git.ozlabs.org Git - ccan/blob - ccan/tdb/tools/replay_trace.c

   1 #include <ccan/tdb/tdb.h>
   2 #include <ccan/grab_file/grab_file.h>
   3 #include <ccan/hash/hash.h>
   4 #include <ccan/talloc/talloc.h>
   5 #include <ccan/str_talloc/str_talloc.h>
   6 #include <ccan/str/str.h>
   7 #include <ccan/list/list.h>
   8 #include <err.h>
   9 #include <ctype.h>
  10 #include <string.h>
  11 #include <unistd.h>
  12 #include <sys/types.h>
  13 #include <sys/wait.h>
  14 #include <sys/time.h>
  15 #include <errno.h>
  16 #include <signal.h>
  17 #include <assert.h>
  18
  19 #define STRINGIFY2(x) #x
  20 #define STRINGIFY(x) STRINGIFY2(x)
  21
  22 /* Avoid mod by zero */
  23 static unsigned int total_keys = 1;
  24
  25 #define DEBUG_DEPS 1
  26
  27 /* Traversals block transactions in the current implementation. */
  28 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
  29
  30 struct pipe {
  31         int fd[2];
  32 };
  33 static struct pipe *pipes;
  34
  35 static void __attribute__((noreturn)) fail(const char *filename,
  36                                            unsigned int line,
  37                                            const char *fmt, ...)
  38 {
  39         va_list ap;
  40
  41         va_start(ap, fmt);
  42         fprintf(stderr, "%s:%u: FAIL: ", filename, line);
  43         vfprintf(stderr, fmt, ap);
  44         fprintf(stderr, "\n");
  45         va_end(ap);
  46         exit(1);
  47 }
  48
  49 /* Try or die. */
  50 #define try(expr, expect)                                               \
  51         do {                                                            \
  52                 int ret = (expr);                                       \
  53                 if (ret != (expect))                                    \
  54                         fail(filename[file], i+1,                       \
  55                              STRINGIFY(expr) "= %i", ret);              \
  56         } while (0)
  57
  58 /* Try or imitate results. */
  59 #define unreliable(expr, expect, force, undo)                           \
  60         do {                                                            \
  61                 int ret = expr;                                         \
  62                 if (ret != expect) {                                    \
  63                         fprintf(stderr, "%s:%u: %s gave %i not %i",     \
  64                                 filename[file], i+1, STRINGIFY(expr),   \
  65                                 ret, expect);                           \
  66                         if (expect == 0)                                \
  67                                 force;                                  \
  68                         else                                            \
  69                                 undo;                                   \
  70                 }                                                       \
  71         } while (0)
  72
  73 static bool key_eq(TDB_DATA a, TDB_DATA b)
  74 {
  75         if (a.dsize != b.dsize)
  76                 return false;
  77         return memcmp(a.dptr, b.dptr, a.dsize) == 0;
  78 }
  79
  80 /* This is based on the hash algorithm from gdbm */
  81 static unsigned int hash_key(TDB_DATA *key)
  82 {
  83         uint32_t value; /* Used to compute the hash value.  */
  84         uint32_t   i;   /* Used to cycle through random values. */
  85
  86         /* Set the initial value from the key size. */
  87         for (value = 0x238F13AF ^ key->dsize, i=0; i < key->dsize; i++)
  88                 value = (value + (key->dptr[i] << (i*5 % 24)));
  89
  90         return (1103515243 * value + 12345);
  91 }
  92
  93 enum op_type {
  94         OP_TDB_LOCKALL,
  95         OP_TDB_LOCKALL_MARK,
  96         OP_TDB_LOCKALL_UNMARK,
  97         OP_TDB_LOCKALL_NONBLOCK,
  98         OP_TDB_UNLOCKALL,
  99         OP_TDB_LOCKALL_READ,
 100         OP_TDB_LOCKALL_READ_NONBLOCK,
 101         OP_TDB_UNLOCKALL_READ,
 102         OP_TDB_CHAINLOCK,
 103         OP_TDB_CHAINLOCK_NONBLOCK,
 104         OP_TDB_CHAINLOCK_MARK,
 105         OP_TDB_CHAINLOCK_UNMARK,
 106         OP_TDB_CHAINUNLOCK,
 107         OP_TDB_CHAINLOCK_READ,
 108         OP_TDB_CHAINUNLOCK_READ,
 109         OP_TDB_PARSE_RECORD,
 110         OP_TDB_EXISTS,
 111         OP_TDB_STORE,
 112         OP_TDB_APPEND,
 113         OP_TDB_GET_SEQNUM,
 114         OP_TDB_WIPE_ALL,
 115         OP_TDB_TRANSACTION_START,
 116         OP_TDB_TRANSACTION_CANCEL,
 117         OP_TDB_TRANSACTION_COMMIT,
 118         OP_TDB_TRAVERSE_READ_START,
 119         OP_TDB_TRAVERSE_START,
 120         OP_TDB_TRAVERSE_END,
 121         OP_TDB_TRAVERSE,
 122         OP_TDB_FIRSTKEY,
 123         OP_TDB_NEXTKEY,
 124         OP_TDB_FETCH,
 125         OP_TDB_DELETE,
 126 };
 127
 128 struct op {
 129         unsigned int serial;
 130         enum op_type op;
 131         TDB_DATA key;
 132         TDB_DATA data;
 133         int ret;
 134
 135         /* Who is waiting for us? */
 136         struct list_head post;
 137         /* What are we waiting for? */
 138         struct list_head pre;
 139
 140         /* If I'm part of a group (traverse/transaction) where is
 141          * start?  (Otherwise, 0) */
 142         unsigned int group_start;
 143
 144         union {
 145                 int flag; /* open and store */
 146                 struct {  /* append */
 147                         TDB_DATA pre;
 148                         TDB_DATA post;
 149                 } append;
 150                 unsigned int group_len; /* transaction/traverse start */
 151         };
 152 };
 153
 154 static unsigned char hex_char(const char *filename, unsigned int line, char c)
 155 {
 156         c = toupper(c);
 157         if (c >= 'A' && c <= 'F')
 158                 return c - 'A' + 10;
 159         if (c >= '0' && c <= '9')
 160                 return c - '0';
 161         fail(filename, line, "invalid hex character '%c'", c);
 162 }
 163
 164 /* TDB data is <size>:<%02x>* */
 165 static TDB_DATA make_tdb_data(const void *ctx,
 166                               const char *filename, unsigned int line,
 167                               const char *word)
 168 {
 169         TDB_DATA data;
 170         unsigned int i;
 171         const char *p;
 172
 173         if (streq(word, "NULL"))
 174                 return tdb_null;
 175
 176         data.dsize = atoi(word);
 177         data.dptr = talloc_array(ctx, unsigned char, data.dsize);
 178         p = strchr(word, ':');
 179         if (!p)
 180                 fail(filename, line, "invalid tdb data '%s'", word);
 181         p++;
 182         for (i = 0; i < data.dsize; i++)
 183                 data.dptr[i] = hex_char(filename, line, p[i*2])*16
 184                         + hex_char(filename, line, p[i*2+1]);
 185
 186         return data;
 187 }
 188
