s3:events: Call all ready fd event handlers on each iteration of the main loop

Previously, only one fd handler was being called per main message loop
in all smbd child processes.

In the case where multiple fds are available for reading the fd
corresponding to the event closest to the beginning of the event list
would be run.  Obviously this is arbitrary and could cause unfairness.

Usually, the first event fd is the network socket, meaning heavy load
of client requests can starve out other fd events such as oplock
or notify upcalls from the kernel.

In this patch, I have changed the behavior of run_events() to unset
any fd that it has already called a handler function, as well
as decrement the number of fds that were returned from select().
This allows the caller of run_events() to iterate it, until all
available fds have been handled.

I then changed the main loop in smbd child processes to iterate
run_events().  This way, all available fds are handled on each wake
of select, while still checking for timed or signalled events between
each handler function call.  I also added an explicit check for
EINTR from select(), which previously was masked by the fact that
run_events() would handle any signal event before the return code
was checked.

This required a signature change to run_events() but all other callers
should have no change in their behavior.  I also fixed a bug in
run_events() where it could be called with a selrtn value of -1,
doing unecessary looping through the fd_event list when no fds were
available.

Also, remove the temporary echo handler hack, as all fds should be
treated fairly now.

1 files changed, 16 insertions, 19 deletions

diff --git a/source3/smbd/process.c b/source3/smbd/process.c
index a484dfd3f2..8d36dd3c89 100644
--- a/source3/smbd/process.c
+++ b/source3/smbd/process.c
@@ -949,7 +949,7 @@ void smbd_setup_sig_hup_handler(struct tevent_context *ev,
 static NTSTATUS smbd_server_connection_loop_once(struct smbd_server_connection *conn)
 {
 	fd_set r_fds, w_fds;
-	int selrtn;
+	int selrtn = 0;
 	struct timeval to;
 	int maxfd = 0;
 
@@ -977,7 +977,7 @@ static NTSTATUS smbd_server_connection_loop_once(struct smbd_server_connection *
 	}
 
 	/* Process a signal and timed events now... */
-	if (run_events(smbd_event_context(), 0, NULL, NULL)) {
+	if (run_events(smbd_event_context(), &selrtn, NULL, NULL)) {
 		return NT_STATUS_RETRY;
 	}
 
@@ -994,26 +994,23 @@ static NTSTATUS smbd_server_connection_loop_once(struct smbd_server_connection *
 
 	/* Check if error */
 	if (selrtn == -1) {
-		/* something is wrong. Maybe the socket is dead? */
-		return map_nt_error_from_unix(errno);
+		if (errno == EINTR)
+			return NT_STATUS_RETRY;
+		else
+			/* Maybe the socket is dead? */
+			return map_nt_error_from_unix(errno);
 	}
 
-        if ((conn->smb1.echo_handler.trusted_fd != -1)
-	    && FD_ISSET(conn->sock, &r_fds)
-	    && FD_ISSET(conn->smb1.echo_handler.trusted_fd, &r_fds)) {
-		/*
-		 * Prefer to read pending requests from the echo handler. To
-		 * quote Jeremy (da70f8ab1): This is a hack of monstrous
-		 * proportions...
-		 */
-		FD_CLR(conn->sock, &r_fds);
-        }
-
-	if (run_events(smbd_event_context(), selrtn, &r_fds, &w_fds)) {
-		return NT_STATUS_RETRY;
-	}
+	/* Process events until all available fds have been handled.
+	 * This allows for fair round-robin handling of all available fds
+	 * on each select() wakeup, while still maintaining responsiveness
+	 * by re-checking for signal and timed events between the handling
+	 * of each ready fd. */
+	do {
+		run_events(smbd_event_context(), &selrtn, &r_fds, &w_fds);
+	} while (selrtn > 0);
 
-	/* Did we timeout ? */
+	/* Processed all fds or timed out */
 	if (selrtn == 0) {
 		return NT_STATUS_RETRY;
 	}