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\t\t[-t trunc_max] [-T] [-v] [-w write_max] <filename> ...\n",
  78                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",
  83                DEFAULT_ITER);
  84         printf("\t-s seed: random seed to use, default {current time}\n");
  85         printf("\t-t trunc_max: maximum size of truncate, default %u bytes\n",
  86                TRUNC_SIZE_MAX);
  87         printf("\t-T: 'classic' truncates (on file 0)\n");
  88         printf("\t-w write_max: maximum size of write, default %u bytes\n",
  89                WRITE_SIZE_MAX);
  90         printf("\t-W: 'classic' writes (on rank 0, file 0)\n");
  91         printf("\t-v: run in verbose mode (repeat for more verbosity)\n");
  92         printf("\tfilename for each mountpoint of same filesystem on a node\n");
  93         printf("\b%s must be run with at least 2 processes\n", prog);
  94
  95         MPI_Finalize();
  96         exit(1);
  97 }
  98
  99 /* Print process rank, loop count, message, and exit (i.e. a fatal error) */
 100 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 101 __attribute__ ((format (printf, 4, 5)));
 102
 103 void rprintf(int rank, int loop, int error, const char *fmt, ...)
 104 {
 105         va_list ap;
 106
 107         printf("r=%2u", rank);
 108         if (loop >= 0)
 109                 printf(" l=%04u", loop);
 110         if (error != 0)
 111                 printf(" %s", hostname);
 112         printf(": ");
 113
 114         va_start(ap, fmt);
 115
 116         vprintf(fmt, ap);
 117
 118         if (error != 0)
 119                 MPI_Abort(MPI_COMM_WORLD, error);
 120 }
 121
 122 int main(int argc, char *argv[])
 123 {
 124         int n, nloops = DEFAULT_ITER;
 125         int nfnames = 0, ifnames, fd;
 126         int rank = -1, nproc, ret;
 127         unsigned int write_max = WRITE_SIZE_MAX;
 128         unsigned int append_max = APPEND_SIZE_MAX;
 129         unsigned int write_size = 0, append_size = 0, trunc_size = 0;
 130         unsigned int trunc_max = 0, trunc_offset = 0;
 131         char *append_buf;
 132         char *write_buf;
 133         char *read_buf = NULL;
 134         char *trunc_buf = NULL;
 135         int seed = time(0);
 136         int done;
 137         int error;
 138         int verbose = 0;
 139         int classic_check = 0, classic_trunc = 0, classic_write = 0;
 140         char write_char = 'A', append_char = 'a';
 141         char *fnames[FNAMES_MAX], *end;
 142         char *prog = "write_append_truncate";
 143         int c;
 144
 145         error = MPI_Init(&argc, &argv);
 146         if (error != MPI_SUCCESS)
 147                 printf("%s: MPI_Init failed: %d\n", prog, error);
 148         else if (verbose > 2)
 149                 printf("%s: MPI_Init succeeded\n", prog);
 150
 151         prog = strrchr(argv[0], '/');
 152         if (!prog)
 153                 prog = argv[0];
 154         else
 155                 prog++;
 156
 157         while ((c = getopt(argc, argv, "a:cCn:s:t:Tvw:W")) != -1) {
 158                 switch (c) {
 159                 case 'a':
 160                         append_max = strtoul(optarg, &end, 0);
 161                         if (append_max < 2 || *end) {
 162                                 fprintf(stderr, "%s: bad append option '%s'\n",
 163                                         prog, optarg);
 164                                 usage(prog);
 165                         }
 166                         break;
 167                 case 'C':
 168                         classic_check++;
 169                         break;
 170                 case 'n':
 171                         nloops = strtoul(optarg, &end, 0);
 172                         if (nloops == 0 || *end) {
 173                                 fprintf(stderr, "%s: bad nloops option '%s'\n",
 174                                         prog, optarg);
 175                                 usage(prog);
 176                         }
 177                         break;
 178                 case 's':
 179                         seed = strtoul(optarg, &end, 0);
 180                         if (*end) {
 181                                 fprintf(stderr, "%s: bad seed option '%s'\n",
 182                                         prog, optarg);
 183                                 usage(prog);
 184                         }
 185                         break;
 186                 case 't':
 187                         trunc_max = strtoul(optarg, &end, 0);
 188                         if (*end) {
 189                                 fprintf(stderr,
 190                                         "%s: bad truncate option '%s'\n", prog,
 191                                         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 < 2 || *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)
 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)
 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) +
 263                                append_max;
 264
 265                 fd = open(fnames[0], O_WRONLY | O_CREAT | O_TRUNC, 0666);
 266                 rprintf(rank, -1, fd < 0,
 267                         "create %s, max size: %u, seed %u: %s\n", fnames[0],
 268                         max_size, seed, strerror(errno));
 269                 close(fd);
 270
 271                 trunc_buf = calloc(1, trunc_max ?: append_max);
 272                 if (!trunc_buf)
 273                         rprintf(rank, -1, 1,
 274                                 "%s: error allocating trunc_buf %u\n",
 275                                 prog, trunc_max ?: append_max);
 276
 277                 /* initial write + truncate up + append */
 278                 read_buf = malloc(max_size);
 279                 if (!read_buf)
 280                         rprintf(rank, -1, 1,
 281                                 "%s: error allocating read_buf %u\n",
 282                                 prog, max_size);
 283         }
 284
 285         error = MPI_Barrier(MPI_COMM_WORLD);
 286         if (verbose > 2 || error != MPI_SUCCESS)
 287                 rprintf(rank, -1, error != MPI_SUCCESS,
 288                         "prep MPI_Barrier: %d\n", error);
 289
 290         ifnames = rank % nfnames;
 291         fd = open(fnames[ifnames], O_RDWR | O_APPEND);
 292         if (verbose || fd < 0)
 293                 rprintf(rank, -1, errno, "open '%s' (%u): %s\n",
 294                         fnames[ifnames], ifnames, strerror(errno));
 295
 296         for (n = 0; n < nloops; n++) {
 297                 /* Initialized only to quiet stupid GCC warnings */
 298                 unsigned int append_rank = n, trunc_rank = n + 1;
 299                 unsigned int write_rank = 0;
 300                 unsigned int mpi_shared_vars[6];
 301
 302                 /* reset the environment */
 303                 write_char = 'A' + (n % 26);
 304                 append_char = 'a' + (n % 26);
 305
 306                 if (rank == 0) {
 307                         write_size = (rand() % (write_max - 1)) + 1;
 308                         append_size = (rand() % (append_max - 1)) + 1;
 309                         trunc_size = (append_size == 1) ? 1 :
 310                                       (rand() %
 311                                       ((trunc_max ?: append_size) - 1)) + 1;
 312                         trunc_offset = write_size + trunc_size;
 313
 314                         if (verbose || n % 1000 == 0)
 315                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 316                                         write_char, write_size, write_size,
 317                                         append_char, append_size, append_size,
 318                                         trunc_offset, trunc_offset);
 319
 320                         write_rank = (classic_write ? 0 : rand()) % nproc;
 321                         do {
 322                                 append_rank = (classic_write ? n : rand()) %
 323                                                nproc;
 324                                 /*
 325                                  * We can't allow the append rank be the same
 326                                  * as the classic_trunc trunc_rank, or we will
 327                                  * spin here forever.
 328                                  */
 329                         } while (append_rank == (n + 1) % nproc);
 330                         do {
 331                                 trunc_rank = (classic_trunc ? (n + 1) :
 332                                               rand()) % nproc;
 333                         } while (trunc_rank == append_rank);
 334
 335                         mpi_shared_vars[0] = write_size;
 336                         mpi_shared_vars[1] = append_size;
 337                         mpi_shared_vars[2] = trunc_size;
 338                         mpi_shared_vars[3] = write_rank;
 339                         mpi_shared_vars[4] = append_rank;
 340                         mpi_shared_vars[5] = trunc_rank;
 341                 }
 342
 343                 error = MPI_Bcast(&mpi_shared_vars, 6,
 344                                   MPI_INT, 0, MPI_COMM_WORLD);
 345                 if (verbose > 2 || error != MPI_SUCCESS)
 346                         rprintf(rank, n, error != MPI_SUCCESS,
 347                                 "MPI_Bcast mpi_shared_vars [%u, %u, %u, %u, %u, %u]: %d\n",
 348                                 mpi_shared_vars[0], mpi_shared_vars[1],
 349                                 mpi_shared_vars[2], mpi_shared_vars[3],
 350                                 mpi_shared_vars[4], mpi_shared_vars[5], error);
 351
 352                 if (rank != 0) {
 353                         write_size  = mpi_shared_vars[0];
 354                         append_size = mpi_shared_vars[1];
 355                         trunc_size  = mpi_shared_vars[2];
 356                         write_rank  = mpi_shared_vars[3];
 357                         append_rank = mpi_shared_vars[4];
 358                         trunc_rank  = mpi_shared_vars[5];
 359
 360                         trunc_offset = write_size + trunc_size;
 361                 }
 362
 363                 if (rank == write_rank || rank == 0)
 364                         memset(write_buf, write_char, write_max);
 365
 366                 if (rank == write_rank) {
 367                         ifnames = (classic_write ? 0 : rand()) % nfnames;
 368                         ret = truncate(fnames[ifnames], 0);
 369                         if (verbose > 1 || ret != 0)
 370                                 rprintf(rank, n, ret,
 371                                         "initial truncate %s (%u) @ 0: %s\n",
 372                                         fnames[ifnames], ifnames,
 373                                         strerror(errno));
 374
 375                         done = 0;
 376                         do {
 377                                 ret = write(fd, write_buf + done,
 378                                             write_size - done);
 379                                 if (verbose > 1 || ret < 0) {
 380                                         rprintf(rank, n,
 381                                                 ret < 0 && errno != EINTR,
 382                                                 "write %d/%d @ %d: %s\n",
 383                                                 ret + done, write_size, done,
 384                                                 strerror(errno));
 385                                         if (ret < 0 && errno != EINTR)
 386                                                 break;
 387                                 }
 388                                 if (ret > 0)
 389                                         done += ret;
 390                         } while (done != write_size);
 391                 }
 392
 393                 if (rank == append_rank || rank == 0)
 394                         memset(append_buf, append_char, append_size);
 395
 396                 error = MPI_Barrier(MPI_COMM_WORLD);
 397                 if (verbose > 2 || error != MPI_SUCCESS)
 398                         rprintf(rank, n, error != MPI_SUCCESS,
 399                                 "start MPI_Barrier: %d\n", error);
 400
 401                 /* Do the race */
 402                 if (rank == append_rank) {
 403                         done = 0;
 404                         do {
 405                                 ret = write(fd, append_buf + done,
 406                                             append_size - done);
 407                                 if (ret < 0) {
 408                                         rprintf(rank, n, errno != EINTR,
 409                                                 "append %u/%u: %s\n",
 410                                                 ret + done, append_size,
 411                                                 strerror(errno));
 412                                         if (errno != EINTR)
 413                                                 break;
 414                                 } else if (verbose > 1 || ret != append_size) {
 415                                         rprintf(rank, n, ret != append_size,
 416                                                 "append %u/%u\n",
 417                                                 ret + done, append_size);
 418                                 }
 419                                 if (ret > 0)
 420                                         done += ret;
 421                         } while (done != append_size);
 422                 } else if (rank == trunc_rank) {
 423                         /*
 424                          * XXX: truncating the same file descriptor as the
 425                          *      append on a single node causes this test
 426                          *      to fail currently (2009-02-01).
 427                          */
 428                         ifnames = (classic_trunc ? rank : rand()) % nfnames;
 429                         ret = truncate(fnames[ifnames], trunc_offset);
 430                         if (verbose > 1 || ret != 0)
 431                                 rprintf(rank, n, ret,
 432                                         "truncate %s (%u) @ %u: %s\n",
 433                                         fnames[ifnames], ifnames,
 434                                         trunc_offset, strerror(errno));
 435                 }
 436
 437                 error = MPI_Barrier(MPI_COMM_WORLD);
 438                 if (verbose > 2 || error != MPI_SUCCESS)
 439                         rprintf(rank, n, error != MPI_SUCCESS,
 440                                 "end MPI_Barrier: %d\n", error);
 441
 442                 error = 0;
 443
 444                 /* Check the result */
 445                 if (rank == 0) {
 446                         char *tmp_buf;
 447                         struct stat st = { 0 };
 448
 449                         ifnames = classic_check ? 0 : (rand() % nfnames);
 450                         ret = stat(fnames[ifnames], &st);
 451                         if (verbose > 1 || ret != 0)
 452                                 rprintf(rank, n, ret,
 453                                         "stat %s (%u) size %llu: %s\n",
 454                                         fnames[ifnames], ifnames,
 455                                         (long long)st.st_size, strerror(errno));
 456
 457                         ret = lseek(fd, 0, SEEK_SET);
 458                         if (ret != 0)
 459                                 rprintf(rank, n, ret, "lseek 0: %s\n",
 460                                         strerror(errno));
 461
 462                         done = 0;
 463                         do {
 464                                 ret = read(fd, read_buf + done,
 465                                            st.st_size - done);
 466                                 if (verbose > 1 || ret <= 0) {
 467                                         rprintf(rank, n, ret <= 0,
 468                                                 "read %d/%llu @ %u: %s\n",
 469                                                 ret,
 470                                                 (long long)st.st_size - done,
 471                                                 done, ret != 0 ?
 472                                                 strerror(errno) : "short read");
 473                                 }
 474                                 done += ret;
 475                         } while (done != st.st_size);
 476
 477                         if (memcmp(read_buf, write_buf, write_size)) {
 478                                 rprintf(rank, n, 0,
 479                                         "WRITE bad [0-%d]/[0-%#x] != %c\n",
 480                                         write_size - 1, write_size - 1,
 481                                         write_char);
 482                                 error = 1;
 483                         }
 484
 485                         tmp_buf = read_buf + write_size;
 486
 487                         if (st.st_size == trunc_offset) {
 488                                 /* Check case 1: first append then truncate */
 489                                 int tmp_size, tmp_offset;
 490
 491                                 tmp_size = trunc_size < append_size ?
 492                                                 trunc_size : append_size;
 493                                 tmp_offset = write_size + tmp_size;
 494
 495                                 if (memcmp(tmp_buf, append_buf, tmp_size)) {
 496                                         rprintf(rank, n, 0,
 497                                                 "trunc-after-APPEND bad [%d-%d]/[%#x-%#x] != %c\n",
 498                                                 write_size, tmp_offset - 1,
 499                                                 write_size, tmp_offset - 1,
 500                                                 append_char);
 501                                         error = 1;
 502                                 } else if (trunc_size > append_size &&
 503                                            memcmp(tmp_buf + append_size,
 504                                                   trunc_buf,
 505                                                   trunc_size - append_size)) {
 506                                         rprintf(rank, n, 0,
 507                                                 "TRUNC-after-append bad [%d-%d]/[%#x-%#x] != 0\n",
 508                                                 tmp_offset, trunc_offset - 1,
 509                                                 tmp_offset, trunc_offset - 1);
 510                                         error = 1;
 511                                 }
 512                         } else {
 513                                 int expected_size = trunc_offset + append_size;
 514                                 /* Check case 2: first truncate then append */
 515                                 if (st.st_size != expected_size) {
 516                                         rprintf(rank, n, 0,
 517                                                 "APPEND-after-trunc bad file size %llu != %u\n",
 518                                                 (long long)st.st_size,
 519                                                 expected_size);
 520                                         error = 1;
 521                                 }
 522
 523                                 if (memcmp(tmp_buf, trunc_buf, trunc_size)) {
 524                                         rprintf(rank, n, 0,
 525                                                 "append-after-TRUNC bad [%d-%d]/[%#x-%#x] != 0\n",
 526                                                 write_size, trunc_offset - 1,
 527                                                 write_size, trunc_offset - 1);
 528                                         error = 1;
 529                                 } else if (memcmp(read_buf + trunc_offset,
 530                                                   append_buf, append_size)) {
 531                                         rprintf(rank, n, 0,
 532                                                 "APPEND-after-trunc bad [%d-%d]/[%#x-%#x] != %c\n",
 533                                                 trunc_offset, expected_size - 1,
 534                                                 trunc_offset, expected_size - 1,
 535                                                 append_char);
 536                                         error = 1;
 537                                 }
 538                         }
 539
 540                         if (error == 1) {
 541                                 char command[4096];
 542
 543                                 rprintf(rank, n, 0, STATUS_FMT"\n",
 544                                         write_char, write_size, write_size,
 545                                         append_char, append_size, append_size,
 546                                         trunc_offset, trunc_offset);
 547
 548                                 sprintf(command, "od -Ax -a %s", fnames[0]);
 549                                 ret = system(command);
 550                                 MPI_Abort(MPI_COMM_WORLD, 1);
 551                         }
 552                 }
 553         }
 554
 555         if (rank == 0 || verbose)
 556                 printf("r=%2u n=%4u: "STATUS_FMT"\nPASS\n", rank, n - 1,
 557                        write_char, write_size, write_size, append_char,
 558                        append_size, append_size, trunc_offset, trunc_offset);
 559
 560         close(fd);
 561
 562         if (rank == 0) {
 563                 ifnames = rand() % nfnames;
 564                 ret = unlink(fnames[ifnames]);
 565                 if (ret != 0)
 566                         printf("%s: unlink %s failed: %s\n",
 567                                prog, fnames[ifnames], strerror(errno));
 568         }
 569
 570         MPI_Finalize();
 571         return 0;
 572 }