 static enum rte_proc_type_t
 601 eal_parse_proc_type(const char *arg)
 602 {
 603         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 604                 return RTE_PROC_PRIMARY;
 605         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 606                 return RTE_PROC_SECONDARY;
 607         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 608                 return RTE_PROC_AUTO;
 609
 610         return RTE_PROC_INVALID;
 611 }
 612
 613 /* Parse the argument given in the command line of the application */
 614 static int
 615 eal_parse_args(int argc, char **argv)
 616 {
 617         int opt, ret, i;
 618         char **argvopt;
 619         int option_index;
 620         int coremask_ok = 0;
 621         char *prgname = argv[0];
 622         static struct option lgopts[] = {
 623                 {OPT_NO_HUGE, 0, 0, 0},
 624                 {OPT_NO_PCI, 0, 0, 0},
 625                 {OPT_NO_HPET, 0, 0, 0},
 626                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 627                 {OPT_HUGE_DIR, 1, 0, 0},
 628                 {OPT_NO_SHCONF, 0, 0, 0},
 629                 {OPT_PROC_TYPE, 1, 0, 0},
 630                 {OPT_FILE_PREFIX, 1, 0, 0},
 631                 {OPT_SOCKET_MEM, 1, 0, 0},
 632                 {OPT_PCI_WHITELIST, 1, 0, 0},
 633                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 634                 {OPT_VDEV, 1, 0, 0},
 635                 {OPT_SYSLOG, 1, NULL, 0},
 636                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 637                 {OPT_XEN_DOM0, 0, 0, 0},
 638                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 639                 {0, 0, 0, 0}
 640         };
 641         struct shared_driver *solib;
 642
 643         argvopt = argv;
 644
 645         internal_config.memory = 0;
 646         internal_config.force_nrank = 0;
 647         internal_config.force_nchannel = 0;
 648         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 649         internal_config.hugepage_dir = NULL;
 650         internal_config.force_sockets = 0;
 651         internal_config.syslog_facility = LOG_DAEMON;
 652         internal_config.xen_dom0_support = 0;
 653 #ifdef RTE_LIBEAL_USE_HPET
 654         internal_config.no_hpet = 0;
 655 #else
 656         internal_config.no_hpet = 1;
 657 #endif
 658         /* zero out the NUMA config */
 659         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 660                 internal_config.socket_mem[i] = 0;
 661
 662         /* zero out hugedir descriptors */
 663         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 664                 internal_config.hugepage_info[i].lock_descriptor = -1;
 665
 666         internal_config.vmware_tsc_map = 0;
 667         internal_config.base_virtaddr = 0;
 668
 669         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
 670                                   lgopts, &option_index)) != EOF) {
 671
 672                 switch (opt) {
 673                 /* blacklist */
 674                 case 'b':
 675                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 676                                         optarg) < 0) {
 677                                 eal_usage(prgname);
 678                                 return (-1);
 679                         }
 680                         break;
 681                 /* whitelist */
 682                 case 'w':
 683                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 684                                         optarg) < 0) {
 685                                 eal_usage(prgname);
 686                                 return -1;
 687                         }
 688                         break;
 689                 /* coremask */
 690                 case 'c':
 691                         if (eal_parse_coremask(optarg) < 0) {
 692                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 693                                 eal_usage(prgname);
 694                                 return -1;
 695                         }
 696                         coremask_ok = 1;
 697                         break;
 698                 /* force loading of external driver */
 699                 case 'd':
 700                         solib = malloc(sizeof(*solib));
 701                         if (solib == NULL) {
 702                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 703                                 return -1;
 704                         }
 705                         memset(solib, 0, sizeof(*solib));
 706                         strncpy(solib->name, optarg, PATH_MAX-1);
 707                         solib->name[PATH_MAX-1] = 0;
 708                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 709                         break;
 710                 /* size of memory */
 711                 case 'm':
 712                         internal_config.memory = atoi(optarg);
 713                         internal_config.memory *= 1024ULL;
 714                         internal_config.memory *= 1024ULL;
 715                         break;
 716                 /* force number of channels */
 717                 case 'n':
 718                         internal_config.force_nchannel = atoi(optarg);
 719                         if (internal_config.force_nchannel == 0 ||
 720                             internal_config.force_nchannel > 4) {
 721                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 722                                 eal_usage(prgname);
 723                                 return -1;
 724                         }
 725                         break;
 726                 /* force number of ranks */
 727                 case 'r':
 728                         internal_config.force_nrank = atoi(optarg);
 729                         if (internal_config.force_nrank == 0 ||
 730                             internal_config.force_nrank > 16) {
 731                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 732                                 eal_usage(prgname);
 733                                 return -1;
 734                         }
 735                         break;
 736                 case 'v':
 737                         /* since message is explicitly requested by user, we
 738                          * write message at highest log level so it can always be seen
 739                          * even if info or warning messages are disabled */
 740                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 741                         break;
 742
 743                 /* long options */
 744                 case 0:
 745                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 746                                 internal_config.no_hugetlbfs = 1;
 747                         }
 748                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 749                 #ifdef RTE_LIBRTE_XEN_DOM0
 750                                 internal_config.xen_dom0_support = 1;
 751                 #else
 752                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 753                                         "running on Dom0, please configure"
 754                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 755                                 return -1;
 756                 #endif
 757                         }
 758                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 759                                 internal_config.no_pci = 1;
 760                         }
 761                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 762                                 internal_config.no_hpet = 1;
 763                         }
 764                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 765                                 internal_config.vmware_tsc_map = 1;
 766                         }
 767                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 768                                 internal_config.no_shconf = 1;
 769                         }
 770                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 771                                 internal_config.hugepage_dir = optarg;
 772                         }
 773                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 774                                 internal_config.process_type = eal_parse_proc_type(optarg);
 775                         }
 776                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 777                                 internal_config.hugefile_prefix = optarg;
 778                         }
 779                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 780                                 if (eal_parse_socket_mem(optarg) < 0) {
 781                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 782                                                         OPT_SOCKET_MEM "\n");
 783                                         eal_usage(prgname);
 784                                         return -1;
 785                                 }
 786                         }
 787                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 788                                 printf("The --use-device option is deprecated, please use\n"
 789                                         "--whitelist or --vdev instead.\n");
 790                                 eal_usage(prgname);
 791                                 return -1;
 792                         }
 793                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 794                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 795                                                 optarg) < 0) {
 796                                         eal_usage(prgname);
 797                                         return -1;
 798                                 }
 799                         }
 800                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 801                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 802                                                 optarg) < 0) {
 803                                         eal_usage(prgname);
 804                                         return -1;
 805                                 }
 806                         }
 807                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 808                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 809                                                 optarg) < 0) {
 810                                         eal_usage(prgname);
 811                                         return -1;
 812                                 }
 813                         }
 814                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 815                                 if (eal_parse_syslog(optarg) < 0) {
 816                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 817                                                         OPT_SYSLOG "\n");
 818                                         eal_usage(prgname);
 819                                         return -1;
 820                                 }
 821                         }
 822                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 823                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 824                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 825                                                         OPT_BASE_VIRTADDR "\n");
 826                                         eal_usage(prgname);
 827                                         return -1;
 828                                 }
 829                         }
 830                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 831                                 internal_config.create_uio_dev = 1;
 832                         }
 833                         break;
 834
 835                 default:
 836                         eal_usage(prgname);
 837                         return -1;
 838                 }
 839         }
 840
 841         /* sanity checks */
 842         if (!coremask_ok) {
 843                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 844                 eal_usage(prgname);
 845                 return -1;
 846         }
 847         if (internal_config.process_type == RTE_PROC_AUTO){
 848                 internal_config.process_type = eal_proc_type_detect();
 849         }
 850         if (internal_config.process_type == RTE_PROC_INVALID){
 851                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 852                 eal_usage(prgname);
 853                 return -1;
 854         }
 855         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 856                         internal_config.force_nchannel == 0) {
 857                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 858                 eal_usage(prgname);
 859                 return -1;
 860         }
 861         if (index(internal_config.hugefile_prefix,'%') != NULL){
 862                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 863                 eal_usage(prgname);
 864                 return -1;
 865         }
 866         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 867                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 868                                 "at the same time\n");
 869                 eal_usage(prgname);
 870                 return -1;
 871         }
 872         /* --no-huge doesn't make sense with either -m or --socket-mem */
 873         if (internal_config.no_hugetlbfs &&
 874                         (internal_config.memory > 0 ||
 875                                         internal_config.force_sockets == 1)) {
 876                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 877                                 "together with --no-huge!\n");
 878                 eal_usage(prgname);
 879                 return -1;
 880         }
 881         /* --xen-dom0 doesn't make sense with --socket-mem */
 882         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 883                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 884                                         "together with --xen_dom0!\n");
 885                 eal_usage(prgname);
 886                 return -1;
 887         }
 888
 889         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 890                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 891                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 892                         "[-w] options cannot be used at the same time\n");
 893                 eal_usage(prgname);
 894                 return -1;
 895         }
 896
 897         if (optind >= 0)
 898                 argv[optind-1] = prgname;
 899
 900         /* if no memory amounts were requested, this will result in 0 and
 901          * will be overriden later, right after eal_hugepage_info_init() */
 902         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 903                 internal_config.memory += internal_config.socket_mem[i];
 904
 905         ret = optind-1;
 906         optind = 0; /* reset getopt lib */
 907         return ret;
 908 }
 909
 910 static void
 911 eal_check_mem_on_local_socket(void)
 912 {
 913         const struct rte_memseg *ms;
 914         int i, socket_id;
 915
 916         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 917
 918         ms = rte_eal_get_physmem_layout();
 919
 920         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 921                 if (ms[i].socket_id == socket_id &&
 922                                 ms[i].len > 0)
 923                         return;
 924
 925         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 926                         "memory on local socket!\n");
 927 }
 928
 929 static int
 930 sync_func(__attribute__((unused)) void *arg)
 931 {
 932         return 0;
 933 }
 934
 935 inline static void
 936 rte_eal_mcfg_complete(void)
 937 {
 938         /* ALL shared mem_config related INIT DONE */
 939         if (rte_config.process_type == RTE_PROC_PRIMARY)
 940                 rte_config.mem_config->magic = RTE_MAGIC;
 941 }
 942
 943 /*
 944  * Request iopl priviledge for all RPL, returns 0 on success
 945  */
 946 static int
 947 rte_eal_iopl_init(void)
 948 {
 949         return iopl(HIGHEST_RPL);
 950 }
 951
 952 /* Launch threads, called at application init(). */
 953 int
 954 rte_eal_init(int argc, char **argv)
 955 {
 956         int i, fctret, ret;
 957         pthread_t thread_id;
 958         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 959         struct shared_driver *solib = NULL;
 960         const char *logid;
 961
