diff mbox

xen/x86: ensure copying to L1 guest in update_runstate_area()

Message ID 20170221021115.24069-1-haozhong.zhang@intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Haozhong Zhang Feb. 21, 2017, 2:11 a.m. UTC
For a HVM domain, if a vcpu is in the nested guest mode,
__raw_copy_to_guest() and __copy_to_guest() used by
update_runstate_area() will copy data to L2 guest other than L1 guest.

Besides copying to the wrong address, this bug also causes crash in
the code path:
    context_switch(prev, next)
      _update_runstate_area(next)
        update_runstate_area(next)
	  __raw_copy_to_guest(...)
	    ...
	      __hvm_copy(...)
	        paging_gva_to_gfn(...)
		  nestedhap_walk_L1_p2m(...)
		    nvmx_hap_walk_L1_p2m(..)
		      vmx_vmcs_enter(v) [ v = next ]
		        vmx_vmcs_try_enter(v) [ v = next ]
			  if ( likely(v == current) )
                              return v->arch.hvm_vmx.vmcs_pa == this_cpu(current_vmcs);
vmx_vmcs_try_enter() will fail and trigger the assert in
vmx_vmcs_enter(), if vcpu 'next' is in the nested guest mode and is
being scheduled to another CPU.

This commit temporally clears the nested guest flag before all
__raw_copy_to_guest() and __copy_to_guest() in update_runstate_area(),
and restores the flag after those guest copy operations.

Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com>
---
 xen/arch/x86/domain.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

Comments

Jan Beulich Feb. 21, 2017, 9:15 a.m. UTC | #1
>>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
> For a HVM domain, if a vcpu is in the nested guest mode,
> __raw_copy_to_guest() and __copy_to_guest() used by
> update_runstate_area() will copy data to L2 guest other than L1 guest.
> 
> Besides copying to the wrong address, this bug also causes crash in
> the code path:
>     context_switch(prev, next)
>       _update_runstate_area(next)
>         update_runstate_area(next)
> 	  __raw_copy_to_guest(...)
> 	    ...
> 	      __hvm_copy(...)
> 	        paging_gva_to_gfn(...)
> 		  nestedhap_walk_L1_p2m(...)
> 		    nvmx_hap_walk_L1_p2m(..)
> 		      vmx_vmcs_enter(v) [ v = next ]
> 		        vmx_vmcs_try_enter(v) [ v = next ]
> 			  if ( likely(v == current) )
>                               return v->arch.hvm_vmx.vmcs_pa == 
> this_cpu(current_vmcs);
> vmx_vmcs_try_enter() will fail and trigger the assert in
> vmx_vmcs_enter(), if vcpu 'next' is in the nested guest mode and is
> being scheduled to another CPU.
> 
> This commit temporally clears the nested guest flag before all
> __raw_copy_to_guest() and __copy_to_guest() in update_runstate_area(),
> and restores the flag after those guest copy operations.

Nice catch, but doesn't the same apply to
update_secondary_system_time() then?

> @@ -1931,10 +1932,29 @@ bool_t update_runstate_area(struct vcpu *v)
>      bool_t rc;
>      smap_check_policy_t smap_policy;
>      void __user *guest_handle = NULL;
> +    bool nested_guest_mode = 0;

false

>      if ( guest_handle_is_null(runstate_guest(v)) )
>          return 1;
>  
> +    /*
> +     * Must be before all following __raw_copy_to_guest() and __copy_to_guest().
> +     *
> +     * Otherwise, if 'v' is in the nested guest mode, paging_gva_to_gfn() called
> +     * from __raw_copy_to_guest() and __copy_to_guest() will treat the target
> +     * address as L2 gva, and __raw_copy_to_guest() and __copy_to_guest() will
> +     * consequently copy runstate to L2 guest other than L1 guest.

... rather than ...

> +     *
> +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
> +     * and __copy_to_guest(), and restore the flag after all guest copy.
> +     */
> +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )

has_hvm_container_vcpu()

And why is this HAP-specific?

