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>
12 #include <sys/types.h>
19 #define STRINGIFY2(x) #x
20 #define STRINGIFY(x) STRINGIFY2(x)
22 /* Avoid mod by zero */
23 static unsigned int total_keys = 1;
27 /* Traversals block transactions in the current implementation. */
28 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
33 static struct pipe *pipes;
35 static void __attribute__((noreturn)) fail(const char *filename,
42 fprintf(stderr, "%s:%u: FAIL: ", filename, line);
43 vfprintf(stderr, fmt, ap);
44 fprintf(stderr, "\n");
50 #define try(expr, expect) \
53 if (ret != (expect)) \
54 fail(filename[file], i+1, \
55 STRINGIFY(expr) "= %i", ret); \
58 /* Try or imitate results. */
59 #define unreliable(expr, expect, force, undo) \
62 if (ret != expect) { \
63 fprintf(stderr, "%s:%u: %s gave %i not %i", \
64 filename[file], i+1, STRINGIFY(expr), \
73 static bool key_eq(TDB_DATA a, TDB_DATA b)
75 if (a.dsize != b.dsize)
77 return memcmp(a.dptr, b.dptr, a.dsize) == 0;
80 /* This is based on the hash algorithm from gdbm */
81 static unsigned int hash_key(TDB_DATA *key)
83 uint32_t value; /* Used to compute the hash value. */
84 uint32_t i; /* Used to cycle through random values. */
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)));
90 return (1103515243 * value + 12345);
96 OP_TDB_LOCKALL_UNMARK,
97 OP_TDB_LOCKALL_NONBLOCK,
100 OP_TDB_LOCKALL_READ_NONBLOCK,
101 OP_TDB_UNLOCKALL_READ,
103 OP_TDB_CHAINLOCK_NONBLOCK,
104 OP_TDB_CHAINLOCK_MARK,
105 OP_TDB_CHAINLOCK_UNMARK,
107 OP_TDB_CHAINLOCK_READ,
108 OP_TDB_CHAINUNLOCK_READ,
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,
135 /* Who is waiting for us? */
136 struct list_head post;
137 /* What are we waiting for? */
138 struct list_head pre;
140 /* If I'm part of a group (traverse/transaction) where is
141 * start? (Otherwise, 0) */
142 unsigned int group_start;
145 int flag; /* open and store */
146 struct { /* append */
150 unsigned int group_len; /* transaction/traverse start */
154 static unsigned char hex_char(const char *filename, unsigned int line, char c)
157 if (c >= 'A' && c <= 'F')
159 if (c >= '0' && c <= '9')
161 fail(filename, line, "invalid hex character '%c'", c);
164 /* TDB data is <size>:<%02x>* */
165 static TDB_DATA make_tdb_data(const void *ctx,
166 const char *filename, unsigned int line,
173 if (streq(word, "NULL"))
176 data.dsize = atoi(word);
177 data.dptr = talloc_array(ctx, unsigned char, data.dsize);
178 p = strchr(word, ':');
180 fail(filename, line, "invalid tdb data '%s'", word);
182 for (i = 0; i < data.dsize; i++)
183 data.dptr[i] = hex_char(filename, line, p[i*2])*16
184 + hex_char(filename, line, p[i*2+1]);
189 static void add_op(const char *filename, struct op **op, unsigned int i,
190 unsigned int serial, enum op_type type)
193 *op = talloc_realloc(NULL, *op, struct op, i+1);
196 new->serial = serial;
198 new->group_start = 0;
201 static void op_add_nothing(const char *filename,
202 struct op op[], unsigned int op_num, char *words[])
205 fail(filename, op_num+1, "Expected no arguments");
206 op[op_num].key = tdb_null;
209 static void op_add_key(const char *filename,
210 struct op op[], unsigned int op_num, char *words[])
212 if (words[2] == NULL || words[3])
213 fail(filename, op_num+1, "Expected just a key");
215 op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
216 if (op[op_num].op != OP_TDB_TRAVERSE)
220 static void op_add_key_ret(const char *filename,
221 struct op op[], unsigned int op_num, char *words[])
223 if (!words[2] || !words[3] || !words[4] || words[5]
224 || !streq(words[3], "="))
225 fail(filename, op_num+1, "Expected <key> = <ret>");
226 op[op_num].ret = atoi(words[4]);
227 op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
228 /* May only be a unique key if it fails */
229 if (op[op_num].ret != 0)
233 static void op_add_key_data(const char *filename,
234 struct op op[], unsigned int op_num, char *words[])
236 if (!words[2] || !words[3] || !words[4] || words[5]
237 || !streq(words[3], "="))
238 fail(filename, op_num+1, "Expected <key> = <data>");
