X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Ftools%2Freplay_trace.c;h=bd4b8aa804eeaca2c11626cd659d4c7345ec73f4;hp=61cdfb35d03b83ecd59be827d22d1c651c565911;hb=68e848993dca98439f73dfaf6641f3f85ef89597;hpb=dc251c895ca56528e6f9e5371a5bdfcaa5a92b47

diff --git a/ccan/tdb/tools/replay_trace.c b/ccan/tdb/tools/replay_trace.c
index 61cdfb35..bd4b8aa8 100644
--- a/ccan/tdb/tools/replay_trace.c
+++ b/ccan/tdb/tools/replay_trace.c
@@ -12,6 +12,9 @@
 #include <sys/types.h>
 #include <sys/wait.h>
 #include <sys/time.h>
+#include <errno.h>
+#include <signal.h>
+#include <assert.h>
 
 #define STRINGIFY2(x) #x
 #define STRINGIFY(x) STRINGIFY2(x)
@@ -19,7 +22,7 @@
 /* Avoid mod by zero */
 static unsigned int total_keys = 1;
 
-#define DEBUG_DEPS 1
+/* #define DEBUG_DEPS 1 */
 
 /* Traversals block transactions in the current implementation. */
 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
@@ -48,7 +51,8 @@ static void __attribute__((noreturn)) fail(const char *filename,
 	do {								\
 		int ret = (expr);					\
 		if (ret != (expect))					\
-			fail(filename, i+1, STRINGIFY(expr) "= %i", ret); \
+			fail(filename[file], i+1,			\
+			     STRINGIFY(expr) "= %i", ret);		\
 	} while (0)
 
 /* Try or imitate results. */
@@ -57,7 +61,8 @@ static void __attribute__((noreturn)) fail(const char *filename,
 		int ret = expr;						\
 		if (ret != expect) {					\
 			fprintf(stderr, "%s:%u: %s gave %i not %i",	\
-			      filename, i+1, STRINGIFY(expr), ret, expect); \
+				filename[file], i+1, STRINGIFY(expr),	\
+				ret, expect);				\
 			if (expect == 0)				\
 				force;					\
 			else						\
@@ -129,8 +134,8 @@ struct op {
 
 	/* Who is waiting for us? */
 	struct list_head post;
-	/* How many are we waiting for? */
-	unsigned int pre;
+	/* What are we waiting for? */
+	struct list_head pre;
 
 	/* If I'm part of a group (traverse/transaction) where is
 	 * start?  (Otherwise, 0) */
@@ -138,9 +143,11 @@ struct op {
 
 	union {
 		int flag; /* open and store */
-		struct traverse *trav; /* traverse start */
-		TDB_DATA post_append; /* append */
-		unsigned int transaction_end; /* transaction start */
+		struct {  /* append */
+			TDB_DATA pre;
+			TDB_DATA post;
+		} append;
+		unsigned int group_len; /* transaction/traverse start */
 	};
 };
 
@@ -187,7 +194,6 @@ static void add_op(const char *filename, struct op **op, unsigned int i,
 	new = (*op) + i;
 	new->op = type;
 	new->serial = serial;
-	new->pre = 0;
 	new->ret = 0;
 	new->group_start = 0;
 }
@@ -262,8 +268,14 @@ static void op_add_append(const char *filename,
 
 	op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
 	op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
-	op[op_num].post_append
+
+	op[op_num].append.post
 		= make_tdb_data(op, filename, op_num+1, words[5]);
+
+	/* By subtraction, figure out what previous data was. */
+	op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
+	op[op_num].append.pre.dsize
+		= op[op_num].append.post.dsize - op[op_num].data.dsize;
 	total_keys++;
 }
 
@@ -285,7 +297,7 @@ static void op_add_traverse(const char *filename,
 		fail(filename, op_num+1, "Expect no arguments");
 
 	op[op_num].key = tdb_null;
-	op[op_num].trav = NULL;
+	op[op_num].group_len = 0;
 }
 
 static void op_add_transaction(const char *filename, struct op op[],
@@ -295,34 +307,39 @@ static void op_add_transaction(const char *filename, struct op op[],
 		fail(filename, op_num+1, "Expect no arguments");
 
 	op[op_num].key = tdb_null;
-	op[op_num].transaction_end = 0;
+	op[op_num].group_len = 0;
+}
+
+static int op_transaction_start(struct op op[], unsigned int op_num)
+{
+	unsigned int i;
+
+	for (i = op_num-1; i > 0; i--) {
+		if (op[i].op == OP_TDB_TRANSACTION_START && !op[i].group_len)
+			return i;
+	}
+	return 0;
 }
 
 static void op_analyze_transaction(const char *filename,
 				   struct op op[], unsigned int op_num,
 				   char *words[])
 {
-	int i, start;
+	unsigned int start, i;
 
 	op[op_num].key = tdb_null;
 
 	if (words[2])
 		fail(filename, op_num+1, "Expect no arguments");
 
-	for (i = op_num-1; i >= 0; i--) {
-		if (op[i].op == OP_TDB_TRANSACTION_START &&
-		    !op[i].transaction_end)
-			break;
-	}
-
-	if (i < 0)
+	start = op_transaction_start(op, op_num);
+	if (!start)
 		fail(filename, op_num+1, "no transaction start found");
 
-	start = i;
-	op[start].transaction_end = op_num;
+	op[start].group_len = op_num - start;
 
 	/* This rolls in nested transactions.  I think that's right. */
-	for (i++; i <= op_num; i++)
+	for (i = start; i <= op_num; i++)
 		op[i].group_start = start;
 }
 
@@ -331,50 +348,11 @@ struct traverse_hash {
 	unsigned int index;
 };
 
-/* A traverse is a hash of keys, each one associated with ops. */
-struct traverse {
-	/* How many traversal callouts should I do? */
-	unsigned int num;
-
-	/* Where is traversal end op? */
-	unsigned int end;
-
-	/* For trivial traversals. */
-	struct traverse_hash *hash;
-};
-
-/* A trivial traversal is one which doesn't terminate early and only
- * plays with its own record.  We can reliably replay these even if
- * traverse order changes. */
-static bool is_trivial_traverse(struct op op[], unsigned int end)
-{
-#if 0
-	unsigned int i;
-	TDB_DATA cur = tdb_null;
-
-	if (op[end].ret != 0)
-		return false;
-
-	for (i = 0; i < end; i++) {
-		if (!op[i].key.dptr)
-			continue;
-		if (op[i].op == OP_TDB_TRAVERSE)
-			cur = op[i].key;
-		if (!key_eq(cur, op[i].key))
-			return false;
-	}
-	return true;
-#endif
-	/* With multiple things happening at once, no traverse is trivial. */
-	return false;
-}
-
 static void op_analyze_traverse(const char *filename,
 				struct op op[], unsigned int op_num,
 				char *words[])
 {
 	int i, start;
-	struct traverse *trav = talloc(op, struct traverse);
 
 	op[op_num].key = tdb_null;
 
@@ -386,15 +364,11 @@ static void op_analyze_traverse(const char *filename,
 	} else
 		op[op_num].ret = 0;
 
-	trav->num = 0;
-	trav->end = op_num;
 	for (i = op_num-1; i >= 0; i--) {
-		if (op[i].op == OP_TDB_TRAVERSE)
-			trav->num++;
 		if (op[i].op != OP_TDB_TRAVERSE_READ_START
 		    && op[i].op != OP_TDB_TRAVERSE_START)
 			continue;
-		if (op[i].trav)
+		if (op[i].group_len)
 			continue;
 		break;
 	}
@@ -403,28 +377,10 @@ static void op_analyze_traverse(const char *filename,
 		fail(filename, op_num+1, "no traversal start found");
 
 	start = i;
-	op[start].trav = trav;
+	op[start].group_len = op_num - start;
 
 	for (i = start; i <= op_num; i++)
 		op[i].group_start = start;
-
-	if (is_trivial_traverse(op+i, op_num-i)) {
-		/* Fill in a plentiful hash table. */
-		op[start].trav->hash = talloc_zero_array(op[i].trav,
-							 struct traverse_hash,
-							 trav->num * 2);
-		for (i = start; i < op_num; i++) {
-			unsigned int h;
-			if (op[i].op != OP_TDB_TRAVERSE)
-				continue;
-			h = hash_key(&op[i].key) % (trav->num * 2);
-			while (trav->hash[h].index)
-				h = (h + 1) % (trav->num * 2);
-			trav->hash[h].index = i+1;
-			trav->hash[h].key = op[i].key;
-		}
-	} else
-		trav->hash = NULL;
 }
 
 /* Keep -Wmissing-declarations happy: */
@@ -433,6 +389,101 @@ find_keyword (register const char *str, register unsigned int len);
 
 #include "keywords.c"
 
+struct depend {
+	/* We can have more than one */
+	struct list_node pre_list;
+	struct list_node post_list;
+	unsigned int needs_file;
+	unsigned int needs_opnum;
+	unsigned int satisfies_file;
+	unsigned int satisfies_opnum;
+};
+
+static void check_deps(const char *filename, struct op op[], unsigned int num)
+{
+#ifdef DEBUG_DEPS
+	unsigned int i;
+
+	for (i = 1; i < num; i++)
+		if (!list_empty(&op[i].pre))
+			fail(filename, i+1, "Still has dependencies");
+#endif
+}
+
+static void dump_pre(char *filename[], struct op *op[],
+		     unsigned int file, unsigned int i)
+{
+	struct depend *dep;
+
+	printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
+		op[file][i].serial);
+	list_for_each(&op[file][i].pre, dep, pre_list)
+		printf("    %s:%u (%u)\n",
+		       filename[dep->satisfies_file], dep->satisfies_opnum+1,
+		       op[dep->satisfies_file][dep->satisfies_opnum].serial);
+	check_deps(filename[file], op[file], i);
+}
+
+/* We simply read/write pointers, since we all are children. */
+static bool do_pre(struct tdb_context *tdb,
+		   char *filename[], struct op *op[],
+		   unsigned int file, int pre_fd, unsigned int i,
+		   bool backoff)
+{
+	while (!list_empty(&op[file][i].pre)) {
+		struct depend *dep;
+
+#if DEBUG_DEPS
+		printf("%s:%u:waiting for pre\n", filename[file], i+1);
+		fflush(stdout);
+#endif
+		if (backoff)
+			alarm(2);
+		else
+			alarm(10);
+		while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
+			if (errno == EINTR) {
+				if (backoff) {
+					warnx("%s:%u:avoiding deadlock",
+					      filename[file], i+1);
+					return false;
+				}
+				dump_pre(filename, op, file, i);
+				exit(1);
+			} else
+				errx(1, "Reading from pipe");
+		}
+		alarm(0);
+
+#if DEBUG_DEPS
+		printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
+		       dep->needs_opnum+1, filename[dep->satisfies_file],
+		       dep->satisfies_opnum+1);
+		fflush(stdout);
+#endif
+		/* This could be any op, not just this one. */
+		talloc_free(dep);
+	}
+	return true;
+}
+
+static void do_post(char *filename[], struct op *op[],
+		    unsigned int file, unsigned int i)
+{
+	struct depend *dep;
+
+	list_for_each(&op[file][i].post, dep, post_list) {
+#if DEBUG_DEPS
+		printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
+		       filename[dep->needs_file], dep->needs_opnum+1);
+#endif
+		if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
+		    != sizeof(dep))
+			err(1, "%s:%u failed to tell file %s",
+			    filename[file], i+1, filename[dep->needs_file]);
+	}
+}
+
 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 {
 	return data.dsize;
@@ -440,64 +491,53 @@ static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 
 static unsigned run_ops(struct tdb_context *tdb,
 			int pre_fd,
-			const char *filename,
-			struct op op[],
-			unsigned int start, unsigned int stop);
+			char *filename[],
+			struct op *op[],
+			unsigned int file,
+			unsigned int start, unsigned int stop,
+			bool backoff);
 
 struct traverse_info {
-	struct op *op;
-	const char *filename;
+	struct op **op;
+	char **filename;
+	unsigned file;
 	int pre_fd;
 	unsigned int start;
 	unsigned int i;
 };
 
-/* Trivial case: do whatever they did for this key. */
-static int trivial_traverse(struct tdb_context *tdb,
-			    TDB_DATA key, TDB_DATA data,
-			    void *_tinfo)
-{
-	struct traverse_info *tinfo = _tinfo;
-	struct traverse *trav = tinfo->op[tinfo->start].trav;
-	unsigned int h = hash_key(&key) % (trav->num * 2);
-
-	while (trav->hash[h].index) {
-		if (key_eq(trav->hash[h].key, key)) {
-			run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
-				trav->hash[h].index, trav->end);
-			tinfo->i++;
-			return 0;
-		}
-		h = (h + 1) % (trav->num * 2);
-	}
-	fail(tinfo->filename, tinfo->start + 1, "unexpected traverse key");
-}
-
 /* More complex.  Just do whatever's they did at the n'th entry. */
 static int nontrivial_traverse(struct tdb_context *tdb,
 			       TDB_DATA key, TDB_DATA data,
 			       void *_tinfo)
 {
 	struct traverse_info *tinfo = _tinfo;
-	struct traverse *trav = tinfo->op[tinfo->start].trav;
+	unsigned int trav_len = tinfo->op[tinfo->file][tinfo->start].group_len;
+	bool avoid_deadlock = false;
 
-	if (tinfo->i == trav->end) {
+	if (tinfo->i == tinfo->start + trav_len) {
 		/* This can happen if traverse expects to be empty. */
-		if (tinfo->start + 1 == trav->end)
+		if (trav_len == 1)
 			return 1;
-		fail(tinfo->filename, tinfo->start + 1,
+		fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 		     "traverse did not terminate");
 	}
 
