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                 /* transaction/traverse start/chainlock */
 151                 unsigned int group_len;
 152         };
 153 };
 154
 155 static unsigned char hex_char(const char *filename, unsigned int line, char c)
 156 {
 157         c = toupper(c);
 158         if (c >= 'A' && c <= 'F')
 159                 return c - 'A' + 10;
 160         if (c >= '0' && c <= '9')
 161                 return c - '0';
 162         fail(filename, line, "invalid hex character '%c'", c);
 163 }
 164
 165 /* TDB data is <size>:<%02x>* */
 166 static TDB_DATA make_tdb_data(const void *ctx,
 167                               const char *filename, unsigned int line,
 168                               const char *word)
 169 {
 170         TDB_DATA data;
 171         unsigned int i;
 172         const char *p;
 173
 174         if (streq(word, "NULL"))
 175                 return tdb_null;
 176
 177         data.dsize = atoi(word);
 178         data.dptr = talloc_array(ctx, unsigned char, data.dsize);
 179         p = strchr(word, ':');
 180         if (!p)
 181                 fail(filename, line, "invalid tdb data '%s'", word);
 182         p++;
 183         for (i = 0; i < data.dsize; i++)
 184                 data.dptr[i] = hex_char(filename, line, p[i*2])*16
 185                         + hex_char(filename, line, p[i*2+1]);
 186
 187         return data;
 188 }
 189
 190 static void add_op(const char *filename, struct op **op, unsigned int i,
 191                    unsigned int serial, enum op_type type)
 192 {
 193         struct op *new;
 194         *op = talloc_realloc(NULL, *op, struct op, i+1);
 195         new = (*op) + i;
 196         new->op = type;
 197         new->serial = serial;
 198         new->ret = 0;
 199         new->group_start = 0;
 200 }
 201
 202 static void op_add_nothing(const char *filename,
 203                            struct op op[], unsigned int op_num, char *words[])
 204 {
 205         if (words[2])
 206                 fail(filename, op_num+1, "Expected no arguments");
 207         op[op_num].key = tdb_null;
 208 }
 209
 210 static void op_add_key(const char *filename,
 211                        struct op op[], unsigned int op_num, char *words[])
 212 {
 213         if (words[2] == NULL || words[3])
 214                 fail(filename, op_num+1, "Expected just a key");
 215
 216         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 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 /* We don't record the keys or data for a traverse, as we don't use them. */
 247 static void op_add_traverse(const char *filename,
 248                             struct op op[], unsigned int op_num, char *words[])
 249 {
 250         if (!words[2] || !words[3] || !words[4] || words[5]
 251             || !streq(words[3], "="))
 252                 fail(filename, op_num+1, "Expected <key> = <data>");
 253         op[op_num].key = tdb_null;
 254 }
 255
 256 /* <serial> tdb_store <rec> <rec> <flag> = <ret> */
 257 static void op_add_store(const char *filename,
 258                          struct op op[], unsigned int op_num, char *words[])
 259 {
 260         if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
 261             || words[7] || !streq(words[5], "="))
 262                 fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
 263
 264         op[op_num].flag = strtoul(words[4], NULL, 0);
 265         op[op_num].ret = atoi(words[6]);
 266         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 267         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 268         total_keys++;
 269 }
 270
 271 /* <serial> tdb_append <rec> <rec> = <rec> */
 272 static void op_add_append(const char *filename,
 273                           struct op op[], unsigned int op_num, char *words[])
 274 {
 275         if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
 276             || !streq(words[4], "="))
 277                 fail(filename, op_num+1, "Expect <key> <data> = <rec>");
 278
 279         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 280         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 281
 282         op[op_num].append.post
 283                 = make_tdb_data(op, filename, op_num+1, words[5]);
 284
 285         /* By subtraction, figure out what previous data was. */
 286         op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
 287         op[op_num].append.pre.dsize
 288                 = op[op_num].append.post.dsize - op[op_num].data.dsize;
 289         total_keys++;
 290 }
 291
 292 /* <serial> tdb_get_seqnum = <ret> */
 293 static void op_add_seqnum(const char *filename,
 294                           struct op op[], unsigned int op_num, char *words[])
 295 {
 296         if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
 297                 fail(filename, op_num+1, "Expect = <ret>");
 298
 299         op[op_num].key = tdb_null;
 300         op[op_num].ret = atoi(words[3]);
 301 }
 302
 303 static void op_add_traverse_start(const char *filename,
 304                                   struct op op[],
 305                                   unsigned int op_num, char *words[])
 306 {
 307         if (words[2])
 308                 fail(filename, op_num+1, "Expect no arguments");
 309
 310         op[op_num].key = tdb_null;
 311         op[op_num].group_len = 0;
 312 }
 313
 314 static void op_add_transaction(const char *filename, struct op op[],
 315                                unsigned int op_num, char *words[])
 316 {
 317         if (words[2])
 318                 fail(filename, op_num+1, "Expect no arguments");
 319
 320         op[op_num].key = tdb_null;
 321         op[op_num].group_len = 0;
 322 }
 323
 324 static void op_add_chainlock(const char *filename,
 325                              struct op op[], unsigned int op_num, char *words[])
 326 {
 327         if (words[2] == NULL || words[3])
 328                 fail(filename, op_num+1, "Expected just a key");
 329
 330         /* A chainlock key isn't a key in the normal sense; it doesn't
 331          * have to be in the db at all.  Also, we don't want to hash this op. */
 332         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 333         op[op_num].key = tdb_null;
 334         op[op_num].group_len = 0;
 335 }
 336
 337 static void op_add_chainlock_ret(const char *filename,
 338                                  struct op op[], unsigned int op_num,
 339                                  char *words[])
 340 {
 341         if (!words[2] || !words[3] || !words[4] || words[5]
 342             || !streq(words[3], "="))
 343                 fail(filename, op_num+1, "Expected <key> = <ret>");
 344         op[op_num].ret = atoi(words[4]);
 345         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 346         op[op_num].key = tdb_null;
 347         op[op_num].group_len = 0;
 348         total_keys++;
 349 }
 350
 351 static int op_find_start(struct op op[], unsigned int op_num, enum op_type type)
 352 {
 353         unsigned int i;
 354
 355         for (i = op_num-1; i > 0; i--) {
 356                 if (op[i].op == type && !op[i].group_len)
 357                         return i;
 358         }
 359         return 0;
 360 }
 361
 362 static void op_analyze_transaction(const char *filename,
 363                                    struct op op[], unsigned int op_num,
 364                                    char *words[])
 365 {
 366         unsigned int start, i;
 367
 368         op[op_num].key = tdb_null;
 369
 370         if (words[2])
 371                 fail(filename, op_num+1, "Expect no arguments");
 372
 373         start = op_find_start(op, op_num, OP_TDB_TRANSACTION_START);
 374         if (!start)
 375                 fail(filename, op_num+1, "no transaction start found");
 376
 377         op[start].group_len = op_num - start;
 378
 379         /* This rolls in nested transactions.  I think that's right. */
 380         for (i = start; i <= op_num; i++)
 381                 op[i].group_start = start;
 382 }
 383
 384 /* We treat chainlocks a lot like transactions, even though that's overkill */
