]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb/tools/replay_trace.c
Fix logic bug in optimizer, and fix up DEBUG_DEPS compile.
[ccan] / ccan / tdb / tools / replay_trace.c
index 61cdfb35d03b83ecd59be827d22d1c651c565911..db5a07be1cf7c7dd4b3cd7ae18a5770ca1d5ca63 100644 (file)
@@ -12,6 +12,9 @@
 #include <sys/types.h>
 #include <sys/wait.h>
 #include <sys/time.h>
 #include <sys/types.h>
 #include <sys/wait.h>
 #include <sys/time.h>
+#include <errno.h>
+#include <signal.h>
+#include <assert.h>
 
 #define STRINGIFY2(x) #x
 #define STRINGIFY(x) STRINGIFY2(x)
 
 #define STRINGIFY2(x) #x
 #define STRINGIFY(x) STRINGIFY2(x)
@@ -19,7 +22,7 @@
 /* Avoid mod by zero */
 static unsigned int total_keys = 1;
 
 /* Avoid mod by zero */
 static unsigned int total_keys = 1;
 
-#define DEBUG_DEPS 1
+/* #define DEBUG_DEPS 1 */
 
 /* Traversals block transactions in the current implementation. */
 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
 
 /* Traversals block transactions in the current implementation. */
 #define TRAVERSALS_TAKE_TRANSACTION_LOCK 1
@@ -48,7 +51,8 @@ static void __attribute__((noreturn)) fail(const char *filename,
        do {                                                            \
                int ret = (expr);                                       \
                if (ret != (expect))                                    \
        do {                                                            \
                int ret = (expr);                                       \
                if (ret != (expect))                                    \
-                       fail(filename, i+1, STRINGIFY(expr) "= %i", ret); \
+                       fail(filename[file], i+1,                       \
+                            STRINGIFY(expr) "= %i", ret);              \
        } while (0)
 
 /* Try or imitate results. */
        } while (0)
 
 /* Try or imitate results. */
@@ -57,7 +61,8 @@ static void __attribute__((noreturn)) fail(const char *filename,
                int ret = expr;                                         \
                if (ret != expect) {                                    \
                        fprintf(stderr, "%s:%u: %s gave %i not %i",     \
                int ret = expr;                                         \
                if (ret != expect) {                                    \
                        fprintf(stderr, "%s:%u: %s gave %i not %i",     \
-                             filename, i+1, STRINGIFY(expr), ret, expect); \
+                               filename[file], i+1, STRINGIFY(expr),   \
+                               ret, expect);                           \
                        if (expect == 0)                                \
                                force;                                  \
                        else                                            \
                        if (expect == 0)                                \
                                force;                                  \
                        else                                            \
@@ -129,8 +134,8 @@ struct op {
 
        /* Who is waiting for us? */
        struct list_head post;
 
        /* Who is waiting for us? */
        struct list_head post;
-       /* How many are we waiting for? */
-       unsigned int pre;
+       /* What are we waiting for? */
+       struct list_head pre;
 
        /* If I'm part of a group (traverse/transaction) where is
         * start?  (Otherwise, 0) */
 
        /* If I'm part of a group (traverse/transaction) where is
         * start?  (Otherwise, 0) */
@@ -139,7 +144,10 @@ struct op {
        union {
                int flag; /* open and store */
                struct traverse *trav; /* traverse start */
        union {
                int flag; /* open and store */
                struct traverse *trav; /* traverse start */
-               TDB_DATA post_append; /* append */
+               struct {  /* append */
+                       TDB_DATA pre;
+                       TDB_DATA post;
+               } append;
                unsigned int transaction_end; /* transaction start */
        };
 };
                unsigned int transaction_end; /* transaction start */
        };
 };
@@ -187,7 +195,6 @@ static void add_op(const char *filename, struct op **op, unsigned int i,
        new = (*op) + i;
        new->op = type;
        new->serial = serial;
        new = (*op) + i;
        new->op = type;
        new->serial = serial;
-       new->pre = 0;
        new->ret = 0;
        new->group_start = 0;
 }
        new->ret = 0;
        new->group_start = 0;
 }
@@ -262,8 +269,14 @@ static void op_add_append(const char *filename,
 
        op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
        op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
 
        op[op_num].key = make_tdb_data(op, filename, op_num+1, words[2]);
        op[op_num].data = make_tdb_data(op, filename, op_num+1, words[3]);
-       op[op_num].post_append
+
+       op[op_num].append.post
                = make_tdb_data(op, filename, op_num+1, words[5]);
                = make_tdb_data(op, filename, op_num+1, words[5]);
+
+       /* By subtraction, figure out what previous data was. */
+       op[op_num].append.pre.dptr = op[op_num].append.post.dptr;
+       op[op_num].append.pre.dsize
+               = op[op_num].append.post.dsize - op[op_num].data.dsize;
        total_keys++;
 }
 
        total_keys++;
 }
 
@@ -433,6 +446,88 @@ find_keyword (register const char *str, register unsigned int len);
 
 #include "keywords.c"
 
 
 #include "keywords.c"
 
