On 31 July 2014 22:41, Arnd Bergmann [off-list ref] wrote:

On Thursday 31 July 2014, Haojian Zhuang wrote:

quoted

Maybe I didn't express this well. Let's check the memory layout in hip04.

<0x00000000-10000000, 0x00000000-c0000000> & <0x00000004-0xc0000000,
0x00000003-40000000> are two memory regions.

These memory address is 64-bit. When ARM_LPAE is disabled,
memblock_add() parses memory base address & size with 32-bit. Since
ARCH_PHYS_ADDR_T_64BIT is highly depend on ARM_LPAE.

Then these two memory regions become <0x10000000, 0xc0000000> &
<0xc0000000, 0xffffffff>. Lots of IO space are in <0xe0000000,
0xefffffff>. So memory space conflicts when ARM_LPAE is disabled.
Kernel panic will come since it believes IO space is memory.

Only two solutions are available in below.
1. Use hip04_defconfig to declare ARM_LPAE.
2. Use hi3xxx_defconfig, and only declare 2.9GB memory in DTS file
without ARM_LPAE.

I think solution #1 is better. What's your opinion?

I think it's a bug in the DT parsing code if incorrect memory
regions get added. It's supposed to parse the memory nodes using
"long enough" (u64 or arbitrary-length) data types and then
skip every range that doesn't fit into phys_addr_t.

        Arnd

OK. I'll check them before invoking memblock_add().

Regards
Haojian