On Fri, Jun 08, 2012 at 10:58:50PM +0900, Kim, Jong-Sung wrote:

quoted

From: Minchan Kim [mailto:minchan at kernel.org]
Sent: Tuesday, June 05, 2012 4:12 PM

If we do arm_memblock_steal with a page which is not aligned with section
size, panic can happen during boot by page fault in map_lowmem.

Detail:

1) mdesc->reserve can steal a page which is allocated at 0x1ffff000 by
memblock
   which prefers tail pages of regions.
2) map_lowmem maps 0x00000000 - 0x1fe00000
3) map_lowmem try to map 0x1fe00000 but it's not aligned by section due to

1.

quoted

4) calling alloc_init_pte allocates a new page for new pte by

memblock_alloc

quoted

5) allocated memory for pte is 0x1fffe000 -> it's not mapped yet.
6) memset(ptr, 0, sz) in early_alloc_aligned got PANICed!

May I suggest another simple approach? The first continuous couples of
sections are always safely section-mapped inside alloc_init_section funtion.
So, by limiting memblock_alloc to the end of the first continuous couples of
sections at the start of map_lowmem, map_lowmem can safely memblock_alloc &
memset even if we have one or more section-unaligned memory regions. The
limit can be extended back to arm_lowmem_limit after the map_lowmem is done.

No.  What if the first block of memory is not large enough to handle all
the allocations?

I think the real problem is folk trying to reserve small amounts.  I have
said all reservations must be aligned to 1MB.