+struct depend {
+       /* We can have more than one */
+       struct list_node pre_list;
+       struct list_node post_list;
+       unsigned int needs_file;
+       unsigned int needs_opnum;
+       unsigned int satisfies_file;
+       unsigned int satisfies_opnum;
+};
+
+static void check_deps(const char *filename, struct op op[], unsigned int num)
+{
+#ifdef DEBUG_DEPS
+       unsigned int i;
+
+       for (i = 1; i < num; i++)
+               if (!list_empty(&op[i].pre))
+                       fail(filename, i+1, "Still has dependencies");
+#endif
+}
+
+static void dump_pre(char *filename[], unsigned int file,
+                    struct op op[], 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);
+}
+
+/* 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)
+{
+       while (!list_empty(&op[i].pre)) {
+               struct depend *dep;
+
+#if DEBUG_DEPS
+               printf("%s:%u:waiting for pre\n", filename[file], i+1);
+               fflush(stdout);
+#endif
+               alarm(10);
+               while (read(pre_fd, &dep, sizeof(dep)) != sizeof(dep)) {
+                       if (errno == EINTR) {
+                               dump_pre(filename, file, op, i);
+                               exit(1);
+                       } else
+                               errx(1, "Reading from pipe");
+               }
+               alarm(0);
+
+#if DEBUG_DEPS
+               printf("%s:%u:got pre %u from %s:%u\n", filename[file], i+1,
+                      dep->needs_opnum+1, filename[dep->satisfies_file],
+                      dep->satisfies_opnum+1);
+               fflush(stdout);
+#endif
+               /* This could be any op, not just this one. */
+               talloc_free(dep);
+       }
+}
+
+static void do_post(char *filename[], unsigned int file,
+                   const struct op op[], unsigned int i)
+{
+       struct depend *dep;
+
+       list_for_each(&op[i].post, dep, post_list) {
+#if DEBUG_DEPS
+               printf("%s:%u:sending to file %s:%u\n", filename[file], i+1,
+                      filename[dep->needs_file], dep->needs_opnum+1);
+#endif
+               if (write(pipes[dep->needs_file].fd[1], &dep, sizeof(dep))
+                   != sizeof(dep))
+                       err(1, "%s:%u failed to tell file %s",
+                           filename[file], i+1, filename[dep->needs_file]);
+       }
+}
+
 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 {
        return data.dsize;
 static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 {
        return data.dsize;
@@ -440,13 +535,15 @@ static int get_len(TDB_DATA key, TDB_DATA data, void *private_data)
 
 static unsigned run_ops(struct tdb_context *tdb,
                        int pre_fd,
 
 static unsigned run_ops(struct tdb_context *tdb,
                        int pre_fd,
-                       const char *filename,
+                       char *filename[],
+                       unsigned int file,
                        struct op op[],
                        unsigned int start, unsigned int stop);
 
 struct traverse_info {
        struct op *op;
                        struct op op[],
                        unsigned int start, unsigned int stop);
 
 struct traverse_info {
        struct op *op;
-       const char *filename;
+       char **filename;
+       unsigned file;
        int pre_fd;
        unsigned int start;
        unsigned int i;
        int pre_fd;
        unsigned int start;
        unsigned int i;
@@ -463,14 +560,16 @@ static int trivial_traverse(struct tdb_context *tdb,
 
        while (trav->hash[h].index) {
                if (key_eq(trav->hash[h].key, key)) {
 
        while (trav->hash[h].index) {
                if (key_eq(trav->hash[h].key, key)) {
-                       run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
-                               trav->hash[h].index, trav->end);
+                       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);
        }
                        tinfo->i++;
                        return 0;
                }
                h = (h + 1) % (trav->num * 2);
        }
-       fail(tinfo->filename, tinfo->start + 1, "unexpected traverse key");
+       fail(tinfo->filename[tinfo->file], tinfo->start + 1,
+            "unexpected traverse key");
 }
 
 /* More complex.  Just do whatever's they did at the n'th entry. */
 }
 
 /* More complex.  Just do whatever's they did at the n'th entry. */
@@ -485,17 +584,17 @@ static int nontrivial_traverse(struct tdb_context *tdb,
                /* This can happen if traverse expects to be empty. */
                if (tinfo->start + 1 == trav->end)
                        return 1;
                /* This can happen if traverse expects to be empty. */
                if (tinfo->start + 1 == trav->end)
                        return 1;
-               fail(tinfo->filename, tinfo->start + 1,
+               fail(tinfo->filename[tinfo->file], tinfo->start + 1,
                     "traverse did not terminate");
        }
 
        if (tinfo->op[tinfo->i].op != OP_TDB_TRAVERSE)
                     "traverse did not terminate");
        }
 
        if (tinfo->op[tinfo->i].op != OP_TDB_TRAVERSE)
-               fail(tinfo->filename, tinfo->start + 1,
+               fail(tinfo->filename[tinfo->file], tinfo->start + 1,
                     "%s:%u:traverse terminated early");
 
        /* Run any normal ops. */
                     "%s:%u:traverse terminated early");
 
