X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Ftools%2Freplay_trace.c;h=67809d82093e57765cd0ac5e12dfc4644b869799;hp=0d53f1fea71cb03ea6266adb0417e561a047cc90;hb=a57ff6cbf6ab6209b201df20e1eecf96f40ae15d;hpb=46fc04374fd1ad3d1f3ff44e39e08cea16688c66

diff --git a/ccan/tdb/tools/replay_trace.c b/ccan/tdb/tools/replay_trace.c
index 0d53f1fe..67809d82 100644
--- a/ccan/tdb/tools/replay_trace.c
+++ b/ccan/tdb/tools/replay_trace.c
@@ -15,6 +15,7 @@
 #include <errno.h>
 #include <signal.h>
 #include <assert.h>
+#include <fcntl.h>
 
 #define STRINGIFY2(x) #x
 #define STRINGIFY(x) STRINGIFY2(x)
@@ -31,6 +32,7 @@ struct pipe {
 	int fd[2];
 };
 static struct pipe *pipes;
+static int backoff_fd = -1;
 
 static void __attribute__((noreturn)) fail(const char *filename,
 					   unsigned int line,
@@ -119,6 +121,7 @@ enum op_type {
 	OP_TDB_TRAVERSE_START,
 	OP_TDB_TRAVERSE_END,
 	OP_TDB_TRAVERSE,
+	OP_TDB_TRAVERSE_END_EARLY,
 	OP_TDB_FIRSTKEY,
 	OP_TDB_NEXTKEY,
 	OP_TDB_FETCH,
@@ -126,8 +129,8 @@ enum op_type {
 };
 
 struct op {
-	unsigned int serial;
-	enum op_type op;
+	unsigned int seqnum;
+	enum op_type type;
 	TDB_DATA key;
 	TDB_DATA data;
 	int ret;
@@ -143,15 +146,20 @@ struct op {
 
 	union {
 		int flag; /* open and store */
-		struct traverse *trav; /* traverse start */
 		struct {  /* append */
 			TDB_DATA pre;
 			TDB_DATA post;
 		} append;
-		unsigned int transaction_end; /* transaction start */
+		/* transaction/traverse start/chainlock */
+		unsigned int group_len;
 	};
 };
 
+struct op_desc {
+	unsigned int file;
+	unsigned int op_num;
+};
+
 static unsigned char hex_char(const char *filename, unsigned int line, char c)
 {
 	c = toupper(c);
@@ -188,13 +196,13 @@ static TDB_DATA make_tdb_data(const void *ctx,
 }
 
 static void add_op(const char *filename, struct op **op, unsigned int i,
-		   unsigned int serial, enum op_type type)
+		   unsigned int seqnum, enum op_type type)
 {
 	struct op *new;
 	*op = talloc_realloc(NULL, *op, struct op, i+1);
 	new = (*op) + i;
-	new->op = type;
-	new->serial = serial;
+	new->type = type;
+	new->seqnum = seqnum;
 	new->ret = 0;
 	new->group_start = 0;
 }
@@ -214,8 +222,7 @@ static void op_add_key(const char *filename,
 		fail(filename, op_num+1, "Expected just a key");
 
 	op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
-	if (op[op_num].op != OP_TDB_TRAVERSE)
-		total_keys++;
+	total_keys++;
 }
 
 static void op_add_key_ret(const char *filename,
@@ -244,7 +251,27 @@ static void op_add_key_data(const char *filename,
 		total_keys++;
 }
 
-/* <serial> tdb_store <rec> <rec> <flag> = <ret> */
+/* We don't record the keys or data for a traverse, as we don't use them. */
+static void op_add_traverse(const char *filename,
+			    struct op op[], unsigned int op_num, char *words[])
+{
+	if (!words[2] || !words[3] || !words[4] || words[5]
+	    || !streq(words[3], "="))
+		fail(filename, op_num+1, "Expected <key> = <data>");
+	op[op_num].key = tdb_null;
+}
+
+/* Full traverse info is useful for debugging, but changing it to
+ * "traversefn" without the data makes the traces *much* smaller! */
+static void op_add_traversefn(const char *filename,
+			    struct op op[], unsigned int op_num, char *words[])
+{
+	if (words[2])
+		fail(filename, op_num+1, "Expected no values");
+	op[op_num].key = tdb_null;
+}
+
+/* <seqnum> tdb_store <rec> <rec> <flag> = <ret> */
 static void op_add_store(const char *filename,
 			 struct op op[], unsigned int op_num, char *words[])
 {
@@ -259,7 +286,7 @@ static void op_add_store(const char *filename,
 	total_keys++;
 }
 
-/* <serial> tdb_append <rec> <rec> = <rec> */
+/* <seqnum> tdb_append <rec> <rec> = <rec> */
 static void op_add_append(const char *filename,
 			  struct op op[], unsigned int op_num, char *words[])
 {
@@ -280,7 +307,7 @@ static void op_add_append(const char *filename,
 	total_keys++;
 }
 
-/* <serial> tdb_get_seqnum = <ret> */
+/* <seqnum> tdb_get_seqnum = <ret> */
 static void op_add_seqnum(const char *filename,
 			  struct op op[], unsigned int op_num, char *words[])
 {
@@ -291,14 +318,15 @@ static void op_add_seqnum(const char *filename,
 	op[op_num].ret = atoi(words[3]);
 }
 