 962         if (!rte_atomic32_test_and_set(&run_once))
 963                 return -1;
 964
 965         logid = strrchr(argv[0], '/');
 966         logid = strdup(logid ? logid + 1: argv[0]);
 967
 968         thread_id = pthread_self();
 969
 970         if (rte_eal_log_early_init() < 0)
 971                 rte_panic("Cannot init early logs\n");
 972
 973         if (rte_eal_cpu_init() < 0)
 974                 rte_panic("Cannot detect lcores\n");
 975
 976         fctret = eal_parse_args(argc, argv);
 977         if (fctret < 0)
 978                 exit(1);
 979
 980         if (internal_config.no_hugetlbfs == 0 &&
 981                         internal_config.process_type != RTE_PROC_SECONDARY &&
 982                         internal_config.xen_dom0_support == 0 &&
 983                         eal_hugepage_info_init() < 0)
 984                 rte_panic("Cannot get hugepage information\n");
 985
 986         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 987                 if (internal_config.no_hugetlbfs)
 988                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 989                 else
 990                         internal_config.memory = eal_get_hugepage_mem_size();
 991         }
 992
 993         if (internal_config.vmware_tsc_map == 1) {
 994 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 995                 rte_cycles_vmware_tsc_map = 1;
 996                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 997                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 998 #else
 999                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
1000                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1001 #endif
1002         }
1003
1004         rte_srand(rte_rdtsc());
1005
1006         rte_config_init();
1007
1008         if (rte_eal_iopl_init() == 0)
1009                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1010
1011         if (rte_eal_pci_init() < 0)
1012                 rte_panic("Cannot init PCI\n");
1013
1014 #ifdef RTE_LIBRTE_IVSHMEM
1015         if (rte_eal_ivshmem_init() < 0)
1016                 rte_panic("Cannot init IVSHMEM\n");
1017 #endif
1018
1019         if (rte_eal_memory_init() < 0)
1020                 rte_panic("Cannot init memory\n");
1021
1022         /* the directories are locked during eal_hugepage_info_init */
1023         eal_hugedirs_unlock();
1024
1025         if (rte_eal_memzone_init() < 0)
1026                 rte_panic("Cannot init memzone\n");
1027
1028         if (rte_eal_tailqs_init() < 0)
1029                 rte_panic("Cannot init tail queues for objects\n");
1030
1031 #ifdef RTE_LIBRTE_IVSHMEM
1032         if (rte_eal_ivshmem_obj_init() < 0)
1033                 rte_panic("Cannot init IVSHMEM objects\n");
1034 #endif
1035
1036         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1037                 rte_panic("Cannot init logs\n");
1038
1039         if (rte_eal_alarm_init() < 0)
1040                 rte_panic("Cannot init interrupt-handling thread\n");
1041
1042         if (rte_eal_intr_init() < 0)
1043                 rte_panic("Cannot init interrupt-handling thread\n");
1044
1045         if (rte_eal_timer_init() < 0)
1046                 rte_panic("Cannot init HPET or TSC timers\n");
1047
1048         eal_check_mem_on_local_socket();
1049
1050         rte_eal_mcfg_complete();
1051
1052         TAILQ_FOREACH(solib, &solib_list, next) {
1053                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1054                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1055                 if (solib->lib_handle == NULL)
1056                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1057         }
1058
1059         eal_thread_init_master(rte_config.master_lcore);
1060
1061         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1062                 rte_config.master_lcore, (int)thread_id);
1063
1064         if (rte_eal_vdev_init() < 0)
1065                 rte_panic("Cannot init virtual devices\n");
1066
1067         RTE_LCORE_FOREACH_SLAVE(i) {
1068
1069                 /*
1070                  * create communication pipes between master thread
1071                  * and children
1072                  */
1073                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1074                         rte_panic("Cannot create pipe\n");
1075                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1076                         rte_panic("Cannot create pipe\n");
1077
1078                 lcore_config[i].state = WAIT;
1079
1080                 /* create a thread for each lcore */
1081                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1082                                      eal_thread_loop, NULL);
1083                 if (ret != 0)
1084                         rte_panic("Cannot create thread\n");
1085         }
1086
1087         /*
1088          * Launch a dummy function on all slave lcores, so that master lcore
1089          * knows they are all ready when this function returns.
1090          */
1091         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1092         rte_eal_mp_wait_lcore();
1093
1094         return fctret;
1095 }
1096
1097 /* get core role */
1098 enum rte_lcore_role_t
1099 rte_eal_lcore_role(unsigned lcore_id)
1100 {
1101         return (rte_config.lcore_role[lcore_id]);
1102 }
1103
1104 enum rte_proc_type_t
1105 rte_eal_process_type(void)
1106 {
1107         return (rte_config.process_type);
1108 }
1109
1110 int rte_eal_has_hugepages(void)
1111 {
1112         return ! internal_config.no_hugetlbfs;
1113 }