lustre/tests/mpi/mdsrate.c

   1 /*
   2  * 2003, Copyright, Hewlett-Packard Development Compnay, LP.
   3  *
   4  * Developed under the sponsorship of the U.S. Government
   5  *     under Subcontract No. B514193
   6  */
   7
   8 /*
   9  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  10  * Use is subject to license terms.
  11  *
  12  * Copyright (c) 2012, 2015, Intel Corporation.
  13  */
  14
  15 #include <stdio.h>
  16 #include <getopt.h>
  17 #include <libgen.h>
  18 #include <sys/types.h>
  19 #include <sys/stat.h>
  20 #include <time.h>
  21 #include <limits.h>
  22 #include <errno.h>
  23 #include <string.h>
  24 #include <fcntl.h>
  25 #include <unistd.h>
  26 #include <stdlib.h>
  27 #include <stdarg.h>
  28 #include <signal.h>
  29 #include <sys/ioctl.h>
  30 #include <dirent.h>
  31 #include <sys/xattr.h>
  32
  33 #include "mpi.h"
  34
  35 /* lustre */
  36 #include <lustre/lustreapi.h>        /* for O_LOV_DELAY_CREATE */
  37
  38 #define CHECK_COUNT 10000
  39 #define DISPLAY_COUNT (CHECK_COUNT * 10)
  40 #define DISPLAY_TIME 100
  41
  42 enum {
  43         CREATE          = 'c',
  44         LOOKUP          = 'l',
  45         MKNOD           = 'm',
  46         OPEN            = 'o',
  47         STAT            = 's',
  48         UNLINK          = 'u',
  49         BEGIN           = 'b',
  50         ITERS           = 'i',
  51         TIME            = 't',
  52         DIRFMT          = 'd',
  53         NDIRS           = 'D',
  54         FILEFMT         = 'f',
  55         NFILES          = 'F',
  56         NOEXCL          = 'X',
  57         STRIPES         = 'S',
  58         SEED            = 'r',
  59         SEEDFILE        = 'R',
  60         RANDOM          = 'A',
  61         READDIR         = 'B',
  62         RECREATE        = 'C',
  63         SETXATTR        = 'x',
  64         SMALLWRITE      = 'w',
  65         IGNORE          = 'E',
  66         VERBOSE         = 'V',
  67         DEBUG           = 'v',
  68         HELP            = 'h',
  69         MNT             = 'M',
  70         MNTCOUNT        = 'N',
  71         MDTCOUNT        = 'T',
  72 };
  73
  74 struct option longOpts[] = {
  75         {"create",              0, NULL, CREATE     },
  76         {"lookup",              0, NULL, LOOKUP     },
  77         {"mknod",               0, NULL, MKNOD      },
  78         {"open",                0, NULL, OPEN       },
  79         {"stat",                0, NULL, STAT       },
  80         {"unlink",              0, NULL, UNLINK     },
  81         {"begin",               1, NULL, BEGIN      },
  82         {"iters",               1, NULL, ITERS      },
  83         {"time",                1, NULL, TIME       },   /* seconds */
  84         {"dirfmt",              1, NULL, DIRFMT     },
  85         {"ndirs",               1, NULL, NDIRS      },
  86         {"filefmt",             1, NULL, FILEFMT    },
  87         {"nfiles",              1, NULL, NFILES     },
  88         {"noexcl",              0, NULL, NOEXCL     },
  89         {"stripes",             1, NULL, STRIPES    },
  90         {"seed",                1, NULL, SEED       },
  91         {"seedfile",            1, NULL, SEEDFILE   },
  92         {"random_order",        0, NULL, RANDOM     },
  93         {"readdir_order",       0, NULL, READDIR    },
  94         {"recreate",            0, NULL, RECREATE   },
  95         {"setxattr",            0, NULL, SETXATTR   },
  96         {"smallwrite",          0, NULL, SMALLWRITE },
  97         {"ignore",              0, NULL, IGNORE     },
  98         {"verbose",             0, NULL, VERBOSE    },
  99         {"debug",               0, NULL, DEBUG      },
 100         {"help",                0, NULL, HELP       },
 101         {"mdtcount",            1, NULL, MDTCOUNT   },
 102         {"mntcount",            1, NULL, MNTCOUNT   },
 103         {"mntfmt",              1, NULL, MNT        },
 104         { 0,                    0, NULL, 0          }
 105 };
 106
 107 int foo1, foo2;
 108
