Is the slow access to memory mapped via remap_pfn_range a MTRR / PAT issue?

Issue

• Slow access to memory mapped via remap_pfn_range
• Customer is using remap_pfn_range within a kernel module and is reporting that regions of a certain size or larger display slow access times.
• We suspect the issue might have to do with caching being disabled on certain regions. The customer's experience is consistent with http://www.linuxquestions.org/questions/showthread.php?p=4952240#post4952240.
• we have a message from our online community concerning access to memory mapped via remap_pfn_range. Here a short description what they do:
  1. at boot time we limit the amount of memory used by Linux (grub parameter mem=...)
  2. we have a small device driver that maps the memory outside the Linux area via remap_pfn_range
• The above works OK in RHEL5 and in RHEL6. But in RHEL6, on some systems, RAM access is much slower (factor [4..5]).
• According to my understanding the slowness sets in when the size of the remapped memory is above 3GB / reaches 4GB:

map area size    Read time (usec)    Write time (usec)
1 GB                  1.22e-3                  1.00e-3
2 GB                  1.22e-3                  1.00e-3
3 GB                  1.22e-3                  1.00e-3
4 GB                  4.81e-2                  2.01e-2

• Is this behaviour to be expected? Can it be influenced by tuning some parameter?