> @@ -1971,6 +1991,9 @@ bool_t update_runstate_area(struct vcpu *v)
>  
>      smap_policy_change(v, smap_policy);
>  
> +    if ( nested_guest_mode )

Can we have an unlikely() here please?

Jan
Haozhong Zhang Feb. 22, 2017, 1:28 a.m. UTC | #2
On 02/21/17 02:15 -0700, Jan Beulich wrote:
> >>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
> > For a HVM domain, if a vcpu is in the nested guest mode,
> > __raw_copy_to_guest() and __copy_to_guest() used by
> > update_runstate_area() will copy data to L2 guest other than L1 guest.
> > 
> > Besides copying to the wrong address, this bug also causes crash in
> > the code path:
> >     context_switch(prev, next)
> >       _update_runstate_area(next)
> >         update_runstate_area(next)
> > 	  __raw_copy_to_guest(...)
> > 	    ...
> > 	      __hvm_copy(...)
> > 	        paging_gva_to_gfn(...)
> > 		  nestedhap_walk_L1_p2m(...)
> > 		    nvmx_hap_walk_L1_p2m(..)
> > 		      vmx_vmcs_enter(v) [ v = next ]
> > 		        vmx_vmcs_try_enter(v) [ v = next ]
> > 			  if ( likely(v == current) )
> >                               return v->arch.hvm_vmx.vmcs_pa == 
> > this_cpu(current_vmcs);
> > vmx_vmcs_try_enter() will fail and trigger the assert in
> > vmx_vmcs_enter(), if vcpu 'next' is in the nested guest mode and is
> > being scheduled to another CPU.
> > 
> > This commit temporally clears the nested guest flag before all
> > __raw_copy_to_guest() and __copy_to_guest() in update_runstate_area(),
> > and restores the flag after those guest copy operations.
> 
> Nice catch, but doesn't the same apply to
> update_secondary_system_time() then?
>

Yes, I'll apply the same change to update_secondary_system_time().

> > @@ -1931,10 +1932,29 @@ bool_t update_runstate_area(struct vcpu *v)
> >      bool_t rc;
> >      smap_check_policy_t smap_policy;
> >      void __user *guest_handle = NULL;
> > +    bool nested_guest_mode = 0;
> 
> false

will change

> 
> >      if ( guest_handle_is_null(runstate_guest(v)) )
> >          return 1;
> >  
> > +    /*
> > +     * Must be before all following __raw_copy_to_guest() and __copy_to_guest().
> > +     *
> > +     * Otherwise, if 'v' is in the nested guest mode, paging_gva_to_gfn() called
> > +     * from __raw_copy_to_guest() and __copy_to_guest() will treat the target
> > +     * address as L2 gva, and __raw_copy_to_guest() and __copy_to_guest() will
> > +     * consequently copy runstate to L2 guest other than L1 guest.
> 
> ... rather than ...

ditto

> 
> > +     *
> > +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
> > +     * and __copy_to_guest(), and restore the flag after all guest copy.
> > +     */
> > +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
> 
> has_hvm_container_vcpu()

Nested HVM is only available to HVM domain, so I think is_hvm_vcpu(v) is enough.

> 
> And why is this HAP-specific?
>

IIUC, nested HVM relies on HAP.

> > @@ -1971,6 +1991,9 @@ bool_t update_runstate_area(struct vcpu *v)
> >  
> >      smap_policy_change(v, smap_policy);
> >  
> > +    if ( nested_guest_mode )
> 
> Can we have an unlikely() here please?

will add

Thanks,
Haozhong
Haozhong Zhang Feb. 22, 2017, 2:20 a.m. UTC | #3
On 02/22/17 09:28 +0800, Haozhong Zhang wrote:
> On 02/21/17 02:15 -0700, Jan Beulich wrote:
> > >>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
[..] 
> > > +     *
> > > +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
> > > +     * and __copy_to_guest(), and restore the flag after all guest copy.
> > > +     */
> > > +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
> >

I think it would be clearer to use nestedhvm_enabled() here.