-	if (tinfo->op[tinfo->i].op != OP_TDB_TRAVERSE)
-		fail(tinfo->filename, tinfo->start + 1,
+	if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
+		fail(tinfo->filename[tinfo->file], tinfo->start + 1,
 		     "%s:%u:traverse terminated early");
 
+#if TRAVERSALS_TAKE_TRANSACTION_LOCK
+	avoid_deadlock = true;
+#endif
+
 	/* Run any normal ops. */
 	tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
-			   tinfo->i+1, trav->end);
+			   tinfo->file, tinfo->i+1, tinfo->start + trav_len,
+			   avoid_deadlock);
 
-	if (tinfo->i == trav->end)
+	/* We backed off, or we hit OP_TDB_TRAVERSE_END. */
+	if (tinfo->op[tinfo->file][tinfo->i].op != OP_TDB_TRAVERSE)
 		return 1;
 
 	return 0;
@@ -505,188 +545,155 @@ static int nontrivial_traverse(struct tdb_context *tdb,
 
 static unsigned op_traverse(struct tdb_context *tdb,
 			    int pre_fd,
-			    const char *filename,
+			    char *filename[],
+			    unsigned int file,
 			    int (*traversefn)(struct tdb_context *,
 					      tdb_traverse_func, void *),
-			    struct op op[],
+			    struct op *op[],
 			    unsigned int start)
 {
-	struct traverse *trav = op[start].trav;
-	struct traverse_info tinfo = { op, filename, pre_fd, start, start+1 };
-
-	/* Trivial case. */
-	if (trav->hash) {
-		int ret = traversefn(tdb, trivial_traverse, &tinfo);
-		if (ret != trav->num)
-			fail(filename, start+1, "short traversal %i", ret);
-		return trav->end;
-	}
+	struct traverse_info tinfo = { op, filename, file, pre_fd,
+				       start, start+1 };
 
 	traversefn(tdb, nontrivial_traverse, &tinfo);
 
 	/* Traversing in wrong order can have strange effects: eg. if
 	 * original traverse went A (delete A), B, we might do B
 	 * (delete A).  So if we have ops left over, we do it now. */
-	while (tinfo.i != trav->end) {
-		if (op[tinfo.i].op == OP_TDB_TRAVERSE)
+	while (tinfo.i != start + op[file][start].group_len) {
+		if (op[file][tinfo.i].op == OP_TDB_TRAVERSE)
 			tinfo.i++;
 		else
-			tinfo.i = run_ops(tdb, pre_fd, filename, op,
-					  tinfo.i, trav->end);
+			tinfo.i = run_ops(tdb, pre_fd, filename, op, file,
+					  tinfo.i,
+					  start + op[file][start].group_len,
+					  false);
 	}
-	return trav->end;
-}
-
-struct depend {
-	/* We can have more than one */
-	struct list_node list;
-	unsigned int file;
-	unsigned int op;
-};
-
-static void do_pre(const char *filename, int pre_fd,
-		   struct op op[], unsigned int i)
-{
-	while (op[i].pre != 0) {
-		unsigned int opnum;
 
-#if DEBUG_DEPS
-		printf("%s:%u:waiting for pre\n", filename, i+1);
-		fflush(stdout);
-#endif
-		if (read(pre_fd, &opnum, sizeof(opnum)) != sizeof(opnum))
-			errx(1, "Reading from pipe");
-
-#if DEBUG_DEPS
-		printf("%s:%u:got pre %u (%u)\n", filename, i+1, opnum+1,
-		       op[opnum].pre);
-		fflush(stdout);
-#endif
-		/* This could be any op, not just this one. */
-		if (op[opnum].pre == 0)
-			errx(1, "Got extra notification for op line %u",
-			     opnum + 1);
-		if (opnum < i)
-			errx(1, "%s:%u: got late notification for line %u",
-			     filename, i + 1, opnum + 1);
-		op[opnum].pre--;
-	}
+	return tinfo.i;
 }
 
-static void do_post(const char *filename, const struct op op[], unsigned int i)
+static void break_out(int sig)
 {
-	struct depend *dep;
-
-	list_for_each(&op[i].post, dep, list) {
-#if DEBUG_DEPS
-		printf("%s:%u:sending %u to file %u\n", filename, i+1,
-		       dep->op+1, dep->file);
-#endif
-		if (write(pipes[dep->file].fd[1], &dep->op, sizeof(dep->op))
-		    != sizeof(dep->op))
-			err(1, "Failed to tell file %u", dep->file);
-	}
 }
 
 static __attribute__((noinline))
 unsigned run_ops(struct tdb_context *tdb,
 		 int pre_fd,
-		 const char *filename,
-		 struct op op[], unsigned int start, unsigned int stop)
+		 char *filename[],
+		 struct op *op[],
+		 unsigned int file,
+		 unsigned int start, unsigned int stop,
+		 bool backoff)
 {
 	unsigned int i;
+	struct sigaction sa;
 
+	sa.sa_handler = break_out;
+	sa.sa_flags = 0;
+
+	sigaction(SIGALRM, &sa, NULL);
 	for (i = start; i < stop; i++) {
-		do_pre(filename, pre_fd, op, i);
+		if (!do_pre(tdb, filename, op, file, pre_fd, i, backoff))
+			return i;
 
-		switch (op[i].op) {
+		switch (op[file][i].op) {
 		case OP_TDB_LOCKALL:
-			try(tdb_lockall(tdb), op[i].ret);
+			try(tdb_lockall(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_LOCKALL_MARK:
-			try(tdb_lockall_mark(tdb), op[i].ret);
+			try(tdb_lockall_mark(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_LOCKALL_UNMARK:
-			try(tdb_lockall_unmark(tdb), op[i].ret);
+			try(tdb_lockall_unmark(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_LOCKALL_NONBLOCK:
-			unreliable(tdb_lockall_nonblock(tdb), op[i].ret,
+			unreliable(tdb_lockall_nonblock(tdb), op[file][i].ret,
 				   tdb_lockall(tdb), tdb_unlockall(tdb));
 			break;
 		case OP_TDB_UNLOCKALL:
-			try(tdb_unlockall(tdb), op[i].ret);
+			try(tdb_unlockall(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_LOCKALL_READ:
-			try(tdb_lockall_read(tdb), op[i].ret);
+			try(tdb_lockall_read(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_LOCKALL_READ_NONBLOCK:
-			unreliable(tdb_lockall_read_nonblock(tdb), op[i].ret,
+			unreliable(tdb_lockall_read_nonblock(tdb),
+				   op[file][i].ret,
 				   tdb_lockall_read(tdb),
 				   tdb_unlockall_read(tdb));
 			break;
 		case OP_TDB_UNLOCKALL_READ:
-			try(tdb_unlockall_read(tdb), op[i].ret);
+			try(tdb_unlockall_read(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_CHAINLOCK:
-			try(tdb_chainlock(tdb, op[i].key), op[i].ret);
+			try(tdb_chainlock(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_CHAINLOCK_NONBLOCK:
-			unreliable(tdb_chainlock_nonblock(tdb, op[i].key),
-				   op[i].ret,
-				   tdb_chainlock(tdb, op[i].key),
-				   tdb_chainunlock(tdb, op[i].key));
+			unreliable(tdb_chainlock_nonblock(tdb, op[file][i].key),
+				   op[file][i].ret,
+				   tdb_chainlock(tdb, op[file][i].key),
+				   tdb_chainunlock(tdb, op[file][i].key));
 			break;
 		case OP_TDB_CHAINLOCK_MARK:
-			try(tdb_chainlock_mark(tdb, op[i].key), op[i].ret);
+			try(tdb_chainlock_mark(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_CHAINLOCK_UNMARK:
-			try(tdb_chainlock_unmark(tdb, op[i].key), op[i].ret);
+			try(tdb_chainlock_unmark(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_CHAINUNLOCK:
-			try(tdb_chainunlock(tdb, op[i].key), op[i].ret);
+			try(tdb_chainunlock(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_CHAINLOCK_READ:
-			try(tdb_chainlock_read(tdb, op[i].key), op[i].ret);
+			try(tdb_chainlock_read(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_CHAINUNLOCK_READ:
-			try(tdb_chainunlock_read(tdb, op[i].key), op[i].ret);
+			try(tdb_chainunlock_read(tdb, op[file][i].key),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_PARSE_RECORD:
-			try(tdb_parse_record(tdb, op[i].key, get_len, NULL),
-			    op[i].ret);
+			try(tdb_parse_record(tdb, op[file][i].key, get_len,
+					     NULL),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_EXISTS:
-			try(tdb_exists(tdb, op[i].key), op[i].ret);
+			try(tdb_exists(tdb, op[file][i].key), op[file][i].ret);
 			break;
 		case OP_TDB_STORE:
-			try(tdb_store(tdb, op[i].key, op[i].data, op[i].flag),
-			    op[i].ret < 0 ? op[i].ret : 0);
+			try(tdb_store(tdb, op[file][i].key, op[file][i].data,
+				      op[file][i].flag),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_APPEND:
-			try(tdb_append(tdb, op[i].key, op[i].data),
-			    op[i].ret < 0 ? op[i].ret : 0);
+			try(tdb_append(tdb, op[file][i].key, op[file][i].data),
+			    op[file][i].ret);
 			break;
 		case OP_TDB_GET_SEQNUM:
-			try(tdb_get_seqnum(tdb), op[i].ret);
+			try(tdb_get_seqnum(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_WIPE_ALL:
-			try(tdb_wipe_all(tdb), op[i].ret);
+			try(tdb_wipe_all(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_TRANSACTION_START:
-			try(tdb_transaction_start(tdb), op[i].ret);
+			try(tdb_transaction_start(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_TRANSACTION_CANCEL:
-			try(tdb_transaction_cancel(tdb), op[i].ret);
+			try(tdb_transaction_cancel(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_TRANSACTION_COMMIT:
-			try(tdb_transaction_commit(tdb), op[i].ret);
+			try(tdb_transaction_commit(tdb), op[file][i].ret);
 			break;
 		case OP_TDB_TRAVERSE_READ_START:
-			i = op_traverse(tdb, pre_fd, filename,
+			i = op_traverse(tdb, pre_fd, filename, file,
 					tdb_traverse_read, op, i);
 			break;
 		case OP_TDB_TRAVERSE_START:
-			i = op_traverse(tdb, pre_fd, filename,
+			i = op_traverse(tdb, pre_fd, filename, file,
 					tdb_traverse, op, i);
 			break;
 		case OP_TDB_TRAVERSE:
@@ -694,31 +701,50 @@ unsigned run_ops(struct tdb_context *tdb,
 			 * done our ops. */
 			return i;
 		case OP_TDB_TRAVERSE_END:
-			fail(filename, i+1, "unepxected end traverse");
+			fail(filename[file], i+1, "unexpected end traverse");
 		/* FIXME: These must be treated like traverse. */
 		case OP_TDB_FIRSTKEY:
-			if (!key_eq(tdb_firstkey(tdb), op[i].data))
-				fail(filename, i+1, "bad firstkey");
+			if (!key_eq(tdb_firstkey(tdb), op[file][i].data))
+				fail(filename[file], i+1, "bad firstkey");
 			break;
 		case OP_TDB_NEXTKEY:
-			if (!key_eq(tdb_nextkey(tdb, op[i].key), op[i].data))
-				fail(filename, i+1, "bad nextkey");
+			if (!key_eq(tdb_nextkey(tdb, op[file][i].key),
+				    op[file][i].data))
+				fail(filename[file], i+1, "bad nextkey");
 			break;
 		case OP_TDB_FETCH: {
-			TDB_DATA f = tdb_fetch(tdb, op[i].key);
-			if (!key_eq(f, op[i].data))
-				fail(filename, i+1, "bad fetch %u", f.dsize);
+			TDB_DATA f = tdb_fetch(tdb, op[file][i].key);
+			if (!key_eq(f, op[file][i].data))
+				fail(filename[file], i+1, "bad fetch %u",
+				     f.dsize);
 			break;
 		}
 		case OP_TDB_DELETE:
-			try(tdb_delete(tdb, op[i].key), op[i].ret);
+			try(tdb_delete(tdb, op[file][i].key), op[file][i].ret);
 			break;
 		}
-		do_post(filename, op, i);
+		do_post(filename, op, file, i);
 	}
 	return i;
 }
 
+/* tdbtorture, in particular, can do a tdb_close with a transaction in
+ * progress. */
+static struct op *maybe_cancel_transaction(const char *filename,
+					   struct op *op, unsigned int *num)
+{
+	unsigned int start = op_transaction_start(op, *num);
+
+	if (start) {
+		char *words[] = { "<unknown>", "tdb_close", NULL };
+		add_op(filename, &op, *num, op[start].serial,
+		       OP_TDB_TRANSACTION_CANCEL);
+		op_analyze_transaction(filename, op, *num, words);
+		(*num)++;
+	}
+	return op;
+}
+
 static struct op *load_tracefile(const char *filename, unsigned int *num,
 				 unsigned int *hashsize,
 				 unsigned int *tdb_flags,
@@ -761,7 +787,8 @@ static struct op *load_tracefile(const char *filename, unsigned int *num,
 					     "lines after tdb_close");
 				*num = i;
 				talloc_free(lines);
-				return op;
+				return maybe_cancel_transaction(filename,
+								op, num);
 			}
 			fail(filename, i+1, "Unknown operation '%s'", words[1]);
 		}
@@ -774,7 +801,7 @@ static struct op *load_tracefile(const char *filename, unsigned int *num,
 	      filename, i);
 	talloc_free(lines);
 	*num = i - 1;
-	return op;
+	return maybe_cancel_transaction(filename, op, num);
 }
 
 /* We remember all the keys we've ever seen, and who has them. */
@@ -789,16 +816,134 @@ struct keyinfo {
 	struct key_user *user;
 };
 
