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 #include <fcntl.h>
  19
  20 #define STRINGIFY2(x) #x
  21 #define STRINGIFY(x) STRINGIFY2(x)
  22
  23 /* Avoid mod by zero */
  24 static unsigned int total_keys = 1;
  25
  26 /* #define DEBUG_DEPS 1 */
  27
  28 /* Traversals block transactions in the current implementation. */
  29 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
  30
  31 struct pipe {
  32         int fd[2];
  33 };
  34 static struct pipe *pipes;
  35 static int backoff_fd = -1;
  36
  37 static void __attribute__((noreturn)) fail(const char *filename,
  38                                            unsigned int line,
  39                                            const char *fmt, ...)
  40 {
  41         va_list ap;
  42
  43         va_start(ap, fmt);
  44         fprintf(stderr, "%s:%u: FAIL: ", filename, line);
  45         vfprintf(stderr, fmt, ap);
  46         fprintf(stderr, "\n");
  47         va_end(ap);
  48         exit(1);
  49 }
  50
  51 /* Try or die. */
  52 #define try(expr, expect)                                               \
  53         do {                                                            \
  54                 int ret = (expr);                                       \
  55                 if (ret != (expect))                                    \
  56                         fail(filename[file], i+1,                       \
  57                              STRINGIFY(expr) "= %i", ret);              \
  58         } while (0)
  59
  60 /* Try or imitate results. */
  61 #define unreliable(expr, expect, force, undo)                           \
  62         do {                                                            \
  63                 int ret = expr;                                         \
  64                 if (ret != expect) {                                    \
  65                         fprintf(stderr, "%s:%u: %s gave %i not %i",     \
  66                                 filename[file], i+1, STRINGIFY(expr),   \
  67                                 ret, expect);                           \
  68                         if (expect == 0)                                \
  69                                 force;                                  \
  70                         else                                            \
  71                                 undo;                                   \
  72                 }                                                       \
  73         } while (0)
  74
  75 static bool key_eq(TDB_DATA a, TDB_DATA b)
  76 {
  77         if (a.dsize != b.dsize)
  78                 return false;
  79         return memcmp(a.dptr, b.dptr, a.dsize) == 0;
  80 }
  81
  82 /* This is based on the hash algorithm from gdbm */
  83 static unsigned int hash_key(TDB_DATA *key)
  84 {
  85         uint32_t value; /* Used to compute the hash value.  */
  86         uint32_t   i;   /* Used to cycle through random values. */
  87
  88         /* Set the initial value from the key size. */
  89         for (value = 0x238F13AF ^ key->dsize, i=0; i < key->dsize; i++)
  90                 value = (value + (key->dptr[i] << (i*5 % 24)));
  91
  92         return (1103515243 * value + 12345);
  93 }
  94
  95 enum op_type {
  96         OP_TDB_LOCKALL,
  97         OP_TDB_LOCKALL_MARK,
  98         OP_TDB_LOCKALL_UNMARK,
  99         OP_TDB_LOCKALL_NONBLOCK,
 100         OP_TDB_UNLOCKALL,
 101         OP_TDB_LOCKALL_READ,
 102         OP_TDB_LOCKALL_READ_NONBLOCK,
 103         OP_TDB_UNLOCKALL_READ,
 104         OP_TDB_CHAINLOCK,
 105         OP_TDB_CHAINLOCK_NONBLOCK,
 106         OP_TDB_CHAINLOCK_MARK,
 107         OP_TDB_CHAINLOCK_UNMARK,
 108         OP_TDB_CHAINUNLOCK,
 109         OP_TDB_CHAINLOCK_READ,
 110         OP_TDB_CHAINUNLOCK_READ,
 111         OP_TDB_PARSE_RECORD,
 112         OP_TDB_EXISTS,
 113         OP_TDB_STORE,
 114         OP_TDB_APPEND,
 115         OP_TDB_GET_SEQNUM,
 116         OP_TDB_WIPE_ALL,
 117         OP_TDB_TRANSACTION_START,
 118         OP_TDB_TRANSACTION_CANCEL,
 119         OP_TDB_TRANSACTION_PREPARE_COMMIT,
 120         OP_TDB_TRANSACTION_COMMIT,
 121         OP_TDB_TRAVERSE_READ_START,
 122         OP_TDB_TRAVERSE_START,
 123         OP_TDB_TRAVERSE_END,
 124         OP_TDB_TRAVERSE,
 125         OP_TDB_TRAVERSE_END_EARLY,
 126         OP_TDB_FIRSTKEY,
 127         OP_TDB_NEXTKEY,
 128         OP_TDB_FETCH,
 129         OP_TDB_DELETE,
 130 };
 131
 132 struct op {
 133         unsigned int seqnum;
 134         enum op_type type;
 135         TDB_DATA key;
 136         TDB_DATA data;
 137         int ret;
 138
 139         /* Who is waiting for us? */
 140         struct list_head post;
 141         /* What are we waiting for? */
 142         struct list_head pre;
 143
 144         /* If I'm part of a group (traverse/transaction) where is
 145          * start?  (Otherwise, 0) */
 146         unsigned int group_start;
 147
 148         union {
 149                 int flag; /* open and store */
 150                 struct {  /* append */
 151                         TDB_DATA pre;
 152                         TDB_DATA post;
 153                 } append;
 154                 /* transaction/traverse start/chainlock */
 155                 unsigned int group_len;
 156         };
 157 };
 158
 159 struct op_desc {
 160         unsigned int file;
 161         unsigned int op_num;
 162 };
 163
 164 static unsigned char hex_char(const char *filename, unsigned int line, char c)
 165 {
 166         c = toupper(c);
 167         if (c >= 'A' && c <= 'F')
 168                 return c - 'A' + 10;
 169         if (c >= '0' && c <= '9')
 170                 return c - '0';
 171         fail(filename, line, "invalid hex character '%c'", c);
 172 }
 173
 174 /* TDB data is <size>:<%02x>* */
 175 static TDB_DATA make_tdb_data(const void *ctx,
 176                               const char *filename, unsigned int line,
 177                               const char *word)
 178 {
 179         TDB_DATA data;
 180         unsigned int i;
 181         const char *p;
 182
 183         if (streq(word, "NULL"))
 184                 return tdb_null;
 185
 186         data.dsize = atoi(word);
 187         data.dptr = talloc_array(ctx, unsigned char, data.dsize);
 188         p = strchr(word, ':');
 189         if (!p)
 190                 fail(filename, line, "invalid tdb data '%s'", word);
 191         p++;
 192         for (i = 0; i < data.dsize; i++)
 193                 data.dptr[i] = hex_char(filename, line, p[i*2])*16
 194                         + hex_char(filename, line, p[i*2+1]);
 195
 196         return data;
 197 }
 198
 199 static void add_op(const char *filename, struct op **op, unsigned int i,
 200                    unsigned int seqnum, enum op_type type)
 201 {
 202         struct op *new;
 203         *op = talloc_realloc(NULL, *op, struct op, i+1);
 204         new = (*op) + i;
 205         new->type = type;
 206         new->seqnum = seqnum;
 207         new->ret = 0;
 208         new->group_start = 0;
 209 }
 210
 211 static void op_add_nothing(const char *filename,
 212                            struct op op[], unsigned int op_num, char *words[])
 213 {
 214         if (words[2])
 215                 fail(filename, op_num+1, "Expected no arguments");
 216         op[op_num].key = tdb_null;
 217 }
 218
 219 static void op_add_key(const char *filename,
 220                        struct op op[], unsigned int op_num, char *words[])
 221 {
 222         if (words[2] == NULL || words[3])
 223                 fail(filename, op_num+1, "Expected just a key");
 224
 225         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 226         total_keys++;
 227 }
 228
 229 static void op_add_key_ret(const char *filename,
 230                            struct op op[], unsigned int op_num, char *words[])
 231 {
 232         if (!words[2] || !words[3] || !words[4] || words[5]
 233             || !streq(words[3], "="))
 234                 fail(filename, op_num+1, "Expected <key> = <ret>");
 235         op[op_num].ret = atoi(words[4]);
 236         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 237         /* May only be a unique key if it fails */
 238         if (op[op_num].ret != 0)
 239                 total_keys++;
 240 }
 241
 242 static void op_add_key_data(const char *filename,
 243                             struct op op[], unsigned int op_num, char *words[])
 244 {
 245         if (!words[2] || !words[3] || !words[4] || words[5]
 246             || !streq(words[3], "="))
 247                 fail(filename, op_num+1, "Expected <key> = <data>");
 248         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 249         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[4]);
 250         /* May only be a unique key if it fails */
 251         if (!op[op_num].data.dptr)
 252                 total_keys++;
 253 }
 254
 255 /* We don't record the keys or data for a traverse, as we don't use them. */
 256 static void op_add_traverse(const char *filename,
 257                             struct op op[], unsigned int op_num, char *words[])
 258 {
 259         if (!words[2] || !words[3] || !words[4] || words[5]
 260             || !streq(words[3], "="))
 261                 fail(filename, op_num+1, "Expected <key> = <data>");
 262         op[op_num].key = tdb_null;
 263 }
 264
 265 /* Full traverse info is useful for debugging, but changing it to
 266  * "traversefn" without the data makes the traces *much* smaller! */
 267 static void op_add_traversefn(const char *filename,
 268                             struct op op[], unsigned int op_num, char *words[])
 269 {
 270         if (words[2])
 271                 fail(filename, op_num+1, "Expected no values");
 272         op[op_num].key = tdb_null;
 273 }
 274
 275 /* <seqnum> tdb_store <rec> <rec> <flag> = <ret> */
 276 static void op_add_store(const char *filename,
 277                          struct op op[], unsigned int op_num, char *words[])
 278 {
 279         if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
 280             || words[7] || !streq(words[5], "="))
 281                 fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
 282
 283         op[op_num].flag = strtoul(words[4], NULL, 0);
 284         op[op_num].ret = atoi(words[6]);
 285         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 286         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 287         total_keys++;
 288 }
 289
 290 /* <seqnum> tdb_append <rec> <rec> = <rec> */
 291 static void op_add_append(const char *filename,
 292                           struct op op[], unsigned int op_num, char *words[])
 293 {
 294         if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
 295             || !streq(words[4], "="))
 296                 fail(filename, op_num+1, "Expect <key> <data> = <rec>");
 297
 298         op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 299         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 300
 301         op[op_num].append.post
 302                 = make_tdb_data(op, filename, op_num+1, words[5]);
 303
 304         /* By subtraction, figure out what previous data was. */
 305         op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
 306         op[op_num].append.pre.dsize
 307                 = op[op_num].append.post.dsize - op[op_num].data.dsize;
 308         total_keys++;
 309 }
 310
 311 /* <seqnum> tdb_get_seqnum = <ret> */
 312 static void op_add_seqnum(const char *filename,
 313                           struct op op[], unsigned int op_num, char *words[])
 314 {
 315         if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
 316                 fail(filename, op_num+1, "Expect = <ret>");
 317
 318         op[op_num].key = tdb_null;
 319         op[op_num].ret = atoi(words[3]);
 320 }
 321
 322 static void op_add_traverse_start(const char *filename,
 323                                   struct op op[],
 324                                   unsigned int op_num, char *words[])
 325 {
 326         if (words[2])
 327                 fail(filename, op_num+1, "Expect no arguments");
 328
 329         op[op_num].key = tdb_null;
 330         op[op_num].group_len = 0;
 331 }
 332
 333 static void op_add_transaction(const char *filename, struct op op[],
 334                                unsigned int op_num, char *words[])
 335 {
 336         if (words[2])
 337                 fail(filename, op_num+1, "Expect no arguments");
 338
 339         op[op_num].key = tdb_null;
 340         op[op_num].group_len = 0;
 341 }
 342
 343 static void op_add_chainlock(const char *filename,
 344                              struct op op[], unsigned int op_num, char *words[])
 345 {
 346         if (words[2] == NULL || words[3])
 347                 fail(filename, op_num+1, "Expected just a key");
 348
 349         /* A chainlock key isn't a key in the normal sense; it doesn't
 350          * have to be in the db at all.  Also, we don't want to hash this op. */
 351         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 352         op[op_num].key = tdb_null;
 353         op[op_num].group_len = 0;
 354 }
 355
 356 static void op_add_chainlock_ret(const char *filename,
 357                                  struct op op[], unsigned int op_num,
 358                                  char *words[])
 359 {
 360         if (!words[2] || !words[3] || !words[4] || words[5]
 361             || !streq(words[3], "="))
 362                 fail(filename, op_num+1, "Expected <key> = <ret>");
 363         op[op_num].ret = atoi(words[4]);
 364         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 365         op[op_num].key = tdb_null;
 366         op[op_num].group_len = 0;
 367         total_keys++;
 368 }
 369
 370 static int op_find_start(struct op op[], unsigned int op_num, enum op_type type)
 371 {
 372         unsigned int i;
 373
 374         for (i = op_num-1; i > 0; i--) {
 375                 if (op[i].type == type && !op[i].group_len)
 376                         return i;
 377         }
 378         return 0;
 379 }
 380
 381 static void op_analyze_transaction(const char *filename,
 382                                    struct op op[], unsigned int op_num,
 383                                    char *words[])
 384 {
 385         unsigned int start, i;
 386
 387         op[op_num].key = tdb_null;
 388
 389         if (words[2])
 390                 fail(filename, op_num+1, "Expect no arguments");
 391
 392         start = op_find_start(op, op_num, OP_TDB_TRANSACTION_START);
 393         if (!start)
 394                 fail(filename, op_num+1, "no transaction start found");
 395
 396         op[start].group_len = op_num - start;
 397
 398         /* This rolls in nested transactions.  I think that's right. */
 399         for (i = start; i <= op_num; i++)
 400                 op[i].group_start = start;
 401 }
 402
 403 /* We treat chainlocks a lot like transactions, even though that's overkill */
 404 static void op_analyze_chainlock(const char *filename,
 405                                  struct op op[], unsigned int op_num,
 406                                  char *words[])
 407 {
 408         unsigned int i, start;
 409
 410         if (words[2] == NULL || words[3])
 411                 fail(filename, op_num+1, "Expected just a key");
 412
 413         op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
 414         op[op_num].key = tdb_null;
 415         total_keys++;
 416
 417         start = op_find_start(op, op_num, OP_TDB_CHAINLOCK);
 418         if (!start)
 419                 start = op_find_start(op, op_num, OP_TDB_CHAINLOCK_READ);
 420         if (!start)
 421                 fail(filename, op_num+1, "no initial chainlock found");
 422
 423         /* FIXME: We'd have to do something clever to make this work
 424          * vs. deadlock. */
 425         if (!key_eq(op[start].data, op[op_num].data))
 426                 fail(filename, op_num+1, "nested chainlock calls?");
 427
 428         op[start].group_len = op_num - start;
 429         for (i = start; i <= op_num; i++)
 430                 op[i].group_start = start;
 431 }
 432
 433 static void op_analyze_traverse(const char *filename,
 434                                 struct op op[], unsigned int op_num,
 435                                 char *words[])
 436 {
 437         int i, start;
 438
 439         op[op_num].key = tdb_null;
 440
 441         /* = %u means traverse function terminated. */
 442         if (words[2]) {
 443                 if (!streq(words[2], "=") || !words[3] || words[4])
 444                         fail(filename, op_num+1, "expect = <num>");
 445                 op[op_num].ret = atoi(words[3]);
 446         } else
 447                 op[op_num].ret = 0;
 448
 449         start = op_find_start(op, op_num, OP_TDB_TRAVERSE_START);
 450         if (!start)
 451                 start = op_find_start(op, op_num, OP_TDB_TRAVERSE_READ_START);
 452         if (!start)
 453                 fail(filename, op_num+1, "no traversal start found");
 454
 455         op[start].group_len = op_num - start;
 456
 457         /* Don't roll in nested traverse/chainlock */
 458         for (i = start; i <= op_num; i++)
 459                 if (!op[i].group_start)
 460                         op[i].group_start = start;
 461 }
 462
 463 /* Keep -Wmissing-declarations happy: */
 464 const struct op_table *
 465 find_keyword (register const char *str, register unsigned int len);
 466
 467 #include "keywords.c"
 468
 469 struct depend {
 470         /* We can have more than one */
 471         struct list_node pre_list;
 472         struct list_node post_list;
 473         struct op_desc needs;
 474         struct op_desc prereq;
 475 };
 476
 477 static void check_deps(const char *filename, struct op op[], unsigned int num)
 478 {
 479 #ifdef DEBUG_DEPS
 480         unsigned int i;
 481
 482         for (i = 1; i < num; i++)
 483                 if (!list_empty(&op[i].pre))
 484                         fail(filename, i+1, "Still has dependencies");
 485 #endif
 486 }
 487
 488 static void dump_pre(char *filename[], struct op *op[],
 489                      unsigned int file, unsigned int i)
 490 {
 491         struct depend *dep;
 492
 493         printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
 494                 op[file][i].seqnum);
 495         list_for_each(&op[file][i].pre, dep, pre_list)
 496                 printf("    %s:%u (%u)\n",
 497                        filename[dep->prereq.file], dep->prereq.op_num+1,
 498                        op[dep->prereq.file][dep->prereq.op_num].seqnum);
 499         check_deps(filename[file], op[file], i);
 500 }
 501
 502 /* We simply read/write pointers, since we all are children. */
 503 static bool do_pre(struct tdb_context *tdb,
