Don't cache the socket-IO while connection setup

The file_no of the socket IO can change while connecting.
This can happen when alternative hosts are tried,
while GSS authentication
and when falling back to unencrypted in sslmode:prefer .
Therefore expire the socket IO at each connect_poll and reset_poll call.

Caching the IO previosly led to occasional errors kind of:
  Errno::EBADF: Bad file descriptor

With the recreation of an IO object per connect_poll the fileno can change in the TcpGateScheduler when running on Windows.
I didn't dig deeper why this happens, but it fails in spec
  "with a Fiber scheduler connects several times concurrently"
and sometimes in other specs.

ext/pg_connection.c

@@ -475,9 +475,7 @@ pgconn_connect_poll(VALUE self)
 	PostgresPollingStatusType status;
 	status = gvl_PQconnectPoll(pg_get_pgconn(self));
 
-	if ( status == PGRES_POLLING_FAILED ) {
-		pgconn_close_socket_io(self);
-	}
+	pgconn_close_socket_io(self);
 
 	return INT2FIX((int)status);
 }
@@ -556,9 +554,7 @@ pgconn_reset_poll(VALUE self)
 	PostgresPollingStatusType status;
 	status = gvl_PQresetPoll(pg_get_pgconn(self));
 
-	if ( status == PGRES_POLLING_FAILED ) {
-		pgconn_close_socket_io(self);
-	}
+	pgconn_close_socket_io(self);
 
 	return INT2FIX((int)status);
 }

lib/pg/connection.rb

@@ -612,9 +612,6 @@ class PG::Connection
 	alias async_cancel cancel
 
 	private def async_connect_or_reset(poll_meth)
-		# Now grab a reference to the underlying socket so we know when the connection is established
-		socket = socket_io
-
 		# Track the progress of the connection, waiting for the socket to become readable/writable before polling it
 		poll_status = PG::PGRES_POLLING_WRITING
 		until poll_status == PG::PGRES_POLLING_OK ||
@@ -623,11 +620,11 @@ class PG::Connection
 			# If the socket needs to read, wait 'til it becomes readable to poll again
 			case poll_status
 			when PG::PGRES_POLLING_READING
-				socket.wait_readable
+				socket_io.wait_readable
 
 			# ...and the same for when the socket needs to write
 			when PG::PGRES_POLLING_WRITING
-				socket.wait_writable
+				socket_io.wait_writable
 			end
 
 			# Check to see if it's finished or failed yet

sample/async_api.rb

@@ -27,10 +27,6 @@ conn = PG::Connection.connect_start( :dbname => 'test' ) or
 abort "Connection failed: %s" % [ conn.error_message ] if
 	conn.status == PG::CONNECTION_BAD
 
-# Now grab a reference to the underlying socket so we know when the
-# connection is established
-socket = conn.socket_io
-
 # Track the progress of the connection, waiting for the socket to become readable/writable
 # before polling it
 poll_status = PG::PGRES_POLLING_WRITING
@@ -41,13 +37,13 @@ until poll_status == PG::PGRES_POLLING_OK ||
 	case poll_status
 	when PG::PGRES_POLLING_READING
 		output_progress "  waiting for socket to become readable"
-		select( [socket], nil, nil, TIMEOUT ) or
+		select( [conn.socket_io], nil, nil, TIMEOUT ) or
 			raise "Asynchronous connection timed out!"
 
 	# ...and the same for when the socket needs to write
 	when PG::PGRES_POLLING_WRITING
 		output_progress "  waiting for socket to become writable"
-		select( nil, [socket], nil, TIMEOUT ) or
+		select( nil, [conn.socket_io], nil, TIMEOUT ) or
 			raise "Asynchronous connection timed out!"
 	end
 
@@ -85,7 +81,7 @@ loop do
 	# Buffer any incoming data on the socket until a full result is ready.
 	conn.consume_input
 	while conn.is_busy
-		select( [socket], nil, nil, TIMEOUT ) or
+		select( [conn.socket_io], nil, nil, TIMEOUT ) or
 			raise "Timeout waiting for query response."
 		conn.consume_input
 	end

spec/helpers/tcp_gate_scheduler.rb

@@ -277,7 +277,7 @@ class TcpGateScheduler < Scheduler
 
 			# compare and store the fileno for debugging
 			if conn.observed_fd && conn.observed_fd != io.fileno
-				raise "observed fd changed: old:#{conn.observed_fd} new:#{io.fileno}"
+				puts "observed fd changed: old:#{conn.observed_fd} new:#{io.fileno}"
 			end
 			conn.observed_fd = io.fileno
 

spec/pg/connection_spec.rb

@@ -378,6 +378,33 @@ describe PG::Connection do
 			res = @conn2.query("SELECT 4")
 		end
 
+		it "can work with changing IO while connection setup" do
+			# The file_no of the socket IO can change while connecting.
+			# This can happen when alternative hosts are tried,
+			# while GSS authentication
+			# and when falling back to unencrypted in sslmode:prefer
+
+			# Consume some file descriptors and free them while the connection is established.
+			pipes = 100.times.map{ IO.pipe }
+			Thread.new do
+				pipes.reverse_each do |ios|
+					ios.each(&:close)
+					sleep 0.01
+				end
+			end
+
+			# Connect with SSL, but use a wrong client cert, so that connection is aborted.
+			# A second connection is then started with a new IO.
+			# And since the pipes above were freed in the concurrent thread above, there is a high chance that it's a lower file descriptor than before.
+			conn = PG.connect( @conninfo + " sslcert=tmp_test_specs/data/ruby-pg-ca-cert" )
+
+			# The new connection should work even when the file descriptor has changed.
+			res = conn.exec("SELECT 1")
+			expect( res.values ).to eq([["1"]])
+
+			conn.close
+		end
+
 		it "can use conn.reset_start to restart the connection" do
 			ios = IO.pipe
 			conn = described_class.connect_start( @conninfo )