ltp/fsx.c

   1 /*
   2  *      Copyright (C) 1991, NeXT Computer, Inc.  All Rights Reserverd.
   3  *
   4  *      File:   fsx.c
   5  *      Author: Avadis Tevanian, Jr.
   6  *
   7  *      File system exerciser.
   8  *
   9  *      Rewritten 8/98 by Conrad Minshall.
  10  *
  11  *      Small changes to work under Linux -- davej.
  12  *
  13  *      Checks for mmap last-page zero fill.
  14  */
  15
  16 #include "global.h"
  17
  18 #include <limits.h>
  19 #include <time.h>
  20 #include <strings.h>
  21 #include <sys/file.h>
  22 #include <sys/mman.h>
  23 #include <stdbool.h>
  24 #ifdef HAVE_ERR_H
  25 #include <err.h>
  26 #endif
  27 #include <signal.h>
  28 #include <stdio.h>
  29 #include <stddef.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32 #include <stdarg.h>
  33 #include <errno.h>
  34 #ifdef AIO
  35 #include <libaio.h>
  36 #endif
  37
  38 #ifndef MAP_FILE
  39 # define MAP_FILE 0
  40 #endif
  41
  42 #define NUMPRINTCOLUMNS 32      /* # columns of data to print on each line */
  43
  44 /* Operation flags */
  45
  46 enum opflags { FL_NONE = 0, FL_SKIPPED = 1, FL_CLOSE_OPEN = 2, FL_KEEP_SIZE = 4 };
  47
  48 /*
  49  *      A log entry is an operation and a bunch of arguments.
  50  */
  51
  52 struct log_entry {
  53         int     operation;
  54         int     nr_args;
  55         int     args[4];
  56         enum opflags flags;
  57 };
  58
  59 #define LOGSIZE 10000
  60
  61 struct log_entry        oplog[LOGSIZE]; /* the log */
  62 int                     logptr = 0;     /* current position in log */
  63 int                     logcount = 0;   /* total ops */
  64
  65 /*
  66  * The operation matrix is complex due to conditional execution of different
  67  * features. Hence when we come to deciding what operation to run, we need to
  68  * be careful in how we select the different operations. The active operations
  69  * are mapped to numbers as follows:
  70  *
  71  *                      lite    !lite   integrity
  72  * READ:                0       0       0
  73  * WRITE:               1       1       1
  74  * MAPREAD:             2       2       2
  75  * MAPWRITE:            3       3       3
  76  * TRUNCATE:            -       4       4
  77  * FALLOCATE:           -       5       5
  78  * PUNCH HOLE:          -       6       6
  79  * ZERO RANGE:          -       7       7
  80  * COLLAPSE RANGE:      -       8       8
  81  * FSYNC:               -       -       9
  82  *
  83  * When mapped read/writes are disabled, they are simply converted to normal
  84  * reads and writes. When fallocate/fpunch calls are disabled, they are
  85  * skipped.
  86  *
  87  * Because of the "lite" version, we also need to have different "maximum
  88  * operation" defines to allow the ops to be selected correctly based on the
  89  * mode being run.
  90  */
  91
  92 enum {
  93         /* common operations */
  94         OP_READ = 0,
  95         OP_WRITE,
  96         OP_MAPREAD,
  97         OP_MAPWRITE,
  98         OP_MAX_LITE,
  99
 100         /* !lite operations */
 101         OP_TRUNCATE = OP_MAX_LITE,
 102         OP_FALLOCATE,
 103         OP_PUNCH_HOLE,
 104         OP_ZERO_RANGE,
 105         OP_COLLAPSE_RANGE,
 106         OP_INSERT_RANGE,
 107         OP_CLONE_RANGE,
 108         OP_DEDUPE_RANGE,
 109         OP_MAX_FULL,
 110
 111         /* integrity operations */
 112         OP_FSYNC = OP_MAX_FULL,
 113         OP_MAX_INTEGRITY,
 114 };
 115
 116 #undef PAGE_SIZE
 117 #define PAGE_SIZE       getpagesize()
 118 #undef PAGE_MASK
 119 #define PAGE_MASK       (PAGE_SIZE - 1)
 120
 121 char    *original_buf;                  /* a pointer to the original data */
 122 char    *good_buf;                      /* a pointer to the correct data */
 123 char    *temp_buf;                      /* a pointer to the current data */
 124 char    *fname;                         /* name of our test file */
 125 char    *bname;                         /* basename of our test file */
 126 char    *logdev;                        /* -i flag */
 127 char    *logid;                         /* -j flag */
 128 char    dname[1024];                    /* -P flag */
 129 char    goodfile[PATH_MAX];
 130 int     dirpath = 0;                    /* -P flag */
 131 int     fd;                             /* fd for our test file */
 132
 133 blksize_t       block_size = 0;
 134 off_t           file_size = 0;
 135 off_t           biggest = 0;
 136 unsigned long   testcalls = 0;          /* calls to function "test" */
 137
 138 unsigned long   simulatedopcount = 0;   /* -b flag */
 139 int     closeprob = 0;                  /* -c flag */
 140 int     debug = 0;                      /* -d flag */
 141 unsigned long   debugstart = 0;         /* -D flag */
 142 char    filldata = 0;                   /* -g flag */
 143 int     flush = 0;                      /* -f flag */
 144 int     do_fsync = 0;                   /* -y flag */
 145 unsigned long   maxfilelen = 256 * 1024;        /* -l flag */
 146 int     sizechecks = 1;                 /* -n flag disables them */
 147 int     maxoplen = 64 * 1024;           /* -o flag */
 148 int     quiet = 0;                      /* -q flag */
 149 unsigned long progressinterval = 0;     /* -p flag */
 150 int     readbdy = 1;                    /* -r flag */
 151 int     style = 0;                      /* -s flag */
 152 int     prealloc = 0;                   /* -x flag */
 153 int     truncbdy = 1;                   /* -t flag */
 154 int     writebdy = 1;                   /* -w flag */
 155 long    monitorstart = -1;              /* -m flag */
 156 long    monitorend = -1;                /* -m flag */
 157 int     lite = 0;                       /* -L flag */
 158 long    numops = -1;                    /* -N flag */
 159 int     randomoplen = 1;                /* -O flag disables it */
 160 int     seed = 1;                       /* -S flag */
 161 int     mapped_writes = 1;              /* -W flag disables */
 162 int     fallocate_calls = 1;            /* -F flag disables */
 163 int     keep_size_calls = 1;            /* -K flag disables */
 164 int     punch_hole_calls = 1;           /* -H flag disables */
 165 int     zero_range_calls = 1;           /* -z flag disables */
 166 int     collapse_range_calls = 1;       /* -C flag disables */
 167 int     insert_range_calls = 1;         /* -I flag disables */
 168 int     mapped_reads = 1;               /* -R flag disables it */
 169 int     check_file = 0;                 /* -X flag enables */
 170 int     clone_range_calls = 1;          /* -J flag disables */
 171 int     dedupe_range_calls = 1;         /* -B flag disables */
 172 int     integrity = 0;                  /* -i flag */
 173 int     fsxgoodfd = 0;
 174 int     o_direct;                       /* -Z */
 175 int     aio = 0;
 176 int     mark_nr = 0;
 177
 178 int page_size;
 179 int page_mask;
 180 int mmap_mask;
 181 #ifdef AIO
 182 int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
 183 #define READ 0
 184 #define WRITE 1
 185 #define fsxread(a,b,c,d)        aio_rw(READ, a,b,c,d)
 186 #define fsxwrite(a,b,c,d)       aio_rw(WRITE, a,b,c,d)
 187 #else
 188 #define fsxread(a,b,c,d)        read(a,b,c)
 189 #define fsxwrite(a,b,c,d)       write(a,b,c)
 190 #endif
 191
 192 const char *replayops = NULL;
 193 const char *recordops = NULL;
 194 FILE *  fsxlogf = NULL;
 195 FILE *  replayopsf = NULL;
 196 char opsfile[PATH_MAX];
 197 int badoff = -1;
 198 int closeopen = 0;
 199
 200 static void *round_ptr_up(void *ptr, unsigned long align, unsigned long offset)
 201 {
 202         unsigned long ret = (unsigned long)ptr;
 203
 204         ret = ((ret + align - 1) & ~(align - 1));
 205         ret += offset;
 206         return (void *)ret;
 207 }
 208
 209 void
 210 vwarnc(int code, const char *fmt, va_list ap)
 211 {
 212         if (logid)
 213                 fprintf(stderr, "%s: ", logid);
 214         fprintf(stderr, "fsx: ");
 215         if (fmt != NULL) {
 216                 vfprintf(stderr, fmt, ap);
 217                 fprintf(stderr, ": ");
 218         }
 219         fprintf(stderr, "%s\n", strerror(code));
 220 }
 221
 222 void
 223 warn(const char * fmt, ...)  {
 224         va_list ap;
 225         va_start(ap, fmt);
 226         vwarnc(errno, fmt, ap);
 227         va_end(ap);
 228 }
 229
 230 void
 231 prt(const char *fmt, ...)
