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 traverse *trav; /* traverse start */
 147                 struct {  /* append */
 148                         TDB_DATA pre;
 149                         TDB_DATA post;
 150                 } append;
 151                 unsigned int transaction_end; /* transaction start */
 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         if (op[op_num].op != OP_TDB_TRAVERSE)
 218                 total_keys++;
 219 }
 220
 221 static void op_add_key_ret(const char *filename,
 222                            struct op op[], unsigned int op_num, char *words[])
 223 {
 224         if (!words[2] || !words[3] || !words[4] || words[5]
 225             || !streq(words[3], "="))
 226                 fail(filename, op_num+1, "Expected <key> = <ret>");
 227         op[op_num].ret = atoi(words[4]);
 228         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 229         /* May only be a unique key if it fails */
 230         if (op[op_num].ret != 0)
 231                 total_keys++;
 232 }
 233
 234 static void op_add_key_data(const char *filename,
 235                             struct op op[], unsigned int op_num, char *words[])
 236 {
 237         if (!words[2] || !words[3] || !words[4] || words[5]
 238             || !streq(words[3], "="))
 239                 fail(filename, op_num+1, "Expected <key> = <data>");
 240         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 241         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[4]);
 242         /* May only be a unique key if it fails */
 243         if (!op[op_num].data.dptr)
 244                 total_keys++;
 245 }
 246
 247 /* <serial> tdb_store <rec> <rec> <flag> = <ret> */
 248 static void op_add_store(const char *filename,
 249                          struct op op[], unsigned int op_num, char *words[])
 250 {
 251         if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
 252             || words[7] || !streq(words[5], "="))
 253                 fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
 254
 255         op[op_num].flag = strtoul(words[4], NULL, 0);
 256         op[op_num].ret = atoi(words[6]);
 257         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 258         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 259         total_keys++;
 260 }
 261
 262 /* <serial> tdb_append <rec> <rec> = <rec> */
 263 static void op_add_append(const char *filename,
 264                           struct op op[], unsigned int op_num, char *words[])
 265 {
 266         if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
 267             || !streq(words[4], "="))
 268                 fail(filename, op_num+1, "Expect <key> <data> = <rec>");
 269
 270         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 271         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 272
 273         op[op_num].append.post
 274                 = make_tdb_data(op, filename, op_num+1, words[5]);
 275
 276         /* By subtraction, figure out what previous data was. */
 277         op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
 278         op[op_num].append.pre.dsize
 279                 = op[op_num].append.post.dsize - op[op_num].data.dsize;
 280         total_keys++;
 281 }
 282
 283 /* <serial> tdb_get_seqnum = <ret> */
 284 static void op_add_seqnum(const char *filename,
 285                           struct op op[], unsigned int op_num, char *words[])
 286 {
 287         if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
 288                 fail(filename, op_num+1, "Expect = <ret>");
 289
 290         op[op_num].key = tdb_null;
 291         op[op_num].ret = atoi(words[3]);
 292 }
 293
 294 static void op_add_traverse(const char *filename,
 295                             struct op op[], unsigned int op_num, char *words[])
 296 {
 297         if (words[2])
 298                 fail(filename, op_num+1, "Expect no arguments");
 299
 300         op[op_num].key = tdb_null;
 301         op[op_num].trav = NULL;
 302 }
 303
 304 static void op_add_transaction(const char *filename, 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].transaction_end = 0;
 312 }
 313
 314 static void op_analyze_transaction(const char *filename,
 315                                    struct op op[], unsigned int op_num,
 316                                    char *words[])
 317 {
 318         int i, start;
 319
 320         op[op_num].key = tdb_null;
 321
 322         if (words[2])
 323                 fail(filename, op_num+1, "Expect no arguments");
 324
 325         for (i = op_num-1; i >= 0; i--) {
 326                 if (op[i].op == OP_TDB_TRANSACTION_START &&
 327                     !op[i].transaction_end)
 328                         break;
 329         }
 330
 331         if (i < 0)
 332                 fail(filename, op_num+1, "no transaction start found");
 333
 334         start = i;
 335         op[start].transaction_end = op_num;
 336
 337         /* This rolls in nested transactions.  I think that's right. */
 338         for (i++; i <= op_num; i++)
 339                 op[i].group_start = start;
 340 }
 341
 342 struct traverse_hash {
 343         TDB_DATA key;
 344         unsigned int index;
 345 };
 346
 347 /* A traverse is a hash of keys, each one associated with ops. */
 348 struct traverse {
 349         /* How many traversal callouts should I do? */
 350         unsigned int num;
 351
 352         /* Where is traversal end op? */
 353         unsigned int end;
 354
 355         /* For trivial traversals. */
 356         struct traverse_hash *hash;
 357 };
 358
 359 /* A trivial traversal is one which doesn't terminate early and only
 360  * plays with its own record.  We can reliably replay these even if
 361  * traverse order changes. */
 362 static bool is_trivial_traverse(struct op op[], unsigned int end)
 363 {
 364 #if 0
 365         unsigned int i;
 366         TDB_DATA cur = tdb_null;
 367
 368         if (op[end].ret != 0)
 369                 return false;
 370
 371         for (i = 0; i < end; i++) {
 372                 if (!op[i].key.dptr)
 373                         continue;
 374                 if (op[i].op == OP_TDB_TRAVERSE)
 375                         cur = op[i].key;
 376                 if (!key_eq(cur, op[i].key))
 377                         return false;
 378         }
 379         return true;
 380 #endif
 381         /* With multiple things happening at once, no traverse is trivial. */
 382         return false;
 383 }
 384
 385 static void op_analyze_traverse(const char *filename,
 386                                 struct op op[], unsigned int op_num,
 387                                 char *words[])
 388 {
