Re: [Xen-devel] [PATCH 4/8] x86/vm-event/monitor: turn monitor_write_data.do_write into enum

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 04.07.16 at 16:24, <czuzu@xxxxxxxxxxxxxxx> wrote:
> On 7/4/2016 5:08 PM, Jan Beulich wrote:
>>>>> On 04.07.16 at 15:21, <czuzu@xxxxxxxxxxxxxxx> wrote:
>>> On 7/4/2016 4:07 PM, Jan Beulich wrote:
>>>>>>> On 04.07.16 at 14:47, <czuzu@xxxxxxxxxxxxxxx> wrote:
>>>>> On 7/4/2016 3:37 PM, Jan Beulich wrote:
>>>>>>>>> On 30.06.16 at 20:44, <czuzu@xxxxxxxxxxxxxxx> wrote:
>>>>>>> After trapping a control-register write vm-event and -until- deciding 
>>>>>>> if that
>>>>>>> write is to be permitted or not (VM_EVENT_FLAG_DENY) and doing the 
>>>>>>> actual
>>>>>>> write,
>>>>>>> there cannot and should not be another trapped control-register write 
>>>>>>> event.
>>>>>> Is that true even for the case where full register state gets updated
>>>>>> for a vCPU?
>>>>> AFAIK, the full register state cannot be updated _at once_, that is:
>>>>> after each trapped register update monitor_write_data must _always_ be
>>>>> handled _before reentering the vCPU_.
>>>> I'm thinking about operations like VCPUOP_initialise here. Of course
>>>> operations on current can't possibly update more than one register
>>>> at a time.
>>> Yes but those register update operations happen outside the vm-event
>>> subsystem, i.e. in those cases the registers get updated directly, not
>>> by setting bits in monitor_write_data.
>>> Only hypervisor-trapped register updates (e.g. hvm_set_cr0 w/ may_defer
>>> parameter = 1) which are deferred because of hvm_event_crX returning
>>> true use monitor_write_data.
>> That's why I had added "Is that updating-all-context case of no
>> interest to a monitoring application, now and forever?" After all I
>> gave VCPUOP_initialise as an example because the guest itself
>> can invoke it, and so I had assumed this to be of interest to a
>> monitoring app.
> 
> Hmm, didn't have in mind the fact that the guest can do those kind of 
> operations, I suppose that applies to PV guests, right? (in which case 
> this scenario would be triggered by an in-guest PV driver, right?)

A driver wouldn't normally do such things, unless the OS is really
highly modularized (far beyond what Linux allows). But yes, this
is a PV operation, but equally applies to PVHv2 (see commit
192df6f912 ["x86: allow HVM guests to use hypercalls to bring up
vCPUs"]) and by extension/generalization HVM.

Jan


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