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 #define OPT_VFIO_INTR    "vfio-intr"
 103
 104 #define RTE_EAL_BLACKLIST_SIZE  0x100
 105
 106 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 107
 108 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 109
 110 #define HIGHEST_RPL 3
 111
 112 #define BITS_PER_HEX 4
 113
 114 /* Allow the application to print its usage message too if set */
 115 static rte_usage_hook_t rte_application_usage_hook = NULL;
 116
 117 TAILQ_HEAD(shared_driver_list, shared_driver);
 118
 119 /* Definition for shared object drivers. */
 120 struct shared_driver {
 121         TAILQ_ENTRY(shared_driver) next;
 122
 123         char    name[PATH_MAX];
 124         void*   lib_handle;
 125 };
 126
 127 /* List of external loadable drivers */
 128 static struct shared_driver_list solib_list =
 129 TAILQ_HEAD_INITIALIZER(solib_list);
 130
 131 /* early configuration structure, when memory config is not mmapped */
 132 static struct rte_mem_config early_mem_config;
 133
 134 /* define fd variable here, because file needs to be kept open for the
 135  * duration of the program, as we hold a write lock on it in the primary proc */
 136 static int mem_cfg_fd = -1;
 137
 138 static struct flock wr_lock = {
 139                 .l_type = F_WRLCK,
 140                 .l_whence = SEEK_SET,
 141                 .l_start = offsetof(struct rte_mem_config, memseg),
 142                 .l_len = sizeof(early_mem_config.memseg),
 143 };
 144
 145 /* Address of global and public configuration */
 146 static struct rte_config rte_config = {
 147                 .mem_config = &early_mem_config,
 148 };
 149
 150 /* internal configuration (per-core) */
 151 struct lcore_config lcore_config[RTE_MAX_LCORE];
 152
 153 /* internal configuration */
 154 struct internal_config internal_config;
 155
 156 /* used by rte_rdtsc() */
 157 int rte_cycles_vmware_tsc_map;
 158
 159 /* Return a pointer to the configuration structure */
 160 struct rte_config *
 161 rte_eal_get_configuration(void)
 162 {
 163         return &rte_config;
 164 }
 165
 166 /* parse a sysfs (or other) file containing one integer value */
 167 int
 168 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 169 {
 170         FILE *f;
 171         char buf[BUFSIZ];
 172         char *end = NULL;
 173
 174         if ((f = fopen(filename, "r")) == NULL) {
 175                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 176                         __func__, filename);
 177                 return -1;
 178         }
 179
 180         if (fgets(buf, sizeof(buf), f) == NULL) {
 181                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 182                         __func__, filename);
 183                 fclose(f);
 184                 return -1;
 185         }
 186         *val = strtoul(buf, &end, 0);
 187         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 188                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 189                                 __func__, filename);
 190                 fclose(f);
 191                 return -1;
 192         }
 193         fclose(f);
 194         return 0;
 195 }
 196
 197
 198 /* create memory configuration in shared/mmap memory. Take out
 199  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 200  * This means we never close the file while running (auto-close on exit).