 389         int i, start;
 390         struct traverse *trav = talloc(op, struct traverse);
 391
 392         op[op_num].key = tdb_null;
 393
 394         /* = %u means traverse function terminated. */
 395         if (words[2]) {
 396                 if (!streq(words[2], "=") || !words[3] || words[4])
 397                         fail(filename, op_num+1, "expect = <num>");
 398                 op[op_num].ret = atoi(words[3]);
 399         } else
 400                 op[op_num].ret = 0;
 401
 402         trav->num = 0;
 403         trav->end = op_num;
 404         for (i = op_num-1; i >= 0; i--) {
 405                 if (op[i].op == OP_TDB_TRAVERSE)
 406                         trav->num++;
 407                 if (op[i].op != OP_TDB_TRAVERSE_READ_START
 408                     && op[i].op != OP_TDB_TRAVERSE_START)
 409                         continue;
 410                 if (op[i].trav)
 411                         continue;
 412                 break;
 413         }
 414
 415         if (i < 0)
 416                 fail(filename, op_num+1, "no traversal start found");
 417
 418         start = i;
 419         op[start].trav = trav;
 420
 421         for (i = start; i <= op_num; i++)
 422                 op[i].group_start = start;
 423
 424         if (is_trivial_traverse(op+i, op_num-i)) {
 425                 /* Fill in a plentiful hash table. */
 426                 op[start].trav->hash = talloc_zero_array(op[i].trav,
 427                                                          struct traverse_hash,
 428                                                          trav->num * 2);
 429                 for (i = start; i < op_num; i++) {
 430                         unsigned int h;
 431                         if (op[i].op != OP_TDB_TRAVERSE)
 432                                 continue;
 433                         h = hash_key(&op[i].key) % (trav->num * 2);
 434                         while (trav->hash[h].index)
 435                                 h = (h + 1) % (trav->num * 2);
 436                         trav->hash[h].index = i+1;
 437                         trav->hash[h].key = op[i].key;
 438                 }
 439         } else
 440                 trav->hash = NULL;
 441 }
 442
 443 /* Keep -Wmissing-declarations happy: */
 444 const struct op_table *
 445 find_keyword (register const char *str, register unsigned int len);
 446
 447 #include "keywords.c"
 448
 449 struct depend {
 450         /* We can have more than one */
 451         struct list_node pre_list;
 452         struct list_node post_list;
 453         unsigned int needs_file;
 454         unsigned int needs_opnum;
 455         unsigned int satisfies_file;
 456         unsigned int satisfies_opnum;
 457 };
 458
 459 static void check_deps(const char *filename, struct op op[], unsigned int num)
 460 {
 461 #ifdef DEBUG_DEPS
 462         unsigned int i;
 463
 464         for (i = 1; i < num; i++)
 465                 if (!list_empty(&op[i].pre))
 466                         fail(filename, i+1, "Still has dependencies");
 467 #endif
 468 }
 469
 470 static void dump_pre(char *filename[], unsigned int file,
 471                      struct op op[], unsigned int i)
 472 {
 473         struct depend *dep;
 474
 475         printf("%s:%u still waiting for:\n", filename[file], i+1);
 476         list_for_each(&op[i].pre, dep, pre_list)
 477                 printf("    %s:%u\n",
 478                        filename[dep->satisfies_file], dep->satisfies_opnum+1);
 479         check_deps(filename[file], op, i);
 480 }
 481
 482 /* We simply read/write pointers, since we all are children. */
 483 static void do_pre(char *filename[], unsigned int file, int pre_fd,
 484                    struct op op[], unsigned int i)
 485 {
 486         while (!list_empty(&op[i].pre)) {
 487                 struct depend *dep;
 488
 489 #if DEBUG_DEPS
 490                 printf("%s:%u:waiting for pre\n", filename[file], i+1);
 491                 fflush(stdout);
 492 #endif
 493                 alarm(10);
 494                 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
 495                         if (errno == EINTR) {
 496                                 dump_pre(filename, file, op, i);
 497                                 exit(1);
 498                         } else
 499                                 errx(1, "Reading from pipe");
 500                 }
 501                 alarm(0);
 502
 503 #if DEBUG_DEPS
 504                 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
 505                        dep->needs_opnum+1, filename[dep->satisfies_file],
 506                        dep->satisfies_opnum+1);
 507                 fflush(stdout);
 508 #endif
 509                 /* This could be any op, not just this one. */
 510                 talloc_free(dep);
 511         }
 512 }
 513
 514 static void do_post(char *filename[], unsigned int file,
 515                     const struct op op[], unsigned int i)
 516 {
 517         struct depend *dep;
 518
 519         list_for_each(&op[i].post, dep, post_list) {
 520 #if DEBUG_DEPS
 521                 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
 522                        filename[dep->needs_file], dep->needs_opnum+1);
 523 #endif
 524                 if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
 525                     != sizeof(dep))
 526                         err(1, "%s:%u failed to tell file %s",
 527                             filename[file], i+1, filename[dep->needs_file]);
 528         }
 529 }
 530
 531 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 532 {
 533         return data.dsize;
 534 }
 535
 536 static unsigned run_ops(struct tdb_context *tdb,
 537                         int pre_fd,
 538                         char *filename[],
 539                         unsigned int file,
 540                         struct op op[],
 541                         unsigned int start, unsigned int stop);
 542
 543 struct traverse_info {
 544         struct op *op;
 545         char **filename;
 546         unsigned file;
 547         int pre_fd;
 548         unsigned int start;
 549         unsigned int i;
 550 };
 551
 552 /* Trivial case: do whatever they did for this key. */
 553 static int trivial_traverse(struct tdb_context *tdb,
 554                             TDB_DATA key, TDB_DATA data,
 555                             void *_tinfo)
 556 {
 557         struct traverse_info *tinfo = _tinfo;
 558         struct traverse *trav = tinfo->op[tinfo->start].trav;
 559         unsigned int h = hash_key(&key) % (trav->num * 2);
 560
 561         while (trav->hash[h].index) {
 562                 if (key_eq(trav->hash[h].key, key)) {
