[linux-sh:03439] sh-linux mmap problem

Hi all,

I have been working on sh4-Linux. I have got some
register (4 Byte each) in my SRAM that is in P2 area (non cacheable).
To Read write
those registers, I have mapped the pages containing registers in user
space. When I access these registers everything goes OK except for 2
registers lodged at address 0xb8000030 and 0xb8000038. This page is
mapped as write only, problem persist even if the page is mapped as
read-write(read is OK write causes problem). Whenever I try to write
anything on these registers, screen goes blank.

When I tried to debug out the problem, I got the following observation:
When I write something(even a byte) on virtual memory, at least 32
bytes of physical memory is refreshed.

say old data is:(L is Location (address), Old and New are Data)
*L1= Old1, *L2= Old2,  *L3=Old3 .....................*L32=Old 32

and I overwrite Old2 with New2, In physical memory L2 = Old2 should happen.

But actual picture is:
*L1 = Old1
*L2 = New2
*L3 = Old3
.
.
.
*L32=Old32

that is all locations are refreshed. If I write some data at
0xb8000030, memory locations from 0xb8000020 to 0xb800003f are
refreshed). Address 0xb8000020 contains the video reset register and
if it is overwritten, even by the same data, video display goes blank.

Please help if anybody can reply.
Code I am using is(for read and write):
[Compiler used is sh4-linux-gcc, generates no warning in the code when
compiled with options -Wall -Wcast-align]

physical address stored in a variable pa = 0xb8000000

{
   fd = open ("/dev/mem", O_RDWR | O_SYNC);
   if (fd == -1) return fd;
   map_base = mmap(0, MAP_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED,
fd, pa & ~MAP_MASK);
   if(map_base[tmp] == (void *) -1)
   {
        fclose(fd);
        return -1;
   }
   addr = map_base + (pa & MAP_MASK);
   buf = (char *)addr;
   buf[30] = 'A';
   msync (addr, MAP_SIZE, MS_SYNC);
   return 0;
}

I have been digging out the sh4 manual to see if there is some setting
problem in hardware.
Data is reaching at proper place (Physical location) because:
1. All other registers are working well.
2. To solve the problem temporarily and continue the development, I
have lodged register (by changing FPGA data) at some other place
(0xa7000030, still in P2 area, still write only registers), and the
board is working well.

But as per final specification and our requirement I need to shift it
back to where they were when problem appeared.
I use /dev/mem for just mapping (physical to user space). I want to
know who copies the data from mapped (virtual) memory actual
(physical) memory. Whether it is some daemon in OS which continuously
copies the data from virtual (mapped) memory to physical whenever
there is some change in virtual memory (Done by OS) or mapping does
some entry in MMU part of CPU so what ever we try to write on mapped
area is written directly over translated (physical) memory locations
(Done by CPU).
Screen still (after changing FPGA data) goes blank when I try to
writing on video reset (0xb8000020) register, or even any memory
location from 0xb8000020 to 0xb800003f.

Being from electronics stream, I'm not very good in operating systems.
I'm just learning OS. Please tell if you know about any such daemon or
the file(s) of kernel responsible for it, so that I may recompile it.
I'm still looking if it is a hardware issue.

One more thing I want to ask is: can I use DMA to transfer this 4 byte
of data, as DMA has direct access to physical address space.

Thanks in advance.

Regards,
Mohit