[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] Deprecating HVM_PARAM_PAE_ENABLED

On Tue, Jul 10, 2018 at 12:45:39PM +0100, Andrew Cooper wrote:
> On 10/07/18 09:04, Wei Liu wrote:
> > On Mon, Jul 09, 2018 at 05:54:44PM +0100, Andrew Cooper wrote:
> >> Hello,
> >>
> >> (I feel as if I've sent this email before, but I can't find any record
> >> of it).
> >>
> >> The parameter was introduced in e661d66f51 (2006) and used in Xen, but
> >> removed from the hypervisor in c/s 5f14a87ce (2008) when CPUID
> >> calculations where moved from the hypervisor into libxc.
> >>
> >> However, the field has always been propagated in the migration stream. 
> >> AFAICT, the only use for the HVM_PARAM is as a function parameter to
> >> xc_cpuid_apply_policy(), which is a very very expensive way of passing a
> >> function parameter!
> >>
> >> Another curiosity is that HVM and PVH guests treat the toplevel xl.cfg
> >> pae boolean differently.  HVM honour the request, while PVH blindly
> >> ignores it and sets it to true.  There is nothing wrong (in principle)
> >> with a PVH non-PAE guest, so the boolean should be honoured IMO.
> > Yes I think it should be honoured in PVH.
> >
> >> A separate usability niggle is that there is no interlink between the
> >> toplevel pae boolean and passing some custom cpuid= configuration.  The
> >> latter appears to take priority but only due to the order in which libxl
> >> processes the data.
> >>
> > I think we either make them interact properly or document which one
> > takes precedence.
> Actually, I'd opt for formally deprecating the toplevel boolean.  We
> will have to leave it working, but we should discourage its use.
> In the past, it was used to hide PSE36 from 32bit HVM guests, as the
> shadow code doesn't support PSE36 (This questionable use has since been
> fixed differently).  It is also my understanding that shadowing 2-on-3
> was easier than shadowing 3-on-3 back in the days of the 32bit
> hypervisor build, so hiding the PAE bit was a legitimate performance
> tradeoff.
> However, PAE is a direct prerequisite for NX and 64bit support, so I
> don't expect anyone has used this option in the past 10 years.

Deprecating it is fine. We still need to document which takes precedence
if user happens to specify top-level option and CPUID policy, or
document its behaviour is implementation defined.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.