-static bool changes_db(const struct op *op)
+static const TDB_DATA must_not_exist;
+static const TDB_DATA must_exist;
+static const TDB_DATA not_exists_or_empty;
+
+/* NULL means doesn't care if it exists or not, &must_exist means
+ * it must exist but we don't care what, &must_not_exist means it must
+ * not exist, otherwise the data it needs. */
+static const TDB_DATA *needs(const struct op *op)
 {
-	if (op->ret != 0)
-		return false;
+	switch (op->op) {
+	/* FIXME: Pull forward deps, since we can deadlock */
+	case OP_TDB_CHAINLOCK:
+	case OP_TDB_CHAINLOCK_NONBLOCK:
+	case OP_TDB_CHAINLOCK_MARK:
+	case OP_TDB_CHAINLOCK_UNMARK:
+	case OP_TDB_CHAINUNLOCK:
+	case OP_TDB_CHAINLOCK_READ:
+	case OP_TDB_CHAINUNLOCK_READ:
+		return NULL;
+
+	case OP_TDB_APPEND:
+		if (op->append.pre.dsize == 0)
+			return &not_exists_or_empty;
+		return &op->append.pre;
+
+	case OP_TDB_STORE:
+		if (op->flag == TDB_INSERT) {
+			if (op->ret < 0)
+				return &must_exist;
+			else
+				return &must_not_exist;
+		} else if (op->flag == TDB_MODIFY) {
+			if (op->ret < 0)
+				return &must_not_exist;
+			else
+				return &must_exist;
+		}
+		/* No flags?  Don't care */
+		return NULL;
+
+	case OP_TDB_EXISTS:
+		if (op->ret == 1)
+			return &must_exist;
+		else
+			return &must_not_exist;
+
+	case OP_TDB_PARSE_RECORD:
+		if (op->ret < 0)
+			return &must_not_exist;
+		return &must_exist;
+
+	/* FIXME: handle these. */
+	case OP_TDB_WIPE_ALL:
+	case OP_TDB_FIRSTKEY:
+	case OP_TDB_NEXTKEY:
+	case OP_TDB_GET_SEQNUM:
+	case OP_TDB_TRAVERSE:
+	case OP_TDB_TRANSACTION_COMMIT:
+	case OP_TDB_TRANSACTION_CANCEL:
+	case OP_TDB_TRANSACTION_START:
+		return NULL;
+
+	case OP_TDB_FETCH:
+		if (!op->data.dptr)
+			return &must_not_exist;
+		return &op->data;
+
+	case OP_TDB_DELETE:
+		if (op->ret < 0)
+			return &must_not_exist;
+		return &must_exist;
+
+	default:
+		errx(1, "Unexpected op %i", op->op);
+	}
+	
+}
+
+static bool is_transaction(const struct op *op)
+{
+	return op->op == OP_TDB_TRANSACTION_START;
+}
+
+/* What's the data after this op?  pre if nothing changed. */
+static const TDB_DATA *gives(const TDB_DATA *key, const TDB_DATA *pre,
+			     const struct op *op)
+{
+	if (is_transaction(op)) {
+		unsigned int i;
+
+		/* Cancelled transactions don't change anything. */
+		if (op[op->group_len].op == OP_TDB_TRANSACTION_CANCEL)
+			return pre;
+		assert(op[op->group_len].op == OP_TDB_TRANSACTION_COMMIT);
+
+		for (i = 1; i < op->group_len; i++) {
+			/* This skips nested transactions, too */
+			if (op[i].op != OP_TDB_TRAVERSE
+			    && key_eq(op[i].key, *key))
+				pre = gives(key, pre, &op[i]);
+		}
+		return pre;
+	}
+
+	/* Failed ops don't change state of db. */
+	if (op->ret < 0)
+		return pre;
+
+	if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
+		return &tdb_null;
+
+	if (op->op == OP_TDB_APPEND)
+		return &op->append.post;
+
+	if (op->op == OP_TDB_STORE)
+		return &op->data;
+
+	return pre;
+}
+
+static bool in_transaction(const struct op op[], unsigned int i)
+{
+	return op[i].group_start && is_transaction(&op[op[i].group_start]);
+}
 
-	return op->op == OP_TDB_STORE
-		|| op->op == OP_TDB_APPEND
-		|| op->op == OP_TDB_WIPE_ALL
-		|| op->op == OP_TDB_TRANSACTION_COMMIT
-		|| op->op == OP_TDB_DELETE;
+static bool in_traverse(const struct op op[], unsigned int i)
+{
+	return op[i].group_start && !is_transaction(&op[op[i].group_start]);
 }
 
 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
@@ -807,30 +952,12 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 	unsigned int i, j, h;
 	struct keyinfo *hash;
 
-	/* Gcc nexted function extension.  How cool is this? */
-	int compare_user_serial(const void *_a, const void *_b)
-	{
-		const struct key_user *a = _a, *b = _b;
-		int ret = op[a->file][a->op_num].serial
-			- op[b->file][b->op_num].serial;
-
-		/* Fetches don't inc serial, so we put changes first. */
-		if (ret == 0) {
-			if (changes_db(&op[a->file][a->op_num])
-			    && !changes_db(&op[b->file][b->op_num]))
-				return -1;
-			if (changes_db(&op[b->file][b->op_num])
-			    && !changes_db(&op[a->file][a->op_num]))
-				return 1;
-		}
-		return ret;
-	}
-
 	hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
 	for (i = 0; i < num; i++) {
 		for (j = 1; j < num_ops[i]; j++) {
 			/* We can't do this on allocation, due to realloc. */
 			list_head_init(&op[i][j].post);
+			list_head_init(&op[i][j].pre);
 
 			if (!op[i][j].key.dptr)
 				continue;
@@ -856,80 +983,414 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 			hash[h].user = talloc_realloc(hash, hash[h].user,
 						     struct key_user,
 						     hash[h].num_users+1);
-			hash[h].user[hash[h].num_users].op_num = j;
+
+			/* If it's in a transaction, it's the transaction which
+			 * matters from an analysis POV. */
+			if (in_transaction(op[i], j)) {
+				unsigned start = op[i][j].group_start;
+
+				/* Don't include twice. */
+				if (hash[h].num_users
+				    && hash[h].user[hash[h].num_users-1].file
+					== i
+				    && hash[h].user[hash[h].num_users-1].op_num
+					== start)
+					continue;
+
+				hash[h].user[hash[h].num_users].op_num = start;
+			} else
+				hash[h].user[hash[h].num_users].op_num = j;
 			hash[h].user[hash[h].num_users].file = i;
 			hash[h].num_users++;
 		}
 	}
 
-	/* Now sort into seqnum order. */
-	for (h = 0; h < total_keys * 2; h++)
-		qsort(hash[h].user, hash[h].num_users, sizeof(hash[h].user[0]),
-		      compare_user_serial);
-
 	return hash;
 }
 