 109 char   shortOpts[128];
 110 int    myrank = -1;
 111 int    nthreads = -1;
 112 char * prog;
 113 char   hostname[512] = "unknown";
 114 char   mode;
 115 char * cmd;
 116 int    openflags = O_RDWR|O_CREAT|O_EXCL;
 117 int    ndirs = 1;
 118 char * dirfmt;
 119 char   dir[PATH_MAX];
 120 char   mkdir_cmd[PATH_MAX+14];
 121 int    dirthreads;
 122 int    dirnum;
 123 DIR *  directory;
 124 struct dirent *dir_entry;
 125 int    nfiles;
 126 char   filefmt[PATH_MAX];
 127 char   filename[PATH_MAX];
 128 char   path[PATH_MAX];
 129 int    stripes = -1;
 130 int    begin;
 131 int    beginsave;
 132 int    end;
 133 int    iters;
 134 int    seconds;
 135 int    alarm_caught;
 136 struct sigaction act;
 137 int    order = RANDOM;
 138 int    seed;
 139 int    recreate;
 140 int    ignore;
 141 int    verbose;
 142 int    debug;
 143 struct stat statbuf;
 144 bool   with_xattr;
 145 char   xattrname[] = "user.mdsrate";
 146 char   xattrbuf[4096];
 147 /* max xattr name + value length is block size, use 4000 here to avoid ENOSPC */
 148 int    xattrlen = 4000;
 149 bool   smallwrite;
 150 int    mnt_count = -1;
 151 int    mdt_count = 1;
 152 char  *mntfmt;
 153
 154 #define dmesg if (debug) printf
 155
 156 #define DISPLAY_PROGRESS() {                                                \
 157         if (verbose && (nops % CHECK_COUNT == 0)) {                         \
 158                 curTime = MPI_Wtime();                                      \
 159                 interval = curTime - lastTime;                              \
 160                 if (interval > DISPLAY_TIME || nops % DISPLAY_COUNT == 0) { \
 161                         rate = (double)(nops - lastOps)/interval;           \
 162                         printf("Rank %d: %.2f %ss/sec %.2f secs "           \
 163                                "(total: %d %ss %.2f secs)\n",               \
 164                                myrank, rate, cmd, interval,                 \
 165                                nops, cmd, curTime - startTime);             \
 166                         lastOps = nops;                                     \
 167                         lastTime = curTime;                                 \
 168                 }                                                           \
 169         }                                                                   \
 170 }
 171
 172 char *usage_msg = "usage: %s\n"
 173                   "    { --create [ --noexcl | --setxattr | --smallwrite ] |\n"
 174                   "      --lookup | --mknod [ --setxattr ] | --open |\n"
 175                   "      --stat | --unlink [ --recreate ] [ --ignore ] |\n"
 176                   "      --setxattr }\n"
 177                   "    [ --help ] [ --verbose ] [ --debug ]\n"
 178                   "    { [ --begin <num> ] --nfiles <num> }\n"
 179                   "    [ --iters <num> ] [ --time <secs> ]\n"
 180                   "    [ --dirfmt <str> ] [ --ndirs  <num> ]\n"
 181                   "    [ --filefmt <str> ] [ --stripes <num> ]\n"
 182                   "    [ --random_order [--seed <num> | --seedfile <file>] ]\n"
 183                   "    [ --readdir_order ] [ --mntfmt <str> ]\n"
 184                   "    [ --mntcount <num> ] [ --mdtcount <num> ]\n"
 185                   "    [ --setxattr ] }\n";
 186
 187 static void
 188 usage(FILE *stream, char *fmt, ...)
 189 {
 190         if (myrank == 0) {
 191                 if (fmt != NULL) {
 192                         va_list       ap;
 193
 194                         fprintf(stream, "%s: ", prog);
 195                         va_start(ap, fmt);
 196                         vfprintf(stderr, fmt, ap);
 197                         va_end(ap);
 198                 }
 199                 fprintf(stream, usage_msg, prog);
 200         }
 201
 202         MPI_Finalize();
 203         exit(stream == stderr);
 204 }
 205
 206 /* Print process myrank and message, and exit (i.e. a fatal error) */
 207 static int
 208 fatal(int rank, const char *fmt, ...)