-static void op_add_traverse(const char *filename,
-			    struct op op[], unsigned int op_num, char *words[])
+static void op_add_traverse_start(const char *filename,
+				  struct op op[],
+				  unsigned int op_num, char *words[])
 {
 	if (words[2])
 		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[],
@@ -308,78 +336,97 @@ 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 void op_add_chainlock(const char *filename,
+			     struct op op[], unsigned int op_num, char *words[])
+{
+	if (words[2] == NULL || words[3])
+		fail(filename, op_num+1, "Expected just a key");
+
+	/* A chainlock key isn't a key in the normal sense; it doesn't
+	 * have to be in the db at all.  Also, we don't want to hash this op. */
+	op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
+	op[op_num].key = tdb_null;
+	op[op_num].group_len = 0;
+}
+
+static void op_add_chainlock_ret(const char *filename,
+				 struct op op[], unsigned int op_num,
+				 char *words[])
+{
+	if (!words[2] || !words[3] || !words[4] || words[5]
+	    || !streq(words[3], "="))
+		fail(filename, op_num+1, "Expected <key> = <ret>");
+	op[op_num].ret = atoi(words[4]);
+	op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
+	op[op_num].key = tdb_null;
+	op[op_num].group_len = 0;
+	total_keys++;
+}
+
+static int op_find_start(struct op op[], unsigned int op_num, enum op_type type)
+{
+	unsigned int i;
+
+	for (i = op_num-1; i > 0; i--) {
+		if (op[i].type == type && !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_find_start(op, op_num, OP_TDB_TRANSACTION_START);
+	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;
 }
 
-struct traverse_hash {
-	TDB_DATA key;
-	unsigned int index;
-};
+/* We treat chainlocks a lot like transactions, even though that's overkill */
+static void op_analyze_chainlock(const char *filename,
+				 struct op op[], unsigned int op_num,
+				 char *words[])
+{
+	unsigned int i, start;
 
-/* A traverse is a hash of keys, each one associated with ops. */
-struct traverse {
-	/* How many traversal callouts should I do? */
-	unsigned int num;
+	if (words[2] == NULL || words[3])
+		fail(filename, op_num+1, "Expected just a key");
 
-	/* Where is traversal end op? */
-	unsigned int end;
+	op[op_num].data = make_tdb_data(op, filename, op_num+1, words[2]);
+	op[op_num].key = tdb_null;
+	total_keys++;
 
-	/* For trivial traversals. */
-	struct traverse_hash *hash;
-};
+	start = op_find_start(op, op_num, OP_TDB_CHAINLOCK);
+	if (!start)
+		start = op_find_start(op, op_num, OP_TDB_CHAINLOCK_READ);
+	if (!start)
+		fail(filename, op_num+1, "no initial chainlock found");
 
-/* 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;
+	/* FIXME: We'd have to do something clever to make this work
+	 * vs. deadlock. */
+	if (!key_eq(op[start].data, op[op_num].data))
+		fail(filename, op_num+1, "nested chainlock calls?");
 
-	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;
+	op[start].group_len = op_num - start;
+	for (i = start; i <= op_num; i++)
+		op[i].group_start = start;
 }
 
 static void op_analyze_traverse(const char *filename,
@@ -387,7 +434,6 @@ static void op_analyze_traverse(const char *filename,
 				char *words[])
 {
 	int i, start;
-	struct traverse *trav = talloc(op, struct traverse);
 
 	op[op_num].key = tdb_null;
 
@@ -399,45 +445,18 @@ 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)
-			continue;
-		break;
-	}
-
-	if (i < 0)
+	start = op_find_start(op, op_num, OP_TDB_TRAVERSE_START);
+	if (!start)
+		start = op_find_start(op, op_num, OP_TDB_TRAVERSE_READ_START);
+	if (!start)
 		fail(filename, op_num+1, "no traversal start found");
 
-	start = i;
-	op[start].trav = trav;
+	op[start].group_len = op_num - start;
 
+	/* Don't roll in nested traverse/chainlock */
 	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;
+		if (!op[i].group_start)
+			op[i].group_start = start;
 }
 
 /* Keep -Wmissing-declarations happy: */
@@ -450,10 +469,8 @@ 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;
+	struct op_desc needs;
+	struct op_desc prereq;
 };
 
 static void check_deps(const char *filename, struct op op[], unsigned int num)
@@ -467,33 +484,49 @@ static void check_deps(const char *filename, struct op op[], unsigned int num)
 #endif
 }
 
-static void dump_pre(char *filename[], unsigned int file,
-		     struct op op[], unsigned int i)
+static void dump_pre(char *filename[], struct op *op[],
+		     unsigned int file, unsigned int i)
 {
 	struct depend *dep;
 
-	printf("%s:%u still waiting for:\n", filename[file], i+1);
-	list_for_each(&op[i].pre, dep, pre_list)
-		printf("    %s:%u\n",
-		       filename[dep->satisfies_file], dep->satisfies_opnum+1);
-	check_deps(filename[file], op, i);
+	printf("%s:%u (%u) still waiting for:\n", filename[file], i+1,
+		op[file][i].seqnum);
+	list_for_each(&op[file][i].pre, dep, pre_list)
+		printf("    %s:%u (%u)\n",
+		       filename[dep->prereq.file], dep->prereq.op_num+1,
+		       op[dep->prereq.file][dep->prereq.op_num].seqnum);
+	check_deps(filename[file], op[file], i);
 }
 