 563                         run_ops(tdb, tinfo->pre_fd, tinfo->filename,
 564                                 tinfo->file, tinfo->op, trav->hash[h].index,
 565                                 trav->end);
 566                         tinfo->i++;
 567                         return 0;
 568                 }
 569                 h = (h + 1) % (trav->num * 2);
 570         }
 571         fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 572              "unexpected traverse key");
 573 }
 574
 575 /* More complex.  Just do whatever's they did at the n'th entry. */
 576 static int nontrivial_traverse(struct tdb_context *tdb,
 577                                TDB_DATA key, TDB_DATA data,
 578                                void *_tinfo)
 579 {
 580         struct traverse_info *tinfo = _tinfo;
 581         struct traverse *trav = tinfo->op[tinfo->start].trav;
 582
 583         if (tinfo->i == trav->end) {
 584                 /* This can happen if traverse expects to be empty. */
 585                 if (tinfo->start + 1 == trav->end)
 586                         return 1;
 587                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 588                      "traverse did not terminate");
 589         }
 590
 591         if (tinfo->op[tinfo->i].op != OP_TDB_TRAVERSE)
 592                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 593                      "%s:%u:traverse terminated early");
 594
 595         /* Run any normal ops. */
 596         tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->file,
 597                            tinfo->op, tinfo->i+1, trav->end);
 598
 599         if (tinfo->i == trav->end)
 600                 return 1;
 601
 602         return 0;
 603 }
 604
 605 static unsigned op_traverse(struct tdb_context *tdb,
 606                             int pre_fd,
 607                             char *filename[],
 608                             unsigned int file,
 609                             int (*traversefn)(struct tdb_context *,
 610                                               tdb_traverse_func, void *),
 611                             struct op op[],
 612                             unsigned int start)
 613 {
 614         struct traverse *trav = op[start].trav;
 615         struct traverse_info tinfo = { op, filename, file, pre_fd,
 616                                        start, start+1 };
 617
 618         /* Trivial case. */
 619         if (trav->hash) {
 620                 int ret = traversefn(tdb, trivial_traverse, &tinfo);
 621                 if (ret != trav->num)
 622                         fail(filename[file], start+1,
 623                              "short traversal %i", ret);
 624                 return trav->end;
 625         }
 626
 627         traversefn(tdb, nontrivial_traverse, &tinfo);
 628
 629         /* Traversing in wrong order can have strange effects: eg. if
 630          * original traverse went A (delete A), B, we might do B
 631          * (delete A).  So if we have ops left over, we do it now. */
 632         while (tinfo.i != trav->end) {
 633                 if (op[tinfo.i].op == OP_TDB_TRAVERSE)
 634                         tinfo.i++;
 635                 else
 636                         tinfo.i = run_ops(tdb, pre_fd, filename, file, op,
 637                                           tinfo.i, trav->end);
 638         }
 639
 640         return trav->end;
 641 }
 642
 643 static void break_out(int sig)
 644 {
 645 }
 646
 647 static __attribute__((noinline))
 648 unsigned run_ops(struct tdb_context *tdb,
 649                  int pre_fd,
 650                  char *filename[],
 651                  unsigned int file,
 652                  struct op op[], unsigned int start, unsigned int stop)
 653 {
 654         unsigned int i;
 655         struct sigaction sa;
 656
 657         sa.sa_handler = break_out;
 658         sa.sa_flags = 0;
 659
 660         sigaction(SIGALRM, &sa, NULL);
 661         for (i = start; i < stop; i++) {
 662                 do_pre(filename, file, pre_fd, op, i);
 663
 664                 switch (op[i].op) {
 665                 case OP_TDB_LOCKALL:
 666                         try(tdb_lockall(tdb), op[i].ret);
 667                         break;
 668                 case OP_TDB_LOCKALL_MARK:
 669                         try(tdb_lockall_mark(tdb), op[i].ret);
 670                         break;
 671                 case OP_TDB_LOCKALL_UNMARK:
 672                         try(tdb_lockall_unmark(tdb), op[i].ret);
 673                         break;
 674                 case OP_TDB_LOCKALL_NONBLOCK:
 675                         unreliable(tdb_lockall_nonblock(tdb), op[i].ret,
 676                                    tdb_lockall(tdb), tdb_unlockall(tdb));
 677                         break;
 678                 case OP_TDB_UNLOCKALL:
 679                         try(tdb_unlockall(tdb), op[i].ret);
 680                         break;
 681                 case OP_TDB_LOCKALL_READ:
 682                         try(tdb_lockall_read(tdb), op[i].ret);
 683                         break;
 684                 case OP_TDB_LOCKALL_READ_NONBLOCK:
 685                         unreliable(tdb_lockall_read_nonblock(tdb), op[i].ret,
 686                                    tdb_lockall_read(tdb),
 687                                    tdb_unlockall_read(tdb));
 688                         break;
 689                 case OP_TDB_UNLOCKALL_READ:
 690                         try(tdb_unlockall_read(tdb), op[i].ret);
 691                         break;
 692                 case OP_TDB_CHAINLOCK:
 693                         try(tdb_chainlock(tdb, op[i].key), op[i].ret);
 694                         break;
 695                 case OP_TDB_CHAINLOCK_NONBLOCK:
 696                         unreliable(tdb_chainlock_nonblock(tdb, op[i].key),
 697                                    op[i].ret,
 698                                    tdb_chainlock(tdb, op[i].key),
 699                                    tdb_chainunlock(tdb, op[i].key));
 700                         break;
 701                 case OP_TDB_CHAINLOCK_MARK:
 702                         try(tdb_chainlock_mark(tdb, op[i].key), op[i].ret);
 703                         break;
 704                 case OP_TDB_CHAINLOCK_UNMARK:
 705                         try(tdb_chainlock_unmark(tdb, op[i].key), op[i].ret);
 706                         break;
 707                 case OP_TDB_CHAINUNLOCK:
 708                         try(tdb_chainunlock(tdb, op[i].key), op[i].ret);
 709                         break;
 710                 case OP_TDB_CHAINLOCK_READ:
 711                         try(tdb_chainlock_read(tdb, op[i].key), op[i].ret);
 712                         break;