 189 static void add_op(const char *filename, struct op **op, unsigned int i,
 190                    unsigned int serial, enum op_type type)
 191 {
 192         struct op *new;
 193         *op = talloc_realloc(NULL, *op, struct op, i+1);
 194         new = (*op) + i;
 195         new->op = type;
 196         new->serial = serial;
 197         new->ret = 0;
 198         new->group_start = 0;
 199 }
 200
 201 static void op_add_nothing(const char *filename,
 202                            struct op op[], unsigned int op_num, char *words[])
 203 {
 204         if (words[2])
 205                 fail(filename, op_num+1, "Expected no arguments");
 206         op[op_num].key = tdb_null;
 207 }
 208
 209 static void op_add_key(const char *filename,
 210                        struct op op[], unsigned int op_num, char *words[])
 211 {
 212         if (words[2] == NULL || words[3])
 213                 fail(filename, op_num+1, "Expected just a key");
 214
 215         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 216         if (op[op_num].op != OP_TDB_TRAVERSE)
 217                 total_keys++;
 218 }
 219
 220 static void op_add_key_ret(const char *filename,
 221                            struct op op[], unsigned int op_num, char *words[])
 222 {
 223         if (!words[2] || !words[3] || !words[4] || words[5]
 224             || !streq(words[3], "="))
 225                 fail(filename, op_num+1, "Expected <key> = <ret>");
 226         op[op_num].ret = atoi(words[4]);
 227         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 228         /* May only be a unique key if it fails */
 229         if (op[op_num].ret != 0)
 230                 total_keys++;
 231 }
 232
 233 static void op_add_key_data(const char *filename,
 234                             struct op op[], unsigned int op_num, char *words[])
 235 {
 236         if (!words[2] || !words[3] || !words[4] || words[5]
 237             || !streq(words[3], "="))
 238                 fail(filename, op_num+1, "Expected <key> = <data>");
 239         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 240         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[4]);
 241         /* May only be a unique key if it fails */
 242         if (!op[op_num].data.dptr)
 243                 total_keys++;
 244 }
 245
 246 /* <serial> tdb_store <rec> <rec> <flag> = <ret> */
 247 static void op_add_store(const char *filename,
 248                          struct op op[], unsigned int op_num, char *words[])
 249 {
 250         if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
 251             || words[7] || !streq(words[5], "="))
 252                 fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
 253
 254         op[op_num].flag = strtoul(words[4], NULL, 0);
 255         op[op_num].ret = atoi(words[6]);
 256         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 257         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 258         total_keys++;
 259 }
 260
 261 /* <serial> tdb_append <rec> <rec> = <rec> */
 262 static void op_add_append(const char *filename,
 263                           struct op op[], unsigned int op_num, char *words[])
 264 {
 265         if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
 266             || !streq(words[4], "="))
 267                 fail(filename, op_num+1, "Expect <key> <data> = <rec>");
 268
 269         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 270         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 271
 272         op[op_num].append.post
 273                 = make_tdb_data(op, filename, op_num+1, words[5]);
 274
 275         /* By subtraction, figure out what previous data was. */
 276         op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
 277         op[op_num].append.pre.dsize
 278                 = op[op_num].append.post.dsize - op[op_num].data.dsize;
 279         total_keys++;
 280 }
 281
 282 /* <serial> tdb_get_seqnum = <ret> */
 283 static void op_add_seqnum(const char *filename,
 284                           struct op op[], unsigned int op_num, char *words[])
 285 {
 286         if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
 287                 fail(filename, op_num+1, "Expect = <ret>");
 288
 289         op[op_num].key = tdb_null;
 290         op[op_num].ret = atoi(words[3]);
 291 }
 292
 293 static void op_add_traverse(const char *filename,
 294                             struct op op[], unsigned int op_num, char *words[])
 295 {
 296         if (words[2])
 297                 fail(filename, op_num+1, "Expect no arguments");
 298
 299         op[op_num].key = tdb_null;
 300         op[op_num].group_len = 0;
 301 }
 302
 303 static void op_add_transaction(const char *filename, struct op op[],
 304                                unsigned int op_num, char *words[])
 305 {
 306         if (words[2])
 307                 fail(filename, op_num+1, "Expect no arguments");
 308
 309         op[op_num].key = tdb_null;
 310         op[op_num].group_len = 0;
 311 }
 312
 313 static void op_analyze_transaction(const char *filename,
 314                                    struct op op[], unsigned int op_num,
 315                                    char *words[])
 316 {
 317         int i, start;
 318
 319         op[op_num].key = tdb_null;
 320
 321         if (words[2])
 322                 fail(filename, op_num+1, "Expect no arguments");
 323
 324         for (i = op_num-1; i >= 0; i--) {
 325                 if (op[i].op == OP_TDB_TRANSACTION_START && !op[i].group_len)
 326                         break;
 327         }
 328
 329         if (i < 0)
 330                 fail(filename, op_num+1, "no transaction start found");
 331
 332         start = i;
 333         op[start].group_len = op_num - i;
 334
 335         /* This rolls in nested transactions.  I think that's right. */
 336         for (i++; i <= op_num; i++)
 337                 op[i].group_start = start;
 338 }
 339
 340 struct traverse_hash {
 341         TDB_DATA key;
 342         unsigned int index;
 343 };
 344
 345 static void op_analyze_traverse(const char *filename,
 346                                 struct op op[], unsigned int op_num,
 347                                 char *words[])
 348 {
 349         int i, start;
 350
 351         op[op_num].key = tdb_null;
 352
 353         /* = %u means traverse function terminated. */
 354         if (words[2]) {
 355                 if (!streq(words[2], "=") || !words[3] || words[4])
 356                         fail(filename, op_num+1, "expect = <num>");
 357                 op[op_num].ret = atoi(words[3]);
 358         } else
 359                 op[op_num].ret = 0;
 360
 361         for (i = op_num-1; i >= 0; i--) {
 362                 if (op[i].op != OP_TDB_TRAVERSE_READ_START
 363                     && op[i].op != OP_TDB_TRAVERSE_START)
 364                         continue;
