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 #include <rte_dev.h>
  79
  80 #include "eal_private.h"
  81 #include "eal_thread.h"
  82 #include "eal_internal_cfg.h"
  83 #include "eal_filesystem.h"
  84 #include "eal_hugepages.h"
  85
  86 #define OPT_HUGE_DIR    "huge-dir"
  87 #define OPT_PROC_TYPE   "proc-type"
  88 #define OPT_NO_SHCONF   "no-shconf"
  89 #define OPT_NO_HPET     "no-hpet"
  90 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  91 #define OPT_NO_PCI      "no-pci"
  92 #define OPT_NO_HUGE     "no-huge"
  93 #define OPT_FILE_PREFIX "file-prefix"
  94 #define OPT_SOCKET_MEM  "socket-mem"
  95 #define OPT_USE_DEVICE  "use-device"
  96 #define OPT_PCI_WHITELIST "pci-whitelist"
  97 #define OPT_PCI_BLACKLIST "pci-blacklist"
  98 #define OPT_VDEV        "vdev"
  99 #define OPT_SYSLOG      "syslog"
 100 #define OPT_BASE_VIRTADDR   "base-virtaddr"
 101 #define OPT_XEN_DOM0    "xen-dom0"
 102 #define OPT_CREATE_UIO_DEV "create-uio-dev"
 103 #define OPT_VFIO_INTR    "vfio-intr"
 104
 105 #define RTE_EAL_BLACKLIST_SIZE  0x100
 106
 107 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 108
 109 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 110
 111 #define HIGHEST_RPL 3
 112
 113 #define BITS_PER_HEX 4
 114
 115 /* Allow the application to print its usage message too if set */
 116 static rte_usage_hook_t rte_application_usage_hook = NULL;
 117
 118 TAILQ_HEAD(shared_driver_list, shared_driver);
 119
 120 /* Definition for shared object drivers. */
 121 struct shared_driver {
 122         TAILQ_ENTRY(shared_driver) next;
 123
 124         char    name[PATH_MAX];
 125         void*   lib_handle;
 126 };
 127
 128 /* List of external loadable drivers */
 129 static struct shared_driver_list solib_list =
 130 TAILQ_HEAD_INITIALIZER(solib_list);
 131
 132 /* early configuration structure, when memory config is not mmapped */
 133 static struct rte_mem_config early_mem_config;
 134
 135 /* define fd variable here, because file needs to be kept open for the
 136  * duration of the program, as we hold a write lock on it in the primary proc */
 137 static int mem_cfg_fd = -1;
 138
 139 static struct flock wr_lock = {
 140                 .l_type = F_WRLCK,
 141                 .l_whence = SEEK_SET,
 142                 .l_start = offsetof(struct rte_mem_config, memseg),
 143                 .l_len = sizeof(early_mem_config.memseg),
 144 };
 145
 146 /* Address of global and public configuration */
 147 static struct rte_config rte_config = {
 148                 .mem_config = &early_mem_config,
 149 };
 150
 151 /* internal configuration (per-core) */
 152 struct lcore_config lcore_config[RTE_MAX_LCORE];
 153
 154 /* internal configuration */
 155 struct internal_config internal_config;
 156
 157 /* used by rte_rdtsc() */
 158 int rte_cycles_vmware_tsc_map;
 159
 160 /* Return a pointer to the configuration structure */
 161 struct rte_config *
 162 rte_eal_get_configuration(void)
 163 {
 164         return &rte_config;
 165 }
 166
 167 /* parse a sysfs (or other) file containing one integer value */
 168 int
 169 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 170 {
 171         FILE *f;
 172         char buf[BUFSIZ];
 173         char *end = NULL;
 174
 175         if ((f = fopen(filename, "r")) == NULL) {
 176                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 177                         __func__, filename);
 178                 return -1;
 179         }
 180
 181         if (fgets(buf, sizeof(buf), f) == NULL) {
 182                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 183                         __func__, filename);
 184                 fclose(f);
 185                 return -1;
 186         }
 187         *val = strtoul(buf, &end, 0);
 188         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 189                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 190                                 __func__, filename);
 191                 fclose(f);
 192                 return -1;
 193         }
 194         fclose(f);
 195         return 0;
 196 }
 197
 198
 199 /* create memory configuration in shared/mmap memory. Take out
 200  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 201  * This means we never close the file while running (auto-close on exit).