 713                 case OP_TDB_CHAINUNLOCK_READ:
 714                         try(tdb_chainunlock_read(tdb, op[i].key), op[i].ret);
 715                         break;
 716                 case OP_TDB_PARSE_RECORD:
 717                         try(tdb_parse_record(tdb, op[i].key, get_len, NULL),
 718                             op[i].ret);
 719                         break;
 720                 case OP_TDB_EXISTS:
 721                         try(tdb_exists(tdb, op[i].key), op[i].ret);
 722                         break;
 723                 case OP_TDB_STORE:
 724                         try(tdb_store(tdb, op[i].key, op[i].data, op[i].flag),
 725                             op[i].ret);
 726                         break;
 727                 case OP_TDB_APPEND:
 728                         try(tdb_append(tdb, op[i].key, op[i].data), op[i].ret);
 729                         break;
 730                 case OP_TDB_GET_SEQNUM:
 731                         try(tdb_get_seqnum(tdb), op[i].ret);
 732                         break;
 733                 case OP_TDB_WIPE_ALL:
 734                         try(tdb_wipe_all(tdb), op[i].ret);
 735                         break;
 736                 case OP_TDB_TRANSACTION_START:
 737                         try(tdb_transaction_start(tdb), op[i].ret);
 738                         break;
 739                 case OP_TDB_TRANSACTION_CANCEL:
 740                         try(tdb_transaction_cancel(tdb), op[i].ret);
 741                         break;
 742                 case OP_TDB_TRANSACTION_COMMIT:
 743                         try(tdb_transaction_commit(tdb), op[i].ret);
 744                         break;
 745                 case OP_TDB_TRAVERSE_READ_START:
 746                         i = op_traverse(tdb, pre_fd, filename, file,
 747                                         tdb_traverse_read, op, i);
 748                         break;
 749                 case OP_TDB_TRAVERSE_START:
 750                         i = op_traverse(tdb, pre_fd, filename, file,
 751                                         tdb_traverse, op, i);
 752                         break;
 753                 case OP_TDB_TRAVERSE:
 754                         /* Terminate: we're in a traverse, and we've
 755                          * done our ops. */
 756                         return i;
 757                 case OP_TDB_TRAVERSE_END:
 758                         fail(filename[file], i+1, "unepxected end traverse");
 759                 /* FIXME: These must be treated like traverse. */
 760                 case OP_TDB_FIRSTKEY:
 761                         if (!key_eq(tdb_firstkey(tdb), op[i].data))
 762                                 fail(filename[file], i+1, "bad firstkey");
 763                         break;
 764                 case OP_TDB_NEXTKEY:
 765                         if (!key_eq(tdb_nextkey(tdb, op[i].key), op[i].data))
 766                                 fail(filename[file], i+1, "bad nextkey");
 767                         break;
 768                 case OP_TDB_FETCH: {
 769                         TDB_DATA f = tdb_fetch(tdb, op[i].key);
 770                         if (!key_eq(f, op[i].data))
 771                                 fail(filename[file], i+1, "bad fetch %u",
 772                                      f.dsize);
 773                         break;
 774                 }
 775                 case OP_TDB_DELETE:
 776                         try(tdb_delete(tdb, op[i].key), op[i].ret);
 777                         break;
 778                 }
 779                 do_post(filename, file, op, i);
 780         }
 781         return i;
 782 }
 783
 784 static struct op *load_tracefile(const char *filename, unsigned int *num,
 785                                  unsigned int *hashsize,
 786                                  unsigned int *tdb_flags,
 787                                  unsigned int *open_flags)
 788 {
 789         unsigned int i;
 790         struct op *op = talloc_array(NULL, struct op, 1);
 791         char **words;
 792         char **lines;
 793         char *file;
 794
 795         file = grab_file(NULL, filename, NULL);
 796         if (!file)
 797                 err(1, "Reading %s", filename);
 798
 799         lines = strsplit(file, file, "\n", NULL);
 800         if (!lines[0])
 801                 errx(1, "%s is empty", filename);
 802
 803         words = strsplit(lines, lines[0], " ", NULL);
 804         if (!streq(words[1], "tdb_open"))
 805                 fail(filename, 1, "does not start with tdb_open");
 806
 807         *hashsize = atoi(words[2]);
 808         *tdb_flags = strtoul(words[3], NULL, 0);
 809         *open_flags = strtoul(words[4], NULL, 0);
 810
 811         for (i = 1; lines[i]; i++) {
 812                 const struct op_table *opt;
 813
 814                 words = strsplit(lines, lines[i], " ", NULL);
 815                 if (!words[0] || !words[1])
 816                         fail(filename, i+1, "Expected serial number and op");
 817
 818                 opt = find_keyword(words[1], strlen(words[1]));
 819                 if (!opt) {
 820                         if (streq(words[1], "tdb_close")) {
 821                                 if (lines[i+1])
 822                                         fail(filename, i+2,
 823                                              "lines after tdb_close");
 824                                 *num = i;
 825                                 talloc_free(lines);
 826                                 return op;
 827                         }
 828                         fail(filename, i+1, "Unknown operation '%s'", words[1]);
 829                 }
 830
 831                 add_op(filename, &op, i, atoi(words[0]), opt->type);
 832                 opt->enhance_op(filename, op, i, words);
 833         }
 834
 835         fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
 836               filename, i);
 837         talloc_free(lines);
 838         *num = i - 1;
 839         return op;
 840 }
 841
 842 /* We remember all the keys we've ever seen, and who has them. */
 843 struct key_user {
 844         unsigned int file;
 845         unsigned int op_num;
 846 };
 847
 848 struct keyinfo {
 849         TDB_DATA key;
 850         unsigned int num_users;
 851         struct key_user *user;
 852 };
 853
 854 static const TDB_DATA must_not_exist;
 855 static const TDB_DATA must_exist;
 856 static const TDB_DATA not_exists_or_empty;
 857
 858 /* NULL means doesn't care if it exists or not, &must_exist means
 859  * it must exist but we don't care what, &must_not_exist means it must
 860  * not exist, otherwise the data it needs. */
 861 static const TDB_DATA *needs(const struct op *op)
 862 {
 863         switch (op->op) {