        /* Run any normal ops. */
-       tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->op,
-                          tinfo->i+1, trav->end);
+       tinfo->i = run_ops(tdb, tinfo->pre_fd, tinfo->filename, tinfo->file,
+                          tinfo->op, tinfo->i+1, trav->end);
 
        if (tinfo->i == trav->end)
                return 1;
 
        if (tinfo->i == trav->end)
                return 1;
@@ -505,20 +604,23 @@ static int nontrivial_traverse(struct tdb_context *tdb,
 
 static unsigned op_traverse(struct tdb_context *tdb,
                            int pre_fd,
 
 static unsigned op_traverse(struct tdb_context *tdb,
                            int pre_fd,
-                           const char *filename,
+                           char *filename[],
+                           unsigned int file,
                            int (*traversefn)(struct tdb_context *,
                                              tdb_traverse_func, void *),
                            struct op op[],
                            unsigned int start)
 {
        struct traverse *trav = op[start].trav;
                            int (*traversefn)(struct tdb_context *,
                                              tdb_traverse_func, void *),
                            struct op op[],
                            unsigned int start)
 {
        struct traverse *trav = op[start].trav;
-       struct traverse_info tinfo = { op, filename, pre_fd, start, start+1 };
+       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)
 
        /* Trivial case. */
        if (trav->hash) {
                int ret = traversefn(tdb, trivial_traverse, &tinfo);
                if (ret != trav->num)
-                       fail(filename, start+1, "short traversal %i", ret);
+                       fail(filename[file], start+1,
+                            "short traversal %i", ret);
                return trav->end;
        }
 
                return trav->end;
        }
 
@@ -531,73 +633,33 @@ static unsigned op_traverse(struct tdb_context *tdb,
                if (op[tinfo.i].op == OP_TDB_TRAVERSE)
                        tinfo.i++;
                else
                if (op[tinfo.i].op == OP_TDB_TRAVERSE)
                        tinfo.i++;
                else
-                       tinfo.i = run_ops(tdb, pre_fd, filename, op,
+                       tinfo.i = run_ops(tdb, pre_fd, filename, file, op,
                                          tinfo.i, trav->end);
        }
                                          tinfo.i, trav->end);
        }
-       return trav->end;
-}
-
-struct depend {
-       /* We can have more than one */
-       struct list_node list;
-       unsigned int file;
-       unsigned int op;
-};
-
-static void do_pre(const char *filename, int pre_fd,
-                  struct op op[], unsigned int i)
-{
-       while (op[i].pre != 0) {
-               unsigned int opnum;
-
-#if DEBUG_DEPS
-               printf("%s:%u:waiting for pre\n", filename, i+1);
-               fflush(stdout);
-#endif
-               if (read(pre_fd, &opnum, sizeof(opnum)) != sizeof(opnum))
-                       errx(1, "Reading from pipe");
 
 
-#if DEBUG_DEPS
-               printf("%s:%u:got pre %u (%u)\n", filename, i+1, opnum+1,
-                      op[opnum].pre);
-               fflush(stdout);
-#endif
-               /* This could be any op, not just this one. */
-               if (op[opnum].pre == 0)
-                       errx(1, "Got extra notification for op line %u",
-                            opnum + 1);
-               if (opnum < i)
-                       errx(1, "%s:%u: got late notification for line %u",
-                            filename, i + 1, opnum + 1);
-               op[opnum].pre--;
-       }
+       return trav->end;
 }
 
 }
 
-static void do_post(const char *filename, const struct op op[], unsigned int i)
+static void break_out(int sig)
 {
 {
-       struct depend *dep;
-
-       list_for_each(&op[i].post, dep, list) {
-#if DEBUG_DEPS
-               printf("%s:%u:sending %u to file %u\n", filename, i+1,
-                      dep->op+1, dep->file);
-#endif
-               if (write(pipes[dep->file].fd[1], &dep->op, sizeof(dep->op))
-                   != sizeof(dep->op))
-                       err(1, "Failed to tell file %u", dep->file);
-       }
 }
 
 static __attribute__((noinline))
 unsigned run_ops(struct tdb_context *tdb,
                 int pre_fd,
 }
 
 static __attribute__((noinline))
 unsigned run_ops(struct tdb_context *tdb,
                 int pre_fd,
-                const char *filename,
+                char *filename[],
+                unsigned int file,
                 struct op op[], unsigned int start, unsigned int stop)
 {
        unsigned int i;
                 struct op op[], unsigned int start, unsigned int stop)
 {
        unsigned int i;
+       struct sigaction sa;
+
+       sa.sa_handler = break_out;
+       sa.sa_flags = 0;
 