 209 {
 210         if (rank == myrank) {
 211                 va_list       ap;
 212
 213                 fprintf(stderr, "rank %d: ", rank);
 214                 va_start(ap, fmt);
 215                 vfprintf(stderr, fmt, ap);
 216                 va_end(ap);
 217         }
 218
 219         MPI_Abort(MPI_COMM_WORLD, 1);
 220         exit(1);
 221 }
 222
 223 static void
 224 sigalrm_handler(int signum)
 225 {
 226         alarm_caught++;
 227 }
 228
 229 /* HAVE_LLAPI_FILE_LOOKUP is defined by liblustreapi.h if this function is
 230  * defined therein.  Otherwise we can do the equivalent operation via ioctl
 231  * if we have access to a complete lustre build tree to get the various
 232  * definitions - then compile with USE_MDC_LOOKUP defined. */
 233 #if defined(HAVE_LLAPI_FILE_LOOKUP)
 234 #define HAVE_MDC_LOOKUP
 235 #elif defined(USE_MDC_LOOKUP)
 236 #include <config.h>
 237 #include <lustre_ioctl.h>
 238
 239 int llapi_file_lookup(int dirfd, const char *name)
 240 {
 241         struct obd_ioctl_data data = { 0 };
 242         char rawbuf[8192];
 243         char *buf = rawbuf;
 244         int rc;
 245
 246         if (dirfd < 0 || name == NULL)
 247                 return -EINVAL;
 248
 249         data.ioc_version = OBD_IOCTL_VERSION;
 250         data.ioc_len = sizeof(data);
 251         data.ioc_inlbuf1 = name;
 252         data.ioc_inllen1 = strlen(name) + 1;
 253
 254         rc = obd_ioctl_pack(&data, &buf, sizeof(rawbuf));
 255         if (rc) {
 256                 fatal(myrank, "ioctl_pack failed: rc = %d\n", rc);
 257                 return rc;
 258         }
 259
 260         return ioctl(fd, IOC_MDC_LOOKUP, buf);
 261 }
 262 #define HAVE_MDC_LOOKUP
 263 #endif
 264
 265 static void
 266 process_args(int argc, char *argv[])
 267 {
 268         char   *cp, *endptr;
 269         int    i, index, offset, tmpend, rc;
 270         char   tmp[16];
 271         FILE * seed_file;
 272         struct option *opt;
 273
 274         setbuf(stdout, 0);
 275         setbuf(stderr, 0);
 276         prog = basename(argv[0]);
 277         strcpy(filefmt, "f%d");
 278         gethostname(hostname, sizeof(hostname));
 279
 280         /* auto create shortOpts rather than maintaining a static string. */
 281         for (opt = longOpts, cp = shortOpts; opt->name != NULL; opt++, cp++) {
 282                 *cp = opt->val;
 283                 if (opt->has_arg)
 284                         *++cp = ':';
 285         }
 286
 287         while ((rc = getopt_long(argc,argv, shortOpts, longOpts,&index)) != -1) {
 288                 switch (rc) {
 289                 case OPEN:
 290                         openflags &= ~(O_CREAT|O_EXCL);
 291                 case CREATE:
 292 #ifdef HAVE_MDC_LOOKUP
 293                 case LOOKUP:
 294 #endif
 295                 case MKNOD:
 296                 case STAT:
 297                 case UNLINK:
 298                         if (cmd != NULL) {
 299                                 fatal(0, "Invalid - more than one operation "
 300                                            "specified: --%s\n",
 301                                         longOpts[index].name);
 302                         }
 303                         mode = rc;
 304                         cmd = (char *)longOpts[index].name;
 305                         break;
 306                 case NOEXCL:
 307                         if (mode != CREATE && mode != MKNOD) {
 308                                 usage(stderr, "--noexcl only applies to "
 309                                               "--create or --mknod.\n");
 310                         }
 311                         openflags &= ~O_EXCL;
 312                         break;
 313                 case RECREATE:
 314                         if (mode != UNLINK) {
 315                                 usage(stderr, "--recreate only makes sense"
 316                                               "with --unlink.\n");
 317                         }
 318                         recreate++;
 319                         break;
 320                 case SETXATTR:
 321                         if (cmd == NULL) {
 322                                 mode = SETXATTR;
 323                                 cmd = (char *)longOpts[index].name;
 324                         } else if (mode == CREATE || mode == MKNOD) {
 325                                 with_xattr = true;
 326                         } else {
 327                                 usage(stderr, "--setxattr only makes sense "
 328                                       "with --create, --mknod or alone.\n");
 329                         }
 330                         break;
 331                 case SMALLWRITE:
 332                         if (mode != CREATE)
 333                                 usage(stderr, "--smallwrite only applies to "
 334                                               "--create.\n");
 335                         smallwrite = true;
 336                         break;
 337                 case BEGIN:
 338                         begin = strtol(optarg, &endptr, 0);
 339                         if ((*endptr != 0) || (begin < 0)) {
 340                                 fatal(0, "Invalid --start value.\n");
 341                         }
 342                         break;
 343                 case ITERS:
 344                         iters = strtol(optarg, &endptr, 0);
 345                         if ((*endptr != 0) || (iters <= 0)) {
 346                                 fatal(0, "Invalid --iters value.\n");
 347                         }
 348                         if (mode != LOOKUP && mode != OPEN) {
 349                                 usage(stderr, "--iters only makes sense with "
 350                                               "--lookup or --open.\n");
 351                         }
 352                         break;
 353                 case TIME:
 354                         seconds = strtol(optarg, &endptr, 0);
 355                         if ((*endptr != 0) || (seconds <= 0)) {