 504                    char *filename[], struct op *op[],
 505                    unsigned int file, int pre_fd, unsigned int i,
 506                    bool backoff)
 507 {
 508         while (!list_empty(&op[file][i].pre)) {
 509                 struct depend *dep;
 510
 511 #if DEBUG_DEPS
 512                 printf("%s:%u:waiting for pre\n", filename[file], i+1);
 513                 fflush(stdout);
 514 #endif
 515                 if (backoff)
 516                         alarm(2);
 517                 else
 518                         alarm(10);
 519                 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
 520                         if (errno == EINTR) {
 521                                 if (backoff) {
 522                                         struct op_desc desc = { file,i };
 523                                         warnx("%s:%u:avoiding deadlock",
 524                                               filename[file], i+1);
 525                                         if (write(backoff_fd, &desc,
 526                                                   sizeof(desc)) != sizeof(desc))
 527                                                 err(1, "writing backoff_fd");
 528                                         return false;
 529                                 }
 530                                 dump_pre(filename, op, file, i);
 531                                 exit(1);
 532                         } else
 533                                 errx(1, "Reading from pipe");
 534                 }
 535                 alarm(0);
 536
 537 #if DEBUG_DEPS
 538                 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
 539                        dep->needs.op_num+1, filename[dep->prereq.file],
 540                        dep->prereq.op_num+1);
 541                 fflush(stdout);
 542 #endif
 543                 /* This could be any op, not just this one. */
 544                 talloc_free(dep);
 545         }
 546         return true;
 547 }
 548
 549 static void do_post(char *filename[], struct op *op[],
 550                     unsigned int file, unsigned int i)
 551 {
 552         struct depend *dep;
 553
 554         list_for_each(&op[file][i].post, dep, post_list) {
 555 #if DEBUG_DEPS
 556                 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
 557                        filename[dep->needs.file], dep->needs.op_num+1);
 558 #endif
 559                 if (write(pipes[dep->needs.file].fd[1], &dep, sizeof(dep))
 560                     != sizeof(dep))
 561                         err(1, "%s:%u failed to tell file %s",
 562                             filename[file], i+1, filename[dep->needs.file]);
 563         }
 564 }
 565
 566 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 567 {
 568         return data.dsize;
 569 }
 570
 571 static unsigned run_ops(struct tdb_context *tdb,
 572                         int pre_fd,
 573                         char *filename[],
 574                         struct op *op[],
 575                         unsigned int file,
 576                         unsigned int start, unsigned int stop,
 577                         bool backoff);
 578
 579 struct traverse_info {
 580         struct op **op;
 581         char **filename;
 582         unsigned file;
 583         int pre_fd;
 584         unsigned int start;
 585         unsigned int i;
 586 };
 587
 588 /* More complex.  Just do whatever's they did at the n'th entry. */
 589 static int nontrivial_traverse(struct tdb_context *tdb,
 590                                TDB_DATA key, TDB_DATA data,
 591                                void *_tinfo)
 592 {
 593         struct traverse_info *tinfo = _tinfo;
 594         unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
 595         bool avoid_deadlock = false;
 596
 597         if (tinfo->i == tinfo->start + trav_len) {
 598                 /* This can happen if traverse expects to be empty. */
 599                 if (trav_len == 1)
 600                         return 1;
 601                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 602                      "traverse did not terminate");
 603         }
 604
 605         if (tinfo->op[tinfo->file][tinfo->i].type != OP_TDB_TRAVERSE)
 606                 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 607                      "%s:%u:traverse terminated early");
 608
 609 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
 610         avoid_deadlock = true;
 611 #endif
 612
 613         /* Run any normal ops. */
 614         tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
 615                            tinfo->file, tinfo->i+1, tinfo->start + trav_len,
 616                            avoid_deadlock);
 617
 618         /* We backed off, or we hit OP_TDB_TRAVERSE_END/EARLY. */
 619         if (tinfo->op[tinfo->file][tinfo->i].type != OP_TDB_TRAVERSE)
 620                 return 1;
 621
 622         return 0;
 623 }
 624
 625 static unsigned op_traverse(struct tdb_context *tdb,
 626                             int pre_fd,
 627                             char *filename[],
 628                             unsigned int file,
 629                             int (*traversefn)(struct tdb_context *,
 630                                               tdb_traverse_func, void *),
 631                             struct op *op[],
 632                             unsigned int start)
 633 {
 634         struct traverse_info tinfo = { op, filename, file, pre_fd,
 635                                        start, start+1 };
 636
 637         traversefn(tdb, nontrivial_traverse, &tinfo);
 638
 639         /* Traversing in wrong order can have strange effects: eg. if
 640          * original traverse went A (delete A), B, we might do B
 641          * (delete A).  So if we have ops left over, we do it now. */
 642         while (tinfo.i != start + op[file][start].group_len) {
 643                 if (op[file][tinfo.i].type == OP_TDB_TRAVERSE
 644                     || op[file][tinfo.i].type == OP_TDB_TRAVERSE_END_EARLY)
 645                         tinfo.i++;
 646                 else
 647                         tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
 648                                           tinfo.i,
 649                                           start + op[file][start].group_len,
 650                                           false);
 651         }
 652
 653         return tinfo.i;
 654 }
 655
 656 static void break_out(int sig)
 657 {
 658 }
 659
 660 static __attribute__((noinline))
 661 unsigned run_ops(struct tdb_context *tdb,
 662                  int pre_fd,
 663                  char *filename[],
 664                  struct op *op[],
 665                  unsigned int file,
 666                  unsigned int start, unsigned int stop,
 667                  bool backoff)
 668 {
 669         unsigned int i;
 670         struct sigaction sa;
 671
 672         sa.sa_handler = break_out;
 673         sa.sa_flags = 0;
 674
 675         sigaction(SIGALRM, &sa, NULL);
 676         for (i = start; i < stop; i++) {
 677                 if (!do_pre(tdb, filename, op, file, pre_fd, i, backoff))
 678                         return i;
 679
 680                 switch (op[file][i].type) {
 681                 case OP_TDB_LOCKALL:
 682                         try(tdb_lockall(tdb), op[file][i].ret);
 683                         break;
 684                 case OP_TDB_LOCKALL_MARK:
 685                         try(tdb_lockall_mark(tdb), op[file][i].ret);
 686                         break;
 687                 case OP_TDB_LOCKALL_UNMARK:
 688                         try(tdb_lockall_unmark(tdb), op[file][i].ret);
 689                         break;
 690                 case OP_TDB_LOCKALL_NONBLOCK:
 691                         unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
 692                                    tdb_lockall(tdb), tdb_unlockall(tdb));
 693                         break;
 694                 case OP_TDB_UNLOCKALL:
 695                         try(tdb_unlockall(tdb), op[file][i].ret);
 696                         break;
 697                 case OP_TDB_LOCKALL_READ:
 698                         try(tdb_lockall_read(tdb), op[file][i].ret);
 699                         break;
 700                 case OP_TDB_LOCKALL_READ_NONBLOCK:
 701                         unreliable(tdb_lockall_read_nonblock(tdb),
 702                                    op[file][i].ret,
 703                                    tdb_lockall_read(tdb),
 704                                    tdb_unlockall_read(tdb));
 705                         break;
 706                 case OP_TDB_UNLOCKALL_READ:
 707                         try(tdb_unlockall_read(tdb), op[file][i].ret);
 708                         break;
 709                 case OP_TDB_CHAINLOCK:
 710                         try(tdb_chainlock(tdb, op[file][i].key),
 711                             op[file][i].ret);
 712                         break;
 713                 case OP_TDB_CHAINLOCK_NONBLOCK:
 714                         unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
 715                                    op[file][i].ret,
 716                                    tdb_chainlock(tdb, op[file][i].key),
 717                                    tdb_chainunlock(tdb, op[file][i].key));
 718                         break;
 719                 case OP_TDB_CHAINLOCK_MARK:
 720                         try(tdb_chainlock_mark(tdb, op[file][i].key),
 721                             op[file][i].ret);
 722                         break;
 723                 case OP_TDB_CHAINLOCK_UNMARK:
 724                         try(tdb_chainlock_unmark(tdb, op[file][i].key),
 725                             op[file][i].ret);
 726                         break;
 727                 case OP_TDB_CHAINUNLOCK:
 728                         try(tdb_chainunlock(tdb, op[file][i].key),
 729                             op[file][i].ret);
 730                         break;
 731                 case OP_TDB_CHAINLOCK_READ:
 732                         try(tdb_chainlock_read(tdb, op[file][i].key),
 733                             op[file][i].ret);
 734                         break;
 735                 case OP_TDB_CHAINUNLOCK_READ:
 736                         try(tdb_chainunlock_read(tdb, op[file][i].key),
 737                             op[file][i].ret);
 738                         break;
 739                 case OP_TDB_PARSE_RECORD:
 740                         try(tdb_parse_record(tdb, op[file][i].key, get_len,
 741                                              NULL),
 742                             op[file][i].ret);
 743                         break;
 744                 case OP_TDB_EXISTS:
 745                         try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
 746                         break;
 747                 case OP_TDB_STORE:
 748                         try(tdb_store(tdb, op[file][i].key, op[file][i].data,
 749                                       op[file][i].flag),
 750                             op[file][i].ret);
 751                         break;
 752                 case OP_TDB_APPEND:
 753                         try(tdb_append(tdb, op[file][i].key, op[file][i].data),
 754                             op[file][i].ret);
 755                         break;
 756                 case OP_TDB_GET_SEQNUM:
 757                         try(tdb_get_seqnum(tdb), op[file][i].ret);
 758                         break;
 759                 case OP_TDB_WIPE_ALL:
 760                         try(tdb_wipe_all(tdb), op[file][i].ret);
 761                         break;
 762                 case OP_TDB_TRANSACTION_START:
 763                         try(tdb_transaction_start(tdb), op[file][i].ret);
 764                         break;
 765                 case OP_TDB_TRANSACTION_CANCEL:
 766                         try(tdb_transaction_cancel(tdb), op[file][i].ret);
 767                         break;
 768                 case OP_TDB_TRANSACTION_PREPARE_COMMIT:
 769                         try(tdb_transaction_prepare_commit(tdb),
 770                             op[file][i].ret);
 771                         break;
 772                 case OP_TDB_TRANSACTION_COMMIT:
 773                         try(tdb_transaction_commit(tdb), op[file][i].ret);
 774                         break;
 775                 case OP_TDB_TRAVERSE_READ_START:
 776                         i = op_traverse(tdb, pre_fd, filename, file,
 777                                         tdb_traverse_read, op, i);
 778                         break;
 779                 case OP_TDB_TRAVERSE_START:
 780                         i = op_traverse(tdb, pre_fd, filename, file,
 781                                         tdb_traverse, op, i);
 782                         break;
 783                 case OP_TDB_TRAVERSE:
 784                 case OP_TDB_TRAVERSE_END_EARLY:
 785                         /* Terminate: we're in a traverse, and we've
 786                          * done our ops. */
 787                         return i;
 788                 case OP_TDB_TRAVERSE_END:
 789                         fail(filename[file], i+1, "unexpected end traverse");
 790                 /* FIXME: These must be treated like traverse. */
 791                 case OP_TDB_FIRSTKEY:
 792                         if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
 793                                 fail(filename[file], i+1, "bad firstkey");
 794                         break;
 795                 case OP_TDB_NEXTKEY:
 796                         if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
 797                                     op[file][i].data))
 798                                 fail(filename[file], i+1, "bad nextkey");
 799                         break;
 800                 case OP_TDB_FETCH: {
 801                         TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
 802                         if (!key_eq(f, op[file][i].data))
 803                                 fail(filename[file], i+1, "bad fetch %u",
 804                                      f.dsize);
 805                         break;
 806                 }
 807                 case OP_TDB_DELETE:
 808                         try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
 809                         break;
 810                 }
 811                 do_post(filename, op, file, i);
 812         }
 813         return i;
 814 }
 815
 816 /* tdbtorture, in particular, can do a tdb_close with a transaction in
 817  * progress. */
 818 static struct op *maybe_cancel_transaction(const char *filename,
 819                                            struct op *op, unsigned int *num)
 820 {
 821         unsigned int start = op_find_start(op, *num, OP_TDB_TRANSACTION_START);
 822
 823         if (start) {
 824                 char *words[] = { "<unknown>", "tdb_close", NULL };
 825                 add_op(filename, &op, *num, op[start].seqnum,
 826                        OP_TDB_TRANSACTION_CANCEL);
 827                 op_analyze_transaction(filename, op, *num, words);
 828                 (*num)++;
 829         }
 830         return op;
 831 }
 832
 833 static struct op *load_tracefile(const char *filename, unsigned int *num,
 834                                  unsigned int *hashsize,
 835                                  unsigned int *tdb_flags,
 836                                  unsigned int *open_flags)
 837 {
 838         unsigned int i;
 839         struct op *op = talloc_array(NULL, struct op, 1);
 840         char **words;
 841         char **lines;
 842         char *file;
 843
 844         file = grab_file(NULL, filename, NULL);
 845         if (!file)
 846                 err(1, "Reading %s", filename);
 847
 848         lines = strsplit(file, file, "\n", NULL);
 849         if (!lines[0])
 850                 errx(1, "%s is empty", filename);
 851
 852         words = strsplit(lines, lines[0], " ", NULL);
 853         if (!streq(words[1], "tdb_open"))
 854                 fail(filename, 1, "does not start with tdb_open");
 855
 856         *hashsize = atoi(words[2]);
 857         *tdb_flags = strtoul(words[3], NULL, 0);
 858         *open_flags = strtoul(words[4], NULL, 0);
 859
 860         for (i = 1; lines[i]; i++) {
 861                 const struct op_table *opt;
 862
 863                 words = strsplit(lines, lines[i], " ", NULL);
 864                 if (!words[0] || !words[1])
 865                         fail(filename, i+1, "Expected seqnum number and op");
 866
 867                 opt = find_keyword(words[1], strlen(words[1]));
 868                 if (!opt) {
 869                         if (streq(words[1], "tdb_close")) {
 870                                 if (lines[i+1])
 871                                         fail(filename, i+2,
 872                                              "lines after tdb_close");
 873                                 *num = i;
 874                                 talloc_free(lines);
 875                                 return maybe_cancel_transaction(filename,
 876                                                                 op, num);
 877                         }
 878                         fail(filename, i+1, "Unknown operation '%s'", words[1]);
 879                 }
 880
 881                 add_op(filename, &op, i, atoi(words[0]), opt->type);
 882                 opt->enhance_op(filename, op, i, words);
 883         }
 884
 885         fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
 886               filename, i);
 887         talloc_free(lines);
 888         *num = i - 1;
 889         return maybe_cancel_transaction(filename, op, num);
 890 }
 891
 892 /* We remember all the keys we've ever seen, and who has them. */
 893 struct keyinfo {
 894         TDB_DATA key;
 895         unsigned int num_users;
 896         struct op_desc *user;
 897 };
 898
 899 static const TDB_DATA must_not_exist;
 900 static const TDB_DATA must_exist;
 901 static const TDB_DATA not_exists_or_empty;
 902
 903 /* NULL means doesn't care if it exists or not, &must_exist means
 904  * it must exist but we don't care what, &must_not_exist means it must
 905  * not exist, otherwise the data it needs. */
 906 static const TDB_DATA *needs(const struct op *op)
 907 {
 908         switch (op->type) {
 909         /* FIXME: Pull forward deps, since we can deadlock */
 910         case OP_TDB_CHAINLOCK:
 911         case OP_TDB_CHAINLOCK_NONBLOCK:
 912         case OP_TDB_CHAINLOCK_MARK:
 913         case OP_TDB_CHAINLOCK_UNMARK:
 914         case OP_TDB_CHAINUNLOCK:
 915         case OP_TDB_CHAINLOCK_READ:
 916         case OP_TDB_CHAINUNLOCK_READ:
 917                 return NULL;
 918
 919         case OP_TDB_APPEND:
