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 void op_analyze_transaction(const char *filename,
314 struct op op[], unsigned int op_num,
319 op[op_num].key = tdb_null;
322 fail(filename, op_num+1, "Expect no arguments");
324 for (i = op_num-1; i >= 0; i--) {
325 if (op[i].op == OP_TDB_TRANSACTION_START && !op[i].group_len)
330 fail(filename, op_num+1, "no transaction start found");
333 op[start].group_len = op_num - i;
335 /* This rolls in nested transactions. I think that's right. */
336 for (i++; i <= op_num; i++)
337 op[i].group_start = start;
340 struct traverse_hash {
345 static void op_analyze_traverse(const char *filename,
346 struct op op[], unsigned int op_num,
351 op[op_num].key = tdb_null;
353 /* = %u means traverse function terminated. */
355 if (!streq(words[2], "=") || !words[3] || words[4])
356 fail(filename, op_num+1, "expect = <num>");
357 op[op_num].ret = atoi(words[3]);
361 for (i = op_num-1; i >= 0; i--) {
362 if (op[i].op != OP_TDB_TRAVERSE_READ_START
363 && op[i].op != OP_TDB_TRAVERSE_START)
371 fail(filename, op_num+1, "no traversal start found");
374 op[start].group_len = op_num - start;
376 for (i = start; i <= op_num; i++)
377 op[i].group_start = start;
380 /* Keep -Wmissing-declarations happy: */
381 const struct op_table *
382 find_keyword (register const char *str, register unsigned int len);
384 #include "keywords.c"
387 /* We can have more than one */
388 struct list_node pre_list;
389 struct list_node post_list;
390 unsigned int needs_file;
391 unsigned int needs_opnum;
392 unsigned int satisfies_file;
393 unsigned int satisfies_opnum;
396 static void check_deps(const char *filename, struct op op[], unsigned int num)
401 for (i = 1; i < num; i++)
402 if (!list_empty(&op[i].pre))
403 fail(filename, i+1, "Still has dependencies");
407 static void dump_pre(char *filename[], struct op *op[],
408 unsigned int file, unsigned int i)
412 printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
414 list_for_each(&op[file][i].pre, dep, pre_list)
415 printf(" %s:%u (%u)\n",
416 filename[dep->satisfies_file], dep->satisfies_opnum+1,
417 op[dep->satisfies_file][dep->satisfies_opnum].serial);
418 check_deps(filename[file], op[file], i);
421 /* We simply read/write pointers, since we all are children. */
422 static void do_pre(struct tdb_context *tdb,
423 char *filename[], struct op *op[],
424 unsigned int file, int pre_fd, unsigned int i)
426 while (!list_empty(&op[file][i].pre)) {
430 printf("%s:%u:waiting for pre\n", filename[file], i+1);
434 while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
435 if (errno == EINTR) {
436 dump_pre(filename, op, file, i);
439 errx(1, "Reading from pipe");
444 printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
445 dep->needs_opnum+1, filename[dep->satisfies_file],
446 dep->satisfies_opnum+1);
449 /* This could be any op, not just this one. */
454 static void do_post(char *filename[], struct op *op[],
455 unsigned int file, unsigned int i)
459 list_for_each(&op[file][i].post, dep, post_list) {
461 printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
462 filename[dep->needs_file], dep->needs_opnum+1);
464 if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
466 err(1, "%s:%u failed to tell file %s",
467 filename[file], i+1, filename[dep->needs_file]);
471 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
476 static unsigned run_ops(struct tdb_context *tdb,
481 unsigned int start, unsigned int stop);
483 struct traverse_info {
492 /* More complex. Just do whatever's they did at the n'th entry. */
493 static int nontrivial_traverse(struct tdb_context *tdb,
494 TDB_DATA key, TDB_DATA data,
497 struct traverse_info *tinfo = _tinfo;