 201  * We also don't lock the whole file, so that in future we can use read-locks
 202  * on other parts, e.g. memzones, to detect if there are running secondary
 203  * processes. */
 204 static void
 205 rte_eal_config_create(void)
 206 {
 207         void *rte_mem_cfg_addr;
 208         int retval;
 209
 210         const char *pathname = eal_runtime_config_path();
 211
 212         if (internal_config.no_shconf)
 213                 return;
 214
 215         if (mem_cfg_fd < 0){
 216                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 217                 if (mem_cfg_fd < 0)
 218                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 219         }
 220
 221         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 222         if (retval < 0){
 223                 close(mem_cfg_fd);
 224                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 225         }
 226
 227         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 228         if (retval < 0){
 229                 close(mem_cfg_fd);
 230                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 231                                 "process running?\n", pathname);
 232         }
 233
 234         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 235                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 236
 237         if (rte_mem_cfg_addr == MAP_FAILED){
 238                 rte_panic("Cannot mmap memory for rte_config\n");
 239         }
 240         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 241         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 242 }
 243
 244 /* attach to an existing shared memory config */
 245 static void
 246 rte_eal_config_attach(void)
 247 {
 248         void *rte_mem_cfg_addr;
 249         const char *pathname = eal_runtime_config_path();
 250
 251         if (internal_config.no_shconf)
 252                 return;
 253
 254         if (mem_cfg_fd < 0){
 255                 mem_cfg_fd = open(pathname, O_RDWR);
 256                 if (mem_cfg_fd < 0)
 257                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 258         }
 259
 260         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 261                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 262         close(mem_cfg_fd);
 263         if (rte_mem_cfg_addr == MAP_FAILED)
 264                 rte_panic("Cannot mmap memory for rte_config\n");
 265
 266         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 267 }
 268
 269 /* Detect if we are a primary or a secondary process */
 270 static enum rte_proc_type_t
 271 eal_proc_type_detect(void)
 272 {
 273         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 274         const char *pathname = eal_runtime_config_path();
 275
 276         /* if we can open the file but not get a write-lock we are a secondary
 277          * process. NOTE: if we get a file handle back, we keep that open
 278          * and don't close it to prevent a race condition between multiple opens */
 279         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 280                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 281                 ptype = RTE_PROC_SECONDARY;
 282
 283         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 284                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 285
 286         return ptype;
 287 }
 288
 289 /* Sets up rte_config structure with the pointer to shared memory config.*/
 290 static void
 291 rte_config_init(void)
 292 {
 293         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 294                         eal_proc_type_detect() : /* for auto, detect the type */
 295                         internal_config.process_type; /* otherwise use what's already set */
 296
 297         switch (rte_config.process_type){
 298         case RTE_PROC_PRIMARY:
 299                 rte_eal_config_create();
 300                 break;
 301         case RTE_PROC_SECONDARY:
 302                 rte_eal_config_attach();
 303                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 304                 break;
 305         case RTE_PROC_AUTO:
 306         case RTE_PROC_INVALID:
 307                 rte_panic("Invalid process type\n");
 308         }
 309 }
 310
 311 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 312 static void
 313 eal_hugedirs_unlock(void)
 314 {
 315         int i;
 316
 317         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 318         {
 319                 /* skip uninitialized */
 320                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 321                         continue;
 322                 /* unlock hugepage file */
 323                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 324                 close(internal_config.hugepage_info[i].lock_descriptor);
 325                 /* reset the field */
 326                 internal_config.hugepage_info[i].lock_descriptor = -1;
 327         }
 328 }
 329
 330 /* display usage */
 331 static void
 332 eal_usage(const char *prgname)
 333 {
 334         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 335                "[--proc-type primary|secondary|auto] \n\n"
 336                "EAL options:\n"
 337                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 338                "  -n NUM       : Number of memory channels\n"
 339                    "  -v           : Display version information on startup\n"
 340                "  -d LIB.so    : add driver (can be used multiple times)\n"
 341                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 342                "  -r NUM       : force number of memory ranks (don't detect)\n"
 343                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 344                            "without hugetlbfs\n"
 345                "  --"OPT_SYSLOG"     : set syslog facility\n"
 346                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 347                    "                 sockets (use comma separated values)\n"
 348                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 349                "  --"OPT_PROC_TYPE"  : type of this process\n"
 350                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 351                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 352                "               Prevent EAL from using this PCI device. The argument\n"
 353                "               format is <domain:bus:devid.func>.\n"
 354                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 355                "               Only use the specified PCI devices. The argument format\n"
 356                "               is <[domain:]bus:devid.func>. This option can be present\n"
 357                "               several times (once per device).\n"
 358                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 359                "  --"OPT_VDEV": add a virtual device.\n"
 360                "               The argument format is <driver><id>[,key=val,...]\n"
 361                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 362                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 363                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 364                "  --"OPT_VFIO_INTR": specify desired interrupt mode for VFIO "
 365                            "(legacy|msi|msix)\n"
 366                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 367                "\nEAL options for DEBUG use only:\n"
 368                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 369                "  --"OPT_NO_PCI"   : disable pci\n"
 370                "  --"OPT_NO_HPET"  : disable hpet\n"
 371                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 372                "\n",
 373                prgname);
 374         /* Allow the application to print its usage message too if hook is set */
 375         if ( rte_application_usage_hook ) {
 376                 printf("===== Application Usage =====\n\n");
 377                 rte_application_usage_hook(prgname);
 378         }
 379 }
 380
 381 /* Set a per-application usage message */
 382 rte_usage_hook_t
 383 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 384 {
 385         rte_usage_hook_t        old_func;
 386
 387         /* Will be NULL on the first call to denote the last usage routine. */
 388         old_func                                        = rte_application_usage_hook;
 389         rte_application_usage_hook      = usage_func;
 390
 391         return old_func;
 392 }
 393
 394 /*
 395  * Parse the coremask given as argument (hexadecimal string) and fill
 396  * the global configuration (core role and core count) with the parsed
 397  * value.