+static bool satisfies(const TDB_DATA *key, const TDB_DATA *data,
+		      const struct op *op)
+{
+	const TDB_DATA *need = NULL;
+
+	if (is_transaction(op)) {
+		unsigned int i;
+
+		/* Look through for an op in this transaction which
+		 * needs this key. */
+		for (i = 1; i < op->group_len; i++) {
+			if (op[i].op != OP_TDB_TRAVERSE
+			    && key_eq(op[i].key, *key)) {
+				need = needs(&op[i]);
+				/* tdb_exists() is special: there might be
+				 * something in the transaction with more
+				 * specific requirements.  Other ops don't have
+				 * specific requirements (eg. store or delete),
+				 * but they change the value so we can't get
+				 * more information from future ops. */
+				if (op[i].op != OP_TDB_EXISTS)
+					break;
+			}
+		}
+	} else
+		need = needs(op);
+
+	/* Don't need anything?  Cool. */
+	if (!need)
+		return true;
+
+	/* This should be tdb_null or a real value. */
+	assert(data != &must_exist);
+	assert(data != &must_not_exist);
+	assert(data != &not_exists_or_empty);
+
+	/* Must not exist?  data must not exist. */
+	if (need == &must_not_exist)
+		return data == &tdb_null;
+
+	/* Must exist? */
+	if (need == &must_exist)
+		return data != &tdb_null;
+
+	/* Either noexist or empty. */
+	if (need == &not_exists_or_empty)
+		return data->dsize == 0;
+
+	/* Needs something specific. */
+	return key_eq(*data, *need);
+}
+
+static void move_to_front(struct key_user res[], unsigned off, unsigned elem)
+{
+	if (elem != off) {
+		struct key_user tmp = res[elem];
+		memmove(res + off + 1, res + off, (elem - off)*sizeof(res[0]));
+		res[off] = tmp;
+	}
+}
+
+static void restore_to_pos(struct key_user res[], unsigned off, unsigned elem)
+{
+	if (elem != off) {
+		struct key_user tmp = res[off];
+		memmove(res + off, res + off + 1, (elem - off)*sizeof(res[0]));
+		res[elem] = tmp;
+	}
+}
+
+static bool sort_deps(char *filename[], struct op *op[],
+		      struct key_user res[],
+		      unsigned off, unsigned num,
+		      const TDB_DATA *key, const TDB_DATA *data,
+		      unsigned num_files, unsigned fuzz)
+{
+	unsigned int i, files_done;
+	struct op *this_op;
+	bool done[num_files];
+
+	/* Does this make serial numbers go backwards?  Allow a little fuzz. */
+	if (off > 0) {
+		int serial1 = op[res[off-1].file][res[off-1].op_num].serial;
+		int serial2 = op[res[off].file][res[off].op_num].serial;
+
+		if (serial1 - serial2 > (int)fuzz) {
+#if DEBUG_DEPS
+			printf("Serial jump too far (%u -> %u)\n",
+			       serial1, serial2);
+#endif
+			return false;
+		}
+	}
+
+	/* One or none left?  We're sorted. */
+	if (off + 1 >= num)
+		return true;
+
+	memset(done, 0, sizeof(done));
+
+	/* Since ops within a trace file are ordered, we just need to figure
+	 * out which file to try next.  Since we don't take into account
+	 * inter-key relationships (which exist by virtue of trace file order),
+	 * we minimize the chance of harm by trying to keep in serial order. */
+	for (files_done = 0, i = off; i < num && files_done < num_files; i++) {
+		if (done[res[i].file])
+			continue;
+
+		this_op = &op[res[i].file][res[i].op_num];
+
+		/* Is what we have good enough for this op? */
+		if (satisfies(key, data, this_op)) {
+			move_to_front(res, off, i);
+			if (sort_deps(filename, op, res, off+1, num,
+				      key, gives(key, data, this_op),
+				      num_files, fuzz))
+				return true;
+			restore_to_pos(res, off, i);
+		}
+		done[res[i].file] = true;
+		files_done++;
+	}
+
+	/* No combination worked. */
+	return false;
+}
+
+static void check_dep_sorting(struct key_user user[], unsigned num_users,
+			      unsigned num_files)
+{
+#if DEBUG_DEPS
+	unsigned int i;
+	unsigned minima[num_files];
+
+	memset(minima, 0, sizeof(minima));
+	for (i = 0; i < num_users; i++) {
+		assert(minima[user[i].file] < user[i].op_num);
+		minima[user[i].file] = user[i].op_num;
+	}
+#endif
+}
+
+/* All these ops happen on the same key.  Which comes first?
+ *
+ * This can happen both because read ops or failed write ops don't
+ * change serial number, and also due to race since we access the
+ * number unlocked (the race can cause less detectable ordering problems,
+ * in which case we'll deadlock and report: fix manually in that case).
+ */
+static void figure_deps(char *filename[], struct op *op[],
+			const TDB_DATA *key, struct key_user user[],
+			unsigned num_users, unsigned num_files)
+{
+	/* We assume database starts empty. */
+	const struct TDB_DATA *data = &tdb_null;
+	unsigned int fuzz;
+
+	/* We prefer to keep strict serial order if possible: it's the
+	 * most likely.  We get more lax if that fails. */
+	for (fuzz = 0; fuzz < 100; fuzz = (fuzz + 1)*2) {
+		if (sort_deps(filename, op, user, 0, num_users, key, data,
+			      num_files, fuzz))
+			break;
+	}
+
+	if (fuzz >= 100)
+		fail(filename[user[0].file], user[0].op_num+1,
+		     "Could not resolve inter-dependencies");
+
+	check_dep_sorting(user, num_users, num_files);
+}
+
+static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
+		     unsigned int num)
+{
+	unsigned int h;
+
+	/* Gcc nexted function extension.  How cool is this? */
+	int compare_serial(const void *_a, const void *_b)
+	{
+		const struct key_user *a = _a, *b = _b;
+
+		/* First, maintain order within any trace file. */
+		if (a->file == b->file)
+			return a->op_num - b->op_num;
+
+		/* Otherwise, arrange by serial order. */
+		return op[a->file][a->op_num].serial
+			- op[b->file][b->op_num].serial;
+	}
+
+	/* Now sort into serial order. */
+	for (h = 0; h < total_keys * 2; h++) {
+		struct key_user *user = hash[h].user;
+
+		qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
+		figure_deps(filename, op, &hash[h].key, user, hash[h].num_users,
+			    num);
+	}
+}
+
+static int destroy_depend(struct depend *dep)
+{
+	list_del(&dep->pre_list);
+	list_del(&dep->post_list);
+	return 0;
+}
+
 static void add_dependency(void *ctx,
 			   struct op *op[],
+			   char *filename[],
 			   unsigned int needs_file,
 			   unsigned int needs_opnum,
 			   unsigned int satisfies_file,
 			   unsigned int satisfies_opnum)
 {
-	struct depend *post;
-	unsigned int needs_start, sat_start;
+	struct depend *dep;
 
-	needs_start = op[needs_file][needs_opnum].group_start;
-	sat_start = op[satisfies_file][satisfies_opnum].group_start;
+	/* We don't depend on ourselves. */
+	if (needs_file == satisfies_file) {
+		assert(satisfies_opnum < needs_opnum);
+		return;
+	}
+
+#if DEBUG_DEPS
+	printf("%s:%u: depends on %s:%u\n",
+	       filename[needs_file], needs_opnum+1,
+	       filename[satisfies_file], satisfies_opnum+1);
+#endif
 
-	/* If needs is in a transaction, we need it before start. */
-	if (needs_start) {
-		switch (op[needs_file][needs_start].op) {
-		case OP_TDB_TRANSACTION_START:
 #if TRAVERSALS_TAKE_TRANSACTION_LOCK
-		case OP_TDB_TRAVERSE_START:
-		case OP_TDB_TRAVERSE_READ_START:
+	/* If something in a traverse depends on something in another
+	 * traverse/transaction, it creates a dependency between the
+	 * two groups. */
+	if ((in_traverse(op[satisfies_file], satisfies_opnum)
+	     && op[needs_file][needs_opnum].group_start)
+	    || (in_traverse(op[needs_file], needs_opnum)
+		&& op[satisfies_file][satisfies_opnum].group_start)) {
+		unsigned int sat;
+
+		/* We are satisfied by end of group. */
+		sat = op[satisfies_file][satisfies_opnum].group_start;
+		satisfies_opnum = sat + op[satisfies_file][sat].group_len;
+		/* And we need that done by start of our group. */
+		needs_opnum = op[needs_file][needs_opnum].group_start;
+	}
+
+	/* There is also this case:
+	 *  <traverse> <read foo> ...
+	 *  <transaction> ... </transaction> <create foo>
+	 * Where if we start the traverse then wait, we could block
+	 * the transaction and deadlock.
+	 *
+	 * We try to address this by ensuring that where seqnum indicates it's
+	 * possible, we wait for <create foo> before *starting* traverse.
+	 */
+	else if (in_traverse(op[needs_file], needs_opnum)) {
+		struct op *need = &op[needs_file][needs_opnum];
+		if (op[needs_file][need->group_start].serial >
+		    op[satisfies_file][satisfies_opnum].serial) {
+			needs_opnum = need->group_start;
+		}
+	}
 #endif
-			needs_opnum = needs_start;
-#ifdef DEBUG_DEPS
-			printf("  -> Back to %u\n", needs_start+1);
-			fflush(stdout);
+
+ 	/* If you depend on a transaction, you actually depend on it ending. */
+ 	if (is_transaction(&op[satisfies_file][satisfies_opnum])) {
+ 		satisfies_opnum
+ 			+= op[satisfies_file][satisfies_opnum].group_len;
+#if DEBUG_DEPS
+		printf("-> Actually end of transaction %s:%u\n",
+		       filename[satisfies_file], satisfies_opnum+1);
 #endif
-			break;
-		default:
-			break;
+ 	} else
+		/* We should never create a dependency from middle of
+		 * a transaction. */
+ 		assert(!in_transaction(op[satisfies_file], satisfies_opnum)
+		       || op[satisfies_file][satisfies_opnum].op
+ 		       == OP_TDB_TRANSACTION_COMMIT
+ 		       || op[satisfies_file][satisfies_opnum].op
+ 		       == OP_TDB_TRANSACTION_CANCEL);
+
+	assert(op[needs_file][needs_opnum].op != OP_TDB_TRAVERSE);
+	assert(op[satisfies_file][satisfies_opnum].op != OP_TDB_TRAVERSE);
+
+	dep = talloc(ctx, struct depend);
+	dep->needs_file = needs_file;
+	dep->needs_opnum = needs_opnum;
+	dep->satisfies_file = satisfies_file;
+	dep->satisfies_opnum = satisfies_opnum;
+	list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
+	list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
+	talloc_set_destructor(dep, destroy_depend);
+}
+
+static bool changes_db(const TDB_DATA *key, const struct op *op)
+{
+	return gives(key, NULL, op) != NULL;
+}
+
+static void depend_on_previous(struct op *op[],
+			       char *filename[],
+			       unsigned int num,
+			       struct key_user user[],
+			       unsigned int i,
+			       int prev)
+{
+	bool deps[num];
+	int j;
+
+	if (i == 0)
+		return;
+
+	if (prev == i - 1) {
+		/* Just depend on previous. */
+		add_dependency(NULL, op, filename,
+			       user[i].file, user[i].op_num,
+			       user[prev].file, user[prev].op_num);
+		return;
+	}
+
+	/* We have to wait for the readers.  Find last one in *each* file. */
+	memset(deps, 0, sizeof(deps));
+	deps[user[i].file] = true;
+	for (j = i - 1; j > prev; j--) {
+		if (!deps[user[j].file]) {
+			add_dependency(NULL, op, filename,
+				       user[i].file, user[i].op_num,
+				       user[j].file, user[j].op_num);
+			deps[user[j].file] = true;
 		}
 	}
+}
 
-	/* If satisfies is in a transaction, we wait until after commit. */
-	/* FIXME: If transaction is cancelled, don't need dependency. */
-	if (sat_start) {
-		if (op[satisfies_file][sat_start].op
-		    == OP_TDB_TRANSACTION_START) {
-			satisfies_opnum
-				= op[satisfies_file][sat_start].transaction_end;
-#ifdef DEBUG_DEPS
-			printf("  -> Depends on %u\n", satisfies_opnum+1);
-			fflush(stdout);
-#endif
+/* This is simple, but not complete.  We don't take into account
+ * indirect dependencies. */
+static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
+				  unsigned int num)
+{
+	unsigned int i, j;
+
+	/* There can only be one real dependency on each file */
+	for (i = 0; i < num; i++) {
+		for (j = 1; j < num_ops[i]; j++) {
+			struct depend *dep, *next;
+			struct depend *prev[num];
+
+			memset(prev, 0, sizeof(prev));
+
+			list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
+				if (!prev[dep->satisfies_file]) {
+					prev[dep->satisfies_file] = dep;
+					continue;
+				}
+				if (prev[dep->satisfies_file]->satisfies_opnum
+				    < dep->satisfies_opnum) {
+					talloc_free(prev[dep->satisfies_file]);
+					prev[dep->satisfies_file] = dep;
+				} else
+					talloc_free(dep);
+			}
 		}
 	}
 
-	post = talloc(ctx, struct depend);
-	post->file = needs_file;
-	post->op = needs_opnum;
-	list_add(&op[satisfies_file][satisfies_opnum].post, &post->list);
+	for (i = 0; i < num; i++) {
+		int deps[num];
 
-	op[needs_file][needs_opnum].pre++;
-#ifdef DEBUG_DEPS
-	if (op[needs_file][needs_opnum].pre > 1) {
-		printf("   (File %u opnum %u hash %u needs)\n",
-		       needs_file, needs_opnum+1,
-		       op[needs_file][needs_opnum].pre);
-		fflush(stdout);
+		for (j = 0; j < num; j++)
+			deps[j] = -1;
+
+		for (j = 1; j < num_ops[i]; j++) {
+			struct depend *dep, *next;
+
+			list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
+				if (deps[dep->satisfies_file]
+				    >= (int)dep->satisfies_opnum)
+					talloc_free(dep);
+				else
+					deps[dep->satisfies_file]
+						= dep->satisfies_opnum;
+			}
+		}
 	}
-#endif
 }
 
 static void derive_dependencies(char *filename[],
@@ -937,35 +1398,44 @@ static void derive_dependencies(char *filename[],
 				unsigned int num)
 {
 	struct keyinfo *hash;
-	unsigned int i;
+	unsigned int h, i;
 
 	/* Create hash table for faster key lookup. */
 	hash = hash_ops(op, num_ops, num);
 
-	/* We make the naive assumption that two ops on the same key
-	 * have to be ordered; it's overkill. */
-	for (i = 0; i < total_keys * 2; i++) {
-		unsigned int j;
+	/* Sort them by serial number. */
+	sort_ops(hash, filename, op, num);
+
+	/* Create dependencies back to the last change, rather than
+	 * creating false dependencies by naively making each one
+	 * depend on the previous.  This has two purposes: it makes
+	 * later optimization simpler, and it also avoids deadlock with
+	 * same sequence number ops inside traversals (if one
+	 * traversal doesn't write anything, two ops can have the same
+	 * sequence number yet we can create a traversal dependency
+	 * the other way). */
+	for (h = 0; h < total_keys * 2; h++) {
+		int prev = -1;
+
+		if (hash[h].num_users < 2)
+			continue;
 
-		for (j = 1; j < hash[i].num_users; j++) {
-			/* We don't depend on ourselves. */
-			if (hash[i].user[j].file == hash[i].user[j-1].file)
-				continue;
-#if DEBUG_DEPS
-			printf("%s:%u: depends on %s:%u\n",
-			       filename[hash[i].user[j].file],
-			       hash[i].user[j].op_num+1,
-			       filename[hash[i].user[j-1].file],
-			       hash[i].user[j-1].op_num+1);
-			fflush(stdout);
-#endif
-			add_dependency(hash, op,
-				       hash[i].user[j].file,
-				       hash[i].user[j].op_num,
-				       hash[i].user[j-1].file,
-				       hash[i].user[j-1].op_num);
+		for (i = 0; i < hash[h].num_users; i++) {
+			if (changes_db(&hash[h].key, &op[hash[h].user[i].file]
+				       [hash[h].user[i].op_num])) {
+				depend_on_previous(op, filename, num,
+						   hash[h].user, i, prev);
+				prev = i;
+			} else if (prev >= 0)
+				add_dependency(hash, op, filename,
+					       hash[h].user[i].file,
+					       hash[h].user[i].op_num,
+					       hash[h].user[prev].file,
+					       hash[h].user[prev].op_num);
 		}
 	}
+
+	optimize_dependencies(op, num_ops, num);
 }
 
 int main(int argc, char *argv[])
@@ -975,28 +1445,40 @@ int main(int argc, char *argv[])
 	struct op *op[argc];
 	int fds[2];
 	char c;
+	bool ok = true;
 
 	if (argc < 3)
 		errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
 
 	pipes = talloc_array(NULL, struct pipe, argc - 2);
 	for (i = 0; i < argc - 2; i++) {
+		printf("Loading tracefile %s...", argv[2+i]);
+		fflush(stdout);
 		op[i] = load_tracefile(argv[2+i], &num_ops[i], &hashsize[i],
 				       &tdb_flags[i], &open_flags[i]);
 		if (pipe(pipes[i].fd) != 0)
 			err(1, "creating pipe");
+		printf("done\n");
 	}
 
+	printf("Calculating inter-dependencies...");
+	fflush(stdout);
 	derive_dependencies(argv+2, op, num_ops, i);
+	printf("done\n");
 
 	/* Don't fork for single arg case: simple debugging. */
 	if (argc == 3) {
 		struct tdb_context *tdb;
-		tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
-				  open_flags[0], 0600,
-				  NULL, hash_key);
-		run_ops(tdb, pipes[0].fd[0], argv[2],
-			op[0], 1, num_ops[0]);
+		tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0]|TDB_NOSYNC,
+				  open_flags[0], 0600, NULL, hash_key);
+		printf("Single threaded run...");
+		fflush(stdout);
+
+		run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1, num_ops[0],
+			false);
+		check_deps(argv[2], op[0], num_ops[0]);
+
+		printf("done\n");
 		exit(0);
 	}
 
@@ -1011,17 +1493,18 @@ int main(int argc, char *argv[])
 			err(1, "fork failed");
 		case 0:
 			close(fds[1]);
-			tdb = tdb_open_ex(argv[1], hashsize[i], tdb_flags[i],
-					  open_flags[i], 0600,
-					  NULL, hash_key);
+			tdb = tdb_open_ex(argv[1], hashsize[i],
+					  tdb_flags[i]|TDB_NOSYNC,
+					  open_flags[i], 0600, NULL, hash_key);
 			if (!tdb)
 				err(1, "Opening tdb %s", argv[1]);
 
 			/* This catches parent exiting. */
 			if (read(fds[0], &c, 1) != 1)
 				exit(1);
-			run_ops(tdb, pipes[i].fd[0], argv[2+i],
-				op[i], 1, num_ops[i]);
+			run_ops(tdb, pipes[i].fd[0], argv+2, op, i, 1,
+				num_ops[i], false);
+			check_deps(argv[2+i], op[i], num_ops[i]);
 			exit(0);
 		default:
 			break;
@@ -1031,6 +1514,8 @@ int main(int argc, char *argv[])
 	/* Let everything settle. */
 	sleep(1);
 
+	printf("Starting run...");
+	fflush(stdout);
 	gettimeofday(&start, NULL);
 	/* Tell them all to go!  Any write of sufficient length will do. */
 	if (write(fds[1], hashsize, i) != i)
@@ -1039,12 +1524,18 @@ int main(int argc, char *argv[])
 	for (i = 0; i < argc - 2; i++) {
 		int status;
 		wait(&status);
-		if (!WIFEXITED(status))
-			errx(1, "Child died with signal %i", WTERMSIG(status));
-		if (WEXITSTATUS(status) != 0)
-			errx(1, "Child died with error code");
+		if (!WIFEXITED(status)) {
+			warnx("Child died with signal %i", WTERMSIG(status));
+			ok = false;
+		} else if (WEXITSTATUS(status) != 0)
+			/* Assume child spat out error. */
+			ok = false;
 	}
+	if (!ok)
+		exit(1);
+
 	gettimeofday(&end, NULL);
+	printf("done\n");
 
 	end.tv_sec -= start.tv_sec;
 	printf("Time replaying: %lu usec\n",