 232 {
 233         va_list args;
 234
 235         if (logid)
 236                 fprintf(stdout, "%s: ", logid);
 237         va_start(args, fmt);
 238         vfprintf(stdout, fmt, args);
 239         va_end(args);
 240         if (fsxlogf) {
 241                 va_start(args, fmt);
 242                 vfprintf(fsxlogf, fmt, args);
 243                 va_end(args);
 244         }
 245 }
 246
 247 void
 248 prterr(const char *prefix)
 249 {
 250         prt("%s%s%s\n", prefix, prefix ? ": " : "", strerror(errno));
 251 }
 252
 253
 254 static const char *op_names[] = {
 255         [OP_READ] = "read",
 256         [OP_WRITE] = "write",
 257         [OP_MAPREAD] = "mapread",
 258         [OP_MAPWRITE] = "mapwrite",
 259         [OP_TRUNCATE] = "truncate",
 260         [OP_FALLOCATE] = "fallocate",
 261         [OP_PUNCH_HOLE] = "punch_hole",
 262         [OP_ZERO_RANGE] = "zero_range",
 263         [OP_COLLAPSE_RANGE] = "collapse_range",
 264         [OP_INSERT_RANGE] = "insert_range",
 265         [OP_CLONE_RANGE] = "clone_range",
 266         [OP_DEDUPE_RANGE] = "dedupe_range",
 267         [OP_FSYNC] = "fsync",
 268 };
 269
 270 static const char *op_name(int operation)
 271 {
 272         if (operation >= 0 &&
 273             operation < sizeof(op_names) / sizeof(op_names[0]))
 274                 return op_names[operation];
 275         return NULL;
 276 }
 277
 278 static int op_code(const char *name)
 279 {
 280         int i;
 281
 282         for (i = 0; i < sizeof(op_names) / sizeof(op_names[0]); i++)
 283                 if (op_names[i] && strcmp(name, op_names[i]) == 0)
 284                         return i;
 285         return -1;
 286 }
 287
 288 void
 289 log5(int operation, int arg0, int arg1, int arg2, enum opflags flags)
 290 {
 291         struct log_entry *le;
 292
 293         le = &oplog[logptr];
 294         le->operation = operation;
 295         if (closeopen)
 296                 flags |= FL_CLOSE_OPEN;
 297         le->args[0] = arg0;
 298         le->args[1] = arg1;
 299         le->args[2] = arg2;
 300         le->args[3] = file_size;
 301         le->nr_args = 4;
 302         le->flags = flags;
 303         logptr++;
 304         logcount++;
 305         if (logptr >= LOGSIZE)
 306                 logptr = 0;
 307 }
 308
 309 void
 310 log4(int operation, int arg0, int arg1, enum opflags flags)
 311 {
 312         struct log_entry *le;
 313
 314         le = &oplog[logptr];
 315         le->operation = operation;
 316         if (closeopen)
 317                 flags |= FL_CLOSE_OPEN;
 318         le->args[0] = arg0;
 319         le->args[1] = arg1;
 320         le->args[2] = file_size;
 321         le->nr_args = 3;
 322         le->flags = flags;
 323         logptr++;
 324         logcount++;
 325         if (logptr >= LOGSIZE)
 326                 logptr = 0;
 327 }
 328
 329 void
 330 logdump(void)
 331 {
 332         FILE    *logopsf;
 333         int     i, count, down;
 334         struct log_entry        *lp;
 335
 336         prt("LOG DUMP (%d total operations):\n", logcount);
 337
 338         logopsf = fopen(opsfile, "w");
 339         if (!logopsf)
 340                 prterr(opsfile);
 341
 342         if (logcount < LOGSIZE) {
 343                 i = 0;
 344                 count = logcount;
 345         } else {
 346                 i = logptr;
 347                 count = LOGSIZE;
 348         }
 349         for ( ; count > 0; count--) {
 350                 bool overlap, overlap2;
 351                 int opnum;
 352
 353                 opnum = i+1 + (logcount/LOGSIZE)*LOGSIZE;
 354                 prt("%d(%3d mod 256): ", opnum, opnum%256);
 355                 lp = &oplog[i];
 356
 357                 overlap = badoff >= lp->args[0] &&
 358                           badoff < lp->args[0] + lp->args[1];
 359
 360                 if (lp->flags & FL_SKIPPED) {
 361                         prt("SKIPPED (no operation)");
 362                         goto skipped;
 363                 }
 364
 365                 switch (lp->operation) {
 366                 case OP_MAPREAD:
 367                         prt("MAPREAD  0x%x thru 0x%x\t(0x%x bytes)",
 368                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 369                             lp->args[1]);
 370                         if (overlap)
 371                                 prt("\t***RRRR***");
 372                         break;
 373                 case OP_MAPWRITE:
 374                         prt("MAPWRITE 0x%x thru 0x%x\t(0x%x bytes)",
 375                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 376                             lp->args[1]);
 377                         if (overlap)
 378                                 prt("\t******WWWW");
 379                         break;
 380                 case OP_READ:
 381                         prt("READ     0x%x thru 0x%x\t(0x%x bytes)",
 382                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 383                             lp->args[1]);
 384                         if (overlap)
 385                                 prt("\t***RRRR***");
 386                         break;
 387                 case OP_WRITE:
 388                         prt("WRITE    0x%x thru 0x%x\t(0x%x bytes)",
 389                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 390                             lp->args[1]);
 391                         if (lp->args[0] > lp->args[2])
 392                                 prt(" HOLE");
 393                         else if (lp->args[0] + lp->args[1] > lp->args[2])
 394                                 prt(" EXTEND");
 395                         overlap = (badoff >= lp->args[0] ||
 396                                    badoff >=lp->args[2]) &&
 397                                   badoff < lp->args[0] + lp->args[1];
 398                         if (overlap)
 399                                 prt("\t***WWWW");
 400                         break;
 401                 case OP_TRUNCATE:
 402                         down = lp->args[1] < lp->args[2];
 403                         prt("TRUNCATE %s\tfrom 0x%x to 0x%x",
 404                             down ? "DOWN" : "UP", lp->args[2], lp->args[1]);
 405                         overlap = badoff >= lp->args[1 + !down] &&
 406                                   badoff < lp->args[1 + !!down];
 407                         if (overlap)
 408                                 prt("\t******WWWW");
 409                         break;
 410                 case OP_FALLOCATE:
 411                         /* 0: offset 1: length 2: where alloced */
 412                         prt("FALLOC   0x%x thru 0x%x\t(0x%x bytes) ",
 413                                 lp->args[0], lp->args[0] + lp->args[1],
 414                                 lp->args[1]);
 415                         if (lp->args[0] + lp->args[1] <= lp->args[2])
 416                                 prt("INTERIOR");
 417                         else if (lp->flags & FL_KEEP_SIZE)
 418                                 prt("PAST_EOF");
 419                         else
 420                                 prt("EXTENDING");
 421                         if (overlap)
 422                                 prt("\t******FFFF");
 423                         break;
 424                 case OP_PUNCH_HOLE:
 425                         prt("PUNCH    0x%x thru 0x%x\t(0x%x bytes)",
 426                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 427                             lp->args[1]);
 428                         if (overlap)
 429                                 prt("\t******PPPP");
 430                         break;
 431                 case OP_ZERO_RANGE:
 432                         prt("ZERO     0x%x thru 0x%x\t(0x%x bytes)",
 433                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 434                             lp->args[1]);
 435                         if (overlap)
 436                                 prt("\t******ZZZZ");
 437                         break;
 438                 case OP_COLLAPSE_RANGE:
 439                         prt("COLLAPSE 0x%x thru 0x%x\t(0x%x bytes)",
 440                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 441                             lp->args[1]);
 442                         if (overlap)
 443                                 prt("\t******CCCC");
 444                         break;
 445                 case OP_INSERT_RANGE:
 446                         prt("INSERT 0x%x thru 0x%x\t(0x%x bytes)",
 447                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 448                             lp->args[1]);
 449                         if (overlap)
 450                                 prt("\t******IIII");
 451                         break;
 452                 case OP_CLONE_RANGE:
 453                         prt("CLONE 0x%x thru 0x%x\t(0x%x bytes) to 0x%x thru 0x%x",
 454                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 455                             lp->args[1],
 456                             lp->args[2], lp->args[2] + lp->args[1] - 1);
 457                         overlap2 = badoff >= lp->args[2] &&
 458                                   badoff < lp->args[2] + lp->args[1];
 459                         if (overlap && overlap2)
 460                                 prt("\tJJJJ**JJJJ");
 461                         else if (overlap)
 462                                 prt("\tJJJJ******");
 463                         else if (overlap2)
 464                                 prt("\t******JJJJ");
 465                         break;
 466                 case OP_DEDUPE_RANGE:
 467                         prt("DEDUPE 0x%x thru 0x%x\t(0x%x bytes) to 0x%x thru 0x%x",
 468                             lp->args[0], lp->args[0] + lp->args[1] - 1,
 469                             lp->args[1],
 470                             lp->args[2], lp->args[2] + lp->args[1] - 1);
 471                         overlap2 = badoff >= lp->args[2] &&
 472                                   badoff < lp->args[2] + lp->args[1];
 473                         if (overlap && overlap2)
 474                                 prt("\tBBBB**BBBB");
 475                         else if (overlap)
 476                                 prt("\tBBBB******");
 477                         else if (overlap2)
 478                                 prt("\t******BBBB");
 479                         break;
 480                 case OP_FSYNC:
 481                         prt("FSYNC");
 482                         break;
 483                 default:
 484                         prt("BOGUS LOG ENTRY (operation code = %d)!",
 485                             lp->operation);
 486                         continue;
 487                 }
 488
 489             skipped:
 490                 if (lp->flags & FL_CLOSE_OPEN)
 491                         prt("\n\t\tCLOSE/OPEN");
 492                 prt("\n");
 493                 i++;
 494                 if (i == LOGSIZE)
 495                         i = 0;
 496
 497                 if (logopsf) {
 498                         int j;
 499
 500                         if (lp->flags & FL_SKIPPED)
 501                                 fprintf(logopsf, "skip ");
 502                         fprintf(logopsf, "%s", op_name(lp->operation));
 503                         for (j = 0; j < lp->nr_args; j++)
 504                                 fprintf(logopsf, " 0x%x", lp->args[j]);
 505                         if (lp->flags & FL_KEEP_SIZE)
 506                                 fprintf(logopsf, " keep_size");
 507                         if (lp->flags & FL_CLOSE_OPEN)
 508                                 fprintf(logopsf, " close_open");
 509                         if (overlap)
 510                                 fprintf(logopsf, " *");
 511                         fprintf(logopsf, "\n");
 512                 }
 513         }
 514
 515         if (logopsf) {
 516                 if (fclose(logopsf) != 0)
 517                         prterr(opsfile);
 518                 else
 519                         prt("Log of operations saved to \"%s\"; "
 520                             "replay with --replay-ops\n",
 521                             opsfile);
 522         }
 523 }
 524
 525
 526 void
 527 save_buffer(char *buffer, off_t bufferlength, int fd)
 528 {
 529         off_t ret;
 530         ssize_t byteswritten;
 531
 532         if (fd <= 0 || bufferlength == 0)
 533                 return;
 534
 535         if (bufferlength > SSIZE_MAX) {
 536                 prt("fsx flaw: overflow in save_buffer\n");
 537                 exit(67);
 538         }
 539         if (lite) {
 540                 off_t size_by_seek = lseek(fd, (off_t)0, SEEK_END);
 541                 if (size_by_seek == (off_t)-1)
 542                         prterr("save_buffer: lseek eof");
 543                 else if (bufferlength > size_by_seek) {
 544                         warn("save_buffer: .fsxgood file too short... will save 0x%llx bytes instead of 0x%llx\n", (unsigned long long)size_by_seek,
 545                              (unsigned long long)bufferlength);
 546                         bufferlength = size_by_seek;
 547                 }
 548         }
 549
 550         ret = lseek(fd, (off_t)0, SEEK_SET);
 551         if (ret == (off_t)-1)
 552                 prterr("save_buffer: lseek 0");
 553
 554         byteswritten = write(fd, buffer, (size_t)bufferlength);
 555         if (byteswritten != bufferlength) {
 556                 if (byteswritten == -1)
 557                         prterr("save_buffer write");
 558                 else
 559                         warn("save_buffer: short write, 0x%x bytes instead of 0x%llx\n",
 560                              (unsigned)byteswritten,
 561                              (unsigned long long)bufferlength);
 562         }
 563 }
 564
 565
 566 void
 567 report_failure(int status)
 568 {
 569         logdump();
 570
 571         if (fsxgoodfd) {
 572                 if (good_buf) {
 573                         save_buffer(good_buf, file_size, fsxgoodfd);
 574                         prt("Correct content saved for comparison\n");
 575                         prt("(maybe hexdump \"%s\" vs \"%s\")\n",
 576                             fname, goodfile);
 577                 }
 578                 close(fsxgoodfd);
 579         }
 580         exit(status);
 581 }
 582
 583
 584 #define short_at(cp) ((unsigned short)((*((unsigned char *)(cp)) << 8) | \
 585                                         *(((unsigned char *)(cp)) + 1)))
 586
 587 void
 588 mark_log(void)
 589 {
 590         char command[256];
 591         int ret;
 592
 593         snprintf(command, 256, "dmsetup message %s 0 mark %s.mark%d", logdev,
 594                  bname, mark_nr);
 595         ret = system(command);
 596         if (ret) {
 597                 prterr("dmsetup mark failed");
 598                 exit(211);
 599         }
 600 }
 601
 602 void
 603 dump_fsync_buffer(void)
 604 {
 605         char fname_buffer[PATH_MAX];
 606         int good_fd;
 607
 608         if (!good_buf)
 609                 return;
 610
 611         snprintf(fname_buffer, sizeof(fname_buffer), "%s%s.mark%d", dname,
 612                  bname, mark_nr);
 613         good_fd = open(fname_buffer, O_WRONLY|O_CREAT|O_TRUNC, 0666);
 614         if (good_fd < 0) {
 615                 prterr(fname_buffer);
 616                 exit(212);
 617         }
 618
 619         save_buffer(good_buf, file_size, good_fd);
 620         close(good_fd);
 621         prt("Dumped fsync buffer to %s\n", fname_buffer + dirpath);
 622 }
 623
 624 void
 625 check_buffers(char *buf, unsigned offset, unsigned size)
 626 {
 627         unsigned char c, t;
 628         unsigned i = 0;
 629         unsigned n = 0;
 630         unsigned op = 0;
 631         unsigned bad = 0;
 632
 633         if (memcmp(good_buf + offset, buf, size) != 0) {
 634                 prt("READ BAD DATA: offset = 0x%x, size = 0x%x, fname = %s\n",
 635                     offset, size, fname);
 636                 prt("OFFSET\tGOOD\tBAD\tRANGE\n");
 637                 while (size > 0) {
 638                         c = good_buf[offset];
 639                         t = buf[i];
 640                         if (c != t) {
 641                                 if (n < 16) {
 642                                         bad = short_at(&buf[i]);
 643                                         prt("0x%05x\t0x%04x\t0x%04x", offset,
 644                                             short_at(&good_buf[offset]), bad);
 645                                         op = buf[offset & 1 ? i+1 : i];
 646                                         prt("\t0x%05x\n", n);
 647                                         if (op)
 648                                                 prt("operation# (mod 256) for "
 649                                                   "the bad data may be %u\n",
 650                                                 ((unsigned)op & 0xff));
 651                                         else
 652                                                 prt("operation# (mod 256) for "
 653                                                   "the bad data unknown, check"
 654                                                   " HOLE and EXTEND ops\n");
 655                                 }
 656                                 n++;
 657                                 badoff = offset;
 658                         }
 659                         offset++;
 660                         i++;
 661                         size--;
 662                 }
 663                 report_failure(110);
 664         }
 665 }
 666
 667
 668 void
 669 check_size(void)
 670 {
 671         struct stat     statbuf;
 672         off_t   size_by_seek;
 673
 674         if (fstat(fd, &statbuf)) {
 675                 prterr("check_size: fstat");
 676                 statbuf.st_size = -1;
 677         }
 678         size_by_seek = lseek(fd, (off_t)0, SEEK_END);
 679         if (file_size != statbuf.st_size || file_size != size_by_seek) {
 680                 prt("Size error: expected 0x%llx stat 0x%llx seek 0x%llx\n",
 681                     (unsigned long long)file_size,
 682                     (unsigned long long)statbuf.st_size,
 683                     (unsigned long long)size_by_seek);
 684                 report_failure(120);
 685         }
 686 }
 687
 688
 689 void
 690 check_trunc_hack(void)
 691 {
 692         struct stat statbuf;
 693         off_t offset = file_size + (off_t)100000;
 694
 695         if (ftruncate(fd, file_size))
 696                 goto ftruncate_err;
 697         if (ftruncate(fd, offset))
 698                 goto ftruncate_err;
 699         fstat(fd, &statbuf);
 700         if (statbuf.st_size != offset) {
 701                 prt("no extend on truncate! not posix!\n");
 702                 exit(130);
 703         }
 704         if (ftruncate(fd, file_size)) {
 705 ftruncate_err:
 706                 prterr("check_trunc_hack: ftruncate");
 707                 exit(131);
 708         }
 709 }
 710
 711 void
 712 doflush(unsigned offset, unsigned size)
 713 {
 714         unsigned pg_offset;
 715         unsigned map_size;
 716         char    *p;
 717
 718         if (o_direct == O_DIRECT)
 719                 return;
 720
 721         pg_offset = offset & mmap_mask;
 722         map_size  = pg_offset + size;
 723
 724         if ((p = (char *)mmap(0, map_size, PROT_READ | PROT_WRITE,
 725                               MAP_FILE | MAP_SHARED, fd,
 726                               (off_t)(offset - pg_offset))) == (char *)-1) {
 727                 prterr("doflush: mmap");
 728                 report_failure(202);
 729         }
 730         if (msync(p, map_size, MS_INVALIDATE) != 0) {
 731                 prterr("doflush: msync");
 732                 report_failure(203);
 733         }
 734         if (munmap(p, map_size) != 0) {
 735                 prterr("doflush: munmap");
 736                 report_failure(204);
 737         }
 738 }
 739
 740 void
 741 doread(unsigned offset, unsigned size)
 742 {
 743         off_t ret;
 744         unsigned iret;
 745
 746         offset -= offset % readbdy;
 747         if (o_direct)
 748                 size -= size % readbdy;
 749         if (size == 0) {
 750                 if (!quiet && testcalls > simulatedopcount && !o_direct)
 751                         prt("skipping zero size read\n");
 752                 log4(OP_READ, offset, size, FL_SKIPPED);
 753                 return;
 754         }
 755         if (size + offset > file_size) {
 756                 if (!quiet && testcalls > simulatedopcount)
 757                         prt("skipping seek/read past end of file\n");
 758                 log4(OP_READ, offset, size, FL_SKIPPED);
 759                 return;
 760         }
 761
 762         log4(OP_READ, offset, size, FL_NONE);
 763
 764         if (testcalls <= simulatedopcount)
 765                 return;
 766
 767         if (!quiet &&
 768                 ((progressinterval && testcalls % progressinterval == 0)  ||
 769                 (debug &&
 770                        (monitorstart == -1 ||
 771                         (offset + size > monitorstart &&
 772                         (monitorend == -1 || offset <= monitorend))))))
 773                 prt("%lu read\t0x%x thru\t0x%x\t(0x%x bytes)\n", testcalls,
 774                     offset, offset + size - 1, size);
 775         ret = lseek(fd, (off_t)offset, SEEK_SET);
 776         if (ret == (off_t)-1) {
 777                 prterr("doread: lseek");
 778                 report_failure(140);
 779         }
 780         iret = fsxread(fd, temp_buf, size, offset);
 781         if (iret != size) {
 782                 if (iret == -1)
 783                         prterr("doread: read");
 784                 else
 785                         prt("short read: 0x%x bytes instead of 0x%x\n",
 786                             iret, size);
 787                 report_failure(141);
 788         }
 789         check_buffers(temp_buf, offset, size);
 790 }
 791
 792 void
 793 check_eofpage(char *s, unsigned offset, char *p, int size)
 794 {
 795         unsigned long last_page, should_be_zero;
 796
 797         if (offset + size <= (file_size & ~page_mask))
 798                 return;
 799         /*
 800          * we landed in the last page of the file
 801          * test to make sure the VM system provided 0's
 802          * beyond the true end of the file mapping
 803          * (as required by mmap def in 1996 posix 1003.1)
 804          */
 805         last_page = ((unsigned long)p + (offset & page_mask) + size) & ~page_mask;
 806
 807         for (should_be_zero = last_page + (file_size & page_mask);
 808              should_be_zero < last_page + page_size;
 809              should_be_zero++)
 810                 if (*(char *)should_be_zero) {
 811                         prt("Mapped %s: non-zero data past EOF (0x%llx) page offset 0x%x is 0x%04x\n",
 812                             s, file_size - 1, should_be_zero & page_mask,
 813                             short_at(should_be_zero));
 814                         report_failure(205);
 815                 }
 816 }
 817
 818 void
 819 check_contents(void)
 820 {
 821         static char *check_buf;
 822         unsigned offset = 0;
 823         unsigned size = file_size;
 824         unsigned map_offset;
 825         unsigned map_size;
 826         char *p;
 827         off_t ret;
 828         unsigned iret;
 829
 830         if (!check_buf) {
 831                 check_buf = (char *) malloc(maxfilelen + writebdy);
 832                 assert(check_buf != NULL);
 833                 check_buf = round_ptr_up(check_buf, writebdy, 0);
 834                 memset(check_buf, '\0', maxfilelen);
 835         }
 836
 837         if (o_direct)
 838                 size -= size % readbdy;
 839         if (size == 0)
 840                 return;
 841
 842         ret = lseek(fd, (off_t)offset, SEEK_SET);
 843         if (ret == (off_t)-1) {
 844                 prterr("doread: lseek");
 845                 report_failure(140);
 846         }
 847
 848         iret = fsxread(fd, check_buf, size, offset);
 849         if (iret != size) {
 850                 if (iret == -1)
 851                         prterr("check_contents: read");
 852                 else
 853                         prt("short check read: 0x%x bytes instead of 0x%x\n",
 854                             iret, size);
 855                 report_failure(141);
 856         }
 857         check_buffers(check_buf, offset, size);
 858
 859         /* Map eof page, check it */
 860         map_offset = size - (size & PAGE_MASK);
 861         if (map_offset == size)
 862                 map_offset -= PAGE_SIZE;
 863         map_size  = size - map_offset;
 864
 865         p = mmap(0, map_size, PROT_READ, MAP_SHARED, fd, map_offset);
 866         if (p == MAP_FAILED) {
 867                 prterr("check_contents: mmap");
 868                 report_failure(190);
 869         }
 870         check_eofpage("check_contents", map_offset, p, map_size);
 871
 872         if (munmap(p, map_size) != 0) {
 873                 prterr("check_contents: munmap");
 874                 report_failure(191);
 875         }
 876 }
 877
 878 void
 879 domapread(unsigned offset, unsigned size)
 880 {
 881         unsigned pg_offset;
 882         unsigned map_size;
