Re: [Cbe-oss-dev] [patch 3/5] cell: updated driver for DDR2 memory on AXON

[Carsten Otte <cotte@de.ibm.com>]

Arnd Bergmann wrote:
> What will actually happen if you try to mount an axonram device with ext2?
> I suppose mount should fail with a proper error code if the block size is
> larger than 4kb, but does that happen?
> 
> If you have a 4k block size axonram device, the ext2 really should work
> using XIP as expected, including the mmap() operation.
> 
Absolutely. What's the symptom if you try mount -o xip? I am willing 
to fix any bugs that show in the ext2/xip code in that scenario...

>> We have the direct_access method here only because it is needed for the
>> azfs file-system, which we recommend to use for accessing the Axon's RAM
>> rather then ext2 or any other buffered file-systems.
> 
> I think I've understood what the problem is now. The generic xip code
> assumes that the data to be mapped has a 'struct page' in mem_map,
> and that it's part of the linear kernel mapping. This is actually
> true on s390, where it is currently used, but not for us. The point
> on s390 is that the kernel virtual address is _identical_ to the
> physical address, so it may never have shown up as a problem.
> 
> I'll talk to Carsten about this, he already had plans to remove
> the need for struct page from the filemap_xip infrastructure.
Removing the need for struct page is indeed impossible with current 
memory management because of the fact that VM_PFNMAP vmas have the 
downside that they require a linear physical image of the vma. Ext2 
does not comply to that requriement, a given file may be spread all 
over a media.