lustre/tests/mpi/write_append_truncate.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * lustre/tests/write_append_truncate.c
  35  *
  36  * Each loop does 3 things:
  37  *   - truncate file to zero (not via ftruncate though, to test O_APPEND)
  38  *   - write a "chunk" of data (should be at file offset 0 after truncate)
  39  *   - on each of two threads either append or truncate-up the file
  40  *
  41  * If the truncate happened first, we should have a hole in the file.
  42  * If the append happened first, we should have truncated the file down.
  43  *
  44  * WRITE_SIZE_MAX and APPEND_SIZE_MAX are large enough to cross a stripe.
  45  *
  46  * compile: mpicc -g -Wall -o write_append_truncate write_append_truncate.c
  47  * run:     mpirun -np 2 -machlist <hostlist file> write_append_truncate <file>
  48  *  or:     pdsh -w <two hosts> write_append_truncate <file>
  49  *  or:     prun -n 2 [-N 2] write_append_truncate <file>
  50  */
  51
  52 #include <stdlib.h>
  53 #include <stdio.h>
  54 #include <stdarg.h>
  55 #include <sys/types.h>
  56 #include <sys/stat.h>
  57 #include <fcntl.h>
  58 #include <errno.h>
  59 #include <time.h>
  60 #include <string.h>
  61 #include <unistd.h>
  62 #include <getopt.h>
  63 #include "mpi.h"
  64
  65 #define DEFAULT_ITER    10000
  66
  67 #define WRITE_SIZE_MAX  1234567
  68 #define APPEND_SIZE_MAX 1234567
  69 #define TRUNC_SIZE_MAX  1234567
  70
  71 #define STATUS_FMT "WR %c %7d/%#08x, AP %c %7d/%#08x, TR@ %7d/%#08x"
  72
  73 #define HOSTNAME_SIZE 50
  74 char hostname[HOSTNAME_SIZE];
  75 #define FNAMES_MAX 256
  76
  77 void usage(char *prog)
  78 {
  79         printf("usage: %s [-a append_max] [-C] [-n nloops] [-s seed]\n"
  80                "\t\t[-t trunc_max] [-T] [-v] [-w write_max] <filename> ...\n", prog);
  81         printf("\t-a append_max: maximum size of append, default %u bytes\n",
  82                APPEND_SIZE_MAX);
  83         printf("\t-C: 'classic' checks (on file 0)\n");
  84         printf("\t-n nloops: count of loops to run, default %u\n",DEFAULT_ITER);
  85         printf("\t-s seed: random seed to use, default {current time}\n");
  86         printf("\t-t trunc_max: maximum size of truncate, default %u bytes\n",
  87                TRUNC_SIZE_MAX);
  88         printf("\t-T: 'classic' truncates (on file 0)\n");
  89         printf("\t-w write_max: maximum size of write, default %u bytes\n",
  90                WRITE_SIZE_MAX);
  91         printf("\t-W: 'classic' writes (on rank 0, file 0)\n");
  92         printf("\t-v: run in verbose mode (repeat for more verbosity)\n");
  93         printf("\tfilename for each mountpoint of same filesystem on a node\n");
  94         printf("\b%s must be run with at least 2 processes\n", prog);
  95
  96         MPI_Finalize();
  97         exit(1);
  98 }
  99
 100 /* Print process rank, loop count, message, and exit (i.e. a fatal error) */
 101 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 102 __attribute__ ((format (printf, 4, 5)));
 103
 104 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 105 {
 106         va_list       ap;
 107
 108         printf("r=%2u", rank);
 109         if (loop >= 0)
 110                 printf(" l=%04u", loop);
 111         if (error != 0)
 112                 printf(" %s", hostname);
 113         printf(": ");
 114
 115         va_start(ap, fmt);
 116
 117         vprintf(fmt, ap);
 118
 119         if (error != 0)
 120                 MPI_Abort(MPI_COMM_WORLD, error);
 121 }
 122
 123 int main(int argc, char *argv[])
 124 {
 125         int n, nloops = DEFAULT_ITER;
 126         int nfnames = 0, ifnames, fd;
 127         int rank = -1, nproc, ret;
 128         unsigned write_max = WRITE_SIZE_MAX;
 129         unsigned append_max = APPEND_SIZE_MAX;
 130         unsigned write_size = 0, append_size = 0, trunc_size = 0;
 131         unsigned trunc_max = 0, trunc_offset = 0;
 132         char *append_buf;
 133         char *write_buf;
 134         char *read_buf = NULL;
 135         char *trunc_buf = NULL;
 136         int seed = time(0);
 137         int done;
 138         int error;
 139         int verbose = 0;
 140         int classic_check = 0, classic_trunc = 0, classic_write = 0;
 141         char write_char = 'A', append_char = 'a';
 142         char *fnames[FNAMES_MAX], *end;
 143         char *prog = "write_append_truncate";
 144         int c;
 145
 146         error = MPI_Init(&argc, &argv);
 147         if (error != MPI_SUCCESS)
 148                 printf("%s: MPI_Init failed: %d\n", prog, error);
 149         else if (verbose > 2)
 150                 printf("%s: MPI_Init succeeded\n", prog);
 151
 152         prog = strrchr(argv[0], '/');
 153         if (prog == NULL)
 154                 prog = argv[0];
 155         else
 156                 prog++;
 157
 158         while ((c = getopt(argc, argv, "a:cCn:s:t:Tvw:W")) != -1) {
 159                 switch(c) {
 160                 case 'a':
 161                         append_max = strtoul(optarg, &end, 0);
 162                         if (append_max == 0 || *end) {