> > has_hvm_container_vcpu()
> 
> Nested HVM is only available to HVM domain, so I think is_hvm_vcpu(v) is enough.
> 
> > 
> > And why is this HAP-specific?
> >
> 
> IIUC, nested HVM relies on HAP.

For nested SVM, I find the following check in hvmop_set_param():
    case HVM_PARAM_NESTEDHVM:
        if ( cpu_has_svm && !paging_mode_hap(d) && a.value )
            rc = -EINVAL;

I don't find the similar check for nested VMX here and in vvmx.c.
Though L1 HVM domain w/ nestedhvm=1 and hap=0 can boot up on Intel
machine (because of the lack of above check?), starting L2 guest does
crash L1 at the very beginning and L0 Xen reports the following debug
messages:

(XEN) realmode.c:111:d18v9 Failed to emulate insn.
(XEN) Real-mode emulation failed: d18v9 Real @ f000:0000fff0 ->
(XEN) domain_crash called from realmode.c:157
(XEN) Domain 18 (vcpu#9) crashed on cpu#29:
(XEN) ----[ Xen-4.9-unstable  x86_64  debug=y   Not tainted ]----
(XEN) CPU:    29
(XEN) RIP:    f000:[<000000000000fff0>]
(XEN) RFLAGS: 0000000000000002   CONTEXT: hvm guest (d18v9)
(XEN) rax: 0000000000000000   rbx: 0000000000000000   rcx: 0000000000000000
(XEN) rdx: 0000000000000f61   rsi: 0000000000000000   rdi: 0000000000000000
(XEN) rbp: 0000000000000000   rsp: 0000000000000000   r8:  0000000000000000
(XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
(XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
(XEN) r15: 0000000000000000   cr0: 0000000000000030   cr4: 0000000000002050
(XEN) cr3: 00000000feffc000   cr2: 0000000000000000
(XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: f000

I haven't dug into this problem, but I suspect there would be other
bugs when using nested VMX w/o HAP. Maybe we should add a similar check in
hvmop_set_param() for nested VMX as well.

Kevin, any comments?

Thanks,
Haozhong
Tian, Kevin Feb. 22, 2017, 3:15 a.m. UTC | #4
> From: Zhang, Haozhong
> Sent: Wednesday, February 22, 2017 10:21 AM
> >
> > >
> > > And why is this HAP-specific?
> > >
> >
> > IIUC, nested HVM relies on HAP.
> 
> For nested SVM, I find the following check in hvmop_set_param():
>     case HVM_PARAM_NESTEDHVM:
>         if ( cpu_has_svm && !paging_mode_hap(d) && a.value )
>             rc = -EINVAL;
> 
> I don't find the similar check for nested VMX here and in vvmx.c.
> Though L1 HVM domain w/ nestedhvm=1 and hap=0 can boot up on Intel
> machine (because of the lack of above check?), starting L2 guest does
> crash L1 at the very beginning and L0 Xen reports the following debug
> messages:
> 
> (XEN) realmode.c:111:d18v9 Failed to emulate insn.
> (XEN) Real-mode emulation failed: d18v9 Real @ f000:0000fff0 ->
> (XEN) domain_crash called from realmode.c:157
> (XEN) Domain 18 (vcpu#9) crashed on cpu#29:
> (XEN) ----[ Xen-4.9-unstable  x86_64  debug=y   Not tainted ]----
> (XEN) CPU:    29
> (XEN) RIP:    f000:[<000000000000fff0>]
> (XEN) RFLAGS: 0000000000000002   CONTEXT: hvm guest (d18v9)
> (XEN) rax: 0000000000000000   rbx: 0000000000000000   rcx: 0000000000000000
> (XEN) rdx: 0000000000000f61   rsi: 0000000000000000   rdi: 0000000000000000
> (XEN) rbp: 0000000000000000   rsp: 0000000000000000   r8:  0000000000000000
> (XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
> (XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
> (XEN) r15: 0000000000000000   cr0: 0000000000000030   cr4: 0000000000002050
> (XEN) cr3: 00000000feffc000   cr2: 0000000000000000
> (XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: f000
> 
> I haven't dug into this problem, but I suspect there would be other
> bugs when using nested VMX w/o HAP. Maybe we should add a similar check in
> hvmop_set_param() for nested VMX as well.
> 
> Kevin, any comments?
> 