 202  * We also don't lock the whole file, so that in future we can use read-locks
 203  * on other parts, e.g. memzones, to detect if there are running secondary
 204  * processes. */
 205 static void
 206 rte_eal_config_create(void)
 207 {
 208         void *rte_mem_cfg_addr;
 209         int retval;
 210
 211         const char *pathname = eal_runtime_config_path();
 212
 213         if (internal_config.no_shconf)
 214                 return;
 215
 216         if (mem_cfg_fd < 0){
 217                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 218                 if (mem_cfg_fd < 0)
 219                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 220         }
 221
 222         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 223         if (retval < 0){
 224                 close(mem_cfg_fd);
 225                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 226         }
 227
 228         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 229         if (retval < 0){
 230                 close(mem_cfg_fd);
 231                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 232                                 "process running?\n", pathname);
 233         }
 234
 235         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 236                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 237
 238         if (rte_mem_cfg_addr == MAP_FAILED){
 239                 rte_panic("Cannot mmap memory for rte_config\n");
 240         }
 241         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 242         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 243 }
 244
 245 /* attach to an existing shared memory config */
 246 static void
 247 rte_eal_config_attach(void)
 248 {
 249         void *rte_mem_cfg_addr;
 250         const char *pathname = eal_runtime_config_path();
 251
 252         if (internal_config.no_shconf)
 253                 return;
 254
 255         if (mem_cfg_fd < 0){
 256                 mem_cfg_fd = open(pathname, O_RDWR);
 257                 if (mem_cfg_fd < 0)
 258                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 259         }
 260
 261         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 262                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 263         close(mem_cfg_fd);
 264         if (rte_mem_cfg_addr == MAP_FAILED)
 265                 rte_panic("Cannot mmap memory for rte_config\n");
 266
 267         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 268 }
 269
 270 /* Detect if we are a primary or a secondary process */
 271 static enum rte_proc_type_t
 272 eal_proc_type_detect(void)
 273 {
 274         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 275         const char *pathname = eal_runtime_config_path();
 276
 277         /* if we can open the file but not get a write-lock we are a secondary
 278          * process. NOTE: if we get a file handle back, we keep that open
 279          * and don't close it to prevent a race condition between multiple opens */
 280         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 281                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 282                 ptype = RTE_PROC_SECONDARY;
 283
 284         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 285                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 286
 287         return ptype;
 288 }
 289
 290 /* Sets up rte_config structure with the pointer to shared memory config.*/
 291 static void
 292 rte_config_init(void)
 293 {
 294         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 295                         eal_proc_type_detect() : /* for auto, detect the type */
 296                         internal_config.process_type; /* otherwise use what's already set */
 297
 298         switch (rte_config.process_type){
 299         case RTE_PROC_PRIMARY:
 300                 rte_eal_config_create();
 301                 break;
 302         case RTE_PROC_SECONDARY:
 303                 rte_eal_config_attach();
 304                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 305                 break;
 306         case RTE_PROC_AUTO:
 307         case RTE_PROC_INVALID:
 308                 rte_panic("Invalid process type\n");
 309         }
 310 }
 311
 312 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 313 static void
 314 eal_hugedirs_unlock(void)
 315 {
 316         int i;
 317
 318         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 319         {
 320                 /* skip uninitialized */
 321                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 322                         continue;
 323                 /* unlock hugepage file */
 324                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 325                 close(internal_config.hugepage_info[i].lock_descriptor);
 326                 /* reset the field */
 327                 internal_config.hugepage_info[i].lock_descriptor = -1;
 328         }
 329 }
 330
 331 /* display usage */
 332 static void
 333 eal_usage(const char *prgname)
 334 {
 335         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 336                "[--proc-type primary|secondary|auto] \n\n"
 337                "EAL options:\n"
 338                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 339                "  -n NUM       : Number of memory channels\n"
 340                    "  -v           : Display version information on startup\n"
 341                "  -d LIB.so    : add driver (can be used multiple times)\n"
 342                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 343                "  -r NUM       : force number of memory ranks (don't detect)\n"
 344                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 345                            "without hugetlbfs\n"
 346                "  --"OPT_SYSLOG"     : set syslog facility\n"
 347                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 348                    "                 sockets (use comma separated values)\n"
 349                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 350                "  --"OPT_PROC_TYPE"  : type of this process\n"
 351                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 352                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 353                "               Prevent EAL from using this PCI device. The argument\n"
 354                "               format is <domain:bus:devid.func>.\n"
 355                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 356                "               Only use the specified PCI devices. The argument format\n"
 357                "               is <[domain:]bus:devid.func>. This option can be present\n"
 358                "               several times (once per device).\n"
 359                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 360                "  --"OPT_VDEV": add a virtual device.\n"
 361                "               The argument format is <driver><id>[,key=val,...]\n"
 362                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 363                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 364                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 365                "  --"OPT_VFIO_INTR": specify desired interrupt mode for VFIO "
 366                            "(legacy|msi|msix)\n"
 367                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 368                "\nEAL options for DEBUG use only:\n"
 369                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 370                "  --"OPT_NO_PCI"   : disable pci\n"
 371                "  --"OPT_NO_HPET"  : disable hpet\n"
 372                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 373                "\n",
 374                prgname);
 375         /* Allow the application to print its usage message too if hook is set */
 376         if ( rte_application_usage_hook ) {
 377                 printf("===== Application Usage =====\n\n");
 378                 rte_application_usage_hook(prgname);
 379         }
 380 }
 381
 382 /* Set a per-application usage message */
 383 rte_usage_hook_t
 384 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 385 {
 386         rte_usage_hook_t        old_func;
 387
 388         /* Will be NULL on the first call to denote the last usage routine. */
 389         old_func                                        = rte_application_usage_hook;
 390         rte_application_usage_hook      = usage_func;
 391
 392         return old_func;
 393 }
 394
 395 /*
 396  * Parse the coremask given as argument (hexadecimal string) and fill
 397  * the global configuration (core role and core count) with the parsed
 398  * value.
 399  */
 400 static int xdigit2val(unsigned char c)
 401 {
 402         int val;
 403         if(isdigit(c))
 404                 val = c - '0';
 405         else if(isupper(c))
 406                 val = c - 'A' + 10;
 407         else
 408                 val = c - 'a' + 10;
 409         return val;
 410 }
 411 static int
 412 eal_parse_coremask(const char *coremask)
 413 {
 414         struct rte_config *cfg = rte_eal_get_configuration();
 415         int i, j, idx = 0 ;
 416         unsigned count = 0;
 417         char c;
 418         int val;
 419
 420         if (coremask == NULL)
 421                 return -1;
 422         /* Remove all blank characters ahead and after .
 423          * Remove 0x/0X if exists.
 424          */
 425         while (isblank(*coremask))
 426                 coremask++;
 427         if (coremask[0] == '0' && ((coremask[1] == 'x')
 428                 ||  (coremask[1] == 'X')) )
 429                 coremask += 2;
 430         i = strnlen(coremask, PATH_MAX);
 431         while ((i > 0) && isblank(coremask[i - 1]))
 432                 i--;
 433         if (i == 0)
 434                 return -1;
 435
 436         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 437                 c = coremask[i];
 438                 if (isxdigit(c) == 0) {
 439                         /* invalid characters */
 440                         return (-1);
 441                 }
 442                 val = xdigit2val(c);
 443                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 444                         if((1 << j) & val) {
 445                                 if (!lcore_config[idx].detected) {
 446                                         RTE_LOG(ERR, EAL, "lcore %u "
 447                                                 "unavailable\n", idx);
 448                                         return -1;
 449                                 }
 450                                 cfg->lcore_role[idx] = ROLE_RTE;
 451                                 if(count == 0)
 452                                         cfg->master_lcore = idx;
 453                                 count++;
 454                         } else  {
 455                                 cfg->lcore_role[idx] = ROLE_OFF;
 456                         }
 457                 }
 458         }
 459         for(; i >= 0; i--)
 460                 if(coremask[i] != '0')
 461                         return -1;
 462         for(; idx < RTE_MAX_LCORE; idx++)
 463                 cfg->lcore_role[idx] = ROLE_OFF;
 464         if(count == 0)
 465                 return -1;
 466         /* Update the count of enabled logical cores of the EAL configuration */
 467         cfg->lcore_count = count;
 468         return 0;
 469 }
 470
 471 static int
 472 eal_parse_syslog(const char *facility)
 473 {
 474         int i;
 475         static struct {
 476                 const char *name;
 477                 int value;
 478         } map[] = {
 479                 { "auth", LOG_AUTH },
 480                 { "cron", LOG_CRON },
 481                 { "daemon", LOG_DAEMON },
 482                 { "ftp", LOG_FTP },
 483                 { "kern", LOG_KERN },
 484                 { "lpr", LOG_LPR },
 485                 { "mail", LOG_MAIL },
 486                 { "news", LOG_NEWS },
 487                 { "syslog", LOG_SYSLOG },
 488                 { "user", LOG_USER },
 489                 { "uucp", LOG_UUCP },
 490                 { "local0", LOG_LOCAL0 },
 491                 { "local1", LOG_LOCAL1 },
 492                 { "local2", LOG_LOCAL2 },
 493                 { "local3", LOG_LOCAL3 },
 494                 { "local4", LOG_LOCAL4 },
 495                 { "local5", LOG_LOCAL5 },
 496                 { "local6", LOG_LOCAL6 },
 497                 { "local7", LOG_LOCAL7 },
 498                 { NULL, 0 }
 499         };
 500
 501         for (i = 0; map[i].name; i++) {
 502                 if (!strcmp(facility, map[i].name)) {
 503                         internal_config.syslog_facility = map[i].value;
 504                         return 0;
 505                 }
 506         }
 507         return -1;
 508 }
 509
 510 static int
 511 eal_parse_socket_mem(char *socket_mem)
 512 {
 513         char * arg[RTE_MAX_NUMA_NODES];
 514         char *end;
 515         int arg_num, i, len;
 516         uint64_t total_mem = 0;
 517
 518         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 519         if (len == SOCKET_MEM_STRLEN) {
 520                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 521                 return -1;
 522         }
 523
 524         /* all other error cases will be caught later */
 525         if (!isdigit(socket_mem[len-1]))
 526                 return -1;
 527
 528         /* split the optarg into separate socket values */
 529         arg_num = rte_strsplit(socket_mem, len,
 530                         arg, RTE_MAX_NUMA_NODES, ',');
 531
 532         /* if split failed, or 0 arguments */
 533         if (arg_num <= 0)
 534                 return -1;
 535
 536         internal_config.force_sockets = 1;
 537
 538         /* parse each defined socket option */
 539         errno = 0;
 540         for (i = 0; i < arg_num; i++) {
 541                 end = NULL;
 542                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 543
 544                 /* check for invalid input */
 545                 if ((errno != 0)  ||
 546                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 547                         return -1;
 548                 internal_config.socket_mem[i] *= 1024ULL;
 549                 internal_config.socket_mem[i] *= 1024ULL;
 550                 total_mem += internal_config.socket_mem[i];
 551         }
 552
 553         /* check if we have a positive amount of total memory */
 554         if (total_mem == 0)
 555                 return -1;
 556
 557         return 0;
 558 }
 559
 560 static int
 561 eal_parse_base_virtaddr(const char *arg)
 562 {
 563         char *end;
 564         uint64_t addr;
 565
 566         errno = 0;
 567         addr = strtoull(arg, &end, 16);
 568
 569         /* check for errors */
 570         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 571                 return -1;
 572
 573         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 574 #ifndef RTE_ARCH_X86_64
 575         if (addr >= UINTPTR_MAX)
 576                 return -1;
 577 #endif
 578
 579         /* align the addr on 2M boundary */
 580         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 581                                                            RTE_PGSIZE_2M);
 582
 583         return 0;
 584 }
 585
 586 static int
 587 eal_parse_vfio_intr(const char *mode)
 588 {
 589         unsigned i;
 590         static struct {
 591                 const char *name;
 592                 enum rte_intr_mode value;
 593         } map[] = {
 594                 { "legacy", RTE_INTR_MODE_LEGACY },
 595                 { "msi", RTE_INTR_MODE_MSI },
 596                 { "msix", RTE_INTR_MODE_MSIX },
 597         };
 598
 599         for (i = 0; i < RTE_DIM(map); i++) {
 600                 if (!strcmp(mode, map[i].name)) {
 601                         internal_config.vfio_intr_mode = map[i].value;
 602                         return 0;
 603                 }
 604         }
 605         return -1;
 606 }
 607
 608 static inline size_t
 609 eal_get_hugepage_mem_size(void)
 610 {
 611         uint64_t size = 0;
 612         unsigned i, j;
 613
 614         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 615                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 616                 if (hpi->hugedir != NULL) {
 617                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 618                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 619                         }
 620                 }
 621         }
 622
 623         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 624 }
 625
 626 static enum rte_proc_type_t
 627 eal_parse_proc_type(const char *arg)
 628 {
 629         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 630                 return RTE_PROC_PRIMARY;
 631         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 632                 return RTE_PROC_SECONDARY;
 633         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 634                 return RTE_PROC_AUTO;
 635
 636         return RTE_PROC_INVALID;
 637 }
 638
 639 /* Parse the argument given in the command line of the application */
 640 static int
 641 eal_parse_args(int argc, char **argv)
 642 {
 643         int opt, ret, i;
 644         char **argvopt;
 645         int option_index;
 646         int coremask_ok = 0;
 647         char *prgname = argv[0];
 648         static struct option lgopts[] = {
 649                 {OPT_NO_HUGE, 0, 0, 0},
 650                 {OPT_NO_PCI, 0, 0, 0},
 651                 {OPT_NO_HPET, 0, 0, 0},
 652                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 653                 {OPT_HUGE_DIR, 1, 0, 0},
 654                 {OPT_NO_SHCONF, 0, 0, 0},
 655                 {OPT_PROC_TYPE, 1, 0, 0},
 656                 {OPT_FILE_PREFIX, 1, 0, 0},
 657                 {OPT_SOCKET_MEM, 1, 0, 0},
 658                 {OPT_PCI_WHITELIST, 1, 0, 0},
 659                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 660                 {OPT_VDEV, 1, 0, 0},
 661                 {OPT_SYSLOG, 1, NULL, 0},
 662                 {OPT_VFIO_INTR, 1, NULL, 0},
 663                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 664                 {OPT_XEN_DOM0, 0, 0, 0},
 665                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 666                 {0, 0, 0, 0}
 667         };
 668         struct shared_driver *solib;
 669
 670         argvopt = argv;
 671
 672         internal_config.memory = 0;
 673         internal_config.force_nrank = 0;
 674         internal_config.force_nchannel = 0;
 675         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 676         internal_config.hugepage_dir = NULL;
 677         internal_config.force_sockets = 0;
 678         internal_config.syslog_facility = LOG_DAEMON;
 679         internal_config.xen_dom0_support = 0;
 680         /* if set to NONE, interrupt mode is determined automatically */
 681         internal_config.vfio_intr_mode = RTE_INTR_MODE_NONE;
 682 #ifdef RTE_LIBEAL_USE_HPET
 683         internal_config.no_hpet = 0;
 684 #else
 685         internal_config.no_hpet = 1;
 686 #endif
 687         /* zero out the NUMA config */
 688         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 689                 internal_config.socket_mem[i] = 0;
 690
 691         /* zero out hugedir descriptors */
 692         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 693                 internal_config.hugepage_info[i].lock_descriptor = -1;
 694
 695         internal_config.vmware_tsc_map = 0;
 696         internal_config.base_virtaddr = 0;
 697
 698         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
 699                                   lgopts, &option_index)) != EOF) {
 700
 701                 switch (opt) {
 702                 /* blacklist */
 703                 case 'b':
 704                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 705                                         optarg) < 0) {
 706                                 eal_usage(prgname);
 707                                 return (-1);
 708                         }
 709                         break;
 710                 /* whitelist */
 711                 case 'w':
 712                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 713                                         optarg) < 0) {
 714                                 eal_usage(prgname);
 715                                 return -1;
 716                         }
 717                         break;
 718                 /* coremask */
 719                 case 'c':
 720                         if (eal_parse_coremask(optarg) < 0) {
 721                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 722                                 eal_usage(prgname);
 723                                 return -1;
 724                         }
 725                         coremask_ok = 1;
 726                         break;
 727                 /* force loading of external driver */
 728                 case 'd':
 729                         solib = malloc(sizeof(*solib));
 730                         if (solib == NULL) {
 731                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 732                                 return -1;
 733                         }
 734                         memset(solib, 0, sizeof(*solib));
 735                         strncpy(solib->name, optarg, PATH_MAX-1);
 736                         solib->name[PATH_MAX-1] = 0;
 737                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 738                         break;
 739                 /* size of memory */
 740                 case 'm':
 741                         internal_config.memory = atoi(optarg);
 742                         internal_config.memory *= 1024ULL;
 743                         internal_config.memory *= 1024ULL;
 744                         break;
 745                 /* force number of channels */
 746                 case 'n':
 747                         internal_config.force_nchannel = atoi(optarg);
 748                         if (internal_config.force_nchannel == 0 ||
 749                             internal_config.force_nchannel > 4) {
 750                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 751                                 eal_usage(prgname);
 752                                 return -1;
 753                         }
 754                         break;
 755                 /* force number of ranks */
 756                 case 'r':
 757                         internal_config.force_nrank = atoi(optarg);
 758                         if (internal_config.force_nrank == 0 ||
 759                             internal_config.force_nrank > 16) {
