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

Re: [Xen-devel] Regression building HVM domains following "x86: add bitmap of enabled emulated devices"



On Thu, 2015-11-12 at 10:07 +0000, Andrew Cooper wrote:
> On 12/11/15 09:25, Ian Campbell wrote:
> > On Wed, 2015-11-11 at 21:07 +0000, Andrew Cooper wrote:
> > > On 11/11/2015 20:57, Konrad Rzeszutek Wilk wrote:
> > > > On Wed, Nov 11, 2015 at 07:13:25PM +0000, Andrew Cooper wrote:
> > > > > Hello,
> > > > > 
> > > > > Xapi uses the Ocaml stub_xc_domain_create() which uses
> > > > > xc_domain_create().ÂÂxc_domain_create() itself zeros the arch
> > > > > configuration but passes flags straight through.
> > > > Ooops.
> > > > > As a result of c/s 171946ab "x86: add bitmap of enabled emulated
> > > > > devices", xc_domain_create() can no longer be used to construct
> > > > > HVM
> > > > > domains, failing the hypervisor-side sanity check.
> > > > > 
> > > > > Needless to say, this has put a dent in XenServer's automated
> > > > > testing.
> > > > > 
> > > > > 
> > > > > There are a couple of options, but neither of them are fantastic.
> > > > > 
> > > > > 1) Have xc_domain_create() fill in XEN_X86_EMU_ALL based on
> > > > > XEN_DOMCTL_CDF_hvm_guest and XEN_DOMCTL_CDF_pvh_guest
> > > > > 
> > > > > or
> > > > > 
> > > > > 2) Mandate that all callers provide a valid arch configuration,
> > > > > essentially turning xc_domain_create() into
> > > > > xc_domain_create_config()
> > > > > 
> > > > > 
> > > > > Longterm, what is the plan wrt guest construction?ÂÂWith my x86
> > > > > maintainership hat on, I don't want to keep
> > > > > XEN_DOMCTL_CDF_pvh_guest
> > > > > in
> > > > > the interface, so I do not like 1) as an option.
> > > > > 
> > > > > `git grep` indicates that the 3 users of xc_domain_create() are
> > > > > the
> > > > > Ocaml/Python stubs and init-xenstore-domain.c which only
> > > > > constructs a
> > > > > PV
> > > > > guest (which bypasses the issue), whereas libxl uses
> > > > > xc_domain_create_config().ÂÂ(For the python stubs, I expect this
> > > > > will
> > > > > hit Oracle who are still using Xend to my knowledge).
> > > > We are moving to 'xl'.. and there are no Xend bits anymore.
> > > > > Option 2) is a better alternative, but will have a knock-on
> > > > > effect
> > > > > for
> > > > > downstream consumers of the stubs.
> > > > But aren't xc_* calls not-release-stable?
> > > They are indeed not, which offers the option to change the API.
> > > 
> > > > Here is a third idea::
> > > > 
> > > > ÂMake 'xc_domain_create' call 'xc_domain_create_config'. The
> > > > xc_domain_create
> > > > Âwould synthesis the flags and we would put an 'deprecated' flag on
> > > > it
> > > > Â(whatever that means?) and remove 'xc_domain_create' in 4.7?
> > > This is option 1.ÂÂxc_domain_create() already calls
> > > xc_domain_create_config() but with a zeroed arch configuration.
> > > 
> > > The issue is that modifying xc_domain_create() will preclude their
> > > construction of DMLite domains.
> > I think that's ok, we would essentially be declaring that
> > xc_domain_create_config() is the formally supported interface which can
> > do
> > everything while xc_domain_create() is essentially a compat shim which
> > can
> > only do things up to a certain point.
> > 
> > Users who want more would need to switch to the _config variant.
> > 
> > I'm half considering suggesting removing xc_domain_create(), renaming
> > xc_domain_create_config() without the _config() and having it do some
> > compat thing if the passed config==NULL, such that existing callers of
> > xc_domain_create() just need to add NULL to retain their current
> > behaviour.
> 
> That sounds best, with one modification.
> 
> In the case that NULL is passed, there should be some default filled in
> locally (as currently done for ARM).ÂÂFor x86, this should now be flags
> & hvm => X86_EMU_ALL, so they retain their ability to build HMV guests
> by passing a NULL config.

Yes, that was the "some compat thing" I was thinking of.

Ian.


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

 


Rackspace

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