I don't recall the initial version of vvmx support - there is a possibility
of not supporting HAP at that time which was introduced later. It's
probably the reason why we don't limit it to HAP=1. But I'm OK to
add similar check now - given HAP=0 is broken and  less efficient
than HAP=1 in nested situation.

Thanks
Kevin
Jan Beulich Feb. 22, 2017, 7:46 a.m. UTC | #5
>>> On 22.02.17 at 03:20, <haozhong.zhang@intel.com> wrote:
> On 02/22/17 09:28 +0800, Haozhong Zhang wrote:
>> On 02/21/17 02:15 -0700, Jan Beulich wrote:
>> > >>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
> [..] 
>> > > +     *
>> > > +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
>> > > +     * and __copy_to_guest(), and restore the flag after all guest copy.
>> > > +     */
>> > > +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
>> >
> 
> I think it would be clearer to use nestedhvm_enabled() here.

Indeed - that would eliminate all these open coding of assumption
which may be valid at present, but not down the road.

>> > And why is this HAP-specific?
>> >
>> 
>> IIUC, nested HVM relies on HAP.
> 
> For nested SVM, I find the following check in hvmop_set_param():
>     case HVM_PARAM_NESTEDHVM:
>         if ( cpu_has_svm && !paging_mode_hap(d) && a.value )
>             rc = -EINVAL;
> 
> I don't find the similar check for nested VMX here and in vvmx.c.
> Though L1 HVM domain w/ nestedhvm=1 and hap=0 can boot up on Intel
> machine (because of the lack of above check?), starting L2 guest does
> crash L1 at the very beginning and L0 Xen reports the following debug
> messages:
> 
> (XEN) realmode.c:111:d18v9 Failed to emulate insn.
> (XEN) Real-mode emulation failed: d18v9 Real @ f000:0000fff0 ->
> (XEN) domain_crash called from realmode.c:157
> (XEN) Domain 18 (vcpu#9) crashed on cpu#29:
> (XEN) ----[ Xen-4.9-unstable  x86_64  debug=y   Not tainted ]----
> (XEN) CPU:    29
> (XEN) RIP:    f000:[<000000000000fff0>]
> (XEN) RFLAGS: 0000000000000002   CONTEXT: hvm guest (d18v9)
> (XEN) rax: 0000000000000000   rbx: 0000000000000000   rcx: 0000000000000000
> (XEN) rdx: 0000000000000f61   rsi: 0000000000000000   rdi: 0000000000000000
> (XEN) rbp: 0000000000000000   rsp: 0000000000000000   r8:  0000000000000000
> (XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
> (XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
> (XEN) r15: 0000000000000000   cr0: 0000000000000030   cr4: 0000000000002050
> (XEN) cr3: 00000000feffc000   cr2: 0000000000000000
> (XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: f000

Now that's of course quite odd: The instruction at that address
ought to be a direct far branch, which I think the emulator is able
to deal with. Which leaves the possibility of it fetching the wrong
bytes (which sadly don't get shown above).

Jan
Haozhong Zhang Feb. 23, 2017, 8 a.m. UTC | #6
On 02/22/17 00:46 -0700, Jan Beulich wrote:
> >>> On 22.02.17 at 03:20, <haozhong.zhang@intel.com> wrote:
> > On 02/22/17 09:28 +0800, Haozhong Zhang wrote:
> >> On 02/21/17 02:15 -0700, Jan Beulich wrote:
> >> > >>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
> > [..] 
> >> > > +     *
> >> > > +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
> >> > > +     * and __copy_to_guest(), and restore the flag after all guest copy.
> >> > > +     */
> >> > > +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
> >> >
> > 
> > I think it would be clearer to use nestedhvm_enabled() here.
> 
> Indeed - that would eliminate all these open coding of assumption
> which may be valid at present, but not down the road.
> 
> >> > And why is this HAP-specific?
> >> >
> >> 
> >> IIUC, nested HVM relies on HAP.
> > 
> > For nested SVM, I find the following check in hvmop_set_param():
> >     case HVM_PARAM_NESTEDHVM:
> >         if ( cpu_has_svm && !paging_mode_hap(d) && a.value )
> >             rc = -EINVAL;
> > 
> > I don't find the similar check for nested VMX here and in vvmx.c.
> > Though L1 HVM domain w/ nestedhvm=1 and hap=0 can boot up on Intel
> > machine (because of the lack of above check?), starting L2 guest does
> > crash L1 at the very beginning and L0 Xen reports the following debug
> > messages:
> > 
> > (XEN) realmode.c:111:d18v9 Failed to emulate insn.
> > (XEN) Real-mode emulation failed: d18v9 Real @ f000:0000fff0 ->
> > (XEN) domain_crash called from realmode.c:157
> > (XEN) Domain 18 (vcpu#9) crashed on cpu#29:
> > (XEN) ----[ Xen-4.9-unstable  x86_64  debug=y   Not tainted ]----
> > (XEN) CPU:    29
> > (XEN) RIP:    f000:[<000000000000fff0>]
> > (XEN) RFLAGS: 0000000000000002   CONTEXT: hvm guest (d18v9)
> > (XEN) rax: 0000000000000000   rbx: 0000000000000000   rcx: 0000000000000000
> > (XEN) rdx: 0000000000000f61   rsi: 0000000000000000   rdi: 0000000000000000
> > (XEN) rbp: 0000000000000000   rsp: 0000000000000000   r8:  0000000000000000
> > (XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
> > (XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
> > (XEN) r15: 0000000000000000   cr0: 0000000000000030   cr4: 0000000000002050
> > (XEN) cr3: 00000000feffc000   cr2: 0000000000000000
> > (XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: f000
> 
> Now that's of course quite odd: The instruction at that address
> ought to be a direct far branch, which I think the emulator is able
> to deal with. Which leaves the possibility of it fetching the wrong
> bytes (which sadly don't get shown above).

Yes, the emulator fails at fetching the instruction from L2
guest. It's at the vm entry to L2 guest after handling the L1
vmlaunch. The code path from vmx_do_vmentry is shown as below:
    vmx_asm_do_vmentry
    vmx_realmode
    vmx_realmode_emulate_one
    hvm_emulate_one
    _hvm_emulate_one
    x86_emulate
    x86_decode
    insn_fetch_type
    ...
    hvmemul_insn_fetch
    __hvmemul_read
    hvm_fetch_from_guest_linear
    __hvm_copy(addr = 0xfffffff0, flags = HVMCOPY_from_guest | HVMCOPY_linear)
    3117    if ( flags & HVMCOPY_linear )
            {
    3119        gfn = paging_gva_to_gfn(v, addr, &pfec);
                ...
    3133        gpa |= gfn << PAGE_SHIHT;
            }
            ...
    3151    page = get_page_from_gfn(v->domain, gfn, &p2mt, P2M_UNSHARE);
    3153    if ( !page )
    3154        return HVMCOPY_bad_gfn_to_mfn;

When hap is not enabled, paging_gva_to_gfn at line 3119 uses hostp2m
to translate the *L2* gpa rather than walking through the nested
EPT, so line 3119 actually translates from *L1* gpa 0xfffffff0 to L1
gfn 0xfffff. The p2m type at gfn 0xfffff is p2m_mmio_dm, so
get_page_from_gfn() at line 3151 fails.

I don't know whether nested VMX with non-hap ever worked or not, but
it does not now. I'll include a patch to enforce the hap requirement
for nested vmx.

Haozhong
Andrew Cooper Feb. 23, 2017, 1:08 p.m. UTC | #7
On 23/02/17 08:00, Haozhong Zhang wrote:
> On 02/22/17 00:46 -0700, Jan Beulich wrote:
>>>>> On 22.02.17 at 03:20, <haozhong.zhang@intel.com> wrote:
>>> On 02/22/17 09:28 +0800, Haozhong Zhang wrote:
>>>> On 02/21/17 02:15 -0700, Jan Beulich wrote:
>>>>>>>> On 21.02.17 at 03:11, <haozhong.zhang@intel.com> wrote:
>>> [..] 
>>>>>> +     *
>>>>>> +     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
>>>>>> +     * and __copy_to_guest(), and restore the flag after all guest copy.
>>>>>> +     */
>>>>>> +    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
>>> I think it would be clearer to use nestedhvm_enabled() here.
>> Indeed - that would eliminate all these open coding of assumption
>> which may be valid at present, but not down the road.
>>
>>>>> And why is this HAP-specific?
>>>>>
>>>> IIUC, nested HVM relies on HAP.
>>> For nested SVM, I find the following check in hvmop_set_param():
>>>     case HVM_PARAM_NESTEDHVM:
>>>         if ( cpu_has_svm && !paging_mode_hap(d) && a.value )
>>>             rc = -EINVAL;
>>>
>>> I don't find the similar check for nested VMX here and in vvmx.c.
>>> Though L1 HVM domain w/ nestedhvm=1 and hap=0 can boot up on Intel
>>> machine (because of the lack of above check?), starting L2 guest does
>>> crash L1 at the very beginning and L0 Xen reports the following debug
>>> messages:
>>>
>>> (XEN) realmode.c:111:d18v9 Failed to emulate insn.
>>> (XEN) Real-mode emulation failed: d18v9 Real @ f000:0000fff0 ->
>>> (XEN) domain_crash called from realmode.c:157
>>> (XEN) Domain 18 (vcpu#9) crashed on cpu#29:
>>> (XEN) ----[ Xen-4.9-unstable  x86_64  debug=y   Not tainted ]----
>>> (XEN) CPU:    29
>>> (XEN) RIP:    f000:[<000000000000fff0>]
>>> (XEN) RFLAGS: 0000000000000002   CONTEXT: hvm guest (d18v9)
>>> (XEN) rax: 0000000000000000   rbx: 0000000000000000   rcx: 0000000000000000
>>> (XEN) rdx: 0000000000000f61   rsi: 0000000000000000   rdi: 0000000000000000
>>> (XEN) rbp: 0000000000000000   rsp: 0000000000000000   r8:  0000000000000000
>>> (XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
>>> (XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
>>> (XEN) r15: 0000000000000000   cr0: 0000000000000030   cr4: 0000000000002050
>>> (XEN) cr3: 00000000feffc000   cr2: 0000000000000000
>>> (XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: f000
>> Now that's of course quite odd: The instruction at that address
>> ought to be a direct far branch, which I think the emulator is able
>> to deal with. Which leaves the possibility of it fetching the wrong
>> bytes (which sadly don't get shown above).
> Yes, the emulator fails at fetching the instruction from L2
> guest. It's at the vm entry to L2 guest after handling the L1
> vmlaunch. The code path from vmx_do_vmentry is shown as below:
>     vmx_asm_do_vmentry
>     vmx_realmode
>     vmx_realmode_emulate_one
>     hvm_emulate_one
>     _hvm_emulate_one
>     x86_emulate
>     x86_decode
>     insn_fetch_type
>     ...
>     hvmemul_insn_fetch
>     __hvmemul_read
>     hvm_fetch_from_guest_linear
>     __hvm_copy(addr = 0xfffffff0, flags = HVMCOPY_from_guest | HVMCOPY_linear)
>     3117    if ( flags & HVMCOPY_linear )
>             {
>     3119        gfn = paging_gva_to_gfn(v, addr, &pfec);
>                 ...
>     3133        gpa |= gfn << PAGE_SHIHT;
>             }
>             ...
>     3151    page = get_page_from_gfn(v->domain, gfn, &p2mt, P2M_UNSHARE);
>     3153    if ( !page )
>     3154        return HVMCOPY_bad_gfn_to_mfn;
>
> When hap is not enabled, paging_gva_to_gfn at line 3119 uses hostp2m
> to translate the *L2* gpa rather than walking through the nested
> EPT, so line 3119 actually translates from *L1* gpa 0xfffffff0 to L1
> gfn 0xfffff. The p2m type at gfn 0xfffff is p2m_mmio_dm, so
> get_page_from_gfn() at line 3151 fails.
>
> I don't know whether nested VMX with non-hap ever worked or not, but
> it does not now. I'll include a patch to enforce the hap requirement
> for nested vmx.

