On Sat, Oct 29, 2011 at 01:04:30PM +0200, Marco Stornelli wrote:About the ioremap problem. It seems there is a problem on some ARM arch
to use ioremap (ramoops driver) to remap a piece of RAM even if it's not
used by kernel (reserved at boot with mem option, Bryan can you
It's all very simple.
We have three major 'memory types' - 'normal memory' which must be used
for things like RAM that we execute code from and use atomic operations
within. This can be prefetched and reordered at will.
'device memory' is for devices, which tighter restrictions on reordering
and guarantees concerning reads and writes.
'strongly ordered memory' is much like device memory.
It is absolutely not permitted to map the same physical addresses with
different types - this is a stronger requirement than getting the cache
attributes the same.
System memory is mapped using 'normal memory' - obviously because we need
to be able to execute code and have working atomic operations throughout
Now, ioremap creates device memory - because its main function is to
dynamically map memory regions in devices.
Now think: if we have system memory mapped as 'normal memory', and then
we try to use ioremap() to remap some of that memory, that will create
a new 'device memory' mapping with the existing 'normal memory' mapping
still present. Now look at the paragraph 'It is absolutely not permitted'
and realise that the requirements for the architecture are being violated
if we permitted this to occur.
Also realise that if that condition is violated, 'unpredictable behaviour'
will occur - not to the extent that the CPU will hang, but it could cause
data errors which could influence overall system stability.
So, the whole idea of using plain ioremap() with system memory is one
that is just completely unsupportable on ARM without _first_ removing
memory from the system mapping, which in turn means updating the page
tables for every task in the system.