 385 static void op_analyze_chainlock(const char *filename,
 386                                  struct op op[], unsigned int op_num,
 387                                  char *words[])
 388 {
 389         unsigned int i, start;
 390
 391         if (words[2] == NULL || words[3])
 392                 fail(filename, op_num+1, "Expected just a key");
 393
 394         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 395         op[op_num].key = tdb_null;
 396         total_keys++;
 397
 398         start = op_find_start(op, op_num, OP_TDB_CHAINLOCK);
 399         if (!start)
 400                 start = op_find_start(op, op_num, OP_TDB_CHAINLOCK_READ);
 401         if (!start)
 402                 fail(filename, op_num+1, "no initial chainlock found");
 403
 404         /* FIXME: We'd have to do something clever to make this work
 405          * vs. deadlock. */
 406         if (!key_eq(op[start].data, op[op_num].data))
 407                 fail(filename, op_num+1, "nested chainlock calls?");
 408
 409         op[start].group_len = op_num - start;
 410         for (i = start; i <= op_num; i++)
 411                 op[i].group_start = start;
 412 }
 413
 414 static void op_analyze_traverse(const char *filename,
 415                                 struct op op[], unsigned int op_num,
 416                                 char *words[])
 417 {
 418         int i, start;
 419
 420         op[op_num].key = tdb_null;
 421
 422         /* = %u means traverse function terminated. */
 423         if (words[2]) {
 424                 if (!streq(words[2], "=") || !words[3] || words[4])
 425                         fail(filename, op_num+1, "expect = <num>");
 426                 op[op_num].ret = atoi(words[3]);
 427         } else
 428                 op[op_num].ret = 0;
 429
 430         start = op_find_start(op, op_num, OP_TDB_TRAVERSE_START);
 431         if (!start)
 432                 start = op_find_start(op, op_num, OP_TDB_TRAVERSE_READ_START);
 433         if (!start)
 434                 fail(filename, op_num+1, "no traversal start found");
 435
 436         op[start].group_len = op_num - start;
 437
 438         /* Don't roll in nested traverse/chainlock */
 439         for (i = start; i <= op_num; i++)
 440                 if (!op[i].group_start)
 441                         op[i].group_start = start;
 442 }
 443
 444 /* Keep -Wmissing-declarations happy: */
 445 const struct op_table *
 446 find_keyword (register const char *str, register unsigned int len);
 447
 448 #include "keywords.c"
 449
 450 struct depend {
 451         /* We can have more than one */
 452         struct list_node pre_list;
 453         struct list_node post_list;
 454         unsigned int needs_file;
 455         unsigned int needs_opnum;
 456         unsigned int satisfies_file;
 457         unsigned int satisfies_opnum;
 458 };
 459
 460 static void check_deps(const char *filename, struct op op[], unsigned int num)
 461 {
 462 #ifdef DEBUG_DEPS
 463         unsigned int i;
 464
 465         for (i = 1; i < num; i++)
 466                 if (!list_empty(&op[i].pre))
 467                         fail(filename, i+1, "Still has dependencies");
 468 #endif
 469 }
 470
 471 static void dump_pre(char *filename[], struct op *op[],
 472                      unsigned int file, unsigned int i)
 473 {
 474         struct depend *dep;
 475
 476         printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
 477                 op[file][i].serial);
 478         list_for_each(&op[file][i].pre, dep, pre_list)
 479                 printf("    %s:%u (%u)\n",
 480                        filename[dep->satisfies_file], dep->satisfies_opnum+1,
 481                        op[dep->satisfies_file][dep->satisfies_opnum].serial);
 482         check_deps(filename[file], op[file], i);
 483 }
 484
 485 /* We simply read/write pointers, since we all are children. */
 486 static bool do_pre(struct tdb_context *tdb,
 487                    char *filename[], struct op *op[],
 488                    unsigned int file, int pre_fd, unsigned int i,
 489                    bool backoff)
 490 {
 491         while (!list_empty(&op[file][i].pre)) {
 492                 struct depend *dep;
 493
 494 #if DEBUG_DEPS
 495                 printf("%s:%u:waiting for pre\n", filename[file], i+1);
 496                 fflush(stdout);
 497 #endif
 498                 if (backoff)
 499                         alarm(2);
 500                 else
 501                         alarm(10);
 502                 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
 503                         if (errno == EINTR) {
 504                                 if (backoff) {
 505                                         warnx("%s:%u:avoiding deadlock",
 506                                               filename[file], i+1);
 507                                         return false;
 508                                 }
 509                                 dump_pre(filename, op, file, i);
 510                                 exit(1);
 511                         } else
 512                                 errx(1, "Reading from pipe");
 513                 }
 514                 alarm(0);
 515
 516 #if DEBUG_DEPS
 517                 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
 518                        dep->needs_opnum+1, filename[dep->satisfies_file],
 519                        dep->satisfies_opnum+1);
 520                 fflush(stdout);
 521 #endif
 522                 /* This could be any op, not just this one. */
 523                 talloc_free(dep);
 524         }
 525         return true;
 526 }
 527
 528 static void do_post(char *filename[], struct op *op[],
 529                     unsigned int file, unsigned int i)
 530 {
 531         struct depend *dep;
 532
 533         list_for_each(&op[file][i].post, dep, post_list) {
 534 #if DEBUG_DEPS
 535                 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
 536                        filename[dep->needs_file], dep->needs_opnum+1);
 537 #endif
 538                 if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
 539                     != sizeof(dep))
 540                         err(1, "%s:%u failed to tell file %s",
 541                             filename[file], i+1, filename[dep->needs_file]);
 542         }
 543 }
 544
 545 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 546 {
 547         return data.dsize;
 548 }
 549
 550 static unsigned run_ops(struct tdb_context *tdb,
 551                         int pre_fd,
 552                         char *filename[],
 553                         struct op *op[],
 554                         unsigned int file,
 555                         unsigned int start, unsigned int stop,
 556                         bool backoff);
 557
 558 struct traverse_info {
 559         struct op **op;
 560         char **filename;
 561         unsigned file;
 562         int pre_fd;
 563         unsigned int start;
 564         unsigned int i;
 565 };
 566
 567 /* More complex.  Just do whatever's they did at the n'th entry. */
 568 static int nontrivial_traverse(struct tdb_context *tdb,
 569                                TDB_DATA key, TDB_DATA data,
 570                                void *_tinfo)
 571 {
 572         struct traverse_info *tinfo = _tinfo;
 573         unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
 574         bool avoid_deadlock = false;
 575
 576         if (tinfo->i == tinfo->start + trav_len) {
 577                 /* This can happen if traverse expects to be empty. */
 578                 if (trav_len == 1)
 579                         return 1;
 580                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 581                      "traverse did not terminate");
 582         }
 583
 584         if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
 585                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 586                      "%s:%u:traverse terminated early");
 587
 588 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
 589         avoid_deadlock = true;
 590 #endif
 591
 592         /* Run any normal ops. */
 593         tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
 594                            tinfo->file, tinfo->i+1, tinfo->start + trav_len,
 595                            avoid_deadlock);
 596
 597         /* We backed off, or we hit OP_TDB_TRAVERSE_END. */
 598         if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
 599                 return 1;
 600
 601         return 0;
 602 }
 603
 604 static unsigned op_traverse(struct tdb_context *tdb,
 605                             int pre_fd,
 606                             char *filename[],
 607                             unsigned int file,
 608                             int (*traversefn)(struct tdb_context *,
 609                                               tdb_traverse_func, void *),
 610                             struct op *op[],
 611                             unsigned int start)
 612 {
 613         struct traverse_info tinfo = { op, filename, file, pre_fd,
 614                                        start, start+1 };
 615
 616         traversefn(tdb, nontrivial_traverse, &tinfo);
 617
 618         /* Traversing in wrong order can have strange effects: eg. if
 619          * original traverse went A (delete A), B, we might do B
 620          * (delete A).  So if we have ops left over, we do it now. */
 621         while (tinfo.i != start + op[file][start].group_len) {
 622                 if (op[file][tinfo.i].op == OP_TDB_TRAVERSE)
 623                         tinfo.i++;
 624                 else
 625                         tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
 626                                           tinfo.i,
 627                                           start + op[file][start].group_len,
 628                                           false);
 629         }
 630
 631         return tinfo.i;
 632 }
 633
 634 static void break_out(int sig)
 635 {
 636 }
 637
 638 static __attribute__((noinline))
 639 unsigned run_ops(struct tdb_context *tdb,
 640                  int pre_fd,
 641                  char *filename[],
 642                  struct op *op[],
 643                  unsigned int file,
 644                  unsigned int start, unsigned int stop,
 645                  bool backoff)
 646 {
 647         unsigned int i;
 648         struct sigaction sa;
 649
 650         sa.sa_handler = break_out;
 651         sa.sa_flags = 0;
 652
 653         sigaction(SIGALRM, &sa, NULL);
 654         for (i = start; i < stop; i++) {
 655                 if (!do_pre(tdb, filename, op, file, pre_fd, i, backoff))
 656                         return i;
 657
 658                 switch (op[file][i].op) {
 659                 case OP_TDB_LOCKALL:
 660                         try(tdb_lockall(tdb), op[file][i].ret);
 661                         break;
 662                 case OP_TDB_LOCKALL_MARK:
 663                         try(tdb_lockall_mark(tdb), op[file][i].ret);
 664                         break;
 665                 case OP_TDB_LOCKALL_UNMARK:
 666                         try(tdb_lockall_unmark(tdb), op[file][i].ret);
 667                         break;
 668                 case OP_TDB_LOCKALL_NONBLOCK:
 669                         unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
 670                                    tdb_lockall(tdb), tdb_unlockall(tdb));
 671                         break;
 672                 case OP_TDB_UNLOCKALL:
 673                         try(tdb_unlockall(tdb), op[file][i].ret);
 674                         break;
 675                 case OP_TDB_LOCKALL_READ:
 676                         try(tdb_lockall_read(tdb), op[file][i].ret);
 677                         break;
 678                 case OP_TDB_LOCKALL_READ_NONBLOCK:
 679                         unreliable(tdb_lockall_read_nonblock(tdb),
 680                                    op[file][i].ret,
 681                                    tdb_lockall_read(tdb),
 682                                    tdb_unlockall_read(tdb));
 683                         break;
 684                 case OP_TDB_UNLOCKALL_READ:
 685                         try(tdb_unlockall_read(tdb), op[file][i].ret);
 686                         break;
 687                 case OP_TDB_CHAINLOCK:
 688                         try(tdb_chainlock(tdb, op[file][i].key),
 689                             op[file][i].ret);
 690                         break;
 691                 case OP_TDB_CHAINLOCK_NONBLOCK:
 692                         unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
 693                                    op[file][i].ret,
 694                                    tdb_chainlock(tdb, op[file][i].key),
 695                                    tdb_chainunlock(tdb, op[file][i].key));
 696                         break;
 697                 case OP_TDB_CHAINLOCK_MARK:
 698                         try(tdb_chainlock_mark(tdb, op[file][i].key),
 699                             op[file][i].ret);
 700                         break;
 701                 case OP_TDB_CHAINLOCK_UNMARK:
 702                         try(tdb_chainlock_unmark(tdb, op[file][i].key),
 703                             op[file][i].ret);
 704                         break;
 705                 case OP_TDB_CHAINUNLOCK:
 706                         try(tdb_chainunlock(tdb, op[file][i].key),
 707                             op[file][i].ret);
 708                         break;
 709                 case OP_TDB_CHAINLOCK_READ:
 710                         try(tdb_chainlock_read(tdb, op[file][i].key),
 711                             op[file][i].ret);
 712                         break;
 713                 case OP_TDB_CHAINUNLOCK_READ:
 714                         try(tdb_chainunlock_read(tdb, op[file][i].key),
 715                             op[file][i].ret);
 716                         break;
 717                 case OP_TDB_PARSE_RECORD:
 718                         try(tdb_parse_record(tdb, op[file][i].key, get_len,
 719                                              NULL),
 720                             op[file][i].ret);
 721                         break;
 722                 case OP_TDB_EXISTS:
 723                         try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
 724                         break;
 725                 case OP_TDB_STORE:
 726                         try(tdb_store(tdb, op[file][i].key, op[file][i].data,
 727                                       op[file][i].flag),
 728                             op[file][i].ret);
 729                         break;
 730                 case OP_TDB_APPEND:
 731                         try(tdb_append(tdb, op[file][i].key, op[file][i].data),
 732                             op[file][i].ret);
 733                         break;
 734                 case OP_TDB_GET_SEQNUM:
 735                         try(tdb_get_seqnum(tdb), op[file][i].ret);
 736                         break;
 737                 case OP_TDB_WIPE_ALL:
 738                         try(tdb_wipe_all(tdb), op[file][i].ret);
 739                         break;
 740                 case OP_TDB_TRANSACTION_START:
 741                         try(tdb_transaction_start(tdb), op[file][i].ret);
 742                         break;
 743                 case OP_TDB_TRANSACTION_CANCEL:
 744                         try(tdb_transaction_cancel(tdb), op[file][i].ret);
 745                         break;
 746                 case OP_TDB_TRANSACTION_COMMIT:
 747                         try(tdb_transaction_commit(tdb), op[file][i].ret);
 748                         break;
 749                 case OP_TDB_TRAVERSE_READ_START:
 750                         i = op_traverse(tdb, pre_fd, filename, file,
 751                                         tdb_traverse_read, op, i);
 752                         break;
 753                 case OP_TDB_TRAVERSE_START:
 754                         i = op_traverse(tdb, pre_fd, filename, file,
 755                                         tdb_traverse, op, i);
 756                         break;
 757                 case OP_TDB_TRAVERSE:
 758                         /* Terminate: we're in a traverse, and we've
 759                          * done our ops. */
 760                         return i;
 761                 case OP_TDB_TRAVERSE_END:
 762                         fail(filename[file], i+1, "unexpected end traverse");
 763                 /* FIXME: These must be treated like traverse. */
 764                 case OP_TDB_FIRSTKEY:
 765                         if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
 766                                 fail(filename[file], i+1, "bad firstkey");
 767                         break;
 768                 case OP_TDB_NEXTKEY:
 769                         if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
 770                                     op[file][i].data))
 771                                 fail(filename[file], i+1, "bad nextkey");
 772                         break;
 773                 case OP_TDB_FETCH: {
 774                         TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
 775                         if (!key_eq(f, op[file][i].data))
 776                                 fail(filename[file], i+1, "bad fetch %u",
 777                                      f.dsize);
 778                         break;
 779                 }
 780                 case OP_TDB_DELETE:
 781                         try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
 782                         break;
 783                 }
 784                 do_post(filename, op, file, i);
 785         }
 786         return i;
 787 }
 788
 789 /* tdbtorture, in particular, can do a tdb_close with a transaction in
 790  * progress. */
 791 static struct op *maybe_cancel_transaction(const char *filename,
 792                                            struct op *op, unsigned int *num)
 793 {
 794         unsigned int start = op_find_start(op, *num, OP_TDB_TRANSACTION_START);
 795
 796         if (start) {
 797                 char *words[] = { "<unknown>", "tdb_close", NULL };
 798                 add_op(filename, &op, *num, op[start].serial,
 799                        OP_TDB_TRANSACTION_CANCEL);
 800                 op_analyze_transaction(filename, op, *num, words);
 801                 (*num)++;
 802         }
 803         return op;
 804 }
 805
 806 static struct op *load_tracefile(const char *filename, unsigned int *num,
 807                                  unsigned int *hashsize,
 808                                  unsigned int *tdb_flags,
 809                                  unsigned int *open_flags)
 810 {
 811         unsigned int i;
 812         struct op *op = talloc_array(NULL, struct op, 1);
 813         char **words;
 814         char **lines;
 815         char *file;
 816
 817         file = grab_file(NULL, filename, NULL);
 818         if (!file)
 819                 err(1, "Reading %s", filename);
 820
 821         lines = strsplit(file, file, "\n", NULL);
 822         if (!lines[0])
 823                 errx(1, "%s is empty", filename);
 824
 825         words = strsplit(lines, lines[0], " ", NULL);
 826         if (!streq(words[1], "tdb_open"))
 827                 fail(filename, 1, "does not start with tdb_open");
 828
 829         *hashsize = atoi(words[2]);
 830         *tdb_flags = strtoul(words[3], NULL, 0);
 831         *open_flags = strtoul(words[4], NULL, 0);
 832
 833         for (i = 1; lines[i]; i++) {
 834                 const struct op_table *opt;
 835
 836                 words = strsplit(lines, lines[i], " ", NULL);
 837                 if (!words[0] || !words[1])
 838                         fail(filename, i+1, "Expected serial number and op");
 839
 840                 opt = find_keyword(words[1], strlen(words[1]));
 841                 if (!opt) {
 842                         if (streq(words[1], "tdb_close")) {
 843                                 if (lines[i+1])
 844                                         fail(filename, i+2,
 845                                              "lines after tdb_close");
 846                                 *num = i;
 847                                 talloc_free(lines);
 848                                 return maybe_cancel_transaction(filename,
 849                                                                 op, num);
 850                         }
 851                         fail(filename, i+1, "Unknown operation '%s'", words[1]);
 852                 }
 853
 854                 add_op(filename, &op, i, atoi(words[0]), opt->type);
 855                 opt->enhance_op(filename, op, i, words);
 856         }
 857
 858         fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
 859               filename, i);
 860         talloc_free(lines);
 861         *num = i - 1;
 862         return maybe_cancel_transaction(filename, op, num);
 863 }
 864
 865 /* We remember all the keys we've ever seen, and who has them. */
 866 struct key_user {
 867         unsigned int file;
 868         unsigned int op_num;
 869 };
 870
 871 struct keyinfo {
 872         TDB_DATA key;
 873         unsigned int num_users;
 874         struct key_user *user;
 875 };
 876
 877 static const TDB_DATA must_not_exist;
 878 static const TDB_DATA must_exist;
 879 static const TDB_DATA not_exists_or_empty;
 880
 881 /* NULL means doesn't care if it exists or not, &must_exist means
 882  * it must exist but we don't care what, &must_not_exist means it must
 883  * not exist, otherwise the data it needs. */
 884 static const TDB_DATA *needs(const struct op *op)
 885 {
 886         switch (op->op) {
 887         /* FIXME: Pull forward deps, since we can deadlock */
 888         case OP_TDB_CHAINLOCK:
 889         case OP_TDB_CHAINLOCK_NONBLOCK:
 890         case OP_TDB_CHAINLOCK_MARK:
 891         case OP_TDB_CHAINLOCK_UNMARK:
 892         case OP_TDB_CHAINUNLOCK:
 893         case OP_TDB_CHAINLOCK_READ:
 894         case OP_TDB_CHAINUNLOCK_READ:
 895                 return NULL;
 896
 897         case OP_TDB_APPEND:
 898                 if (op->append.pre.dsize == 0)
 899                         return &not_exists_or_empty;
 900                 return &op->append.pre;
 901
 902         case OP_TDB_STORE:
 903                 if (op->flag == TDB_INSERT) {
 904                         if (op->ret < 0)
 905                                 return &must_exist;
 906                         else
 907                                 return &must_not_exist;
 908                 } else if (op->flag == TDB_MODIFY) {
 909                         if (op->ret < 0)
 910                                 return &must_not_exist;
 911                         else
 912                                 return &must_exist;
 913                 }
 914                 /* No flags?  Don't care */
 915                 return NULL;
 916
 917         case OP_TDB_EXISTS:
 918                 if (op->ret == 1)
 919                         return &must_exist;
 920                 else
 921                         return &must_not_exist;
 922
 923         case OP_TDB_PARSE_RECORD:
 924                 if (op->ret < 0)
 925                         return &must_not_exist;
 926                 return &must_exist;
 927
 928         /* FIXME: handle these. */
 929         case OP_TDB_WIPE_ALL:
 930         case OP_TDB_FIRSTKEY:
 931         case OP_TDB_NEXTKEY:
 932         case OP_TDB_GET_SEQNUM:
 933         case OP_TDB_TRAVERSE:
 934         case OP_TDB_TRANSACTION_COMMIT:
 935         case OP_TDB_TRANSACTION_CANCEL:
 936         case OP_TDB_TRANSACTION_START:
 937                 return NULL;
 938
 939         case OP_TDB_FETCH:
 940                 if (!op->data.dptr)
 941                         return &must_not_exist;
 942                 return &op->data;
 943
 944         case OP_TDB_DELETE:
 945                 if (op->ret < 0)
 946                         return &must_not_exist;
 947                 return &must_exist;
 948
 949         default:
 950                 errx(1, "Unexpected op %i", op->op);
 951         }
 952
 953 }
 954
 955 static bool starts_transaction(const struct op *op)
 956 {
 957         return op->op == OP_TDB_TRANSACTION_START;
 958 }
 959
 960 static bool in_transaction(const struct op op[], unsigned int i)
 961 {
 962         return op[i].group_start && starts_transaction(&op[op[i].group_start]);
 963 }
 964
 965 static bool successful_transaction(const struct op *op)