 356                                 fatal(0, "Invalid --time value.\n");
 357                         }
 358                         break;
 359                 case DIRFMT:
 360                         if (strlen(optarg) > (PATH_MAX - 16)) {
 361                                 fatal(0, "--dirfmt too long\n");
 362                         }
 363                         dirfmt = optarg;
 364                         break;
 365                 case NDIRS:
 366                         ndirs = strtol(optarg, &endptr, 0);
 367                         if ((*endptr != 0) || (ndirs <= 0)) {
 368                                 fatal(0, "Invalid --ndirs value.\n");
 369                         }
 370                         if ((ndirs > nthreads) &&
 371                             ((mode == CREATE) || (mode == MKNOD))) {
 372                                 fatal(0, "--ndirs=%d must be less than or "
 373                                       "equal to the number of threads (%d).\n",
 374                                       ndirs, nthreads);
 375                         }
 376                         break;
 377                 case FILEFMT:
 378                         if (strlen(optarg) > 4080) {
 379                                 fatal(0, "--filefmt too long\n");
 380                         }
 381
 382                         /* Use %%d where you want the file # in the name. */
 383                         sprintf(filefmt, optarg, myrank);
 384                         break;
 385                 case NFILES:
 386                         nfiles = strtol(optarg, &endptr, 0);
 387                         if ((*endptr != 0) || (nfiles <= 0)) {
 388                                 fatal(0, "Invalid --nfiles value.\n");
 389                         }
 390                         break;
 391                 case STRIPES:
 392                         stripes = strtol(optarg, &endptr, 0);
 393                         if ((*endptr != 0) || (stripes < 0)) {
 394                                 fatal(0, "Invalid --stripes value.\n");
 395                         }
 396
 397                         if (stripes == 0) {
 398                                 openflags |= O_LOV_DELAY_CREATE;
 399                         } else {
 400                                 fatal(0, "non-zero --stripes value "
 401                                          "not yet supported.\n");
 402                         }
 403
 404                         break;
 405                 case SEED:
 406                         seed = strtoul(optarg, &endptr, 0);
 407                         if (*endptr) {
 408                                 fatal(0, "bad --seed option %s\n", optarg);
 409                         }
 410                         break;
 411                 case SEEDFILE:
 412                         seed_file = fopen(optarg, "r");
 413                         if (!seed_file) {
 414                               fatal(myrank, "fopen(%s) error: %s\n",
 415                                       optarg, strerror(errno));
 416                         }
 417
 418                         for (i = -1; fgets(tmp, 16, seed_file) != NULL;) {
 419                                 if (++i == myrank)
 420                                         break;
 421                         }
 422
 423                         if (i == myrank) {
 424                                 rc = sscanf(tmp, "%d", &seed);
 425                                 if ((rc != 1) || (seed < 0)) {
 426                                         fatal(myrank, "Invalid seed value '%s' "
 427                                               "at line %d in %s.\n",
 428                                               tmp, i, optarg);
 429                                 }
 430                         } else {
 431                                 fatal(myrank, "File '%s' too short. Does not "
 432                                       "contain a seed for thread %d.\n",
 433                                       optarg, myrank);
 434                         }
 435
 436                         fclose(seed_file);
 437                         break;
 438                 case RANDOM:
 439                 case READDIR:
 440                         if (mode != LOOKUP && mode != OPEN)  {
 441                                 fatal(0, "--%s can only be specified with "
 442                                          "--lookup, or --open.\n",
 443                                       (char *)longOpts[index].name);
 444                         }
 445                         order = rc;
 446                         break;
 447                 case IGNORE:
 448                         ++ignore;
 449                         break;
 450                 case DEBUG:
 451                         ++debug;
 452                 case VERBOSE:
 453                         ++verbose;
 454                         break;
 455                 case HELP:
 456                         usage(stdout, NULL);
 457                         break;
 458                 case MNT:
 459                         if (strlen(optarg) > (PATH_MAX - 16))
 460                                 fatal(0, "--mnt too long\n");
 461                         mntfmt = optarg;
 462                         break;
 463                 case MNTCOUNT:
 464                         mnt_count = strtol(optarg, &endptr, 0);
 465                         if ((*endptr != 0) || (mnt_count <= 0)) {
 466                                 fatal(0, "Invalid --mnt_count value %s.\n",
 467                                       optarg);
 468                         }
 469                         break;
 470                 case MDTCOUNT:
 471                         mdt_count = strtol(optarg, &endptr, 0);
 472                         if ((*endptr != 0) || (mdt_count <= 0)) {
 473                                 fatal(0, "Invalid --mdt_count value %s.\n",
 474                                       optarg);
 475                         }
 476                         break;