 398  */
 399 static int xdigit2val(unsigned char c)
 400 {
 401         int val;
 402         if(isdigit(c))
 403                 val = c - '0';
 404         else if(isupper(c))
 405                 val = c - 'A' + 10;
 406         else
 407                 val = c - 'a' + 10;
 408         return val;
 409 }
 410 static int
 411 eal_parse_coremask(const char *coremask)
 412 {
 413         struct rte_config *cfg = rte_eal_get_configuration();
 414         int i, j, idx = 0 ;
 415         unsigned count = 0;
 416         char c;
 417         int val;
 418
 419         if (coremask == NULL)
 420                 return -1;
 421         /* Remove all blank characters ahead and after .
 422          * Remove 0x/0X if exists.
 423          */
 424         while (isblank(*coremask))
 425                 coremask++;
 426         if (coremask[0] == '0' && ((coremask[1] == 'x')
 427                 ||  (coremask[1] == 'X')) )
 428                 coremask += 2;
 429         i = strnlen(coremask, PATH_MAX);
 430         while ((i > 0) && isblank(coremask[i - 1]))
 431                 i--;
 432         if (i == 0)
 433                 return -1;
 434
 435         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 436                 c = coremask[i];
 437                 if (isxdigit(c) == 0) {
 438                         /* invalid characters */
 439                         return (-1);
 440                 }
 441                 val = xdigit2val(c);
 442                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 443                         if((1 << j) & val) {
 444                                 if (!lcore_config[idx].detected) {
 445                                         RTE_LOG(ERR, EAL, "lcore %u "
 446                                                 "unavailable\n", idx);
 447                                         return -1;
 448                                 }
 449                                 cfg->lcore_role[idx] = ROLE_RTE;
 450                                 if(count == 0)
 451                                         cfg->master_lcore = idx;
 452                                 count++;
 453                         } else  {
 454                                 cfg->lcore_role[idx] = ROLE_OFF;
 455                         }
 456                 }
 457         }
 458         for(; i >= 0; i--)
 459                 if(coremask[i] != '0')
 460                         return -1;
 461         for(; idx < RTE_MAX_LCORE; idx++)
 462                 cfg->lcore_role[idx] = ROLE_OFF;
 463         if(count == 0)
 464                 return -1;
 465         /* Update the count of enabled logical cores of the EAL configuration */
 466         cfg->lcore_count = count;
 467         return 0;
 468 }
 469
 470 static int
 471 eal_parse_syslog(const char *facility)
 472 {
 473         int i;
 474         static struct {
 475                 const char *name;
 476                 int value;
 477         } map[] = {
 478                 { "auth", LOG_AUTH },
 479                 { "cron", LOG_CRON },
 480                 { "daemon", LOG_DAEMON },
 481                 { "ftp", LOG_FTP },
 482                 { "kern", LOG_KERN },
 483                 { "lpr", LOG_LPR },
 484                 { "mail", LOG_MAIL },
 485                 { "news", LOG_NEWS },
 486                 { "syslog", LOG_SYSLOG },
 487                 { "user", LOG_USER },
 488                 { "uucp", LOG_UUCP },
 489                 { "local0", LOG_LOCAL0 },
 490                 { "local1", LOG_LOCAL1 },
 491                 { "local2", LOG_LOCAL2 },
 492                 { "local3", LOG_LOCAL3 },
 493                 { "local4", LOG_LOCAL4 },
 494                 { "local5", LOG_LOCAL5 },
 495                 { "local6", LOG_LOCAL6 },
 496                 { "local7", LOG_LOCAL7 },
 497                 { NULL, 0 }
 498         };
 499
 500         for (i = 0; map[i].name; i++) {
 501                 if (!strcmp(facility, map[i].name)) {
 502                         internal_config.syslog_facility = map[i].value;
 503                         return 0;
 504                 }
 505         }
 506         return -1;
 507 }
 508
 509 static int
 510 eal_parse_socket_mem(char *socket_mem)
 511 {
 512         char * arg[RTE_MAX_NUMA_NODES];
 513         char *end;
 514         int arg_num, i, len;
 515         uint64_t total_mem = 0;
 516
 517         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 518         if (len == SOCKET_MEM_STRLEN) {
 519                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 520                 return -1;
 521         }
 522
 523         /* all other error cases will be caught later */
 524         if (!isdigit(socket_mem[len-1]))
 525                 return -1;
 526
 527         /* split the optarg into separate socket values */
 528         arg_num = rte_strsplit(socket_mem, len,
 529                         arg, RTE_MAX_NUMA_NODES, ',');
 530
 531         /* if split failed, or 0 arguments */
 532         if (arg_num <= 0)
 533                 return -1;
 534
 535         internal_config.force_sockets = 1;
 536
 537         /* parse each defined socket option */
 538         errno = 0;
 539         for (i = 0; i < arg_num; i++) {
 540                 end = NULL;
 541                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 542
 543                 /* check for invalid input */
 544                 if ((errno != 0)  ||
 545                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 546                         return -1;
 547                 internal_config.socket_mem[i] *= 1024ULL;
 548                 internal_config.socket_mem[i] *= 1024ULL;
 549                 total_mem += internal_config.socket_mem[i];
 550         }
 551
 552         /* check if we have a positive amount of total memory */
 553         if (total_mem == 0)
 554                 return -1;
 555
 556         return 0;
 557 }
 558
 559 static int
 560 eal_parse_base_virtaddr(const char *arg)
 561 {
 562         char *end;
 563         uint64_t addr;
 564
 565         errno = 0;
 566         addr = strtoull(arg, &end, 16);
 567
 568         /* check for errors */
 569         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 570                 return -1;
 571
 572         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 573 #ifndef RTE_ARCH_X86_64
 574         if (addr >= UINTPTR_MAX)
 575                 return -1;
 576 #endif
 577
 578         /* align the addr on 2M boundary */
 579         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 580                                                            RTE_PGSIZE_2M);
 581
 582         return 0;
 583 }
 584
 585 static int
 586 eal_parse_vfio_intr(const char *mode)
 587 {
 588         unsigned i;
 589         static struct {
 590                 const char *name;
 591                 enum rte_intr_mode value;
 592         } map[] = {
 593                 { "legacy", RTE_INTR_MODE_LEGACY },
 594                 { "msi", RTE_INTR_MODE_MSI },
 595                 { "msix", RTE_INTR_MODE_MSIX },
 596         };
 597
 598         for (i = 0; i < RTE_DIM(map); i++) {
 599                 if (!strcmp(mode, map[i].name)) {
 600                         internal_config.vfio_intr_mode = map[i].value;
 601                         return 0;
 602                 }
 603         }
 604         return -1;
 605 }
 606
 607 static inline size_t
 608 eal_get_hugepage_mem_size(void)
 609 {
 610         uint64_t size = 0;
 611         unsigned i, j;
 612
 613         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 614                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 615                 if (hpi->hugedir != NULL) {
 616                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 617                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 618                         }
 619                 }
 620         }
 621
 622         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 623 }
 624
 625 static enum rte_proc_type_t
 626 eal_parse_proc_type(const char *arg)
 627 {
 628         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 629                 return RTE_PROC_PRIMARY;
 630         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 631                 return RTE_PROC_SECONDARY;
 632         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 633                 return RTE_PROC_AUTO;
 634
 635         return RTE_PROC_INVALID;
 636 }
 637
 638 /* Parse the argument given in the command line of the application */
 639 static int
 640 eal_parse_args(int argc, char **argv)
 641 {
 642         int opt, ret, i;
 643         char **argvopt;
 644         int option_index;
 645         int coremask_ok = 0;
 646         char *prgname = argv[0];
 647         static struct option lgopts[] = {
 648                 {OPT_NO_HUGE, 0, 0, 0},
 649                 {OPT_NO_PCI, 0, 0, 0},
 650                 {OPT_NO_HPET, 0, 0, 0},
 651                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 652                 {OPT_HUGE_DIR, 1, 0, 0},
 653                 {OPT_NO_SHCONF, 0, 0, 0},
 654                 {OPT_PROC_TYPE, 1, 0, 0},
 655                 {OPT_FILE_PREFIX, 1, 0, 0},
 656                 {OPT_SOCKET_MEM, 1, 0, 0},
 657                 {OPT_PCI_WHITELIST, 1, 0, 0},
 658                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 659                 {OPT_VDEV, 1, 0, 0},
 660                 {OPT_SYSLOG, 1, NULL, 0},
 661                 {OPT_VFIO_INTR, 1, NULL, 0},
 662                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 663                 {OPT_XEN_DOM0, 0, 0, 0},
 664                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 665                 {0, 0, 0, 0}
 666         };
 667         struct shared_driver *solib;
 668
 669         argvopt = argv;
 670
 671         internal_config.memory = 0;
 672         internal_config.force_nrank = 0;
 673         internal_config.force_nchannel = 0;
 674         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 675         internal_config.hugepage_dir = NULL;
 676         internal_config.force_sockets = 0;
 677         internal_config.syslog_facility = LOG_DAEMON;
 678         internal_config.xen_dom0_support = 0;
 679         /* if set to NONE, interrupt mode is determined automatically */
