Re: [Qemu-devel] [PATCH 6/6] pc: limit 64 bit hole to 2G by default

[Michael Roth]

Quoting Andreas Färber (2013-07-28 05:17:47)
> Am 28.07.2013 11:11, schrieb Michael S. Tsirkin:
> > On Sun, Jul 28, 2013 at 10:21:56AM +0200, Igor Mammedov wrote:
> >> On Sun, 28 Jul 2013 10:57:12 +0300
> >> "Michael S. Tsirkin" <address@hidden> wrote:
> >>
> >>> On Sun, Jul 28, 2013 at 09:29:13AM +0200, Igor Mammedov wrote:
> >>>>  
> >>>>      info = g_malloc(sizeof *info);
> >>>> -    info->w32_min = cpu_to_le64(guest_info->pci_info.w32.begin);
> >>>> -    info->w32_max = cpu_to_le64(guest_info->pci_info.w32.end);
> >>>> -    info->w64_min = cpu_to_le64(guest_info->pci_info.w64.begin);
> >>>> -    info->w64_max = cpu_to_le64(guest_info->pci_info.w64.end);
> >>>> +    info->w32_min = cpu_to_le64(object_property_get_int(pci_info,
> >>>> +                                "pci_hole_start", NULL));
> >>>> +    info->w32_max = cpu_to_le64(object_property_get_int(pci_info,
> >>>> +                                "pci_hole_end", NULL));
> >>>
> >>> Looks wrong.
> >>> object_property_get_int returns a signed int64.
> >>> w32 is unsigned.
> >>> Happens to work but I think we need an explicit API.
> 
> That's how QAPI works internally today for any uint64 visitor/property.
> uint64_t is cast to int64_t and back in visitors.
> 
> So I'd hope something like
> uint64_t val = (uint64_t) object_property_get_int()
> would work equally well - CC'ing Michael.

It actually depends on the 'wire'/'serialized' encoding of the underlying
visitor implementation. In this case we're 'serializing' into a QObject
via QmpOutputVisitor, where all integers are actually stored as a
QInt/int64 anyway, so this cast is unavoidable in this case regardless
of what QAPI interface we use: we'll always end up storing as an int64,
requiring us to re-cast to get back to original value. Best we can
achieve is burying the cast deeper (or significantly re-working how
QObject/QInt works)

There is additional bounds checking performed prior to serialization
if the serialized type is less than 64 bits though, so we'd probably
want to add fixed-width accessors if we found ourselves in a situation
where we needed to cast to a smaller datatype than the original.

It's also worth noting that visiting uint64 types using int64 visitor
interfaces isn't universally guaranteed to work: for certain visitor
implementations the serialized encodings may not be compatible with
one another. But in this case we should be good.

> 
> >> I guess it's copy-past error s/cpu_to_le64/cpu_to_le32/
> > 
> > Not these are 64 bit values, but they need to be
> > unsigned not signed.
> > > >> but not need for extra API, with fixed property definition
> >> i.e. s/UINT64/UNIT32/ property set code will take care about limits.
> > 
> > If you replace these with UINT32 you won't be able to
> > specify values >4G.
> > 
> >>> Property names are hard-coded string literals.
> >>> Please add macros to set and get them
> >>> so we can avoid duplicating code.
> >>> E.g.
> >> sure.
> >>
> >>>
> >>> #define PCI_HOST_PROPS...
> >>> static inline get_
> [...]
> >>>> @@ -629,6 +648,15 @@ static const char 
> >>>> *i440fx_pcihost_root_bus_path(PCIHostState *host_bridge,
> >>>>      return "0000";
> >>>>  }
> >>>>  
> >>>> +static Property i440fx_props[] = {
> >>>> +    DEFINE_PROP_UINT64("pci_hole64_start", I440FXState, 
> >>>> pci_info.w64.begin, 0),
> >>>> +    DEFINE_PROP_UINT64("pci_hole64_end", I440FXState, pci_info.w64.end, 
> >>>> 0),
> >>>> +    DEFINE_PROP_UINT64("pci_hole_start", I440FXState, 
> >>>> pci_info.w32.begin, 0),
> >>>> +    DEFINE_PROP_UINT64("pci_hole_end", I440FXState, pci_info.w32.end,
> >>>> +                       IO_APIC_DEFAULT_ADDRESS),
> >>>> +    DEFINE_PROP_END_OF_LIST(),
> >>>> +};
> >>>> +
> >>>
> >>> So we have 4 properties. One of them pci_hole64_end
> >>> is supposed to be set to a value.
> >>> Others should not be touched under any circuimstances.
> >>> Of course if you query properties you have no way
> >>> to know which is which and what are the legal values.
> >>> Ouch.
> >> read-only properties are possible but we would have to drop
> >> usage DEFINE_PROP_UINT64 of and explicitly use only setter in PropertyInfo,
> > 
> > Or add DEFINE_PROP_UINT64_RO for this?
> > 
> >> user better not to touch these properties since they are mostly internal 
> >> API.
> >> but if we say it's internal properties then enforcing read-only might be
> >> overkill.
> >> For user friendly property "pci_hole64_size" would be nice to have.
> > 
> > So at the moment I do
> > 
> > qemu -device i440FX-pcihost,help
> > 
> > and this will get all properties.
> > 
> > If we add some properties that user can not set
> > they should not appear in this output.
> [snip]
> 
> Igor, you can simply use dynamic properties with NULL as setter argument
> for object_property_add*() to achieve that effect.
> 
> Andreas
> 
> -- 
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
> GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg