lustre/utils/llverfs.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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 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/utils/llverfs.c
  33  *
  34  * Filesystem Verification Tool.
  35  * This program tests the correct operation of large filesystems and
  36  * the underlying block storage device(s).
  37  * This tool have two working modes
  38  * 1. full mode
  39  * 2. partial mode
  40  *
  41  * In full mode, the program creates a subdirectory in the test
  42  * filesystem, writes n(files_in_dir, default=32) large(4GB) files to
  43  * the directory with the test pattern at the start of each 4kb block.
  44  * The test pattern contains timestamp, relative file offset and per
  45  * file unique identifier(inode number).  This continues until the
  46  * whole filesystem is full and then the tool verifies that the data
  47  * in all of the test files is correct.
  48  *
  49  * In partial mode, the tool creates test directories with the
  50  * EXT3_TOPDIR_FL flag set (if supported) to spread the directory data
  51  * around the block device instead of localizing it in a single place.
  52  * The number of directories equals to the number of block groups in the
  53  * filesystem (e.g. 65536 directories for 8TB ext3/ext4 filesystem) and
  54  * then writes a single 1MB file in each directory. The tool then verifies
  55  * that the data in each file is correct.
  56  */
  57
  58 #ifndef _GNU_SOURCE
  59 #define _GNU_SOURCE
  60 #endif
  61 #ifndef LUSTRE_UTILS
  62 #define LUSTRE_UTILS
  63 #endif
  64 #ifndef _LARGEFILE64_SOURCE
  65 #define _LARGEFILE64_SOURCE
  66 #endif
  67 #ifndef _FILE_OFFSET_BITS
  68 #define _FILE_OFFSET_BITS 64
  69 #endif
  70
  71 #include <features.h>
  72 #include <stdlib.h>
  73 #include <stdio.h>
  74 #include <string.h>
  75 #include <ctype.h>
  76 #include <fcntl.h>
  77 #include <unistd.h>
  78 #include <limits.h>
  79 #include <errno.h>
  80 #include <fcntl.h>
  81 #include <getopt.h>
  82 #include <time.h>
  83 #include <dirent.h>
  84 #include <mntent.h>
  85 #include <sys/types.h>
  86 #include <sys/stat.h>
  87 #include <sys/vfs.h>
  88 #include <sys/time.h>
  89 #include <gnu/stubs.h>
  90 #include <gnu/stubs.h>
  91
  92 #ifdef HAVE_EXT2FS_EXT2FS_H
  93 #  include <e2p/e2p.h>
  94 #  include <ext2fs/ext2fs.h>
  95 #else
  96 #  ifndef EXT2_TOPDIR_FL
  97 #    define EXT2_TOPDIR_FL              0x00020000 /* Top of directory tree */
  98 #  endif
  99 static int fsetflags(const char *path, unsigned int flag)
 100 {
 101         char cmd[PATH_MAX + 128];
 102         int rc;
 103
 104         if (flag != EXT2_TOPDIR_FL) {
 105                 rc = EOPNOTSUPP;
 106                 goto out;
 107         }
 108
 109         snprintf(cmd, sizeof(cmd), "chattr +T %s", path);
 110
 111         rc = system(cmd);
 112         if (rc > 0) {
 113                 rc = WEXITSTATUS(rc);
 114 out:
 115                 errno = rc;
 116         }
 117
 118         return rc;
 119 }
 120 #endif
 121
 122 #define ONE_MB (1024 * 1024)
 123 #define ONE_GB ((unsigned long long)(1024 * 1024 * 1024))
 124 #define BLOCKSIZE 4096
 125
 126 /* Structure for writing test pattern */
 127 struct block_data {
 128         unsigned long long bd_offset;
 129         unsigned long long bd_time;
 130         unsigned long long bd_inode;
 131 };
 132 static char *progname;              /* name by which this program was run. */
 133 static unsigned verbose = 1;        /* prints offset in kB, operation rate */
 134 static int readoption;              /* run test in read-only (verify) mode */
 135 static int writeoption;             /* run test in write_only mode */
 136 char *testdir;                      /* name of device to be tested. */
 137 static unsigned full = 1;           /* flag to full check */
 138 static int error_count;             /* number of IO errors hit during run */
 139 char filecount[PATH_MAX];           /* file with total number of files written*/
 140 static unsigned long num_files;     /* Total number of files for read/write */
 141 static loff_t file_size = 4*ONE_GB; /* Size of each file */
 142 static unsigned files_in_dir = 32;  /* number of files in each directioy */
 143 static unsigned num_dirs = 30000;   /* total number of directories */
 144 const int dirmode = S_IRWXU | S_IRGRP | S_IXGRP | S_IROTH | S_IXOTH;
 145 static int isatty_flag;
 146 static int perms =  S_IRWXU | S_IRGRP | S_IROTH;
 147
 148 static struct option const long_opts[] = {
 149         { .val = 'c',   .name = "chunksize",    .has_arg = required_argument },
 150         { .val = 'h',   .name = "help",         .has_arg = no_argument },
 151         { .val = 'l',   .name = "long",         .has_arg = no_argument },
 152         { .val = 'l',   .name = "full",         .has_arg = no_argument },
 153         { .val = 'o',   .name = "offset",       .has_arg = required_argument },
 154         { .val = 'p',   .name = "partial",      .has_arg = required_argument },
 155         { .val = 'q',   .name = "quiet",        .has_arg = required_argument },
 156         { .val = 'r',   .name = "read",         .has_arg = no_argument },
 157         { .val = 's',   .name = "filesize",     .has_arg = no_argument },
 158         { .val = 't',   .name = "timestamp",    .has_arg = required_argument },
 159         { .val = 'v',   .name = "verbose",      .has_arg = no_argument },
 160         { .val = 'w',   .name = "write",        .has_arg = no_argument },
 161         { .name = NULL } };
 162
 163 /*
 164  * Usages: displays help information, whenever user supply --help option in
 165  * command or enters incorrect command line.
 166  */
 167 void usage(int status)
 168 {
 169         if (status != 0) {
 170                 printf("\nUsage: %s [OPTION]... <filesystem path> ...\n",
 171                        progname);
 172                 printf("Filesystem verification tool.\n"
 173                        "\t-t {seconds}, --timestamp,  set test time"
 174                        "(default=current time())\n"
 175                        "\t-o {offset}, --offset, directory starting offset"
 176                        " from which tests should start\n"
 177                        "\t-r, --read, run in verify mode\n"
 178                        "\t-w, --write, run in test-pattern mode, default=rw\n"
 179                        "\t-v, --verbose\n"
 180                        "\t-p, --partial, for partial check (1MB files)\n"
 181                        "\t-l, --long, --full check (4GB file with 4k blocks)\n"
 182                        "\t-c, --chunksize, IO chunk size in MB (default=1)\n"
 183                        "\t-s, --filesize, file size in MB (default=4096)\n"
 184                        "\t-h, --help, display this help and exit\n");
 185         }
 186         exit(status);
 187 }
 188
 189 /*
 190  * open_file: Opens file in specified mode and returns fd.
 191  */
 192 static int open_file(const char *file, int flag)
 193 {
 194         int fd = open(file, flag, perms);
 195         if (fd < 0) {
 196                 fprintf(stderr, "\n%s: Open '%s' failed:%s\n",
 197                         progname, file, strerror(errno));
 198         }
 199         return (fd);
 200 }
 201
 202 /*
 203  * Verify_chunk: Verifies test pattern in each 4kB (BLOCKSIZE) is correct.
 204  * Returns 0 if test offset and timestamp is correct otherwise 1.
 205  */
 206 int verify_chunk(char *chunk_buf, const size_t chunksize,
 207                  unsigned long long chunk_off, const unsigned long long time_st,
 208                  const unsigned long long inode_st, const char *file)
 209 {
 210         struct block_data *bd;
 211         char *chunk_end;
 212
 213         for (chunk_end = chunk_buf + chunksize - sizeof(*bd);
 214              (char *)chunk_buf < chunk_end;
 215              chunk_buf += BLOCKSIZE, chunk_off += BLOCKSIZE) {
 216                 bd = (struct block_data *)chunk_buf;
 217                 if ((bd->bd_offset == chunk_off) && (bd->bd_time == time_st) &&
 218                     (bd->bd_inode == inode_st))
 219                         continue;
 220                 fprintf(stderr, "\n%s: verify %s failed offset/timestamp/inode "
 221                         "%llu/%llu/%llu: found %llu/%llu/%llu instead\n",
 222                         progname, file, chunk_off, time_st, inode_st,
 223                         bd->bd_offset, bd->bd_time, bd->bd_inode);
 224                 return 1;
 225         }
 226         return 0;
 227 }
 228
 229 /*
 230  * fill_chunk: Fills the chunk with current or user specified timestamp
 231  * and offset. The test pattern is filled at the beginning of
 232  * each 4kB(BLOCKSIZE) blocks in chunk_buf.
 233  */
 234 void fill_chunk(char *chunk_buf, size_t chunksize, loff_t chunk_off,
 235                 const time_t time_st, const ino_t inode_st)
 236 {
 237         struct block_data *bd;
 238         char *chunk_end;
 239
 240         for (chunk_end = chunk_buf + chunksize - sizeof(*bd);
 241              (char *)chunk_buf < chunk_end;
 242              chunk_buf += BLOCKSIZE, chunk_off += BLOCKSIZE) {
 243                 bd = (struct block_data *)chunk_buf;
 244                 bd->bd_offset = chunk_off;
 245                 bd->bd_time = time_st;
 246                 bd->bd_inode = inode_st;
 247         }
 248 }
 249
 250 /*
 251  * Write a chunk to disk, handling errors, interrupted writes, etc.
 252  *
 253  * If there is an IO error hit during the write, it is possible that
 254  * this will just show up as a short write, and a subsequent write
 255  * will return the actual error.  We want to continue in the face of
 256  * minor media errors so that we can validate the whole device if
 257  * possible, but if there are many errors we don't want to loop forever.
 258  *
 259  * The error count will be returned upon exit to ensure that the
 260  * media errors are detected even if nobody is looking at the output.
 261  *
 262  * Returns 0 on success, or -ve errno on failure.
 263  */
 264 int write_retry(int fd, const char *chunk_buf, size_t nrequested,
 265                 unsigned long long offset, const char *file)
 266 {
 267         long nwritten;
 268
 269 retry:
 270         nwritten = write(fd, chunk_buf, nrequested);
 271         if (nwritten < 0) {
 272                 if (errno != ENOSPC) {
 273                         fprintf(stderr, "\n%s: write %s@%llu+%zi failed: %s\n",
 274                                 progname, file, offset, nrequested,
 275                                 strerror(errno));
 276                         if (error_count++ < 100)
 277                                 return 0;
 278                 }
 279                 return -errno;
 280         }
 281         if (nwritten < nrequested) {
 282                 fprintf(stderr, "\n%s: write %s@%llu+%zi short: %ld written\n",
 283                         progname, file, offset, nrequested, nwritten);
 284                 offset += nwritten;
 285                 chunk_buf += nwritten;
 286                 nrequested -= nwritten;
 287                 goto retry;
 288         }
 289
 290         return 0;
 291 }
 292
 293 /*
 294  * write_chunks: write the chunk_buf on the device. The number of write
 295  * operations are based on the parameters write_end, offset, and chunksize.
 296  *
 297  * Returns 0 on success, or -ve error number on failure.
 298  */
 299 int write_chunks(int fd, unsigned long long offset,unsigned long long write_end,
 300                  char *chunk_buf, size_t chunksize, const time_t time_st,
 301                  const ino_t inode_st, const char *file)
 302 {
 303         unsigned long long stride;
 304
 305         stride = full ? chunksize : (ONE_GB - chunksize);
 306         for (offset = offset & ~(chunksize - 1); offset < write_end;
 307              offset += stride) {
 308                 int ret;
 309
 310                 if (stride != chunksize && lseek64(fd, offset, SEEK_SET) < 0) {
 311                         fprintf(stderr, "\n%s: lseek66(%s+%llu) failed: %s\n",
 312                                 progname, file, offset, strerror(errno));
 313                         return -errno;
 314                 }
 315                 if (offset + chunksize > write_end)
 316                         chunksize = write_end - offset;
 317                 if (!full && offset > chunksize) {
 318                         fill_chunk(chunk_buf, chunksize, offset, time_st,
 319                                    inode_st);
 320                         ret = write_retry(fd, chunk_buf, chunksize,offset,file);
 321                         if (ret < 0)
 322                                 return ret;
 323                         offset += chunksize;
 324                         if (offset + chunksize > write_end)
 325                                 chunksize = write_end - offset;
 326                 }
 327                 fill_chunk(chunk_buf, chunksize, offset, time_st, inode_st);
 328                 ret = write_retry(fd, chunk_buf, chunksize, offset, file);
 329                 if (ret < 0)
 330                         return ret;
 331         }
 332         return 0;
 333 }
 334
 335 /*
 336  * read_chunk: reads the chunk_buf from the device. The number of read
 337  * operations are based on the parameters read_end, offset, and chunksize.
 338  */
 339 int read_chunks(int fd, unsigned long long offset, unsigned long long read_end,
 340                 char *chunk_buf, size_t chunksize, const time_t time_st,
 341                 const ino_t inode_st, const char *file)
 342 {
 343         unsigned long long stride;
 344
 345         stride = full ? chunksize : (ONE_GB - chunksize);
 346         for (offset = offset & ~(chunksize - 1); offset < read_end;
 347              offset += stride) {
 348                 ssize_t nread;
 349
 350                 if (stride != chunksize && lseek64(fd, offset, SEEK_SET) < 0) {
 351                         fprintf(stderr, "\n%s: lseek64(%s+%llu) failed: %s\n",
 352                                 progname, file, offset, strerror(errno));
 353                         return 1;
 354                 }
 355                 if (offset + chunksize > read_end)
 356                         chunksize = read_end - offset;
 357
 358                 if (!full && offset > chunksize) {
 359                         nread = read(fd, chunk_buf, chunksize);
 360                         if (nread < 0) {
 361                                 fprintf(stderr,"\n%s: read %s@%llu+%zi failed: "
 362                                         "%s\n", progname, file, offset,
 363                                         chunksize, strerror(errno));
 364                                 error_count++;
 365                                 return 1;
 366                         }
 367                         if (nread < chunksize) {
 368                                 fprintf(stderr, "\n%s: read %s@%llu+%zi short: "
 369                                         "%zi read\n", progname, file, offset,
 370                                         chunksize, nread);
 371                                 error_count++;
 372                         }
 373                         if (verify_chunk(chunk_buf, nread, offset, time_st,
 374                                          inode_st, file) != 0) {
 375                                 return 1;
 376                         }
 377                         offset += chunksize;
 378
 379                         /* Need to reset position after read error */
 380                         if (nread < chunksize &&
 381                             lseek64(fd, offset, SEEK_SET) == -1) {
 382                                 fprintf(stderr,
 383                                         "\n%s: lseek64(%s@%llu) failed: %s\n",
 384                                         progname, file, offset,strerror(errno));
 385                                 return 1;
 386                         }
 387                         if (offset + chunksize >= read_end)
 388                                 chunksize = read_end - offset;
 389                 }
 390                 nread = read(fd, chunk_buf, chunksize);
 391                 if (nread < 0) {
 392                         fprintf(stderr, "\n%s: read %s@%llu+%zi failed: %s\n",
 393                                 progname, file, offset, chunksize,
 394                                 strerror(errno));
 395                         error_count++;
 396                         return 1;
 397                 }
 398                 if (nread < chunksize) {
 399                         fprintf(stderr, "\n%s: read %s@%llu+%zi short: "
 400                                 "%zi read\n", progname, file, offset,
 401                                 chunksize, nread);
 402                         error_count++;
 403                 }
 404
 405                 if (verify_chunk(chunk_buf, nread, offset, time_st,
 406                                  inode_st, file) != 0) {
 407                         return 1;
 408                 }
 409         }
 410         return 0;
 411 }
 412
 413 /*
 414  * new_file: prepares new filename using file counter and current dir.
 415  */
 416 char *new_file(char *tempfile, char *cur_dir, int file_num)
 417 {
 418         snprintf(tempfile, PATH_MAX, "%s/file%03d", cur_dir, file_num);
 419         return tempfile;
 420 }
 421
 422 /*
 423  * new_dir: prepares new dir name using dir counters.
 424  */
 425 char *new_dir(char *tempdir, int dir_num)
 426 {
 427         snprintf(tempdir, PATH_MAX, "%s/llverfs_dir%05d", testdir, dir_num);
 428         return tempdir;
 429 }
 430
 431 /*
 432  * calc_total_bytes: calculates total bytes that need to be
 433  * written into or read from the filesystem.
 434  */
 435 static unsigned long long calc_total_bytes(const char *op)
 436 {
 437         unsigned long long total_bytes = 0;
 438         struct statfs64 statbuf;
 439
 440         if (full) {
 441                 if (statfs64(testdir, &statbuf) == 0) {
 442                         if (strcmp(op, "write") == 0)
 443                                 total_bytes = (unsigned long long)
 444                                         (statbuf.f_bavail * statbuf.f_bsize);
 445                         else if (strcmp(op, "read") == 0)
 446                                 total_bytes = (unsigned long long)
 447                                         (statbuf.f_blocks * statbuf.f_bsize);
 448                         else {
 449                                 fprintf(stderr, "\n%s: invalid operation: %s\n",
 450                                         progname, op);
 451                                 return -1;
 452                         }
 453                 } else {
 454                         fprintf(stderr, "\n%s: unable to stat %s: %s\n",
 455                                 progname, testdir, strerror(errno));
 456                         return -errno;
 457                 }
 458         } else {
 459                 total_bytes = num_dirs * files_in_dir * file_size;
 460         }
 461
 462         return total_bytes;
 463 }
 464
 465 /*
 466  * show_rate: displays the current read/write file name and performance,
 467  * along with an estimate of how long the whole read/write operation
 468  * will continue.