 /* We simply read/write pointers, since we all are children. */
-static void do_pre(char *filename[], unsigned int file, int pre_fd,
-		   struct op op[], unsigned int i)
+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[i].pre)) {
+	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
-		alarm(10);
+		if (backoff)
+			alarm(2);
+		else
+			alarm(10);
 		while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
 			if (errno == EINTR) {
-				dump_pre(filename, file, op, i);
+				if (backoff) {
+					struct op_desc desc = { file,i };
+					warnx("%s:%u:avoiding deadlock",
+					      filename[file], i+1);
+					if (write(backoff_fd, &desc,
+						  sizeof(desc)) != sizeof(desc))
+						err(1, "writing backoff_fd");
+					return false;
+				}
+				dump_pre(filename, op, file, i);
 				exit(1);
 			} else
 				errx(1, "Reading from pipe");
@@ -502,28 +535,30 @@ static void do_pre(char *filename[], unsigned int file, int pre_fd,
 
 #if DEBUG_DEPS
 		printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
-		       dep->needs_op, dep->satisfies_file, dep->satisfies_op+1);
+		       dep->needs.op_num+1, filename[dep->prereq.file],
+		       dep->prereq.op_num+1);
 		fflush(stdout);
 #endif
 		/* This could be any op, not just this one. */
 		talloc_free(dep);
 	}
+	return true;
 }
 
-static void do_post(char *filename[], unsigned int file,
-		    const struct op op[], unsigned int i)
+static void do_post(char *filename[], struct op *op[],
+		    unsigned int file, unsigned int i)
 {
 	struct depend *dep;
 
-	list_for_each(&op[i].post, dep, post_list) {
+	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);
+		       filename[dep->needs.file], dep->needs.op_num+1);
 #endif
-		if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
+		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]);
+			    filename[file], i+1, filename[dep->needs.file]);
 	}
 }
 
@@ -535,12 +570,13 @@ static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 static unsigned run_ops(struct tdb_context *tdb,
 			int pre_fd,
 			char *filename[],
+			struct op *op[],
 			unsigned int file,
-			struct op op[],
-			unsigned int start, unsigned int stop);
+			unsigned int start, unsigned int stop,
+			bool backoff);
 
 struct traverse_info {
-	struct op *op;
+	struct op **op;
 	char **filename;
 	unsigned file;
 	int pre_fd;
@@ -548,54 +584,38 @@ struct traverse_info {
 	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->file, tinfo->op, trav->hash[h].index,
-				trav->end);
-			tinfo->i++;
-			return 0;
-		}
-		h = (h + 1) % (trav->num * 2);
-	}
-	fail(tinfo->filename[tinfo->file], 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->file], tinfo->start + 1,
 		     "traverse did not terminate");
 	}
 
-	if (tinfo->op[tinfo->i].op != OP_TDB_TRAVERSE)
+	if (tinfo->op[tinfo->file][tinfo->i].type != 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->file,
-			   tinfo->op, tinfo->i+1, trav->end);
+	tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
+			   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/EARLY. */
+	if (tinfo->op[tinfo->file][tinfo->i].type != OP_TDB_TRAVERSE)
 		return 1;
 
 	return 0;
@@ -607,36 +627,29 @@ static unsigned op_traverse(struct tdb_context *tdb,
 			    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, file, pre_fd,
 				       start, start+1 };
 
-	/* Trivial case. */
-	if (trav->hash) {
-		int ret = traversefn(tdb, trivial_traverse, &tinfo);
-		if (ret != trav->num)
-			fail(filename[file], start+1,
-			     "short traversal %i", ret);
-		return trav->end;
-	}
-
 	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].type == OP_TDB_TRAVERSE
+		    || op[file][tinfo.i].type == OP_TDB_TRAVERSE_END_EARLY)
 			tinfo.i++;
 		else
-			tinfo.i = run_ops(tdb, pre_fd, filename, file, 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;
+	return tinfo.i;
 }
 
 static void break_out(int sig)
@@ -647,8 +660,10 @@ static __attribute__((noinline))
 unsigned run_ops(struct tdb_context *tdb,
 		 int pre_fd,
 		 char *filename[],
+		 struct op *op[],
 		 unsigned int file,
-		 struct op op[], unsigned int start, unsigned int stop)
+		 unsigned int start, unsigned int stop,
+		 bool backoff)
 {
 	unsigned int i;
 	struct sigaction sa;
@@ -658,88 +673,99 @@ unsigned run_ops(struct tdb_context *tdb,
 
 	sigaction(SIGALRM, &sa, NULL);
 	for (i = start; i < stop; i++) {
-		do_pre(filename, file, 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].type) {
 		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);
+			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);
+			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, file,
@@ -750,36 +776,55 @@ unsigned run_ops(struct tdb_context *tdb,
 					tdb_traverse, op, i);
 			break;
 		case OP_TDB_TRAVERSE:
+		case OP_TDB_TRAVERSE_END_EARLY:
 			/* Terminate: we're in a traverse, and we've
 			 * done our ops. */
 			return i;
 		case OP_TDB_TRAVERSE_END:
-			fail(filename[file], 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))
+			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))
+			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))
+			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, file, 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_find_start(op, *num, OP_TDB_TRANSACTION_START);
+
+	if (start) {
+		char *words[] = { "<unknown>", "tdb_close", NULL };
+		add_op(filename, &op, *num, op[start].seqnum,
+		       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,
@@ -812,7 +857,7 @@ static struct op *load_tracefile(const char *filename, unsigned int *num,
 
 		words = strsplit(lines, lines[i], " ", NULL);
 		if (!words[0] || !words[1])
-			fail(filename, i+1, "Expected serial number and op");
+			fail(filename, i+1, "Expected seqnum number and op");
 	       
 		opt = find_keyword(words[1], strlen(words[1]));
 		if (!opt) {
@@ -822,7 +867,8 @@ 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]);
 		}
@@ -835,19 +881,14 @@ 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. */
-struct key_user {
-	unsigned int file;
-	unsigned int op_num;
-};
-
 struct keyinfo {
 	TDB_DATA key;
 	unsigned int num_users;
-	struct key_user *user;
+	struct op_desc *user;
 };
 
 static const TDB_DATA must_not_exist;
@@ -859,7 +900,7 @@ static const TDB_DATA not_exists_or_empty;
  * not exist, otherwise the data it needs. */
 static const TDB_DATA *needs(const struct op *op)
 {
-	switch (op->op) {
+	switch (op->type) {
 	/* FIXME: Pull forward deps, since we can deadlock */
 	case OP_TDB_CHAINLOCK:
 	case OP_TDB_CHAINLOCK_NONBLOCK:
@@ -923,25 +964,82 @@ static const TDB_DATA *needs(const struct op *op)
 		return &must_exist;
 
 	default:
-		errx(1, "Unexpected op %i", op->op);
+		errx(1, "Unexpected op type %i", op->type);
 	}
 	
 }
 
+static bool starts_transaction(const struct op *op)
+{
+	return op->type == OP_TDB_TRANSACTION_START;
+}
+
+static bool in_transaction(const struct op op[], unsigned int i)
+{
+	return op[i].group_start && starts_transaction(&op[op[i].group_start]);
+}
+
+static bool successful_transaction(const struct op *op)
+{
+	return starts_transaction(op)
+		&& op[op->group_len].type == OP_TDB_TRANSACTION_COMMIT;
+}
+
+static bool starts_traverse(const struct op *op)
+{
+	return op->type == OP_TDB_TRAVERSE_START
+		|| op->type == OP_TDB_TRAVERSE_READ_START;
+}
+
+static bool in_traverse(const struct op op[], unsigned int i)
+{
+	return op[i].group_start && starts_traverse(&op[op[i].group_start]);
+}
+
+static bool starts_chainlock(const struct op *op)
+{
+	return op->type == OP_TDB_CHAINLOCK_READ
+		|| op->type == OP_TDB_CHAINLOCK;
+}
+
+static bool in_chainlock(const struct op op[], unsigned int i)
+{
+	return op[i].group_start && starts_chainlock(&op[op[i].group_start]);
+}
+
 /* What's the data after this op?  pre if nothing changed. */
-static const TDB_DATA *gives(const struct op *op, const TDB_DATA *pre)
+static const TDB_DATA *gives(const TDB_DATA *key, const TDB_DATA *pre,
+			     const struct op *op)
 {
+	if (starts_transaction(op) || starts_chainlock(op)) {
+		unsigned int i;
+
+		/* Cancelled transactions don't change anything. */
+		if (op[op->group_len].type == OP_TDB_TRANSACTION_CANCEL)
+			return pre;
+		assert(op[op->group_len].type == OP_TDB_TRANSACTION_COMMIT
+		       || op[op->group_len].type == OP_TDB_CHAINUNLOCK_READ
+		       || op[op->group_len].type == OP_TDB_CHAINUNLOCK);
+
+		for (i = 1; i < op->group_len; i++) {
+			/* This skips nested transactions, too */
+			if (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)
+	if (op->type == OP_TDB_DELETE || op->type == OP_TDB_WIPE_ALL)
 		return &tdb_null;
 
-	if (op->op == OP_TDB_APPEND)
+	if (op->type == OP_TDB_APPEND)
 		return &op->append.post;
 
-	if (op->op == OP_TDB_STORE)
+	if (op->type == OP_TDB_STORE)
 		return &op->data;
 
 	return pre;
@@ -963,11 +1061,6 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 			if (!op[i][j].key.dptr)
 				continue;
 
-			/* We don't wait for traverse keys */
-			/* FIXME: We should, for trivial traversals. */
-			if (op[i][j].op == OP_TDB_TRAVERSE)
-				continue;
-
 			h = hash_key(&op[i][j].key) % (total_keys * 2);
 			while (!key_eq(hash[h].key, op[i][j].key)) {
 				if (!hash[h].key.dptr) {
@@ -982,9 +1075,26 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 				op[i][j].key.dptr = hash[h].key.dptr;
 			}
 			hash[h].user = talloc_realloc(hash, hash[h].user,
-						     struct key_user,
+						     struct op_desc,
 						     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)
+			    || in_chainlock(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++;
 		}
@@ -993,8 +1103,32 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 	return hash;
 }
 
-static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
+static bool satisfies(const TDB_DATA *key, const TDB_DATA *data,
+		      const struct op *op)
 {
+	const TDB_DATA *need = NULL;
+
+	if (starts_transaction(op) || starts_chainlock(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 (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].type != OP_TDB_EXISTS)
+					break;
+			}
+		}
+	} else
+		need = needs(op);
+
 	/* Don't need anything?  Cool. */
 	if (!need)
 		return true;
@@ -1004,18 +1138,13 @@ static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
 	assert(data != &must_not_exist);
 	assert(data != &not_exists_or_empty);
 
-	/* must_not_exist == must_not_exist, must_exist == must_exist, or
-	   not_exists_or_empty == not_exists_or_empty. */
-	if (data->dsize == need->dsize && data->dptr == need->dptr)
-		return true;
-
 	/* Must not exist?  data must not exist. */
 	if (need == &must_not_exist)
-		return data->dptr == NULL;
+		return data == &tdb_null;
 
 	/* Must exist? */
 	if (need == &must_exist)
-		return data->dptr != NULL;
+		return data != &tdb_null;
 
 	/* Either noexist or empty. */
 	if (need == &not_exists_or_empty)
@@ -1025,54 +1154,72 @@ static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
 	return key_eq(*data, *need);
 }
 
-static void move_to_front(struct key_user res[], unsigned int elem)
+static void move_to_front(struct op_desc res[], unsigned off, unsigned elem)
 {
-	if (elem != 0) {
-		struct key_user tmp = res[elem];
-		memmove(res + 1, res, elem*sizeof(res[0]));
-		res[0] = tmp;
+	if (elem != off) {
+		struct op_desc 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 int elem)
+static void restore_to_pos(struct op_desc res[], unsigned off, unsigned elem)
 {
-	if (elem != 0) {
-		struct key_user tmp = res[0];
-		memmove(res, res + 1, elem*sizeof(res[0]));
+	if (elem != off) {
+		struct op_desc 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 num,
-		      const TDB_DATA *data, unsigned num_files)
+		      struct op_desc 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];
 
-	/* Nothing left?  We're sorted. */
-	if (num == 0)
+	/* None left?  We're sorted. */
+	if (off == num)
 		return true;
 
+	/* Does this make sequence number go backwards?  Allow a little fuzz. */
+	if (off > 0) {
+		int seqnum1 = op[res[off-1].file][res[off-1].op_num].seqnum;
+		int seqnum2 = op[res[off].file][res[off].op_num].seqnum;
+
+		if (seqnum1 - seqnum2 > (int)fuzz) {
+#if DEBUG_DEPS
+			printf("Seqnum jump too far (%u -> %u)\n",
+			       seqnum1, seqnum2);
+#endif
+			return false;
+		}
+	}
+
 	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 = 0; i < num && files_done < num_files; i++) {
+	 * we minimize the chance of harm by trying to keep in seqnum 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(data, needs(this_op))) {
-			move_to_front(res, i);
-			if (sort_deps(filename, op, res+1, num-1,
-				      gives(this_op, data), num_files))
+		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, i);
+			restore_to_pos(res, off, i);
 		}
 		done[res[i].file] = true;
 		files_done++;
@@ -1082,7 +1229,7 @@ static bool sort_deps(char *filename[], struct op *op[],
 	return false;
 }
 
-static void check_dep_sorting(struct key_user user[], unsigned num_users,
+static void check_dep_sorting(struct op_desc user[], unsigned num_users,
 			      unsigned num_files)
 {
 #if DEBUG_DEPS
@@ -1097,21 +1244,30 @@ static void check_dep_sorting(struct key_user user[], unsigned num_users,
 #endif
 }
 
-/* All these ops have the same serial number.  Which comes first?
+/* 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
+ * change sequence 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[],
-			struct key_user user[], unsigned num_users,
-			unsigned num_files)
+			const TDB_DATA *key, struct op_desc 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 seqnum 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 (!sort_deps(filename, op, user, num_users, data, num_files))
+	if (fuzz >= 100)
 		fail(filename[user[0].file], user[0].op_num+1,
 		     "Could not resolve inter-dependencies");
 
@@ -1124,25 +1280,39 @@ static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
 	unsigned int h;
 
 	/* Gcc nexted function extension.  How cool is this? */
-	int compare_serial(const void *_a, const void *_b)
+	int compare_seqnum(const void *_a, const void *_b)
 	{
-		const struct key_user *a = _a, *b = _b;
+		const struct op_desc *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;
+		/* Otherwise, arrange by seqnum order. */
+		if (op[a->file][a->op_num].seqnum !=
+		    op[b->file][b->op_num].seqnum)
+			return op[a->file][a->op_num].seqnum
+				- op[b->file][b->op_num].seqnum;
+
+		/* Cancelled transactions are assumed to happen first. */
+		if (starts_transaction(&op[a->file][a->op_num])
+		    && !successful_transaction(&op[a->file][a->op_num]))
+			return -1;
+		if (starts_transaction(&op[b->file][b->op_num])
+		    && !successful_transaction(&op[b->file][b->op_num]))
+			return 1;
+
+		/* No idea. */
+		return 0;
 	}
 
-	/* Now sort into serial order. */
+	/* Now sort into seqnum order. */
 	for (h = 0; h < total_keys * 2; h++) {
-		struct key_user *user = hash[h].user;
+		struct op_desc *user = hash[h].user;
 
-		qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
-		figure_deps(filename, op, user, hash[h].num_users, num);
+		qsort(user, hash[h].num_users, sizeof(user[0]), compare_seqnum);
+		figure_deps(filename, op, &hash[h].key, user, hash[h].num_users,
+			    num);
 	}
 }
 
@@ -1156,134 +1326,99 @@ static int destroy_depend(struct depend *dep)
 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)
+			   const struct op_desc *needs,
+			   const struct op_desc *prereq)
 {
 	struct depend *dep;
-	unsigned int needs_start, sat_start;
 
