Re: [PATCH 00 of 36] x86/paravirt: groundwork for 64-bit Xen support

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
>> "paravirt/x86_64: move __PAGE_OFFSET to leave a space for hypervisor"
>>
>> This moves __PAGE_OFFSET up by 16 GDT slots, from 0xffff810000000000 
>> to 0xffff880000000000.  I have no general justification for this: the 
>> specific reason is that Xen claims the first 16 kernel GDT slots for 
>> itself, and we must move up the mapping to make room.  In the process 
>> I parameterised the compile-time construction of the initial 
>> pagetables in head_64.S to cope with it.
>>     
>
> This reduces native kernel max memory support from around 127 TB to 
> around 120 TB. We also limit the Xen hypervisor to ~7 TB of physical 
> memory - is that wise in the long run? Sure, current CPUs support 40 
> physical bits [1 TB] for now so it's all theoretical at this moment.
>
> my guess is that CPU makers will first extend the physical lines all the 
> way up to 46-47 bits before they are willing to touch the logical model 
> and extend the virtual space beyond 48 bits (47 bits of that available 
> to kernel-space in practice - i.e. 128 TB).
>
> So eventually, in a few years, we'll feel some sort of crunch when the # 
> of physical lines approaches the # of logical bits - just like when 
> 32-bit felt a crunch when physical lines went to 31 and beyond.
>   

There's no inherent reason why Xen itself needs to be able to have all 
memory mapped at once.  32-bit Xen doesn't and can survive quite 
happily.  It's certainly nice to be able to access anything directly, 
but it's just a performance optimisation.  In practice, the guest 
generally has almost everything interesting mapped anyway, and Xen 
maintains a recursive mapping of the pagetable to make its access to the 
pagetable very efficient, so it's only when a hypercall is doing 
something to an unmapped page that there's an issue.

The main limitation the hole-size imposes is the max size of the machine 
to physical map.  That uses 8bytes/page, and reserves 256GB of space for 
it, meaning that the current limit is 2^47 bytes - but there's another 
256GB of reserved and unused space next to it, so that could be easily 
extended to 2^48 if that really becomes an issue.

> That should be fine too - and probably useful for 64-bit kmemcheck 
> support as well.
>
> To further increase the symmetry between 64-bit and 32-bit, could you 
> please also activate the mem=nopentium switch on 64-bit to allow the 
> forcing of a non-PSE native 64-bit bootup? (Obviously not a good idea 
> normally, as it wastes 0.1% of RAM and increases PTE related CPU cache 
> footprint and TLB overhead, but it is useful for debugging.)
>   

OK.  Though it might be an idea to add "nopse" and start deprecating 
nopentium.

> a few other risk areas:
>
> - the vmalloc-sync changes. Are you absolutely sure that it does not
>   matter for performance?
>   

Oh, I didn't mean to include that one.  I think it's probably safe (from 
both the performance and correctness stands), but it's not necessary for 
64-bit Xen.

> - "The 32-bit early_ioremap will work equally well for 64-bit, so just
>    use it." Famous last words ;-)
>
> Anyway, that's all theory - i'll try out your patchset in -tip to see 
> what breaks in practice ;-)
>   

Yep, thanks,

    J