]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb/tools/replay_trace.c
compiler: use everywhere.
[ccan] / ccan / tdb / tools / replay_trace.c
index 0d53f1fea71cb03ea6266adb0417e561a047cc90..b42fb7b1b5e61e43858679a0b21e64f2ac66122f 100644 (file)
 #include <errno.h>
 #include <signal.h>
 #include <assert.h>
+#include <fcntl.h>
 
 #define STRINGIFY2(x) #x
 #define STRINGIFY(x) STRINGIFY2(x)
 
+static bool quiet = false;
+
 /* Avoid mod by zero */
 static unsigned int total_keys = 1;
 
+/* All the wipe_all ops. */
+static struct op_desc *wipe_alls = NULL;
+static unsigned int num_wipe_alls = 0;
+
 /* #define DEBUG_DEPS 1 */
 
 /* Traversals block transactions in the current implementation. */
@@ -31,6 +38,7 @@ 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,
@@ -114,20 +122,23 @@ enum op_type {
        OP_TDB_WIPE_ALL,
        OP_TDB_TRANSACTION_START,
        OP_TDB_TRANSACTION_CANCEL,
+       OP_TDB_TRANSACTION_PREPARE_COMMIT,
        OP_TDB_TRANSACTION_COMMIT,
        OP_TDB_TRAVERSE_READ_START,
        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,
        OP_TDB_DELETE,
+       OP_TDB_REPACK,
 };
 
 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 +154,20 @@ 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);
@@ -188,90 +204,111 @@ 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;
 }
 
-static void op_add_nothing(const char *filename,
-                          struct op op[], unsigned int op_num, char *words[])
+static void op_add_nothing(char *filename[], struct op op[],
+                          unsigned file, unsigned op_num, char *words[])
 {
        if (words[2])
-               fail(filename, op_num+1, "Expected no arguments");
+               fail(filename[file], op_num+1, "Expected no arguments");
        op[op_num].key = tdb_null;
 }
 
-static void op_add_key(const char *filename,
-                      struct op op[], unsigned int op_num, char *words[])
+static void op_add_key(char *filename[], struct op op[],
+                      unsigned file, unsigned op_num, char *words[])
 {
        if (words[2] == NULL || words[3])
-               fail(filename, op_num+1, "Expected just a key");
+               fail(filename[file], 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++;
+       op[op_num].key = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       total_keys++;
 }
 
-static void op_add_key_ret(const char *filename,
-                          struct op op[], unsigned int op_num, char *words[])
+static void op_add_key_ret(char *filename[], struct op op[],
+                          unsigned file, unsigned op_num, char *words[])
 {
        if (!words[2] || !words[3] || !words[4] || words[5]
            || !streq(words[3], "="))
-               fail(filename, op_num+1, "Expected <key> = <ret>");
+               fail(filename[file], op_num+1, "Expected <key> = <ret>");
        op[op_num].ret = atoi(words[4]);
-       op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
+       op[op_num].key = make_tdb_data(op, filename[file], op_num+1, words[2]);
        /* May only be a unique key if it fails */
        if (op[op_num].ret != 0)
                total_keys++;
 }
 
-static void op_add_key_data(const char *filename,
-                           struct op op[], unsigned int op_num, char *words[])
+static void op_add_key_data(char *filename[], struct op op[],
+                           unsigned file, unsigned 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 = make_tdb_data(op, filename, op_num+1, words[2]);
-       op[op_num].data = make_tdb_data(op, filename, op_num+1, words[4]);
-       /* May only be a unique key if it fails */
+               fail(filename[file], op_num+1, "Expected <key> = <data>");
+       op[op_num].key = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       op[op_num].data = make_tdb_data(op, filename[file], op_num+1, words[4]);
+       /* Likely only be a unique key if it fails */
        if (!op[op_num].data.dptr)
                total_keys++;
+       else if (random() % 2)
+               total_keys++;
+}
+
+/* We don't record the keys or data for a traverse, as we don't use them. */
+static void op_add_traverse(char *filename[], struct op op[],
+                           unsigned file, unsigned op_num, char *words[])
+{
+       if (!words[2] || !words[3] || !words[4] || words[5]
+           || !streq(words[3], "="))
+               fail(filename[file], op_num+1, "Expected <key> = <data>");
+       op[op_num].key = tdb_null;
 }
 
-/* <serial> tdb_store <rec> <rec> <flag> = <ret> */
-static void op_add_store(const char *filename,
-                        struct op op[], unsigned int op_num, char *words[])
+/* 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(char *filename[], struct op op[],
+                           unsigned file, unsigned op_num, char *words[])
+{
+       if (words[2])
+               fail(filename[file], op_num+1, "Expected no values");
+       op[op_num].key = tdb_null;
+}
+
+/* <seqnum> tdb_store <rec> <rec> <flag> = <ret> */
+static void op_add_store(char *filename[], struct op op[],
+                        unsigned file, unsigned op_num, char *words[])
 {
        if (!words[2] || !words[3] || !words[4] || !words[5] || !words[6]
            || words[7] || !streq(words[5], "="))
-               fail(filename, op_num+1, "Expect <key> <data> <flag> = <ret>");
+               fail(filename[file], op_num+1, "Expect <key> <data> <flag> = <ret>");
 
        op[op_num].flag = strtoul(words[4], NULL, 0);
        op[op_num].ret = atoi(words[6]);
-       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].key = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       op[op_num].data = make_tdb_data(op, filename[file], op_num+1, words[3]);
        total_keys++;
 }
 
-/* <serial> tdb_append <rec> <rec> = <rec> */
-static void op_add_append(const char *filename,
-                         struct op op[], unsigned int op_num, char *words[])
+/* <seqnum> tdb_append <rec> <rec> = <rec> */
+static void op_add_append(char *filename[], struct op op[],
+                         unsigned file, unsigned op_num, char *words[])
 {
        if (!words[2] || !words[3] || !words[4] || !words[5] || words[6]
            || !streq(words[4], "="))
-               fail(filename, op_num+1, "Expect <key> <data> = <rec>");
+               fail(filename[file], op_num+1, "Expect <key> <data> = <rec>");
 
-       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].key = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       op[op_num].data = make_tdb_data(op, filename[file], op_num+1, words[3]);
 
        op[op_num].append.post
-               = make_tdb_data(op, filename, op_num+1, words[5]);
+               = make_tdb_data(op, filename[file], 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;
@@ -280,164 +317,165 @@ static void op_add_append(const char *filename,
        total_keys++;
 }
 
-/* <serial> tdb_get_seqnum = <ret> */
-static void op_add_seqnum(const char *filename,
-                         struct op op[], unsigned int op_num, char *words[])
+/* <seqnum> tdb_get_seqnum = <ret> */
+static void op_add_seqnum(char *filename[], struct op op[],
+                         unsigned file, unsigned op_num, char *words[])
 {
        if (!words[2] || !words[3] || words[4] || !streq(words[2], "="))
-               fail(filename, op_num+1, "Expect = <ret>");
+               fail(filename[file], op_num+1, "Expect = <ret>");
 
        op[op_num].key = tdb_null;
        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(char *filename[], struct op op[],
+                                 unsigned file, unsigned op_num, char *words[])
 {
        if (words[2])
-               fail(filename, op_num+1, "Expect no arguments");
+               fail(filename[file], 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[],
-                              unsigned int op_num, char *words[])
+static void op_add_transaction(char *filename[], struct op op[],
+                              unsigned file, unsigned op_num, char *words[])
 {
        if (words[2])
-               fail(filename, op_num+1, "Expect no arguments");
+               fail(filename[file], 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_analyze_transaction(const char *filename,
-                                  struct op op[], unsigned int op_num,
-                                  char *words[])
+static void op_add_chainlock(char *filename[], struct op op[],
+                            unsigned file, unsigned op_num, char *words[])
 {
-       int i, start;
+       if (words[2] == NULL || words[3])
+               fail(filename[file], 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[file], op_num+1, words[2]);
        op[op_num].key = tdb_null;
+       op[op_num].group_len = 0;
+}
+
+static void op_add_chainlock_ret(char *filename[], struct op op[],
+                                unsigned file, unsigned op_num, char *words[])
+{
+       if (!words[2] || !words[3] || !words[4] || words[5]
+           || !streq(words[3], "="))
+               fail(filename[file], op_num+1, "Expected <key> = <ret>");
+       op[op_num].ret = atoi(words[4]);
+       op[op_num].data = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       op[op_num].key = tdb_null;
+       op[op_num].group_len = 0;
+       total_keys++;
+}
 
