Re: [BUG 2.6.31-rc1] HIGHMEM64G causes hang in PCI init on 32-bit x86

["H. Peter Anvin" <hpa@zytor.com>]

Yinghai Lu wrote:
>>>
>> OK.  This is starting to make sense.  I suspect this is a similar issue
>> as 3b0fde0fac19c180317eb0601b3504083f4b9bf5 addresses, which is that the
>> e820 code assumes -- and I don't see any exception to that in
>> 45fbe3ee01b8e463b28c2751b5dcc0cbdc142d90 -- that iomem_resource covers
>> the entire 64-bit address space that e820 knows.  I wonder what happens
>> with "interestingly shaped" memory above 4 GB if resource_size_t is 32
>> bits with that code.
>>
>> In terms of address space assignment, an alternate implementation of the
>> address space cap is to mark it reserved; that would unfortunately
>> result in an ugly turd at the end of /proc/iomem, but that can be
>> addressed if need be, too.
> 
> always enable 64bit resource for 32bit too?
> 

That would address the problem if combined with the "alternative
implementation" that I described below, but I'm not sure how well it
would go over, especially since the 32-bit x86 world is increasingly
getting concentrated on the very-resource-starved end of the computing
spectrum.

The bottom-line problem is the same: e820, and the e820 allocator, can
describe address space that lies outside our real range of possible
address space.  What to do with that is easy -- it should simply be
ignored -- but it does lead to oddball sequencing issues.  In that
sense, reserving a chunk of address space at the end is cleaner, but
that doesn't address the issue of what happens with a 32-bit
resource_size_t.

Unfortunately, my attempts at reproducing the problem locally has failed
so far.

	-hpa

-- 
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.