feature flags use

feature flags use

[Jan Beulich]

Following the need to allocate a new feature flag I soon realized that
currently there is not even a place to store them - this to me is
particularly puzzling for XENFEAT_pae_pgdir_above_4gb, as I would
have expected that the CR3 handling in PAE mode gets adjusted
according to whether the guest has that flag set. Any reason for that
(other than *assuming* that guests without that flag will always pass
the low 12 bits of CR3 clear)? Namely, what's the purpose of having
the guest specify that flag if it doesn't really change anything?

So, in turn, it seems to me that I'll first have to create a domctl to
allow setting feature flags reported by the guest (and manually fill
them for dom0), and the tools then would have to follow. The
problematic part with this is that this must happen early enough to
make sure the hv didn't use any of the flag values, yet - much like
the (un-)setting of the compatibility mode flag, and that the hv has
to verify the call isn't made too late.

Such a new domctl could certainly be combined with the one setting
the mode (native or compat) of the (pv) guest - would that be
considered desirable?

Jan

Re: feature flags use

[Gerd Hoffmann]

  Hi,

> Following the need to allocate a new feature flag I soon realized that
> currently there is not even a place to store them - this to me is
> particularly puzzling for XENFEAT_pae_pgdir_above_4gb, as I would
> have expected that the CR3 handling in PAE mode gets adjusted
> according to whether the guest has that flag set. Any reason for that
> (other than *assuming* that guests without that flag will always pass
> the low 12 bits of CR3 clear)? Namely, what's the purpose of having
> the guest specify that flag if it doesn't really change anything?

Works the other way around:  The flags are xen-specfic, not guest
specific.  The guest has to check whenever xen supports pgd's above 4GB
before placing them there.

cheers,
  Gerd

-- 
Gerd Hoffmann <kraxel@suse.de>
http://www.suse.de/~kraxel/julika-dora.jpeg

Re: feature flags use

[Jan Beulich]

>Works the other way around:  The flags are xen-specfic, not guest
>specific.  The guest has to check whenever xen supports pgd's above 4GB
>before placing them there.

By itself, this makes sense, but what's the purpose of the guest advertising
it supports this via the feature string then?

Jan

Re: feature flags use

[Gerd Hoffmann]

Jan Beulich wrote:
>> Works the other way around:  The flags are xen-specfic, not guest
>> specific.  The guest has to check whenever xen supports pgd's above 4GB
>> before placing them there.
> 
> By itself, this makes sense, but what's the purpose of the guest advertising
> it supports this via the feature string then?

Dunno.  Well, in theory the domain builder would be able to take care
then and allow placing the initial pgd above 4GB only if the guest (and
xen, but that is implicit due to xen and tools must match) can handle
it.  In practice the domain builder doesn't.

cheers,
  Gerd

-- 
Gerd Hoffmann <kraxel@suse.de>
http://www.suse.de/~kraxel/julika-dora.jpeg

Re: feature flags use

[Jan Beulich]

>That flag is treated oddly. Actually we detect if the kernel uses Elf notes
>or whether it has the string PAE=yes[extended]. In either case we notify the
>hypervisor via *vm_assist*! This then decides whether or not we advertise
>the XENFEAT flag to the domU.

I was talking about XENFEAT_pae_pgdir_above_4gb - this flag appears to be
be useless.
Also, regardless of whether you advertise support for the feature to a domU,
the code inside Xen doesn't treat the two cases (kernel with and without this
flag set) differently.

>Anyway, what feature flag do you want to add? Is this to distinguish PV PAE
>guest from PV 64-bit guest?

At this point I was thinking about that hypervisor_virt_start change that the
kernel needs to advertise it can deal with.

Jan

RE: feature flags use

[Ian Pratt]

> >That flag is treated oddly. Actually we detect if the kernel uses Elf
> notes
> >or whether it has the string PAE=yes[extended]. In either case we
notify
> the
> >hypervisor via *vm_assist*! This then decides whether or not we
advertise
> >the XENFEAT flag to the domU.
> 
> I was talking about XENFEAT_pae_pgdir_above_4gb - this flag appears to
be
> be useless.
> Also, regardless of whether you advertise support for the feature to a
> domU,
> the code inside Xen doesn't treat the two cases (kernel with and
without
> this
> flag set) differently.

The xc_linux_restore code in the tools library needs to be able to read
the feature flag from Xen.

Ian