 
+       sigaction(SIGALRM, &sa, NULL);
        for (i = start; i < stop; i++) {
        for (i = start; i < stop; i++) {
-               do_pre(filename, pre_fd, op, i);
+               do_pre(filename, file, pre_fd, op, i);
 
                switch (op[i].op) {
                case OP_TDB_LOCKALL:
 
                switch (op[i].op) {
                case OP_TDB_LOCKALL:
@@ -660,11 +722,10 @@ unsigned run_ops(struct tdb_context *tdb,
                        break;
                case OP_TDB_STORE:
                        try(tdb_store(tdb, op[i].key, op[i].data, op[i].flag),
                        break;
                case OP_TDB_STORE:
                        try(tdb_store(tdb, op[i].key, op[i].data, op[i].flag),
-                           op[i].ret < 0 ? op[i].ret : 0);
+                           op[i].ret);
                        break;
                case OP_TDB_APPEND:
                        break;
                case OP_TDB_APPEND:
-                       try(tdb_append(tdb, op[i].key, op[i].data),
-                           op[i].ret < 0 ? op[i].ret : 0);
+                       try(tdb_append(tdb, op[i].key, op[i].data), op[i].ret);
                        break;
                case OP_TDB_GET_SEQNUM:
                        try(tdb_get_seqnum(tdb), op[i].ret);
                        break;
                case OP_TDB_GET_SEQNUM:
                        try(tdb_get_seqnum(tdb), op[i].ret);
@@ -682,11 +743,11 @@ unsigned run_ops(struct tdb_context *tdb,
                        try(tdb_transaction_commit(tdb), op[i].ret);
                        break;
                case OP_TDB_TRAVERSE_READ_START:
                        try(tdb_transaction_commit(tdb), op[i].ret);
                        break;
                case OP_TDB_TRAVERSE_READ_START:
-                       i = op_traverse(tdb, pre_fd, filename,
+                       i = op_traverse(tdb, pre_fd, filename, file,
                                        tdb_traverse_read, op, i);
                        break;
                case OP_TDB_TRAVERSE_START:
                                        tdb_traverse_read, op, i);
                        break;
                case OP_TDB_TRAVERSE_START:
-                       i = op_traverse(tdb, pre_fd, filename,
+                       i = op_traverse(tdb, pre_fd, filename, file,
                                        tdb_traverse, op, i);
                        break;
                case OP_TDB_TRAVERSE:
                                        tdb_traverse, op, i);
                        break;
                case OP_TDB_TRAVERSE:
@@ -694,27 +755,28 @@ unsigned run_ops(struct tdb_context *tdb,
                         * done our ops. */
                        return i;
                case OP_TDB_TRAVERSE_END:
                         * done our ops. */
                        return i;
                case OP_TDB_TRAVERSE_END:
-                       fail(filename, i+1, "unepxected end traverse");
+                       fail(filename[file], i+1, "unepxected end traverse");
                /* FIXME: These must be treated like traverse. */
                case OP_TDB_FIRSTKEY:
                        if (!key_eq(tdb_firstkey(tdb), op[i].data))
                /* FIXME: These must be treated like traverse. */
                case OP_TDB_FIRSTKEY:
                        if (!key_eq(tdb_firstkey(tdb), op[i].data))
-                               fail(filename, i+1, "bad firstkey");
+                               fail(filename[file], i+1, "bad firstkey");
                        break;
                case OP_TDB_NEXTKEY:
                        if (!key_eq(tdb_nextkey(tdb, op[i].key), op[i].data))
                        break;
                case OP_TDB_NEXTKEY:
                        if (!key_eq(tdb_nextkey(tdb, op[i].key), op[i].data))
-                               fail(filename, i+1, "bad nextkey");
+                               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))
                        break;
                case OP_TDB_FETCH: {
                        TDB_DATA f = tdb_fetch(tdb, op[i].key);
                        if (!key_eq(f, op[i].data))
-                               fail(filename, i+1, "bad fetch %u", f.dsize);
+                               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);
                        break;
                }
                        break;
                }
                case OP_TDB_DELETE:
                        try(tdb_delete(tdb, op[i].key), op[i].ret);
                        break;
                }
-               do_post(filename, op, i);
+               do_post(filename, file, op, i);
        }
        return i;
 }
        }
        return i;
 }
@@ -789,16 +851,101 @@ struct keyinfo {
        struct key_user *user;
 };
 
        struct key_user *user;
 };
 
-static bool changes_db(const struct op *op)
+static const TDB_DATA must_not_exist;
+static const TDB_DATA must_exist;
+static const TDB_DATA not_exists_or_empty;
+
+/* NULL means doesn't care if it exists or not, &must_exist means
+ * it must exist but we don't care what, &must_not_exist means it must
+ * not exist, otherwise the data it needs. */
+static const TDB_DATA *needs(const struct op *op)
 {
 {
-       if (op->ret != 0)
-               return false;
+       switch (op->op) {
+       /* FIXME: Pull forward deps, since we can deadlock */
+       case OP_TDB_CHAINLOCK:
+       case OP_TDB_CHAINLOCK_NONBLOCK:
+       case OP_TDB_CHAINLOCK_MARK:
+       case OP_TDB_CHAINLOCK_UNMARK:
+       case OP_TDB_CHAINUNLOCK:
+       case OP_TDB_CHAINLOCK_READ:
+       case OP_TDB_CHAINUNLOCK_READ:
+               return NULL;
+
+       case OP_TDB_APPEND:
+               if (op->append.pre.dsize == 0)
+                       return &not_exists_or_empty;
+               return &op->append.pre;
+
+       case OP_TDB_STORE:
+               if (op->flag == TDB_INSERT) {
+                       if (op->ret < 0)
+                               return &must_exist;
+                       else
+                               return &must_not_exist;
+               } else if (op->flag == TDB_MODIFY) {
+                       if (op->ret < 0)
+                               return &must_not_exist;
+                       else
+                               return &must_exist;
+               }
+               /* No flags?  Don't care */
+               return NULL;
+
+       case OP_TDB_EXISTS:
+               if (op->ret == 1)
+                       return &must_exist;
+               else
+                       return &must_not_exist;
+
+       case OP_TDB_PARSE_RECORD:
+               if (op->ret < 0)
+                       return &must_not_exist;
+               return &must_exist;
+
+       /* FIXME: handle these. */
+       case OP_TDB_WIPE_ALL:
+       case OP_TDB_FIRSTKEY:
+       case OP_TDB_NEXTKEY:
+       case OP_TDB_GET_SEQNUM:
+       case OP_TDB_TRAVERSE:
+       case OP_TDB_TRANSACTION_COMMIT:
+       case OP_TDB_TRANSACTION_CANCEL:
+       case OP_TDB_TRANSACTION_START:
+               return NULL;
+
+       case OP_TDB_FETCH:
+               if (!op->data.dptr)
+                       return &must_not_exist;
+               return &op->data;
+
+       case OP_TDB_DELETE:
+               if (op->ret < 0)
+                       return &must_not_exist;
+               return &must_exist;
+
+       default:
+               errx(1, "Unexpected op %i", op->op);
+       }
+       
+}
+
+/* What's the data after this op?  pre if nothing changed. */
+static const TDB_DATA *gives(const struct op *op, const TDB_DATA *pre)
+{
+       /* Failed ops don't change state of db. */
+       if (op->ret < 0)
+               return pre;
+
+       if (op->op == OP_TDB_DELETE || op->op == OP_TDB_WIPE_ALL)
+               return &tdb_null;
+
+       if (op->op == OP_TDB_APPEND)
+               return &op->append.post;
+
+       if (op->op == OP_TDB_STORE)
+               return &op->data;
 
 
-       return op->op == OP_TDB_STORE
-               || op->op == OP_TDB_APPEND
-               || op->op == OP_TDB_WIPE_ALL
-               || op->op == OP_TDB_TRANSACTION_COMMIT
-               || op->op == OP_TDB_DELETE;
+       return pre;
 }
 
 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
 }
 
 static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
@@ -807,30 +954,12 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
        unsigned int i, j, h;
        struct keyinfo *hash;
 
        unsigned int i, j, h;
        struct keyinfo *hash;
 
-       /* Gcc nexted function extension.  How cool is this? */
-       int compare_user_serial(const void *_a, const void *_b)
-       {
-               const struct key_user *a = _a, *b = _b;
-               int ret = op[a->file][a->op_num].serial
-                       - op[b->file][b->op_num].serial;
-
-               /* Fetches don't inc serial, so we put changes first. */
-               if (ret == 0) {
-                       if (changes_db(&op[a->file][a->op_num])
-                           && !changes_db(&op[b->file][b->op_num]))
-                               return -1;
-                       if (changes_db(&op[b->file][b->op_num])
-                           && !changes_db(&op[a->file][a->op_num]))
-                               return 1;
-               }
-               return ret;
-       }
-
        hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
        for (i = 0; i < num; i++) {
                for (j = 1; j < num_ops[i]; j++) {
                        /* We can't do this on allocation, due to realloc. */
                        list_head_init(&op[i][j].post);
        hash = talloc_zero_array(op[0], struct keyinfo, total_keys*2);
        for (i = 0; i < num; i++) {
                for (j = 1; j < num_ops[i]; j++) {
                        /* We can't do this on allocation, due to realloc. */
                        list_head_init(&op[i][j].post);
+                       list_head_init(&op[i][j].pre);
 
                        if (!op[i][j].key.dptr)
                                continue;
 
                        if (!op[i][j].key.dptr)
                                continue;
@@ -862,24 +991,192 @@ static struct keyinfo *hash_ops(struct op *op[], unsigned int num_ops[],
                }
        }
 
                }
        }
 
-       /* Now sort into seqnum order. */
-       for (h = 0; h < total_keys * 2; h++)
-               qsort(hash[h].user, hash[h].num_users, sizeof(hash[h].user[0]),
-                     compare_user_serial);
-
        return hash;
 }
 
        return hash;
 }
 
+static bool satisfies(const TDB_DATA *data, const TDB_DATA *need)
+{
+       /* Don't need anything?  Cool. */
+       if (!need)
+               return true;
+
+       /* This should be tdb_null or a real value. */
+       assert(data != &must_exist);
+       assert(data != &must_not_exist);
+       assert(data != &not_exists_or_empty);
+
+       /* must_not_exist == 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;
+
+       /* Must exist? */
+       if (need == &must_exist)
+               return data->dptr != NULL;
+
+       /* Either noexist or empty. */
+       if (need == &not_exists_or_empty)
+               return data->dsize == 0;
+
+       /* Needs something specific. */
+       return key_eq(*data, *need);
+}
+
+static void move_to_front(struct key_user res[], unsigned int elem)
+{
+       if (elem != 0) {
+               struct key_user tmp = res[elem];
+               memmove(res + 1, res, elem*sizeof(res[0]));
+               res[0] = tmp;
+       }
+}
+
+static void restore_to_pos(struct key_user res[], unsigned int elem)
+{
+       if (elem != 0) {
+               struct key_user tmp = res[0];
+               memmove(res, res + 1, elem*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)
+{
+       unsigned int i, files_done;
+       struct op *this_op;
+       bool done[num_files];
+
+       /* Nothing left?  We're sorted. */
+       if (num == 0)
+               return true;
+
+       memset(done, 0, sizeof(done));
+
+       /* Since ops within a trace file are ordered, we just need to figure
+        * out which file to try next.  Since we don't take into account
+        * inter-key relationships (which exist by virtue of trace file order),
+        * we minimize the chance of harm by trying to keep in serial order. */
+       for (files_done = 0, i = 0; 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))
+                               return true;
+                       restore_to_pos(res, i);
+               }
+               done[res[i].file] = true;
+               files_done++;
+       }
+
+       /* No combination worked. */
+       return false;
+}
+
+static void check_dep_sorting(struct key_user user[], unsigned num_users,
+                             unsigned num_files)
+{
+#if DEBUG_DEPS
+       unsigned int i;
+       unsigned minima[num_files];
+
+       memset(minima, 0, sizeof(minima));
+       for (i = 0; i < num_users; i++) {
+               assert(minima[user[i].file] < user[i].op_num);
+               minima[user[i].file] = user[i].op_num;
+       }
+#endif
+}
+
+/* All these ops have the same serial number.  Which comes first?
+ *
+ * This can happen both because read ops or failed write ops don't
+ * change serial number, and also due to race since we access the
+ * number unlocked (the race can cause less detectable ordering problems,
+ * in which case we'll deadlock and report: fix manually in that case).
+ */
+static void figure_deps(char *filename[], struct op *op[],
+                       struct key_user user[], unsigned num_users,
+                       unsigned num_files)
+{
+       /* We assume database starts empty. */
+       const struct TDB_DATA *data = &tdb_null;
+
+       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");
+
+       check_dep_sorting(user, num_users, num_files);
+}
+
+static void sort_ops(struct keyinfo hash[], char *filename[], struct op *op[],
+                    unsigned int num)
+{
+       unsigned int h;
+
+       /* Gcc nexted function extension.  How cool is this? */
+       int compare_serial(const void *_a, const void *_b)
+       {
+               const struct key_user *a = _a, *b = _b;
+
+               /* First, maintain order within any trace file. */
+               if (a->file == b->file)
+                       return a->op_num - b->op_num;
+
+               /* Otherwise, arrange by serial order. */
+               return op[a->file][a->op_num].serial
+                       - op[b->file][b->op_num].serial;
+       }
+
+       /* Now sort into serial order. */
+       for (h = 0; h < total_keys * 2; h++) {
+               struct key_user *user = hash[h].user;
+
+               qsort(user, hash[h].num_users, sizeof(user[0]), compare_serial);
+               figure_deps(filename, op, user, hash[h].num_users, num);
+       }
+}
+
+static int destroy_depend(struct depend *dep)
+{
+       list_del(&dep->pre_list);
+       list_del(&dep->post_list);
+       return 0;
+}
+
 static void add_dependency(void *ctx,
                           struct op *op[],
 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)
 {
                           unsigned int needs_file,
                           unsigned int needs_opnum,
                           unsigned int satisfies_file,
                           unsigned int satisfies_opnum)
 {
-       struct depend *post;
+       struct depend *dep;
        unsigned int needs_start, sat_start;
 
        unsigned int needs_start, sat_start;
 
+       /* We don't depend on ourselves. */
+       if (needs_file == satisfies_file) {
+               assert(satisfies_opnum < needs_opnum);
+               return;
+       }
+
+#if DEBUG_DEPS
+       printf("%s:%u: depends on %s:%u\n",
+              filename[needs_file], needs_opnum+1,
+              filename[satisfies_file], satisfies_opnum+1);
+#endif
+
        needs_start = op[needs_file][needs_opnum].group_start;
        sat_start = op[satisfies_file][satisfies_opnum].group_start;
 
        needs_start = op[needs_file][needs_opnum].group_start;
        sat_start = op[satisfies_file][satisfies_opnum].group_start;
 
@@ -887,10 +1184,6 @@ static void add_dependency(void *ctx,
        if (needs_start) {
                switch (op[needs_file][needs_start].op) {
                case OP_TDB_TRANSACTION_START:
        if (needs_start) {
                switch (op[needs_file][needs_start].op) {
                case OP_TDB_TRANSACTION_START:
-#if TRAVERSALS_TAKE_TRANSACTION_LOCK
-               case OP_TDB_TRAVERSE_START:
-               case OP_TDB_TRAVERSE_READ_START:
-#endif
                        needs_opnum = needs_start;
 #ifdef DEBUG_DEPS
                        printf("  -> Back to %u\n", needs_start+1);
                        needs_opnum = needs_start;
 #ifdef DEBUG_DEPS
                        printf("  -> Back to %u\n", needs_start+1);
@@ -916,20 +1209,161 @@ static void add_dependency(void *ctx,
                }
        }
 
                }
        }
 
-       post = talloc(ctx, struct depend);
-       post->file = needs_file;
-       post->op = needs_opnum;
-       list_add(&op[satisfies_file][satisfies_opnum].post, &post->list);
+       dep = talloc(ctx, struct depend);
+       dep->needs_file = needs_file;
+       dep->needs_opnum = needs_opnum;
+       dep->satisfies_file = satisfies_file;
+       dep->satisfies_opnum = satisfies_opnum;
+       list_add(&op[satisfies_file][satisfies_opnum].post, &dep->post_list);
+       list_add(&op[needs_file][needs_opnum].pre, &dep->pre_list);
+       talloc_set_destructor(dep, destroy_depend);
+}
 
 
-       op[needs_file][needs_opnum].pre++;
-#ifdef DEBUG_DEPS
-       if (op[needs_file][needs_opnum].pre > 1) {
-               printf("   (File %u opnum %u hash %u needs)\n",
-                      needs_file, needs_opnum+1,
-                      op[needs_file][needs_opnum].pre);
-               fflush(stdout);
+#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)
+{
+       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;
+}
+
+static void depend_on_previous(struct op *op[],
+                              char *filename[],
+                              unsigned int num,
+                              struct key_user user[],
+                              unsigned int i,
+                              int prev)
+{
+       bool deps[num];
+       int j;
+
+       if (i == 0)
+               return;
+
+       if (prev == i - 1) {
+               /* Just depend on previous. */
+               add_dependency(NULL, op, filename,
+                              user[i].file, user[i].op_num,
+                              user[prev].file, user[prev].op_num);
+               return;
+       }
+
+       /* We have to wait for the readers.  Find last one in *each* file. */
+       memset(deps, 0, sizeof(deps));
+       deps[user[i].file] = true;
+       for (j = i - 1; j > prev; j--) {
+               if (!deps[user[j].file]) {
+                       add_dependency(NULL, op, filename,
+                                      user[i].file, user[i].op_num,
+                                      user[j].file, user[j].op_num);
+                       deps[user[j].file] = true;
+               }
+       }
+}
+
+/* This is simple, but not complete.  We don't take into account
+ * indirect dependencies. */
+static void optimize_dependencies(struct op *op[], unsigned int num_ops[],
+                                 unsigned int num)
+{
+       unsigned int i, j;
+
+       /* There can only be one real dependency on each file */
+       for (i = 0; i < num; i++) {
+               for (j = 1; j < num_ops[i]; j++) {
+                       struct depend *dep, *next;
+                       struct depend *prev[num];
+
+                       memset(prev, 0, sizeof(prev));
+
+                       list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
+                               if (!prev[dep->satisfies_file]) {
+                                       prev[dep->satisfies_file] = dep;
+                                       continue;
+                               }
+                               if (prev[dep->satisfies_file]->satisfies_opnum
+                                   < dep->satisfies_opnum) {
+                                       talloc_free(prev[dep->satisfies_file]);
+                                       prev[dep->satisfies_file] = dep;
+                               } else
+                                       talloc_free(dep);
+                       }
+               }
+       }
+
+       for (i = 0; i < num; i++) {
+               int deps[num];
+
+               for (j = 0; j < num; j++)
+                       deps[j] = -1;
+
+               for (j = 1; j < num_ops[i]; j++) {
+                       struct depend *dep, *next;
+
+                       list_for_each_safe(&op[i][j].pre, dep, next, pre_list) {
+                               if (deps[dep->satisfies_file]
+                                   >= (int)dep->satisfies_opnum)
+                                       talloc_free(dep);
+                               else
+                                       deps[dep->satisfies_file]
+                                               = dep->satisfies_opnum;
+                       }
+               }
        }
        }
-#endif
 }
 
 static void derive_dependencies(char *filename[],
 }
 
 static void derive_dependencies(char *filename[],
@@ -937,35 +1371,48 @@ static void derive_dependencies(char *filename[],
                                unsigned int num)
 {
        struct keyinfo *hash;
                                unsigned int num)
 {
        struct keyinfo *hash;
-       unsigned int i;
+       unsigned int h, i;
 
        /* Create hash table for faster key lookup. */
        hash = hash_ops(op, num_ops, num);
 
 
        /* Create hash table for faster key lookup. */
        hash = hash_ops(op, num_ops, num);
 