498 unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
500 if (tinfo->i == tinfo->start + trav_len) {
501 /* This can happen if traverse expects to be empty. */
504 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
505 "traverse did not terminate");
508 if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
509 fail(tinfo->filename[tinfo->file], tinfo->start + 1,
510 "%s:%u:traverse terminated early");
512 /* Run any normal ops. */
513 tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
514 tinfo->file, tinfo->i+1, tinfo->start + trav_len);
516 if (tinfo->i == tinfo->start + trav_len)
522 static unsigned op_traverse(struct tdb_context *tdb,
526 int (*traversefn)(struct tdb_context *,
527 tdb_traverse_func, void *),
531 struct traverse_info tinfo = { op, filename, file, pre_fd,
534 traversefn(tdb, nontrivial_traverse, &tinfo);
536 /* Traversing in wrong order can have strange effects: eg. if
537 * original traverse went A (delete A), B, we might do B
538 * (delete A). So if we have ops left over, we do it now. */
539 while (tinfo.i != start + op[file][start].group_len) {
540 if (op[file][tinfo.i].op == OP_TDB_TRAVERSE)
543 tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
545 start + op[file][start].group_len);
551 static void break_out(int sig)
555 static __attribute__((noinline))
556 unsigned run_ops(struct tdb_context *tdb,
561 unsigned int start, unsigned int stop)
566 sa.sa_handler = break_out;
569 sigaction(SIGALRM, &sa, NULL);
570 for (i = start; i < stop; i++) {
571 do_pre(tdb, filename, op, file, pre_fd, i);
573 switch (op[file][i].op) {
575 try(tdb_lockall(tdb), op[file][i].ret);
577 case OP_TDB_LOCKALL_MARK:
578 try(tdb_lockall_mark(tdb), op[file][i].ret);
580 case OP_TDB_LOCKALL_UNMARK:
581 try(tdb_lockall_unmark(tdb), op[file][i].ret);
583 case OP_TDB_LOCKALL_NONBLOCK:
584 unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
585 tdb_lockall(tdb), tdb_unlockall(tdb));
587 case OP_TDB_UNLOCKALL:
588 try(tdb_unlockall(tdb), op[file][i].ret);
590 case OP_TDB_LOCKALL_READ:
591 try(tdb_lockall_read(tdb), op[file][i].ret);
593 case OP_TDB_LOCKALL_READ_NONBLOCK:
594 unreliable(tdb_lockall_read_nonblock(tdb),
596 tdb_lockall_read(tdb),
597 tdb_unlockall_read(tdb));
599 case OP_TDB_UNLOCKALL_READ:
600 try(tdb_unlockall_read(tdb), op[file][i].ret);
602 case OP_TDB_CHAINLOCK:
603 try(tdb_chainlock(tdb, op[file][i].key),
606 case OP_TDB_CHAINLOCK_NONBLOCK:
607 unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
609 tdb_chainlock(tdb, op[file][i].key),
610 tdb_chainunlock(tdb, op[file][i].key));
612 case OP_TDB_CHAINLOCK_MARK:
613 try(tdb_chainlock_mark(tdb, op[file][i].key),
616 case OP_TDB_CHAINLOCK_UNMARK:
617 try(tdb_chainlock_unmark(tdb, op[file][i].key),
620 case OP_TDB_CHAINUNLOCK:
621 try(tdb_chainunlock(tdb, op[file][i].key),
624 case OP_TDB_CHAINLOCK_READ:
625 try(tdb_chainlock_read(tdb, op[file][i].key),
628 case OP_TDB_CHAINUNLOCK_READ:
629 try(tdb_chainunlock_read(tdb, op[file][i].key),
632 case OP_TDB_PARSE_RECORD:
633 try(tdb_parse_record(tdb, op[file][i].key, get_len,
638 try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
641 try(tdb_store(tdb, op[file][i].key, op[file][i].data,
646 try(tdb_append(tdb, op[file][i].key, op[file][i].data),
649 case OP_TDB_GET_SEQNUM:
650 try(tdb_get_seqnum(tdb), op[file][i].ret);
652 case OP_TDB_WIPE_ALL:
653 try(tdb_wipe_all(tdb), op[file][i].ret);
655 case OP_TDB_TRANSACTION_START:
656 try(tdb_transaction_start(tdb), op[file][i].ret);
658 case OP_TDB_TRANSACTION_CANCEL:
659 try(tdb_transaction_cancel(tdb), op[file][i].ret);
661 case OP_TDB_TRANSACTION_COMMIT:
662 try(tdb_transaction_commit(tdb), op[file][i].ret);
664 case OP_TDB_TRAVERSE_READ_START:
665 i = op_traverse(tdb, pre_fd, filename, file,