 477                 default:
 478                         usage(stderr, "unrecognized option: '%c'.\n", optopt);
 479                 }
 480         }
 481
 482         if (optind < argc) {
 483                 usage(stderr, "too many arguments %d >= %d.\n", optind, argc);
 484         }
 485
 486         if ((mnt_count != -1 && mntfmt == NULL) ||
 487             (mnt_count == -1 && mntfmt != NULL)) {
 488                 usage(stderr, "mnt_count and mntfmt must be specified at the "
 489                              "same time\n");
 490         }
 491
 492         if (mode == CREATE || mode == MKNOD || mode == UNLINK ||
 493             mode == STAT || mode == SETXATTR) {
 494                 if (seconds != 0) {
 495                         if (nfiles == 0)
 496                                 nfiles = INT_MAX;
 497                 } else if (nfiles == 0) {
 498                         usage(stderr, "--nfiles or --time must be specified "
 499                                       "with %s.\n", cmd);
 500                 }
 501         } else if (mode == LOOKUP || mode == OPEN) {
 502                 if (seconds != 0) {
 503                         if (iters == 0)
 504                                 iters = INT_MAX;
 505                 } else if (iters == 0) {
 506                         usage(stderr, "--iters or --time must be specifed "
 507                                       "with %s.\n", cmd);
 508                 }
 509
 510                 if (nfiles == 0) {
 511                         usage(stderr, "--nfiles must be specifed with --%s.\n",
 512                               cmd);
 513                 }
 514
 515                 if (seed == 0) {
 516                         int fd = open("/dev/urandom", O_RDONLY);
 517
 518                         if (fd >= 0) {
 519                                 if (read(fd, &seed, sizeof(seed)) <
 520                                     sizeof(seed))
 521                                         seed = time(0);
 522                                 close(fd);
 523                         } else {
 524                                 seed = time(0);
 525                         }
 526                 }
 527
 528                 srand(seed);
 529
 530                 dmesg("%s: rank %d seed %d (%s).\n", prog, myrank, seed,
 531                       (order == RANDOM) ? "random_order" : "readdir_order");
 532         } else {
 533                 usage(stderr, "one --create, --mknod, --open, --stat,"
 534 #ifdef HAVE_MDC_LOOKUP
 535                       " --lookup,"
 536 #endif
 537                       " --unlink or --setxattr must be specifed.");
 538         }
 539
 540         /* support for multiple threads in a dir, set begin/end appropriately.*/
 541         dirnum = myrank % ndirs;
 542         dirthreads = nthreads / ndirs;
 543         if (nthreads > (ndirs * dirthreads + dirnum))
 544                 ++dirthreads;
 545
 546         offset = myrank / ndirs;
 547
 548         tmpend = begin + nfiles - 1;
 549         if (tmpend <= 0)
 550                 tmpend = INT_MAX;
 551
 552         end = begin + (nfiles / dirthreads) * dirthreads + offset;
 553         if ((end > tmpend) || (end <= 0))
 554                 end -= dirthreads;
 555
 556         /* make sure mnt_count <= nthreads, otherwise it might div 0 in
 557          * the following test */
 558         if (mnt_count > nthreads)
 559                 mnt_count = nthreads;
 560
 561         begin += offset;
 562         if (begin < 0)
 563                 begin = INT_MAX;
 564
 565         beginsave = begin;
 566
 567         dmesg("%d: iters %d nfiles %d time %d begin %d end %d dirthreads %d."
 568               "\n", myrank, iters, nfiles, seconds, begin, end, dirthreads);
 569
 570         if (dirfmt == NULL) {
 571                 strcpy(dir, ".");
 572         } else {
 573                 int dir_len = 0;
 574
 575                 if (mntfmt != NULL) {
 576                         sprintf(dir, mntfmt, (myrank / (nthreads/mnt_count)));
 577                         strcat(dir, "/");
 578                         dir_len = strlen(dir);
 579                 }
 580                 sprintf(dir + dir_len, dirfmt, dirnum);
 581
 582                 if (mdt_count > 1) {
 583                         struct stat sb;
 584                         if (stat(dir, &sb) == 0) {
 585                                 if (!S_ISDIR(sb.st_mode))
 586                                         fatal(myrank, "'%s' is not dir\n", dir);
 587                         } else if (errno == ENOENT) {
 588                                 sprintf(mkdir_cmd, "lfs mkdir -i %d %s",
 589                                         myrank % mdt_count, dir);
 590                         } else {
 591                                 fatal(myrank, "'%s' stat failed\n", dir);
 592                         }
 593                 } else {
 594                         sprintf(mkdir_cmd, "mkdir -p %s", dir);
 595                 }
 596
 597                 dmesg("%d: %s\n", myrank, mkdir_cmd);
 598 #ifdef _LIGHTWEIGHT_KERNEL
 599                 printf("NOTICE: not running system(%s)\n", mkdir_cmd);
 600 #else
 601                 rc = system(mkdir_cmd);
 602                 if (rc)
 603                         fatal(myrank, "'%s' failed.\n", mkdir_cmd);
 604 #endif
 605
 606                 rc = chdir(dir);
 607                 if (rc) {
 608                         fatal(myrank, "unable to chdir to '%s'.\n", dir);
 609                 }
 610         }
 611 }
 612
 613 static inline char *next_file()
 614 {
 615         if (order == RANDOM) {
 616                 sprintf(filename, filefmt, random() % nfiles);
 617                 return(filename);
 618         }
 619
 620         /* readdir order */
 621
 622         dir_entry = readdir(directory);
 623         if (dir_entry == NULL) {
 624                 rewinddir(directory);
 625                 while ((dir_entry = readdir(directory)) != NULL) {
 626                         if (dir_entry->d_name[0] != '.')