 864         /* FIXME: Pull forward deps, since we can deadlock */
 865         case OP_TDB_CHAINLOCK:
 866         case OP_TDB_CHAINLOCK_NONBLOCK:
 867         case OP_TDB_CHAINLOCK_MARK:
 868         case OP_TDB_CHAINLOCK_UNMARK:
 869         case OP_TDB_CHAINUNLOCK:
 870         case OP_TDB_CHAINLOCK_READ:
 871         case OP_TDB_CHAINUNLOCK_READ:
 872                 return NULL;
 873
 874         case OP_TDB_APPEND:
 875                 if (op->append.pre.dsize == 0)
 876                         return &not_exists_or_empty;
 877                 return &op->append.pre;
 878
 879         case OP_TDB_STORE:
 880                 if (op->flag == TDB_INSERT) {
 881                         if (op->ret < 0)
 882                                 return &must_exist;
 883                         else
 884                                 return &must_not_exist;
 885                 } else if (op->flag == TDB_MODIFY) {
 886                         if (op->ret < 0)
 887                                 return &must_not_exist;
 888                         else
 889                                 return &must_exist;
 890                 }
 891                 /* No flags?  Don't care */
 892                 return NULL;
 893
 894         case OP_TDB_EXISTS:
 895                 if (op->ret == 1)
 896                         return &must_exist;
 897                 else
 898                         return &must_not_exist;
 899
 900         case OP_TDB_PARSE_RECORD:
 901                 if (op->ret < 0)
 902                         return &must_not_exist;
 903                 return &must_exist;
 904
 905         /* FIXME: handle these. */
 906         case OP_TDB_WIPE_ALL:
 907         case OP_TDB_FIRSTKEY:
 908         case OP_TDB_NEXTKEY:
 909         case OP_TDB_GET_SEQNUM:
 910         case OP_TDB_TRAVERSE:
 911         case OP_TDB_TRANSACTION_COMMIT:
 912         case OP_TDB_TRANSACTION_CANCEL:
 913         case OP_TDB_TRANSACTION_START:
 914                 return NULL;
 915
 916         case OP_TDB_FETCH:
 917                 if (!op->data.dptr)
 918                         return &must_not_exist;
 919                 return &op->data;
 920
 921         case OP_TDB_DELETE:
 922                 if (op->ret < 0)
 923                         return &must_not_exist;
 924                 return &must_exist;
 925
 926         default:
 927                 errx(1, "Unexpected op %i", op->op);
 928         }
 929
 930 }
 931
 932 /* What's the data after this op?  pre if nothing changed. */
 933 static const TDB_DATA *gives(const struct op *op, const TDB_DATA *pre)
 934 {
 935         /* Failed ops don't change state of db. */
 936         if (op->ret < 0)
 937                 return pre;
 938
 939         if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
 940                 return &tdb_null;
 941
 942         if (op->op == OP_TDB_APPEND)
 943                 return &op->append.post;
 944
 945         if (op->op == OP_TDB_STORE)
 946                 return &op->data;
 947
 948         return pre;
 949 }
 950
 951 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 952                                 unsigned int num)
 953 {
 954         unsigned int i, j, h;
 955         struct keyinfo *hash;
 956
 957         hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
 958         for (i = 0; i < num; i++) {
 959                 for (j = 1; j < num_ops[i]; j++) {
 960                         /* We can't do this on allocation, due to realloc. */
 961                         list_head_init(&op[i][j].post);
 962                         list_head_init(&op[i][j].pre);
 963
 964                         if (!op[i][j].key.dptr)
 965                                 continue;
 966
 967                         /* We don't wait for traverse keys */
 968                         /* FIXME: We should, for trivial traversals. */
 969                         if (op[i][j].op == OP_TDB_TRAVERSE)
 970                                 continue;
 971
 972                         h = hash_key(&op[i][j].key) % (total_keys * 2);
 973                         while (!key_eq(hash[h].key, op[i][j].key)) {
 974                                 if (!hash[h].key.dptr) {
 975                                         hash[h].key = op[i][j].key;
 976                                         break;
 977                                 }
 978                                 h = (h + 1) % (total_keys * 2);
 979                         }
 980                         /* Might as well save some memory if we can. */
 981                         if (op[i][j].key.dptr != hash[h].key.dptr) {
 982                                 talloc_free(op[i][j].key.dptr);
 983                                 op[i][j].key.dptr = hash[h].key.dptr;
 984                         }
 985                         hash[h].user = talloc_realloc(hash, hash[h].user,
 986                                                      struct key_user,
 987                                                      hash[h].num_users+1);
 988                         hash[h].user[hash[h].num_users].op_num = j;
 989                         hash[h].user[hash[h].num_users].file = i;
 990                         hash[h].num_users++;
 991                 }
 992         }
 993
 994         return hash;
 995 }
 996
 997 static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
 998 {
 999         /* Don't need anything?  Cool. */
1000         if (!need)
1001                 return true;
1002
1003         /* This should be tdb_null or a real value. */
1004         assert(data != &must_exist);
1005         assert(data != &must_not_exist);
1006         assert(data != &not_exists_or_empty);
1007
1008         /* must_not_exist == must_not_exist, must_exist == must_exist, or
1009            not_exists_or_empty == not_exists_or_empty. */
1010         if (data->dsize == need->dsize && data->dptr == need->dptr)
1011                 return true;
1012
1013         /* Must not exist?  data must not exist. */
1014         if (need == &must_not_exist)
1015                 return data->dptr == NULL;
1016
1017         /* Must exist? */
1018         if (need == &must_exist)
1019                 return data->dptr != NULL;
1020
1021         /* Either noexist or empty. */
1022         if (need == &not_exists_or_empty)
1023                 return data->dsize == 0;
1024
1025         /* Needs something specific. */
1026         return key_eq(*data, *need);
1027 }
1028
1029 static void move_to_front(struct key_user res[], unsigned int elem)
1030 {
1031         if (elem != 0) {
1032                 struct key_user tmp = res[elem];
1033                 memmove(res + 1, res, elem*sizeof(res[0]));
1034                 res[0] = tmp;
1035         }
1036 }
1037
1038 static void restore_to_pos(struct key_user res[], unsigned int elem)
1039 {
1040         if (elem != 0) {
1041                 struct key_user tmp = res[0];
1042                 memmove(res, res + 1, elem*sizeof(res[0]));
1043                 res[elem] = tmp;
1044         }
1045 }
1046
1047 static bool sort_deps(char *filename[], struct op *op[],
1048                       struct key_user res[], unsigned num,
1049                       const TDB_DATA *data, unsigned num_files)
1050 {
1051         unsigned int i, files_done;
1052         struct op *this_op;
1053         bool done[num_files];
1054
1055         /* Nothing left?  We're sorted. */
1056         if (num == 0)
1057                 return true;
1058
1059         memset(done, 0, sizeof(done));
1060
1061         /* Since ops within a trace file are ordered, we just need to figure
1062          * out which file to try next.  Since we don't take into account
1063          * inter-key relationships (which exist by virtue of trace file order),
1064          * we minimize the chance of harm by trying to keep in serial order. */
1065         for (files_done = 0, i = 0; i < num && files_done < num_files; i++) {
1066                 if (done[res[i].file])
1067                         continue;
1068
1069                 this_op = &op[res[i].file][res[i].op_num];
1070                 /* Is what we have good enough for this op? */
1071                 if (satisfies(data, needs(this_op))) {
1072                         move_to_front(res, i);
1073                         if (sort_deps(filename, op, res+1, num-1,
1074                                       gives(this_op, data), num_files))
1075                                 return true;
1076                         restore_to_pos(res, i);
1077                 }
1078                 done[res[i].file] = true;
1079                 files_done++;
1080         }
1081
1082         /* No combination worked. */
1083         return false;
1084 }
1085
1086 static void check_dep_sorting(struct key_user user[], unsigned num_users,
1087                               unsigned num_files)
1088 {
1089 #if DEBUG_DEPS
1090         unsigned int i;
1091         unsigned minima[num_files];
1092
1093         memset(minima, 0, sizeof(minima));
1094         for (i = 0; i < num_users; i++) {
1095                 assert(minima[user[i].file] < user[i].op_num);
1096                 minima[user[i].file] = user[i].op_num;
1097         }
1098 #endif
1099 }
1100
1101 /* All these ops have the same serial number.  Which comes first?
1102  *
1103  * This can happen both because read ops or failed write ops don't
1104  * change serial number, and also due to race since we access the
1105  * number unlocked (the race can cause less detectable ordering problems,
1106  * in which case we'll deadlock and report: fix manually in that case).
1107  */
1108 static void figure_deps(char *filename[], struct op *op[],
1109                         struct key_user user[], unsigned num_users,
1110                         unsigned num_files)
1111 {
1112         /* We assume database starts empty. */
1113         const struct TDB_DATA *data = &tdb_null;
1114
1115         if (!sort_deps(filename, op, user, num_users, data, num_files))
1116                 fail(filename[user[0].file], user[0].op_num+1,
1117                      "Could not resolve inter-dependencies");
1118
1119         check_dep_sorting(user, num_users, num_files);
1120 }
1121
1122 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1123                      unsigned int num)
1124 {
1125         unsigned int h;
1126
1127         /* Gcc nexted function extension.  How cool is this? */
1128         int compare_serial(const void *_a, const void *_b)
1129         {
1130                 const struct key_user *a = _a, *b = _b;
1131
1132                 /* First, maintain order within any trace file. */
1133                 if (a->file == b->file)
1134                         return a->op_num - b->op_num;
1135
1136                 /* Otherwise, arrange by serial order. */
1137                 return op[a->file][a->op_num].serial
1138                         - op[b->file][b->op_num].serial;
1139         }
1140
1141         /* Now sort into serial order. */
1142         for (h = 0; h < total_keys * 2; h++) {
1143                 struct key_user *user = hash[h].user;
1144
1145                 qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
1146                 figure_deps(filename, op, user, hash[h].num_users, num);
1147         }
1148 }
1149
1150 static int destroy_depend(struct depend *dep)
1151 {
1152         list_del(&dep->pre_list);
1153         list_del(&dep->post_list);
1154         return 0;
1155 }
1156
1157 static void add_dependency(void *ctx,
1158                            struct op *op[],
1159                            char *filename[],
1160                            unsigned int needs_file,
1161                            unsigned int needs_opnum,
1162                            unsigned int satisfies_file,
1163                            unsigned int satisfies_opnum)
1164 {
1165         struct depend *dep;
1166         unsigned int needs_start, sat_start;
1167
1168         /* We don't depend on ourselves. */
1169         if (needs_file == satisfies_file) {
1170                 assert(satisfies_opnum < needs_opnum);
1171                 return;
1172         }
1173
1174 #if DEBUG_DEPS
1175         printf("%s:%u: depends on %s:%u\n",
1176                filename[needs_file], needs_opnum+1,
1177                filename[satisfies_file], satisfies_opnum+1);
1178 #endif
1179
1180         needs_start = op[needs_file][needs_opnum].group_start;
1181         sat_start = op[satisfies_file][satisfies_opnum].group_start;
1182
1183         /* If needs is in a transaction, we need it before start. */
1184         if (needs_start) {
1185                 switch (op[needs_file][needs_start].op) {
1186                 case OP_TDB_TRANSACTION_START:
1187                         needs_opnum = needs_start;
1188 #ifdef DEBUG_DEPS
1189                         printf("  -> Back to %u\n", needs_start+1);
1190                         fflush(stdout);
1191 #endif
1192                         break;
1193                 default:
1194                         break;
1195                 }
1196         }
1197
1198         /* If satisfies is in a transaction, we wait until after commit. */
1199         /* FIXME: If transaction is cancelled, don't need dependency. */
1200         if (sat_start) {
1201                 if (op[satisfies_file][sat_start].op
1202                     == OP_TDB_TRANSACTION_START) {
1203                         satisfies_opnum