666 tdb_traverse_read, op, i);
668 case OP_TDB_TRAVERSE_START:
669 i = op_traverse(tdb, pre_fd, filename, file,
670 tdb_traverse, op, i);
672 case OP_TDB_TRAVERSE:
673 /* Terminate: we're in a traverse, and we've
676 case OP_TDB_TRAVERSE_END:
677 fail(filename[file], i+1, "unexpected end traverse");
678 /* FIXME: These must be treated like traverse. */
679 case OP_TDB_FIRSTKEY:
680 if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
681 fail(filename[file], i+1, "bad firstkey");
684 if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
686 fail(filename[file], i+1, "bad nextkey");
689 TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
690 if (!key_eq(f, op[file][i].data))
691 fail(filename[file], i+1, "bad fetch %u",
696 try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
699 do_post(filename, op, file, i);
704 static struct op *load_tracefile(const char *filename, unsigned int *num,
705 unsigned int *hashsize,
706 unsigned int *tdb_flags,
707 unsigned int *open_flags)
710 struct op *op = talloc_array(NULL, struct op, 1);
715 file = grab_file(NULL, filename, NULL);
717 err(1, "Reading %s", filename);
719 lines = strsplit(file, file, "\n", NULL);
721 errx(1, "%s is empty", filename);
723 words = strsplit(lines, lines[0], " ", NULL);
724 if (!streq(words[1], "tdb_open"))
725 fail(filename, 1, "does not start with tdb_open");
727 *hashsize = atoi(words[2]);
728 *tdb_flags = strtoul(words[3], NULL, 0);
729 *open_flags = strtoul(words[4], NULL, 0);
731 for (i = 1; lines[i]; i++) {
732 const struct op_table *opt;
734 words = strsplit(lines, lines[i], " ", NULL);
735 if (!words[0] || !words[1])
736 fail(filename, i+1, "Expected serial number and op");
738 opt = find_keyword(words[1], strlen(words[1]));
740 if (streq(words[1], "tdb_close")) {
743 "lines after tdb_close");
748 fail(filename, i+1, "Unknown operation '%s'", words[1]);
751 add_op(filename, &op, i, atoi(words[0]), opt->type);
752 opt->enhance_op(filename, op, i, words);
755 fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
762 /* We remember all the keys we've ever seen, and who has them. */
770 unsigned int num_users;
771 struct key_user *user;
774 static const TDB_DATA must_not_exist;
775 static const TDB_DATA must_exist;
776 static const TDB_DATA not_exists_or_empty;
778 /* NULL means doesn't care if it exists or not, &must_exist means
779 * it must exist but we don't care what, &must_not_exist means it must
780 * not exist, otherwise the data it needs. */
781 static const TDB_DATA *needs(const struct op *op)
784 /* FIXME: Pull forward deps, since we can deadlock */
785 case OP_TDB_CHAINLOCK:
786 case OP_TDB_CHAINLOCK_NONBLOCK:
787 case OP_TDB_CHAINLOCK_MARK:
788 case OP_TDB_CHAINLOCK_UNMARK:
789 case OP_TDB_CHAINUNLOCK:
790 case OP_TDB_CHAINLOCK_READ:
791 case OP_TDB_CHAINUNLOCK_READ:
795 if (op->append.pre.dsize == 0)
796 return ¬_exists_or_empty;
797 return &op->append.pre;
800 if (op->flag == TDB_INSERT) {
804 return &must_not_exist;
805 } else if (op->flag == TDB_MODIFY) {
807 return &must_not_exist;
811 /* No flags? Don't care */
818 return &must_not_exist;
820 case OP_TDB_PARSE_RECORD:
822 return &must_not_exist;
825 /* FIXME: handle these. */
826 case OP_TDB_WIPE_ALL:
827 case OP_TDB_FIRSTKEY:
829 case OP_TDB_GET_SEQNUM:
830 case OP_TDB_TRAVERSE:
831 case OP_TDB_TRANSACTION_COMMIT:
832 case OP_TDB_TRANSACTION_CANCEL:
833 case OP_TDB_TRANSACTION_START:
838 return &must_not_exist;
843 return &must_not_exist;
847 errx(1, "Unexpected op %i", op->op);
852 /* What's the data after this op? pre if nothing changed. */
853 static const TDB_DATA *gives(const struct op *op, const TDB_DATA *pre)
855 /* Failed ops don't change state of db. */
859 if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
862 if (op->op == OP_TDB_APPEND)
863 return &op->append.post;
865 if (op->op == OP_TDB_STORE)
871 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
874 unsigned int i, j, h;
875 struct keyinfo *hash;
877 hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
878 for (i = 0; i < num; i++) {
879 for (j = 1; j < num_ops[i]; j++) {
880 /* We can't do this on allocation, due to realloc. */
881 list_head_init(&op[i][j].post);
882 list_head_init(&op[i][j].pre);
884 if (!op[i][j].key.dptr)
887 /* We don't wait for traverse keys */
888 /* FIXME: We should, for trivial traversals. */
889 if (op[i][j].op == OP_TDB_TRAVERSE)
892 h = hash_key(&op[i][j].key) % (total_keys * 2);
893 while (!key_eq(hash[h].key, op[i][j].key)) {
894 if (!hash[h].key.dptr) {
895 hash[h].key = op[i][j].key;
898 h = (h + 1) % (total_keys * 2);
900 /* Might as well save some memory if we can. */
901 if (op[i][j].key.dptr != hash[h].key.dptr) {
902 talloc_free(op[i][j].key.dptr);
903 op[i][j].key.dptr = hash[h].key.dptr;
905 hash[h].user = talloc_realloc(hash, hash[h].user,
907 hash[h].num_users+1);
908 hash[h].user[hash[h].num_users].op_num = j;
909 hash[h].user[hash[h].num_users].file = i;
917 static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
919 /* Don't need anything? Cool. */
923 /* This should be tdb_null or a real value. */
924 assert(data != &must_exist);
925 assert(data != &must_not_exist);
926 assert(data != ¬_exists_or_empty);
928 /* must_not_exist == must_not_exist, must_exist == must_exist, or
929 not_exists_or_empty == not_exists_or_empty. */
930 if (data->dsize == need->dsize && data->dptr == need->dptr)
933 /* Must not exist? data must not exist. */
934 if (need == &must_not_exist)
935 return data->dptr == NULL;
938 if (need == &must_exist)
939 return data->dptr != NULL;
941 /* Either noexist or empty. */
942 if (need == ¬_exists_or_empty)
943 return data->dsize == 0;
945 /* Needs something specific. */
946 return key_eq(*data, *need);
949 static void move_to_front(struct key_user res[], unsigned int elem)
952 struct key_user tmp = res[elem];
953 memmove(res + 1, res, elem*sizeof(res[0]));
958 static void restore_to_pos(struct key_user res[], unsigned int elem)
961 struct key_user tmp = res[0];
962 memmove(res, res + 1, elem*sizeof(res[0]));
967 static bool sort_deps(char *filename[], struct op *op[],
968 struct key_user res[], unsigned num,
969 const TDB_DATA *data, unsigned num_files)
971 unsigned int i, files_done;
973 bool done[num_files];
975 /* Nothing left? We're sorted. */
979 memset(done, 0, sizeof(done));
981 /* Since ops within a trace file are ordered, we just need to figure
982 * out which file to try next. Since we don't take into account
983 * inter-key relationships (which exist by virtue of trace file order),
984 * we minimize the chance of harm by trying to keep in serial order. */
985 for (files_done = 0, i = 0; i < num && files_done < num_files; i++) {
986 if (done[res[i].file])
989 this_op = &op[res[i].file][res[i].op_num];
990 /* Is what we have good enough for this op? */
991 if (satisfies(data, needs(this_op))) {
992 move_to_front(res, i);
993 if (sort_deps(filename, op, res+1, num-1,
994 gives(this_op, data), num_files))
996 restore_to_pos(res, i);
998 done[res[i].file] = true;
1002 /* No combination worked. */
1006 static void check_dep_sorting(struct key_user user[], unsigned num_users,
1011 unsigned minima[num_files];
1013 memset(minima, 0, sizeof(minima));
1014 for (i = 0; i < num_users; i++) {
1015 assert(minima[user[i].file] < user[i].op_num);
1016 minima[user[i].file] = user[i].op_num;
1021 /* All these ops have the same serial number. Which comes first?
1023 * This can happen both because read ops or failed write ops don't
1024 * change serial number, and also due to race since we access the
1025 * number unlocked (the race can cause less detectable ordering problems,
1026 * in which case we'll deadlock and report: fix manually in that case).
1028 static void figure_deps(char *filename[], struct op *op[],
1029 struct key_user user[], unsigned num_users,
1032 /* We assume database starts empty. */
1033 const struct TDB_DATA *data = &tdb_null;
1035 if (!sort_deps(filename, op, user, num_users, data, num_files))
1036 fail(filename[user[0].file], user[0].op_num+1,
1037 "Could not resolve inter-dependencies");
1039 check_dep_sorting(user, num_users, num_files);
1042 static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
1047 /* Gcc nexted function extension. How cool is this? */
1048 int compare_serial(const void *_a, const void *_b)
1050 const struct key_user *a = _a, *b = _b;
1052 /* First, maintain order within any trace file. */
1053 if (a->file == b->file)
1054 return a->op_num - b->op_num;
1056 /* Otherwise, arrange by serial order. */
1057 return op[a->file][a->op_num].serial
1058 - op[b->file][b->op_num].serial;
1061 /* Now sort into serial order. */
1062 for (h = 0; h < total_keys * 2; h++) {
1063 struct key_user *user = hash[h].user;
1065 qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
1066 figure_deps(filename, op, user, hash[h].num_users, num);
1070 static int destroy_depend(struct depend *dep)
1072 list_del(&dep->pre_list);
1073 list_del(&dep->post_list);
1077 static void add_dependency(void *ctx,
1080 unsigned int needs_file,
1081 unsigned int needs_opnum,
1082 unsigned int satisfies_file,
1083 unsigned int satisfies_opnum)
1086 unsigned int needs_start, sat_start;
1088 /* We don't depend on ourselves. */
1089 if (needs_file == satisfies_file) {
1090 assert(satisfies_opnum < needs_opnum);
1095 printf("%s:%u: depends on %s:%u\n",
1096 filename[needs_file], needs_opnum+1,
1097 filename[satisfies_file], satisfies_opnum+1);
1100 needs_start = op[needs_file][needs_opnum].group_start;
1101 sat_start = op[satisfies_file][satisfies_opnum].group_start;
1103 /* If needs is in a transaction, we need it before start. */
1105 switch (op[needs_file][needs_start].op) {
1106 case OP_TDB_TRANSACTION_START:
1107 needs_opnum = needs_start;
1109 printf(" -> Back to %u\n", needs_start+1);
1118 /* If satisfies is in a transaction, we wait until after commit. */
1119 /* FIXME: If transaction is cancelled, don't need dependency. */
1121 if (op[satisfies_file][sat_start].op
1122 == OP_TDB_TRANSACTION_START) {
1123 satisfies_opnum = sat_start
1124 + op[satisfies_file][sat_start].group_len;
1126 printf(" -> Depends on %u\n", satisfies_opnum+1);
1132 assert(op[needs_file][needs_opnum].op != OP_TDB_TRAVERSE);
1133 assert(op[satisfies_file][satisfies_opnum].op != OP_TDB_TRAVERSE);
1135 dep = talloc(ctx, struct depend);
1136 dep->needs_file = needs_file;
1137 dep->needs_opnum = needs_opnum;
1138 dep->satisfies_file = satisfies_file;
1139 dep->satisfies_opnum = satisfies_opnum;
1140 list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
1141 list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
1142 talloc_set_destructor(dep, destroy_depend);
1145 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1146 struct traverse_dep {
1148 unsigned int op_num;
1151 /* Traversals can deadlock against each other, and transactions. Force
1153 static void make_traverse_depends(char *filename[],
1154 struct op *op[], unsigned int num_ops[],
1157 unsigned int i, j, num_traversals = 0;
1158 struct traverse_dep *dep;
1160 /* Sort by which one runs first. */
1161 int compare_traverse_dep(const void *_a, const void *_b)
1163 const struct traverse_dep *ta = _a, *tb = _b;
1164 const struct op *a = &op[ta->file][ta->op_num],
1165 *b = &op[tb->file][tb->op_num];
1167 if (a->serial != b->serial)
1168 return a->serial - b->serial;
1170 /* If they have same serial, it means one didn't make any
1171 * changes. Thus sort by end in that case. */
1172 return a[a->group_len].serial - b[b->group_len].serial;
1175 dep = talloc_array(NULL, struct traverse_dep, 1);
1178 for (i = 0; i < num; i++) {
1179 for (j = 1; j < num_ops[i]; j++) {
1180 /* Transaction or traverse start. */
1181 if (op[i][j].group_start == j) {
1182 dep = talloc_realloc(NULL, dep,
1183 struct traverse_dep,
1185 dep[num_traversals].file = i;
1186 dep[num_traversals].op_num = j;
1191 qsort(dep, num_traversals, sizeof(dep[0]), compare_traverse_dep);
1192 for (i = 1; i < num_traversals; i++) {
1193 /* i depends on end of traverse i-1. */
1194 add_dependency(NULL, op, filename, dep[i].file, dep[i].op_num,
1195 dep[i-1].file, dep[i-1].op_num
1196 + op[dep[i-1].file][dep[i-1].op_num].group_len);
1200 #endif /* TRAVERSALS_TAKE_TRANSACTION_LOCK */
1202 static bool changes_db(const struct op *op)
1204 return gives(op, NULL) != NULL;
1207 static void depend_on_previous(struct op *op[],
1210 struct key_user user[],
1220 if (prev == i - 1) {
1221 /* Just depend on previous. */
1222 add_dependency(NULL, op, filename,
1223 user[i].file, user[i].op_num,
1224 user[prev].file, user[prev].op_num);
1228 /* We have to wait for the readers. Find last one in *each* file. */
1229 memset(deps, 0, sizeof(deps));
1230 deps[user[i].file] = true;
1231 for (j = i - 1; j > prev; j--) {
1232 if (!deps[user[j].file]) {
1233 add_dependency(NULL, op, filename,
1234 user[i].file, user[i].op_num,
1235 user[j].file, user[j].op_num);
1236 deps[user[j].file] = true;
1241 /* This is simple, but not complete. We don't take into account
1242 * indirect dependencies. */
1243 static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
1248 /* There can only be one real dependency on each file */
1249 for (i = 0; i < num; i++) {
1250 for (j = 1; j < num_ops[i]; j++) {
1251 struct depend *dep, *next;
1252 struct depend *prev[num];
1254 memset(prev, 0, sizeof(prev));
1256 list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1257 if (!prev[dep->satisfies_file]) {
1258 prev[dep->satisfies_file] = dep;
1261 if (prev[dep->satisfies_file]->satisfies_opnum
1262 < dep->satisfies_opnum) {
1263 talloc_free(prev[dep->satisfies_file]);
1264 prev[dep->satisfies_file] = dep;
1271 for (i = 0; i < num; i++) {
1274 for (j = 0; j < num; j++)
1277 for (j = 1; j < num_ops[i]; j++) {
1278 struct depend *dep, *next;
1280 list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
1281 if (deps[dep->satisfies_file]
1282 >= (int)dep->satisfies_opnum)
1285 deps[dep->satisfies_file]
1286 = dep->satisfies_opnum;
1292 static void derive_dependencies(char *filename[],
1293 struct op *op[], unsigned int num_ops[],
1296 struct keyinfo *hash;
1299 /* Create hash table for faster key lookup. */
1300 hash = hash_ops(op, num_ops, num);
1302 /* Sort them by serial number. */
1303 sort_ops(hash, filename, op, num);
1305 /* Create dependencies back to the last change, rather than
1306 * creating false dependencies by naively making each one
1307 * depend on the previous. This has two purposes: it makes
1308 * later optimization simpler, and it also avoids deadlock with
1309 * same sequence number ops inside traversals (if one
1310 * traversal doesn't write anything, two ops can have the same
1311 * sequence number yet we can create a traversal dependency
1312 * the other way). */
1313 for (h = 0; h < total_keys * 2; h++) {
1316 if (hash[h].num_users < 2)
1319 for (i = 0; i < hash[h].num_users; i++) {
1320 if (changes_db(&op[hash[h].user[i].file]
1321 [hash[h].user[i].op_num])) {
1322 depend_on_previous(op, filename, num,
1323 hash[h].user, i, prev);
1325 } else if (prev >= 0)
1326 add_dependency(hash, op, filename,
1327 hash[h].user[i].file,
1328 hash[h].user[i].op_num,
1329 hash[h].user[prev].file,
1330 hash[h].user[prev].op_num);
1334 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
1335 make_traverse_depends(filename, op, num_ops, num);
1338 optimize_dependencies(op, num_ops, num);
1341 int main(int argc, char *argv[])
1343 struct timeval start, end;
1344 unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
1345 struct op *op[argc];
1351 errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
1353 pipes = talloc_array(NULL, struct pipe, argc - 2);
1354 for (i = 0; i < argc - 2; i++) {
1355 printf("Loading tracefile %s...", argv[2+i]);
1357 op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
1358 &tdb_flags[i], &open_flags[i]);
1359 if (pipe(pipes[i].fd) != 0)
1360 err(1, "creating pipe");
1364 printf("Calculating inter-dependencies...");
1366 derive_dependencies(argv+2, op, num_ops, i);
1369 /* Don't fork for single arg case: simple debugging. */
1371 struct tdb_context *tdb;
1372 tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
1373 open_flags[0], 0600,
1375 printf("Single threaded run...");
1378 run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0]);
1379 check_deps(argv[2], op[0], num_ops[0]);
1386 err(1, "creating pipe");
1388 for (i = 0; i < argc - 2; i++) {
1389 struct tdb_context *tdb;
1393 err(1, "fork failed");
1396 tdb = tdb_open_ex(argv[1], hashsize[i], tdb_flags[i],
1397 open_flags[i], 0600,
1400 err(1, "Opening tdb %s", argv[1]);
1402 /* This catches parent exiting. */
1403 if (read(fds[0], &c, 1) != 1)
1405 run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
1407 check_deps(argv[2+i], op[i], num_ops[i]);
1414 /* Let everything settle. */
1417 printf("Starting run...");
1419 gettimeofday(&start, NULL);
1420 /* Tell them all to go! Any write of sufficient length will do. */
1421 if (write(fds[1], hashsize, i) != i)
1422 err(1, "Writing to wakeup pipe");
1424 for (i = 0; i < argc - 2; i++) {
1427 if (!WIFEXITED(status)) {
1428 warnx("Child died with signal %i", WTERMSIG(status));
1430 } else if (WEXITSTATUS(status) != 0)
1431 /* Assume child spat out error. */
1437 gettimeofday(&end, NULL);
1440 end.tv_sec -= start.tv_sec;
1441 printf("Time replaying: %lu usec\n",
1442 end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));