 627                                 return(dir_entry->d_name);
 628                 }
 629
 630                 fatal(myrank, "unable to read directory %s (%s).\n",
 631                       dir, strerror(errno));
 632         }
 633
 634         return(dir_entry->d_name);
 635 }
 636
 637 int
 638 main(int argc, char *argv[])
 639 {
 640         int    i, j, fd, rc, nops, lastOps;
 641         int ag_ops = 0;
 642         double ag_interval = 0;
 643         double ag_rate = 0;
 644         double rate, avg_rate, effective_rate;
 645         double startTime, curTime, lastTime, interval;
 646         time_t timestamp;
 647         char * file;
 648
 649         rc = MPI_Init(&argc, &argv);
 650         if (rc != MPI_SUCCESS)
 651                 fatal(myrank, "MPI_Init failed: %d\n", rc);
 652
 653         rc = MPI_Comm_size(MPI_COMM_WORLD, &nthreads);
 654         if (rc != MPI_SUCCESS)
 655                 fatal(myrank, "MPI_Comm_size failed: %d\n", rc);
 656
 657         rc = MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
 658         if (rc != MPI_SUCCESS)
 659                 fatal(myrank, "MPI_Comm_rank failed: %d\n", rc);
 660
 661         process_args(argc, argv);
 662
 663         timestamp = time(0);
 664         if ((myrank == 0) || debug) {
 665                 printf("%d: %s starting at %s",
 666                        myrank, hostname, ctime(&timestamp));
 667         }
 668
 669         /* if we're not measuring creation rates then precreate
 670          * the files we're operating on. */
 671         if ((mode != CREATE) && (mode != MKNOD) && !ignore &&
 672             (mode != UNLINK || recreate)) {
 673                 /* create the files in reverse order. When we encounter
 674                  * a file that already exists, assume the remainder of
 675                  * the files exist to save time. The timed performance
 676                  * test scripts make use of this behavior. */
 677                 for (i = end, j = 0; i >= begin; i -= dirthreads) {
 678                         sprintf(filename, filefmt, i);
 679                         fd = open(filename, openflags, 0644);
 680                         if (fd < 0) {
 681                                 if (errno == EEXIST)
 682                                         break;
 683                                 rc = errno;
 684                                 fatal(myrank, "precreate open(%s) error: %s\n",
 685                                       filename, strerror(rc));
 686                         }
 687                         j++;
 688                         close(fd);
 689                 }
 690                 dmesg("%d: %s pre-created %d files.\n",myrank,hostname,j);
 691
 692                 rc = MPI_Barrier(MPI_COMM_WORLD);
 693                 if (rc != MPI_SUCCESS)
 694                         fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
 695         }
 696
 697         if (order == READDIR) {
 698                 directory = opendir(dir);
 699                 if (directory == NULL) {
 700                         rc = errno;
 701                         fatal(myrank, "opendir(%s) error: %s\n",
 702                               dir, strerror(rc));
 703                 }
 704
 705                 timestamp = time(0);
 706                 j = random() % nfiles;
 707                 dmesg("%d: %s initializing dir offset %u: %s",
 708                       myrank, hostname, j, ctime(&timestamp));
 709
 710                 for (i = 0; i <= j; i++) {
 711                         if ((dir_entry = readdir(directory)) == NULL) {
 712                                 fatal(myrank, "could not read entry number %d "
 713                                       "in directory %s.\n", i, dir);
 714                         }
 715                 }
 716
 717                 timestamp = time(0);
 718                 dmesg("%d: index %d, filename %s, offset %ld: "
 719                       "%s initialization complete: %s",
 720                       myrank, i, dir_entry->d_name, telldir(directory),
 721                       hostname, ctime(&timestamp));
 722         }
 723
 724         if (seconds) {
 725                 act.sa_handler = sigalrm_handler;
 726                 (void)sigemptyset(&act.sa_mask);
 727                 act.sa_flags = 0;
 728                 sigaction(SIGALRM, &act, NULL);
 729                 alarm(seconds);
 730         }
 731
 732         rc = MPI_Barrier(MPI_COMM_WORLD);
 733         if (rc != MPI_SUCCESS)
 734                 fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
 735
 736         startTime = lastTime = MPI_Wtime();
 737         nops = lastOps = 0;
 738
 739         switch (mode) {
 740         case CREATE:
 741                 for (; begin <= end && !alarm_caught; begin += dirthreads) {
 742                         snprintf(filename, sizeof(filename), filefmt, begin);
 743                         fd = open(filename, openflags, 0644);
 744                         if (fd < 0) {
 745                                 rc = errno;
 746                                 if (rc == EINTR && alarm_caught)
 747                                         break;
 748                                 fatal(myrank, "open(%s) error: %s\n",
 749                                       filename, strerror(rc));
 750                         }