 760                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 761                                 eal_usage(prgname);
 762                                 return -1;
 763                         }
 764                         break;
 765                 case 'v':
 766                         /* since message is explicitly requested by user, we
 767                          * write message at highest log level so it can always be seen
 768                          * even if info or warning messages are disabled */
 769                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 770                         break;
 771
 772                 /* long options */
 773                 case 0:
 774                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 775                                 internal_config.no_hugetlbfs = 1;
 776                         }
 777                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 778                 #ifdef RTE_LIBRTE_XEN_DOM0
 779                                 internal_config.xen_dom0_support = 1;
 780                 #else
 781                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 782                                         "running on Dom0, please configure"
 783                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 784                                 return -1;
 785                 #endif
 786                         }
 787                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 788                                 internal_config.no_pci = 1;
 789                         }
 790                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 791                                 internal_config.no_hpet = 1;
 792                         }
 793                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 794                                 internal_config.vmware_tsc_map = 1;
 795                         }
 796                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 797                                 internal_config.no_shconf = 1;
 798                         }
 799                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 800                                 internal_config.hugepage_dir = optarg;
 801                         }
 802                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 803                                 internal_config.process_type = eal_parse_proc_type(optarg);
 804                         }
 805                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 806                                 internal_config.hugefile_prefix = optarg;
 807                         }
 808                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 809                                 if (eal_parse_socket_mem(optarg) < 0) {
 810                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 811                                                         OPT_SOCKET_MEM "\n");
 812                                         eal_usage(prgname);
 813                                         return -1;
 814                                 }
 815                         }
 816                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 817                                 printf("The --use-device option is deprecated, please use\n"
 818                                         "--whitelist or --vdev instead.\n");
 819                                 eal_usage(prgname);
 820                                 return -1;
 821                         }
 822                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 823                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 824                                                 optarg) < 0) {
 825                                         eal_usage(prgname);
 826                                         return -1;
 827                                 }
 828                         }
 829                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 830                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 831                                                 optarg) < 0) {
 832                                         eal_usage(prgname);
 833                                         return -1;
 834                                 }
 835                         }
 836                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 837                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 838                                                 optarg) < 0) {
 839                                         eal_usage(prgname);
 840                                         return -1;
 841                                 }
 842                         }
 843                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 844                                 if (eal_parse_syslog(optarg) < 0) {
 845                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 846                                                         OPT_SYSLOG "\n");
 847                                         eal_usage(prgname);
 848                                         return -1;
 849                                 }
 850                         }
 851                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 852                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 853                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 854                                                         OPT_BASE_VIRTADDR "\n");
 855                                         eal_usage(prgname);
 856                                         return -1;
 857                                 }
 858                         }
 859                         else if (!strcmp(lgopts[option_index].name, OPT_VFIO_INTR)) {
 860                                 if (eal_parse_vfio_intr(optarg) < 0) {
 861                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 862                                                         OPT_VFIO_INTR "\n");
 863                                         eal_usage(prgname);
 864                                         return -1;
 865                                 }
 866                         }
 867                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 868                                 internal_config.create_uio_dev = 1;
 869                         }
 870                         break;
 871
 872                 default:
 873                         eal_usage(prgname);
 874                         return -1;
 875                 }
 876         }
 877
 878         /* sanity checks */
 879         if (!coremask_ok) {
 880                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 881                 eal_usage(prgname);
 882                 return -1;
 883         }
 884         if (internal_config.process_type == RTE_PROC_AUTO){
 885                 internal_config.process_type = eal_proc_type_detect();
 886         }
 887         if (internal_config.process_type == RTE_PROC_INVALID){
 888                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 889                 eal_usage(prgname);
 890                 return -1;
 891         }
 892         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 893                         internal_config.force_nchannel == 0) {
 894                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 895                 eal_usage(prgname);
 896                 return -1;
 897         }
 898         if (index(internal_config.hugefile_prefix,'%') != NULL){
 899                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 900                 eal_usage(prgname);
 901                 return -1;
 902         }
 903         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 904                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 905                                 "at the same time\n");
 906                 eal_usage(prgname);
 907                 return -1;
 908         }
 909         /* --no-huge doesn't make sense with either -m or --socket-mem */
 910         if (internal_config.no_hugetlbfs &&
 911                         (internal_config.memory > 0 ||
 912                                         internal_config.force_sockets == 1)) {
 913                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 914                                 "together with --no-huge!\n");
 915                 eal_usage(prgname);
 916                 return -1;
 917         }
 918         /* --xen-dom0 doesn't make sense with --socket-mem */
 919         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 920                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 921                                         "together with --xen_dom0!\n");
 922                 eal_usage(prgname);
 923                 return -1;
 924         }
 925
 926         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 927                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 928                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 929                         "[-w] options cannot be used at the same time\n");
 930                 eal_usage(prgname);
 931                 return -1;
 932         }
 933
 934         if (optind >= 0)
 935                 argv[optind-1] = prgname;
 936
 937         /* if no memory amounts were requested, this will result in 0 and
 938          * will be overriden later, right after eal_hugepage_info_init() */
 939         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 940                 internal_config.memory += internal_config.socket_mem[i];
 941
 942         ret = optind-1;
 943         optind = 0; /* reset getopt lib */
 944         return ret;
 945 }
 946
 947 static void
 948 eal_check_mem_on_local_socket(void)
 949 {
 950         const struct rte_memseg *ms;
 951         int i, socket_id;
 952
 953         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 954
 955         ms = rte_eal_get_physmem_layout();
 956
 957         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 958                 if (ms[i].socket_id == socket_id &&
 959                                 ms[i].len > 0)
 960                         return;
 961
 962         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 963                         "memory on local socket!\n");
 964 }
 965
 966 static int
 967 sync_func(__attribute__((unused)) void *arg)
 968 {
 969         return 0;
 970 }
 971
 972 inline static void
 973 rte_eal_mcfg_complete(void)
 974 {
 975         /* ALL shared mem_config related INIT DONE */
 976         if (rte_config.process_type == RTE_PROC_PRIMARY)
 977                 rte_config.mem_config->magic = RTE_MAGIC;
 978 }
 979
 980 /*
 981  * Request iopl privilege for all RPL, returns 0 on success
 982  */
 983 static int
 984 rte_eal_iopl_init(void)
 985 {
 986         return iopl(HIGHEST_RPL);
 987 }
 988
 989 /* Launch threads, called at application init(). */
 990 int
 991 rte_eal_init(int argc, char **argv)
 992 {
 993         int i, fctret, ret;
 994         pthread_t thread_id;
 995         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 996         struct shared_driver *solib = NULL;
 997         const char *logid;
 998
 999         if (!rte_atomic32_test_and_set(&run_once))
1000                 return -1;
1001
1002         logid = strrchr(argv[0], '/');
1003         logid = strdup(logid ? logid + 1: argv[0]);
1004