 883         char    *p;
 884
 885         offset -= offset % readbdy;
 886         if (size == 0) {
 887                 if (!quiet && testcalls > simulatedopcount)
 888                         prt("skipping zero size read\n");
 889                 log4(OP_MAPREAD, offset, size, FL_SKIPPED);
 890                 return;
 891         }
 892         if (size + offset > file_size) {
 893                 if (!quiet && testcalls > simulatedopcount)
 894                         prt("skipping seek/read past end of file\n");
 895                 log4(OP_MAPREAD, offset, size, FL_SKIPPED);
 896                 return;
 897         }
 898
 899         log4(OP_MAPREAD, offset, size, FL_NONE);
 900
 901         if (testcalls <= simulatedopcount)
 902                 return;
 903
 904         if (!quiet &&
 905                 ((progressinterval && testcalls % progressinterval == 0) ||
 906                        (debug &&
 907                        (monitorstart == -1 ||
 908                         (offset + size > monitorstart &&
 909                         (monitorend == -1 || offset <= monitorend))))))
 910                 prt("%lu mapread\t0x%x thru\t0x%x\t(0x%x bytes)\n", testcalls,
 911                     offset, offset + size - 1, size);
 912
 913         pg_offset = offset & PAGE_MASK;
 914         map_size  = pg_offset + size;
 915
 916         if ((p = (char *)mmap(0, map_size, PROT_READ, MAP_SHARED, fd,
 917                               (off_t)(offset - pg_offset))) == (char *)-1) {
 918                 prterr("domapread: mmap");
 919                 report_failure(190);
 920         }
 921         memcpy(temp_buf, p + pg_offset, size);
 922
 923         check_eofpage("Read", offset, p, size);
 924
 925         if (munmap(p, map_size) != 0) {
 926                 prterr("domapread: munmap");
 927                 report_failure(191);
 928         }
 929
 930         check_buffers(temp_buf, offset, size);
 931 }
 932
 933
 934 void
 935 gendata(char *original_buf, char *good_buf, unsigned offset, unsigned size)
 936 {
 937         while (size--) {
 938                 if (filldata) {
 939                         good_buf[offset] = filldata;
 940                 } else {
 941                         good_buf[offset] = testcalls % 256;
 942                         if (offset % 2)
 943                                 good_buf[offset] += original_buf[offset];
 944                 }
 945                 offset++;
 946         }
 947 }
 948
 949
 950 void
 951 dowrite(unsigned offset, unsigned size)
 952 {
 953         off_t ret;
 954         unsigned iret;
 955
 956         offset -= offset % writebdy;
 957         if (o_direct)
 958                 size -= size % writebdy;
 959         if (size == 0) {
 960                 if (!quiet && testcalls > simulatedopcount && !o_direct)
 961                         prt("skipping zero size write\n");
 962                 log4(OP_WRITE, offset, size, FL_SKIPPED);
 963                 return;
 964         }
 965
 966         log4(OP_WRITE, offset, size, FL_NONE);
 967
 968         gendata(original_buf, good_buf, offset, size);
 969         if (file_size < offset + size) {
 970                 if (file_size < offset)
 971                         memset(good_buf + file_size, '\0', offset - file_size);
 972                 file_size = offset + size;
 973                 if (lite) {
 974                         warn("Lite file size bug in fsx!");
 975                         report_failure(149);
 976                 }
 977         }
 978
 979         if (testcalls <= simulatedopcount)
 980                 return;
 981
 982         if (!quiet &&
 983                 ((progressinterval && testcalls % progressinterval == 0) ||
 984                        (debug &&
 985                        (monitorstart == -1 ||
 986                         (offset + size > monitorstart &&
 987                         (monitorend == -1 || offset <= monitorend))))))
 988                 prt("%lu write\t0x%x thru\t0x%x\t(0x%x bytes)\n", testcalls,
 989                     offset, offset + size - 1, size);
 990         ret = lseek(fd, (off_t)offset, SEEK_SET);
 991         if (ret == (off_t)-1) {
 992                 prterr("dowrite: lseek");
 993                 report_failure(150);
 994         }
 995         iret = fsxwrite(fd, good_buf + offset, size, offset);
 996         if (iret != size) {
 997                 if (iret == -1)
 998                         prterr("dowrite: write");
 999                 else
1000                         prt("short write: 0x%x bytes instead of 0x%x\n",
1001                             iret, size);
1002                 report_failure(151);
1003         }
1004         if (do_fsync) {
1005                 if (fsync(fd)) {
1006                         prt("fsync() failed: %s\n", strerror(errno));
1007                         report_failure(152);
1008                 }
1009         }
1010         if (flush) {
1011                 doflush(offset, size);
1012         }
1013 }
1014
1015
1016 void
1017 domapwrite(unsigned offset, unsigned size)
1018 {
1019         unsigned pg_offset;
1020         unsigned map_size;
1021         off_t    cur_filesize;
1022         char    *p;
1023
1024         offset -= offset % writebdy;
1025         if (size == 0) {
1026                 if (!quiet && testcalls > simulatedopcount)
1027                         prt("skipping zero size write\n");
1028                 log4(OP_MAPWRITE, offset, size, FL_SKIPPED);
1029                 return;
1030         }
1031         cur_filesize = file_size;
1032
1033         log4(OP_MAPWRITE, offset, size, FL_NONE);
1034
1035         gendata(original_buf, good_buf, offset, size);
1036         if (file_size < offset + size) {
1037                 if (file_size < offset)
1038                         memset(good_buf + file_size, '\0', offset - file_size);
1039                 file_size = offset + size;
1040                 if (lite) {
1041                         warn("Lite file size bug in fsx!");
1042                         report_failure(200);
1043                 }
1044         }
1045
1046         if (testcalls <= simulatedopcount)
1047                 return;
1048
1049         if (!quiet &&
1050                 ((progressinterval && testcalls % progressinterval == 0) ||
1051                        (debug &&
1052                        (monitorstart == -1 ||
1053                         (offset + size > monitorstart &&
1054                         (monitorend == -1 || offset <= monitorend))))))
1055                 prt("%lu mapwrite\t0x%x thru\t0x%x\t(0x%x bytes)\n", testcalls,
1056                     offset, offset + size - 1, size);
1057
1058         if (file_size > cur_filesize) {
1059                 if (ftruncate(fd, file_size) == -1) {
1060                         prterr("domapwrite: ftruncate");
1061                         exit(201);
1062                 }
1063         }
1064         pg_offset = offset & PAGE_MASK;
1065         map_size  = pg_offset + size;
1066
1067         if ((p = (char *)mmap(0, map_size, PROT_READ | PROT_WRITE,
1068                               MAP_FILE | MAP_SHARED, fd,
1069                               (off_t)(offset - pg_offset))) == (char *)-1) {
1070                 prterr("domapwrite: mmap");
1071                 report_failure(202);
1072         }
1073         memcpy(p + pg_offset, good_buf + offset, size);
1074         if (msync(p, map_size, MS_SYNC) != 0) {
1075                 prterr("domapwrite: msync");
1076                 report_failure(203);
1077         }
1078
1079         check_eofpage("Write", offset, p, size);
1080
1081         if (munmap(p, map_size) != 0) {
1082                 prterr("domapwrite: munmap");
1083                 report_failure(204);
1084         }
1085 }
1086
1087
1088 void
1089 dotruncate(unsigned size)
1090 {
1091         int oldsize = file_size;
1092
1093         size -= size % truncbdy;
1094         if (size > biggest) {
1095                 biggest = size;
1096                 if (!quiet && testcalls > simulatedopcount)
1097                         prt("truncating to largest ever: 0x%x\n", size);
1098         }
1099
1100         log4(OP_TRUNCATE, 0, size, FL_NONE);
1101
1102         if (size > file_size)
1103                 memset(good_buf + file_size, '\0', size - file_size);
1104         file_size = size;
1105
1106         if (testcalls <= simulatedopcount)
1107                 return;
1108
1109         if ((progressinterval && testcalls % progressinterval == 0) ||
1110             (debug && (monitorstart == -1 || monitorend == -1 ||
1111                       size <= monitorend)))
1112                 prt("%lu trunc\tfrom 0x%x to 0x%x\n", testcalls, oldsize, size);
1113         if (ftruncate(fd, (off_t)size) == -1) {
1114                 prt("ftruncate1: %x\n", size);
1115                 prterr("dotruncate: ftruncate");
1116                 report_failure(160);
1117         }
1118 }
1119
1120 #ifdef FALLOC_FL_PUNCH_HOLE
1121 void
1122 do_punch_hole(unsigned offset, unsigned length)
1123 {
1124         unsigned end_offset;
1125         int max_offset = 0;
1126         int max_len = 0;
1127         int mode = FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE;
1128
1129         if (length == 0) {
1130                 if (!quiet && testcalls > simulatedopcount)
1131                         prt("skipping zero length punch hole\n");
1132                 log4(OP_PUNCH_HOLE, offset, length, FL_SKIPPED);
1133                 return;
1134         }
1135
1136         if (file_size <= (loff_t)offset) {
1137                 if (!quiet && testcalls > simulatedopcount)
1138                         prt("skipping hole punch off the end of the file\n");
1139                 log4(OP_PUNCH_HOLE, offset, length, FL_SKIPPED);
1140                 return;
1141         }
1142
1143         end_offset = offset + length;
1144
1145         log4(OP_PUNCH_HOLE, offset, length, FL_NONE);
1146
1147         if (testcalls <= simulatedopcount)
1148                 return;
1149
1150         if ((progressinterval && testcalls % progressinterval == 0) ||
1151             (debug && (monitorstart == -1 || monitorend == -1 ||
1152                       end_offset <= monitorend))) {
1153                 prt("%lu punch\tfrom 0x%x to 0x%x, (0x%x bytes)\n", testcalls,
1154                         offset, offset+length, length);
1155         }
1156         if (fallocate(fd, mode, (loff_t)offset, (loff_t)length) == -1) {
1157                 prt("punch hole: 0x%x to 0x%x\n", offset, offset + length);
1158                 prterr("do_punch_hole: fallocate");
1159                 report_failure(161);
1160         }
1161
1162
1163         max_offset = offset < file_size ? offset : file_size;
1164         max_len = max_offset + length <= file_size ? length :
1165                         file_size - max_offset;
1166         memset(good_buf + max_offset, '\0', max_len);
1167 }
1168
1169 #else
1170 void
1171 do_punch_hole(unsigned offset, unsigned length)
1172 {
1173         return;
1174 }
1175 #endif
1176
1177 #ifdef FALLOC_FL_ZERO_RANGE
1178 void
1179 do_zero_range(unsigned offset, unsigned length, int keep_size)
1180 {
1181         unsigned end_offset;
1182         int mode = FALLOC_FL_ZERO_RANGE;
1183
1184         if (length == 0) {
1185                 if (!quiet && testcalls > simulatedopcount)
1186                         prt("skipping zero length zero range\n");
1187                 log4(OP_ZERO_RANGE, offset, length, FL_SKIPPED |
1188                      (keep_size ? FL_KEEP_SIZE : FL_NONE));
1189                 return;
1190         }
1191
1192         end_offset = keep_size ? 0 : offset + length;
1193
1194         if (end_offset > biggest) {
1195                 biggest = end_offset;
1196                 if (!quiet && testcalls > simulatedopcount)
1197                         prt("zero_range to largest ever: 0x%x\n", end_offset);
1198         }
1199
1200         /*
1201          * last arg matches fallocate string array index in logdump:
1202          *      0: allocate past EOF
1203          *      1: extending prealloc
1204          *      2: interior prealloc
1205          */
1206         log4(OP_ZERO_RANGE, offset, length,
1207              keep_size ? FL_KEEP_SIZE : FL_NONE);
1208
1209         if (testcalls <= simulatedopcount)
1210                 return;
1211
1212         if ((progressinterval && testcalls % progressinterval == 0) ||
1213             (debug && (monitorstart == -1 || monitorend == -1 ||
1214                       end_offset <= monitorend))) {
1215                 prt("%lu zero\tfrom 0x%x to 0x%x, (0x%x bytes)\n", testcalls,
1216                         offset, offset+length, length);
1217         }
1218         if (fallocate(fd, mode, (loff_t)offset, (loff_t)length) == -1) {
1219                 prt("zero range: 0x%x to 0x%x\n", offset, offset + length);
1220                 prterr("do_zero_range: fallocate");
1221                 report_failure(161);
1222         }
1223
1224         memset(good_buf + offset, '\0', length);
1225 }
1226
1227 #else
1228 void
1229 do_zero_range(unsigned offset, unsigned length, int keep_size)
1230 {
1231         return;
1232 }
1233 #endif
1234
1235 #ifdef FALLOC_FL_COLLAPSE_RANGE
1236 void
1237 do_collapse_range(unsigned offset, unsigned length)
1238 {
1239         unsigned end_offset;
1240         int mode = FALLOC_FL_COLLAPSE_RANGE;
1241
1242         if (length == 0) {
1243                 if (!quiet && testcalls > simulatedopcount)
1244                         prt("skipping zero length collapse range\n");
1245                 log4(OP_COLLAPSE_RANGE, offset, length, FL_SKIPPED);
1246                 return;
1247         }
1248
1249         end_offset = offset + length;
1250         if ((loff_t)end_offset >= file_size) {
1251                 if (!quiet && testcalls > simulatedopcount)
1252                         prt("skipping collapse range behind EOF\n");
1253                 log4(OP_COLLAPSE_RANGE, offset, length, FL_SKIPPED);
1254                 return;
1255         }
1256
1257         log4(OP_COLLAPSE_RANGE, offset, length, FL_NONE);
1258
1259         if (testcalls <= simulatedopcount)
1260                 return;
1261
1262         if ((progressinterval && testcalls % progressinterval == 0) ||
1263             (debug && (monitorstart == -1 || monitorend == -1 ||
1264                       end_offset <= monitorend))) {
1265                 prt("%lu collapse\tfrom 0x%x to 0x%x, (0x%x bytes)\n", testcalls,
1266                         offset, offset+length, length);
1267         }
1268         if (fallocate(fd, mode, (loff_t)offset, (loff_t)length) == -1) {
1269                 prt("collapse range: 0x%x to 0x%x\n", offset, offset + length);
1270                 prterr("do_collapse_range: fallocate");
1271                 report_failure(161);
1272         }
1273
1274         memmove(good_buf + offset, good_buf + end_offset,
1275                 file_size - end_offset);
1276         file_size -= length;
1277 }
1278
1279 #else
1280 void
1281 do_collapse_range(unsigned offset, unsigned length)
1282 {
1283         return;
1284 }
1285 #endif
1286
1287 #ifdef FALLOC_FL_INSERT_RANGE
1288 void
1289 do_insert_range(unsigned offset, unsigned length)
1290 {
1291         unsigned end_offset;
1292         int mode = FALLOC_FL_INSERT_RANGE;
1293
1294         if (length == 0) {
1295                 if (!quiet && testcalls > simulatedopcount)
1296                         prt("skipping zero length insert range\n");
1297                 log4(OP_INSERT_RANGE, offset, length, FL_SKIPPED);
1298                 return;
1299         }
1300
1301         if ((loff_t)offset >= file_size) {
1302                 if (!quiet && testcalls > simulatedopcount)
1303                         prt("skipping insert range behind EOF\n");
1304                 log4(OP_INSERT_RANGE, offset, length, FL_SKIPPED);
1305                 return;
1306         }
1307
1308         log4(OP_INSERT_RANGE, offset, length, FL_NONE);
1309
1310         if (testcalls <= simulatedopcount)
1311                 return;
1312
1313         end_offset = offset + length;
1314         if ((progressinterval && testcalls % progressinterval == 0) ||
1315             (debug && (monitorstart == -1 || monitorend == -1 ||
1316                       end_offset <= monitorend))) {
1317                 prt("%lu insert\tfrom 0x%x to 0x%x, (0x%x bytes)\n", testcalls,
1318                         offset, offset+length, length);
1319         }
1320         if (fallocate(fd, mode, (loff_t)offset, (loff_t)length) == -1) {
1321                 prt("insert range: 0x%x to 0x%x\n", offset, offset + length);
1322                 prterr("do_insert_range: fallocate");
1323                 report_failure(161);
1324         }
1325
1326         memmove(good_buf + end_offset, good_buf + offset,
1327                 file_size - offset);
1328         memset(good_buf + offset, '\0', length);
1329         file_size += length;
1330 }
1331
1332 #else
1333 void
1334 do_insert_range(unsigned offset, unsigned length)
1335 {
1336         return;
1337 }
1338 #endif
1339
1340 #ifdef FICLONERANGE
1341 int
1342 test_clone_range(void)
1343 {
1344         struct file_clone_range fcr = {
1345                 .src_fd = fd,
1346         };
1347
1348         if (ioctl(fd, FICLONERANGE, &fcr) &&
1349             (errno = EOPNOTSUPP || errno == ENOTTY)) {
1350                 if (!quiet)
1351                         fprintf(stderr,
1352                                 "main: filesystem does not support "
1353                                 "clone range, disabling!\n");
1354                 return 0;
1355         }
1356
1357         return 1;
1358 }
1359
1360 void
1361 do_clone_range(unsigned offset, unsigned length, unsigned dest)
1362 {
1363         struct file_clone_range fcr = {
1364                 .src_fd = fd,
1365                 .src_offset = offset,
1366                 .src_length = length,
1367                 .dest_offset = dest,
1368         };
1369
1370         if (length == 0) {
1371                 if (!quiet && testcalls > simulatedopcount)
1372                         prt("skipping zero length clone range\n");
1373                 log5(OP_CLONE_RANGE, offset, length, dest, FL_SKIPPED);
1374                 return;
1375         }
1376
1377         if ((loff_t)offset >= file_size) {
1378                 if (!quiet && testcalls > simulatedopcount)
1379                         prt("skipping clone range behind EOF\n");
1380                 log5(OP_CLONE_RANGE, offset, length, dest, FL_SKIPPED);
1381                 return;
1382         }
1383
1384         if (dest + length > biggest) {
1385                 biggest = dest + length;
1386                 if (!quiet && testcalls > simulatedopcount)
1387                         prt("cloning to largest ever: 0x%x\n", dest + length);
1388         }
1389
1390         log5(OP_CLONE_RANGE, offset, length, dest, FL_NONE);
1391
1392         if (testcalls <= simulatedopcount)
1393                 return;
1394
1395         if ((progressinterval && testcalls % progressinterval == 0) ||
1396             (debug && (monitorstart == -1 || monitorend == -1 ||
1397                        dest <= monitorstart || dest + length <= monitorend))) {
1398                 prt("%lu clone\tfrom 0x%x to 0x%x, (0x%x bytes) at 0x%x\n",
1399                         testcalls, offset, offset+length, length, dest);
1400         }
1401
1402         if (ioctl(fd, FICLONERANGE, &fcr) == -1) {
1403                 prt("clone range: 0x%x to 0x%x at 0x%x\n", offset,
1404                                 offset + length, dest);
1405                 prterr("do_clone_range: FICLONERANGE");
1406                 report_failure(161);
1407         }
1408
1409         memcpy(good_buf + dest, good_buf + offset, length);
1410         if (dest > file_size)
1411                 memset(good_buf + file_size, '\0', dest - file_size);
1412         if (dest + length > file_size)
1413                 file_size = dest + length;
1414 }
1415
1416 #else
1417 int
1418 test_clone_range(void)
1419 {
1420         return 0;
1421 }
1422
1423 void
1424 do_clone_range(unsigned offset, unsigned length, unsigned dest)
1425 {
1426         return;
1427 }
1428 #endif
1429
1430 #ifdef FIDEDUPERANGE
1431 int
1432 test_dedupe_range(void)
1433 {
1434         struct file_dedupe_range *fdr;
1435         off_t new_len;
1436         int error;
1437         int ret = 1;
1438
1439         /* Alloc memory */
1440         fdr = calloc(sizeof(struct file_dedupe_range_info) +
1441                      sizeof(struct file_dedupe_range), 1);
1442         if (!fdr) {
1443                 prterr("do_dedupe_range: malloc");
1444                 report_failure(161);
1445         }
1446
1447         /* Make sure we have at least two blocks */
1448         new_len = block_size * 2;
1449         if (file_size < new_len && ftruncate(fd, new_len)) {
1450                 warn("main: ftruncate");
1451                 exit(132);
1452         }
1453
1454         /* Try to dedupe them */
1455         fdr->src_length = block_size;
1456         fdr->dest_count = 1;
1457         fdr->info[0].dest_fd = fd;
1458         fdr->info[0].dest_offset = block_size;
1459
1460         if (ioctl(fd, FIDEDUPERANGE, fdr))
1461                 error = errno;
1462         else if (fdr->info[0].status < 0)
1463                 error = -fdr->info[0].status;
1464         else
1465                 error = 0;
1466
1467         if (error == EOPNOTSUPP || error == ENOTTY) {
1468                 if (!quiet)
1469                         fprintf(stderr,
1470                                 "main: filesystem does not support "
1471                                 "dedupe range, disabling!\n");
1472                 ret = 0;
1473         }
1474
1475         /* Put the file back the way it was. */
1476         if (file_size < new_len && ftruncate(fd, file_size)) {
1477                 warn("main: ftruncate");
1478                 exit(132);
1479         }
1480
1481         free(fdr);
1482         return ret;
1483 }
1484
1485 void
1486 do_dedupe_range(unsigned offset, unsigned length, unsigned dest)
1487 {
1488         struct file_dedupe_range *fdr;
1489
1490         if (length == 0) {
1491                 if (!quiet && testcalls > simulatedopcount)
1492                         prt("skipping zero length dedupe range\n");
1493                 log5(OP_DEDUPE_RANGE, offset, length, dest, FL_SKIPPED);
1494                 return;
1495         }
1496
1497         if ((loff_t)offset >= file_size) {
1498                 if (!quiet && testcalls > simulatedopcount)
1499                         prt("skipping dedupe range behind EOF\n");
1500                 log5(OP_DEDUPE_RANGE, offset, length, dest, FL_SKIPPED);
1501                 return;
1502         }
1503
1504         log5(OP_DEDUPE_RANGE, offset, length, dest, FL_NONE);
1505
1506         if (testcalls <= simulatedopcount)
1507                 return;
1508
1509         if ((progressinterval && testcalls % progressinterval == 0) ||
1510             (debug && (monitorstart == -1 || monitorend == -1 ||
1511                        dest <= monitorstart || dest + length <= monitorend))) {
1512                 prt("%lu dedupe\tfrom 0x%x to 0x%x, (0x%x bytes) at 0x%x\n",
1513                         testcalls, offset, offset+length, length, dest);
1514         }
1515
1516         /* Alloc memory */
1517         fdr = calloc(sizeof(struct file_dedupe_range_info) +
1518                      sizeof(struct file_dedupe_range), 1);
1519         if (!fdr) {
1520                 prterr("do_dedupe_range: malloc");
1521                 report_failure(161);
1522         }
1523
1524         /* Dedupe data blocks */
1525         fdr->src_offset = offset;
1526         fdr->src_length = length;
1527         fdr->dest_count = 1;
1528         fdr->info[0].dest_fd = fd;
1529         fdr->info[0].dest_offset = dest;
1530
1531         if (ioctl(fd, FIDEDUPERANGE, fdr) == -1) {
1532                 prt("dedupe range: 0x%x to 0x%x at 0x%x\n", offset,
1533                                 offset + length, dest);
1534                 prterr("do_dedupe_range(0): FIDEDUPERANGE");
1535                 report_failure(161);
1536         } else if (fdr->info[0].status < 0) {
1537                 errno = -fdr->info[0].status;
1538                 prt("dedupe range: 0x%x to 0x%x at 0x%x\n", offset,
1539                                 offset + length, dest);
1540                 prterr("do_dedupe_range(1): FIDEDUPERANGE");
1541                 report_failure(161);
1542         }
1543
1544         free(fdr);
1545 }
1546
1547 #else
1548 int
1549 test_dedupe_range(void)
1550 {
1551         return 0;
1552 }
1553
1554 void
1555 do_dedupe_range(unsigned offset, unsigned length, unsigned dest)
1556 {
1557         return;
1558 }
1559 #endif
1560
1561 #ifdef HAVE_LINUX_FALLOC_H
1562 /* fallocate is basically a no-op unless extending, then a lot like a truncate */
1563 void
1564 do_preallocate(unsigned offset, unsigned length, int keep_size)
1565 {
1566         unsigned end_offset;
1567
1568         if (length == 0) {
1569                 if (!quiet && testcalls > simulatedopcount)
1570                         prt("skipping zero length fallocate\n");
1571                 log4(OP_FALLOCATE, offset, length, FL_SKIPPED |
1572                      (keep_size ? FL_KEEP_SIZE : FL_NONE));
1573                 return;
1574         }
1575
1576         end_offset = keep_size ? 0 : offset + length;
1577
1578         if (end_offset > biggest) {
1579                 biggest = end_offset;
1580                 if (!quiet && testcalls > simulatedopcount)
1581                         prt("fallocating to largest ever: 0x%x\n", end_offset);
1582         }
1583
1584         /*
1585          * last arg matches fallocate string array index in logdump:
1586          *      0: allocate past EOF
1587          *      1: extending prealloc
1588          *      2: interior prealloc
1589          */
1590         log4(OP_FALLOCATE, offset, length,
1591              keep_size ? FL_KEEP_SIZE : FL_NONE);
1592
1593         if (end_offset > file_size) {
1594                 memset(good_buf + file_size, '\0', end_offset - file_size);
1595                 file_size = end_offset;
1596         }
1597
1598         if (testcalls <= simulatedopcount)
1599                 return;
1600
1601         if ((progressinterval && testcalls % progressinterval == 0) ||
1602             (debug && (monitorstart == -1 || monitorend == -1 ||
1603                       end_offset <= monitorend)))
1604                 prt("%lu falloc\tfrom 0x%x to 0x%x (0x%x bytes)\n", testcalls,
1605                                 offset, offset + length, length);
1606         if (fallocate(fd, keep_size ? FALLOC_FL_KEEP_SIZE : 0, (loff_t)offset, (loff_t)length) == -1) {
1607                 prt("fallocate: 0x%x to 0x%x\n", offset, offset + length);
1608                 prterr("do_preallocate: fallocate");
1609                 report_failure(161);
1610         }
1611 }
1612 #else
1613 void
1614 do_preallocate(unsigned offset, unsigned length, int keep_size)
1615 {
1616         return;
1617 }
1618 #endif
1619
1620 void
1621 writefileimage()
1622 {
1623         ssize_t iret;
1624
1625         if (lseek(fd, (off_t)0, SEEK_SET) == (off_t)-1) {
1626                 prterr("writefileimage: lseek");
1627                 report_failure(171);
1628         }
1629         iret = write(fd, good_buf, file_size);
1630         if ((off_t)iret != file_size) {
1631                 if (iret == -1)
1632                         prterr("writefileimage: write");
1633                 else
1634                         prt("short write: 0x%x bytes instead of 0x%llx\n",
1635                             iret, (unsigned long long)file_size);
1636                 report_failure(172);
1637         }
1638         if (lite ? 0 : ftruncate(fd, file_size) == -1) {
1639                 prt("ftruncate2: %llx\n", (unsigned long long)file_size);
1640                 prterr("writefileimage: ftruncate");
1641                 report_failure(173);
1642         }
1643 }
1644
1645
1646 void
1647 docloseopen(void)
1648 {
1649         if (testcalls <= simulatedopcount)
1650                 return;
1651
1652         if (debug)
1653                 prt("%lu close/open\n", testcalls);
1654         if (close(fd)) {
1655                 prterr("docloseopen: close");
1656                 report_failure(180);
1657         }
1658         fd = open(fname, O_RDWR|o_direct, 0);
1659         if (fd < 0) {
1660                 prterr("docloseopen: open");
1661                 report_failure(181);
1662         }
1663 }
1664
1665 void
1666 dofsync(void)
1667 {
1668         int ret;
1669
1670         if (testcalls <= simulatedopcount)
1671                 return;
1672         if (debug)
1673                 prt("%lu fsync\n", testcalls);
1674         log4(OP_FSYNC, 0, 0, 0);
1675         ret = fsync(fd);
1676         if (ret < 0) {
1677                 prterr("dofsync");
1678                 report_failure(210);
1679         }
1680         mark_log();
1681         dump_fsync_buffer();
1682         mark_nr++;
1683 }
1684
1685 #define TRIM_OFF(off, size)                     \
1686 do {                                            \
1687         if (size)                               \
1688                 (off) %= (size);                \
1689         else                                    \
1690                 (off) = 0;                      \
1691 } while (0)
1692
1693 #define TRIM_LEN(off, len, size)                \
1694 do {                                            \
1695         if ((off) + (len) > (size))             \
1696                 (len) = (size) - (off);         \
1697 } while (0)
1698
1699 #define TRIM_OFF_LEN(off, len, size)            \
1700 do {                                            \
1701         TRIM_OFF(off, size);                    \
1702         TRIM_LEN(off, len, size);               \
1703 } while (0)
1704
1705 void
1706 cleanup(int sig)
1707 {
1708         if (sig)
1709                 prt("signal %d\n", sig);
1710         prt("testcalls = %lu\n", testcalls);
1711         exit(sig);
1712 }
1713
1714 static int
1715 op_args_count(int operation)
1716 {
1717         switch (operation) {
1718         case OP_CLONE_RANGE:
1719         case OP_DEDUPE_RANGE:
1720                 return 4;
1721         default:
1722                 return 3;
1723         }
1724 }
1725
1726 static int
1727 read_op(struct log_entry *log_entry)
1728 {
1729         char line[256];
1730
1731         memset(log_entry, 0, sizeof(*log_entry));
1732         log_entry->operation = -1;
1733
1734         while (log_entry->operation == -1) {
1735                 char *str;
1736                 int i;
1737
1738                 do {
1739                         if (!fgets(line, sizeof(line), replayopsf)) {
1740                                 if (feof(replayopsf)) {
1741                                         replayopsf = NULL;
1742                                         return 0;
1743                                 }
1744                                 goto fail;
1745                         }
1746                         str = strtok(line, " \t\n");
1747                 } while (!str || str[0] == '#');
1748
1749                 if (strcmp(str, "skip") == 0) {
1750                         log_entry->flags |= FL_SKIPPED;
1751                         str = strtok(NULL, " \t\n");
1752                         if (!str)
1753                                 goto fail;
1754                 }
1755                 log_entry->operation = op_code(str);
1756                 if (log_entry->operation == -1)
1757                         goto fail;
1758                 log_entry->nr_args = op_args_count(log_entry->operation);
1759                 for (i = 0; i < log_entry->nr_args; i++) {
1760                         char *end;
1761
1762                         str = strtok(NULL, " \t\n");
1763                         if (!str)
1764                                 goto fail;
1765                         log_entry->args[i] = strtoul(str, &end, 0);
1766                         if (*end)
1767                                 goto fail;
1768                 }
1769                 while ((str = strtok(NULL, " \t\n"))) {
1770                         if (strcmp(str, "keep_size") == 0)
1771                                 log_entry->flags |= FL_KEEP_SIZE;
1772                         else if (strcmp(str, "close_open") == 0)
1773                                 log_entry->flags |= FL_CLOSE_OPEN;
1774                         else if (strcmp(str, "*") == 0)
1775                                 ;  /* overlap marker; ignore */
1776                         else
1777                                 goto fail;
1778                 }
1779         }
1780         return 1;
1781
1782 fail:
1783         fprintf(stderr, "%s: parse error\n", replayops);
1784         fclose(replayopsf);
1785         replayopsf = NULL;
1786         cleanup(100);  /* doesn't return */
1787         return 0;
1788 }
1789
1790
1791 int
1792 test(void)
1793 {
1794         unsigned long   offset, offset2;
1795         unsigned long   size;
1796         unsigned long   rv;
1797         unsigned long   op;
1798         int             keep_size = 0;
1799
1800         if (simulatedopcount > 0 && testcalls == simulatedopcount)
1801                 writefileimage();
1802
1803         testcalls++;
1804
1805         if (debugstart > 0 && testcalls >= debugstart)
1806                 debug = 1;
1807
1808         if (!quiet && testcalls < simulatedopcount && testcalls % 100000 == 0)
1809                 prt("%lu...\n", testcalls);
1810
1811         if (replayopsf) {
1812                 struct log_entry log_entry;
1813
1814                 while (read_op(&log_entry)) {
1815                         if (log_entry.flags & FL_SKIPPED) {
1816                                 log4(log_entry.operation,
1817                                      log_entry.args[0], log_entry.args[1],
1818                                      log_entry.flags);
1819                                 continue;
1820                         }
1821
1822                         op = log_entry.operation;
1823                         offset = log_entry.args[0];
1824                         size = log_entry.args[1];
1825                         offset2 = log_entry.args[2];
1826                         closeopen = !!(log_entry.flags & FL_CLOSE_OPEN);
1827                         keep_size = !!(log_entry.flags & FL_KEEP_SIZE);
1828                         goto have_op;
1829                 }
1830                 return 0;
1831         }
1832
1833         rv = random();
1834         if (closeprob)
1835                 closeopen = (rv >> 3) < (1 << 28) / closeprob;
1836
1837         offset = random();
1838         offset2 = 0;
1839         size = maxoplen;
1840         if (randomoplen)
1841                 size = random() % (maxoplen + 1);
1842
1843         /* calculate appropriate op to run */
1844         if (lite)
1845                 op = rv % OP_MAX_LITE;
1846         else if (!integrity)
1847                 op = rv % OP_MAX_FULL;
1848         else
1849                 op = rv % OP_MAX_INTEGRITY;
1850
1851         switch(op) {
1852         case OP_TRUNCATE:
1853                 if (!style)
1854                         size = random() % maxfilelen;
1855                 break;
1856         case OP_FALLOCATE:
1857                 if (fallocate_calls && size && keep_size_calls)
1858                         keep_size = random() % 2;
1859                 break;
1860         case OP_ZERO_RANGE:
1861                 if (zero_range_calls && size && keep_size_calls)
1862                         keep_size = random() % 2;
1863                 break;
1864         case OP_CLONE_RANGE:
1865                 TRIM_OFF_LEN(offset, size, file_size);
1866                 offset = offset & ~(block_size - 1);
1867                 size = size & ~(block_size - 1);
1868                 do {
1869                         offset2 = random();
1870                         TRIM_OFF(offset2, maxfilelen);
1871                         offset2 = offset2 & ~(block_size - 1);
1872                 } while (llabs(offset2 - offset) < size ||
1873                          offset2 + size > maxfilelen);
1874                 break;
1875         case OP_DEDUPE_RANGE:
1876                 {
1877                         int tries = 0;
1878
1879                         TRIM_OFF_LEN(offset, size, file_size);
1880                         offset = offset & ~(block_size - 1);
1881                         size = size & ~(block_size - 1);
1882                         do {
1883                                 if (tries++ >= 30) {
1884                                         size = 0;
1885                                         break;
1886                                 }
1887                                 offset2 = random();
1888                                 TRIM_OFF(offset2, file_size);
1889                                 offset2 = offset2 & ~(block_size - 1);
1890                         } while (llabs(offset2 - offset) < size ||
1891                                  offset2 + size > file_size);
1892                         break;
1893                 }
1894         }
1895
1896 have_op:
1897
1898         switch (op) {
1899         case OP_MAPREAD:
1900                 if (!mapped_reads)
1901                         op = OP_READ;
1902                 break;
1903         case OP_MAPWRITE:
1904                 if (!mapped_writes)
1905                         op = OP_WRITE;
1906                 break;
1907         case OP_FALLOCATE:
1908                 if (!fallocate_calls) {
1909                         log4(OP_FALLOCATE, offset, size, FL_SKIPPED);
1910                         goto out;
1911                 }
1912                 break;
1913         case OP_PUNCH_HOLE:
1914                 if (!punch_hole_calls) {
1915                         log4(OP_PUNCH_HOLE, offset, size, FL_SKIPPED);
1916                         goto out;
1917                 }
1918                 break;
1919         case OP_ZERO_RANGE:
1920                 if (!zero_range_calls) {
