On Thu, 28 Jun 2012, Russell King - ARM Linux wrote:

Err, I don't think you understand what's going on here.

The sequence is:

1. setup the initial mappings so we can run the kernel in virtual space.
2. provide the memory areas to memblock
3. ask the platform to reserve whatever memory it wants from memblock
   [this means using memblock_reserve or arm_memblock_steal).  The
   reserved memory is *not* expected to be mapped at this point, and is
   therefore inaccessible.
4. Setup the lowmem mappings.

I do understand that pretty well so far.

And when we're setting up the lowmem mappings, we do *not* expect to
create any non-section page mappings, which again means we have no reason
to use the memblock allocator to obtain memory that we want to immediately
use.

And why does this have to remain so?

So I don't know where you're claim of being "fragile" is coming from.

It doesn't come from anything you've described so far. It comes from 
those previous attempts at lifting this limitation.  I think that my 
proposal is much less fragile than the other ones.

What is fragile is people wanting to use arm_memblock_steal() without
following the rules for it I layed down.

What about enhancing your rules if the technical limitations they were 
based on are lifted?


Nicolas