Re: [Qemu-devel] [RFC v5 86/86] 440fx: fix PAM, PCI holes

[Avi Kivity <avi@redhat.com>]

On 07/25/2011 04:07 PM, Anthony Liguori wrote:
> On 07/20/2011 11:50 AM, Avi Kivity wrote:
>> The current implementation of PAM and the PCI holes is broken in several
>> ways:
>>
>>    - PCI BARs are not restricted to the PCI hole (a BAR may hide memory)
>
> Technically, a BAR can be mapped to any non-RAM memory location.

I understood TOM (Top Of Memory) to be fixed - can't find a register for 
it - but maybe I misread the spec.

>
>>    - PCI devices do not respect PAM (if a PCI device maps a region while
>>      PAM maps the region to RAM, the request will be honored)
>
> I assume you mean SMM shadowing, right?  PAM doesn't cover an area 
> that's ever forwarded to the PCI bus.

No, PAM.  And all of the PAM address space is forwarded to the PCI bus 
(the ROMs are satisfied by PIIX which is a PCI device).

Here at least the spec is clear, see 3.2.18.

>
>> This patch fixes things by introducing a pci address space, and using
>> memory region aliases to represent PAM regions, SMRAM, and PCI holes.
>>
>> The memory hierarchy looks something like
>>
>> system_memory
>>   |
>>   +--- low memory alias (0-0xe0000000)
>
> According to the spec, PCI memory doesn't start at e00... but rather 
> at the top of RAM.  In fact, this is what the spec says:
>
> "The address range from the top of main DRAM to 4 Gbytes (top of 
> physical memory space supported by the 440FX PCIset) is normally 
> mapped to PCI. The PMC forwards all accesses within this address range 
> to PCI.
> There are two sub-ranges within this address range defined as APIC 
> Configuration Space and High BIOS Address Range."
>
> So the right thing to do is to forward all accesses from 
> low_memory_memsize ... 4GB to the PCI bus.
>

We could do that.  However what happens if we implement memory hotplug?

Maybe we should implement memory hotplug on a newer chipset anyway.

-- 
error compiling committee.c: too many arguments to function