[PATCH] WIP: HACK: LPAE, BOOTMEM and NO_BOOTMEM

[Yinghai Lu]

On Sat, Jun 29, 2013 at 10:57 AM, Tejun Heo <tj at kernel.org> wrote:
> ( Expanding cc list, original thread is at
>   http://thread.gmane.org/gmane.linux.kernel/1518046 )
>
> Hello,
>
> On Sat, Jun 29, 2013 at 06:21:24PM +0100, Russell King - ARM Linux wrote:
>> Unfortunately, that has not been true on ARM - it's very common for
>> there to be an offset on physical memory, sometimes of the order of
>> 3GB or more.  This is because on reset, ARMs start executing the code
>> at physical address zero, which therefore can't be RAM - and there's
>> a desire to avoid complex switching games in hardware to temporarily
>> map ROM there instead of RAM.
>>
>> On these SoCs which Santosh is working on, the main physical memory
>> mapping is above 4GB, with just a small alias below 4GB to allow the
>> system to boot without the MMU being on, as they may have more than
>> 4GB of RAM.  As I understand it, the small alias below 4GB is not
>> suitable for use as a "lowmem" mapping.

is that 32bit ARM or 64bit ARM?

>
> Ah, okay, so the @limit which is in physical address can be over 4GB
> even for lowmem mappings and alloc_bootmem takes them in ulongs,
> urghhh....
>
> Given that still about half of the archs aren't using memblock yet, I
> think there are three options.
>
> 1. Converting all bootmem interface to use physaddr_t.  But that's
>    what memblock is.
>
> 2. Introducing new interface.  Easier right now but the danger there
>    is that it might end up duplicating most of alloc_bootmem()
>    interface anyway and we'll have yet another variant of early mem
>    allocator to enjoy.
>
> 3. Make all generic code use memblock interface instead of bootmem and
>    implement memblock wrapper on archs which don't use memblock yet.
>    We'll probably need to sort out different combinations of
>    HAVE_MEMBLOCK and NO_BOOTMEM.  If this is doable, it probably is
>    the most future proof way.  While it adds new memblock interface
>    built on top of bootmem, it would also allow removing the bootmem
>    interface built on top of memblock - ie. nobootmem.c, which
>    probably is what we should have done from the beginning.
>
> What do you guys think?

2. looks more simple.
but will use alloc_memblock as interface.

We don't need to use __alloc_memory_core_early() directly, right?

Thanks

Yinghai