Fix a race condition in winbind leading to a panic

In winbind, we do multiple events in one select round. This needs fixing, but
as long as we're still using it, for efficiency reasons we need to do that.

What can happen is the following: We have outgoing data pending for a client,
thus

	state->fd_event.flags == EVENT_FD_WRITE

Now a new client comes in, we go through the list of clients to find an idle
one. The detection for idle clients in remove_idle_client does not take the
pending data into account. We close the socket that has pending outgoing data,
the accept(2) one syscall later gives us the same socket.

In new_connection(), we do a setup_async_read, setting up a read fde. The
select from before however had found the socket (that we had already closed!!)
to be writable. In rw_callback we only want to see a readable flag, and we
panic in the SMB_ASSERT(flags == EVENT_FD_READ).

Found using

bin/smbtorture //127.0.0.1/tmp -U% -N 500 -o 2 local-wbclient

Volker

(commit 68c5c6df in master)

diff --git a/source/winbindd/winbindd.c b/source/winbindd/winbindd.c
index f333e3766e00ae31ed6f947c3e16376a048f08b9..9005e6754d911357ab91e7025db547876bf6ba09 100644 (file)
--- a/source/winbindd/winbindd.c
+++ b/source/winbindd/winbindd.c
@@ -778,6 +778,7 @@ static bool remove_idle_client(void)
 
        for (state = winbindd_client_list(); state; state = state->next) {
                if (state->response.result != WINBINDD_PENDING &&
+                   state->fd_event.flags == EVENT_FD_READ &&
                    !state->getpwent_state && !state->getgrent_state) {
                        nidle++;
                        if (!last_access || state->last_access < last_access) {