Re: [Xen-devel] [PATCH v2 2/2] hvmloader, pci: Don't try to relocate memory if 64-bit BAR is bigger than ~2GB

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 29.09.16 at 01:48, <konrad.wilk@xxxxxxxxxx> wrote:
> @@ -265,11 +266,30 @@ void pci_setup(void)
>              bars[i].devfn   = devfn;
>              bars[i].bar_reg = bar_reg;
>              bars[i].bar_sz  = bar_sz;
> +            bars[i].above_4gb = false;
>  
>              if ( ((bar_data & PCI_BASE_ADDRESS_SPACE) ==
>                    PCI_BASE_ADDRESS_SPACE_MEMORY) ||
>                   (bar_reg == PCI_ROM_ADDRESS) )
> -                mmio_total += bar_sz;
> +            {
> +                /*
> +                 * If bigger than 2GB minus emulated devices BAR space and
> +                 * APIC space, then don't try to put under 4GB.
> +                 */
> +                if ( is_64bar && (mmio_total >= GB(2) || bar_sz >=
> +                     (GB(2) - HVM_BELOW_4G_MMIO_LENGTH - mmio_total)) )

As mentioned in the reply to your earlier mail already, the
subtraction of mmio_total here is risking wrap through zero (the
>= GB(2) check doesn't fully guard against that).

Furthermore you're now making behavior dependent on the order
devices appear on the bus: The same device appearing early may
get its BAR placed below 4Gb whereas when it appears late, it'll
get placed high. IOW I think this needs further refinement: We
should in a first pass place only 32-bit BARs. In a second pass we
can then see which 64-bit BARs still fit (and I think we then ought
to prefer small ones). Which means we should presumably account
32- and 64-bit BARs here independent of any other considerations,
deferring the decision which 64-bit ones to place low until after this
first pass.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel