Fix the "too many fcntl locks" scalability problem raised by tridge.

I've now tested this in daemon mode and also on xinetd and I'm pretty
sure it's working.
Jeremy.
(This used to be commit 14dee038019b11300466b148c53515fc76e5e870)

4 files changed, 62 insertions, 35 deletions

diff --git a/source3/smbd/server.c b/source3/smbd/server.c
index be59e92cd7..1de33739b2 100644
--- a/source3/smbd/server.c
+++ b/source3/smbd/server.c
@@ -405,10 +405,10 @@ static BOOL open_sockets_smbd(BOOL is_daemon, BOOL interactive, const char *smb_
 				   done correctly in the process.  */
 				reset_globals_after_fork();
 
-				/* tdb needs special fork handling */
+				/* tdb needs special fork handling - remove CLEAR_IF_FIRST flags */
 				if (tdb_reopen_all() == -1) {
 					DEBUG(0,("tdb_reopen_all failed.\n"));
-					return False;
+					smb_panic("tdb_reopen_all failed.");
 				}
 
 				return True; 
@@ -809,9 +809,27 @@ void build_options(BOOL screen);
 	if (is_daemon)
 		pidfile_create("smbd");
 
+	/* Setup all the TDB's - including CLEAR_IF_FIRST tdb's. */
 	if (!message_init())
 		exit(1);
 
+	if (!session_init())
+		exit(1);
+
+	if (conn_tdb_ctx() == NULL)
+		exit(1);
+
+	if (!locking_init(0))
+		exit(1);
+
+	if (!share_info_db_init())
+		exit(1);
+
+	namecache_enable();
+
+	if (!init_registry())
+		exit(1);
+
 	if (!print_backend_init())
 		exit(1);
 
@@ -832,17 +850,6 @@ void build_options(BOOL screen);
 	 * everything after this point is run after the fork()
 	 */ 
 
-	namecache_enable();
-
-	if (!locking_init(0))
-		exit(1);
-
-	if (!share_info_db_init())
-		exit(1);
-
-	if (!init_registry())
-		exit(1);
-
 	/* Initialise the password backed before the global_sam_sid
 	   to ensure that we fetch from ldap before we make a domain sid up */
 
@@ -891,4 +898,3 @@ void build_options(BOOL screen);
 	exit_server("normal exit");
 	return(0);
 }
-
diff --git a/source3/smbd/session.c b/source3/smbd/session.c
index a811a6e305..61118f13dd 100644
--- a/source3/smbd/session.c
+++ b/source3/smbd/session.c
@@ -28,6 +28,22 @@
 #include "includes.h"
 
 static TDB_CONTEXT *tdb;
+
+BOOL session_init(void)
+{
+	if (tdb)
+		return True;
+
+	tdb = tdb_open_ex(lock_path("sessionid.tdb"), 0, TDB_CLEAR_IF_FIRST|TDB_DEFAULT, 
+		       O_RDWR | O_CREAT, 0644, smbd_tdb_log);
+	if (!tdb) {
+		DEBUG(1,("session_init: failed to open sessionid tdb\n"));
+		return False;
+	}
+
+	return True;
+}
+
 /* called when a session is created */
 BOOL session_claim(user_struct *vuser)
 {
@@ -52,14 +68,8 @@ BOOL session_claim(user_struct *vuser)
 		return True;
 	}
 
-	if (!tdb) {
-		tdb = tdb_open_ex(lock_path("sessionid.tdb"), 0, TDB_CLEAR_IF_FIRST|TDB_DEFAULT, 
-			       O_RDWR | O_CREAT, 0644, smbd_tdb_log);
-		if (!tdb) {
-			DEBUG(1,("session_claim: failed to open sessionid tdb\n"));
-			return False;
-		}
-	}
+	if (!session_init())
+		return False;
 
 	ZERO_STRUCT(sessionid);
 
@@ -190,7 +200,7 @@ void session_yield(user_struct *vuser)
 
 static BOOL session_traverse(int (*fn)(TDB_CONTEXT *, TDB_DATA, TDB_DATA, void *), void *state)
 {
-	if (!tdb) {
+	if (!session_init()) {
 		DEBUG(3, ("No tdb opened\n"));
 		return False;
 	}
@@ -238,4 +248,3 @@ int list_sessions(struct sessionid **session_list)
 	*session_list = sesslist.sessions;
 	return sesslist.count;
 }
-		
diff --git a/source3/tdb/tdb.c b/source3/tdb/tdb.c
index 7ad39175ac..7b10cfb377 100644
--- a/source3/tdb/tdb.c
+++ b/source3/tdb/tdb.c
@@ -1705,7 +1705,7 @@ TDB_CONTEXT *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
 {
 	TDB_CONTEXT *tdb;
 	struct stat st;
-	int rev = 0, locked;
+	int rev = 0, locked = 0;
 	unsigned char *vp;
 	u32 vertest;
 
@@ -1763,8 +1763,8 @@ TDB_CONTEXT *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
 	}
 
 	/* we need to zero database if we are the only one with it open */
-	if ((locked = (tdb_brlock(tdb, ACTIVE_LOCK, F_WRLCK, F_SETLK, 0) == 0))
-	    && (tdb_flags & TDB_CLEAR_IF_FIRST)) {
+	if ((tdb_flags & TDB_CLEAR_IF_FIRST) &&
+		(locked = (tdb_brlock(tdb, ACTIVE_LOCK, F_WRLCK, F_SETLK, 0) == 0))) {
 		open_flags |= O_CREAT;
 		if (ftruncate(tdb->fd, 0) == -1) {
 			TDB_LOG((tdb, 0, "tdb_open_ex: "
@@ -1837,10 +1837,19 @@ TDB_CONTEXT *tdb_open_ex(const char *name, int hash_size, int tdb_flags,
 				 name, strerror(errno)));
 			goto fail;
 		}
+
 	}
-	/* leave this lock in place to indicate it's in use */
-	if (tdb_brlock(tdb, ACTIVE_LOCK, F_RDLCK, F_SETLKW, 0) == -1)
-		goto fail;
+
+	/* We always need to do this if the CLEAR_IF_FIRST flag is set, even if
+	   we didn't get the initial exclusive lock as we need to let all other
+	   users know we're using it. */
+
+	if (tdb_flags & TDB_CLEAR_IF_FIRST) {
+		/* leave this lock in place to indicate it's in use */
+		if (tdb_brlock(tdb, ACTIVE_LOCK, F_RDLCK, F_SETLKW, 0) == -1)
+			goto fail;
+	}
+
 
  internal:
 	/* Internal (memory-only) databases skip all the code above to
@@ -2018,12 +2027,14 @@ void tdb_logging_function(TDB_CONTEXT *tdb, void (*fn)(TDB_CONTEXT *, int , cons
 }
 
 
-/* reopen a tdb - this is used after a fork to ensure that we have an independent
+/* reopen a tdb - this can be used after a fork to ensure that we have an independent
    seek pointer from our parent and to re-establish locks */
 int tdb_reopen(TDB_CONTEXT *tdb)
 {
 	struct stat st;
 
+	if (tdb->flags & TDB_INTERNAL)
+		return 0; /* Nothing to do. */
 	if (tdb_munmap(tdb) != 0) {
 		TDB_LOG((tdb, 0, "tdb_reopen: munmap failed (%s)\n", strerror(errno)));
 		goto fail;
@@ -2044,7 +2055,7 @@ int tdb_reopen(TDB_CONTEXT *tdb)
 		goto fail;
 	}
 	tdb_mmap(tdb);
-	if (tdb_brlock(tdb, ACTIVE_LOCK, F_RDLCK, F_SETLKW, 0) == -1) {
+	if ((tdb->flags & TDB_CLEAR_IF_FIRST) && (tdb_brlock(tdb, ACTIVE_LOCK, F_RDLCK, F_SETLKW, 0) == -1)) {
 		TDB_LOG((tdb, 0, "tdb_reopen: failed to obtain active lock\n"));
 		goto fail;
 	}
@@ -2062,7 +2073,10 @@ int tdb_reopen_all(void)
 	TDB_CONTEXT *tdb;
 
 	for (tdb=tdbs; tdb; tdb = tdb->next) {
-		if (tdb_reopen(tdb) != 0) return -1;
+		/* Ensure no clear-if-first. */
+		tdb->flags &= ~TDB_CLEAR_IF_FIRST;
+		if (tdb_reopen(tdb) != 0)
+			return -1;
 	}
 
 	return 0;
diff --git a/source3/tdb/tdbutil.c b/source3/tdb/tdbutil.c
index bae9a8b9d4..304bf9c816 100644
--- a/source3/tdb/tdbutil.c
+++ b/source3/tdb/tdbutil.c
@@ -826,5 +826,3 @@ void tdb_search_list_free(TDB_LIST_NODE* node)
 		node = next_node;
 	};
 }
-
-