1005         thread_id = pthread_self();
1006
1007         if (rte_eal_log_early_init() < 0)
1008                 rte_panic("Cannot init early logs\n");
1009
1010         if (rte_eal_cpu_init() < 0)
1011                 rte_panic("Cannot detect lcores\n");
1012
1013         fctret = eal_parse_args(argc, argv);
1014         if (fctret < 0)
1015                 exit(1);
1016
1017         if (internal_config.no_hugetlbfs == 0 &&
1018                         internal_config.process_type != RTE_PROC_SECONDARY &&
1019                         internal_config.xen_dom0_support == 0 &&
1020                         eal_hugepage_info_init() < 0)
1021                 rte_panic("Cannot get hugepage information\n");
1022
1023         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
1024                 if (internal_config.no_hugetlbfs)
1025                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
1026                 else
1027                         internal_config.memory = eal_get_hugepage_mem_size();
1028         }
1029
1030         if (internal_config.vmware_tsc_map == 1) {
1031 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
1032                 rte_cycles_vmware_tsc_map = 1;
1033                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
1034                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
1035 #else
1036                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
1037                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1038 #endif
1039         }
1040
1041         rte_srand(rte_rdtsc());
1042
1043         rte_config_init();
1044
1045         if (rte_eal_iopl_init() == 0)
1046                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1047
1048         if (rte_eal_pci_init() < 0)
1049                 rte_panic("Cannot init PCI\n");
1050
1051 #ifdef RTE_LIBRTE_IVSHMEM
1052         if (rte_eal_ivshmem_init() < 0)
1053                 rte_panic("Cannot init IVSHMEM\n");
1054 #endif
1055
1056         if (rte_eal_memory_init() < 0)
1057                 rte_panic("Cannot init memory\n");
1058
1059         /* the directories are locked during eal_hugepage_info_init */
1060         eal_hugedirs_unlock();
1061
1062         if (rte_eal_memzone_init() < 0)
1063                 rte_panic("Cannot init memzone\n");
1064
1065         if (rte_eal_tailqs_init() < 0)
1066                 rte_panic("Cannot init tail queues for objects\n");
1067
1068 #ifdef RTE_LIBRTE_IVSHMEM
1069         if (rte_eal_ivshmem_obj_init() < 0)
1070                 rte_panic("Cannot init IVSHMEM objects\n");
1071 #endif
1072
1073         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1074                 rte_panic("Cannot init logs\n");
1075
1076         if (rte_eal_alarm_init() < 0)
1077                 rte_panic("Cannot init interrupt-handling thread\n");
1078
1079         if (rte_eal_intr_init() < 0)
1080                 rte_panic("Cannot init interrupt-handling thread\n");
1081
1082         if (rte_eal_timer_init() < 0)
1083                 rte_panic("Cannot init HPET or TSC timers\n");
1084
1085         eal_check_mem_on_local_socket();
1086
1087         rte_eal_mcfg_complete();
1088
1089         TAILQ_FOREACH(solib, &solib_list, next) {
1090                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1091                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1092                 if (solib->lib_handle == NULL)
1093                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1094         }
1095
1096         eal_thread_init_master(rte_config.master_lcore);
1097
1098         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1099                 rte_config.master_lcore, (int)thread_id);
1100
1101         if (rte_eal_dev_init(PMD_INIT_PRE_PCI_PROBE) < 0)
1102                 rte_panic("Cannot init pmd devices\n");
1103
1104         RTE_LCORE_FOREACH_SLAVE(i) {
1105
1106                 /*
1107                  * create communication pipes between master thread
1108                  * and children
1109                  */
1110                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1111                         rte_panic("Cannot create pipe\n");
1112                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1113                         rte_panic("Cannot create pipe\n");
1114
1115                 lcore_config[i].state = WAIT;
1116
1117                 /* create a thread for each lcore */
1118                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1119                                      eal_thread_loop, NULL);
1120                 if (ret != 0)
1121                         rte_panic("Cannot create thread\n");
1122         }
1123
1124         /*
1125          * Launch a dummy function on all slave lcores, so that master lcore
1126          * knows they are all ready when this function returns.
1127          */
1128         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1129         rte_eal_mp_wait_lcore();
1130
1131         /* Probe & Initialize PCI devices */
1132         if (rte_eal_pci_probe())
1133                         rte_panic("Cannot probe PCI\n");
1134
1135         /* Initialize any outstanding devices */
1136         if (rte_eal_dev_init(PMD_INIT_POST_PCI_PROBE) < 0)
1137                 rte_panic("Cannot init pmd devices\n");
1138
1139         return fctret;
1140 }
1141
1142 /* get core role */
1143 enum rte_lcore_role_t
1144 rte_eal_lcore_role(unsigned lcore_id)
1145 {
1146         return (rte_config.lcore_role[lcore_id]);
1147 }
1148
1149 enum rte_proc_type_t
1150 rte_eal_process_type(void)
1151 {
1152         return (rte_config.process_type);
1153 }
1154
1155 int rte_eal_has_hugepages(void)
1156 {
1157         return ! internal_config.no_hugetlbfs;
1158 }