Re: [PATCH] x86: Construct 32 bit boot time page tables in native format.

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

Jeremy Fitzhardinge wrote:
> H. Peter Anvin wrote:
>> No, if Xen wasn't an issue there wouldn't be anything to do for the 
>> PAE case at all (since the PGD is trivial.)
>>
>> Copying PMDs is more or less an analogous case of the !PAE case, once 
>> the allocation is already done.  The allocation should be trivial 
>> though, since this would be a one-time thing.
> 
> I think we're in vehement agreement here.  In either case, its just a 
> matter of something like:
> 
>    memcpy(pgd, &pgd[USER_PTRS_PER_PGD], sizeof(pgd_t) * 
> KERNEL_PTRS_PER_PGD);
>     
> which would work for both PAE and non-PAE.
> 
>>> It would be easy enough to add some code on xen side to look for pmd 
>>> aliases when using/pinning a pagetable, and allocate'n'copy a new pmd 
>>> page as needed.  That way the core code can ignore the issue.
>>
>> As much as I'd rather see Xen fixing this than having it continue to 
>> impact the kernel, I presume it will take some time to flush the 
>> broken hypervisors out? 
> 
> Sorry, I was unclear.  I meant in the purely Xen-specific parts of the 
> kernel (arch/x86/xen).  It wouldn't require a hypervisor change.
> 

Oh, that makes that option much more viable and probably preferrable.

	-hpa