 365                 if (op[i].group_len)
 366                         continue;
 367                 break;
 368         }
 369
 370         if (i < 0)
 371                 fail(filename, op_num+1, "no traversal start found");
 372
 373         start = i;
 374         op[start].group_len = op_num - start;
 375
 376         for (i = start; i <= op_num; i++)
 377                 op[i].group_start = start;
 378 }
 379
 380 /* Keep -Wmissing-declarations happy: */
 381 const struct op_table *
 382 find_keyword (register const char *str, register unsigned int len);
 383
 384 #include "keywords.c"
 385
 386 struct depend {
 387         /* We can have more than one */
 388         struct list_node pre_list;
 389         struct list_node post_list;
 390         unsigned int needs_file;
 391         unsigned int needs_opnum;
 392         unsigned int satisfies_file;
 393         unsigned int satisfies_opnum;
 394 };
 395
 396 static void check_deps(const char *filename, struct op op[], unsigned int num)
 397 {
 398 #ifdef DEBUG_DEPS
 399         unsigned int i;
 400
 401         for (i = 1; i < num; i++)
 402                 if (!list_empty(&op[i].pre))
 403                         fail(filename, i+1, "Still has dependencies");
 404 #endif
 405 }
 406
 407 static void dump_pre(char *filename[], struct op *op[],
 408                      unsigned int file, unsigned int i)
 409 {
 410         struct depend *dep;
 411
 412         printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
 413                 op[file][i].serial);
 414         list_for_each(&op[file][i].pre, dep, pre_list)
 415                 printf("    %s:%u (%u)\n",
 416                        filename[dep->satisfies_file], dep->satisfies_opnum+1,
 417                        op[dep->satisfies_file][dep->satisfies_opnum].serial);
 418         check_deps(filename[file], op[file], i);
 419 }
 420
 421 /* We simply read/write pointers, since we all are children. */
 422 static void do_pre(struct tdb_context *tdb,
 423                    char *filename[], struct op *op[],
 424                    unsigned int file, int pre_fd, unsigned int i)
 425 {
 426         while (!list_empty(&op[file][i].pre)) {
 427                 struct depend *dep;
 428
 429 #if DEBUG_DEPS
 430                 printf("%s:%u:waiting for pre\n", filename[file], i+1);
 431                 fflush(stdout);
 432 #endif
 433                 alarm(10);
 434                 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
 435                         if (errno == EINTR) {
 436                                 dump_pre(filename, op, file, i);
 437                                 exit(1);
 438                         } else
 439                                 errx(1, "Reading from pipe");
 440                 }
 441                 alarm(0);
 442
 443 #if DEBUG_DEPS
 444                 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
 445                        dep->needs_opnum+1, filename[dep->satisfies_file],
 446                        dep->satisfies_opnum+1);
 447                 fflush(stdout);
 448 #endif
 449                 /* This could be any op, not just this one. */
 450                 talloc_free(dep);
 451         }
 452 }
 453
 454 static void do_post(char *filename[], struct op *op[],
 455                     unsigned int file, unsigned int i)
 456 {
 457         struct depend *dep;
 458
 459         list_for_each(&op[file][i].post, dep, post_list) {
 460 #if DEBUG_DEPS
 461                 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
 462                        filename[dep->needs_file], dep->needs_opnum+1);
 463 #endif
 464                 if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
 465                     != sizeof(dep))
 466                         err(1, "%s:%u failed to tell file %s",
 467                             filename[file], i+1, filename[dep->needs_file]);
 468         }
 469 }
 470
 471 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 472 {
 473         return data.dsize;
 474 }
 475
 476 static unsigned run_ops(struct tdb_context *tdb,
 477                         int pre_fd,
 478                         char *filename[],
 479                         struct op *op[],
 480                         unsigned int file,
 481                         unsigned int start, unsigned int stop);
 482
 483 struct traverse_info {
 484         struct op **op;
 485         char **filename;
 486         unsigned file;
 487         int pre_fd;
 488         unsigned int start;
 489         unsigned int i;
 490 };
 491
 492 /* More complex.  Just do whatever's they did at the n'th entry. */
 493 static int nontrivial_traverse(struct tdb_context *tdb,
 494                                TDB_DATA key, TDB_DATA data,
 495                                void *_tinfo)
 496 {
 497         struct traverse_info *tinfo = _tinfo;
 498         unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
 499
 500         if (tinfo->i == tinfo->start + trav_len) {
 501                 /* This can happen if traverse expects to be empty. */
 502                 if (trav_len == 1)
 503                         return 1;
 504                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 505                      "traverse did not terminate");
 506         }
 507
 508         if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
 509                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 510                      "%s:%u:traverse terminated early");
 511
 512         /* Run any normal ops. */
 513         tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
 514                            tinfo->file, tinfo->i+1, tinfo->start + trav_len);
 515
 516         if (tinfo->i == tinfo->start + trav_len)
 517                 return 1;
 518
 519         return 0;
 520 }
 521
 522 static unsigned op_traverse(struct tdb_context *tdb,
 523                             int pre_fd,
 524                             char *filename[],
 525                             unsigned int file,
 526                             int (*traversefn)(struct tdb_context *,
 527                                               tdb_traverse_func, void *),
 528                             struct op *op[],
 529                             unsigned int start)
 530 {
 531         struct traverse_info tinfo = { op, filename, file, pre_fd,
 532                                        start, start+1 };
 533
 534         traversefn(tdb, nontrivial_traverse, &tinfo);
 535
 536         /* Traversing in wrong order can have strange effects: eg. if
 537          * original traverse went A (delete A), B, we might do B
 538          * (delete A).  So if we have ops left over, we do it now. */
 539         while (tinfo.i != start + op[file][start].group_len) {
 540                 if (op[file][tinfo.i].op == OP_TDB_TRAVERSE)
 541                         tinfo.i++;