 920                 if (op->append.pre.dsize == 0)
 921                         return &not_exists_or_empty;
 922                 return &op->append.pre;
 923
 924         case OP_TDB_STORE:
 925                 if (op->flag == TDB_INSERT) {
 926                         if (op->ret < 0)
 927                                 return &must_exist;
 928                         else
 929                                 return &must_not_exist;
 930                 } else if (op->flag == TDB_MODIFY) {
 931                         if (op->ret < 0)
 932                                 return &must_not_exist;
 933                         else
 934                                 return &must_exist;
 935                 }
 936                 /* No flags?  Don't care */
 937                 return NULL;
 938
 939         case OP_TDB_EXISTS:
 940                 if (op->ret == 1)
 941                         return &must_exist;
 942                 else
 943                         return &must_not_exist;
 944
 945         case OP_TDB_PARSE_RECORD:
 946                 if (op->ret < 0)
 947                         return &must_not_exist;
 948                 return &must_exist;
 949
 950         /* FIXME: handle these. */
 951         case OP_TDB_WIPE_ALL:
 952         case OP_TDB_FIRSTKEY:
 953         case OP_TDB_NEXTKEY:
 954         case OP_TDB_GET_SEQNUM:
 955         case OP_TDB_TRAVERSE:
 956         case OP_TDB_TRANSACTION_COMMIT:
 957         case OP_TDB_TRANSACTION_CANCEL:
 958         case OP_TDB_TRANSACTION_START:
 959                 return NULL;
 960
 961         case OP_TDB_FETCH:
 962                 if (!op->data.dptr)
 963                         return &must_not_exist;
 964                 return &op->data;
 965
 966         case OP_TDB_DELETE:
 967                 if (op->ret < 0)
 968                         return &must_not_exist;
 969                 return &must_exist;
 970
 971         default:
 972                 errx(1, "Unexpected op type %i", op->type);
 973         }
 974
 975 }
 976
 977 static bool starts_transaction(const struct op *op)
 978 {
 979         return op->type == OP_TDB_TRANSACTION_START;
 980 }
 981
 982 static bool in_transaction(const struct op op[], unsigned int i)
 983 {
 984         return op[i].group_start && starts_transaction(&op[op[i].group_start]);
 985 }
 986
 987 static bool successful_transaction(const struct op *op)
 988 {
 989         return starts_transaction(op)
 990                 && op[op->group_len].type == OP_TDB_TRANSACTION_COMMIT;
 991 }
 992
 993 static bool starts_traverse(const struct op *op)
 994 {
 995         return op->type == OP_TDB_TRAVERSE_START
 996                 || op->type == OP_TDB_TRAVERSE_READ_START;
 997 }
 998
 999 static bool in_traverse(const struct op op[], unsigned int i)
1000 {
1001         return op[i].group_start && starts_traverse(&op[op[i].group_start]);
1002 }
1003
1004 static bool starts_chainlock(const struct op *op)
1005 {
1006         return op->type == OP_TDB_CHAINLOCK_READ
1007                 || op->type == OP_TDB_CHAINLOCK;
1008 }
1009
1010 static bool in_chainlock(const struct op op[], unsigned int i)
1011 {
1012         return op[i].group_start && starts_chainlock(&op[op[i].group_start]);
1013 }
1014
1015 /* What's the data after this op?  pre if nothing changed. */
1016 static const TDB_DATA *gives(const TDB_DATA *key, const TDB_DATA *pre,
1017                              const struct op *op)
1018 {
1019         if (starts_transaction(op) || starts_chainlock(op)) {
1020                 unsigned int i;
1021
1022                 /* Cancelled transactions don't change anything. */
1023                 if (op[op->group_len].type == OP_TDB_TRANSACTION_CANCEL)
1024                         return pre;
1025                 assert(op[op->group_len].type == OP_TDB_TRANSACTION_COMMIT
1026                        || op[op->group_len].type == OP_TDB_CHAINUNLOCK_READ
1027                        || op[op->group_len].type == OP_TDB_CHAINUNLOCK);
1028
1029                 for (i = 1; i < op->group_len; i++) {
1030                         /* This skips nested transactions, too */
1031                         if (key_eq(op[i].key, *key))
1032                                 pre = gives(key, pre, &op[i]);
1033                 }
1034                 return pre;
1035         }
1036
1037         /* Failed ops don't change state of db. */
1038         if (op->ret < 0)
1039                 return pre;
1040
1041         if (op->type == OP_TDB_DELETE || op->type == OP_TDB_WIPE_ALL)
1042                 return &tdb_null;
1043
1044         if (op->type == OP_TDB_APPEND)
1045                 return &op->append.post;
1046
1047         if (op->type == OP_TDB_STORE)
1048                 return &op->data;
1049
1050         return pre;
1051 }
1052
1053 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
1054                                 unsigned int num)
1055 {
1056         unsigned int i, j, h;
1057         struct keyinfo *hash;
1058
1059         hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
1060         for (i = 0; i < num; i++) {
1061                 for (j = 1; j < num_ops[i]; j++) {
1062                         /* We can't do this on allocation, due to realloc. */
1063                         list_head_init(&op[i][j].post);
1064                         list_head_init(&op[i][j].pre);
1065
1066                         if (!op[i][j].key.dptr)
1067                                 continue;
1068
1069                         h = hash_key(&op[i][j].key) % (total_keys * 2);
1070                         while (!key_eq(hash[h].key, op[i][j].key)) {
1071                                 if (!hash[h].key.dptr) {
1072                                         hash[h].key = op[i][j].key;
1073                                         break;
1074                                 }
1075                                 h = (h + 1) % (total_keys * 2);
1076                         }
1077                         /* Might as well save some memory if we can. */
1078                         if (op[i][j].key.dptr != hash[h].key.dptr) {
1079                                 talloc_free(op[i][j].key.dptr);
1080                                 op[i][j].key.dptr = hash[h].key.dptr;
1081                         }
1082                         hash[h].user = talloc_realloc(hash, hash[h].user,
1083                                                      struct op_desc,
1084                                                      hash[h].num_users+1);
1085
1086                         /* If it's in a transaction, it's the transaction which
1087                          * matters from an analysis POV. */
1088                         if (in_transaction(op[i], j)
1089                             || in_chainlock(op[i], j)) {
1090                                 unsigned start = op[i][j].group_start;
1091
1092                                 /* Don't include twice. */
1093                                 if (hash[h].num_users
1094                                     && hash[h].user[hash[h].num_users-1].file
1095                                         == i
1096                                     && hash[h].user[hash[h].num_users-1].op_num
1097                                         == start)
1098                                         continue;
1099
1100                                 hash[h].user[hash[h].num_users].op_num = start;
1101                         } else
1102                                 hash[h].user[hash[h].num_users].op_num = j;
1103                         hash[h].user[hash[h].num_users].file = i;
1104                         hash[h].num_users++;
1105                 }
1106         }
1107
1108         return hash;
1109 }
1110
1111 static bool satisfies(const TDB_DATA *key, const TDB_DATA *data,
1112                       const struct op *op)
1113 {
1114         const TDB_DATA *need = NULL;
1115
1116         if (starts_transaction(op) || starts_chainlock(op)) {
1117                 unsigned int i;
1118
1119                 /* Look through for an op in this transaction which
1120                  * needs this key. */
1121                 for (i = 1; i < op->group_len; i++) {
1122                         if (key_eq(op[i].key, *key)) {
1123                                 need = needs(&op[i]);
1124                                 /* tdb_exists() is special: there might be
1125                                  * something in the transaction with more
1126                                  * specific requirements.  Other ops don't have
1127                                  * specific requirements (eg. store or delete),
1128                                  * but they change the value so we can't get
1129                                  * more information from future ops. */
1130                                 if (op[i].type != OP_TDB_EXISTS)
1131                                         break;
1132                         }
1133                 }
1134         } else
1135                 need = needs(op);
1136
1137         /* Don't need anything?  Cool. */
1138         if (!need)
1139                 return true;
1140
1141         /* This should be tdb_null or a real value. */
1142         assert(data != &must_exist);
1143         assert(data != &must_not_exist);
1144         assert(data != &not_exists_or_empty);
1145
1146         /* Must not exist?  data must not exist. */