+static void op_add_wipe_all(char *filename[], struct op op[],
+                           unsigned file, unsigned op_num, char *words[])
+{
        if (words[2])
-               fail(filename, op_num+1, "Expect no arguments");
+               fail(filename[file], op_num+1, "Expected no arguments");
+       op[op_num].key = tdb_null;
+       wipe_alls = talloc_realloc(NULL, wipe_alls, struct op_desc,
+                                  num_wipe_alls+1);
+       wipe_alls[num_wipe_alls].file = file;
+       wipe_alls[num_wipe_alls].op_num = op_num;
+       num_wipe_alls++;
+}
 
-       for (i = op_num-1; i >= 0; i--) {
-               if (op[i].op == OP_TDB_TRANSACTION_START &&
-                   !op[i].transaction_end)
-                       break;
+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;
+}
 
-       if (i < 0)
-               fail(filename, op_num+1, "no transaction start found");
+static void op_analyze_transaction(char *filename[], struct op op[],
+                                  unsigned file, unsigned op_num,
+                                  char *words[])
+{
+       unsigned int start, i;
+
+       op[op_num].key = tdb_null;
+
+       if (words[2])
+               fail(filename[file], op_num+1, "Expect no arguments");
 
-       start = i;
-       op[start].transaction_end = op_num;
+       start = op_find_start(op, op_num, OP_TDB_TRANSACTION_START);
+       if (!start)
+               fail(filename[file], op_num+1, "no transaction start found");
+
+       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;
-};
-
-/* A traverse is a hash of keys, each one associated with ops. */
-struct traverse {
-       /* How many traversal callouts should I do? */
-       unsigned int num;
+/* We treat chainlocks a lot like transactions, even though that's overkill */
+static void op_analyze_chainlock(char *filename[], struct op op[],
+                                unsigned file, unsigned op_num, char *words[])
+{
+       unsigned int i, start;
 
-       /* Where is traversal end op? */
-       unsigned int end;
+       if (words[2] == NULL || words[3])
+               fail(filename[file], op_num+1, "Expected just a key");
 
-       /* For trivial traversals. */
-       struct traverse_hash *hash;
-};
+       op[op_num].data = make_tdb_data(op, filename[file], op_num+1, words[2]);
+       op[op_num].key = tdb_null;
+       total_keys++;
 
-/* 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;
+       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[file], op_num+1, "no initial chainlock found");
 
-       if (op[end].ret != 0)
-               return false;
+       /* 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[file], op_num+1, "nested chainlock calls?");
 
-       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,
-                               struct op op[], unsigned int op_num,
-                               char *words[])
+static void op_analyze_traverse(char *filename[], struct op op[],
+                               unsigned file, unsigned op_num, char *words[])
 {
        int i, start;
-       struct traverse *trav = talloc(op, struct traverse);
 
        op[op_num].key = tdb_null;
 
        /* = %u means traverse function terminated. */
        if (words[2]) {
                if (!streq(words[2], "=") || !words[3] || words[4])
-                       fail(filename, op_num+1, "expect = <num>");
+                       fail(filename[file], op_num+1, "expect = <num>");
                op[op_num].ret = atoi(words[3]);
        } 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)
-               fail(filename, op_num+1, "no traversal start found");
+       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[file], 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 +488,8 @@ 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 +503,51 @@ 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);
+       if (!quiet) {
+               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 +556,30 @@ 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 +591,13 @@ 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 +605,38 @@ 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 +648,29 @@ 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 +681,10 @@ 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 +694,103 @@ 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_PREPARE_COMMIT:
+                       try(tdb_transaction_prepare_commit(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,
@@ -750,37 +801,63 @@ 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;
+               case OP_TDB_REPACK:
+                       /* We do nothing here: the transaction and traverse are
+                        * traced.  It's in the trace to mark it, since it
+                        * may become unnecessary in future. */
                        break;
                }
-               do_post(filename, file, op, i);
+               do_post(filename, op, file, i);
        }
        return i;
 }
 
-static struct op *load_tracefile(const char *filename, unsigned int *num,
+/* tdbtorture, in particular, can do a tdb_close with a transaction in
+ * progress. */
+static struct op *maybe_cancel_transaction(char *filename[], unsigned int file,
+                                          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[file], &op, *num, op[start].seqnum,
+                      OP_TDB_TRANSACTION_CANCEL);
+               op_analyze_transaction(filename, op, file, *num, words);
+               (*num)++;
+       }
+       return op;
+}
+
+static struct op *load_tracefile(char *filename[],
+                                unsigned int file,
+                                unsigned int *num,
                                 unsigned int *hashsize,
                                 unsigned int *tdb_flags,
                                 unsigned int *open_flags)
@@ -789,19 +866,19 @@ static struct op *load_tracefile(const char *filename, unsigned int *num,
        struct op *op = talloc_array(NULL, struct op, 1);
        char **words;
        char **lines;
-       char *file;
+       char *contents;
 
