Re: [XEN PATCH 1/7] xen: introduce XENFEAT_xenstore_late_init

[Jan Beulich <jbeulich@xxxxxxxx>]

On 11.01.2022 00:08, Stefano Stabellini wrote:
> On Mon, 10 Jan 2022, Jan Beulich wrote:
>> On 08.01.2022 01:49, Stefano Stabellini wrote:
>>> Introduce a new feature flag to signal that xenstore will not be
>>> immediately available at boot time. Instead, xenstore will become
>>> available later, and a notification of xenstore readiness will be
>>> signalled to the guest using the xenstore event channel.
>>
>> In addition to what Julien has said, I'd like to point out that Linux'es
>> xenbus driver already has means to deal with xenstored not being around
>> right away (perhaps because of living in a stubdom which starts in
>> parallel). I therefore wonder whether what you want can't be achieved
>> entirely inside that driver, without any new feature flag.
> 
> The fewer changes to Linux the better but we couldn't find a way to make
> it work with zero changes.
> 
> In a way, the patch for Linux is re-using the existing infrastructure to
> delay initialization, e.g. xenbus_probe_thread to call xenbus_probe
> later.
> 
> However, today there is nothing stopping Linux/HVM to continue
> initialization immediately except for xs_hvm_defer_init_for_callback
> which is not applicable in this case. So we introduced
> XENFEAT_xenstore_late_init.
> 
> As I wrote in another reply, I think we could do without
> XENFEAT_xenstore_late_init if we instead rely on checking for
> HVM_PARAM_STORE_EVTCHN being valid and HVM_PARAM_STORE_PFN being zero.

Just as an aside - as discussed in some other context not so long ago,
HVM_PARAM_*_PFN being zero isn't the best way of expressing "not yet
initialized", and hence you would also want to check against ~0.

> But either way as far as I can tell we need to add a new check to delay
> xenstore initialization in Linux/HVM.

Yes, I can see that a Linux side change might be needed. But isolating
the change to there will be much better than needing to also have a
Xen side change in place.

Jan