 542                 else
 543                         tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
 544                                           tinfo.i,
 545                                           start + op[file][start].group_len);
 546         }
 547
 548         return tinfo.i;
 549 }
 550
 551 static void break_out(int sig)
 552 {
 553 }
 554
 555 static __attribute__((noinline))
 556 unsigned run_ops(struct tdb_context *tdb,
 557                  int pre_fd,
 558                  char *filename[],
 559                  struct op *op[],
 560                  unsigned int file,
 561                  unsigned int start, unsigned int stop)
 562 {
 563         unsigned int i;
 564         struct sigaction sa;
 565
 566         sa.sa_handler = break_out;
 567         sa.sa_flags = 0;
 568
 569         sigaction(SIGALRM, &sa, NULL);
 570         for (i = start; i < stop; i++) {
 571                 do_pre(tdb, filename, op, file, pre_fd, i);
 572
 573                 switch (op[file][i].op) {
 574                 case OP_TDB_LOCKALL:
 575                         try(tdb_lockall(tdb), op[file][i].ret);
 576                         break;
 577                 case OP_TDB_LOCKALL_MARK:
 578                         try(tdb_lockall_mark(tdb), op[file][i].ret);
 579                         break;
 580                 case OP_TDB_LOCKALL_UNMARK:
 581                         try(tdb_lockall_unmark(tdb), op[file][i].ret);
 582                         break;
 583                 case OP_TDB_LOCKALL_NONBLOCK:
 584                         unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
 585                                    tdb_lockall(tdb), tdb_unlockall(tdb));
 586                         break;
 587                 case OP_TDB_UNLOCKALL:
 588                         try(tdb_unlockall(tdb), op[file][i].ret);
 589                         break;
 590                 case OP_TDB_LOCKALL_READ:
 591                         try(tdb_lockall_read(tdb), op[file][i].ret);
 592                         break;
 593                 case OP_TDB_LOCKALL_READ_NONBLOCK:
 594                         unreliable(tdb_lockall_read_nonblock(tdb),
 595                                    op[file][i].ret,
 596                                    tdb_lockall_read(tdb),
 597                                    tdb_unlockall_read(tdb));
 598                         break;
 599                 case OP_TDB_UNLOCKALL_READ:
 600                         try(tdb_unlockall_read(tdb), op[file][i].ret);
 601                         break;
 602                 case OP_TDB_CHAINLOCK:
 603                         try(tdb_chainlock(tdb, op[file][i].key),
 604                             op[file][i].ret);
 605                         break;
 606                 case OP_TDB_CHAINLOCK_NONBLOCK:
 607                         unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
 608                                    op[file][i].ret,
 609                                    tdb_chainlock(tdb, op[file][i].key),
 610                                    tdb_chainunlock(tdb, op[file][i].key));
 611                         break;
 612                 case OP_TDB_CHAINLOCK_MARK:
 613                         try(tdb_chainlock_mark(tdb, op[file][i].key),
 614                             op[file][i].ret);
 615                         break;
 616                 case OP_TDB_CHAINLOCK_UNMARK:
 617                         try(tdb_chainlock_unmark(tdb, op[file][i].key),
 618                             op[file][i].ret);
 619                         break;
 620                 case OP_TDB_CHAINUNLOCK:
 621                         try(tdb_chainunlock(tdb, op[file][i].key),
 622                             op[file][i].ret);
 623                         break;
 624                 case OP_TDB_CHAINLOCK_READ:
 625                         try(tdb_chainlock_read(tdb, op[file][i].key),
 626                             op[file][i].ret);
 627                         break;
 628                 case OP_TDB_CHAINUNLOCK_READ:
 629                         try(tdb_chainunlock_read(tdb, op[file][i].key),
 630                             op[file][i].ret);
 631                         break;
 632                 case OP_TDB_PARSE_RECORD:
 633                         try(tdb_parse_record(tdb, op[file][i].key, get_len,
 634                                              NULL),
 635                             op[file][i].ret);
 636                         break;
 637                 case OP_TDB_EXISTS:
 638                         try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
 639                         break;
 640                 case OP_TDB_STORE:
 641                         try(tdb_store(tdb, op[file][i].key, op[file][i].data,
 642                                       op[file][i].flag),
 643                             op[file][i].ret);
 644                         break;
 645                 case OP_TDB_APPEND:
 646                         try(tdb_append(tdb, op[file][i].key, op[file][i].data),
 647                             op[file][i].ret);
 648                         break;
 649                 case OP_TDB_GET_SEQNUM:
 650                         try(tdb_get_seqnum(tdb), op[file][i].ret);
 651                         break;
 652                 case OP_TDB_WIPE_ALL:
 653                         try(tdb_wipe_all(tdb), op[file][i].ret);
 654                         break;
 655                 case OP_TDB_TRANSACTION_START:
 656                         try(tdb_transaction_start(tdb), op[file][i].ret);
 657                         break;
 658                 case OP_TDB_TRANSACTION_CANCEL:
 659                         try(tdb_transaction_cancel(tdb), op[file][i].ret);
 660                         break;
 661                 case OP_TDB_TRANSACTION_COMMIT:
 662                         try(tdb_transaction_commit(tdb), op[file][i].ret);
 663                         break;
 664                 case OP_TDB_TRAVERSE_READ_START:
 665                         i = op_traverse(tdb, pre_fd, filename, file,
 666                                         tdb_traverse_read, op, i);
 667                         break;
 668                 case OP_TDB_TRAVERSE_START:
 669                         i = op_traverse(tdb, pre_fd, filename, file,
 670                                         tdb_traverse, op, i);
 671                         break;
 672                 case OP_TDB_TRAVERSE:
 673                         /* Terminate: we're in a traverse, and we've
 674                          * done our ops. */
 675                         return i;
 676                 case OP_TDB_TRAVERSE_END:
 677                         fail(filename[file], i+1, "unexpected end traverse");
 678                 /* FIXME: These must be treated like traverse. */
 679                 case OP_TDB_FIRSTKEY:
 680                         if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