 966 {
 967         return starts_transaction(op)
 968                 && op[op->group_len].op == OP_TDB_TRANSACTION_COMMIT;
 969 }
 970
 971 static bool starts_traverse(const struct op *op)
 972 {
 973         return op->op == OP_TDB_TRAVERSE_START
 974                 || op->op == OP_TDB_TRAVERSE_READ_START;
 975 }
 976
 977 static bool in_traverse(const struct op op[], unsigned int i)
 978 {
 979         return op[i].group_start && starts_traverse(&op[op[i].group_start]);
 980 }
 981
 982 static bool starts_chainlock(const struct op *op)
 983 {
 984         return op->op == OP_TDB_CHAINLOCK_READ || op->op == OP_TDB_CHAINLOCK;
 985 }
 986
 987 static bool in_chainlock(const struct op op[], unsigned int i)
 988 {
 989         return op[i].group_start && starts_chainlock(&op[op[i].group_start]);
 990 }
 991
 992 /* What's the data after this op?  pre if nothing changed. */
 993 static const TDB_DATA *gives(const TDB_DATA *key, const TDB_DATA *pre,
 994                              const struct op *op)
 995 {
 996         if (starts_transaction(op) || starts_chainlock(op)) {
 997                 unsigned int i;
 998
 999                 /* Cancelled transactions don't change anything. */
1000                 if (op[op->group_len].op == OP_TDB_TRANSACTION_CANCEL)
1001                         return pre;
1002                 assert(op[op->group_len].op == OP_TDB_TRANSACTION_COMMIT
1003                        || op[op->group_len].op == OP_TDB_CHAINUNLOCK_READ
1004                        || op[op->group_len].op == OP_TDB_CHAINUNLOCK);
1005
1006                 for (i = 1; i < op->group_len; i++) {
1007                         /* This skips nested transactions, too */
1008                         if (key_eq(op[i].key, *key))
1009                                 pre = gives(key, pre, &op[i]);
1010                 }
1011                 return pre;
1012         }
1013
1014         /* Failed ops don't change state of db. */
1015         if (op->ret < 0)
1016                 return pre;
1017
1018         if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
1019                 return &tdb_null;
1020
1021         if (op->op == OP_TDB_APPEND)
1022                 return &op->append.post;
1023
1024         if (op->op == OP_TDB_STORE)
1025                 return &op->data;
1026
1027         return pre;
1028 }
1029
1030 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
1031                                 unsigned int num)
1032 {
1033         unsigned int i, j, h;
1034         struct keyinfo *hash;
1035
1036         hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
1037         for (i = 0; i < num; i++) {
1038                 for (j = 1; j < num_ops[i]; j++) {
1039                         /* We can't do this on allocation, due to realloc. */
1040                         list_head_init(&op[i][j].post);
1041                         list_head_init(&op[i][j].pre);
1042
1043                         if (!op[i][j].key.dptr)
1044                                 continue;
1045
1046                         h = hash_key(&op[i][j].key) % (total_keys * 2);
1047                         while (!key_eq(hash[h].key, op[i][j].key)) {
1048                                 if (!hash[h].key.dptr) {
1049                                         hash[h].key = op[i][j].key;
1050                                         break;
1051                                 }
1052                                 h = (h + 1) % (total_keys * 2);
1053                         }
1054                         /* Might as well save some memory if we can. */
1055                         if (op[i][j].key.dptr != hash[h].key.dptr) {
1056                                 talloc_free(op[i][j].key.dptr);
1057                                 op[i][j].key.dptr = hash[h].key.dptr;
1058                         }
1059                         hash[h].user = talloc_realloc(hash, hash[h].user,
1060                                                      struct key_user,
1061                                                      hash[h].num_users+1);
1062
1063                         /* If it's in a transaction, it's the transaction which
1064                          * matters from an analysis POV. */
1065                         if (in_transaction(op[i], j)
1066                             || in_chainlock(op[i], j)) {
1067                                 unsigned start = op[i][j].group_start;
1068
1069                                 /* Don't include twice. */
1070                                 if (hash[h].num_users
1071                                     && hash[h].user[hash[h].num_users-1].file
1072                                         == i
1073                                     && hash[h].user[hash[h].num_users-1].op_num
1074                                         == start)
1075                                         continue;
1076
1077                                 hash[h].user[hash[h].num_users].op_num = start;
1078                         } else
1079                                 hash[h].user[hash[h].num_users].op_num = j;
1080                         hash[h].user[hash[h].num_users].file = i;
1081                         hash[h].num_users++;
1082                 }
1083         }
1084
1085         return hash;
1086 }
1087
1088 static bool satisfies(const TDB_DATA *key, const TDB_DATA *data,
1089                       const struct op *op)
1090 {
1091         const TDB_DATA *need = NULL;
1092
1093         if (starts_transaction(op) || starts_chainlock(op)) {
1094                 unsigned int i;
1095
1096                 /* Look through for an op in this transaction which
1097                  * needs this key. */
1098                 for (i = 1; i < op->group_len; i++) {
1099                         if (key_eq(op[i].key, *key)) {
1100                                 need = needs(&op[i]);
1101                                 /* tdb_exists() is special: there might be
1102                                  * something in the transaction with more
1103                                  * specific requirements.  Other ops don't have
1104                                  * specific requirements (eg. store or delete),
1105                                  * but they change the value so we can't get
1106                                  * more information from future ops. */
1107                                 if (op[i].op != OP_TDB_EXISTS)
1108                                         break;
1109                         }
1110                 }
1111         } else
1112                 need = needs(op);
1113
1114         /* Don't need anything?  Cool. */
1115         if (!need)
1116                 return true;
1117
1118         /* This should be tdb_null or a real value. */
1119         assert(data != &must_exist);
1120         assert(data != &must_not_exist);
1121         assert(data != &not_exists_or_empty);
1122
1123         /* Must not exist?  data must not exist. */
1124         if (need == &must_not_exist)
1125                 return data == &tdb_null;
1126
1127         /* Must exist? */
1128         if (need == &must_exist)
1129                 return data != &tdb_null;
1130
1131         /* Either noexist or empty. */
1132         if (need == &not_exists_or_empty)
1133                 return data->dsize == 0;
1134
1135         /* Needs something specific. */
1136         return key_eq(*data, *need);
1137 }
1138
1139 static void move_to_front(struct key_user res[], unsigned off, unsigned elem)
1140 {
1141         if (elem != off) {
1142                 struct key_user tmp = res[elem];
1143                 memmove(res + off + 1, res + off, (elem - off)*sizeof(res[0]));
1144                 res[off] = tmp;
1145         }
1146 }
1147
1148 static void restore_to_pos(struct key_user res[], unsigned off, unsigned elem)
1149 {
1150         if (elem != off) {
1151                 struct key_user tmp = res[off];
1152                 memmove(res + off, res + off + 1, (elem - off)*sizeof(res[0]));
1153                 res[elem] = tmp;
1154         }
1155 }
1156
1157 static bool sort_deps(char *filename[], struct op *op[],
1158                       struct key_user res[],
1159                       unsigned off, unsigned num,
1160                       const TDB_DATA *key, const TDB_DATA *data,
1161                       unsigned num_files, unsigned fuzz)
1162 {
1163         unsigned int i, files_done;
1164         struct op *this_op;
1165         bool done[num_files];
1166
1167         /* None left?  We're sorted. */
1168         if (off == num)
1169                 return true;
1170
1171         /* Does this make serial numbers go backwards?  Allow a little fuzz. */
1172         if (off > 0) {
1173                 int serial1 = op[res[off-1].file][res[off-1].op_num].serial;
1174                 int serial2 = op[res[off].file][res[off].op_num].serial;
1175
1176                 if (serial1 - serial2 > (int)fuzz) {
1177 #if DEBUG_DEPS
1178                         printf("Serial jump too far (%u -> %u)\n",
1179                                serial1, serial2);
1180 #endif
1181                         return false;
1182                 }
1183         }
1184
1185         memset(done, 0, sizeof(done));
1186
1187         /* Since ops within a trace file are ordered, we just need to figure
1188          * out which file to try next.  Since we don't take into account
1189          * inter-key relationships (which exist by virtue of trace file order),
1190          * we minimize the chance of harm by trying to keep in serial order. */
1191         for (files_done = 0, i = off; i < num && files_done < num_files; i++) {
1192                 if (done[res[i].file])
1193                         continue;
1194
1195                 this_op = &op[res[i].file][res[i].op_num];
1196
1197                 /* Is what we have good enough for this op? */
1198                 if (satisfies(key, data, this_op)) {
1199                         move_to_front(res, off, i);
1200                         if (sort_deps(filename, op, res, off+1, num,
1201                                       key, gives(key, data, this_op),
1202                                       num_files, fuzz))
1203                                 return true;
1204                         restore_to_pos(res, off, i);
1205                 }
1206                 done[res[i].file] = true;
1207                 files_done++;
1208         }
1209
1210         /* No combination worked. */
1211         return false;
1212 }
1213
1214 static void check_dep_sorting(struct key_user user[], unsigned num_users,
1215                               unsigned num_files)
1216 {
1217 #if DEBUG_DEPS
1218         unsigned int i;
1219         unsigned minima[num_files];
1220
1221         memset(minima, 0, sizeof(minima));
1222         for (i = 0; i < num_users; i++) {
1223                 assert(minima[user[i].file] < user[i].op_num);
1224                 minima[user[i].file] = user[i].op_num;
1225         }
1226 #endif
1227 }
1228
1229 /* All these ops happen on the same key.  Which comes first?
1230  *
1231  * This can happen both because read ops or failed write ops don't
1232  * change serial number, and also due to race since we access the
1233  * number unlocked (the race can cause less detectable ordering problems,
1234  * in which case we'll deadlock and report: fix manually in that case).
1235  */
1236 static void figure_deps(char *filename[], struct op *op[],
1237                         const TDB_DATA *key, struct key_user user[],
1238                         unsigned num_users, unsigned num_files)
1239 {
1240         /* We assume database starts empty. */
1241         const struct TDB_DATA *data = &tdb_null;
1242         unsigned int fuzz;
1243
1244         /* We prefer to keep strict serial order if possible: it's the
1245          * most likely.  We get more lax if that fails. */
1246         for (fuzz = 0; fuzz < 100; fuzz = (fuzz + 1)*2) {
1247                 if (sort_deps(filename, op, user, 0, num_users, key, data,
1248                               num_files, fuzz))
1249                         break;
1250         }
1251
1252         if (fuzz >= 100)
1253                 fail(filename[user[0].file], user[0].op_num+1,
1254                      "Could not resolve inter-dependencies");
1255
1256         check_dep_sorting(user, num_users, num_files);
1257 }
1258
1259 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1260                      unsigned int num)
1261 {
1262         unsigned int h;
1263
1264         /* Gcc nexted function extension.  How cool is this? */
1265         int compare_serial(const void *_a, const void *_b)
1266         {
1267                 const struct key_user *a = _a, *b = _b;
1268
1269                 /* First, maintain order within any trace file. */
1270                 if (a->file == b->file)
1271                         return a->op_num - b->op_num;
1272
1273                 /* Otherwise, arrange by serial order. */
1274                 if (op[a->file][a->op_num].serial !=
1275                     op[b->file][b->op_num].serial)
1276                         return op[a->file][a->op_num].serial
1277                                 - op[b->file][b->op_num].serial;
1278
1279                 /* Cancelled transactions are assumed to happen first. */
1280                 if (starts_transaction(&op[a->file][a->op_num])
1281                     && !successful_transaction(&op[a->file][a->op_num]))
1282                         return -1;
1283                 if (starts_transaction(&op[b->file][b->op_num])
1284                     && !successful_transaction(&op[b->file][b->op_num]))
1285                         return 1;
1286
1287                 /* No idea. */
1288                 return 0;
1289         }
1290
1291         /* Now sort into serial order. */
1292         for (h = 0; h < total_keys * 2; h++) {
1293                 struct key_user *user = hash[h].user;
1294
1295                 qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
1296                 figure_deps(filename, op, &hash[h].key, user, hash[h].num_users,
1297                             num);
1298         }
1299 }
1300
1301 static int destroy_depend(struct depend *dep)
1302 {
1303         list_del(&dep->pre_list);
1304         list_del(&dep->post_list);
1305         return 0;
1306 }
1307
1308 static void add_dependency(void *ctx,
1309                            struct op *op[],
1310                            char *filename[],
1311                            unsigned int needs_file,
1312                            unsigned int needs_opnum,
1313                            unsigned int satisfies_file,
1314                            unsigned int satisfies_opnum)
1315 {
1316         struct depend *dep;
1317
1318         /* We don't depend on ourselves. */
1319         if (needs_file == satisfies_file) {
1320                 assert(satisfies_opnum < needs_opnum);
1321                 return;
1322         }
1323
1324 #if DEBUG_DEPS
1325         printf("%s:%u: depends on %s:%u\n",
1326                filename[needs_file], needs_opnum+1,
1327                filename[satisfies_file], satisfies_opnum+1);
1328 #endif
1329
1330 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1331         /* If something in a traverse depends on something in another
1332          * traverse/transaction, it creates a dependency between the
1333          * two groups. */
1334         if ((in_traverse(op[satisfies_file], satisfies_opnum)
1335              && (starts_transaction(&op[needs_file][needs_opnum])
1336                  || starts_traverse(&op[needs_file][needs_opnum])))
1337             || (in_traverse(op[needs_file], needs_opnum)
1338                 && (starts_transaction(&op[satisfies_file][satisfies_opnum])
1339                     || starts_traverse(&op[satisfies_file][satisfies_opnum])))){
1340                 unsigned int sat;
1341
1342                 /* We are satisfied by end of group. */
1343                 sat = op[satisfies_file][satisfies_opnum].group_start;
1344                 satisfies_opnum = sat + op[satisfies_file][sat].group_len;
1345                 /* And we need that done by start of our group. */
1346                 needs_opnum = op[needs_file][needs_opnum].group_start;
1347         }
1348
1349         /* There is also this case:
1350          *  <traverse> <read foo> ...