239 op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
240 op[op_num].data = make_tdb_data(op, filename, op_num+1, words[4]);
241 /* May only be a unique key if it fails */
242 if (!op[op_num].data.dptr)
246 /* <serial> tdb_store <rec> <rec> <flag> = <ret> */
247 static void op_add_store(const char *filename,
248 struct op op[], unsigned int op_num, char *words[])
250 if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
251 || words[7] || !streq(words[5], "="))
252 fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
254 op[op_num].flag = strtoul(words[4], NULL, 0);
255 op[op_num].ret = atoi(words[6]);
256 op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
257 op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
261 /* <serial> tdb_append <rec> <rec> = <rec> */
262 static void op_add_append(const char *filename,
263 struct op op[], unsigned int op_num, char *words[])
265 if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
266 || !streq(words[4], "="))
267 fail(filename, op_num+1, "Expect <key> <data> = <rec>");
269 op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
270 op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
272 op[op_num].append.post
273 = make_tdb_data(op, filename, op_num+1, words[5]);
275 /* By subtraction, figure out what previous data was. */
276 op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
277 op[op_num].append.pre.dsize
278 = op[op_num].append.post.dsize - op[op_num].data.dsize;
282 /* <serial> tdb_get_seqnum = <ret> */
283 static void op_add_seqnum(const char *filename,
284 struct op op[], unsigned int op_num, char *words[])
286 if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
287 fail(filename, op_num+1, "Expect = <ret>");
289 op[op_num].key = tdb_null;
290 op[op_num].ret = atoi(words[3]);
293 static void op_add_traverse(const char *filename,
294 struct op op[], unsigned int op_num, char *words[])
297 fail(filename, op_num+1, "Expect no arguments");
299 op[op_num].key = tdb_null;
300 op[op_num].group_len = 0;
303 static void op_add_transaction(const char *filename, struct op op[],
304 unsigned int op_num, char *words[])
307 fail(filename, op_num+1, "Expect no arguments");
309 op[op_num].key = tdb_null;
310 op[op_num].group_len = 0;
313 static int op_transaction_start(struct op op[], unsigned int op_num)
317 for (i = op_num-1; i > 0; i--) {
318 if (op[i].op == OP_TDB_TRANSACTION_START && !op[i].group_len)
324 static void op_analyze_transaction(const char *filename,
325 struct op op[], unsigned int op_num,
328 unsigned int start, i;
330 op[op_num].key = tdb_null;
333 fail(filename, op_num+1, "Expect no arguments");
335 start = op_transaction_start(op, op_num);
337 fail(filename, op_num+1, "no transaction start found");
339 op[start].group_len = op_num - start;
341 /* This rolls in nested transactions. I think that's right. */
342 for (i = start; i <= op_num; i++)
343 op[i].group_start = start;
346 struct traverse_hash {
351 static void op_analyze_traverse(const char *filename,
352 struct op op[], unsigned int op_num,
357 op[op_num].key = tdb_null;
359 /* = %u means traverse function terminated. */
361 if (!streq(words[2], "=") || !words[3] || words[4])
362 fail(filename, op_num+1, "expect = <num>");
363 op[op_num].ret = atoi(words[3]);
367 for (i = op_num-1; i >= 0; i--) {
368 if (op[i].op != OP_TDB_TRAVERSE_READ_START
369 && op[i].op != OP_TDB_TRAVERSE_START)
377 fail(filename, op_num+1, "no traversal start found");
380 op[start].group_len = op_num - start;
382 for (i = start; i <= op_num; i++)
383 op[i].group_start = start;
386 /* Keep -Wmissing-declarations happy: */
387 const struct op_table *
388 find_keyword (register const char *str, register unsigned int len);
390 #include "keywords.c"
393 /* We can have more than one */
394 struct list_node pre_list;
395 struct list_node post_list;
396 unsigned int needs_file;
397 unsigned int needs_opnum;
398 unsigned int satisfies_file;
399 unsigned int satisfies_opnum;
402 static void check_deps(const char *filename, struct op op[], unsigned int num)
407 for (i = 1; i < num; i++)
408 if (!list_empty(&op[i].pre))
409 fail(filename, i+1, "Still has dependencies");
413 static void dump_pre(char *filename[], struct op *op[],
414 unsigned int file, unsigned int i)
418 printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
420 list_for_each(&op[file][i].pre, dep, pre_list)
421 printf(" %s:%u (%u)\n",
422 filename[dep->satisfies_file], dep->satisfies_opnum+1,
423 op[dep->satisfies_file][dep->satisfies_opnum].serial);
424 check_deps(filename[file], op[file], i);
427 /* We simply read/write pointers, since we all are children. */
428 static void do_pre(struct tdb_context *tdb,
429 char *filename[], struct op *op[],
430 unsigned int file, int pre_fd, unsigned int i)
432 while (!list_empty(&op[file][i].pre)) {
436 printf("%s:%u:waiting for pre\n", filename[file], i+1);
440 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
441 if (errno == EINTR) {
442 dump_pre(filename, op, file, i);
445 errx(1, "Reading from pipe");
450 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
451 dep->needs_opnum+1, filename[dep->satisfies_file],
452 dep->satisfies_opnum+1);
455 /* This could be any op, not just this one. */
460 static void do_post(char *filename[], struct op *op[],
461 unsigned int file, unsigned int i)
465 list_for_each(&op[file][i].post, dep, post_list) {
467 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
468 filename[dep->needs_file], dep->needs_opnum+1);
470 if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
472 err(1, "%s:%u failed to tell file %s",
473 filename[file], i+1, filename[dep->needs_file]);
477 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
482 static unsigned run_ops(struct tdb_context *tdb,
487 unsigned int start, unsigned int stop);
489 struct traverse_info {
498 /* More complex. Just do whatever's they did at the n'th entry. */
499 static int nontrivial_traverse(struct tdb_context *tdb,
500 TDB_DATA key, TDB_DATA data,
503 struct traverse_info *tinfo = _tinfo;
504 unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
506 if (tinfo->i == tinfo->start + trav_len) {
507 /* This can happen if traverse expects to be empty. */
510 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
511 "traverse did not terminate");
514 if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
515 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
516 "%s:%u:traverse terminated early");
518 /* Run any normal ops. */
519 tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
520 tinfo->file, tinfo->i+1, tinfo->start + trav_len);
522 if (tinfo->i == tinfo->start + trav_len)
528 static unsigned op_traverse(struct tdb_context *tdb,
532 int (*traversefn)(struct tdb_context *,
533 tdb_traverse_func, void *),
537 struct traverse_info tinfo = { op, filename, file, pre_fd,
540 traversefn(tdb, nontrivial_traverse, &tinfo);
542 /* Traversing in wrong order can have strange effects: eg. if
543 * original traverse went A (delete A), B, we might do B
544 * (delete A). So if we have ops left over, we do it now. */
545 while (tinfo.i != start + op[file][start].group_len) {
546 if (op[file][tinfo.i].op == OP_TDB_TRAVERSE)
549 tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
551 start + op[file][start].group_len);
557 static void break_out(int sig)
561 static __attribute__((noinline))
562 unsigned run_ops(struct tdb_context *tdb,
567 unsigned int start, unsigned int stop)
572 sa.sa_handler = break_out;
575 sigaction(SIGALRM, &sa, NULL);
576 for (i = start; i < stop; i++) {
577 do_pre(tdb, filename, op, file, pre_fd, i);
579 switch (op[file][i].op) {
581 try(tdb_lockall(tdb), op[file][i].ret);
583 case OP_TDB_LOCKALL_MARK:
584 try(tdb_lockall_mark(tdb), op[file][i].ret);
586 case OP_TDB_LOCKALL_UNMARK:
587 try(tdb_lockall_unmark(tdb), op[file][i].ret);
589 case OP_TDB_LOCKALL_NONBLOCK:
590 unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
591 tdb_lockall(tdb), tdb_unlockall(tdb));
593 case OP_TDB_UNLOCKALL:
594 try(tdb_unlockall(tdb), op[file][i].ret);
596 case OP_TDB_LOCKALL_READ:
597 try(tdb_lockall_read(tdb), op[file][i].ret);
599 case OP_TDB_LOCKALL_READ_NONBLOCK:
600 unreliable(tdb_lockall_read_nonblock(tdb),
602 tdb_lockall_read(tdb),
603 tdb_unlockall_read(tdb));
605 case OP_TDB_UNLOCKALL_READ:
606 try(tdb_unlockall_read(tdb), op[file][i].ret);
608 case OP_TDB_CHAINLOCK:
609 try(tdb_chainlock(tdb, op[file][i].key),
612 case OP_TDB_CHAINLOCK_NONBLOCK:
613 unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
615 tdb_chainlock(tdb, op[file][i].key),
616 tdb_chainunlock(tdb, op[file][i].key));
618 case OP_TDB_CHAINLOCK_MARK:
619 try(tdb_chainlock_mark(tdb, op[file][i].key),
622 case OP_TDB_CHAINLOCK_UNMARK:
623 try(tdb_chainlock_unmark(tdb, op[file][i].key),
626 case OP_TDB_CHAINUNLOCK:
627 try(tdb_chainunlock(tdb, op[file][i].key),
630 case OP_TDB_CHAINLOCK_READ:
631 try(tdb_chainlock_read(tdb, op[file][i].key),
634 case OP_TDB_CHAINUNLOCK_READ:
635 try(tdb_chainunlock_read(tdb, op[file][i].key),
638 case OP_TDB_PARSE_RECORD:
639 try(tdb_parse_record(tdb, op[file][i].key, get_len,
644 try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
647 try(tdb_store(tdb, op[file][i].key, op[file][i].data,
652 try(tdb_append(tdb, op[file][i].key, op[file][i].data),
655 case OP_TDB_GET_SEQNUM:
656 try(tdb_get_seqnum(tdb), op[file][i].ret);
658 case OP_TDB_WIPE_ALL:
659 try(tdb_wipe_all(tdb), op[file][i].ret);
661 case OP_TDB_TRANSACTION_START:
662 try(tdb_transaction_start(tdb), op[file][i].ret);
664 case OP_TDB_TRANSACTION_CANCEL:
665 try(tdb_transaction_cancel(tdb), op[file][i].ret);
667 case OP_TDB_TRANSACTION_COMMIT:
668 try(tdb_transaction_commit(tdb), op[file][i].ret);
670 case OP_TDB_TRAVERSE_READ_START:
671 i = op_traverse(tdb, pre_fd, filename, file,
672 tdb_traverse_read, op, i);
674 case OP_TDB_TRAVERSE_START:
675 i = op_traverse(tdb, pre_fd, filename, file,
676 tdb_traverse, op, i);
678 case OP_TDB_TRAVERSE:
679 /* Terminate: we're in a traverse, and we've
682 case OP_TDB_TRAVERSE_END:
683 fail(filename[file], i+1, "unexpected end traverse");
684 /* FIXME: These must be treated like traverse. */
685 case OP_TDB_FIRSTKEY:
686 if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
687 fail(filename[file], i+1, "bad firstkey");
690 if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
692 fail(filename[file], i+1, "bad nextkey");
695 TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
696 if (!key_eq(f, op[file][i].data))
697 fail(filename[file], i+1, "bad fetch %u",
702 try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
705 do_post(filename, op, file, i);
710 /* tdbtorture, in particular, can do a tdb_close with a transaction in
712 static struct op *maybe_cancel_transaction(const char *filename,
713 struct op *op, unsigned int *num)
715 unsigned int start = op_transaction_start(op, *num);
718 char *words[] = { "<unknown>", "tdb_close", NULL };
719 add_op(filename, &op, *num, op[start].serial,
720 OP_TDB_TRANSACTION_CANCEL);
721 op_analyze_transaction(filename, op, *num, words);
727 static struct op *load_tracefile(const char *filename, unsigned int *num,
728 unsigned int *hashsize,
729 unsigned int *tdb_flags,
730 unsigned int *open_flags)
733 struct op *op = talloc_array(NULL, struct op, 1);
738 file = grab_file(NULL, filename, NULL);
740 err(1, "Reading %s", filename);
742 lines = strsplit(file, file, "\n", NULL);
744 errx(1, "%s is empty", filename);
746 words = strsplit(lines, lines[0], " ", NULL);
747 if (!streq(words[1], "tdb_open"))
748 fail(filename, 1, "does not start with tdb_open");
750 *hashsize = atoi(words[2]);
751 *tdb_flags = strtoul(words[3], NULL, 0);
752 *open_flags = strtoul(words[4], NULL, 0);
754 for (i = 1; lines[i]; i++) {
755 const struct op_table *opt;
757 words = strsplit(lines, lines[i], " ", NULL);
758 if (!words[0] || !words[1])
759 fail(filename, i+1, "Expected serial number and op");
761 opt = find_keyword(words[1], strlen(words[1]));
763 if (streq(words[1], "tdb_close")) {
766 "lines after tdb_close");
769 return maybe_cancel_transaction(filename,
772 fail(filename, i+1, "Unknown operation '%s'", words[1]);
775 add_op(filename, &op, i, atoi(words[0]), opt->type);
776 opt->enhance_op(filename, op, i, words);
779 fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
783 return maybe_cancel_transaction(filename, op, num);
786 /* We remember all the keys we've ever seen, and who has them. */
794 unsigned int num_users;
795 struct key_user *user;
798 static const TDB_DATA must_not_exist;
799 static const TDB_DATA must_exist;
800 static const TDB_DATA not_exists_or_empty;
802 /* NULL means doesn't care if it exists or not, &must_exist means
803 * it must exist but we don't care what, &must_not_exist means it must
804 * not exist, otherwise the data it needs. */
805 static const TDB_DATA *needs(const struct op *op)
808 /* FIXME: Pull forward deps, since we can deadlock */
809 case OP_TDB_CHAINLOCK:
810 case OP_TDB_CHAINLOCK_NONBLOCK:
811 case OP_TDB_CHAINLOCK_MARK:
812 case OP_TDB_CHAINLOCK_UNMARK:
813 case OP_TDB_CHAINUNLOCK:
814 case OP_TDB_CHAINLOCK_READ:
815 case OP_TDB_CHAINUNLOCK_READ:
819 if (op->append.pre.dsize == 0)
820 return ¬_exists_or_empty;
821 return &op->append.pre;
824 if (op->flag == TDB_INSERT) {
828 return &must_not_exist;
829 } else if (op->flag == TDB_MODIFY) {
831 return &must_not_exist;
835 /* No flags? Don't care */
842 return &must_not_exist;
844 case OP_TDB_PARSE_RECORD:
846 return &must_not_exist;
849 /* FIXME: handle these. */
850 case OP_TDB_WIPE_ALL:
851 case OP_TDB_FIRSTKEY:
853 case OP_TDB_GET_SEQNUM:
854 case OP_TDB_TRAVERSE:
855 case OP_TDB_TRANSACTION_COMMIT:
856 case OP_TDB_TRANSACTION_CANCEL:
857 case OP_TDB_TRANSACTION_START:
862 return &must_not_exist;
867 return &must_not_exist;
871 errx(1, "Unexpected op %i", op->op);
876 static bool is_transaction(const struct op *op)
878 return op->op == OP_TDB_TRANSACTION_START;
881 /* What's the data after this op? pre if nothing changed. */
882 static const TDB_DATA *gives(const TDB_DATA *key, const TDB_DATA *pre,
885 if (is_transaction(op)) {
888 /* Cancelled transactions don't change anything. */
889 if (op[op->group_len].op == OP_TDB_TRANSACTION_CANCEL)
891 assert(op[op->group_len].op == OP_TDB_TRANSACTION_COMMIT);
893 for (i = 1; i < op->group_len; i++) {
894 /* This skips nested transactions, too */
895 if (op[i].op != OP_TDB_TRAVERSE
896 && key_eq(op[i].key, *key))
897 pre = gives(key, pre, &op[i]);
902 /* Failed ops don't change state of db. */
906 if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
909 if (op->op == OP_TDB_APPEND)
910 return &op->append.post;
912 if (op->op == OP_TDB_STORE)
918 static bool in_transaction(const struct op op[], unsigned int i)
920 return op[i].group_start && is_transaction(&op[op[i].group_start]);
923 static bool in_traverse(const struct op op[], unsigned int i)
925 return op[i].group_start && !is_transaction(&op[op[i].group_start]);
928 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
931 unsigned int i, j, h;
932 struct keyinfo *hash;
934 hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
935 for (i = 0; i < num; i++) {
936 for (j = 1; j < num_ops[i]; j++) {
937 /* We can't do this on allocation, due to realloc. */
938 list_head_init(&op[i][j].post);
939 list_head_init(&op[i][j].pre);
941 if (!op[i][j].key.dptr)
944 /* We don't wait for traverse keys */
945 /* FIXME: We should, for trivial traversals. */
946 if (op[i][j].op == OP_TDB_TRAVERSE)
949 h = hash_key(&op[i][j].key) % (total_keys * 2);
950 while (!key_eq(hash[h].key, op[i][j].key)) {
951 if (!hash[h].key.dptr) {
952 hash[h].key = op[i][j].key;
955 h = (h + 1) % (total_keys * 2);
957 /* Might as well save some memory if we can. */
958 if (op[i][j].key.dptr != hash[h].key.dptr) {
959 talloc_free(op[i][j].key.dptr);
960 op[i][j].key.dptr = hash[h].key.dptr;
962 hash[h].user = talloc_realloc(hash, hash[h].user,
964 hash[h].num_users+1);
966 /* If it's in a transaction, it's the transaction which
967 * matters from an analysis POV. */
968 if (in_transaction(op[i], j)) {
969 unsigned start = op[i][j].group_start;
971 /* Don't include twice. */
972 if (hash[h].num_users
973 && hash[h].user[hash[h].num_users-1].file
975 && hash[h].user[hash[h].num_users-1].op_num
979 hash[h].user[hash[h].num_users].op_num = start;
981 hash[h].user[hash[h].num_users].op_num = j;
982 hash[h].user[hash[h].num_users].file = i;
990 static bool satisfies(const TDB_DATA *key, const TDB_DATA *data,
993 const TDB_DATA *need = NULL;
995 if (is_transaction(op)) {
998 /* Look through for an op in this transaction which
1000 for (i = 1; i < op->group_len; i++) {
1001 if (op[i].op != OP_TDB_TRAVERSE
1002 && key_eq(op[i].key, *key)) {
1003 need = needs(&op[i]);
1004 /* tdb_exists() is special: there might be
1005 * something in the transaction with more
1006 * specific requirements. Other ops don't have
1007 * specific requirements (eg. store or delete),
1008 * but they change the value so we can't get
1009 * more information from future ops. */
1010 if (op[i].op != OP_TDB_EXISTS)
1017 /* Don't need anything? Cool. */
1021 /* This should be tdb_null or a real value. */
1022 assert(data != &must_exist);
1023 assert(data != &must_not_exist);
1024 assert(data != ¬_exists_or_empty);
1026 /* must_not_exist == must_not_exist, must_exist == must_exist, or
1027 not_exists_or_empty == not_exists_or_empty. */
1028 if (data->dsize == need->dsize && data->dptr == need->dptr)
1031 /* Must not exist? data must not exist. */
1032 if (need == &must_not_exist)
1033 return data->dptr == NULL;
1036 if (need == &must_exist)
1037 return data->dptr != NULL;
1039 /* Either noexist or empty. */
1040 if (need == ¬_exists_or_empty)
1041 return data->dsize == 0;
1043 /* Needs something specific. */
1044 return key_eq(*data, *need);
1047 static void move_to_front(struct key_user res[], unsigned off, unsigned elem)
1050 struct key_user tmp = res[elem];
1051 memmove(res + off + 1, res + off, (elem - off)*sizeof(res[0]));
1056 static void restore_to_pos(struct key_user res[], unsigned off, unsigned elem)
1059 struct key_user tmp = res[off];
1060 memmove(res + off, res + off + 1, (elem - off)*sizeof(res[0]));
1065 static bool sort_deps(char *filename[], struct op *op[],
1066 struct key_user res[],
1067 unsigned off, unsigned num,
1068 const TDB_DATA *key, const TDB_DATA *data,
1069 unsigned num_files, unsigned fuzz)
1071 unsigned int i, files_done;
1073 bool done[num_files];
1075 /* Does this make serial numbers go backwards? Allow a little fuzz. */
1077 int serial1 = op[res[off-1].file][res[off-1].op_num].serial;
1078 int serial2 = op[res[off].file][res[off].op_num].serial;
1080 if (serial1 - serial2 > (int)fuzz) {
1082 printf("Serial jump too far (%u -> %u)\n",
1089 /* One or none left? We're sorted. */
1093 memset(done, 0, sizeof(done));
1095 /* Since ops within a trace file are ordered, we just need to figure
1096 * out which file to try next. Since we don't take into account
1097 * inter-key relationships (which exist by virtue of trace file order),
1098 * we minimize the chance of harm by trying to keep in serial order. */
1099 for (files_done = 0, i = off; i < num && files_done < num_files; i++) {
1100 if (done[res[i].file])
1103 this_op = &op[res[i].file][res[i].op_num];
1105 /* Is what we have good enough for this op? */
1106 if (satisfies(key, data, this_op)) {
1107 move_to_front(res, off, i);
1108 if (sort_deps(filename, op, res, off+1, num,
1109 key, gives(key, data, this_op),
1112 restore_to_pos(res, off, i);
1114 done[res[i].file] = true;
1118 /* No combination worked. */
1122 static void check_dep_sorting(struct key_user user[], unsigned num_users,
1127 unsigned minima[num_files];
1129 memset(minima, 0, sizeof(minima));
1130 for (i = 0; i < num_users; i++) {
1131 assert(minima[user[i].file] < user[i].op_num);
1132 minima[user[i].file] = user[i].op_num;
1137 /* All these ops happen on the same key. Which comes first?
1139 * This can happen both because read ops or failed write ops don't
1140 * change serial number, and also due to race since we access the
1141 * number unlocked (the race can cause less detectable ordering problems,
1142 * in which case we'll deadlock and report: fix manually in that case).
