Re: [PATCH v3 07/32] xen/x86: fix arch_set_info_guest for HVM guests

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 23/07/15 17:56, Roger Pau Monné wrote:
> El 23/07/15 a les 18.19, Jan Beulich ha escrit:
>>>>> On 23.07.15 at 18:15, <andrew.cooper3@citrix.com> wrote:
>>> On 23/07/15 17:00, Ian Campbell wrote:
>>>> On Thu, 2015-07-23 at 17:48 +0200, Roger Pau Monné wrote:
>>>>> El 23/07/15 a les 17.32, Jan Beulich ha escrit:
>>>>>>>>> On 23.07.15 at 17:10, <roger.pau@citrix.com> wrote:
>>>>>>> IMHO introducing a new structure that gets rid of all the PV-only 
>>>>>>>
>>>>>>> fields seems like a good option:
>>>>>>>
>>>>>>> struct vcpu_hvm_context {
>>>>>>> #define _VGCF_online                   5
>>>>>>> #define VGCF_online                    (1<<_VGCF_online)
>>>>>>>     uint32_t flags;                         /* VGCF_* flags      
>>>>>>>            
>>>>>>> */
>>>>>>>     struct cpu_hvm_user_regs user_regs;     /* User-level CPU 
>>>>>>> registers     
>>>>>>> */
>>>>>>>     /* NB. User pagetable on x86/64 is placed in ctrlreg[1]. */
>>>>>>>     uint32_t ctrlreg[8];                    /* CR0-CR7 (control 
>>>>>>> registers)  
>>>>>>> */
>>>>>>>     uint32_t debugreg[8];                   /* DB0-DB7 (debug 
>>>>>>> registers)    
>>>>>>> */
>>>>>>> };
>>>>>>>
>>>>>>> I'm also seriously considering getting rid of ctrlreg and 
>>>>>>> debugreg.
>>>>>>> Since HVM VCPUs will always be started in 32bit flat mode, it 
>>>>>>> doesn't
>>>>>>> make sense IMHO to have both the 32 and the 64 version of the 
>>>>>>> registers, so cpu_hvm_user_regs is always going to be:
>>>>>>>
>>>>>>> struct cpu_hvm_user_regs {
>>>>>>>     uint32_t ebx;
>>>>>>>     uint32_t ecx;
>>>>>>>     uint32_t edx;
>>>>>>>     uint32_t esi;
>>>>>>>     uint32_t edi;
>>>>>>>     uint32_t ebp;
>>>>>>>     uint32_t eax;
>>>>>>>     uint32_t eip;
>>>>>>>     uint32_t esp;
>>>>>>>     uint32_t eflags;
>>>>>>>     uint16_t cs;
>>>>>>>     uint16_t ss;
>>>>>>>     uint16_t es;
>>>>>>>     uint16_t ds;
>>>>>>>     uint16_t fs;
>>>>>>>     uint16_t gs;
>>>>>>> };
>>>>>>>
>>>>>>> We could however do something similar to what's done in ARM and 
>>>>>>> have
>>>>>>> a union of both the 32 and the 64bit registers in case we want to
>>>>>>> start the vCPU in 64bit mode sometime in the future.
>>>>>> What you gave above is suitable only for VCPUOP_initialise afaict.
>>>>>> Did you intend this to be the case?
>>>>> Certainly not, this should also be used by XEN_DOMCTL_setvcpucontext.
>>>>> I'm afraid I'm missing something obvious, but I don't see why this
>>>>> couldn't be used by XEN_DOMCTL_setvcpucontext TBH, can you please 
>>>>> clarify?
>>>> set/getvcpucontext need to pickle all state for save/restore/migration,
>>>> not just the start of day state.
>>> HVM migration doesn't use set/getvcpucontext.
>>>
>>> I would expect an HVMLite-based-PVH to follow suit and just use
>>> set/gethvmcontext.
>> Oh, right, but that doesn't mean the respective domctl-s should
>> produce unconsumable data, or have to guess how to interpret
>> what gets passed in. I.e. at the very minimum the domctl-s then
>> should be disallowed for PVH if they can#t be made work right.
> Right, the problem is that if I introduce this new vcpu_context
> structure plain HVM guests are also going to use it.
>
> This is not a problem with the VCPU_initialize op, because it was not
> available to HVM guests. But it's going to be a problem for
> XEN_DOMCTL_{get/set}vcpucontext consumers, since the structure
> taken/returned by the hypercall is going to change when the target is a
> HVM guest, and this is a can of worms I might prefer to avoid opening.
>
> One partial solution would be to allow the usage of this new struct only
> for the VCPU_initialize hypercall for HVM guests, and keep using the old
> structures for the domctls, aiming at removing them once PV support is
> dropped from Xen, but that looks half-baked IMHO.

PV and HVM "vcpu state" are fundamentally different because they are
fundamentally different kinds of guest.  Lets not try and mash them
together.

I would go with Jan's suggestion and make
XEN_DOMCTL_{get/set}vcpucontext ineligible for use on HVM guests.

The state of an HVMLite vcpu can already be set with
XEN_DOMCTL_sethvmcontext and a stream containing just a single hvm vcpu
record, and retrieved with XEN_DOMCTL_gethvmcontext_partial.

~Andrew

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