 469  */
 470 void show_rate(char *op, char *filename, const struct timeval *start_time,
 471                const unsigned long long total_bytes,
 472                const unsigned long long curr_bytes)
 473 {
 474         static struct timeval last_time;
 475         static unsigned long long last_bytes;
 476         static char last_op;
 477         struct timeval curr_time;
 478         double curr_delta, overall_delta, curr_rate, overall_rate;
 479         double remain_time;
 480         int remain_hours, remain_minutes, remain_seconds;
 481
 482         if (last_op != op[0]) {
 483                 last_bytes = 0;
 484                 last_time = *start_time;
 485                 last_op = op[0];
 486         }
 487
 488         gettimeofday(&curr_time, NULL);
 489
 490         curr_delta = (curr_time.tv_sec - last_time.tv_sec) +
 491                 (double)(curr_time.tv_usec - last_time.tv_usec) / 1000000;
 492
 493         overall_delta = (curr_time.tv_sec - start_time->tv_sec) +
 494                 (double)(curr_time.tv_usec - start_time->tv_usec) / 1000000;
 495
 496         curr_rate = (curr_bytes - last_bytes) / curr_delta;
 497         overall_rate = curr_bytes / overall_delta;
 498
 499         if (curr_rate == 0) {
 500                 last_time = curr_time;
 501                 return;
 502         }
 503         remain_time = (total_bytes - curr_bytes) / curr_rate;
 504
 505         remain_hours = remain_time / 3600;
 506         remain_minutes = (remain_time - remain_hours * 3600) / 60;
 507         remain_seconds = (remain_time - remain_hours * 3600 -
 508                 remain_minutes * 60);
 509
 510         if (curr_delta > 4 || verbose > 2) {
 511                 if (isatty_flag)
 512                         printf("\r");
 513
 514                 printf("%s: %s, current: %5g MB/s, overall: %5g MB/s, "
 515                        "ETA: %u:%02u:%02u", op, filename,
 516                        curr_rate / ONE_MB, overall_rate / ONE_MB,
 517                        remain_hours, remain_minutes, remain_seconds);
 518
 519                 if (isatty_flag)
 520                         fflush(stdout);
 521                 else
 522                         printf("\n");
 523
 524                 last_time = curr_time;
 525                 last_bytes = curr_bytes;
 526         }
 527 }
 528
 529 /*
 530  * dir_write: This function writes directories and files on device.
 531  * it works for both full and partial modes.
 532  */
 533 static int dir_write(char *chunk_buf, size_t chunksize,
 534                      time_t time_st, unsigned long dir_num)
 535 {
 536         char tempfile[PATH_MAX];
 537         char tempdir[PATH_MAX];
 538         FILE *countfile;
 539         struct stat64 file;
 540         int file_num = 999999999;
 541         ino_t inode_st = 0;
 542         struct timeval start_time;
 543         unsigned long long total_bytes;
 544         unsigned long long curr_bytes = 0;
 545         int rc = 0;
 546
 547         if (!full && fsetflags(testdir, EXT2_TOPDIR_FL))
 548                 fprintf(stderr,
 549                         "\n%s: can't set TOPDIR_FL on %s: %s (ignoring)",
 550                         progname, testdir, strerror(errno));
 551
 552         countfile = fopen(filecount, "w");
 553         if (countfile == NULL) {
 554                 fprintf(stderr, "\n%s: creating %s failed :%s\n",
 555                         progname, filecount, strerror(errno));
 556                 return 5;
 557         }
 558         /* reserve space for the countfile */
 559         if (fprintf(countfile, "%lu", num_files) < 1 ||
 560             fflush(countfile) != 0) {
 561                 fprintf(stderr, "\n%s: writing %s failed :%s\n",
 562                         progname, filecount, strerror(errno));
 563                 rc = 6;
 564                 goto out;
 565         }
 566
 567         /* calculate total bytes that need to be written */
 568         total_bytes = calc_total_bytes("write");
 569         if (total_bytes <= 0) {
 570                 fprintf(stderr, "\n%s: unable to calculate total bytes\n",
 571                         progname);
 572                 rc = 7;
 573                 goto out;
 574         }
 575
 576         if (!full && (dir_num != 0))
 577                 total_bytes -= dir_num * files_in_dir * file_size;
 578
 579         gettimeofday(&start_time, NULL);
 580         for (; dir_num < num_dirs; num_files++, file_num++) {
 581                 int fd, ret;
 582
 583                 if (file_num >= files_in_dir) {
 584                         file_num = 0;
 585                         if (mkdir(new_dir(tempdir, dir_num), dirmode) < 0) {
 586                                 if (errno == ENOSPC)
 587                                         break;
 588                                 if (errno != EEXIST) {
 589                                         fprintf(stderr, "\n%s: mkdir %s : %s\n",
 590                                                 progname, tempdir,
 591                                                 strerror(errno));
 592                                         rc = 1;
 593                                         goto out;
 594                                 }
 595                         }
 596                         dir_num++;
 597                 }
 598
 599                 fd = open_file(new_file(tempfile, tempdir, file_num),
 600                                O_WRONLY | O_CREAT | O_TRUNC | O_LARGEFILE);
 601                 if (fd >= 0) {
 602                         if (fstat64(fd, &file) == 0) {
 603                                 inode_st = file.st_ino;
 604                         } else {
 605                                 fprintf(stderr, "\n%s: write stat '%s': %s",
 606                                         progname, tempfile, strerror(errno));
 607                                 close(fd);
 608                                 break;
 609                         }
 610                 } else {
 611                         break;
 612                 }
 613
 614                 ret = write_chunks(fd, 0, file_size, chunk_buf, chunksize,
 615                                    time_st, inode_st, tempfile);
 616                 close(fd);
 617                 if (ret < 0) {
 618                         if (ret != -ENOSPC) {
 619                                 rc = 1;
 620                                 goto out;
 621                         }
 622                         curr_bytes = total_bytes;
 623                         break;
 624                 }
 625
 626                 curr_bytes += file_size;
 627                 if (verbose > 1)
 628                         show_rate("write", tempfile, &start_time,
 629                                   total_bytes, curr_bytes);
 630
 631                 fseek(countfile, 0, SEEK_SET);
 632                 if (fprintf(countfile, "%lu", num_files) < 1 ||
 633                     fflush(countfile) != 0) {
 634                         fprintf(stderr, "\n%s: writing %s failed :%s\n",
 635                                 progname, filecount, strerror(errno));
 636                 }
 637         }
 638
 639         verbose += 2;
 640         show_rate("write_done", tempfile, &start_time, total_bytes, curr_bytes);
 641         printf("\n");
 642         verbose -= 2;
 643
 644 out:
 645         fclose(countfile);
 646
 647         return rc;
 648 }
 649
 650 /*
 651  * dir_read: This function reads directories and files on device.
 652  * it works for both full and partial modes.
 653  */
 654 static int dir_read(char *chunk_buf, size_t chunksize,
 655                     time_t time_st, unsigned long dir_num)
 656 {
 657         char tempfile[PATH_MAX];
 658         char tempdir[PATH_MAX];
 659         unsigned long count = 0;
 660         struct stat64 file;
 661         int file_num = 0;
 662         ino_t inode_st = 0;
 663         struct timeval start_time;
 664         unsigned long long total_bytes;
 665         unsigned long long curr_bytes = 0;
 666
 667         /* calculate total bytes that need to be read */
 668         total_bytes = calc_total_bytes("read");
 669         if (total_bytes <= 0) {
 670                 fprintf(stderr, "\n%s: unable to calculate total bytes\n",
 671                         progname);
 672                 return 1;
 673         }
 674
 675         if (dir_num != 0)
 676                 total_bytes -= dir_num * files_in_dir * file_size;
 677
 678         gettimeofday(&start_time, NULL);
 679         for (count = 0; count < num_files && dir_num < num_dirs; count++) {
 680                 int fd, ret;
 681
 682                 if (file_num == 0) {
 683                         new_dir(tempdir, dir_num);
 684                         dir_num++;
 685                 }
 686
 687                 fd = open_file(new_file(tempfile, tempdir, file_num),
 688                                O_RDONLY | O_LARGEFILE);
 689                 if (fd >= 0) {
 690                         if (fstat64(fd, &file) == 0) {
 691                                 inode_st = file.st_ino;
 692                         } else {
 693                                 fprintf(stderr, "\n%s: read stat '%s': %s\n",
 694                                         progname, tempfile, strerror(errno));
 695                                 close(fd);
 696                                 return 1;
 697                         }
 698                 } else {
 699                         break;
 700                 }
 701
 702                 if (count == num_files)
 703                         file_size = file.st_size;
 704                 ret = read_chunks(fd, 0, file_size, chunk_buf, chunksize,
 705                                   time_st, inode_st, tempfile);
 706                 close(fd);
 707                 if (ret)
 708                         return 1;
 709