 751
 752                         if (with_xattr) {
 753                                 rc = fsetxattr(fd, xattrname, xattrbuf,
 754                                                xattrlen, XATTR_CREATE);
 755                                 if (rc) {
 756                                         rc = errno;
 757                                         if (rc == EINTR && alarm_caught)
 758                                                 break;
 759                                         fatal(myrank,
 760                                               "setxattr(%s) error: %s\n",
 761                                               filename, strerror(rc));
 762                                 }
 763                         }
 764                         if (smallwrite) {
 765                                 rc = write(fd, xattrbuf, xattrlen);
 766                                 if (rc < 0) {
 767                                         rc = errno;
 768                                         if (rc == EINTR && alarm_caught)
 769                                                 break;
 770                                         fatal(myrank,
 771                                               "write(%s) error: %s\n",
 772                                               filename, strerror(rc));
 773                                 }
 774                         }
 775
 776                         close(fd);
 777                         nops++;
 778                         DISPLAY_PROGRESS();
 779                 }
 780
 781                 dmesg("%d: created %d files, last file '%s'.\n",
 782                       myrank, nops, filename);
 783                 break;
 784 #ifdef HAVE_MDC_LOOKUP
 785         case LOOKUP:
 786                 fd = open(dir, O_RDONLY);
 787                 if (fd < 0) {
 788                         fatal(myrank, "open(dir == '%s') error: %s\n",
 789                               dir, strerror(errno));
 790                 }
 791
 792                 for (; nops < iters && !alarm_caught;) {
 793                         char *filename = next_file();
 794                         rc = llapi_file_lookup(fd, filename);
 795                         if (rc < 0) {
 796                                 if (((rc = errno) == EINTR) && alarm_caught)
 797                                         break;
 798                                 fatal(myrank, "llapi_file_lookup(%s) "
 799                                       "error: %s\n", filename, strerror(rc));
 800                         }
 801
 802                         nops++;
 803                         DISPLAY_PROGRESS();
 804                 }
 805                 break;
 806 #endif
 807         case MKNOD:
 808                 for (; begin <= end && !alarm_caught; begin += dirthreads) {
 809                         snprintf(filename, sizeof(filename), filefmt, begin);
 810                         rc = mknod(filename, S_IFREG | 0644, 0);
 811                         if (rc) {
 812                                 rc = errno;
 813                                 if (rc == EINTR && alarm_caught)
 814                                         break;
 815                                 fatal(myrank, "mknod(%s) error: %s\n",
 816                                       filename, strerror(rc));
 817                         }
 818
 819                         if (with_xattr) {
 820                                 rc = setxattr(filename, xattrname, xattrbuf,
 821                                               xattrlen, XATTR_CREATE);
 822                                 if (rc) {
 823                                         rc = errno;
 824                                         if (rc == EINTR && alarm_caught)
 825                                                 break;
 826                                         fatal(myrank,
 827                                               "setxattr(%s) error: %s\n",
 828                                               filename, strerror(rc));
 829                                 }
 830                         }
 831
 832                         nops++;
 833                         DISPLAY_PROGRESS();
 834                 }
 835                 break;
 836         case OPEN:
 837                 for (; nops < iters && !alarm_caught;) {
 838                         file = next_file();
 839                         if ((fd = open(file, openflags, 0644)) < 0) {
 840                                 if (((rc = errno) == EINTR) && alarm_caught)
 841                                         break;
 842                                 fatal(myrank, "open(%s) error: %s\n",
 843                                       file, strerror(rc));
 844                         }
 845
 846                         close(fd);
 847
 848                         nops++;
 849                         DISPLAY_PROGRESS();
 850                 }
 851                 break;
 852         case STAT:
 853                 for (; begin <= end && !alarm_caught; begin += dirthreads) {
 854                         sprintf(filename, filefmt, begin);
 855                         rc = stat(filename, &statbuf);
 856                         if (rc) {
 857                                 if (((rc = errno) == EINTR) && alarm_caught)
 858                                         break;