 163                                 fprintf(stderr, "%s: bad append option '%s'\n",
 164                                         prog, optarg);
 165                                 usage(prog);
 166                         }
 167                         break;
 168                 case 'C':
 169                         classic_check++;
 170                         break;
 171                 case 'n':
 172                         nloops = strtoul(optarg, &end, 0);
 173                         if (nloops == 0 || *end) {
 174                                 fprintf(stderr, "%s: bad nloops option '%s'\n",
 175                                         prog, optarg);
 176                                 usage(prog);
 177                         }
 178                         break;
 179                 case 's':
 180                         seed = strtoul(optarg, &end, 0);
 181                         if (*end) {
 182                                 fprintf(stderr, "%s: bad seed option '%s'\n",
 183                                         prog, optarg);
 184                                 usage(prog);
 185                         }
 186                         break;
 187                 case 't':
 188                         trunc_max = strtoul(optarg, &end, 0);
 189                         if (*end) {
 190                                 fprintf(stderr,"%s: bad truncate option '%s'\n",
 191                                         prog, optarg);
 192                                 usage(prog);
 193                         }
 194                         break;
 195                 case 'T':
 196                         classic_trunc++;
 197                         break;
 198                 case 'v':
 199                         verbose++;
 200                         break;
 201                 case 'w':
 202                         write_max = strtoul(optarg, &end, 0);
 203                         if (write_max == 0 || *end) {
 204                                 fprintf(stderr, "%s: bad write option '%s'\n",
 205                                         prog, optarg);
 206                                 usage(prog);
 207                         }
 208                         break;
 209                 case 'W':
 210                         classic_write++;
 211                         break;
 212                 default:
 213                         fprintf(stderr, "%s: unknown option '%c'\n", prog, c);
 214                         usage(prog);
 215                 }
 216         }
 217
 218         srand(seed);
 219
 220         if (argc == optind) {
 221                 fprintf(stderr, "%s: missing filename argument\n", prog);
 222                 usage(prog);
 223         }
 224
 225         if (argc > optind + FNAMES_MAX) {
 226                 fprintf(stderr, "%s: too many extra options\n", prog);
 227                 usage(prog);
 228         }
 229
 230         while (optind < argc)
 231                 fnames[nfnames++] = argv[optind++];
 232
 233         error = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
 234         if (verbose > 2 || error != MPI_SUCCESS)
 235                 rprintf(rank, -1, error != MPI_SUCCESS, "MPI_Comm_rank: %d\n",
 236                         error);
 237
 238         error = MPI_Comm_size(MPI_COMM_WORLD, &nproc);
 239         if (verbose > 2 || error != MPI_SUCCESS)
 240                 rprintf(rank, -1, error != MPI_SUCCESS, "MPI_Comm_size: %d\n",
 241                         error);
 242
 243         if (nproc < 2)
 244                 rprintf(rank, -1, 1, "%s: must run with at least 2 processes\n",
 245                         prog);
 246
 247         append_buf = malloc(append_max);
 248         if (append_buf == NULL)
 249                 rprintf(rank, -1, 1,"%s: error allocating append_buf %u\n",
 250                         prog, append_max);
 251
 252         write_buf = malloc(write_max);
 253         if (write_buf == NULL)
 254                 rprintf(rank, -1, 1, "%s: error allocating write_buf %u\n",
 255                         prog, write_max);
 256
 257         if (gethostname(hostname, HOSTNAME_SIZE) < 0)
 258                 rprintf(rank, -1, 1, "%s: gethostname failed: %s\n",
 259                         prog, strerror(errno));
 260
 261         if (rank == 0) {
 262                 int max_size = write_max + (trunc_max ?: append_max)+append_max;
 263
 264                 fd = open(fnames[0], O_WRONLY|O_CREAT|O_TRUNC, 0666);
 265                 rprintf(rank,-1, fd<0, "create %s, max size: %u, seed %u: %s\n",
 266                         fnames[0], max_size, seed, strerror(errno));
 267                 close(fd);
 268
 269                 trunc_buf = calloc(1, trunc_max ?: append_max);
 270                 if (trunc_buf == NULL)
 271                         rprintf(rank,-1,1,"%s: error allocating trunc_buf %u\n",
 272                                 prog, trunc_max ?: append_max);
 273
 274                 /* initial write + truncate up + append */
 275                 read_buf = malloc(max_size);
 276                 if (read_buf == NULL)
 277                         rprintf(rank,-1,1,"%s: error allocating read_buf %u\n",
 278                                 prog, max_size);
 279         }
 280
 281         error = MPI_Barrier(MPI_COMM_WORLD);
 282         if (verbose > 2 || error != MPI_SUCCESS)
 283                 rprintf(rank, -1, error != MPI_SUCCESS,
 284                         "prep MPI_Barrier: %d\n", error);
 285
 286         ifnames = rank % nfnames;
 287         fd = open(fnames[ifnames], O_RDWR | O_APPEND);
 288         if (verbose || fd < 0)
 289                 rprintf(rank, -1, errno, "open '%s' (%u): %s\n",
 290                         fnames[ifnames], ifnames, strerror(errno));
 291
 292         for (n = 0; n < nloops; n++) {
 293                 /* Initialized only to quiet stupid GCC warnings */
 294                 unsigned write_rank = 0, append_rank = n, trunc_rank = n + 1;
 295                 unsigned mpi_shared_vars[6];
 296
 297                 /* reset the environment */
 298                 write_char = 'A' + (n % 26);
 299                 append_char = 'a' + (n % 26);
 300
 301                 if (rank == 0) {
 302                         write_size = (rand() % (write_max - 1)) + 1;
 303                         append_size = (rand() % (append_max - 1)) + 1;
 304                         trunc_size = (rand() % ((trunc_max?: append_size)-1))+1;
 305                         trunc_offset = write_size + trunc_size;
 306
 307                         if (verbose || n % 1000 == 0)
 308                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 309                                         write_char, write_size, write_size,
 310                                         append_char, append_size, append_size,
 311                                         trunc_offset, trunc_offset);
 312
 313                         write_rank = (classic_write ? 0 : rand()) % nproc;
 314                         do {
 315                                 append_rank = (classic_write ? n : rand()) %
 316                                                                 nproc;
 317                                 /* We can't allow the append rank be the same
 318                                  * as the classic_trunc trunc_rank, or we will
 319                                  * spin here forever. */
 320                         } while (append_rank == (n + 1) % nproc);
 321                         do {
 322                                 trunc_rank = (classic_trunc? (n + 1) : rand()) %
 323                                                                 nproc;
 324                         } while (trunc_rank == append_rank);
 325
 326                         mpi_shared_vars[0] = write_size;
 327                         mpi_shared_vars[1] = append_size;
 328                         mpi_shared_vars[2] = trunc_size;
 329                         mpi_shared_vars[3] = write_rank;
 330                         mpi_shared_vars[4] = append_rank;
 331                         mpi_shared_vars[5] = trunc_rank;
 332                 }
 333
 334                 error = MPI_Bcast(&mpi_shared_vars, 6,
 335                                   MPI_INT, 0, MPI_COMM_WORLD);
 336                 if (verbose > 2 || error != MPI_SUCCESS)
 337                         rprintf(rank, n, error != MPI_SUCCESS,
 338                                 "MPI_Bcast mpi_shared_vars"
 339                                 "[%u, %u, %u, %u, %u, %u]: %d\n",
 340                                 mpi_shared_vars[0], mpi_shared_vars[1],
 341                                 mpi_shared_vars[2], mpi_shared_vars[3],
 342                                 mpi_shared_vars[4], mpi_shared_vars[5], error);
 343
 344                 if (rank != 0) {
 345                         write_size  = mpi_shared_vars[0];
 346                         append_size = mpi_shared_vars[1];
 347                         trunc_size  = mpi_shared_vars[2];
 348                         write_rank  = mpi_shared_vars[3];
 349                         append_rank = mpi_shared_vars[4];
 350                         trunc_rank  = mpi_shared_vars[5];
 351
 352                         trunc_offset = write_size + trunc_size;
 353                 }
 354
 355                 if (rank == write_rank || rank == 0)
 356                         memset(write_buf, write_char, write_max);
 357
 358                 if (rank == write_rank) {
 359                         ifnames = (classic_write ? 0 : rand()) % nfnames;
 360                         ret = truncate(fnames[ifnames], 0);
 361                         if (verbose > 1 || ret != 0)
 362                                 rprintf(rank, n, ret,
 363                                         "initial truncate %s (%u) @ 0: %s\n",
 364                                         fnames[ifnames], ifnames,
 365                                         strerror(errno));
 366
 367                         done = 0;
 368                         do {
 369                                 ret = write(fd, write_buf+done,write_size-done);
 370                                 if (verbose > 1 || ret < 0) {
 371                                         rprintf(rank, n,
 372                                                 ret < 0 && errno != EINTR,
 373                                                 "write %d/%d @ %d: %s\n",
 374                                                 ret + done, write_size, done,
 375                                                 strerror(errno));
 376                                         if (ret < 0 && errno != EINTR)
 377                                                 break;
 378                                 }
 379                                 if (ret > 0)
 380                                         done += ret;
 381                         } while (done != write_size);
 382                 }
 383
 384                 if (rank == append_rank || rank == 0)
 385                         memset(append_buf, append_char, append_size);
 386
 387                 error = MPI_Barrier(MPI_COMM_WORLD);
 388                 if (verbose > 2 || error != MPI_SUCCESS)
 389                         rprintf(rank, n, error != MPI_SUCCESS,
 390                                 "start MPI_Barrier: %d\n", error);
 391
 392                 /* Do the race */
 393                 if (rank == append_rank) {
 394                         done = 0;
 395                         do {
 396                                 ret = write(fd, append_buf + done,
 397                                             append_size - done);
 398                                 if (ret < 0) {
 399                                         rprintf(rank, n, errno != EINTR,
 400                                                 "append %u/%u: %s\n",
 401                                                 ret + done, append_size,
 402                                                 strerror(errno));
 403                                         if (errno != EINTR)
 404                                                 break;
 405                                 } else if (verbose > 1 || ret != append_size) {
 406                                         rprintf(rank, n, ret != append_size,
 407                                                 "append %u/%u\n",
 408                                                 ret + done, append_size);
 409                                 }
 410                                 if (ret > 0)
 411                                         done += ret;
 412                         } while (done != append_size);
 413                 } else if (rank == trunc_rank) {
 414                         /* XXX: truncating the same file descriptor as the
 415                          *      append on a single node causes this test
 416                          *      to fail currently (2009-02-01). */
 417                         ifnames = (classic_trunc ? rank : rand()) % nfnames;
 418                         ret = truncate(fnames[ifnames], trunc_offset);
 419                         if (verbose > 1 || ret != 0)
 420                                 rprintf(rank, n, ret,
 421                                         "truncate %s (%u) @ %u: %s\n",
 422                                         fnames[ifnames], ifnames,
 423                                         trunc_offset, strerror(errno));
 424                 }
 425
 426                 error = MPI_Barrier(MPI_COMM_WORLD);
 427                 if (verbose > 2 || error != MPI_SUCCESS)
 428                         rprintf(rank, n, error != MPI_SUCCESS,
 429                                 "end MPI_Barrier: %d\n", error);
 430
 431                 error = 0;
 432
 433                 /* Check the result */
 434                 if (rank == 0) {
 435                         char *tmp_buf;
 436                         struct stat st = { 0 };
 437
 438                         ifnames = classic_check ? 0 : (rand() % nfnames);
 439                         ret = stat(fnames[ifnames], &st);
 440                         if (verbose > 1 || ret != 0)
 441                                 rprintf(rank, n, ret,
 442                                         "stat %s (%u) size %llu: %s\n",
 443                                         fnames[ifnames], ifnames,
 444                                         (long long)st.st_size, strerror(errno));
 445
 446                         ret = lseek(fd, 0, SEEK_SET);
 447                         if (ret != 0)
 448                                 rprintf(rank, n, ret, "lseek 0: %s\n",
 449                                         strerror(errno));
 450
 451                         done = 0;
 452                         do {
 453                                 ret = read(fd, read_buf+done, st.st_size-done);
 454                                 if (verbose > 1 || ret <= 0) {
 455                                         rprintf(rank, n, ret <= 0,
 456                                                 "read %d/%llu @ %u: %s\n",
 457                                                 ret, (long long)st.st_size-done,
 458                                                 done, ret != 0 ?
 459                                                 strerror(errno) : "short read");
 460                                 }
 461                                 done += ret;
 462                         } while (done != st.st_size);
 463
 464                         if (memcmp(read_buf, write_buf, write_size)) {
 465                                 rprintf(rank, n, 0, "WRITE bad "
 466                                         "[0-%d]/[0-%#x] != %c\n",
 467                                         write_size - 1, write_size - 1,
 468                                         write_char);
 469                                 error = 1;
 470                         }
 471
 472                         tmp_buf = read_buf + write_size;
 473
 474                         if (st.st_size == trunc_offset) {
 475                                 /* Check case 1: first append then truncate */
 476                                 int tmp_size, tmp_offset;
 477
 478                                 tmp_size = trunc_size < append_size ?
 479                                                 trunc_size : append_size;
 480                                 tmp_offset = write_size + tmp_size;
 481
 482                                 if (memcmp(tmp_buf, append_buf, tmp_size)) {
 483                                         rprintf(rank, n, 0,"trunc-after-APPEND "
 484                                                 "bad [%d-%d]/[%#x-%#x] != %c\n",
 485                                                 write_size, tmp_offset - 1,
 486                                                 write_size, tmp_offset - 1,
 487                                                 append_char);
 488                                         error = 1;
 489                                 } else if (trunc_size > append_size &&
 490                                            memcmp(tmp_buf+append_size,trunc_buf,
 491                                                   trunc_size - append_size)) {
 492                                         rprintf(rank, n, 0,"TRUNC-after-append "
 493                                                 "bad [%d-%d]/[%#x-%#x] != 0\n",
 494                                                 tmp_offset, trunc_offset - 1,
 495                                                 tmp_offset, trunc_offset - 1);
 496                                         error = 1;
 497                                 }
 498                         } else {
 499                                 int expected_size = trunc_offset + append_size;
 500                                 /* Check case 2: first truncate then append */
 501                                 if (st.st_size != expected_size) {
 502                                         rprintf(rank, n, 0,"APPEND-after-trunc "
 503                                                 "bad file size %llu != %u\n",
 504                                                 (long long)st.st_size,
 505                                                 expected_size);
 506                                         error = 1;
 507                                 }
 508
 509                                 if (memcmp(tmp_buf, trunc_buf, trunc_size)) {
 510                                         rprintf(rank, n, 0,"append-after-TRUNC "
 511                                                 "bad [%d-%d]/[%#x-%#x] != 0\n",
 512                                                 write_size, trunc_offset - 1,
 513                                                 write_size, trunc_offset - 1);
 514                                         error = 1;
 515                                 } else if (memcmp(read_buf + trunc_offset,
 516                                                   append_buf, append_size)) {
 517                                         rprintf(rank, n, 0,"APPEND-after-trunc "
 518                                                 "bad [%d-%d]/[%#x-%#x] != %c\n",
 519                                                 trunc_offset, expected_size - 1,
 520                                                 trunc_offset, expected_size - 1,
 521                                                 append_char);
 522                                         error = 1;
 523                                 }
 524                         }
 525
 526                         if (error == 1) {
 527                                 char command[4096];
 528                                 int rc;
 529
 530                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 531                                         write_char, write_size, write_size,
 532                                         append_char, append_size, append_size,
 533                                         trunc_offset, trunc_offset);
 534
 535                                 sprintf(command, "od -Ax -a %s", fnames[0]);
 536                                 rc = system(command);
 537                                 MPI_Abort(MPI_COMM_WORLD, 1);
 538                         }
 539                 }
 540         }
 541
 542         if (rank == 0 || verbose)
 543                 printf("r=%2u n=%4u: "STATUS_FMT"\nPASS\n", rank, n - 1,
 544                        write_char, write_size, write_size,
 545                        append_char, append_size, append_size,
 546                        trunc_offset, trunc_offset);
 547
 548         close(fd);
 549
 550         if (rank == 0) {
 551                 ifnames = rand() % nfnames;
 552                 ret = unlink(fnames[ifnames]);
 553                 if (ret != 0)
 554                         printf("%s: unlink %s failed: %s\n",
 555                                prog, fnames[ifnames], strerror(errno));
 556         }
 557
 558         MPI_Finalize();
 559         return 0;
 560 }