 681                                 fail(filename[file], i+1, "bad firstkey");
 682                         break;
 683                 case OP_TDB_NEXTKEY:
 684                         if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
 685                                     op[file][i].data))
 686                                 fail(filename[file], i+1, "bad nextkey");
 687                         break;
 688                 case OP_TDB_FETCH: {
 689                         TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
 690                         if (!key_eq(f, op[file][i].data))
 691                                 fail(filename[file], i+1, "bad fetch %u",
 692                                      f.dsize);
 693                         break;
 694                 }
 695                 case OP_TDB_DELETE:
 696                         try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
 697                         break;
 698                 }
 699                 do_post(filename, op, file, i);
 700         }
 701         return i;
 702 }
 703
 704 static struct op *load_tracefile(const char *filename, unsigned int *num,
 705                                  unsigned int *hashsize,
 706                                  unsigned int *tdb_flags,
 707                                  unsigned int *open_flags)
 708 {
 709         unsigned int i;
 710         struct op *op = talloc_array(NULL, struct op, 1);
 711         char **words;
 712         char **lines;
 713         char *file;
 714
 715         file = grab_file(NULL, filename, NULL);
 716         if (!file)
 717                 err(1, "Reading %s", filename);
 718
 719         lines = strsplit(file, file, "\n", NULL);
 720         if (!lines[0])
 721                 errx(1, "%s is empty", filename);
 722
 723         words = strsplit(lines, lines[0], " ", NULL);
 724         if (!streq(words[1], "tdb_open"))
 725                 fail(filename, 1, "does not start with tdb_open");
 726
 727         *hashsize = atoi(words[2]);
 728         *tdb_flags = strtoul(words[3], NULL, 0);
 729         *open_flags = strtoul(words[4], NULL, 0);
 730
 731         for (i = 1; lines[i]; i++) {
 732                 const struct op_table *opt;
 733
 734                 words = strsplit(lines, lines[i], " ", NULL);
 735                 if (!words[0] || !words[1])
 736                         fail(filename, i+1, "Expected serial number and op");
 737
 738                 opt = find_keyword(words[1], strlen(words[1]));
 739                 if (!opt) {
 740                         if (streq(words[1], "tdb_close")) {
 741                                 if (lines[i+1])
 742                                         fail(filename, i+2,
 743                                              "lines after tdb_close");
 744                                 *num = i;
 745                                 talloc_free(lines);
 746                                 return op;
 747                         }
 748                         fail(filename, i+1, "Unknown operation '%s'", words[1]);
 749                 }
 750
 751                 add_op(filename, &op, i, atoi(words[0]), opt->type);
 752                 opt->enhance_op(filename, op, i, words);
 753         }
 754
 755         fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
 756               filename, i);
 757         talloc_free(lines);
 758         *num = i - 1;
 759         return op;
 760 }
 761
 762 /* We remember all the keys we've ever seen, and who has them. */
 763 struct key_user {
 764         unsigned int file;
 765         unsigned int op_num;
 766 };
 767
 768 struct keyinfo {
 769         TDB_DATA key;
 770         unsigned int num_users;
 771         struct key_user *user;
 772 };
 773
 774 static const TDB_DATA must_not_exist;
 775 static const TDB_DATA must_exist;
 776 static const TDB_DATA not_exists_or_empty;
 777
 778 /* NULL means doesn't care if it exists or not, &must_exist means
 779  * it must exist but we don't care what, &must_not_exist means it must
 780  * not exist, otherwise the data it needs. */
 781 static const TDB_DATA *needs(const struct op *op)
 782 {
 783         switch (op->op) {
 784         /* FIXME: Pull forward deps, since we can deadlock */
 785         case OP_TDB_CHAINLOCK:
 786         case OP_TDB_CHAINLOCK_NONBLOCK:
 787         case OP_TDB_CHAINLOCK_MARK:
 788         case OP_TDB_CHAINLOCK_UNMARK:
 789         case OP_TDB_CHAINUNLOCK:
 790         case OP_TDB_CHAINLOCK_READ:
 791         case OP_TDB_CHAINUNLOCK_READ:
 792                 return NULL;
 793
 794         case OP_TDB_APPEND:
 795                 if (op->append.pre.dsize == 0)
 796                         return &not_exists_or_empty;
 797                 return &op->append.pre;
 798
 799         case OP_TDB_STORE:
 800                 if (op->flag == TDB_INSERT) {
 801                         if (op->ret < 0)
 802                                 return &must_exist;
 803                         else
 804                                 return &must_not_exist;
 805                 } else if (op->flag == TDB_MODIFY) {
 806                         if (op->ret < 0)
 807                                 return &must_not_exist;
 808                         else
 809                                 return &must_exist;
 810                 }
 811                 /* No flags?  Don't care */
 812                 return NULL;
 813
 814         case OP_TDB_EXISTS:
 815                 if (op->ret == 1)
 816                         return &must_exist;
 817                 else
 818                         return &must_not_exist;
 819
 820         case OP_TDB_PARSE_RECORD:
 821                 if (op->ret < 0)
 822                         return &must_not_exist;
 823                 return &must_exist;
 824
 825         /* FIXME: handle these. */
 826         case OP_TDB_WIPE_ALL:
 827         case OP_TDB_FIRSTKEY:
 828         case OP_TDB_NEXTKEY:
 829         case OP_TDB_GET_SEQNUM:
 830         case OP_TDB_TRAVERSE:
 831         case OP_TDB_TRANSACTION_COMMIT:
 832         case OP_TDB_TRANSACTION_CANCEL:
 833         case OP_TDB_TRANSACTION_START:
 834                 return NULL;
 835
 836         case OP_TDB_FETCH:
 837                 if (!op->data.dptr)
 838                         return &must_not_exist;
 839                 return &op->data;
 840
 841         case OP_TDB_DELETE:
 842                 if (op->ret < 0)
 843                         return &must_not_exist;
 844                 return &must_exist;
 845
 846         default:
 847                 errx(1, "Unexpected op %i", op->op);
 848         }
 849
 850 }
 851
 852 /* What's the data after this op?  pre if nothing changed. */
 853 static const TDB_DATA *gives(const struct op *op, const TDB_DATA *pre)
 854 {
 855         /* Failed ops don't change state of db. */
 856         if (op->ret < 0)
 857                 return pre;
 858
 859         if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
 860                 return &tdb_null;
 861
 862         if (op->op == OP_TDB_APPEND)
 863                 return &op->append.post;
 864
 865         if (op->op == OP_TDB_STORE)
 866                 return &op->data;
 867
 868         return pre;
 869 }
 870
 871 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 872                                 unsigned int num)
 873 {
 874         unsigned int i, j, h;
 875         struct keyinfo *hash;
 876
 877         hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
 878         for (i = 0; i < num; i++) {
 879                 for (j = 1; j < num_ops[i]; j++) {
 880                         /* We can't do this on allocation, due to realloc. */
 881                         list_head_init(&op[i][j].post);
 882                         list_head_init(&op[i][j].pre);
 883
 884                         if (!op[i][j].key.dptr)
 885                                 continue;
 886
 887                         /* We don't wait for traverse keys */
 888                         /* FIXME: We should, for trivial traversals. */
 889                         if (op[i][j].op == OP_TDB_TRAVERSE)
 890                                 continue;
 891
 892                         h = hash_key(&op[i][j].key) % (total_keys * 2);
 893                         while (!key_eq(hash[h].key, op[i][j].key)) {
 894                                 if (!hash[h].key.dptr) {
 895                                         hash[h].key = op[i][j].key;
 896                                         break;
 897                                 }
 898                                 h = (h + 1) % (total_keys * 2);
 899                         }
 900                         /* Might as well save some memory if we can. */
 901                         if (op[i][j].key.dptr != hash[h].key.dptr) {
 902                                 talloc_free(op[i][j].key.dptr);
 903                                 op[i][j].key.dptr = hash[h].key.dptr;
 904                         }
 905                         hash[h].user = talloc_realloc(hash, hash[h].user,
 906                                                      struct key_user,
 907                                                      hash[h].num_users+1);
 908                         hash[h].user[hash[h].num_users].op_num = j;
 909                         hash[h].user[hash[h].num_users].file = i;
 910                         hash[h].num_users++;
 911                 }
 912         }
 913
 914         return hash;
 915 }
 916
 917 static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
 918 {
 919         /* Don't need anything?  Cool. */
 920         if (!need)
 921                 return true;
 922
 923         /* This should be tdb_null or a real value. */
 924         assert(data != &must_exist);
 925         assert(data != &must_not_exist);
 926         assert(data != &not_exists_or_empty);
 927
 928         /* must_not_exist == must_not_exist, must_exist == must_exist, or
 929            not_exists_or_empty == not_exists_or_empty. */
 930         if (data->dsize == need->dsize && data->dptr == need->dptr)
 931                 return true;
 932
 933         /* Must not exist?  data must not exist. */
 934         if (need == &must_not_exist)
 935                 return data->dptr == NULL;
 936
 937         /* Must exist? */
 938         if (need == &must_exist)
 939                 return data->dptr != NULL;
 940
 941         /* Either noexist or empty. */
 942         if (need == &not_exists_or_empty)
 943                 return data->dsize == 0;
 944
 945         /* Needs something specific. */
 946         return key_eq(*data, *need);
 947 }
 948
 949 static void move_to_front(struct key_user res[], unsigned int elem)
 950 {
 951         if (elem != 0) {
 952                 struct key_user tmp = res[elem];
 953                 memmove(res + 1, res, elem*sizeof(res[0]));
 954                 res[0] = tmp;
 955         }
 956 }
 957
 958 static void restore_to_pos(struct key_user res[], unsigned int elem)
 959 {
 960         if (elem != 0) {
 961                 struct key_user tmp = res[0];
 962                 memmove(res, res + 1, elem*sizeof(res[0]));
 963                 res[elem] = tmp;
 964         }
 965 }
 966
 967 static bool sort_deps(char *filename[], struct op *op[],
 968                       struct key_user res[], unsigned num,
 969                       const TDB_DATA *data, unsigned num_files)
 970 {
 971         unsigned int i, files_done;
 972         struct op *this_op;
 973         bool done[num_files];
 974
 975         /* Nothing left?  We're sorted. */
 976         if (num == 0)
 977                 return true;
 978
 979         memset(done, 0, sizeof(done));
 980
 981         /* Since ops within a trace file are ordered, we just need to figure
 982          * out which file to try next.  Since we don't take into account
 983          * inter-key relationships (which exist by virtue of trace file order),
 984          * we minimize the chance of harm by trying to keep in serial order. */
 985         for (files_done = 0, i = 0; i < num && files_done < num_files; i++) {
 986                 if (done[res[i].file])
 987                         continue;
 988
 989                 this_op = &op[res[i].file][res[i].op_num];
 990                 /* Is what we have good enough for this op? */
 991                 if (satisfies(data, needs(this_op))) {
 992                         move_to_front(res, i);
 993                         if (sort_deps(filename, op, res+1, num-1,
 994                                       gives(this_op, data), num_files))
 995                                 return true;
 996                         restore_to_pos(res, i);
 997                 }
 998                 done[res[i].file] = true;
 999                 files_done++;
1000         }
1001
1002         /* No combination worked. */
1003         return false;
1004 }
1005
1006 static void check_dep_sorting(struct key_user user[], unsigned num_users,
1007                               unsigned num_files)
1008 {
1009 #if DEBUG_DEPS
1010         unsigned int i;
1011         unsigned minima[num_files];
1012
1013         memset(minima, 0, sizeof(minima));
1014         for (i = 0; i < num_users; i++) {
1015                 assert(minima[user[i].file] < user[i].op_num);
1016                 minima[user[i].file] = user[i].op_num;
1017         }
1018 #endif
1019 }
1020
1021 /* All these ops have the same serial number.  Which comes first?
1022  *
1023  * This can happen both because read ops or failed write ops don't
1024  * change serial number, and also due to race since we access the
1025  * number unlocked (the race can cause less detectable ordering problems,
1026  * in which case we'll deadlock and report: fix manually in that case).