1921                         log4(OP_ZERO_RANGE, offset, size, FL_SKIPPED);
1922                         goto out;
1923                 }
1924                 break;
1925         case OP_COLLAPSE_RANGE:
1926                 if (!collapse_range_calls) {
1927                         log4(OP_COLLAPSE_RANGE, offset, size, FL_SKIPPED);
1928                         goto out;
1929                 }
1930                 break;
1931         case OP_INSERT_RANGE:
1932                 if (!insert_range_calls) {
1933                         log4(OP_INSERT_RANGE, offset, size, FL_SKIPPED);
1934                         goto out;
1935                 }
1936                 break;
1937         case OP_CLONE_RANGE:
1938                 if (!clone_range_calls) {
1939                         log5(op, offset, size, offset2, FL_SKIPPED);
1940                         goto out;
1941                 }
1942                 break;
1943         case OP_DEDUPE_RANGE:
1944                 if (!dedupe_range_calls) {
1945                         log5(op, offset, size, offset2, FL_SKIPPED);
1946                         goto out;
1947                 }
1948                 break;
1949         }
1950
1951         switch (op) {
1952         case OP_READ:
1953                 TRIM_OFF_LEN(offset, size, file_size);
1954                 doread(offset, size);
1955                 break;
1956
1957         case OP_WRITE:
1958                 TRIM_OFF_LEN(offset, size, maxfilelen);
1959                 dowrite(offset, size);
1960                 break;
1961
1962         case OP_MAPREAD:
1963                 TRIM_OFF_LEN(offset, size, file_size);
1964                 domapread(offset, size);
1965                 break;
1966
1967         case OP_MAPWRITE:
1968                 TRIM_OFF_LEN(offset, size, maxfilelen);
1969                 domapwrite(offset, size);
1970                 break;
1971
1972         case OP_TRUNCATE:
1973                 dotruncate(size);
1974                 break;
1975
1976         case OP_FALLOCATE:
1977                 TRIM_OFF_LEN(offset, size, maxfilelen);
1978                 do_preallocate(offset, size, keep_size);
1979                 break;
1980
1981         case OP_PUNCH_HOLE:
1982                 TRIM_OFF_LEN(offset, size, file_size);
1983                 do_punch_hole(offset, size);
1984                 break;
1985         case OP_ZERO_RANGE:
1986                 TRIM_OFF_LEN(offset, size, file_size);
1987                 do_zero_range(offset, size, keep_size);
1988                 break;
1989         case OP_COLLAPSE_RANGE:
1990                 TRIM_OFF_LEN(offset, size, file_size - 1);
1991                 offset = offset & ~(block_size - 1);
1992                 size = size & ~(block_size - 1);
1993                 if (size == 0) {
1994                         log4(OP_COLLAPSE_RANGE, offset, size, FL_SKIPPED);
1995                         goto out;
1996                 }
1997                 do_collapse_range(offset, size);
1998                 break;
1999         case OP_INSERT_RANGE:
2000                 TRIM_OFF(offset, file_size);
2001                 TRIM_LEN(file_size, size, maxfilelen);
2002                 offset = offset & ~(block_size - 1);
2003                 size = size & ~(block_size - 1);
2004                 if (size == 0) {
2005                         log4(OP_INSERT_RANGE, offset, size, FL_SKIPPED);
2006                         goto out;
2007                 }
2008                 if (file_size + size > maxfilelen) {
2009                         log4(OP_INSERT_RANGE, offset, size, FL_SKIPPED);
2010                         goto out;
2011                 }
2012
2013                 do_insert_range(offset, size);
2014                 break;
2015         case OP_CLONE_RANGE:
2016                 if (size == 0) {
2017                         log5(OP_CLONE_RANGE, offset, size, offset2, FL_SKIPPED);
2018                         goto out;
2019                 }
2020                 if (offset2 + size > maxfilelen) {
2021                         log5(OP_CLONE_RANGE, offset, size, offset2, FL_SKIPPED);
2022                         goto out;
2023                 }
2024
2025                 do_clone_range(offset, size, offset2);
2026                 break;
2027         case OP_DEDUPE_RANGE:
2028                 if (size == 0) {
2029                         log5(OP_DEDUPE_RANGE, offset, size, offset2, FL_SKIPPED);
2030                         goto out;
2031                 }
2032                 if (offset2 + size > maxfilelen) {
2033                         log5(OP_DEDUPE_RANGE, offset, size, offset2, FL_SKIPPED);
2034                         goto out;
2035                 }
2036
2037                 do_dedupe_range(offset, size, offset2);
2038                 break;
2039         case OP_FSYNC:
2040                 dofsync();
2041                 break;
2042         default:
2043                 prterr("test: unknown operation");
2044                 report_failure(42);
2045                 break;
2046         }
2047
2048         if (check_file && testcalls > simulatedopcount)
2049                 check_contents();
2050
2051 out:
2052         if (sizechecks && testcalls > simulatedopcount)
2053                 check_size();
2054         if (closeopen)
2055                 docloseopen();
2056         return 1;
2057 }
2058
2059
2060 void
2061 usage(void)
2062 {
2063         fprintf(stdout, "usage: %s",
2064                 "fsx [-dknqxABFJLOWZ] [-b opnum] [-c Prob] [-g filldata] [-i logdev] [-j logid] [-l flen] [-m start:end] [-o oplen] [-p progressinterval] [-r readbdy] [-s style] [-t truncbdy] [-w writebdy] [-D startingop] [-N numops] [-P dirpath] [-S seed] fname\n\
2065         -b opnum: beginning operation number (default 1)\n\
2066         -c P: 1 in P chance of file close+open at each op (default infinity)\n\
2067         -d: debug output for all operations\n\
2068         -f flush and invalidate cache after I/O\n\
2069         -g X: write character X instead of random generated data\n\
2070         -i logdev: do integrity testing, logdev is the dm log writes device\n\
2071         -j logid: prefix debug log messsages with this id\n\
2072         -k: do not truncate existing file and use its size as upper bound on file size\n\
2073         -l flen: the upper bound on file size (default 262144)\n\
2074         -m startop:endop: monitor (print debug output) specified byte range (default 0:infinity)\n\
2075         -n: no verifications of file size\n\
2076         -o oplen: the upper bound on operation size (default 65536)\n\
2077         -p progressinterval: debug output at specified operation interval\n\
2078         -q: quieter operation\n\
2079         -r readbdy: 4096 would make reads page aligned (default 1)\n\
2080         -s style: 1 gives smaller truncates (default 0)\n\
2081         -t truncbdy: 4096 would make truncates page aligned (default 1)\n\
2082         -w writebdy: 4096 would make writes page aligned (default 1)\n\
2083         -x: preallocate file space before starting, XFS only (default 0)\n\
2084         -y synchronize changes to a file\n"
2085
2086 #ifdef AIO
2087 "       -A: Use the AIO system calls\n"
2088 #endif
2089 "       -D startingop: debug output starting at specified operation\n"
2090 #ifdef HAVE_LINUX_FALLOC_H
2091 "       -F: Do not use fallocate (preallocation) calls\n"
2092 #endif
2093 #ifdef FALLOC_FL_PUNCH_HOLE
2094 "       -H: Do not use punch hole calls\n"
2095 #endif
2096 #ifdef FALLOC_FL_ZERO_RANGE
2097 "       -z: Do not use zero range calls\n"
2098 #endif
2099 #ifdef FALLOC_FL_COLLAPSE_RANGE
2100 "       -C: Do not use collapse range calls\n"
2101 #endif
2102 #ifdef FALLOC_FL_INSERT_RANGE
2103 "       -I: Do not use insert range calls\n"
2104 #endif
2105 #ifdef FICLONERANGE
2106 "       -J: Do not use clone range calls\n"
2107 #endif
2108 #ifdef FIDEDUPERANGE
2109 "       -B: Do not use dedupe range calls\n"
2110 #endif
2111 "       -L: fsxLite - no file creations & no file size changes\n\
2112         -N numops: total # operations to do (default infinity)\n\
2113         -O: use oplen (see -o flag) for every op (default random)\n\
2114         -P: save .fsxlog .fsxops and .fsxgood files in dirpath (default ./)\n\
2115         -S seed: for random # generator (default 1) 0 gets timestamp\n\
2116         -W: mapped write operations DISabled\n\
2117         -X: Read file and compare to good buffer after every operation.\n\
2118         -R: read() system calls only (mapped reads disabled)\n\
2119         -Z: O_DIRECT (use -R, -W, -r and -w too)\n\
2120         --replay-ops opsfile: replay ops from recorded .fsxops file\n\
2121         --record-ops[=opsfile]: dump ops file also on success. optionally specify ops file name\n\
2122         fname: this filename is REQUIRED (no default)\n");
2123         exit(90);
2124 }
2125
2126
2127 int
2128 getnum(char *s, char **e)
2129 {
2130         int ret;
2131
2132         *e = (char *) 0;
2133         ret = strtol(s, e, 0);
2134         if (*e)
2135                 switch (**e) {
2136                 case 'b':
2137                 case 'B':
2138                         ret *= 512;
2139                         *e = *e + 1;
2140                         break;
2141                 case 'k':
2142                 case 'K':
2143                         ret *= 1024;
2144                         *e = *e + 1;
2145                         break;
2146                 case 'm':
2147                 case 'M':
2148                         ret *= 1024*1024;
2149                         *e = *e + 1;
2150                         break;
2151                 case 'w':
2152                 case 'W':
2153                         ret *= 4;
2154                         *e = *e + 1;
2155                         break;
2156                 }
2157         return (ret);
2158 }
2159
2160 #ifdef AIO
2161
2162 #define QSZ     1024
2163 io_context_t    io_ctx;
2164 struct iocb     iocb;
2165
2166 int aio_setup()
2167 {
2168         int ret;
2169         ret = io_queue_init(QSZ, &io_ctx);
2170         if (ret != 0) {
2171                 fprintf(stderr, "aio_setup: io_queue_init failed: %s\n",
2172                         strerror(ret));
2173                 return(-1);
2174         }
2175         return(0);
2176 }
2177
2178 int
2179 __aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
2180 {
2181         struct io_event event;
2182         static struct timespec ts;
2183         struct iocb *iocbs[] = { &iocb };
2184         int ret;
2185         long res;
2186
2187         if (rw == READ) {
2188                 io_prep_pread(&iocb, fd, buf, len, offset);
2189         } else {
2190                 io_prep_pwrite(&iocb, fd, buf, len, offset);
2191         }
2192
2193         ts.tv_sec = 30;
2194         ts.tv_nsec = 0;
2195         ret = io_submit(io_ctx, 1, iocbs);
2196         if (ret != 1) {
2197                 fprintf(stderr, "errcode=%d\n", ret);
2198                 fprintf(stderr, "aio_rw: io_submit failed: %s\n",
2199                                 strerror(ret));
2200                 goto out_error;
2201         }
2202
2203         ret = io_getevents(io_ctx, 1, 1, &event, &ts);
2204         if (ret != 1) {
2205                 if (ret == 0)
2206                         fprintf(stderr, "aio_rw: no events available\n");
2207                 else {
2208                         fprintf(stderr, "errcode=%d\n", -ret);
2209                         fprintf(stderr, "aio_rw: io_getevents failed: %s\n",
2210                                         strerror(-ret));
2211                 }
2212                 goto out_error;
2213         }
2214         if (len != event.res) {
2215                 /*
2216                  * The b0rked libaio defines event.res as unsigned.
2217                  * However the kernel strucuture has it signed,
2218                  * and it's used to pass negated error value.
2219                  * Till the library is fixed use the temp var.
2220                  */
2221                 res = (long)event.res;
2222                 if (res >= 0)
2223                         fprintf(stderr, "bad io length: %lu instead of %u\n",
2224                                         res, len);
2225                 else {
2226                         fprintf(stderr, "errcode=%ld\n", -res);
2227                         fprintf(stderr, "aio_rw: async io failed: %s\n",
2228                                         strerror(-res));
2229                         ret = res;
2230                         goto out_error;
2231                 }
2232
2233         }
2234         return event.res;
2235
2236 out_error:
2237         /*
2238          * The caller expects error return in traditional libc
2239          * convention, i.e. -1 and the errno set to error.
2240          */
2241         errno = -ret;
2242         return -1;
2243 }
2244
2245 int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
2246 {
2247         int ret;
2248
2249         if (aio) {
2250                 ret = __aio_rw(rw, fd, buf, len, offset);
2251         } else {
2252                 if (rw == READ)
2253                         ret = read(fd, buf, len);
2254                 else
2255                         ret = write(fd, buf, len);
2256         }
2257         return ret;
2258 }
2259
2260 #endif
2261
2262 #define test_fallocate(mode) __test_fallocate(mode, #mode)
2263
2264 int
2265 __test_fallocate(int mode, const char *mode_str)
2266 {
2267 #ifdef HAVE_LINUX_FALLOC_H
2268         int ret = 0;
2269         if (!lite) {
2270                 if (fallocate(fd, mode, file_size, 1) && errno == EOPNOTSUPP) {
2271                         if(!quiet)
2272                                 fprintf(stderr,
2273                                         "main: filesystem does not support "
2274                                         "fallocate mode %s, disabling!\n",
2275                                         mode_str);
2276                 } else {
2277                         ret = 1;
2278                         if (ftruncate(fd, file_size)) {
2279                                 warn("main: ftruncate");
2280                                 exit(132);
2281                         }
2282                 }
2283         }
2284         return ret;
2285 #endif
2286 }
2287
2288 static struct option longopts[] = {
2289         {"replay-ops", required_argument, 0, 256},
2290         {"record-ops", optional_argument, 0, 255},
2291         { }
2292 };
2293
2294 int
2295 main(int argc, char **argv)
2296 {
2297         int     i, style, ch;
2298         char    *endp, *tmp;
2299         char logfile[PATH_MAX];
2300         struct stat statbuf;
2301         int o_flags = O_RDWR|O_CREAT|O_TRUNC;
2302
2303         logfile[0] = 0;
2304         dname[0] = 0;
2305
2306         page_size = getpagesize();
2307         page_mask = page_size - 1;
2308         mmap_mask = page_mask;
2309
2310
2311         setvbuf(stdout, (char *)0, _IOLBF, 0); /* line buffered stdout */
2312
2313         while ((ch = getopt_long(argc, argv,
2314                                  "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:FJKHzCILN:OP:RS:WXZ",
2315                                  longopts, NULL)) != EOF)
2316                 switch (ch) {
2317                 case 'b':
2318                         simulatedopcount = getnum(optarg, &endp);
2319                         if (!quiet)
2320                                 prt("Will begin at operation %ld\n", simulatedopcount);
2321                         if (simulatedopcount == 0)
2322                                 usage();
2323                         simulatedopcount -= 1;
2324                         break;
2325                 case 'c':
2326                         closeprob = getnum(optarg, &endp);
2327                         if (!quiet)
2328                                 prt("Chance of close/open is 1 in %d\n", closeprob);
2329                         if (closeprob <= 0)
2330                                 usage();
2331                         break;
2332                 case 'd':
2333                         debug = 1;
2334                         break;
2335                 case 'f':
2336                         flush = 1;
2337                         break;
2338                 case 'g':
2339                         filldata = *optarg;
2340                         break;
2341                 case 'i':
2342                         integrity = 1;
2343                         logdev = strdup(optarg);
2344                         if (!logdev) {
2345                                 prterr("strdup");
2346                                 exit(101);
2347                         }
2348                         break;
2349                 case 'j':
2350                         logid = strdup(optarg);
2351                         if (!logid) {
2352                                 prterr("strdup");
2353                                 exit(101);
2354                         }
2355                         break;
2356                 case 'k':
2357                         o_flags &= ~O_TRUNC;
2358                         break;
2359                 case 'l':
2360                         maxfilelen = getnum(optarg, &endp);
2361                         if (maxfilelen <= 0)
2362                                 usage();
2363                         break;
2364                 case 'm':
2365                         monitorstart = getnum(optarg, &endp);
2366                         if (monitorstart < 0)
2367                                 usage();
2368                         if (!endp || *endp++ != ':')
2369                                 usage();
2370                         monitorend = getnum(endp, &endp);
2371                         if (monitorend < 0)
2372                                 usage();
2373                         if (monitorend == 0)
2374                                 monitorend = -1; /* aka infinity */
2375                         debug = 1;
2376                 case 'n':
2377                         sizechecks = 0;
2378                         break;
2379                 case 'o':
2380                         maxoplen = getnum(optarg, &endp);
2381                         if (maxoplen <= 0)
2382                                 usage();
2383                         break;
2384                 case 'p':
2385                         progressinterval = getnum(optarg, &endp);
2386                         if (progressinterval == 0)
2387                                 usage();
2388                         break;
2389                 case 'q':
2390                         quiet = 1;
2391                         break;
2392                 case 'r':
2393                         readbdy = getnum(optarg, &endp);
2394                         if (readbdy <= 0)
2395                                 usage();
2396                         break;
2397                 case 's':
2398                         style = getnum(optarg, &endp);
2399                         if (style < 0 || style > 1)
2400                                 usage();
2401                         break;
2402                 case 't':
2403                         truncbdy = getnum(optarg, &endp);
2404                         if (truncbdy <= 0)
2405                                 usage();
2406                         break;
2407                 case 'w':
2408                         writebdy = getnum(optarg, &endp);
2409                         if (writebdy <= 0)
2410                                 usage();
2411                         break;
2412                 case 'x':
2413                         prealloc = 1;
2414                         break;
2415                 case 'y':
2416                         do_fsync = 1;
2417                         break;
2418                 case 'A':
2419                         aio = 1;
2420                         break;
2421                 case 'D':
2422                         debugstart = getnum(optarg, &endp);
2423                         if (debugstart < 1)
2424                                 usage();
2425                         break;
2426                 case 'F':
2427                         fallocate_calls = 0;
2428                         break;
2429                 case 'K':
2430                         keep_size_calls = 0;
2431                         break;
2432                 case 'H':
2433                         punch_hole_calls = 0;
2434                         break;
2435                 case 'z':
2436                         zero_range_calls = 0;
2437                         break;
2438                 case 'C':
2439                         collapse_range_calls = 0;
2440                         break;
2441                 case 'I':
2442                         insert_range_calls = 0;
2443                         break;
2444                 case 'J':
2445                         clone_range_calls = 0;
2446                         break;
2447                 case 'B':
2448                         dedupe_range_calls = 0;
2449                         break;
2450                 case 'L':
2451                         lite = 1;
2452                         o_flags &= ~(O_CREAT|O_TRUNC);
2453                         break;
2454                 case 'N':
2455                         numops = getnum(optarg, &endp);
2456                         if (numops < 0)
2457                                 usage();
2458                         break;
2459                 case 'O':
2460                         randomoplen = 0;
2461                         break;
2462                 case 'P':
2463                         strncpy(dname, optarg, sizeof(dname));
2464                         strcat(dname, "/");
2465                         dirpath = strlen(dname);
2466                         break;
2467                 case 'R':
2468                         mapped_reads = 0;
2469                         break;
2470                 case 'S':
2471                         seed = getnum(optarg, &endp);
2472                         if (seed == 0) {
2473                                 seed = time(0) % 10000;
2474                                 seed += (int)getpid();
2475                         }
2476                         if (seed < 0)
2477                                 usage();
2478                         break;
2479                 case 'W':
2480                         mapped_writes = 0;
2481                         if (!quiet)
2482                                 prt("mapped writes DISABLED\n");
2483                         break;
2484                 case 'X':
2485                         check_file = 1;
2486                         break;
2487                 case 'Z':
2488                         o_direct = O_DIRECT;
2489                         o_flags |= O_DIRECT;
2490                         break;
2491                 case 255:  /* --record-ops */
2492                         if (optarg)
2493                                 strncpy(opsfile, optarg, sizeof(opsfile));
2494                         recordops = opsfile;
2495                         break;
2496                 case 256:  /* --replay-ops */
2497                         replayops = optarg;
2498                         break;
2499                 default:
2500                         usage();
2501                         /* NOTREACHED */
2502                 }
2503         argc -= optind;
2504         argv += optind;
2505         if (argc != 1)
2506                 usage();
2507
2508         if (integrity && !dirpath) {
2509                 fprintf(stderr, "option -i <logdev> requires -P <dirpath>\n");
2510                 usage();
2511         }
2512
2513         fname = argv[0];
2514         tmp = strdup(fname);
2515         if (!tmp) {
2516                 prterr("strdup");
2517                 exit(101);
2518         }
2519         bname = basename(tmp);
2520
2521         signal(SIGHUP,  cleanup);
2522         signal(SIGINT,  cleanup);
2523         signal(SIGPIPE, cleanup);
2524         signal(SIGALRM, cleanup);
2525         signal(SIGTERM, cleanup);
2526         signal(SIGXCPU, cleanup);
2527         signal(SIGXFSZ, cleanup);
2528         signal(SIGVTALRM,       cleanup);
2529         signal(SIGUSR1, cleanup);
2530         signal(SIGUSR2, cleanup);
2531
2532         if (!quiet && seed)
2533                 prt("Seed set to %d\n", seed);
2534         srandom(seed);
2535         fd = open(fname, o_flags, 0666);
2536         if (fd < 0) {
2537                 prterr(fname);
2538                 exit(91);
2539         }
2540         if (fstat(fd, &statbuf)) {
2541                 prterr("check_size: fstat");
2542                 exit(91);
2543         }
2544         block_size = statbuf.st_blksize;
2545 #ifdef XFS
2546         if (prealloc) {
2547                 xfs_flock64_t   resv = { 0 };
2548 #ifdef HAVE_XFS_PLATFORM_DEFS_H
2549                 if (!platform_test_xfs_fd(fd)) {
2550                         prterr(fname);
2551                         fprintf(stderr, "main: cannot prealloc, non XFS\n");
2552                         exit(96);
2553                 }
2554 #endif
2555                 resv.l_len = maxfilelen;
2556                 if ((xfsctl(fname, fd, XFS_IOC_RESVSP, &resv)) < 0) {
2557                         prterr(fname);
2558                         exit(97);
2559                 }
2560         }
2561 #endif
2562
2563         if (dirpath) {
2564                 snprintf(goodfile, sizeof(goodfile), "%s%s.fsxgood", dname, bname);
2565                 snprintf(logfile, sizeof(logfile), "%s%s.fsxlog", dname, bname);
2566                 if (!*opsfile)
2567                         snprintf(opsfile, sizeof(opsfile), "%s%s.fsxops", dname, bname);
2568         } else {
2569                 snprintf(goodfile, sizeof(goodfile), "%s.fsxgood", fname);
2570                 snprintf(logfile, sizeof(logfile), "%s.fsxlog", fname);
2571                 if (!*opsfile)
2572                         snprintf(opsfile, sizeof(opsfile), "%s.fsxops", fname);
2573         }
2574         fsxgoodfd = open(goodfile, O_RDWR|O_CREAT|O_TRUNC, 0666);
2575         if (fsxgoodfd < 0) {
2576                 prterr(goodfile);
2577                 exit(92);
2578         }
2579         fsxlogf = fopen(logfile, "w");
2580         if (fsxlogf == NULL) {
2581                 prterr(logfile);
2582                 exit(93);
2583         }
2584         unlink(opsfile);
2585
2586         if (replayops) {
2587                 replayopsf = fopen(replayops, "r");
2588                 if (!replayopsf) {
2589                         prterr(replayops);
2590                         exit(93);
2591                 }
2592         }
2593
2594 #ifdef AIO
2595         if (aio)
2596                 aio_setup();
2597 #endif
2598
2599         if (!(o_flags & O_TRUNC)) {
2600                 off_t ret;
2601                 file_size = maxfilelen = biggest = lseek(fd, (off_t)0, SEEK_END);
2602                 if (file_size == (off_t)-1) {
2603                         prterr(fname);
2604                         warn("main: lseek eof");
2605                         exit(94);
2606                 }
2607                 ret = lseek(fd, (off_t)0, SEEK_SET);
2608                 if (ret == (off_t)-1) {
2609                         prterr(fname);
2610                         warn("main: lseek 0");
2611                         exit(95);
2612                 }
2613         }
2614         original_buf = (char *) malloc(maxfilelen);
2615         for (i = 0; i < maxfilelen; i++)
2616                 original_buf[i] = random() % 256;
2617         good_buf = (char *) malloc(maxfilelen + writebdy);
2618         good_buf = round_ptr_up(good_buf, writebdy, 0);
2619         memset(good_buf, '\0', maxfilelen);
2620         temp_buf = (char *) malloc(maxoplen + readbdy);
2621         temp_buf = round_ptr_up(temp_buf, readbdy, 0);
2622         memset(temp_buf, '\0', maxoplen);
2623         if (lite) {     /* zero entire existing file */
2624                 ssize_t written;
2625
2626                 written = write(fd, good_buf, (size_t)maxfilelen);
2627                 if (written != maxfilelen) {
2628                         if (written == -1) {
2629                                 prterr(fname);
2630                                 warn("main: error on write");
2631                         } else
2632                                 warn("main: short write, 0x%x bytes instead "
2633                                         "of 0x%lx\n",
2634                                         (unsigned)written,
2635                                         maxfilelen);
2636                         exit(98);
2637                 }
2638         } else {
2639                 ssize_t ret, len = file_size;
2640                 off_t off = 0;
2641
2642                 while (len > 0) {
2643                         ret = read(fd, good_buf + off, len);
2644                         if (ret == -1) {
2645                                 prterr(fname);
2646                                 warn("main: error on read");
2647                                 exit(98);
2648                         }
2649                         len -= ret;
2650                         off += ret;
2651                 }
2652
2653                 check_trunc_hack();
2654         }
2655
2656         if (fallocate_calls)
2657                 fallocate_calls = test_fallocate(0);
2658         if (keep_size_calls)
2659                 keep_size_calls = test_fallocate(FALLOC_FL_KEEP_SIZE);
2660         if (punch_hole_calls)
2661                 punch_hole_calls = test_fallocate(FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE);
2662         if (zero_range_calls)
2663                 zero_range_calls = test_fallocate(FALLOC_FL_ZERO_RANGE);
2664         if (collapse_range_calls)
2665                 collapse_range_calls = test_fallocate(FALLOC_FL_COLLAPSE_RANGE);
2666         if (insert_range_calls)
2667                 insert_range_calls = test_fallocate(FALLOC_FL_INSERT_RANGE);
2668         if (clone_range_calls)
2669                 clone_range_calls = test_clone_range();
2670         if (dedupe_range_calls)
2671                 dedupe_range_calls = test_dedupe_range();
2672
2673         while (numops == -1 || numops--)
2674                 if (!test())
2675                         break;
2676
2677         free(tmp);
2678         if (close(fd)) {
2679                 prterr("close");
2680                 report_failure(99);
2681         }
2682         prt("All %lu operations completed A-OK!\n", testcalls);
2683         if (recordops)
2684                 logdump();
2685
2686         exit(0);
2687         return 0;
2688 }