Fix concurrent map writes.

This commit should fix an issue with concurrent map writes by making the
opts values map in separate connections not reference the same
underlying data.

Author: bjornouderoelink

conn.go

@@ -298,7 +298,13 @@ func (c *Connector) open(ctx context.Context) (cn *conn, err error) {
 	// the user.
 	defer errRecoverNoErrBadConn(&err)
 
-	o := c.opts
+	// Create a new values map (copy). This makes it so maps in different
+	// connections do not reference the same underlying data structure, so it
+	// is safe for multiple connections to concurrently write to their opts.
+	o := make(values)
+	for k, v := range c.opts {
+		o[k] = v
+	}
 
 	bad := &atomic.Value{}
 	bad.Store(false)

conn_go18.go

@@ -129,7 +129,16 @@ func (cn *conn) watchCancel(ctx context.Context) func() {
 }
 
 func (cn *conn) cancel(ctx context.Context) error {
-	c, err := dial(ctx, cn.dialer, cn.opts)
+	// Create a new values map (copy). This makes sure the connection created
+	// in this method cannot write to the same underlying data, which could
+	// cause a concurrent map write panic. This is necessary because cancel
+	// is called from a goroutine in watchCancel.
+	o := make(values)
+	for k, v := range cn.opts {
+		o[k] = v
+	}
+
+	c, err := dial(ctx, cn.dialer, o)
 	if err != nil {
 		return err
 	}
@@ -142,7 +151,7 @@ func (cn *conn) cancel(ctx context.Context) error {
 			c:   c,
 			bad: bad,
 		}
-		err = can.ssl(cn.opts)
+		err = can.ssl(o)
 		if err != nil {
 			return err
 		}