1351          *  <transaction> ... </transaction> <create foo>
1352          * Where if we start the traverse then wait, we could block
1353          * the transaction and deadlock.
1354          *
1355          * We try to address this by ensuring that where seqnum indicates it's
1356          * possible, we wait for <create foo> before *starting* traverse.
1357          */
1358         else if (in_traverse(op[needs_file], needs_opnum)) {
1359                 struct op *need = &op[needs_file][needs_opnum];
1360                 if (op[needs_file][need->group_start].serial >
1361                     op[satisfies_file][satisfies_opnum].serial) {
1362                         needs_opnum = need->group_start;
1363                 }
1364         }
1365 #endif
1366
1367         /* If you depend on a transaction or chainlock, you actually
1368          * depend on it ending. */
1369         if (starts_transaction(&op[satisfies_file][satisfies_opnum])
1370             || starts_chainlock(&op[satisfies_file][satisfies_opnum])) {
1371                 satisfies_opnum
1372                         += op[satisfies_file][satisfies_opnum].group_len;
1373 #if DEBUG_DEPS
1374                 printf("-> Actually end of transaction %s:%u\n",
1375                        filename[satisfies_file], satisfies_opnum+1);
1376 #endif
1377         } else
1378                 /* We should never create a dependency from middle of
1379                  * a transaction. */
1380                 assert(!in_transaction(op[satisfies_file], satisfies_opnum)
1381                        || op[satisfies_file][satisfies_opnum].op
1382                        == OP_TDB_TRANSACTION_COMMIT
1383                        || op[satisfies_file][satisfies_opnum].op
1384                        == OP_TDB_TRANSACTION_CANCEL);
1385
1386         assert(op[needs_file][needs_opnum].op != OP_TDB_TRAVERSE);
1387         assert(op[satisfies_file][satisfies_opnum].op != OP_TDB_TRAVERSE);
1388
1389         dep = talloc(ctx, struct depend);
1390         dep->needs_file = needs_file;
1391         dep->needs_opnum = needs_opnum;
1392         dep->satisfies_file = satisfies_file;
1393         dep->satisfies_opnum = satisfies_opnum;
1394         list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
1395         list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
1396         talloc_set_destructor(dep, destroy_depend);
1397 }
1398
1399 static bool changes_db(const TDB_DATA *key, const struct op *op)
1400 {
1401         return gives(key, NULL, op) != NULL;
1402 }
1403
1404 static void depend_on_previous(struct op *op[],
1405                                char *filename[],
1406                                unsigned int num,
1407                                struct key_user user[],
1408                                unsigned int i,
1409                                int prev)
1410 {
1411         bool deps[num];
1412         int j;
1413
1414         if (i == 0)
1415                 return;
1416
1417         if (prev == i - 1) {
1418                 /* Just depend on previous. */
1419                 add_dependency(NULL, op, filename,
1420                                user[i].file, user[i].op_num,
1421                                user[prev].file, user[prev].op_num);
1422                 return;
1423         }
1424
1425         /* We have to wait for the readers.  Find last one in *each* file. */
1426         memset(deps, 0, sizeof(deps));
1427         deps[user[i].file] = true;
1428         for (j = i - 1; j > prev; j--) {
1429                 if (!deps[user[j].file]) {
1430                         add_dependency(NULL, op, filename,
1431                                        user[i].file, user[i].op_num,
1432                                        user[j].file, user[j].op_num);
1433                         deps[user[j].file] = true;
1434                 }
1435         }
1436 }
1437
1438 /* This is simple, but not complete.  We don't take into account
1439  * indirect dependencies. */
1440 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1441                                   unsigned int num)
1442 {
1443         unsigned int i, j;
1444
1445         /* There can only be one real dependency on each file */
1446         for (i = 0; i < num; i++) {
1447                 for (j = 1; j < num_ops[i]; j++) {
1448                         struct depend *dep, *next;
1449                         struct depend *prev[num];
1450
1451                         memset(prev, 0, sizeof(prev));
1452
1453                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1454                                 if (!prev[dep->satisfies_file]) {
1455                                         prev[dep->satisfies_file] = dep;
1456                                         continue;
1457                                 }
1458                                 if (prev[dep->satisfies_file]->satisfies_opnum
1459                                     < dep->satisfies_opnum) {
1460                                         talloc_free(prev[dep->satisfies_file]);
1461                                         prev[dep->satisfies_file] = dep;
1462                                 } else
1463                                         talloc_free(dep);
1464                         }
1465                 }
1466         }
1467
1468         for (i = 0; i < num; i++) {
1469                 int deps[num];
1470
1471                 for (j = 0; j < num; j++)
1472                         deps[j] = -1;
1473
1474                 for (j = 1; j < num_ops[i]; j++) {
1475                         struct depend *dep, *next;
1476
1477                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1478                                 if (deps[dep->satisfies_file]
1479                                     >= (int)dep->satisfies_opnum)
1480                                         talloc_free(dep);
1481                                 else
1482                                         deps[dep->satisfies_file]
1483                                                 = dep->satisfies_opnum;
1484                         }
1485                 }
1486         }
1487 }
1488
1489 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1490 struct traverse_dep {
1491         unsigned int file;
1492         unsigned int op_num;
1493 };
1494
1495 /* Force an order among the traversals, so they don't deadlock (as much) */
1496 static void make_traverse_depends(char *filename[],
1497                                   struct op *op[], unsigned int num_ops[],
1498                                   unsigned int num)
1499 {
1500         unsigned int i, num_traversals = 0;
1501         int j;
1502         struct traverse_dep *dep;
1503
1504         /* Sort by which one runs first. */
1505         int compare_traverse_dep(const void *_a, const void *_b)
1506         {
1507                 const struct traverse_dep *ta = _a, *tb = _b;
1508                 const struct op *a = &op[ta->file][ta->op_num],
1509                         *b = &op[tb->file][tb->op_num];
1510
1511                 if (a->serial != b->serial)
1512                         return a->serial - b->serial;
1513
1514                 /* If they have same serial, it means one didn't make any
1515                  * changes.  Thus sort by end in that case. */
1516                 return a[a->group_len].serial - b[b->group_len].serial;
1517         }
1518
1519         dep = talloc_array(NULL, struct traverse_dep, 1);
1520
1521         /* Count them. */
1522         for (i = 0; i < num; i++) {
1523                 for (j = 1; j < num_ops[i]; j++) {
1524                         /* Traverse start (ignore those in
1525                          * transactions; they're already covered by
1526                          * transaction dependencies). */
1527                         if (starts_traverse(&op[i][j])
1528                             && !in_transaction(op[i], j)) {
1529                                 dep = talloc_realloc(NULL, dep,
