Re: [PATCH v2 10/23] xen/arm: compile and run xenbus

[Daniel De Graaf <dgdegra@tycho.nsa.gov>]

On 08/08/2012 12:51 PM, Stefano Stabellini wrote:
> On Tue, 7 Aug 2012, Daniel De Graaf wrote:
>> On 08/07/2012 02:21 PM, Konrad Rzeszutek Wilk wrote:
>>> On Mon, Aug 06, 2012 at 03:27:13PM +0100, Stefano Stabellini wrote:
>>>> bind_evtchn_to_irqhandler can legitimately return 0 (irq 0): it is not
>>>> an error.
>>>>
>>>> If Linux is running as an HVM domain and is running as Dom0, use
>>>> xenstored_local_init to initialize the xenstore page and event channel.
>>>>
>>>> Changes in v2:
>>>>
>>>> - refactor xenbus_init.
>>>
>>> Thank you. Lets also CC our friend at NSA who has been doing some work
>>> in that area. Daniel are you OK with this change - will it still make
>>> PV initial domain with with the MiniOS XenBus driver?
>>>
>>> Thanks.
>>
>> That case will work, but what this will break is launching the initial domain
>> with a Xenstore stub domain already running (see below).
>>
>>>>
>>>> Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
>>>> ---
>>>>  drivers/xen/xenbus/xenbus_comms.c |    2 +-
>>>>  drivers/xen/xenbus/xenbus_probe.c |   62 +++++++++++++++++++++++++-----------
>>>>  drivers/xen/xenbus/xenbus_xs.c    |    1 +
>>>>  3 files changed, 45 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/xen/xenbus/xenbus_comms.c b/drivers/xen/xenbus/xenbus_comms.c
>>>> index 52fe7ad..c5aa55c 100644
>>>> --- a/drivers/xen/xenbus/xenbus_comms.c
>>>> +++ b/drivers/xen/xenbus/xenbus_comms.c
>>>> @@ -224,7 +224,7 @@ int xb_init_comms(void)
>>>>  		int err;
>>>>  		err = bind_evtchn_to_irqhandler(xen_store_evtchn, wake_waiting,
>>>>  						0, "xenbus", &xb_waitq);
>>>> -		if (err <= 0) {
>>>> +		if (err < 0) {
>>>>  			printk(KERN_ERR "XENBUS request irq failed %i\n", err);
>>>>  			return err;
>>>>  		}
>>>
>>>> diff --git a/drivers/xen/xenbus/xenbus_probe.c b/drivers/xen/xenbus/xenbus_probe.c
>>>> index b793723..a67ccc0 100644
>>>> --- a/drivers/xen/xenbus/xenbus_probe.c
>>>> +++ b/drivers/xen/xenbus/xenbus_probe.c
>>>> @@ -719,37 +719,61 @@ static int __init xenstored_local_init(void)
>>>>  	return err;
>>>>  }
>>>>  
>>>> +enum xenstore_init {
>>>> +	UNKNOWN,
>>>> +	PV,
>>>> +	HVM,
>>>> +	LOCAL,
>>>> +};
>>>>  static int __init xenbus_init(void)
>>>>  {
>>>>  	int err = 0;
>>>> +	enum xenstore_init usage = UNKNOWN;
>>>> +	uint64_t v = 0;
>>>>  
>>>>  	if (!xen_domain())
>>>>  		return -ENODEV;
>>>>  
>>>>  	xenbus_ring_ops_init();
>>>>  
>>>> -	if (xen_hvm_domain()) {
>>>> -		uint64_t v = 0;
>>>> -		err = hvm_get_parameter(HVM_PARAM_STORE_EVTCHN, &v);
>>>> -		if (err)
>>>> -			goto out_error;
>>>> -		xen_store_evtchn = (int)v;
>>>> -		err = hvm_get_parameter(HVM_PARAM_STORE_PFN, &v);
>>>> -		if (err)
>>>> -			goto out_error;
>>>> -		xen_store_mfn = (unsigned long)v;
>>>> -		xen_store_interface = ioremap(xen_store_mfn << PAGE_SHIFT, PAGE_SIZE);
>>>> -	} else {
>>>> -		xen_store_evtchn = xen_start_info->store_evtchn;
>>>> -		xen_store_mfn = xen_start_info->store_mfn;
>>>> -		if (xen_store_evtchn)
>>>> -			xenstored_ready = 1;
>>>> -		else {
>>>> +	if (xen_pv_domain())
>>>> +		usage = PV;
>>>> +	if (xen_hvm_domain())
>>>> +		usage = HVM;
>>
>> The above is correct for domUs, and is overridden for dom0s:
>>
>>>> +	if (xen_hvm_domain() && xen_initial_domain())
>>>> +		usage = LOCAL;
>>>> +	if (xen_pv_domain() && !xen_start_info->store_evtchn)
>>>> +		usage = LOCAL;
>>
>> Instead of these checks, I think it should just be:
>>
>> if (!xen_start_info->store_evtchn)
>> 	usage = LOCAL;
>>
>> Any domain started after xenstore will have store_evtchn set, so if you don't
>> have this set, you are either going to be running xenstore locally, or will
>> use the ioctl to change it later (and so should still set up everything as if
>> it will be running locally).
> 
> That would be wrong for an HVM dom0 domain (at least on ARM), because
> we don't have a start_info page at all.
> 
> 
>>>> +	if (xen_pv_domain() && xen_start_info->store_evtchn)
>>>> +		xenstored_ready = 1;
>>
>> This part can now just be moved unconditionally into case PV.
> 
> What about:
> 
> if (xen_pv_domain())
>     usage = PV;
> if (xen_hvm_domain())
>     usage = HVM;
> if (!xen_store_evtchn)
>     usage = LOCAL;
> 
> and moving xenstored_ready in case PV, like you suggested.
> 

That looks correct, but you'd need to split up the switch statement in
order to populate xen_store_evtchn before that last condition, which
ends up pretty much eliminating the usage variable.