 680         internal_config.vfio_intr_mode = RTE_INTR_MODE_NONE;
 681 #ifdef RTE_LIBEAL_USE_HPET
 682         internal_config.no_hpet = 0;
 683 #else
 684         internal_config.no_hpet = 1;
 685 #endif
 686         /* zero out the NUMA config */
 687         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 688                 internal_config.socket_mem[i] = 0;
 689
 690         /* zero out hugedir descriptors */
 691         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 692                 internal_config.hugepage_info[i].lock_descriptor = -1;
 693
 694         internal_config.vmware_tsc_map = 0;
 695         internal_config.base_virtaddr = 0;
 696
 697         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
 698                                   lgopts, &option_index)) != EOF) {
 699
 700                 switch (opt) {
 701                 /* blacklist */
 702                 case 'b':
 703                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 704                                         optarg) < 0) {
 705                                 eal_usage(prgname);
 706                                 return (-1);
 707                         }
 708                         break;
 709                 /* whitelist */
 710                 case 'w':
 711                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 712                                         optarg) < 0) {
 713                                 eal_usage(prgname);
 714                                 return -1;
 715                         }
 716                         break;
 717                 /* coremask */
 718                 case 'c':
 719                         if (eal_parse_coremask(optarg) < 0) {
 720                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 721                                 eal_usage(prgname);
 722                                 return -1;
 723                         }
 724                         coremask_ok = 1;
 725                         break;
 726                 /* force loading of external driver */
 727                 case 'd':
 728                         solib = malloc(sizeof(*solib));
 729                         if (solib == NULL) {
 730                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 731                                 return -1;
 732                         }
 733                         memset(solib, 0, sizeof(*solib));
 734                         strncpy(solib->name, optarg, PATH_MAX-1);
 735                         solib->name[PATH_MAX-1] = 0;
 736                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 737                         break;
 738                 /* size of memory */
 739                 case 'm':
 740                         internal_config.memory = atoi(optarg);
 741                         internal_config.memory *= 1024ULL;
 742                         internal_config.memory *= 1024ULL;
 743                         break;
 744                 /* force number of channels */
 745                 case 'n':
 746                         internal_config.force_nchannel = atoi(optarg);
 747                         if (internal_config.force_nchannel == 0 ||
 748                             internal_config.force_nchannel > 4) {
 749                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 750                                 eal_usage(prgname);
 751                                 return -1;
 752                         }
 753                         break;
 754                 /* force number of ranks */
 755                 case 'r':
 756                         internal_config.force_nrank = atoi(optarg);
 757                         if (internal_config.force_nrank == 0 ||
 758                             internal_config.force_nrank > 16) {
 759                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 760                                 eal_usage(prgname);
 761                                 return -1;
 762                         }
 763                         break;
 764                 case 'v':
 765                         /* since message is explicitly requested by user, we
 766                          * write message at highest log level so it can always be seen
 767                          * even if info or warning messages are disabled */
 768                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 769                         break;
 770
 771                 /* long options */
 772                 case 0:
 773                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 774                                 internal_config.no_hugetlbfs = 1;
 775                         }
 776                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 777                 #ifdef RTE_LIBRTE_XEN_DOM0
 778                                 internal_config.xen_dom0_support = 1;
 779                 #else
 780                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 781                                         "running on Dom0, please configure"
 782                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 783                                 return -1;
 784                 #endif