1147         if (need == &must_not_exist)
1148                 return data == &tdb_null;
1149
1150         /* Must exist? */
1151         if (need == &must_exist)
1152                 return data != &tdb_null;
1153
1154         /* Either noexist or empty. */
1155         if (need == &not_exists_or_empty)
1156                 return data->dsize == 0;
1157
1158         /* Needs something specific. */
1159         return key_eq(*data, *need);
1160 }
1161
1162 static void move_to_front(struct op_desc res[], unsigned off, unsigned elem)
1163 {
1164         if (elem != off) {
1165                 struct op_desc tmp = res[elem];
1166                 memmove(res + off + 1, res + off, (elem - off)*sizeof(res[0]));
1167                 res[off] = tmp;
1168         }
1169 }
1170
1171 static void restore_to_pos(struct op_desc res[], unsigned off, unsigned elem)
1172 {
1173         if (elem != off) {
1174                 struct op_desc tmp = res[off];
1175                 memmove(res + off, res + off + 1, (elem - off)*sizeof(res[0]));
1176                 res[elem] = tmp;
1177         }
1178 }
1179
1180 static bool sort_deps(char *filename[], struct op *op[],
1181                       struct op_desc res[],
1182                       unsigned off, unsigned num,
1183                       const TDB_DATA *key, const TDB_DATA *data,
1184                       unsigned num_files, unsigned fuzz)
1185 {
1186         unsigned int i, files_done;
1187         struct op *this_op;
1188         bool done[num_files];
1189
1190         /* None left?  We're sorted. */
1191         if (off == num)
1192                 return true;
1193
1194         /* Does this make sequence number go backwards?  Allow a little fuzz. */
1195         if (off > 0) {
1196                 int seqnum1 = op[res[off-1].file][res[off-1].op_num].seqnum;
1197                 int seqnum2 = op[res[off].file][res[off].op_num].seqnum;
1198
1199                 if (seqnum1 - seqnum2 > (int)fuzz) {
1200 #if DEBUG_DEPS
1201                         printf("Seqnum jump too far (%u -> %u)\n",
1202                                seqnum1, seqnum2);
1203 #endif
1204                         return false;
1205                 }
1206         }
1207
1208         memset(done, 0, sizeof(done));
1209
1210         /* Since ops within a trace file are ordered, we just need to figure
1211          * out which file to try next.  Since we don't take into account
1212          * inter-key relationships (which exist by virtue of trace file order),
1213          * we minimize the chance of harm by trying to keep in seqnum order. */
1214         for (files_done = 0, i = off; i < num && files_done < num_files; i++) {
1215                 if (done[res[i].file])
1216                         continue;
1217
1218                 this_op = &op[res[i].file][res[i].op_num];
1219
1220                 /* Is what we have good enough for this op? */
1221                 if (satisfies(key, data, this_op)) {
1222                         move_to_front(res, off, i);
1223                         if (sort_deps(filename, op, res, off+1, num,
1224                                       key, gives(key, data, this_op),
1225                                       num_files, fuzz))
1226                                 return true;
1227                         restore_to_pos(res, off, i);
1228                 }
1229                 done[res[i].file] = true;
1230                 files_done++;
1231         }
1232
1233         /* No combination worked. */
1234         return false;
1235 }
1236
1237 static void check_dep_sorting(struct op_desc user[], unsigned num_users,
1238                               unsigned num_files)
1239 {
1240 #if DEBUG_DEPS
1241         unsigned int i;
1242         unsigned minima[num_files];
1243
1244         memset(minima, 0, sizeof(minima));
1245         for (i = 0; i < num_users; i++) {
1246                 assert(minima[user[i].file] < user[i].op_num);
1247                 minima[user[i].file] = user[i].op_num;
1248         }
1249 #endif
1250 }
1251
1252 /* All these ops happen on the same key.  Which comes first?
1253  *
1254  * This can happen both because read ops or failed write ops don't
1255  * change sequence number, and also due to race since we access the
1256  * number unlocked (the race can cause less detectable ordering problems,
1257  * in which case we'll deadlock and report: fix manually in that case).
1258  */
1259 static void figure_deps(char *filename[], struct op *op[],
1260                         const TDB_DATA *key, struct op_desc user[],
1261                         unsigned num_users, unsigned num_files)
1262 {
1263         /* We assume database starts empty. */
1264         const struct TDB_DATA *data = &tdb_null;
1265         unsigned int fuzz;
1266
1267         /* We prefer to keep strict seqnum order if possible: it's the
1268          * most likely.  We get more lax if that fails. */
1269         for (fuzz = 0; fuzz < 100; fuzz = (fuzz + 1)*2) {
1270                 if (sort_deps(filename, op, user, 0, num_users, key, data,
1271                               num_files, fuzz))
1272                         break;
1273         }
1274
1275         if (fuzz >= 100)
1276                 fail(filename[user[0].file], user[0].op_num+1,
1277                      "Could not resolve inter-dependencies");
1278
1279         check_dep_sorting(user, num_users, num_files);
1280 }
1281
1282 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1283                      unsigned int num)
1284 {
1285         unsigned int h;
1286
1287         /* Gcc nexted function extension.  How cool is this? */
1288         int compare_seqnum(const void *_a, const void *_b)
1289         {
1290                 const struct op_desc *a = _a, *b = _b;
1291
1292                 /* First, maintain order within any trace file. */
1293                 if (a->file == b->file)
1294                         return a->op_num - b->op_num;
1295
1296                 /* Otherwise, arrange by seqnum order. */
1297                 if (op[a->file][a->op_num].seqnum !=
1298                     op[b->file][b->op_num].seqnum)
1299                         return op[a->file][a->op_num].seqnum
1300                                 - op[b->file][b->op_num].seqnum;
1301
1302                 /* Cancelled transactions are assumed to happen first. */
1303                 if (starts_transaction(&op[a->file][a->op_num])
1304                     && !successful_transaction(&op[a->file][a->op_num]))
1305                         return -1;
1306                 if (starts_transaction(&op[b->file][b->op_num])
1307                     && !successful_transaction(&op[b->file][b->op_num]))
1308                         return 1;
1309
1310                 /* No idea. */
1311                 return 0;
1312         }
1313
1314         /* Now sort into seqnum order. */
1315         for (h = 0; h < total_keys * 2; h++) {
1316                 struct op_desc *user = hash[h].user;
1317
1318                 qsort(user, hash[h].num_users, sizeof(user[0]), compare_seqnum);
1319                 figure_deps(filename, op, &hash[h].key, user, hash[h].num_users,
1320                             num);
1321         }
1322 }
1323
1324 static int destroy_depend(struct depend *dep)
1325 {
1326         list_del(&dep->pre_list);
1327         list_del(&dep->post_list);
1328         return 0;
1329 }
1330
1331 static void add_dependency(void *ctx,
1332                            struct op *op[],
1333                            char *filename[],
1334                            const struct op_desc *needs,
1335                            const struct op_desc *prereq)
1336 {
1337         struct depend *dep;
1338
1339         /* We don't depend on ourselves. */
1340         if (needs->file == prereq->file) {
1341                 assert(prereq->op_num < needs->op_num);
1342                 return;
1343         }
1344
1345 #if DEBUG_DEPS
1346         printf("%s:%u: depends on %s:%u\n",
1347                filename[needs->file], needs->op_num+1,
1348                filename[prereq->file], prereq->op_num+1);
1349 #endif
1350
1351         dep = talloc(ctx, struct depend);
1352         dep->needs = *needs;
1353         dep->prereq = *prereq;
1354
1355 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1356         /* If something in a traverse depends on something in another
1357          * traverse/transaction, it creates a dependency between the
1358          * two groups. */
1359         if ((in_traverse(op[prereq->file], prereq->op_num)
1360              && (starts_transaction(&op[needs->file][needs->op_num])
1361                  || starts_traverse(&op[needs->file][needs->op_num])))
1362             || (in_traverse(op[needs->file], needs->op_num)
1363                 && (starts_transaction(&op[prereq->file][prereq->op_num])
1364                     || starts_traverse(&op[prereq->file][prereq->op_num])))) {
1365                 unsigned int start;
1366
1367                 /* We are satisfied by end of group. */
1368                 start = op[prereq->file][prereq->op_num].group_start;
1369                 dep->prereq.op_num = start + op[prereq->file][start].group_len;
1370                 /* And we need that done by start of our group. */
1371                 dep->needs.op_num = op[needs->file][needs->op_num].group_start;
1372         }
1373
1374         /* There is also this case:
1375          *  <traverse> <read foo> ...