1027  */
1028 static void figure_deps(char *filename[], struct op *op[],
1029                         struct key_user user[], unsigned num_users,
1030                         unsigned num_files)
1031 {
1032         /* We assume database starts empty. */
1033         const struct TDB_DATA *data = &tdb_null;
1034
1035         if (!sort_deps(filename, op, user, num_users, data, num_files))
1036                 fail(filename[user[0].file], user[0].op_num+1,
1037                      "Could not resolve inter-dependencies");
1038
1039         check_dep_sorting(user, num_users, num_files);
1040 }
1041
1042 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1043                      unsigned int num)
1044 {
1045         unsigned int h;
1046
1047         /* Gcc nexted function extension.  How cool is this? */
1048         int compare_serial(const void *_a, const void *_b)
1049         {
1050                 const struct key_user *a = _a, *b = _b;
1051
1052                 /* First, maintain order within any trace file. */
1053                 if (a->file == b->file)
1054                         return a->op_num - b->op_num;
1055
1056                 /* Otherwise, arrange by serial order. */
1057                 return op[a->file][a->op_num].serial
1058                         - op[b->file][b->op_num].serial;
1059         }
1060
1061         /* Now sort into serial order. */
1062         for (h = 0; h < total_keys * 2; h++) {
1063                 struct key_user *user = hash[h].user;
1064
1065                 qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
1066                 figure_deps(filename, op, user, hash[h].num_users, num);
1067         }
1068 }
1069
1070 static int destroy_depend(struct depend *dep)
1071 {
1072         list_del(&dep->pre_list);
1073         list_del(&dep->post_list);
1074         return 0;
1075 }
1076
1077 static void add_dependency(void *ctx,
1078                            struct op *op[],
1079                            char *filename[],
1080                            unsigned int needs_file,
1081                            unsigned int needs_opnum,
1082                            unsigned int satisfies_file,
1083                            unsigned int satisfies_opnum)
1084 {
1085         struct depend *dep;
1086         unsigned int needs_start, sat_start;
1087
1088         /* We don't depend on ourselves. */
1089         if (needs_file == satisfies_file) {
1090                 assert(satisfies_opnum < needs_opnum);
1091                 return;
1092         }
1093
1094 #if DEBUG_DEPS
1095         printf("%s:%u: depends on %s:%u\n",
1096                filename[needs_file], needs_opnum+1,
1097                filename[satisfies_file], satisfies_opnum+1);
1098 #endif
1099
1100         needs_start = op[needs_file][needs_opnum].group_start;
1101         sat_start = op[satisfies_file][satisfies_opnum].group_start;
1102
1103         /* If needs is in a transaction, we need it before start. */
1104         if (needs_start) {
1105                 switch (op[needs_file][needs_start].op) {
1106                 case OP_TDB_TRANSACTION_START:
1107                         needs_opnum = needs_start;
1108 #ifdef DEBUG_DEPS
1109                         printf("  -> Back to %u\n", needs_start+1);
1110                         fflush(stdout);
1111 #endif
1112                         break;
1113                 default:
1114                         break;
1115                 }
1116         }
1117
1118         /* If satisfies is in a transaction, we wait until after commit. */
1119         /* FIXME: If transaction is cancelled, don't need dependency. */
1120         if (sat_start) {
1121                 if (op[satisfies_file][sat_start].op
1122                     == OP_TDB_TRANSACTION_START) {
1123                         satisfies_opnum = sat_start
1124                                 + op[satisfies_file][sat_start].group_len;
1125 #ifdef DEBUG_DEPS
1126                         printf("  -> Depends on %u\n", satisfies_opnum+1);
1127                         fflush(stdout);
1128 #endif
1129                 }
1130         }
1131
1132         assert(op[needs_file][needs_opnum].op != OP_TDB_TRAVERSE);
1133         assert(op[satisfies_file][satisfies_opnum].op != OP_TDB_TRAVERSE);
1134
1135         dep = talloc(ctx, struct depend);
1136         dep->needs_file = needs_file;
1137         dep->needs_opnum = needs_opnum;
1138         dep->satisfies_file = satisfies_file;
1139         dep->satisfies_opnum = satisfies_opnum;
1140         list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
1141         list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
1142         talloc_set_destructor(dep, destroy_depend);
1143 }
1144
1145 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1146 struct traverse_dep {
1147         unsigned int file;
1148         unsigned int op_num;
1149 };
1150
1151 /* Traversals can deadlock against each other, and transactions.  Force
1152  * order. */
1153 static void make_traverse_depends(char *filename[],
1154                                   struct op *op[], unsigned int num_ops[],
1155                                   unsigned int num)
1156 {
1157         unsigned int i, j, num_traversals = 0;
1158         struct traverse_dep *dep;
1159
1160         /* Sort by which one runs first. */
1161         int compare_traverse_dep(const void *_a, const void *_b)
1162         {
1163                 const struct traverse_dep *ta = _a, *tb = _b;
1164                 const struct op *a = &op[ta->file][ta->op_num],
1165                         *b = &op[tb->file][tb->op_num];
1166
1167                 if (a->serial != b->serial)
1168                         return a->serial - b->serial;
1169
1170                 /* If they have same serial, it means one didn't make any
1171                  * changes.  Thus sort by end in that case. */
1172                 return a[a->group_len].serial - b[b->group_len].serial;
1173         }
1174
1175         dep = talloc_array(NULL, struct traverse_dep, 1);
1176
1177         /* Count them. */
1178         for (i = 0; i < num; i++) {
1179                 for (j = 1; j < num_ops[i]; j++) {
1180                         /* Transaction or traverse start. */
1181                         if (op[i][j].group_start == j) {
1182                                 dep = talloc_realloc(NULL, dep,
1183                                                      struct traverse_dep,
1184                                                      num_traversals+1);
1185                                 dep[num_traversals].file = i;
1186                                 dep[num_traversals].op_num = j;
1187                                 num_traversals++;
1188                         }
1189                 }
1190         }
1191         qsort(dep, num_traversals, sizeof(dep[0]), compare_traverse_dep);
1192         for (i = 1; i < num_traversals; i++) {
1193                 /* i depends on end of traverse i-1. */
1194                 add_dependency(NULL, op, filename, dep[i].file, dep[i].op_num,
1195                                dep[i-1].file, dep[i-1].op_num
1196                                + op[dep[i-1].file][dep[i-1].op_num].group_len);
1197         }
1198         talloc_free(dep);
1199 }
1200 #endif /* TRAVERSALS_TAKE_TRANSACTION_LOCK */
1201
1202 static bool changes_db(const struct op *op)
1203 {
1204         return gives(op, NULL) != NULL;
1205 }
1206
1207 static void depend_on_previous(struct op *op[],
1208                                char *filename[],
1209                                unsigned int num,
1210                                struct key_user user[],
1211                                unsigned int i,
1212                                int prev)
1213 {
1214         bool deps[num];
1215         int j;
1216
1217         if (i == 0)
1218                 return;
1219
1220         if (prev == i - 1) {
1221                 /* Just depend on previous. */
1222                 add_dependency(NULL, op, filename,
1223                                user[i].file, user[i].op_num,
1224                                user[prev].file, user[prev].op_num);
1225                 return;
1226         }
1227
1228         /* We have to wait for the readers.  Find last one in *each* file. */
1229         memset(deps, 0, sizeof(deps));
1230         deps[user[i].file] = true;
1231         for (j = i - 1; j > prev; j--) {
1232                 if (!deps[user[j].file]) {
1233                         add_dependency(NULL, op, filename,
1234                                        user[i].file, user[i].op_num,
1235                                        user[j].file, user[j].op_num);
1236                         deps[user[j].file] = true;
1237                 }