 785                         }
 786                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 787                                 internal_config.no_pci = 1;
 788                         }
 789                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 790                                 internal_config.no_hpet = 1;
 791                         }
 792                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 793                                 internal_config.vmware_tsc_map = 1;
 794                         }
 795                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 796                                 internal_config.no_shconf = 1;
 797                         }
 798                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 799                                 internal_config.hugepage_dir = optarg;
 800                         }
 801                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 802                                 internal_config.process_type = eal_parse_proc_type(optarg);
 803                         }
 804                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 805                                 internal_config.hugefile_prefix = optarg;
 806                         }
 807                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 808                                 if (eal_parse_socket_mem(optarg) < 0) {
 809                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 810                                                         OPT_SOCKET_MEM "\n");
 811                                         eal_usage(prgname);
 812                                         return -1;
 813                                 }
 814                         }
 815                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 816                                 printf("The --use-device option is deprecated, please use\n"
 817                                         "--whitelist or --vdev instead.\n");
 818                                 eal_usage(prgname);
 819                                 return -1;
 820                         }
 821                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 822                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 823                                                 optarg) < 0) {
 824                                         eal_usage(prgname);
 825                                         return -1;
 826                                 }
 827                         }
 828                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 829                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 830                                                 optarg) < 0) {
 831                                         eal_usage(prgname);
 832                                         return -1;
 833                                 }
 834                         }
 835                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 836                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 837                                                 optarg) < 0) {
 838                                         eal_usage(prgname);
 839                                         return -1;
 840                                 }
 841                         }
 842                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 843                                 if (eal_parse_syslog(optarg) < 0) {
 844                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 845                                                         OPT_SYSLOG "\n");
 846                                         eal_usage(prgname);
 847                                         return -1;
 848                                 }
 849                         }
 850                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 851                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 852                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 853                                                         OPT_BASE_VIRTADDR "\n");
 854                                         eal_usage(prgname);
 855                                         return -1;
 856                                 }
 857                         }
 858                         else if (!strcmp(lgopts[option_index].name, OPT_VFIO_INTR)) {
 859                                 if (eal_parse_vfio_intr(optarg) < 0) {
 860                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 861                                                         OPT_VFIO_INTR "\n");
 862                                         eal_usage(prgname);
 863                                         return -1;
 864                                 }
 865                         }
 866                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 867                                 internal_config.create_uio_dev = 1;
 868                         }
 869                         break;
 870
 871                 default:
 872                         eal_usage(prgname);