 	/* We don't depend on ourselves. */
-	if (needs_file == satisfies_file) {
-		assert(satisfies_opnum < needs_opnum);
+	if (needs->file == prereq->file) {
+		assert(prereq->op_num < needs->op_num);
 		return;
 	}
 
 #if DEBUG_DEPS
 	printf("%s:%u: depends on %s:%u\n",
-	       filename[needs_file], needs_opnum+1,
-	       filename[satisfies_file], satisfies_opnum+1);
+	       filename[needs->file], needs->op_num+1,
+	       filename[prereq->file], prereq->op_num+1);
 #endif
 
-	needs_start = op[needs_file][needs_opnum].group_start;
-	sat_start = op[satisfies_file][satisfies_opnum].group_start;
+	dep = talloc(ctx, struct depend);
+	dep->needs = *needs;
+	dep->prereq = *prereq;
 
-	/* 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:
-			needs_opnum = needs_start;
-#ifdef DEBUG_DEPS
-			printf("  -> Back to %u\n", needs_start+1);
-			fflush(stdout);
-#endif
-			break;
-		default:
-			break;
-		}
+#if TRAVERSALS_TAKE_TRANSACTION_LOCK
+	/* If something in a traverse depends on something in another
+	 * traverse/transaction, it creates a dependency between the
+	 * two groups. */
+	if ((in_traverse(op[prereq->file], prereq->op_num)
+	     && (starts_transaction(&op[needs->file][needs->op_num])
+		 || starts_traverse(&op[needs->file][needs->op_num])))
+	    || (in_traverse(op[needs->file], needs->op_num)
+		&& (starts_transaction(&op[prereq->file][prereq->op_num])
+		    || starts_traverse(&op[prereq->file][prereq->op_num])))) {
+		unsigned int start;
+
+		/* We are satisfied by end of group. */
+		start = op[prereq->file][prereq->op_num].group_start;
+		dep->prereq.op_num = start + op[prereq->file][start].group_len;
+		/* And we need that done by start of our group. */
+		dep->needs.op_num = op[needs->file][needs->op_num].group_start;
 	}
 
-	/* 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
+	/* 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->op_num)) {
+		struct op *need = &op[needs->file][needs->op_num];
+		if (op[needs->file][need->group_start].seqnum >
+		    op[prereq->file][prereq->op_num].seqnum) {
+			dep->needs.op_num = need->group_start;
 		}
 	}
+#endif
 
-	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);
+ 	/* If you depend on a transaction or chainlock, you actually
+	 * depend on it ending. */
+ 	if (starts_transaction(&op[prereq->file][dep->prereq.op_num])
+	    || starts_chainlock(&op[prereq->file][dep->prereq.op_num])) {
+ 		dep->prereq.op_num
+			+= op[dep->prereq.file][dep->prereq.op_num].group_len;
+#if DEBUG_DEPS
+		printf("-> Actually end of transaction %s:%u\n",
+		       filename[dep->prereq->file], dep->prereq->op_num+1);
+#endif
+ 	} else
+		/* We should never create a dependency from middle of
+		 * a transaction. */
+ 		assert(!in_transaction(op[prereq->file], dep->prereq.op_num)
+		       || op[prereq->file][dep->prereq.op_num].type
+ 		       == OP_TDB_TRANSACTION_COMMIT
+ 		       || op[prereq->file][dep->prereq.op_num].type
+ 		       == OP_TDB_TRANSACTION_CANCEL);
+
+	list_add(&op[dep->prereq.file][dep->prereq.op_num].post,
+		 &dep->post_list);
+	list_add(&op[dep->needs.file][dep->needs.op_num].pre,
+		 &dep->pre_list);
 	talloc_set_destructor(dep, destroy_depend);
 }
 
-#if TRAVERSALS_TAKE_TRANSACTION_LOCK
-struct traverse_dep {
-	unsigned int file;
-	unsigned int op_num;
-	const struct op *op;
-};
-
-/* Sort by which one runs first. */
-static int compare_traverse_dep(const void *_a, const void *_b)
+static bool changes_db(const TDB_DATA *key, const struct op *op)
 {
-	const struct traverse_dep *a = _a, *b = _b;
-	const struct traverse *trava = a->op->trav, *travb = b->op->trav;
-
-	if (a->op->serial != b->op->serial)
-		return a->op->serial - b->op->serial;
-
-	/* If they have same serial, it means one didn't make any changes.
-	 * Thus sort by end in that case. */
-	return a->op[trava->end - a->op_num].serial
-		- b->op[travb->end - b->op_num].serial;
-}
-
-/* Traversals can deadlock against each other.  Force order. */
-static void make_traverse_depends(char *filename[],
-				  struct op *op[], unsigned int num_ops[],
-				  unsigned int num)
-{
-	unsigned int i, j, num_traversals = 0;
-	struct traverse_dep *dep;
-
-	dep = talloc_array(NULL, struct traverse_dep, 1);
-
-	/* Count them. */
-	for (i = 0; i < num; i++) {
-		for (j = 0; j < num_ops[i]; j++) {
-			if (op[i][j].op == OP_TDB_TRAVERSE_START
-			    || op[i][j].op == OP_TDB_TRAVERSE_READ_START) {
-				dep = talloc_realloc(NULL, dep,
-						     struct traverse_dep,
-						     num_traversals+1);
-				dep[num_traversals].file = i;
-				dep[num_traversals].op_num = j;
-				dep[num_traversals].op = &op[i][j];
-				num_traversals++;
-			}
-		}
-	}
-	qsort(dep, num_traversals, sizeof(dep[0]), compare_traverse_dep);
-	for (i = 1; i < num_traversals; i++) {
-		/* i depends on end of traverse i-1. */
-		add_dependency(NULL, op, filename, dep[i].file, dep[i].op_num,
-			       dep[i-1].file, dep[i-1].op->trav->end);
-	}
-	talloc_free(dep);
-}
-#endif /* TRAVERSALS_TAKE_TRANSACTION_LOCK */
-
-static bool changes_db(const struct op *op)
-{
-	return gives(op, NULL) != NULL;
+	return gives(key, NULL, op) != NULL;
 }
 
 static void depend_on_previous(struct op *op[],
 			       char *filename[],
 			       unsigned int num,
-			       struct key_user user[],
+			       struct op_desc user[],
 			       unsigned int i,
 			       int prev)
 {
@@ -1295,9 +1430,7 @@ static void depend_on_previous(struct op *op[],
 
 	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);
+		add_dependency(NULL, op, filename, &user[i], &user[prev]);
 		return;
 	}
 
@@ -1306,9 +1439,7 @@ static void depend_on_previous(struct op *op[],
 	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);
+			add_dependency(NULL, op, filename, &user[i], &user[j]);
 			deps[user[j].file] = true;
 		}
 	}
@@ -1330,14 +1461,14 @@ static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
 			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;
+				if (!prev[dep->prereq.file]) {
+					prev[dep->prereq.file] = dep;
 					continue;
 				}
-				if (prev[dep->satisfies_file]->satisfies_opnum
-				    > dep->satisfies_opnum) {
-					talloc_free(prev[dep->satisfies_file]);
-					prev[dep->satisfies_file] = dep;
+				if (prev[dep->prereq.file]->prereq.op_num
+				    < dep->prereq.op_num) {
+					talloc_free(prev[dep->prereq.file]);
+					prev[dep->prereq.file] = dep;
 				} else
 					talloc_free(dep);
 			}
@@ -1354,17 +1485,130 @@ static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
 			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)
+				if (deps[dep->prereq.file]
+				    >= (int)dep->prereq.op_num)
 					talloc_free(dep);
 				else
-					deps[dep->satisfies_file]
-						= dep->satisfies_opnum;
+					deps[dep->prereq.file]
+						= dep->prereq.op_num;
+			}
+		}
+	}
+}
+
+#if TRAVERSALS_TAKE_TRANSACTION_LOCK
+/* Force an order among the traversals, so they don't deadlock (as much) */
+static void make_traverse_depends(char *filename[],
+				  struct op *op[], unsigned int num_ops[],
+				  unsigned int num)
+{
+	unsigned int i, num_traversals = 0;
+	int j;
+	struct op_desc *desc;
+
+	/* Sort by which one runs first. */
+	int compare_traverse_desc(const void *_a, const void *_b)
+	{
+		const struct op_desc *da = _a, *db = _b;
+		const struct op *a = &op[da->file][da->op_num],
+			*b = &op[db->file][db->op_num];
+
+		if (a->seqnum != b->seqnum)
+			return a->seqnum - b->seqnum;
+
+		/* If they have same seqnum, it means one didn't make any
+		 * changes.  Thus sort by end in that case. */
+		return a[a->group_len].seqnum - b[b->group_len].seqnum;
+	}
+
+	desc = talloc_array(NULL, struct op_desc, 1);
+
+	/* Count them. */
+	for (i = 0; i < num; i++) {
+		for (j = 1; j < num_ops[i]; j++) {
+ 			/* Traverse start (ignore those in
+			 * transactions; they're already covered by
+			 * transaction dependencies). */
+			if (starts_traverse(&op[i][j])
+			    && !in_transaction(op[i], j)) {
+				desc = talloc_realloc(NULL, desc,
+						      struct op_desc,
+						      num_traversals+1);
+				desc[num_traversals].file = i;
+				desc[num_traversals].op_num = j;
+				num_traversals++;
 			}
 		}
 	}
+	qsort(desc, num_traversals, sizeof(desc[0]), compare_traverse_desc);
+
+	for (i = 1; i < num_traversals; i++) {
+		const struct op *prev = &op[desc[i-1].file][desc[i-1].op_num];
+		const struct op *curr = &op[desc[i].file][desc[i].op_num];
+
+		/* Read traverses don't depend on each other (read lock). */
+		if (prev->type == OP_TDB_TRAVERSE_READ_START
+		    && curr->type == OP_TDB_TRAVERSE_READ_START)
+			continue;
+
+		/* Only make dependency if it's clear. */
+		if (compare_traverse_desc(&desc[i], &desc[i-1])) {
+			/* i depends on end of traverse i-1. */
+			struct op_desc end = desc[i-1];
+			end.op_num += prev->group_len;
+			add_dependency(NULL, op, filename, &desc[i], &end);
+		}
+	}
+	talloc_free(desc);
 }
 
+static void set_nonblock(int fd)
+{
+	if (fcntl(fd, F_SETFL, fcntl(fd, F_GETFL)|O_NONBLOCK) != 0)
+		err(1, "Setting pipe nonblocking");
+}
+
+static bool handle_backoff(struct op *op[], int fd)
+{
+	struct op_desc desc;
+	bool handled = false;
+
+	/* Sloppy coding: we assume PIPEBUF never fills. */
+	while (read(fd, &desc, sizeof(desc)) != -1) {
+		unsigned int i;
+		handled = true;
+		for (i = desc.op_num; i > 0; i--) {
+			if (op[desc.file][i].type == OP_TDB_TRAVERSE) {
+				/* We insert a fake end here. */
+				op[desc.file][i].type
+					= OP_TDB_TRAVERSE_END_EARLY;
+				break;
+			} else if (starts_traverse(&op[desc.file][i])) {
+				unsigned int start = i;
+				struct op tmp = op[desc.file][i];
+				/* Move the ops outside traverse. */
+				memmove(&op[desc.file][i],
+					&op[desc.file][i+1],
+					(desc.op_num-i-1) * sizeof(op[0][0]));
+				op[desc.file][desc.op_num] = tmp;
+				while (op[desc.file][i].group_start == start) {
+					op[desc.file][i++].group_start
+						= desc.op_num;
+				}
+				break;
+			}
+		}
+	}
+	return handled;
+}
+
+#else /* !TRAVERSALS_TAKE_TRANSACTION_LOCK */
+static bool handle_backoff(struct op *op[], int fd)
+{
+	return false;
+}
+#endif
+
 static void derive_dependencies(char *filename[],
 				struct op *op[], unsigned int num_ops[],
 				unsigned int num)
@@ -1375,7 +1619,7 @@ static void derive_dependencies(char *filename[],
 	/* Create hash table for faster key lookup. */
 	hash = hash_ops(op, num_ops, num);
 
-	/* Sort them by serial number. */
+	/* Sort them by sequence number. */
 	sort_ops(hash, filename, op, num);
 
 	/* Create dependencies back to the last change, rather than
@@ -1393,17 +1637,15 @@ static void derive_dependencies(char *filename[],
 			continue;
 
 		for (i = 0; i < hash[h].num_users; i++) {
-			if (changes_db(&op[hash[h].user[i].file]
+			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);
+					       &hash[h].user[i],
+					       &hash[h].user[prev]);
 		}
 	}
 
@@ -1414,72 +1656,38 @@ static void derive_dependencies(char *filename[],
 	optimize_dependencies(op, num_ops, num);
 }
 
-int main(int argc, char *argv[])
+static struct timeval run_test(char *argv[],
+			       unsigned int num_ops[],
+			       unsigned int hashsize[],
+			       unsigned int tdb_flags[],
+			       unsigned int open_flags[],
+			       struct op *op[],
+			       int fds[2])
 {
-	struct timeval start, end;
-	unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
-	struct op *op[argc];
-	int fds[2];
-	char c;
+	unsigned int i;
+	struct timeval start, end, diff;
 	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);
-		printf("Single threaded run...");
-		fflush(stdout);
-
-		run_ops(tdb, pipes[0].fd[0], argv+2, 0, op[0], 1, num_ops[0]);
-		check_deps(argv[2], op[0], num_ops[0]);
-
-		printf("done\n");
-		exit(0);
-	}
-
-	if (pipe(fds) != 0)
-		err(1, "creating pipe");
-
-	for (i = 0; i < argc - 2; i++) {
+	for (i = 0; argv[i+2]; i++) {
 		struct tdb_context *tdb;
+		char c;
 
 		switch (fork()) {
 		case -1:
 			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:
@@ -1497,7 +1705,7 @@ int main(int argc, char *argv[])
 	if (write(fds[1], hashsize, i) != i)
 		err(1, "Writing to wakeup pipe");
 
-	for (i = 0; i < argc - 2; i++) {
+	for (i = 0; argv[i + 2]; i++) {
 		int status;
 		wait(&status);
 		if (!WIFEXITED(status)) {
@@ -1513,9 +1721,75 @@ int main(int argc, char *argv[])
 	gettimeofday(&end, NULL);
 	printf("done\n");
 
-	end.tv_sec -= start.tv_sec;
+	if (end.tv_usec < start.tv_usec) {
+		end.tv_usec += 1000000;
+		end.tv_sec--;
+	}
+	diff.tv_sec = end.tv_sec - start.tv_sec;
+	diff.tv_usec = end.tv_usec - start.tv_usec;
+	return diff;
+}
+
+int main(int argc, char *argv[])
+{
+	struct timeval diff;
+	unsigned int i, num_ops[argc], hashsize[argc], tdb_flags[argc], open_flags[argc];
+	struct op *op[argc];
+	int fds[2];
+
+	if (argc < 3)
+		errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
+
+	pipes = talloc_array(NULL, struct pipe, argc - 1);
+	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]|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);
+	}
+
+	if (pipe(fds) != 0)
+		err(1, "creating pipe");
+
+#if TRAVERSALS_TAKE_TRANSACTION_LOCK
+	if (pipe(pipes[argc-2].fd) != 0)
+		err(1, "creating pipe");
+	backoff_fd = pipes[argc-2].fd[1];
+	set_nonblock(pipes[argc-2].fd[1]);
+	set_nonblock(pipes[argc-2].fd[0]);
+#endif
+
+	do {
+		diff = run_test(argv, num_ops, hashsize, tdb_flags, open_flags,
+				op, fds);
+	} while (handle_backoff(op, pipes[argc-2].fd[0]));
+
 	printf("Time replaying: %lu usec\n",
-	       end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));
+	       diff.tv_sec * 1000000UL + diff.tv_usec);
 	
 	exit(0);
 }