mmap performance of reserved top sdram in powerpc platform

[Lonsn]

Hi all:
I have reserved top 128MBytes sdram for PCI DMA usage.  External PCI
master sends data to top sdram and then wake up user space app.
User space use /dev/mem and mmap to access this region and then send
them to network using socket.
I found send speed of memory data in /dev/mem to network is very slow.
(300Mbits/s)
If I malloc memory data in user space and then send to network, the
speed is much higher. (750Mbits/s).
I know the reason is the sdram in /dev/mem marked as un-cached.
How can I set the top sdram as cached to impove performance and be
sure that the data is coherent.
My kernel version is DENX-v2.6.32.

Thanks.