1204                                 = op[satisfies_file][sat_start].transaction_end;
1205 #ifdef DEBUG_DEPS
1206                         printf("  -> Depends on %u\n", satisfies_opnum+1);
1207                         fflush(stdout);
1208 #endif
1209                 }
1210         }
1211
1212         dep = talloc(ctx, struct depend);
1213         dep->needs_file = needs_file;
1214         dep->needs_opnum = needs_opnum;
1215         dep->satisfies_file = satisfies_file;
1216         dep->satisfies_opnum = satisfies_opnum;
1217         list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
1218         list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
1219         talloc_set_destructor(dep, destroy_depend);
1220 }
1221
1222 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1223 struct traverse_dep {
1224         unsigned int file;
1225         unsigned int op_num;
1226         const struct op *op;
1227 };
1228
1229 /* Sort by which one runs first. */
1230 static int compare_traverse_dep(const void *_a, const void *_b)
1231 {
1232         const struct traverse_dep *a = _a, *b = _b;
1233         const struct traverse *trava = a->op->trav, *travb = b->op->trav;
1234
1235         if (a->op->serial != b->op->serial)
1236                 return a->op->serial - b->op->serial;
1237
1238         /* If they have same serial, it means one didn't make any changes.
1239          * Thus sort by end in that case. */
1240         return a->op[trava->end - a->op_num].serial
1241                 - b->op[travb->end - b->op_num].serial;
1242 }
1243
1244 /* Traversals can deadlock against each other.  Force order. */
1245 static void make_traverse_depends(char *filename[],
1246                                   struct op *op[], unsigned int num_ops[],
1247                                   unsigned int num)
1248 {
1249         unsigned int i, j, num_traversals = 0;
1250         struct traverse_dep *dep;
1251
1252         dep = talloc_array(NULL, struct traverse_dep, 1);
1253
1254         /* Count them. */
1255         for (i = 0; i < num; i++) {
1256                 for (j = 0; j < num_ops[i]; j++) {
1257                         if (op[i][j].op == OP_TDB_TRAVERSE_START
1258                             || op[i][j].op == OP_TDB_TRAVERSE_READ_START) {
1259                                 dep = talloc_realloc(NULL, dep,
1260                                                      struct traverse_dep,
1261                                                      num_traversals+1);
1262                                 dep[num_traversals].file = i;
1263                                 dep[num_traversals].op_num = j;
1264                                 dep[num_traversals].op = &op[i][j];
1265                                 num_traversals++;
1266                         }
1267                 }
1268         }
1269         qsort(dep, num_traversals, sizeof(dep[0]), compare_traverse_dep);
1270         for (i = 1; i < num_traversals; i++) {
1271                 /* i depends on end of traverse i-1. */
1272                 add_dependency(NULL, op, filename, dep[i].file, dep[i].op_num,
1273                                dep[i-1].file, dep[i-1].op->trav->end);
1274         }
1275         talloc_free(dep);
1276 }
1277 #endif /* TRAVERSALS_TAKE_TRANSACTION_LOCK */
1278
1279 static bool changes_db(const struct op *op)
1280 {
1281         return gives(op, NULL) != NULL;
1282 }
1283
1284 static void depend_on_previous(struct op *op[],
1285                                char *filename[],
1286                                unsigned int num,
1287                                struct key_user user[],
1288                                unsigned int i,
1289                                int prev)
1290 {
1291         bool deps[num];
1292         int j;
1293
1294         if (i == 0)
1295                 return;
1296
1297         if (prev == i - 1) {
1298                 /* Just depend on previous. */
1299                 add_dependency(NULL, op, filename,
1300                                user[i].file, user[i].op_num,
1301                                user[prev].file, user[prev].op_num);
1302                 return;
1303         }
1304
1305         /* We have to wait for the readers.  Find last one in *each* file. */
1306         memset(deps, 0, sizeof(deps));
1307         deps[user[i].file] = true;
1308         for (j = i - 1; j > prev; j--) {
1309                 if (!deps[user[j].file]) {
1310                         add_dependency(NULL, op, filename,
1311                                        user[i].file, user[i].op_num,
1312                                        user[j].file, user[j].op_num);
1313                         deps[user[j].file] = true;
1314                 }
1315         }
1316 }
1317
1318 /* This is simple, but not complete.  We don't take into account
1319  * indirect dependencies. */
1320 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1321                                   unsigned int num)
1322 {
1323         unsigned int i, j;
1324
1325         /* There can only be one real dependency on each file */
1326         for (i = 0; i < num; i++) {
1327                 for (j = 1; j < num_ops[i]; j++) {
1328                         struct depend *dep, *next;
1329                         struct depend *prev[num];
1330
1331                         memset(prev, 0, sizeof(prev));
1332
1333                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1334                                 if (!prev[dep->satisfies_file]) {
1335                                         prev[dep->satisfies_file] = dep;
1336                                         continue;
1337                                 }
1338                                 if (prev[dep->satisfies_file]->satisfies_opnum
1339                                     < dep->satisfies_opnum) {
1340                                         talloc_free(prev[dep->satisfies_file]);
1341                                         prev[dep->satisfies_file] = dep;
1342                                 } else
1343                                         talloc_free(dep);
1344                         }
1345                 }
1346         }
1347
1348         for (i = 0; i < num; i++) {
1349                 int deps[num];
1350
1351                 for (j = 0; j < num; j++)
1352                         deps[j] = -1;
1353
1354                 for (j = 1; j < num_ops[i]; j++) {
1355                         struct depend *dep, *next;
1356
1357                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1358                                 if (deps[dep->satisfies_file]
1359                                     >= (int)dep->satisfies_opnum)
1360                                         talloc_free(dep);