1238         }
1239 }
1240
1241 /* This is simple, but not complete.  We don't take into account
1242  * indirect dependencies. */
1243 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1244                                   unsigned int num)
1245 {
1246         unsigned int i, j;
1247
1248         /* There can only be one real dependency on each file */
1249         for (i = 0; i < num; i++) {
1250                 for (j = 1; j < num_ops[i]; j++) {
1251                         struct depend *dep, *next;
1252                         struct depend *prev[num];
1253
1254                         memset(prev, 0, sizeof(prev));
1255
1256                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1257                                 if (!prev[dep->satisfies_file]) {
1258                                         prev[dep->satisfies_file] = dep;
1259                                         continue;
1260                                 }
1261                                 if (prev[dep->satisfies_file]->satisfies_opnum
1262                                     < dep->satisfies_opnum) {
1263                                         talloc_free(prev[dep->satisfies_file]);
1264                                         prev[dep->satisfies_file] = dep;
1265                                 } else
1266                                         talloc_free(dep);
1267                         }
1268                 }
1269         }
1270
1271         for (i = 0; i < num; i++) {
1272                 int deps[num];
1273
1274                 for (j = 0; j < num; j++)
1275                         deps[j] = -1;
1276
1277                 for (j = 1; j < num_ops[i]; j++) {
1278                         struct depend *dep, *next;
1279
1280                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1281                                 if (deps[dep->satisfies_file]
1282                                     >= (int)dep->satisfies_opnum)
1283                                         talloc_free(dep);
1284                                 else
1285                                         deps[dep->satisfies_file]
1286                                                 = dep->satisfies_opnum;
1287                         }
1288                 }
1289         }
1290 }
1291
1292 static void derive_dependencies(char *filename[],
1293                                 struct op *op[], unsigned int num_ops[],
1294                                 unsigned int num)
1295 {
1296         struct keyinfo *hash;
1297         unsigned int h, i;
1298
1299         /* Create hash table for faster key lookup. */
1300         hash = hash_ops(op, num_ops, num);
1301
1302         /* Sort them by serial number. */
1303         sort_ops(hash, filename, op, num);
1304
1305         /* Create dependencies back to the last change, rather than
1306          * creating false dependencies by naively making each one
1307          * depend on the previous.  This has two purposes: it makes
1308          * later optimization simpler, and it also avoids deadlock with
1309          * same sequence number ops inside traversals (if one
1310          * traversal doesn't write anything, two ops can have the same
1311          * sequence number yet we can create a traversal dependency
1312          * the other way). */
1313         for (h = 0; h < total_keys * 2; h++) {
1314                 int prev = -1;
1315
1316                 if (hash[h].num_users < 2)
1317                         continue;
1318
1319                 for (i = 0; i < hash[h].num_users; i++) {
1320                         if (changes_db(&op[hash[h].user[i].file]
1321                                        [hash[h].user[i].op_num])) {
1322                                 depend_on_previous(op, filename, num,
1323                                                    hash[h].user, i, prev);
1324                                 prev = i;
1325                         } else if (prev >= 0)
1326                                 add_dependency(hash, op, filename,
1327                                                hash[h].user[i].file,
1328                                                hash[h].user[i].op_num,
1329                                                hash[h].user[prev].file,
1330                                                hash[h].user[prev].op_num);
1331                 }
1332         }
1333
1334 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1335         make_traverse_depends(filename, op, num_ops, num);
1336 #endif
1337
1338         optimize_dependencies(op, num_ops, num);
1339 }
1340
1341 int main(int argc, char *argv[])
1342 {
1343         struct timeval start, end;
1344         unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1345         struct op *op[argc];
1346         int fds[2];
1347         char c;
1348         bool ok = true;
1349
1350         if (argc < 3)
1351                 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1352
1353         pipes = talloc_array(NULL, struct pipe, argc - 2);
1354         for (i = 0; i < argc - 2; i++) {
1355                 printf("Loading tracefile %s...", argv[2+i]);
1356                 fflush(stdout);
1357                 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1358                                        &tdb_flags[i], &open_flags[i]);
1359                 if (pipe(pipes[i].fd) != 0)
1360                         err(1, "creating pipe");
1361                 printf("done\n");
1362         }
1363
1364         printf("Calculating inter-dependencies...");
1365         fflush(stdout);
1366         derive_dependencies(argv+2, op, num_ops, i);
1367         printf("done\n");
1368
1369         /* Don't fork for single arg case: simple debugging. */
1370         if (argc == 3) {
1371                 struct tdb_context *tdb;
1372                 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
1373                                   open_flags[0], 0600,
1374                                   NULL, hash_key);
1375                 printf("Single threaded run...");
1376                 fflush(stdout);
1377
1378                 run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0]);
1379                 check_deps(argv[2], op[0], num_ops[0]);
1380
1381                 printf("done\n");
1382                 exit(0);
1383         }
1384
1385         if (pipe(fds) != 0)
1386                 err(1, "creating pipe");
1387
1388         for (i = 0; i < argc - 2; i++) {
1389                 struct tdb_context *tdb;
1390
1391                 switch (fork()) {
1392                 case -1:
1393                         err(1, "fork failed");
1394                 case 0:
1395                         close(fds[1]);
1396                         tdb = tdb_open_ex(argv[1], hashsize[i], tdb_flags[i],
1397                                           open_flags[i], 0600,
1398                                           NULL, hash_key);
1399                         if (!tdb)
1400                                 err(1, "Opening tdb %s", argv[1]);
1401
1402                         /* This catches parent exiting. */
1403                         if (read(fds[0], &c, 1) != 1)
1404                                 exit(1);
1405                         run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
1406                                 num_ops[i]);
1407                         check_deps(argv[2+i], op[i], num_ops[i]);
1408                         exit(0);
1409                 default:
1410                         break;
1411                 }
1412         }
1413
1414         /* Let everything settle. */
1415         sleep(1);
1416
1417         printf("Starting run...");
1418         fflush(stdout);
1419         gettimeofday(&start, NULL);
1420         /* Tell them all to go!  Any write of sufficient length will do. */
1421         if (write(fds[1], hashsize, i) != i)
1422                 err(1, "Writing to wakeup pipe");
1423
1424         for (i = 0; i < argc - 2; i++) {
1425                 int status;
1426                 wait(&status);
1427                 if (!WIFEXITED(status)) {
1428                         warnx("Child died with signal %i", WTERMSIG(status));
1429                         ok = false;
1430                 } else if (WEXITSTATUS(status) != 0)
1431                         /* Assume child spat out error. */
1432                         ok = false;
1433         }
1434         if (!ok)
1435                 exit(1);
1436
1437         gettimeofday(&end, NULL);
1438         printf("done\n");
1439
1440         end.tv_sec -= start.tv_sec;
1441         printf("Time replaying: %lu usec\n",
1442                end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));
1443
1444         exit(0);
1445 }