We use just single buffer to store name/value headers fields, instead of
allocating memory for each name/value strings.
It is now more than 2 times faster than old one.