-       file = grab_file(NULL, filename, NULL);
-       if (!file)
-               err(1, "Reading %s", filename);
+       contents = grab_file(NULL, filename[file], NULL);
+       if (!contents)
+               err(1, "Reading %s", filename[file]);
 
-       lines = strsplit(file, file, "\n", NULL);
+       lines = strsplit(contents, contents, "\n", NULL);
        if (!lines[0])
-               errx(1, "%s is empty", filename);
+               errx(1, "%s is empty", filename[file]);
 
        words = strsplit(lines, lines[0], " ", NULL);
        if (!streq(words[1], "tdb_open"))
-               fail(filename, 1, "does not start with tdb_open");
+               fail(filename[file], 1, "does not start with tdb_open");
 
        *hashsize = atoi(words[2]);
        *tdb_flags = strtoul(words[3], NULL, 0);
@@ -812,44 +889,83 @@ 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[file], i+1,
+                            "Expected seqnum number and op");
               
                opt = find_keyword(words[1], strlen(words[1]));
                if (!opt) {
                        if (streq(words[1], "tdb_close")) {
                                if (lines[i+1])
-                                       fail(filename, i+2,
+                                       fail(filename[file], i+2,
                                             "lines after tdb_close");
                                *num = i;
                                talloc_free(lines);
-                               return op;
+                               return maybe_cancel_transaction(filename, file,
+                                                               op, num);
                        }
-                       fail(filename, i+1, "Unknown operation '%s'", words[1]);
+                       fail(filename[file], i+1,
+                            "Unknown operation '%s'", words[1]);
                }
 
-               add_op(filename, &op, i, atoi(words[0]), opt->type);
-               opt->enhance_op(filename, op, i, words);
+               add_op(filename[file], &op, i, atoi(words[0]), opt->type);
+               opt->enhance_op(filename, op, file, i, words);
        }
 
-       fprintf(stderr, "%s:%u:last operation is not tdb_close: incomplete?",
-             filename, i);
-       talloc_free(lines);
+       if (!quiet)
+               fprintf(stderr,
+                       "%s:%u:last operation is not tdb_close: incomplete?",
+                       filename[file], i);
+
+       talloc_free(contents);
        *num = i - 1;
-       return op;
+       return maybe_cancel_transaction(filename, file, 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 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]);
+}
+
 static const TDB_DATA must_not_exist;
 static const TDB_DATA must_exist;
 static const TDB_DATA not_exists_or_empty;
@@ -857,9 +973,32 @@ 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)
+static const TDB_DATA *needs(const TDB_DATA *key, const struct op *op)
 {
-       switch (op->op) {
+       /* Look through for an op in this transaction which needs this key. */
+       if (starts_transaction(op) || starts_chainlock(op)) {
+               unsigned int i;
+               const TDB_DATA *need = NULL;
+
+               for (i = 1; i < op->group_len; i++) {
+                       if (key_eq(op[i].key, *key)
+                           || op[i].type == OP_TDB_WIPE_ALL) {
+                               need = needs(key, &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;
+                       }
+               }
+
+               return need;
+       }
+
+       switch (op->type) {
        /* FIXME: Pull forward deps, since we can deadlock */
        case OP_TDB_CHAINLOCK:
        case OP_TDB_CHAINLOCK_NONBLOCK:
@@ -923,30 +1062,79 @@ 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);
        }
        
 }
 
 /* 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)
+                           || op[i].type == OP_TDB_WIPE_ALL)
+                               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;
 }
 