1376          *  <transaction> ... </transaction> <create foo>
1377          * Where if we start the traverse then wait, we could block
1378          * the transaction and deadlock.
1379          *
1380          * We try to address this by ensuring that where seqnum indicates it's
1381          * possible, we wait for <create foo> before *starting* traverse.
1382          */
1383         else if (in_traverse(op[needs->file], needs->op_num)) {
1384                 struct op *need = &op[needs->file][needs->op_num];
1385                 if (op[needs->file][need->group_start].seqnum >
1386                     op[prereq->file][prereq->op_num].seqnum) {
1387                         dep->needs.op_num = need->group_start;
1388                 }
1389         }
1390 #endif
1391
1392         /* If you depend on a transaction or chainlock, you actually
1393          * depend on it ending. */
1394         if (starts_transaction(&op[prereq->file][dep->prereq.op_num])
1395             || starts_chainlock(&op[prereq->file][dep->prereq.op_num])) {
1396                 dep->prereq.op_num
1397                         += op[dep->prereq.file][dep->prereq.op_num].group_len;
1398 #if DEBUG_DEPS
1399                 printf("-> Actually end of transaction %s:%u\n",
1400                        filename[dep->prereq->file], dep->prereq->op_num+1);
1401 #endif
1402         } else
1403                 /* We should never create a dependency from middle of
1404                  * a transaction. */
1405                 assert(!in_transaction(op[prereq->file], dep->prereq.op_num)
1406                        || op[prereq->file][dep->prereq.op_num].type
1407                        == OP_TDB_TRANSACTION_COMMIT
1408                        || op[prereq->file][dep->prereq.op_num].type
1409                        == OP_TDB_TRANSACTION_CANCEL);
1410
1411         list_add(&op[dep->prereq.file][dep->prereq.op_num].post,
1412                  &dep->post_list);
1413         list_add(&op[dep->needs.file][dep->needs.op_num].pre,
1414                  &dep->pre_list);
1415         talloc_set_destructor(dep, destroy_depend);
1416 }
1417
1418 static bool changes_db(const TDB_DATA *key, const struct op *op)
1419 {
1420         return gives(key, NULL, op) != NULL;
1421 }
1422
1423 static void depend_on_previous(struct op *op[],
1424                                char *filename[],
1425                                unsigned int num,
1426                                struct op_desc user[],
1427                                unsigned int i,
1428                                int prev)
1429 {
1430         bool deps[num];
1431         int j;
1432
1433         if (i == 0)
1434                 return;
1435
1436         if (prev == i - 1) {
1437                 /* Just depend on previous. */
1438                 add_dependency(NULL, op, filename, &user[i], &user[prev]);
1439                 return;
1440         }
1441
1442         /* We have to wait for the readers.  Find last one in *each* file. */
1443         memset(deps, 0, sizeof(deps));
1444         deps[user[i].file] = true;
1445         for (j = i - 1; j > prev; j--) {
1446                 if (!deps[user[j].file]) {
1447                         add_dependency(NULL, op, filename, &user[i], &user[j]);
1448                         deps[user[j].file] = true;
1449                 }
1450         }
1451 }
1452
1453 /* This is simple, but not complete.  We don't take into account
1454  * indirect dependencies. */
1455 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1456                                   unsigned int num)
1457 {
1458         unsigned int i, j;
1459
1460         /* There can only be one real dependency on each file */
1461         for (i = 0; i < num; i++) {
1462                 for (j = 1; j < num_ops[i]; j++) {
1463                         struct depend *dep, *next;
1464                         struct depend *prev[num];
1465
1466                         memset(prev, 0, sizeof(prev));
1467
1468                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1469                                 if (!prev[dep->prereq.file]) {
1470                                         prev[dep->prereq.file] = dep;
1471                                         continue;
1472                                 }
1473                                 if (prev[dep->prereq.file]->prereq.op_num
1474                                     < dep->prereq.op_num) {
1475                                         talloc_free(prev[dep->prereq.file]);
1476                                         prev[dep->prereq.file] = dep;
1477                                 } else
1478                                         talloc_free(dep);
1479                         }
1480                 }
1481         }
1482
1483         for (i = 0; i < num; i++) {
1484                 int deps[num];
1485
1486                 for (j = 0; j < num; j++)
1487                         deps[j] = -1;
1488
1489                 for (j = 1; j < num_ops[i]; j++) {
1490                         struct depend *dep, *next;
1491
1492                         list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1493                                 if (deps[dep->prereq.file]
1494                                     >= (int)dep->prereq.op_num)
1495                                         talloc_free(dep);
1496                                 else
1497                                         deps[dep->prereq.file]
1498                                                 = dep->prereq.op_num;
1499                         }
1500                 }
1501         }
1502 }
1503
1504 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1505 /* Force an order among the traversals, so they don't deadlock (as much) */
1506 static void make_traverse_depends(char *filename[],
1507                                   struct op *op[], unsigned int num_ops[],
1508                                   unsigned int num)
1509 {
1510         unsigned int i, num_traversals = 0;
1511         int j;
1512         struct op_desc *desc;
1513
1514         /* Sort by which one runs first. */
1515         int compare_traverse_desc(const void *_a, const void *_b)
1516         {
1517                 const struct op_desc *da = _a, *db = _b;
1518                 const struct op *a = &op[da->file][da->op_num],
1519                         *b = &op[db->file][db->op_num];
1520
1521                 if (a->seqnum != b->seqnum)
1522                         return a->seqnum - b->seqnum;
1523
1524                 /* If they have same seqnum, it means one didn't make any
1525                  * changes.  Thus sort by end in that case. */
1526                 return a[a->group_len].seqnum - b[b->group_len].seqnum;
1527         }
1528
1529         desc = talloc_array(NULL, struct op_desc, 1);
1530
1531         /* Count them. */
1532         for (i = 0; i < num; i++) {
1533                 for (j = 1; j < num_ops[i]; j++) {
1534                         /* Traverse start (ignore those in
1535                          * transactions; they're already covered by
1536                          * transaction dependencies). */
1537                         if (starts_traverse(&op[i][j])
1538                             && !in_transaction(op[i], j)) {
1539                                 desc = talloc_realloc(NULL, desc,
1540                                                       struct op_desc,
1541                                                       num_traversals+1);
1542                                 desc[num_traversals].file = i;
1543                                 desc[num_traversals].op_num = j;
1544                                 num_traversals++;
1545                         }
1546                 }
1547         }
1548         qsort(desc, num_traversals, sizeof(desc[0]), compare_traverse_desc);
1549
1550         for (i = 1; i < num_traversals; i++) {
1551                 const struct op *prev = &op[desc[i-1].file][desc[i-1].op_num];
1552                 const struct op *curr = &op[desc[i].file][desc[i].op_num];
1553
1554                 /* Read traverses don't depend on each other (read lock). */
1555                 if (prev->type == OP_TDB_TRAVERSE_READ_START
1556                     && curr->type == OP_TDB_TRAVERSE_READ_START)
1557                         continue;
1558
1559                 /* Only make dependency if it's clear. */
1560                 if (compare_traverse_desc(&desc[i], &desc[i-1])) {
1561                         /* i depends on end of traverse i-1. */
1562                         struct op_desc end = desc[i-1];
1563                         end.op_num += prev->group_len;
1564                         add_dependency(NULL, op, filename, &desc[i], &end);
1565                 }
1566         }
1567         talloc_free(desc);
1568 }
1569
1570 static void set_nonblock(int fd)
1571 {
1572         if (fcntl(fd, F_SETFL, fcntl(fd, F_GETFL)|O_NONBLOCK) != 0)
1573                 err(1, "Setting pipe nonblocking");
1574 }
1575
1576 static bool handle_backoff(struct op *op[], int fd)
1577 {
1578         struct op_desc desc;
1579         bool handled = false;
1580
1581         /* Sloppy coding: we assume PIPEBUF never fills. */
1582         while (read(fd, &desc, sizeof(desc)) != -1) {
1583                 unsigned int i;
1584                 handled = true;
1585                 for (i = desc.op_num; i > 0; i--) {
1586                         if (op[desc.file][i].type == OP_TDB_TRAVERSE) {
1587                                 /* We insert a fake end here. */
1588                                 op[desc.file][i].type
1589                                         = OP_TDB_TRAVERSE_END_EARLY;
1590                                 break;
1591                         } else if (starts_traverse(&op[desc.file][i])) {
1592                                 unsigned int start = i;
1593                                 struct op tmp = op[desc.file][i];
1594                                 /* Move the ops outside traverse. */
1595                                 memmove(&op[desc.file][i],
1596                                         &op[desc.file][i+1],
1597                                         (desc.op_num-i-1) * sizeof(op[0][0]));
1598                                 op[desc.file][desc.op_num] = tmp;
1599                                 while (op[desc.file][i].group_start == start) {
1600                                         op[desc.file][i++].group_start
1601                                                 = desc.op_num;
1602                                 }
1603                                 break;
1604                         }
1605                 }
1606         }
1607         return handled;
1608 }
1609
1610 #else /* !TRAVERSALS_TAKE_TRANSACTION_LOCK */
1611 static bool handle_backoff(struct op *op[], int fd)
1612 {
1613         return false;
1614 }
1615 #endif
1616
1617 static void derive_dependencies(char *filename[],
1618                                 struct op *op[], unsigned int num_ops[],
1619                                 unsigned int num)
1620 {
1621         struct keyinfo *hash;
1622         unsigned int h, i;
1623
1624         /* Create hash table for faster key lookup. */
1625         hash = hash_ops(op, num_ops, num);
1626
1627         /* Sort them by sequence number. */
1628         sort_ops(hash, filename, op, num);
1629
1630         /* Create dependencies back to the last change, rather than
1631          * creating false dependencies by naively making each one
1632          * depend on the previous.  This has two purposes: it makes
1633          * later optimization simpler, and it also avoids deadlock with
1634          * same sequence number ops inside traversals (if one
1635          * traversal doesn't write anything, two ops can have the same
1636          * sequence number yet we can create a traversal dependency
1637          * the other way). */
1638         for (h = 0; h < total_keys * 2; h++) {
1639                 int prev = -1;
1640
1641                 if (hash[h].num_users < 2)
1642                         continue;
1643
1644                 for (i = 0; i < hash[h].num_users; i++) {
1645                         if (changes_db(&hash[h].key, &op[hash[h].user[i].file]
1646                                        [hash[h].user[i].op_num])) {
1647                                 depend_on_previous(op, filename, num,
1648                                                    hash[h].user, i, prev);
1649                                 prev = i;
1650                         } else if (prev >= 0)
1651                                 add_dependency(hash, op, filename,
1652                                                &hash[h].user[i],
1653                                                &hash[h].user[prev]);
1654                 }
1655         }
1656
1657 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1658         make_traverse_depends(filename, op, num_ops, num);
1659 #endif
1660
1661         optimize_dependencies(op, num_ops, num);
1662 }
1663
1664 static struct timeval run_test(char *argv[],
1665                                unsigned int num_ops[],
1666                                unsigned int hashsize[],
1667                                unsigned int tdb_flags[],
1668                                unsigned int open_flags[],
1669                                struct op *op[],
1670                                int fds[2])
1671 {
1672         unsigned int i;
1673         struct timeval start, end, diff;
1674         bool ok = true;
1675
1676         for (i = 0; argv[i+2]; i++) {
1677                 struct tdb_context *tdb;
1678                 char c;
1679
1680                 switch (fork()) {
1681                 case -1:
1682                         err(1, "fork failed");
1683                 case 0:
1684                         close(fds[1]);
1685                         tdb = tdb_open_ex(argv[1], hashsize[i],
1686                                           tdb_flags[i]|TDB_NOSYNC,
1687                                           open_flags[i], 0600, NULL, hash_key);
1688                         if (!tdb)
1689                                 err(1, "Opening tdb %s", argv[1]);
1690
1691                         /* This catches parent exiting. */
1692                         if (read(fds[0], &c, 1) != 1)
1693                                 exit(1);
1694                         run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
1695                                 num_ops[i], false);
1696                         check_deps(argv[2+i], op[i], num_ops[i]);
1697                         exit(0);
1698                 default:
1699                         break;
1700                 }
1701         }
1702
1703         /* Let everything settle. */
1704         sleep(1);
1705
1706         printf("Starting run...");
1707         fflush(stdout);
1708         gettimeofday(&start, NULL);
1709         /* Tell them all to go!  Any write of sufficient length will do. */
1710         if (write(fds[1], hashsize, i) != i)
1711                 err(1, "Writing to wakeup pipe");
1712
1713         for (i = 0; argv[i + 2]; i++) {
1714                 int status;
1715                 wait(&status);
1716                 if (!WIFEXITED(status)) {
1717                         warnx("Child died with signal %i", WTERMSIG(status));
1718                         ok = false;
1719                 } else if (WEXITSTATUS(status) != 0)
1720                         /* Assume child spat out error. */
1721                         ok = false;
1722         }
1723         if (!ok)
1724                 exit(1);
1725
1726         gettimeofday(&end, NULL);
1727         printf("done\n");
1728
1729         if (end.tv_usec < start.tv_usec) {
1730                 end.tv_usec += 1000000;
1731                 end.tv_sec--;
1732         }
1733         diff.tv_sec = end.tv_sec - start.tv_sec;
1734         diff.tv_usec = end.tv_usec - start.tv_usec;
1735         return diff;
1736 }
1737
1738 int main(int argc, char *argv[])
1739 {
1740         struct timeval diff;
1741         unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1742         struct op *op[argc];
1743         int fds[2];
1744
1745         if (argc < 3)
1746                 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1747
1748         pipes = talloc_array(NULL, struct pipe, argc - 1);
1749         for (i = 0; i < argc - 2; i++) {
1750                 printf("Loading tracefile %s...", argv[2+i]);
1751                 fflush(stdout);
1752                 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1753                                        &tdb_flags[i], &open_flags[i]);
1754                 if (pipe(pipes[i].fd) != 0)
1755                         err(1, "creating pipe");
1756                 printf("done\n");
1757         }
1758
1759         printf("Calculating inter-dependencies...");
1760         fflush(stdout);
1761         derive_dependencies(argv+2, op, num_ops, i);
1762         printf("done\n");
1763
1764         /* Don't fork for single arg case: simple debugging. */
1765         if (argc == 3) {
1766                 struct tdb_context *tdb;
1767                 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0]|TDB_NOSYNC,
1768                                   open_flags[0], 0600, NULL, hash_key);
1769                 printf("Single threaded run...");
1770                 fflush(stdout);
1771
1772                 run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0],
1773                         false);
1774                 check_deps(argv[2], op[0], num_ops[0]);
1775
1776                 printf("done\n");
1777                 exit(0);
1778         }
1779
1780         if (pipe(fds) != 0)
1781                 err(1, "creating pipe");
1782
1783 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1784         if (pipe(pipes[argc-2].fd) != 0)
1785                 err(1, "creating pipe");
1786         backoff_fd = pipes[argc-2].fd[1];
1787         set_nonblock(pipes[argc-2].fd[1]);
1788         set_nonblock(pipes[argc-2].fd[0]);
1789 #endif
1790
1791         do {
1792                 diff = run_test(argv, num_ops, hashsize, tdb_flags, open_flags,
1793                                 op, fds);
1794         } while (handle_backoff(op, pipes[argc-2].fd[0]));
1795
1796         printf("Time replaying: %lu usec\n",
1797                diff.tv_sec * 1000000UL + diff.tv_usec);
1798
1799         exit(0);
1800 }