 859                                 if (((rc = errno) == ENOENT) && ignore)
 860                                         continue;
 861                                 fatal(myrank, "stat(%s) error: %s\n",
 862                                       filename, strerror(rc));
 863                         }
 864
 865                         nops++;
 866                         DISPLAY_PROGRESS();
 867                 }
 868                 break;
 869         case UNLINK:
 870                 for (; begin <= end && !alarm_caught; begin += dirthreads) {
 871                         sprintf(filename, filefmt, begin);
 872                         rc = unlink(filename);
 873                         if (rc) {
 874                                 if (((rc = errno) == EINTR) && alarm_caught)
 875                                         break;
 876                                 if ((rc = errno) == ENOENT) {
 877                                         if (ignore)
 878                                                 continue;
 879                                         /* no more files to unlink */
 880                                         break;
 881                                 }
 882                                 fatal(myrank, "unlink(%s) error: %s\n",
 883                                       filename, strerror(rc));
 884                         }
 885
 886                         nops++;
 887                         DISPLAY_PROGRESS();
 888                 }
 889                 break;
 890         case SETXATTR:
 891                 for (; begin <= end && !alarm_caught; begin += dirthreads) {
 892                         snprintf(filename, sizeof(filename), filefmt, begin);
 893                         rc = setxattr(filename, xattrname, xattrbuf, xattrlen,
 894                                       XATTR_CREATE);
 895                         if (rc) {
 896                                 rc = errno;
 897                                 if (rc == EINTR && alarm_caught)
 898                                         break;
 899                                 if (rc == ENOENT && ignore)
 900                                         continue;
 901                                 fatal(myrank, "setxattr(%s) error: %s\n",
 902                                       filename, strerror(rc));
 903                         }
 904
 905                         nops++;
 906                         DISPLAY_PROGRESS();
 907                 }
 908                 break;
 909         }
 910
 911         rc = MPI_Barrier(MPI_COMM_WORLD);
 912         if (rc != MPI_SUCCESS)
 913                fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
 914         curTime = MPI_Wtime();
 915         interval = curTime - startTime;
 916         rate = (double) (nops) / interval;
 917
 918         rc = MPI_Reduce(&nops, &ag_ops, 1, MPI_INT, MPI_SUM, 0,
 919                         MPI_COMM_WORLD);
 920         if (rc != MPI_SUCCESS) {
 921                 fatal(myrank, "Failure in MPI_Reduce of total ops.\n");
 922         }
 923
 924         rc = MPI_Reduce(&interval, &ag_interval, 1, MPI_DOUBLE, MPI_SUM, 0,
 925                         MPI_COMM_WORLD);
 926         if (rc != MPI_SUCCESS) {
 927                 fatal(myrank, "Failure in MPI_Reduce of total interval.\n");
 928         }
 929
 930         rc = MPI_Reduce(&rate, &ag_rate, 1, MPI_DOUBLE, MPI_SUM, 0,
 931                         MPI_COMM_WORLD);
 932         if (rc != MPI_SUCCESS) {
 933                 fatal(myrank, "Failure in MPI_Reduce of aggregated rate.\n");
 934         }
 935
 936         if (myrank == 0) {
 937                 curTime = MPI_Wtime();
 938                 interval = curTime - startTime;
 939                 effective_rate = (double) ag_ops / interval;
 940                 avg_rate = (double) ag_ops / ag_interval;
 941
 942                 printf("Rate: %.2f eff %.2f aggr %.2f avg client %ss/sec "
 943                        "(total: %d threads %d %ss %d dirs %d threads/dir %.2f secs)\n",
 944                        effective_rate, ag_rate, avg_rate, cmd, nthreads, ag_ops,
 945                        cmd, ndirs, dirthreads, interval);
 946                 if (mode == UNLINK && !recreate && !ignore && ag_ops != nfiles)
 947                         printf("Warning: only unlinked %d files instead of %d"
 948                                "\n", ag_ops, nfiles);
 949         }
 950
 951         if (recreate) {
 952                 for (begin = beginsave; begin <= end; begin += dirthreads) {
 953                         sprintf(filename, filefmt, begin);
 954                         if ((fd = open(filename, openflags, 0644)) < 0) {
 955                                 rc = errno;
 956                                 if (rc == EEXIST)
 957                                         break;
 958                                 fatal(myrank, "recreate open(%s) error: %s\n",
 959                                       filename, strerror(rc));
 960                         }
 961
 962                         close(fd);
 963                 }
 964         }
 965
 966         timestamp = time(0);
 967         if ((myrank == 0) || debug) {
 968                 printf("%d: %s finished at %s",
 969                        myrank, hostname, ctime(&timestamp));
 970         }
 971
 972         MPI_Finalize();
 973         return(0);
 974 }