Re: MMIO ioremap() error with PCI passthrough

[Jeremy Fitzhardinge <jeremy@goop.org>]

Keir Fraser wrote:
> On 1/7/08 14:16, "Andy Burns" <lists.xensource.com@adslpipe.co.uk> wrote:
>
>   
>>> The functions look like they should map the correct range of pages (in this
>>> case only the page covering FEBFF000-FEBFFFFF)
>>>       
>> Built and installed kernel/modules/initrd, rebooted, when it loads the
>> saa7134 driver I see this
>>
>> REMAP: phys=0xfebffc00, len=4096
>> REMAPPFN: addr=0xffffc20000038000, mfn=0xfebff, size=8192
>>
>> So the 1K mapping has been rounded up to 4K (x86_64 page size?) before
>> passing to __ioremap() and then rounded up again to 8K by the time it
>> gets passed to __direct_remap_pfn_range() is that right?
>>     
>
> Well, your analysis is correct, and the size argument to __ioremap() is
> bogus. It shouldn't have been rounded up to 4096 without also rounding down
> the base address. I don't think this would happen with out linux-2.6.18-xen
> tree. In there, ioremap() is defined in include/asm-x86_64/mach-xen/io.h as
> a thin wrapper around __ioremap() which does not modify the size parameter.
>
> So, could be a bug specific to the FC8 kernel. I don't have its sources to
> hand to pinpoint where the size is getting changed from 0x400 to 0x1000. You
> should be able to dig up that detail pretty easily though.

There was a kernel bug in which ioremap would add a 1 page redzone to 
the end of the mapping, and then pass that extra page to iounmap.  It 
has been fixed for a while, but since F8.

    J