Hi Jan,
On 7/4/2016 1:22 PM, Jan Beulich wrote:
On 30.06.16 at 20:43, <cz...@bitdefender.com> wrote:
--- a/xen/arch/x86/hvm/hvm.c
+++ b/xen/arch/x86/hvm/hvm.c
@@ -475,8 +475,6 @@ void hvm_do_resume(struct vcpu *v)
if ( unlikely(v->arch.vm_event) )
{
- struct monitor_write_data *w = &v->arch.vm_event->write_data;
-
if ( v->arch.vm_event->emulate_flags )
{
enum emul_kind kind = EMUL_KIND_NORMAL;
@@ -493,32 +491,10 @@ void hvm_do_resume(struct vcpu *v)
v->arch.vm_event->emulate_flags = 0;
}
-
- if ( w->do_write.msr )
- {
- hvm_msr_write_intercept(w->msr, w->value, 0);
- w->do_write.msr = 0;
- }
-
- if ( w->do_write.cr0 )
- {
- hvm_set_cr0(w->cr0, 0);
- w->do_write.cr0 = 0;
- }
-
- if ( w->do_write.cr4 )
- {
- hvm_set_cr4(w->cr4, 0);
- w->do_write.cr4 = 0;
- }
-
- if ( w->do_write.cr3 )
- {
- hvm_set_cr3(w->cr3, 0);
- w->do_write.cr3 = 0;
- }
}
+ arch_monitor_write_data(v);
Why does this get moved outside the if(), with the same condition
getting added inside the function (inverted for bailing early)?
I left that so because of patch 5/8 - specifically, monitor_write_data
handling shouldn't depend on the vm_event subsystem being initialized.
But you're right, it still does depend on that initialization in this
patch, so I should leave the call inside the if (and remove the check
inside the function) as you suggest and only get it out in 5/8.
Will do that in v2.
@@ -119,6 +156,55 @@ bool_t monitored_msr(const struct domain *d, u32 msr)
return test_bit(msr, bitmap);
}
+static void write_ctrlreg_adjust_traps(struct domain *d)
+{
+ struct vcpu *v;
+ struct arch_vmx_struct *avmx;
+ unsigned int cr3_bitmask;
+ bool_t cr3_vmevent, cr3_ldexit;
+
+ /* Adjust CR3 load-exiting. */
+
+ /* vmx only */
+ ASSERT(cpu_has_vmx);
+
+ /* non-hap domains trap CR3 writes unconditionally */
+ if ( !paging_mode_hap(d) )
+ {
+ for_each_vcpu ( d, v )
+ ASSERT(v->arch.hvm_vmx.exec_control & CPU_BASED_CR3_LOAD_EXITING);
+ return;
+ }
+
+ cr3_bitmask = monitor_ctrlreg_bitmask(VM_EVENT_X86_CR3);
+ cr3_vmevent = !!(d->arch.monitor.write_ctrlreg_enabled & cr3_bitmask);
+
+ for_each_vcpu ( d, v )
+ {
+ avmx = &v->arch.hvm_vmx;
+ cr3_ldexit = !!(avmx->exec_control & CPU_BASED_CR3_LOAD_EXITING);
+
+ if ( cr3_vmevent == cr3_ldexit )
+ continue;
+
+ /*
+ * If CR0.PE=0, CR3 load exiting must remain enabled.
+ * See vmx_update_guest_cr code motion for cr = 0.
+ */
+ if ( cr3_ldexit && !hvm_paging_enabled(v) && !vmx_unrestricted_guest(v)
)
+ continue;
+
+ if ( cr3_vmevent )
+ avmx->exec_control |= CPU_BASED_CR3_LOAD_EXITING;
+ else
+ avmx->exec_control &= ~CPU_BASED_CR3_LOAD_EXITING;
+
+ vmx_vmcs_enter(v);
+ vmx_update_cpu_exec_control(v);
+ vmx_vmcs_exit(v);
+ }
+}
While Razvan gave his ack already, I wonder whether it's really a
good idea to put deeply VMX-specific code outside of a VMX-specific
file.
Jan
Well, a summary of what this function does would sound like: "adjusts
CR3 load-exiting for cr-write monitor vm-events". IMHO that's (monitor)
vm-event specific enough to be placed within the vm-event subsystem.
Could you suggest concretely how this separation would look like? (where
to put this function/parts of it (and what parts), what name should it
have once moved). Another reason this was done (besides avoiding
hackishly doing a CR0 update when we actually need a CR3 update
specifically for a vm-event to happen) is keeping symmetry between
ARM<->X86 in a future patch that would implement monitor CR vm-events
for ARM. In that patch write_ctrlreg_adjust_traps is renamed and
implemented per-architecture, on ARM it would have the same job, i.e.
updating some hypervisor traps (~ vmx execution controls) for CR
vm-events to happen.
On a different note, one thing I forgot to do though is to also move the
following check (instead of completely removing it from
arch_monitor_domctl_event):
if ( VM_EVENT_X86_CR3 == mop->u.mov_to_cr.index )
inside write_ctrlreg_adjust_traps. Will remedy that in v2.
Thanks,
Corneliu.
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel
