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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/tests/write_append_truncate.c
  33  *
  34  * Each loop does 3 things:
  35  *   - truncate file to zero (not via ftruncate though, to test O_APPEND)
  36  *   - write a "chunk" of data (should be at file offset 0 after truncate)
  37  *   - on each of two threads either append or truncate-up the file
  38  *
  39  * If the truncate happened first, we should have a hole in the file.
  40  * If the append happened first, we should have truncated the file down.
  41  *
  42  * WRITE_SIZE_MAX and APPEND_SIZE_MAX are large enough to cross a stripe.
  43  *
  44  * compile: mpicc -g -Wall -o write_append_truncate write_append_truncate.c
  45  * run:     mpirun -np 2 -machlist <hostlist file> write_append_truncate <file>
  46  *  or:     pdsh -w <two hosts> write_append_truncate <file>
  47  *  or:     prun -n 2 [-N 2] write_append_truncate <file>
  48  */
  49
  50 #include <stdlib.h>
  51 #include <stdio.h>
  52 #include <stdarg.h>
  53 #include <sys/types.h>
  54 #include <sys/stat.h>
  55 #include <fcntl.h>
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <string.h>
  59 #include <unistd.h>
  60 #include <getopt.h>
  61 #include "mpi.h"
  62
  63 #define DEFAULT_ITER    10000
  64
  65 #define WRITE_SIZE_MAX  1234567
  66 #define APPEND_SIZE_MAX 1234567
  67 #define TRUNC_SIZE_MAX  1234567
  68
  69 #define STATUS_FMT "WR %c %7d/%#08x, AP %c %7d/%#08x, TR@ %7d/%#08x"
  70
  71 #define HOSTNAME_SIZE 50
  72 char hostname[HOSTNAME_SIZE];
  73 #define FNAMES_MAX 256
  74
  75 void usage(char *prog)
  76 {
  77         printf("usage: %s [-a append_max] [-C] [-n nloops] [-s seed]\n"
  78                "\t\t[-t trunc_max] [-T] [-v] [-w write_max] <filename> ...\n", prog);
  79         printf("\t-a append_max: maximum size of append, default %u bytes\n",
  80                APPEND_SIZE_MAX);
  81         printf("\t-C: 'classic' checks (on file 0)\n");
  82         printf("\t-n nloops: count of loops to run, default %u\n",DEFAULT_ITER);
  83         printf("\t-s seed: random seed to use, default {current time}\n");
  84         printf("\t-t trunc_max: maximum size of truncate, default %u bytes\n",
  85                TRUNC_SIZE_MAX);
  86         printf("\t-T: 'classic' truncates (on file 0)\n");
  87         printf("\t-w write_max: maximum size of write, default %u bytes\n",
  88                WRITE_SIZE_MAX);
  89         printf("\t-W: 'classic' writes (on rank 0, file 0)\n");
  90         printf("\t-v: run in verbose mode (repeat for more verbosity)\n");
  91         printf("\tfilename for each mountpoint of same filesystem on a node\n");
  92         printf("\b%s must be run with at least 2 processes\n", prog);
  93
  94         MPI_Finalize();
  95         exit(1);
  96 }
  97
  98 /* Print process rank, loop count, message, and exit (i.e. a fatal error) */
  99 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 100 __attribute__ ((format (printf, 4, 5)));
 101
 102 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 103 {
 104         va_list       ap;
 105
 106         printf("r=%2u", rank);
 107         if (loop >= 0)
 108                 printf(" l=%04u", loop);
 109         if (error != 0)
 110                 printf(" %s", hostname);
 111         printf(": ");
 112
 113         va_start(ap, fmt);
 114
 115         vprintf(fmt, ap);
 116
 117         if (error != 0)
 118                 MPI_Abort(MPI_COMM_WORLD, error);
 119 }
 120
 121 int main(int argc, char *argv[])
 122 {
 123         int n, nloops = DEFAULT_ITER;
 124         int nfnames = 0, ifnames, fd;
 125         int rank = -1, nproc, ret;
 126         unsigned write_max = WRITE_SIZE_MAX;
 127         unsigned append_max = APPEND_SIZE_MAX;
 128         unsigned write_size = 0, append_size = 0, trunc_size = 0;
 129         unsigned trunc_max = 0, trunc_offset = 0;
 130         char *append_buf;
 131         char *write_buf;
 132         char *read_buf = NULL;
 133         char *trunc_buf = NULL;
 134         int seed = time(0);
 135         int done;
 136         int error;
 137         int verbose = 0;
 138         int classic_check = 0, classic_trunc = 0, classic_write = 0;
 139         char write_char = 'A', append_char = 'a';
 140         char *fnames[FNAMES_MAX], *end;
 141         char *prog = "write_append_truncate";
 142         int c;
 143
 144         error = MPI_Init(&argc, &argv);
 145         if (error != MPI_SUCCESS)
 146                 printf("%s: MPI_Init failed: %d\n", prog, error);
 147         else if (verbose > 2)
 148                 printf("%s: MPI_Init succeeded\n", prog);
 149
 150         prog = strrchr(argv[0], '/');
 151         if (prog == NULL)
 152                 prog = argv[0];
 153         else
 154                 prog++;
 155
 156         while ((c = getopt(argc, argv, "a:cCn:s:t:Tvw:W")) != -1) {
 157                 switch(c) {
 158                 case 'a':
 159                         append_max = strtoul(optarg, &end, 0);
 160                         if (append_max < 2 || *end) {
 161                                 fprintf(stderr, "%s: bad append option '%s'\n",
 162                                         prog, optarg);
 163                                 usage(prog);
 164                         }
 165                         break;
 166                 case 'C':
 167                         classic_check++;
 168                         break;
 169                 case 'n':
 170                         nloops = strtoul(optarg, &end, 0);
 171                         if (nloops == 0 || *end) {
 172                                 fprintf(stderr, "%s: bad nloops option '%s'\n",
 173                                         prog, optarg);
 174                                 usage(prog);
 175                         }
 176                         break;
 177                 case 's':
 178                         seed = strtoul(optarg, &end, 0);
 179                         if (*end) {
 180                                 fprintf(stderr, "%s: bad seed option '%s'\n",
 181                                         prog, optarg);
 182                                 usage(prog);
 183                         }
 184                         break;
 185                 case 't':
 186                         trunc_max = strtoul(optarg, &end, 0);
 187                         if (*end) {
 188                                 fprintf(stderr,"%s: bad truncate option '%s'\n",
 189                                         prog, optarg);
 190                                 usage(prog);
 191                         }
 192                         break;
 193                 case 'T':
 194                         classic_trunc++;
 195                         break;
 196                 case 'v':
 197                         verbose++;
 198                         break;
 199                 case 'w':
 200                         write_max = strtoul(optarg, &end, 0);
 201                         if (write_max < 2 || *end) {
 202                                 fprintf(stderr, "%s: bad write option '%s'\n",
 203                                         prog, optarg);
 204                                 usage(prog);
 205                         }
 206                         break;
 207                 case 'W':
 208                         classic_write++;
 209                         break;
 210                 default:
 211                         fprintf(stderr, "%s: unknown option '%c'\n", prog, c);
 212                         usage(prog);
 213                 }
 214         }
 215
 216         srand(seed);
 217
 218         if (argc == optind) {
 219                 fprintf(stderr, "%s: missing filename argument\n", prog);
 220                 usage(prog);
 221         }
 222
 223         if (argc > optind + FNAMES_MAX) {
 224                 fprintf(stderr, "%s: too many extra options\n", prog);
 225                 usage(prog);
 226         }
 227
 228         while (optind < argc)
 229                 fnames[nfnames++] = argv[optind++];
 230
 231         error = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
 232         if (verbose > 2 || error != MPI_SUCCESS)
 233                 rprintf(rank, -1, error != MPI_SUCCESS, "MPI_Comm_rank: %d\n",
 234                         error);
 235
 236         error = MPI_Comm_size(MPI_COMM_WORLD, &nproc);
 237         if (verbose > 2 || error != MPI_SUCCESS)
 238                 rprintf(rank, -1, error != MPI_SUCCESS, "MPI_Comm_size: %d\n",
 239                         error);
 240
 241         if (nproc < 2)
 242                 rprintf(rank, -1, 1, "%s: must run with at least 2 processes\n",
 243                         prog);
 244
 245         append_buf = malloc(append_max);
 246         if (append_buf == NULL)
 247                 rprintf(rank, -1, 1,"%s: error allocating append_buf %u\n",
 248                         prog, append_max);
 249
 250         write_buf = malloc(write_max);
 251         if (write_buf == NULL)
 252                 rprintf(rank, -1, 1, "%s: error allocating write_buf %u\n",
 253                         prog, write_max);
 254
 255         if (gethostname(hostname, HOSTNAME_SIZE) < 0)
 256                 rprintf(rank, -1, 1, "%s: gethostname failed: %s\n",
 257                         prog, strerror(errno));
 258
 259         if (rank == 0) {
 260                 int max_size = write_max + (trunc_max ?: append_max)+append_max;
 261
 262                 fd = open(fnames[0], O_WRONLY|O_CREAT|O_TRUNC, 0666);
 263                 rprintf(rank,-1, fd<0, "create %s, max size: %u, seed %u: %s\n",
 264                         fnames[0], max_size, seed, strerror(errno));
 265                 close(fd);
 266
 267                 trunc_buf = calloc(1, trunc_max ?: append_max);
 268                 if (trunc_buf == NULL)
 269                         rprintf(rank,-1,1,"%s: error allocating trunc_buf %u\n",
 270                                 prog, trunc_max ?: append_max);
 271
 272                 /* initial write + truncate up + append */
 273                 read_buf = malloc(max_size);
 274                 if (read_buf == NULL)
 275                         rprintf(rank,-1,1,"%s: error allocating read_buf %u\n",
 276                                 prog, max_size);
 277         }
 278
 279         error = MPI_Barrier(MPI_COMM_WORLD);
 280         if (verbose > 2 || error != MPI_SUCCESS)
 281                 rprintf(rank, -1, error != MPI_SUCCESS,
 282                         "prep MPI_Barrier: %d\n", error);
 283
 284         ifnames = rank % nfnames;
 285         fd = open(fnames[ifnames], O_RDWR | O_APPEND);
 286         if (verbose || fd < 0)
 287                 rprintf(rank, -1, errno, "open '%s' (%u): %s\n",
 288                         fnames[ifnames], ifnames, strerror(errno));
 289
 290         for (n = 0; n < nloops; n++) {
 291                 /* Initialized only to quiet stupid GCC warnings */
 292                 unsigned write_rank = 0, append_rank = n, trunc_rank = n + 1;
 293                 unsigned mpi_shared_vars[6];
 294
 295                 /* reset the environment */
 296                 write_char = 'A' + (n % 26);
 297                 append_char = 'a' + (n % 26);
 298
 299                 if (rank == 0) {
 300                         write_size = (rand() % (write_max - 1)) + 1;
 301                         append_size = (rand() % (append_max - 1)) + 1;
 302                         trunc_size = (append_size == 1) ? 1 :
 303                                      (rand() % ((trunc_max?: append_size)-1))+1;
 304                         trunc_offset = write_size + trunc_size;
 305
 306                         if (verbose || n % 1000 == 0)
 307                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 308                                         write_char, write_size, write_size,
 309                                         append_char, append_size, append_size,
 310                                         trunc_offset, trunc_offset);
 311
 312                         write_rank = (classic_write ? 0 : rand()) % nproc;
 313                         do {
 314                                 append_rank = (classic_write ? n : rand()) %
 315                                                                 nproc;
 316                                 /* We can't allow the append rank be the same
 317                                  * as the classic_trunc trunc_rank, or we will
 318                                  * spin here forever. */
 319                         } while (append_rank == (n + 1) % nproc);
 320                         do {
 321                                 trunc_rank = (classic_trunc? (n + 1) : rand()) %
 322                                                                 nproc;
 323                         } while (trunc_rank == append_rank);
 324
 325                         mpi_shared_vars[0] = write_size;
 326                         mpi_shared_vars[1] = append_size;
 327                         mpi_shared_vars[2] = trunc_size;
 328                         mpi_shared_vars[3] = write_rank;
 329                         mpi_shared_vars[4] = append_rank;
 330                         mpi_shared_vars[5] = trunc_rank;
 331                 }
 332
 333                 error = MPI_Bcast(&mpi_shared_vars, 6,
 334                                   MPI_INT, 0, MPI_COMM_WORLD);
 335                 if (verbose > 2 || error != MPI_SUCCESS)
 336                         rprintf(rank, n, error != MPI_SUCCESS,
 337                                 "MPI_Bcast mpi_shared_vars"
 338                                 "[%u, %u, %u, %u, %u, %u]: %d\n",
 339                                 mpi_shared_vars[0], mpi_shared_vars[1],
 340                                 mpi_shared_vars[2], mpi_shared_vars[3],
 341                                 mpi_shared_vars[4], mpi_shared_vars[5], error);
 342
 343                 if (rank != 0) {
 344                         write_size  = mpi_shared_vars[0];
 345                         append_size = mpi_shared_vars[1];
 346                         trunc_size  = mpi_shared_vars[2];
 347                         write_rank  = mpi_shared_vars[3];
 348                         append_rank = mpi_shared_vars[4];
 349                         trunc_rank  = mpi_shared_vars[5];
 350
 351                         trunc_offset = write_size + trunc_size;
 352                 }
 353
 354                 if (rank == write_rank || rank == 0)
 355                         memset(write_buf, write_char, write_max);
 356
 357                 if (rank == write_rank) {
 358                         ifnames = (classic_write ? 0 : rand()) % nfnames;
 359                         ret = truncate(fnames[ifnames], 0);
 360                         if (verbose > 1 || ret != 0)
 361                                 rprintf(rank, n, ret,
 362                                         "initial truncate %s (%u) @ 0: %s\n",
 363                                         fnames[ifnames], ifnames,
 364                                         strerror(errno));
 365
 366                         done = 0;
 367                         do {
 368                                 ret = write(fd, write_buf+done,write_size-done);
 369                                 if (verbose > 1 || ret < 0) {
 370                                         rprintf(rank, n,
 371                                                 ret < 0 && errno != EINTR,
 372                                                 "write %d/%d @ %d: %s\n",
 373                                                 ret + done, write_size, done,
 374                                                 strerror(errno));
 375                                         if (ret < 0 && errno != EINTR)
 376                                                 break;
 377                                 }
 378                                 if (ret > 0)
 379                                         done += ret;
 380                         } while (done != write_size);
 381                 }
 382
 383                 if (rank == append_rank || rank == 0)
 384                         memset(append_buf, append_char, append_size);
 385
 386                 error = MPI_Barrier(MPI_COMM_WORLD);
 387                 if (verbose > 2 || error != MPI_SUCCESS)
 388                         rprintf(rank, n, error != MPI_SUCCESS,
 389                                 "start MPI_Barrier: %d\n", error);
 390
 391                 /* Do the race */
 392                 if (rank == append_rank) {
 393                         done = 0;
 394                         do {
 395                                 ret = write(fd, append_buf + done,
 396                                             append_size - done);
 397                                 if (ret < 0) {
 398                                         rprintf(rank, n, errno != EINTR,
 399                                                 "append %u/%u: %s\n",
 400                                                 ret + done, append_size,
 401                                                 strerror(errno));
 402                                         if (errno != EINTR)
 403                                                 break;
 404                                 } else if (verbose > 1 || ret != append_size) {
 405                                         rprintf(rank, n, ret != append_size,
 406                                                 "append %u/%u\n",
 407                                                 ret + done, append_size);
 408                                 }
 409                                 if (ret > 0)
 410                                         done += ret;
 411                         } while (done != append_size);
 412                 } else if (rank == trunc_rank) {
 413                         /* XXX: truncating the same file descriptor as the
 414                          *      append on a single node causes this test
 415                          *      to fail currently (2009-02-01). */
 416                         ifnames = (classic_trunc ? rank : rand()) % nfnames;
 417                         ret = truncate(fnames[ifnames], trunc_offset);
 418                         if (verbose > 1 || ret != 0)
 419                                 rprintf(rank, n, ret,
 420                                         "truncate %s (%u) @ %u: %s\n",
 421                                         fnames[ifnames], ifnames,
 422                                         trunc_offset, strerror(errno));
 423                 }
 424
 425                 error = MPI_Barrier(MPI_COMM_WORLD);
 426                 if (verbose > 2 || error != MPI_SUCCESS)
 427                         rprintf(rank, n, error != MPI_SUCCESS,
 428                                 "end MPI_Barrier: %d\n", error);
 429
 430                 error = 0;
 431
 432                 /* Check the result */
 433                 if (rank == 0) {
 434                         char *tmp_buf;
 435                         struct stat st = { 0 };
 436
 437                         ifnames = classic_check ? 0 : (rand() % nfnames);
 438                         ret = stat(fnames[ifnames], &st);
 439                         if (verbose > 1 || ret != 0)
 440                                 rprintf(rank, n, ret,
 441                                         "stat %s (%u) size %llu: %s\n",
 442                                         fnames[ifnames], ifnames,
 443                                         (long long)st.st_size, strerror(errno));
 444
 445                         ret = lseek(fd, 0, SEEK_SET);
 446                         if (ret != 0)
 447                                 rprintf(rank, n, ret, "lseek 0: %s\n",
 448                                         strerror(errno));
 449
 450                         done = 0;
 451                         do {
 452                                 ret = read(fd, read_buf+done, st.st_size-done);
 453                                 if (verbose > 1 || ret <= 0) {
 454                                         rprintf(rank, n, ret <= 0,
 455                                                 "read %d/%llu @ %u: %s\n",
 456                                                 ret, (long long)st.st_size-done,
 457                                                 done, ret != 0 ?
 458                                                 strerror(errno) : "short read");
 459                                 }
 460                                 done += ret;
 461                         } while (done != st.st_size);
 462
 463                         if (memcmp(read_buf, write_buf, write_size)) {
 464                                 rprintf(rank, n, 0, "WRITE bad "
 465                                         "[0-%d]/[0-%#x] != %c\n",
 466                                         write_size - 1, write_size - 1,
 467                                         write_char);
 468                                 error = 1;
 469                         }
 470
 471                         tmp_buf = read_buf + write_size;
 472
 473                         if (st.st_size == trunc_offset) {
 474                                 /* Check case 1: first append then truncate */
 475                                 int tmp_size, tmp_offset;
 476
 477                                 tmp_size = trunc_size < append_size ?
 478                                                 trunc_size : append_size;
 479                                 tmp_offset = write_size + tmp_size;
 480
 481                                 if (memcmp(tmp_buf, append_buf, tmp_size)) {
 482                                         rprintf(rank, n, 0,"trunc-after-APPEND "
 483                                                 "bad [%d-%d]/[%#x-%#x] != %c\n",
 484                                                 write_size, tmp_offset - 1,
 485                                                 write_size, tmp_offset - 1,
 486                                                 append_char);
 487                                         error = 1;
 488                                 } else if (trunc_size > append_size &&
 489                                            memcmp(tmp_buf+append_size,trunc_buf,
 490                                                   trunc_size - append_size)) {
 491                                         rprintf(rank, n, 0,"TRUNC-after-append "
 492                                                 "bad [%d-%d]/[%#x-%#x] != 0\n",
 493                                                 tmp_offset, trunc_offset - 1,
 494                                                 tmp_offset, trunc_offset - 1);
 495                                         error = 1;
 496                                 }
 497                         } else {
 498                                 int expected_size = trunc_offset + append_size;
 499                                 /* Check case 2: first truncate then append */
 500                                 if (st.st_size != expected_size) {
 501                                         rprintf(rank, n, 0,"APPEND-after-trunc "
 502                                                 "bad file size %llu != %u\n",
 503                                                 (long long)st.st_size,
 504                                                 expected_size);
 505                                         error = 1;
 506                                 }
 507
 508                                 if (memcmp(tmp_buf, trunc_buf, trunc_size)) {
 509                                         rprintf(rank, n, 0,"append-after-TRUNC "
 510                                                 "bad [%d-%d]/[%#x-%#x] != 0\n",
 511                                                 write_size, trunc_offset - 1,
 512                                                 write_size, trunc_offset - 1);
 513                                         error = 1;
 514                                 } else if (memcmp(read_buf + trunc_offset,
 515                                                   append_buf, append_size)) {
 516                                         rprintf(rank, n, 0,"APPEND-after-trunc "
 517                                                 "bad [%d-%d]/[%#x-%#x] != %c\n",
 518                                                 trunc_offset, expected_size - 1,
 519                                                 trunc_offset, expected_size - 1,
 520                                                 append_char);
 521                                         error = 1;
 522                                 }
 523                         }
 524
 525                         if (error == 1) {
 526                                 char command[4096];
 527
 528                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 529                                         write_char, write_size, write_size,
 530                                         append_char, append_size, append_size,
 531                                         trunc_offset, trunc_offset);
 532
 533                                 sprintf(command, "od -Ax -a %s", fnames[0]);
 534                                 ret = system(command);
 535                                 MPI_Abort(MPI_COMM_WORLD, 1);
 536                         }
 537                 }
 538         }
 539
 540         if (rank == 0 || verbose)
 541                 printf("r=%2u n=%4u: "STATUS_FMT"\nPASS\n", rank, n - 1,
 542                        write_char, write_size, write_size,
 543                        append_char, append_size, append_size,
 544                        trunc_offset, trunc_offset);
 545
 546         close(fd);
 547
 548         if (rank == 0) {
 549                 ifnames = rand() % nfnames;
 550                 ret = unlink(fnames[ifnames]);
 551                 if (ret != 0)
 552                         printf("%s: unlink %s failed: %s\n",
 553                                prog, fnames[ifnames], strerror(errno));
 554         }
 555
 556         MPI_Finalize();
 557         return 0;
 558 }