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

[Yinghai Lu <yinghai@kernel.org>]

H. Peter Anvin wrote:
> Mikael Pettersson wrote:
>> Thanks, 2.6.31-rc1 vanilla (which didn't boot) plus this one does boot.
>> /proc/iomem now looks as follows:
>>
> 
> ... as it should.  So far so good, and this is a real problem.
> 
> However, there is something that really bothers me: *why does this help
> on Mikael's system, which is PAE and therefore has a 64-bit
> resource_size_t*?  This whole patch should be a no-op!  There is still
> something that doesn't make sense.
> 
> The use of "unsigned long" in ram_alignment() will overflow after 2^52
> bytes, but again, that's not the issue here, since the highest "start"
> value we have is (0x2 << 32).
> 
> By process of elimination, the culprit must be round_up(), which reveals
> that the macro definition of round_up() has a *very* sublte behavior
> with mixed types:
> 
> #define round_up(x, y) (((x) + (y) - 1) & ~((y) - 1))
> 
> ram_alignment() returns unsigned long, which becomes (y).  This means
> that the mask word on the right hand of the & gets truncated to 32 bits
> *before* the masking happens -- since ((y) - 1) is still unsigned long,
> inverting it will not set bits [63..32] to on.
> 
> I think this macro is actively dangerous.  Better would be:
> 
> ({ __typeof__(x) __mask = (y)-1;  ((x)+__mask) & ~__mask; })
> 
> ... which is also multiple-inclusion-free at the cost of using gcc
> ({...}) constructs.
> 
> The deep irony in this is that in our particular case is perhaps that
> align_up(x,y)-1 is the same thing as x | (y-1) which would have avoided
> the problem...

agreed, that is why we change round_up to take u64.

wonder if we should kill round_up and use roundup instead.

in include/linux/kernel.h
#define roundup(x, y) ((((x) + ((y) - 1)) / (y)) * (y))

YH