This answers one of my open tickets (although being shadow-related, it
wasn't the highest priority item to look at).

Nested VMX definitely used to work with shadow paging, but I am not
surprised if it has bitrotted in the meantime.  One larger area I have
identified in the current swamp of nested-virt problems is that the
current p2m infrastructure makes it far too easy to make mistakes like
this (e.g. c/s bab2bd8e2)

My idea (probably needs refining) to address the issue is to introduce
new TYPE_SAFE()'s for l2 and l2 gfns, and use the compiler to make it
harder to get this wrong.

~Andrew
diff mbox

Patch

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index 7d3071e..5f0444c 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -50,6 +50,7 @@ 
 #include <asm/mpspec.h>
 #include <asm/ldt.h>
 #include <asm/hvm/hvm.h>
+#include <asm/hvm/nestedhvm.h>
 #include <asm/hvm/support.h>
 #include <asm/hvm/viridian.h>
 #include <asm/debugreg.h>
@@ -1931,10 +1932,29 @@  bool_t update_runstate_area(struct vcpu *v)
     bool_t rc;
     smap_check_policy_t smap_policy;
     void __user *guest_handle = NULL;
+    bool nested_guest_mode = 0;
 
     if ( guest_handle_is_null(runstate_guest(v)) )
         return 1;
 
+    /*
+     * Must be before all following __raw_copy_to_guest() and __copy_to_guest().
+     *
+     * Otherwise, if 'v' is in the nested guest mode, paging_gva_to_gfn() called
+     * from __raw_copy_to_guest() and __copy_to_guest() will treat the target
+     * address as L2 gva, and __raw_copy_to_guest() and __copy_to_guest() will
+     * consequently copy runstate to L2 guest other than L1 guest.
+     *
+     * Therefore, we clear the nested guest flag before __raw_copy_to_guest()
+     * and __copy_to_guest(), and restore the flag after all guest copy.
+     */
+    if ( is_hvm_vcpu(v) && paging_mode_hap(v->domain) )
+    {
+        nested_guest_mode = nestedhvm_is_n2(v);
+        if ( nested_guest_mode )
+            nestedhvm_vcpu_exit_guestmode(v);
+    }
+
     smap_policy = smap_policy_change(v, SMAP_CHECK_ENABLED);
 
     if ( VM_ASSIST(v->domain, runstate_update_flag) )
@@ -1971,6 +1991,9 @@  bool_t update_runstate_area(struct vcpu *v)
 
     smap_policy_change(v, smap_policy);
 
+    if ( nested_guest_mode )
+        nestedhvm_vcpu_enter_guestmode(v);
+
     return rc;
 }