1361                                 else
1362                                         deps[dep->satisfies_file]
1363                                                 = dep->satisfies_opnum;
1364                         }
1365                 }
1366         }
1367 }
1368
1369 static void derive_dependencies(char *filename[],
1370                                 struct op *op[], unsigned int num_ops[],
1371                                 unsigned int num)
1372 {
1373         struct keyinfo *hash;
1374         unsigned int h, i;
1375
1376         /* Create hash table for faster key lookup. */
1377         hash = hash_ops(op, num_ops, num);
1378
1379         /* Sort them by serial number. */
1380         sort_ops(hash, filename, op, num);
1381
1382         /* Create dependencies back to the last change, rather than
1383          * creating false dependencies by naively making each one
1384          * depend on the previous.  This has two purposes: it makes
1385          * later optimization simpler, and it also avoids deadlock with
1386          * same sequence number ops inside traversals (if one
1387          * traversal doesn't write anything, two ops can have the same
1388          * sequence number yet we can create a traversal dependency
1389          * the other way). */
1390         for (h = 0; h < total_keys * 2; h++) {
1391                 int prev = -1;
1392
1393                 if (hash[h].num_users < 2)
1394                         continue;
1395
1396                 for (i = 0; i < hash[h].num_users; i++) {
1397                         if (changes_db(&op[hash[h].user[i].file]
1398                                        [hash[h].user[i].op_num])) {
1399                                 depend_on_previous(op, filename, num,
1400                                                    hash[h].user, i, prev);
1401                                 prev = i;
1402                         } else if (prev >= 0)
1403                                 add_dependency(hash, op, filename,
1404                                                hash[h].user[i].file,
1405                                                hash[h].user[i].op_num,
1406                                                hash[h].user[prev].file,
1407                                                hash[h].user[prev].op_num);
1408                 }
1409         }
1410
1411 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1412         make_traverse_depends(filename, op, num_ops, num);
1413 #endif
1414
1415         optimize_dependencies(op, num_ops, num);
1416 }
1417
1418 int main(int argc, char *argv[])
1419 {
1420         struct timeval start, end;
1421         unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1422         struct op *op[argc];
1423         int fds[2];
1424         char c;
1425         bool ok = true;
1426
1427         if (argc < 3)
1428                 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1429
1430         pipes = talloc_array(NULL, struct pipe, argc - 2);
1431         for (i = 0; i < argc - 2; i++) {
1432                 printf("Loading tracefile %s...", argv[2+i]);
1433                 fflush(stdout);
1434                 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1435                                        &tdb_flags[i], &open_flags[i]);
1436                 if (pipe(pipes[i].fd) != 0)
1437                         err(1, "creating pipe");
1438                 printf("done\n");
1439         }
1440
1441         printf("Calculating inter-dependencies...");
1442         fflush(stdout);
1443         derive_dependencies(argv+2, op, num_ops, i);
1444         printf("done\n");
1445
1446         /* Don't fork for single arg case: simple debugging. */
1447         if (argc == 3) {
1448                 struct tdb_context *tdb;
1449                 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
1450                                   open_flags[0], 0600,
1451                                   NULL, hash_key);
1452                 printf("Single threaded run...");
1453                 fflush(stdout);
1454
1455                 run_ops(tdb, pipes[0].fd[0], argv+2, 0, op[0], 1, num_ops[0]);
1456                 check_deps(argv[2], op[0], num_ops[0]);
1457
1458                 printf("done\n");
1459                 exit(0);
1460         }
1461
1462         if (pipe(fds) != 0)
1463                 err(1, "creating pipe");
1464
1465         for (i = 0; i < argc - 2; i++) {
1466                 struct tdb_context *tdb;
1467
1468                 switch (fork()) {
1469                 case -1:
1470                         err(1, "fork failed");
1471                 case 0:
1472                         close(fds[1]);
1473                         tdb = tdb_open_ex(argv[1], hashsize[i], tdb_flags[i],
1474                                           open_flags[i], 0600,
1475                                           NULL, hash_key);
1476                         if (!tdb)
1477                                 err(1, "Opening tdb %s", argv[1]);
1478
1479                         /* This catches parent exiting. */
1480                         if (read(fds[0], &c, 1) != 1)
1481                                 exit(1);
1482                         run_ops(tdb, pipes[i].fd[0], argv+2, i, op[i], 1,
1483                                 num_ops[i]);
1484                         check_deps(argv[2+i], op[i], num_ops[i]);
1485                         exit(0);
1486                 default:
1487                         break;
1488                 }
1489         }
1490
1491         /* Let everything settle. */
1492         sleep(1);
1493
1494         printf("Starting run...");
1495         fflush(stdout);
1496         gettimeofday(&start, NULL);
1497         /* Tell them all to go!  Any write of sufficient length will do. */
1498         if (write(fds[1], hashsize, i) != i)
1499                 err(1, "Writing to wakeup pipe");
1500
1501         for (i = 0; i < argc - 2; i++) {
1502                 int status;
1503                 wait(&status);
1504                 if (!WIFEXITED(status)) {
1505                         warnx("Child died with signal %i", WTERMSIG(status));
1506                         ok = false;
1507                 } else if (WEXITSTATUS(status) != 0)
1508                         /* Assume child spat out error. */
1509                         ok = false;
1510         }
1511         if (!ok)
1512                 exit(1);
1513
1514         gettimeofday(&end, NULL);
1515         printf("done\n");
1516
1517         end.tv_sec -= start.tv_sec;
1518         printf("Time replaying: %lu usec\n",
1519                end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));
1520
1521         exit(0);
1522 }