 710                 curr_bytes += file_size;
 711                 if (verbose > 1)
 712                         show_rate("read", tempfile, &start_time,
 713                                   total_bytes, curr_bytes);
 714
 715                 if (++file_num >= files_in_dir)
 716                         file_num = 0;
 717         }
 718         verbose += 2;
 719         show_rate("read_done", tempfile, &start_time, total_bytes, curr_bytes);
 720         printf("\n");
 721         verbose -= 2;
 722
 723         return 0;
 724 }
 725
 726 int main(int argc, char **argv)
 727 {
 728         time_t time_st = 0;             /* Default timestamp */
 729         size_t chunksize = ONE_MB;      /* IO chunk size(defailt=1MB) */
 730         char *chunk_buf;                /* chunk buffer */
 731         int error = 0;
 732         FILE *countfile = NULL;
 733         unsigned long dir_num = 0, dir_num_orig = 0;/* starting directory */
 734         int c;
 735
 736         progname = strrchr(argv[0], '/') ? strrchr(argv[0], '/') + 1 : argv[0];
 737         while ((c = getopt_long(argc, argv, "c:hlo:pqrs:t:vw",
 738                                       long_opts, NULL)) != -1) {
 739                 switch (c) {
 740                 case 'c':
 741                         chunksize = strtoul(optarg, NULL, 0) * ONE_MB;
 742                         if (chunksize == 0) {
 743                                 fprintf(stderr, "%s: bad chunk size '%s'\n",
 744                                         optarg, progname);
 745                                 return -1;
 746                         }
 747                         break;
 748                 case 'l':
 749                         full = 1;
 750                         break;
 751                 case 'o': /* offset */
 752                         dir_num = strtoul(optarg, NULL, 0);
 753                         break;
 754                 case 'p':
 755                         file_size = ONE_MB;
 756                         chunksize = ONE_MB;
 757                         files_in_dir = 1;
 758                         full = 0;
 759                         break;
 760                 case 'q':
 761                         verbose = 0;
 762                         break;
 763                 case 'r':
 764                         readoption = 1;
 765                         break;
 766                 case 's':
 767                         file_size = strtoul(optarg, NULL, 0) * ONE_MB;
 768                         if (file_size == 0) {
 769                                 fprintf(stderr, "%s: bad file size '%s'\n",
 770                                         optarg, progname);
 771                                 return -1;
 772                         }
 773                         break;
 774                 case 't':
 775                         time_st = (time_t)strtoul(optarg, NULL, 0);
 776                         break;
 777                 case 'v':
 778                         verbose++;
 779                         break;
 780                 case 'w':
 781                         writeoption = 1;
 782                         break;
 783
 784                 case 'h':
 785                 default:
 786                         usage(1);
 787                         return 0;
 788                 }
 789         }
 790         testdir = argv[optind];
 791
 792         if (!testdir) {
 793                 fprintf(stderr, "%s: pathname not given\n", progname);
 794                 usage(1);
 795                 return -1;
 796         }
 797         if (!readoption && !writeoption) {
 798                 readoption = 1;
 799                 writeoption = 1;
 800         }
 801         if (!time_st)
 802                 (void) time(&time_st);
 803         printf("Timestamp: %lu\n", (unsigned long )time_st);
 804         isatty_flag = isatty(STDOUT_FILENO);
 805
 806         if (!full) {
 807 #ifdef HAVE_EXT2FS_EXT2FS_H
 808                 struct mntent *tempmnt;
 809                 FILE *fp = NULL;
 810                 ext2_filsys fs;
 811
 812                 if ((fp = setmntent("/etc/mtab", "r")) == NULL) {
 813                         fprintf(stderr, "%s: fail to open /etc/mtab in read mode :%s\n",
 814                                 progname, strerror(errno));
 815                         goto guess;
 816                 }
 817
 818                 /* find device name using filesystem */
 819                 while ((tempmnt = getmntent(fp)) != NULL) {
 820                         if (strcmp(tempmnt->mnt_dir, testdir) == 0)
 821                                 break;
 822                 }
 823
 824                 if (tempmnt == NULL) {
 825                         fprintf(stderr, "%s: no device found for '%s'\n",
 826                                 progname, testdir);
 827                         endmntent(fp);
 828                         goto guess;
 829                 }
 830
 831                 if (ext2fs_open(tempmnt->mnt_fsname, 0, 0, 0,