1144 static void figure_deps(char *filename[], struct op *op[],
1145 const TDB_DATA *key, struct key_user user[],
1146 unsigned num_users, unsigned num_files)
1148 /* We assume database starts empty. */
1149 const struct TDB_DATA *data = &tdb_null;
1152 /* We prefer to keep strict serial order if possible: it's the
1153 * most likely. We get more lax if that fails. */
1154 for (fuzz = 0; fuzz < 100; fuzz = (fuzz + 1)*2) {
1155 if (sort_deps(filename, op, user, 0, num_users, key, data,
1161 fail(filename[user[0].file], user[0].op_num+1,
1162 "Could not resolve inter-dependencies");
1164 check_dep_sorting(user, num_users, num_files);
1167 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1172 /* Gcc nexted function extension. How cool is this? */
1173 int compare_serial(const void *_a, const void *_b)
1175 const struct key_user *a = _a, *b = _b;
1177 /* First, maintain order within any trace file. */
1178 if (a->file == b->file)
1179 return a->op_num - b->op_num;
1181 /* Otherwise, arrange by serial order. */
1182 return op[a->file][a->op_num].serial
1183 - op[b->file][b->op_num].serial;
1186 /* Now sort into serial order. */
1187 for (h = 0; h < total_keys * 2; h++) {
1188 struct key_user *user = hash[h].user;
1190 qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
1191 figure_deps(filename, op, &hash[h].key, user, hash[h].num_users,
1196 static int destroy_depend(struct depend *dep)
1198 list_del(&dep->pre_list);
1199 list_del(&dep->post_list);
1203 static void add_dependency(void *ctx,
1206 unsigned int needs_file,
1207 unsigned int needs_opnum,
1208 unsigned int satisfies_file,
1209 unsigned int satisfies_opnum)
1213 /* We don't depend on ourselves. */
1214 if (needs_file == satisfies_file) {
1215 assert(satisfies_opnum < needs_opnum);
1220 printf("%s:%u: depends on %s:%u\n",
1221 filename[needs_file], needs_opnum+1,
1222 filename[satisfies_file], satisfies_opnum+1);
1225 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1226 /* If something in a traverse depends on something in another
1227 * traverse/transaction, it creates a dependency between the
1229 if ((in_traverse(op[satisfies_file], satisfies_opnum)
1230 && op[needs_file][needs_opnum].group_start)
1231 || (in_traverse(op[needs_file], needs_opnum)
1232 && op[satisfies_file][satisfies_opnum].group_start)) {
1235 /* We are satisfied by end of group. */
1236 sat = op[satisfies_file][satisfies_opnum].group_start;
1237 satisfies_opnum = sat + op[satisfies_file][sat].group_len;
1238 /* And we need that done by start of our group. */
1239 needs_opnum = op[needs_file][needs_opnum].group_start;
1242 /* There is also this case:
1243 * <traverse> <read foo> ...
1244 * <transaction> ... </transaction> <create foo>
1245 * Where if we start the traverse then wait, we could block
1246 * the transaction and deadlock.
1248 * We try to address this by ensuring that where seqnum indicates it's
1249 * possible, we wait for <create foo> before *starting* traverse.
1251 else if (in_traverse(op[needs_file], needs_opnum)) {
1252 struct op *need = &op[needs_file][needs_opnum];
1253 if (op[needs_file][need->group_start].serial <
1254 op[satisfies_file][satisfies_opnum].serial) {
1255 needs_opnum = need->group_start;
1260 /* If you depend on a transaction, you actually depend on it ending. */
1261 if (is_transaction(&op[satisfies_file][satisfies_opnum])) {
1263 += op[satisfies_file][satisfies_opnum].group_len;
1265 printf("-> Actually end of transaction %s:%u\n",
1266 filename[satisfies_file], satisfies_opnum+1);
1269 /* We should never create a dependency from middle of
1271 assert(!in_transaction(op[satisfies_file], satisfies_opnum)
1272 || op[satisfies_file][satisfies_opnum].op
1273 == OP_TDB_TRANSACTION_COMMIT
1274 || op[satisfies_file][satisfies_opnum].op
1275 == OP_TDB_TRANSACTION_CANCEL);
1277 assert(op[needs_file][needs_opnum].op != OP_TDB_TRAVERSE);
1278 assert(op[satisfies_file][satisfies_opnum].op != OP_TDB_TRAVERSE);
1280 dep = talloc(ctx, struct depend);
1281 dep->needs_file = needs_file;
1282 dep->needs_opnum = needs_opnum;
1283 dep->satisfies_file = satisfies_file;
1284 dep->satisfies_opnum = satisfies_opnum;
1285 list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
1286 list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
1287 talloc_set_destructor(dep, destroy_depend);
1290 static bool changes_db(const TDB_DATA *key, const struct op *op)
1292 return gives(key, NULL, op) != NULL;
1295 static void depend_on_previous(struct op *op[],
1298 struct key_user user[],
1308 if (prev == i - 1) {
1309 /* Just depend on previous. */
1310 add_dependency(NULL, op, filename,
1311 user[i].file, user[i].op_num,
1312 user[prev].file, user[prev].op_num);
1316 /* We have to wait for the readers. Find last one in *each* file. */
1317 memset(deps, 0, sizeof(deps));
1318 deps[user[i].file] = true;
1319 for (j = i - 1; j > prev; j--) {
1320 if (!deps[user[j].file]) {
1321 add_dependency(NULL, op, filename,
1322 user[i].file, user[i].op_num,
1323 user[j].file, user[j].op_num);
1324 deps[user[j].file] = true;
1329 /* This is simple, but not complete. We don't take into account
1330 * indirect dependencies. */
1331 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1336 /* There can only be one real dependency on each file */
1337 for (i = 0; i < num; i++) {
1338 for (j = 1; j < num_ops[i]; j++) {
1339 struct depend *dep, *next;
1340 struct depend *prev[num];
1342 memset(prev, 0, sizeof(prev));
1344 list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1345 if (!prev[dep->satisfies_file]) {
1346 prev[dep->satisfies_file] = dep;
1349 if (prev[dep->satisfies_file]->satisfies_opnum
1350 < dep->satisfies_opnum) {
1351 talloc_free(prev[dep->satisfies_file]);
1352 prev[dep->satisfies_file] = dep;
1359 for (i = 0; i < num; i++) {
1362 for (j = 0; j < num; j++)
1365 for (j = 1; j < num_ops[i]; j++) {
1366 struct depend *dep, *next;
1368 list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1369 if (deps[dep->satisfies_file]
1370 >= (int)dep->satisfies_opnum)
1373 deps[dep->satisfies_file]
1374 = dep->satisfies_opnum;
1380 static void derive_dependencies(char *filename[],
1381 struct op *op[], unsigned int num_ops[],
1384 struct keyinfo *hash;
1387 /* Create hash table for faster key lookup. */
1388 hash = hash_ops(op, num_ops, num);
1390 /* Sort them by serial number. */
1391 sort_ops(hash, filename, op, num);
1393 /* Create dependencies back to the last change, rather than
1394 * creating false dependencies by naively making each one
1395 * depend on the previous. This has two purposes: it makes
1396 * later optimization simpler, and it also avoids deadlock with
1397 * same sequence number ops inside traversals (if one
1398 * traversal doesn't write anything, two ops can have the same
1399 * sequence number yet we can create a traversal dependency
1400 * the other way). */
1401 for (h = 0; h < total_keys * 2; h++) {
1404 if (hash[h].num_users < 2)
1407 for (i = 0; i < hash[h].num_users; i++) {
1408 if (changes_db(&hash[h].key, &op[hash[h].user[i].file]
1409 [hash[h].user[i].op_num])) {
1410 depend_on_previous(op, filename, num,
1411 hash[h].user, i, prev);
1413 } else if (prev >= 0)
1414 add_dependency(hash, op, filename,
1415 hash[h].user[i].file,
1416 hash[h].user[i].op_num,
1417 hash[h].user[prev].file,
1418 hash[h].user[prev].op_num);
1422 optimize_dependencies(op, num_ops, num);
1425 int main(int argc, char *argv[])
1427 struct timeval start, end;
1428 unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1429 struct op *op[argc];
1435 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1437 pipes = talloc_array(NULL, struct pipe, argc - 2);
1438 for (i = 0; i < argc - 2; i++) {
1439 printf("Loading tracefile %s...", argv[2+i]);
1441 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1442 &tdb_flags[i], &open_flags[i]);
1443 if (pipe(pipes[i].fd) != 0)
1444 err(1, "creating pipe");
1448 printf("Calculating inter-dependencies...");
1450 derive_dependencies(argv+2, op, num_ops, i);
1453 /* Don't fork for single arg case: simple debugging. */
1455 struct tdb_context *tdb;
1456 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
1457 open_flags[0], 0600,
1459 printf("Single threaded run...");
1462 run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0]);
1463 check_deps(argv[2], op[0], num_ops[0]);
1470 err(1, "creating pipe");
1472 for (i = 0; i < argc - 2; i++) {
1473 struct tdb_context *tdb;
1477 err(1, "fork failed");
1480 tdb = tdb_open_ex(argv[1], hashsize[i], tdb_flags[i],
1481 open_flags[i], 0600,
1484 err(1, "Opening tdb %s", argv[1]);
1486 /* This catches parent exiting. */
1487 if (read(fds[0], &c, 1) != 1)
1489 run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
1491 check_deps(argv[2+i], op[i], num_ops[i]);
1498 /* Let everything settle. */
1501 printf("Starting run...");
1503 gettimeofday(&start, NULL);
1504 /* Tell them all to go! Any write of sufficient length will do. */
1505 if (write(fds[1], hashsize, i) != i)
1506 err(1, "Writing to wakeup pipe");
1508 for (i = 0; i < argc - 2; i++) {
1511 if (!WIFEXITED(status)) {
1512 warnx("Child died with signal %i", WTERMSIG(status));
1514 } else if (WEXITSTATUS(status) != 0)
1515 /* Assume child spat out error. */
1521 gettimeofday(&end, NULL);
1524 end.tv_sec -= start.tv_sec;
1525 printf("Time replaying: %lu usec\n",
1526 end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));