On Wed, Sep 16, 2020 at 09:12:26AM -0300, Jason Gunthorpe wrote: [...] > > You mean the driver uses a different path to the HW which ahs that > > overhead, not that MMIOs have that overhead right ? > > The different path has overhead of doing extra useless MMIOs because > they don't combine For my own information, this is IB user space driver code, correct ? It tries to mmap buffer as WC and if it succeeds write into it in an optimized fashion (that is just pure overhead on platforms where normal NC memory - ie WC on arm64 - does not do what the _architecture_ defines it should). Lorenzo