-       /* We make the naive assumption that two ops on the same key
-        * have to be ordered; it's overkill. */
-       for (i = 0; i < total_keys * 2; i++) {
-               unsigned int j;
+       /* Sort them by serial number. */
+       sort_ops(hash, filename, op, num);
+
+       /* Create dependencies back to the last change, rather than
+        * creating false dependencies by naively making each one
+        * depend on the previous.  This has two purposes: it makes
+        * later optimization simpler, and it also avoids deadlock with
+        * same sequence number ops inside traversals (if one
+        * traversal doesn't write anything, two ops can have the same
+        * sequence number yet we can create a traversal dependency
+        * the other way). */
+       for (h = 0; h < total_keys * 2; h++) {
+               int prev = -1;
+
+               if (hash[h].num_users < 2)
+                       continue;
 
 
-               for (j = 1; j < hash[i].num_users; j++) {
-                       /* We don't depend on ourselves. */
-                       if (hash[i].user[j].file == hash[i].user[j-1].file)
-                               continue;
-#if DEBUG_DEPS
-                       printf("%s:%u: depends on %s:%u\n",
-                              filename[hash[i].user[j].file],
-                              hash[i].user[j].op_num+1,
-                              filename[hash[i].user[j-1].file],
-                              hash[i].user[j-1].op_num+1);
-                       fflush(stdout);
-#endif
-                       add_dependency(hash, op,
-                                      hash[i].user[j].file,
-                                      hash[i].user[j].op_num,
-                                      hash[i].user[j-1].file,
-                                      hash[i].user[j-1].op_num);
+               for (i = 0; i < hash[h].num_users; i++) {
+                       if (changes_db(&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);
                }
        }
                }
        }
+
+#if TRAVERSALS_TAKE_TRANSACTION_LOCK
+       make_traverse_depends(filename, op, num_ops, num);
+#endif
+
+       optimize_dependencies(op, num_ops, num);
 }
 
 int main(int argc, char *argv[])
 }
 
 int main(int argc, char *argv[])
@@ -975,19 +1422,26 @@ int main(int argc, char *argv[])
        struct op *op[argc];
        int fds[2];
        char c;
        struct op *op[argc];
        int fds[2];
        char c;
+       bool ok = true;
 
        if (argc < 3)
                errx(1, "Usage: %s <tdbfile> <tracefile>...", argv[0]);
 
        pipes = talloc_array(NULL, struct pipe, argc - 2);
        for (i = 0; i < argc - 2; i++) {
 
        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");
                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);
        derive_dependencies(argv+2, op, num_ops, i);
+       printf("done\n");
 
        /* Don't fork for single arg case: simple debugging. */
        if (argc == 3) {
 
        /* Don't fork for single arg case: simple debugging. */
        if (argc == 3) {
@@ -995,8 +1449,13 @@ int main(int argc, char *argv[])
                tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
                                  open_flags[0], 0600,
                                  NULL, hash_key);
                tdb = tdb_open_ex(argv[1], hashsize[0], tdb_flags[0],
                                  open_flags[0], 0600,
                                  NULL, hash_key);
-               run_ops(tdb, pipes[0].fd[0], argv[2],
-                       op[0], 1, num_ops[0]);
+               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);
        }
 
                exit(0);
        }
 
@@ -1020,8 +1479,9 @@ int main(int argc, char *argv[])
                        /* This catches parent exiting. */
                        if (read(fds[0], &c, 1) != 1)
                                exit(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, i, op[i], 1,
+                               num_ops[i]);
+                       check_deps(argv[2+i], op[i], num_ops[i]);
                        exit(0);
                default:
                        break;
                        exit(0);
                default:
                        break;
@@ -1031,6 +1491,8 @@ int main(int argc, char *argv[])
        /* Let everything settle. */
        sleep(1);
 
        /* Let everything settle. */
        sleep(1);
 
+       printf("Starting run...");
+       fflush(stdout);
        gettimeofday(&start, NULL);
        /* Tell them all to go!  Any write of sufficient length will do. */
        if (write(fds[1], hashsize, i) != i)
        gettimeofday(&start, NULL);
        /* Tell them all to go!  Any write of sufficient length will do. */
        if (write(fds[1], hashsize, i) != i)
@@ -1039,12 +1501,18 @@ int main(int argc, char *argv[])
        for (i = 0; i < argc - 2; i++) {
                int status;
                wait(&status);
        for (i = 0; i < argc - 2; i++) {
                int status;
                wait(&status);
-               if (!WIFEXITED(status))
-                       errx(1, "Child died with signal %i", WTERMSIG(status));
-               if (WEXITSTATUS(status) != 0)
-                       errx(1, "Child died with error code");
+               if (!WIFEXITED(status)) {
+                       warnx("Child died with signal %i", WTERMSIG(status));
+                       ok = false;
+               } else if (WEXITSTATUS(status) != 0)
+                       /* Assume child spat out error. */
+                       ok = false;
        }
        }
+       if (!ok)
+               exit(1);
+
        gettimeofday(&end, NULL);
        gettimeofday(&end, NULL);
+       printf("done\n");
 
        end.tv_sec -= start.tv_sec;
        printf("Time replaying: %lu usec\n",
 
        end.tv_sec -= start.tv_sec;
        printf("Time replaying: %lu usec\n",