 832                                 unix_io_manager, &fs)) {
 833                         fprintf(stderr, "%s: unable to open ext3 fs on '%s'\n",
 834                                 progname, testdir);
 835                         endmntent(fp);
 836                         goto guess;
 837                 }
 838                 endmntent(fp);
 839
 840                 num_dirs = (fs->super->s_blocks_count +
 841                             fs->super->s_blocks_per_group - 1) /
 842                         fs->super->s_blocks_per_group;
 843                 if (verbose)
 844                         printf("ext3 block groups: %u, fs blocks: %u "
 845                                "blocks per group: %u\n",
 846                                num_dirs, fs->super->s_blocks_count,
 847                                fs->super->s_blocks_per_group);
 848                 ext2fs_close(fs);
 849 #else
 850                 goto guess;
 851 #endif
 852                 if (0) { /* ugh */
 853                         struct statfs64 statbuf;
 854 guess:
 855                         /*
 856                          * Most extN filesystems are formatted with 128MB/group
 857                          * (32k bitmap = 4KB blocksize * 8 bits/block) * 4KB,
 858                          * so this is a relatively safe default (somewhat more
 859                          * or less doesn't make a huge difference for testing).
 860                          *
 861                          * We want to create one directory per group, together
 862                          * with the "TOPDIR" feature, so that the directories
 863                          * are spread across the whole block device.
 864                          */
 865                         if (statfs64(testdir, &statbuf) == 0) {
 866                                 num_dirs = 1 + (long long)statbuf.f_blocks *
 867                                         statbuf.f_bsize / (128ULL * ONE_MB);
 868                                 if (verbose)
 869                                         printf("dirs: %u, fs blocks: %llu\n",
 870                                                num_dirs,
 871                                                (long long)statbuf.f_blocks);
 872                         } else {
 873                                 fprintf(stderr, "%s: unable to stat '%s': %s\n",
 874                                         progname, testdir, strerror(errno));
 875                                 if (verbose)
 876                                         printf("dirs: %u\n", num_dirs);
 877                         }
 878                 }
 879         }
 880         chunk_buf = (char *)calloc(chunksize, 1);
 881         if (chunk_buf == NULL) {
 882                 fprintf(stderr, "Memory allocation failed for chunk_buf\n");
 883                 return 4;
 884         }
 885         snprintf(filecount, sizeof(filecount), "%s/%s.filecount",
 886                  testdir, progname);
 887         if (writeoption) {
 888                 (void)mkdir(testdir, dirmode);
 889
 890                 unlink(filecount);
 891                 if (dir_num != 0) {
 892                         num_files = dir_num * files_in_dir;
 893                         if (verbose)
 894                                 printf("\n%s: %lu files already written\n",
 895                                        progname, num_files);
 896                 }
 897                 if (dir_write(chunk_buf, chunksize, time_st, dir_num)) {
 898                         error = 3;
 899                         goto out;
 900                 }
 901                 dir_num = dir_num_orig;
 902         }
 903         if (readoption) {
 904                 if (!writeoption) {
 905                         countfile = fopen(filecount, "r");
 906                         if (countfile == NULL ||
 907                             fscanf(countfile, "%lu", &num_files) != 1 ||
 908                             num_files == 0) {
 909                                 fprintf(stderr, "\n%s: reading %s failed :%s\n",
 910                                         progname, filecount, strerror(errno));
 911                                 num_files = num_dirs * files_in_dir;
 912                         } else {
 913                                 num_files -= (dir_num * files_in_dir);
 914                         }
 915                         if (countfile)
 916                                 fclose(countfile);
 917                 }
 918                 if (dir_read(chunk_buf, chunksize, time_st, dir_num)) {
 919                         fprintf(stderr, "\n%s: Data verification failed\n",
 920                                 progname) ;
 921                         error = 2;
 922                         goto out;
 923                 }
 924         }
 925         error = error_count;
 926 out:
 927         free(chunk_buf);
 928         return error;
 929 }