1530                                                      struct traverse_dep,
1531                                                      num_traversals+1);
1532                                 dep[num_traversals].file = i;
1533                                 dep[num_traversals].op_num = j;
1534                                 num_traversals++;
1535                         }
1536                 }
1537         }
1538         qsort(dep, num_traversals, sizeof(dep[0]), compare_traverse_dep);
1539
1540         for (i = 1; i < num_traversals; i++) {
1541                 const struct op *prev = &op[dep[i-1].file][dep[i-1].op_num];
1542                 const struct op *curr = &op[dep[i].file][dep[i].op_num];
1543
1544                 /* Read traverses don't depend on each other (read lock). */
1545                 if (prev->op == OP_TDB_TRAVERSE_READ_START
1546                     && curr->op == OP_TDB_TRAVERSE_READ_START)
1547                         continue;
1548
1549                 /* Only make dependency if it's clear. */
1550                 if (compare_traverse_dep(&dep[i], &dep[i-1])) {
1551                         /* i depends on end of traverse i-1. */
1552                         add_dependency(NULL, op, filename,
1553                                        dep[i].file, dep[i].op_num,
1554                                        dep[i-1].file, dep[i-1].op_num
1555                                        + prev->group_len);
1556                 }
1557         }
1558         talloc_free(dep);
1559 }
1560 #endif
1561
1562 static void derive_dependencies(char *filename[],
1563                                 struct op *op[], unsigned int num_ops[],
1564                                 unsigned int num)
1565 {
1566         struct keyinfo *hash;
1567         unsigned int h, i;
1568
1569         /* Create hash table for faster key lookup. */
1570         hash = hash_ops(op, num_ops, num);
1571
1572         /* Sort them by serial number. */
1573         sort_ops(hash, filename, op, num);
1574
1575         /* Create dependencies back to the last change, rather than
1576          * creating false dependencies by naively making each one
1577          * depend on the previous.  This has two purposes: it makes
1578          * later optimization simpler, and it also avoids deadlock with
1579          * same sequence number ops inside traversals (if one
1580          * traversal doesn't write anything, two ops can have the same
1581          * sequence number yet we can create a traversal dependency
1582          * the other way). */
1583         for (h = 0; h < total_keys * 2; h++) {
1584                 int prev = -1;
1585
1586                 if (hash[h].num_users < 2)
1587                         continue;
1588
1589                 for (i = 0; i < hash[h].num_users; i++) {
1590                         if (changes_db(&hash[h].key, &op[hash[h].user[i].file]
1591                                        [hash[h].user[i].op_num])) {
1592                                 depend_on_previous(op, filename, num,
1593                                                    hash[h].user, i, prev);
1594                                 prev = i;
1595                         } else if (prev >= 0)
1596                                 add_dependency(hash, op, filename,
1597                                                hash[h].user[i].file,
1598                                                hash[h].user[i].op_num,
1599                                                hash[h].user[prev].file,
1600                                                hash[h].user[prev].op_num);
1601                 }
1602         }
1603
1604 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1605         make_traverse_depends(filename, op, num_ops, num);
1606 #endif
1607
1608         optimize_dependencies(op, num_ops, num);
1609 }
1610
1611 int main(int argc, char *argv[])
1612 {
1613         struct timeval start, end;
1614         unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1615         struct op *op[argc];
1616         int fds[2];
1617         char c;
1618         bool ok = true;
1619
1620         if (argc < 3)
1621                 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1622
1623         pipes = talloc_array(NULL, struct pipe, argc - 2);
1624         for (i = 0; i < argc - 2; i++) {
1625                 printf("Loading tracefile %s...", argv[2+i]);
1626                 fflush(stdout);
1627                 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1628                                        &tdb_flags[i], &open_flags[i]);
1629                 if (pipe(pipes[i].fd) != 0)
1630                         err(1, "creating pipe");
1631                 printf("done\n");
1632         }
1633
1634         printf("Calculating inter-dependencies...");
1635         fflush(stdout);
1636         derive_dependencies(argv+2, op, num_ops, i);
1637         printf("done\n");
1638
1639         /* Don't fork for single arg case: simple debugging. */
1640         if (argc == 3) {
1641                 struct tdb_context *tdb;
1642                 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0]|TDB_NOSYNC,
1643                                   open_flags[0], 0600, NULL, hash_key);
1644                 printf("Single threaded run...");
1645                 fflush(stdout);
1646
1647                 run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0],
1648                         false);
1649                 check_deps(argv[2], op[0], num_ops[0]);
1650
1651                 printf("done\n");
1652                 exit(0);
1653         }
1654
1655         if (pipe(fds) != 0)
1656                 err(1, "creating pipe");
1657
1658         for (i = 0; i < argc - 2; i++) {
1659                 struct tdb_context *tdb;
1660
1661                 switch (fork()) {
1662                 case -1:
1663                         err(1, "fork failed");
1664                 case 0:
1665                         close(fds[1]);
1666                         tdb = tdb_open_ex(argv[1], hashsize[i],
1667                                           tdb_flags[i]|TDB_NOSYNC,
1668                                           open_flags[i], 0600, NULL, hash_key);
1669                         if (!tdb)
1670                                 err(1, "Opening tdb %s", argv[1]);
1671
1672                         /* This catches parent exiting. */
1673                         if (read(fds[0], &c, 1) != 1)
1674                                 exit(1);
1675                         run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
1676                                 num_ops[i], false);
1677                         check_deps(argv[2+i], op[i], num_ops[i]);
1678                         exit(0);
1679                 default:
1680                         break;
1681                 }
1682         }
1683
1684         /* Let everything settle. */
1685         sleep(1);
1686
1687         printf("Starting run...");
1688         fflush(stdout);
1689         gettimeofday(&start, NULL);
1690         /* Tell them all to go!  Any write of sufficient length will do. */
1691         if (write(fds[1], hashsize, i) != i)
1692                 err(1, "Writing to wakeup pipe");
1693
1694         for (i = 0; i < argc - 2; i++) {
1695                 int status;
1696                 wait(&status);
1697                 if (!WIFEXITED(status)) {
1698                         warnx("Child died with signal %i", WTERMSIG(status));
1699                         ok = false;
1700                 } else if (WEXITSTATUS(status) != 0)
1701                         /* Assume child spat out error. */
1702                         ok = false;
1703         }
1704         if (!ok)
1705                 exit(1);
1706
1707         gettimeofday(&end, NULL);
1708         printf("done\n");
1709
1710         end.tv_sec -= start.tv_sec;
1711         printf("Time replaying: %lu usec\n",
1712                end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));
1713
1714         exit(0);
1715 }