Re: [Xen-devel] xc_hvm_inject_trap() races

[Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx>]

On 11/07/2016 04:59 PM, Jan Beulich wrote:
>>>> On 07.11.16 at 15:34, <rcojocaru@xxxxxxxxxxxxxxx> wrote:
>> My proposal was simply something along the lines of:
>>
>> diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
>> index 704fd64..58f5ae4 100644
>> --- a/xen/arch/x86/hvm/hvm.c
>> +++ b/xen/arch/x86/hvm/hvm.c
>> @@ -535,8 +535,22 @@ void hvm_do_resume(struct vcpu *v)
>>      /* Inject pending hw/sw trap */
>>      if ( v->arch.hvm_vcpu.inject_trap.vector != -1 )
>>      {
>> -        hvm_inject_trap(&v->arch.hvm_vcpu.inject_trap);
>> +        unsigned int success = 0;
>> +
>> +        /* Check for already pending interrupts (races). */
>> +        if ( !hvm_event_pending(v) )
>> +        {
>> +            hvm_inject_trap(&v->arch.hvm_vcpu.inject_trap);
>> +            success = 1;
>> +        }
>> +
>>          v->arch.hvm_vcpu.inject_trap.vector = -1;
>> +
>> +        hvm_monitor_injection_result(v->arch.hvm_vcpu.inject_trap.vector,
>> +                                     v->arch.hvm_vcpu.inject_trap.type,
>> +
>> v->arch.hvm_vcpu.inject_trap.error_code,
>> +                                     v->arch.hvm_vcpu.inject_trap.cr2,
>> +                                     success);
>>      }
>>  }
>>
> 
> But you realize that injection isn't really VM event related; it's an
> independent interface. Hence my "too special cased"complaint.
> What if the event I did propose would be a one shot one, which
> you enable right before (or after) doing your event injection (it
> could also be an injection flag)? You'd then see whether the
> next event the vCPU gets is the one you tried to inject.

Not only do I realize that, but the irony is that that's been my initial
reply to Tamas' suggestion. :) Unfortunately after looking carefully at
all the alternatives, this one has turned out to be the simplest and
most effective one - since it's not possible to know if the injection
will succeed after xc_hvm_inject_trap() returns, the only way to know is
if the application can be somehow notified asynchronously when the
hypervisor knows, and the simplest way to do that is via vm_event.

The one-shot vm_event does sound reasonable. I could set a flag
per-VCPU, basically similar to v->arch.hvm_vcpu.inject_trap.vector, and
fire a single event from hvm_inject_trap() if it's set (then unset it) -
the flag would be set via an xc_monitor_next_interrupt() call in libxc.
If nobody objects, I'll test that and see how it goes.


Thanks,
Razvan

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