 873                         return -1;
 874                 }
 875         }
 876
 877         /* sanity checks */
 878         if (!coremask_ok) {
 879                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 880                 eal_usage(prgname);
 881                 return -1;
 882         }
 883         if (internal_config.process_type == RTE_PROC_AUTO){
 884                 internal_config.process_type = eal_proc_type_detect();
 885         }
 886         if (internal_config.process_type == RTE_PROC_INVALID){
 887                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 888                 eal_usage(prgname);
 889                 return -1;
 890         }
 891         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 892                         internal_config.force_nchannel == 0) {
 893                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 894                 eal_usage(prgname);
 895                 return -1;
 896         }
 897         if (index(internal_config.hugefile_prefix,'%') != NULL){
 898                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 899                 eal_usage(prgname);
 900                 return -1;
 901         }
 902         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 903                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 904                                 "at the same time\n");
 905                 eal_usage(prgname);
 906                 return -1;
 907         }
 908         /* --no-huge doesn't make sense with either -m or --socket-mem */
 909         if (internal_config.no_hugetlbfs &&
 910                         (internal_config.memory > 0 ||
 911                                         internal_config.force_sockets == 1)) {
 912                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 913                                 "together with --no-huge!\n");
 914                 eal_usage(prgname);
 915                 return -1;
 916         }
 917         /* --xen-dom0 doesn't make sense with --socket-mem */
 918         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 919                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 920                                         "together with --xen_dom0!\n");
 921                 eal_usage(prgname);
 922                 return -1;
 923         }
 924
 925         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 926                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 927                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 928                         "[-w] options cannot be used at the same time\n");
 929                 eal_usage(prgname);
 930                 return -1;
 931         }
 932
 933         if (optind >= 0)
 934                 argv[optind-1] = prgname;
 935
 936         /* if no memory amounts were requested, this will result in 0 and
 937          * will be overriden later, right after eal_hugepage_info_init() */
 938         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 939                 internal_config.memory += internal_config.socket_mem[i];
 940
 941         ret = optind-1;
 942         optind = 0; /* reset getopt lib */
 943         return ret;
 944 }
 945
 946 static void
 947 eal_check_mem_on_local_socket(void)
 948 {
 949         const struct rte_memseg *ms;
 950         int i, socket_id;
 951
 952         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 953
 954         ms = rte_eal_get_physmem_layout();
 955
 956         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 957                 if (ms[i].socket_id == socket_id &&
 958                                 ms[i].len > 0)
 959                         return;
 960
 961         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 962                         "memory on local socket!\n");
 963 }
 964
 965 static int
 966 sync_func(__attribute__((unused)) void *arg)
 967 {
 968         return 0;
 969 }
 970
 971 inline static void
 972 rte_eal_mcfg_complete(void)
 973 {
 974         /* ALL shared mem_config related INIT DONE */
 975         if (rte_config.process_type == RTE_PROC_PRIMARY)
 976                 rte_config.mem_config->magic = RTE_MAGIC;
 977 }
 978
 979 /*
 980  * Request iopl privilege for all RPL, returns 0 on success
 981  */
 982 static int
 983 rte_eal_iopl_init(void)
 984 {
 985         return iopl(HIGHEST_RPL);
 986 }
 987
 988 /* Launch threads, called at application init(). */
 989 int
 990 rte_eal_init(int argc, char **argv)
 991 {
 992         int i, fctret, ret;
 993         pthread_t thread_id;
 994         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 995         struct shared_driver *solib = NULL;
 996         const char *logid;
 997
 998         if (!rte_atomic32_test_and_set(&run_once))
 999                 return -1;
1000
1001         logid = strrchr(argv[0], '/');
1002         logid = strdup(logid ? logid + 1: argv[0]);
1003
1004         thread_id = pthread_self();
1005
1006         if (rte_eal_log_early_init() < 0)
1007                 rte_panic("Cannot init early logs\n");
1008
1009         if (rte_eal_cpu_init() < 0)
1010                 rte_panic("Cannot detect lcores\n");
1011
1012         fctret = eal_parse_args(argc, argv);
1013         if (fctret < 0)
1014                 exit(1);
1015
1016         if (internal_config.no_hugetlbfs == 0 &&
1017                         internal_config.process_type != RTE_PROC_SECONDARY &&
1018                         internal_config.xen_dom0_support == 0 &&
1019                         eal_hugepage_info_init() < 0)
1020                 rte_panic("Cannot get hugepage information\n");
1021
1022         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
1023                 if (internal_config.no_hugetlbfs)
1024                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
1025                 else
1026                         internal_config.memory = eal_get_hugepage_mem_size();
1027         }
1028
1029         if (internal_config.vmware_tsc_map == 1) {
1030 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
1031                 rte_cycles_vmware_tsc_map = 1;
1032                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
1033                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
1034 #else
1035                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
1036                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1037 #endif
1038         }
1039
1040         rte_srand(rte_rdtsc());
1041
1042         rte_config_init();
1043
1044         if (rte_eal_iopl_init() == 0)
1045                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1046
1047         if (rte_eal_pci_init() < 0)
1048                 rte_panic("Cannot init PCI\n");
1049
1050 #ifdef RTE_LIBRTE_IVSHMEM
1051         if (rte_eal_ivshmem_init() < 0)
1052                 rte_panic("Cannot init IVSHMEM\n");
1053 #endif
1054
1055         if (rte_eal_memory_init() < 0)
1056                 rte_panic("Cannot init memory\n");
1057
1058         /* the directories are locked during eal_hugepage_info_init */
1059         eal_hugedirs_unlock();
1060
1061         if (rte_eal_memzone_init() < 0)
1062                 rte_panic("Cannot init memzone\n");
1063
1064         if (rte_eal_tailqs_init() < 0)
1065                 rte_panic("Cannot init tail queues for objects\n");
1066
1067 #ifdef RTE_LIBRTE_IVSHMEM
1068         if (rte_eal_ivshmem_obj_init() < 0)
1069                 rte_panic("Cannot init IVSHMEM objects\n");
1070 #endif
1071
1072         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1073                 rte_panic("Cannot init logs\n");
1074
1075         if (rte_eal_alarm_init() < 0)
1076                 rte_panic("Cannot init interrupt-handling thread\n");
1077
1078         if (rte_eal_intr_init() < 0)
1079                 rte_panic("Cannot init interrupt-handling thread\n");
1080
1081         if (rte_eal_timer_init() < 0)
1082                 rte_panic("Cannot init HPET or TSC timers\n");
1083
1084         eal_check_mem_on_local_socket();
1085
1086         rte_eal_mcfg_complete();
1087
1088         TAILQ_FOREACH(solib, &solib_list, next) {
1089                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1090                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1091                 if (solib->lib_handle == NULL)
1092                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1093         }
1094
1095         eal_thread_init_master(rte_config.master_lcore);
1096
1097         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1098                 rte_config.master_lcore, (int)thread_id);
1099
1100         if (rte_eal_dev_init() < 0)
1101                 rte_panic("Cannot init pmd devices\n");
1102
1103         RTE_LCORE_FOREACH_SLAVE(i) {
1104
1105                 /*
1106                  * create communication pipes between master thread
1107                  * and children
1108                  */
1109                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1110                         rte_panic("Cannot create pipe\n");
1111                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1112                         rte_panic("Cannot create pipe\n");
1113
1114                 lcore_config[i].state = WAIT;
1115
1116                 /* create a thread for each lcore */
1117                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1118                                      eal_thread_loop, NULL);
1119                 if (ret != 0)
1120                         rte_panic("Cannot create thread\n");
1121         }
1122
1123         /*
1124          * Launch a dummy function on all slave lcores, so that master lcore
1125          * knows they are all ready when this function returns.
1126          */
1127         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1128         rte_eal_mp_wait_lcore();
1129
1130         return fctret;
1131 }
1132
1133 /* get core role */
1134 enum rte_lcore_role_t
1135 rte_eal_lcore_role(unsigned lcore_id)
1136 {
1137         return (rte_config.lcore_role[lcore_id]);
1138 }
1139
1140 enum rte_proc_type_t
1141 rte_eal_process_type(void)
1142 {
1143         return (rte_config.process_type);
1144 }
1145
1146 int rte_eal_has_hugepages(void)
1147 {
1148         return ! internal_config.no_hugetlbfs;
1149 }