+static void add_hash_user(struct keyinfo *hash,
+                         unsigned int h,
+                         struct op *op[],
+                         unsigned int file,
+                         unsigned int op_num)
+{
+       hash[h].user = talloc_realloc(hash, hash[h].user,
+                                     struct op_desc, hash[h].num_users+1);
+
+       /* If it's in a transaction, it's the transaction which
+        * matters from an analysis POV. */
+       if (in_transaction(op[file], op_num)
+           || in_chainlock(op[file], op_num)) {
+               unsigned i;
+
+               op_num = op[file][op_num].group_start;
+
+               /* Don't include twice. */
+               for (i = 0; i < hash[h].num_users; i++) {
+                       if (hash[h].user[i].file == file
+                           && hash[h].user[i].op_num == op_num)
+                               return;
+               }
+       }
+       hash[h].user[hash[h].num_users].op_num = op_num;
+       hash[h].user[hash[h].num_users].file = file;
+       hash[h].num_users++;
+}
+
 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
                                unsigned int num)
 {
@@ -963,11 +1151,6 @@ 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) {
@@ -981,20 +1164,29 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
                                talloc_free(op[i][j].key.dptr);
                                op[i][j].key.dptr = hash[h].key.dptr;
                        }
-                       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;
-                       hash[h].user[hash[h].num_users].file = i;
-                       hash[h].num_users++;
+
+                       add_hash_user(hash, h, op, i, j);
                }
        }
 
+       /* Any wipe all entries need adding to all hash entries. */
+       for (h = 0; h < total_keys*2; h++) {
+               if (!hash[h].num_users)
+                       continue;
+
+               for (i = 0; i < num_wipe_alls; i++)
+                       add_hash_user(hash, h, op,
+                                     wipe_alls[i].file, wipe_alls[i].op_num);
+       }
+
        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 = needs(key, op);
+
        /* Don't need anything?  Cool. */
        if (!need)
                return true;
@@ -1004,18 +1196,13 @@ 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 +1212,72 @@ 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 +1287,7 @@ 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
@@ -1097,52 +1302,109 @@ 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)
+static bool figure_deps(char *filename[], struct op *op[],
+                       const TDB_DATA *key, const TDB_DATA *data,
+                       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;
 
-       if (!sort_deps(filename, op, user, num_users, data, num_files))
-               fail(filename[user[0].file], user[0].op_num+1,
-                    "Could not resolve inter-dependencies");
+       /* 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 (fuzz >= 100)
+               return false;
 
        check_dep_sorting(user, num_users, num_files);
+       return true;
 }
 
-static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
+/* We're having trouble sorting out dependencies for this key.  Assume that it's
+ * a pre-existing record in the db, so determine a likely value. */
+static const TDB_DATA *preexisting_data(char *filename[], struct op *op[],
+                                       const TDB_DATA *key,
+                                       struct op_desc *user,
+                                       unsigned int num_users)
+{
+       unsigned int i;
+       const TDB_DATA *data;
+
+       for (i = 0; i < num_users; i++) {
+               data = needs(key, &op[user->file][user->op_num]);
+               if (data && data != &must_not_exist) {
+                       if (!quiet)
+                               printf("%s:%u: needs pre-existing record\n",
+                                      filename[user->file], user->op_num+1);
+                       return data;
+               }
+       }
+       return &tdb_null;
+}
+
+static void sort_ops(struct tdb_context *tdb,
+                    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)
+       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;
-
-               qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
-               figure_deps(filename, op, user, hash[h].num_users, num);
+               struct op_desc *user = hash[h].user;
+
+               qsort(user, hash[h].num_users, sizeof(user[0]), compare_seqnum);
+               if (!figure_deps(filename, op, &hash[h].key, &tdb_null, user,
+                                hash[h].num_users, num)) {
+                       const TDB_DATA *data;
+
+                       data = preexisting_data(filename, op, &hash[h].key,
+                                               user, hash[h].num_users);
+                       /* Give the first op what it wants: does that help? */
+                       if (!figure_deps(filename, op, &hash[h].key, data, user,
+                                        hash[h].num_users, num))
+                               fail(filename[user[0].file], user[0].op_num+1,
+                                    "Could not resolve inter-dependencies");
+                       if (tdb_store(tdb, hash[h].key, *data, TDB_INSERT) != 0)
+                               errx(1, "Could not store initial value");
+               }
        }
 }
 
@@ -1156,134 +1418,99 @@ 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 +1522,7 @@ 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 +1531,7 @@ 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 +1553,14 @@ 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);
                        }
@@ -1354,18 +1577,132 @@ 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[],
+static void derive_dependencies(struct tdb_context *tdb,
+                               char *filename[],
                                struct op *op[], unsigned int num_ops[],
                                unsigned int num)
 {
@@ -1375,8 +1712,8 @@ 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_ops(hash, filename, op, num);
+       /* Sort them by sequence number. */
+       sort_ops(tdb, hash, filename, op, num);
 
        /* Create dependencies back to the last change, rather than
         * creating false dependencies by naively making each one
@@ -1393,17 +1730,15 @@ 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 +1749,37 @@ 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(argv[1], hashsize[i],
+                                      tdb_flags[i], open_flags[i], 0600);
                        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:
@@ -1490,14 +1790,15 @@ int main(int argc, char *argv[])
        /* Let everything settle. */
        sleep(1);
 
-       printf("Starting run...");
+       if (!quiet)
+               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)
                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)) {
@@ -1511,11 +1812,144 @@ int main(int argc, char *argv[])
                exit(1);
 
        gettimeofday(&end, NULL);
-       printf("done\n");
+       if (!quiet)
+               printf("done\n");
+
+       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;
+}
+
+static void init_tdb(struct tdb_context *master_tdb,
+                    const char *name, unsigned int hashsize)
+{
+       TDB_DATA key, data;
+       struct tdb_context *tdb;
+
+       tdb = tdb_open(name, hashsize, TDB_CLEAR_IF_FIRST|TDB_NOSYNC,
+                      O_CREAT|O_TRUNC|O_RDWR, 0600);
+       if (!tdb)
+               errx(1, "opening tdb %s", name);
+
+       for (key = tdb_firstkey(master_tdb);
+            key.dptr;
+            key = tdb_nextkey(master_tdb, key)) {
+               data = tdb_fetch(master_tdb, key);
+               if (tdb_store(tdb, key, data, TDB_INSERT) != 0)
+                       errx(1, "Failed to store initial key");
+       }
+       tdb_close(tdb);
+}
+
+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];
+       struct tdb_context *master;
+       unsigned int runs = 1;
+
+       if (argc < 3)
+               errx(1, "Usage: %s [--quiet] [-n <number>] <tdbfile> <tracefile>...", argv[0]);
+
+       if (streq(argv[1], "--quiet")) {
+               quiet = true;
+               argv++;
+               argc--;
+       }
+       if (streq(argv[1], "-n")) {
+               runs = atoi(argv[2]);
+               argv += 2;
+               argc -= 2;
+       }
+
+       pipes = talloc_array(NULL, struct pipe, argc - 1);
+       for (i = 0; i < argc - 2; i++) {
+               if (!quiet)
+                       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");
+               /* Don't truncate, or clear if first: we do that. */
+               open_flags[i] &= ~(O_TRUNC);
+               tdb_flags[i] &= ~(TDB_CLEAR_IF_FIRST);
+               /* Open NOSYNC, to save time. */
+               tdb_flags[i] |= TDB_NOSYNC;
+               if (!quiet)
+                       printf("done\n");
+       }
+
+       /* Dependency may figure we need to create seed records. */
+       master = tdb_open(NULL, 0, TDB_INTERNAL, O_RDWR, 0);
+       if (!quiet) {
+               printf("Calculating inter-dependencies...");
+               fflush(stdout);
+       }
+       derive_dependencies(master, argv+2, op, num_ops, i);
+       if (!quiet)
+               printf("done\n");
+
+       for (i = 0; i < runs; i++) {
+               init_tdb(master, argv[1], hashsize[0]);
+
+               /* Don't fork for single arg case: simple debugging. */
+               if (argc == 3) {
+                       struct timeval start, end;
+                       struct tdb_context *tdb;
+
+                       tdb = tdb_open(argv[1], hashsize[0], tdb_flags[0],
+                                      open_flags[0], 0600);
+                       if (!quiet) {
+                               printf("Single threaded run...");
+                               fflush(stdout);
+                       }
+                       gettimeofday(&start, NULL);
+
+                       run_ops(tdb, pipes[0].fd[0], argv+2, op, 0, 1,
+                               num_ops[0], false);
+                       gettimeofday(&end, NULL);
+                       if (!quiet)
+                               printf("done\n");
+                       tdb_close(tdb);
+
+                       check_deps(argv[2], op[0], num_ops[0]);
+                       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;
+                       goto print_time;
+               }
+
+               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]));
+
+       print_time:
+               if (!quiet)
+                       printf("Time replaying: ");
+               printf("%lu usec\n", diff.tv_sec * 1000000UL + diff.tv_usec);
+       }
 
-       end.tv_sec -= start.tv_sec;
-       printf("Time replaying: %lu usec\n",
-              end.tv_sec * 1000000UL + (end.tv_usec - start.tv_usec));
-       
        exit(0);
 }