Re: [PATCH] kvm-pr: manage illegal instructions

[Thomas Huth]

On 18.05.2016 12:53, Thomas Huth wrote:
> On 18.05.2016 12:18, Thomas Huth wrote:
>> On 17.05.2016 19:49, Laurent Vivier wrote:
>>>
>>>
>>> On 17/05/2016 10:37, Alexander Graf wrote:
>>>> On 05/17/2016 10:35 AM, Laurent Vivier wrote:
>>>>>
>>>>> On 12/05/2016 16:23, Laurent Vivier wrote:
>>>>>>
>>>>>> On 12/05/2016 11:27, Alexander Graf wrote:
>>>>>>> On 05/12/2016 11:10 AM, Laurent Vivier wrote:
>>>>>>>> On 11/05/2016 13:49, Alexander Graf wrote:
>>>>>>>>> On 05/11/2016 01:14 PM, Laurent Vivier wrote:
>>>>>>>>>> On 11/05/2016 12:35, Alexander Graf wrote:
>>>>>>>>>>> On 03/15/2016 09:18 PM, Laurent Vivier wrote:
>>>>>>>>>>>> While writing some instruction tests for kvm-unit-tests for
>>>>>>>>>>>> powerpc,
>>>>>>>>>>>> I've found that illegal instructions are not managed correctly
>>>>>>>>>>>> with
>>>>>>>>>>>> kvm-pr,
>>>>>>>>>>>> while it is fine with kvm-hv.
>>>>>>>>>>>>
>>>>>>>>>>>> When an illegal instruction (like ".long 0") is processed by
>>>>>>>>>>>> kvm-pr,
>>>>>>>>>>>> the kernel logs are filled with:
>>>>>>>>>>>>
>>>>>>>>>>>>          Couldn't emulate instruction 0x00000000 (op 0 xop 0)
>>>>>>>>>>>>          kvmppc_handle_exit_pr: emulation at 700 failed (00000000)
>>>>>>>>>>>>
>>>>>>>>>>>> While the exception handler receives an interrupt for each
>>>>>>>>>>>> instruction
>>>>>>>>>>>> executed after the illegal instruction.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Laurent Vivier <lvivier@xxxxxxxxxx>
>>>>>>>>>>>> ---
>>>>>>>>>>>>      arch/powerpc/kvm/book3s_emulate.c | 4 +++-
>>>>>>>>>>>>      1 file changed, 3 insertions(+), 1 deletion(-)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/arch/powerpc/kvm/book3s_emulate.c
>>>>>>>>>>>> b/arch/powerpc/kvm/book3s_emulate.c
>>>>>>>>>>>> index 2afdb9c..4ee969d 100644
>>>>>>>>>>>> --- a/arch/powerpc/kvm/book3s_emulate.c
>>>>>>>>>>>> +++ b/arch/powerpc/kvm/book3s_emulate.c
>>>>>>>>>>>> @@ -99,7 +99,6 @@ int kvmppc_core_emulate_op_pr(struct kvm_run
>>>>>>>>>>>> *run,
>>>>>>>>>>>> struct kvm_vcpu *vcpu,
>>>>>>>>>>>>            switch (get_op(inst)) {
>>>>>>>>>>>>          case 0:
>>>>>>>>>>>> -        emulated = EMULATE_FAIL;
>>>>>>>>>>>>              if ((kvmppc_get_msr(vcpu) & MSR_LE) &&
>>>>>>>>>>>>                  (inst == swab32(inst_sc))) {
>>>>>>>>>>>>                  /*
>>>>>>>>>>>> @@ -112,6 +111,9 @@ int kvmppc_core_emulate_op_pr(struct kvm_run
>>>>>>>>>>>> *run,
>>>>>>>>>>>> struct kvm_vcpu *vcpu,
>>>>>>>>>>>>                  kvmppc_set_gpr(vcpu, 3, EV_UNIMPLEMENTED);
>>>>>>>>>>>>                  kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
>>>>>>>>>>>>                  emulated = EMULATE_DONE;
>>>>>>>>>>>> +        } else {
>>>>>>>>>>>> +            kvmppc_core_queue_program(vcpu, SRR1_PROGILL);
>>>>>>>>>>> But isn't that exactly what the semantic of EMULATE_FAIL is?
>>>>>>>>>>> Fixing it
>>>>>>>>>>> up in book3s_emulate.c is definitely the wrong spot.
>>>>>>>>>>>
>>>>>>>>>>> So what is the problem you're trying to solve? Is the SRR0 at the
>>>>>>>>>>> wrong
>>>>>>>>>>> spot or are the log messages the problem?
>>>>>>>>>> No, the problem is the host kernel logs are filled by the message
>>>>>>>>>> and
>>>>>>>>>> the execution hangs. And the host becomes unresponsiveness, even
>>>>>>>>>> after
>>>>>>>>>> the end of the tests.
>>>>>>>>>>
>>>>>>>>>> Please, try to run kvm-unit-tests (the emulator test) on a KVM-PR
>>>>>>>>>> host,
>>>>>>>>>> and check the kernel logs (dmesg), then try to ssh to the host...
>>>>>>>>> Ok, so the log messages are the problem. Please fix the message
>>>>>>>>> output
>>>>>>>>> then - or remove it altogether. Or if you like, create a module
>>>>>>>>> parameter that allows you to emit them.
>>>>>>>>>
>>>>>>>>> I personally think the best solution would be to just convert the
>>>>>>>>> message into a trace point.
>>>>>>>>>
>>>>>>>>> While at it, please see whether the guest can trigger similar host
>>>>>>>>> log
>>>>>>>>> output excess in other code paths.
>>>>>>>> The problem is not really with the log messages: they are
>>>>>>>> consequence of
>>>>>>>> the bug I try to fix.
>>>>>>>>
>>>>>>>> What happens is once kvm_pr decodes an invalid instruction all the
>>>>>>>> valid
>>>>>>>> following instructions trigger a Program exception to the guest
>>>>>>>> (but are
>>>>>>>> executed correctly). It has no real consequence on big machine like
>>>>>>>> POWER8, except that the guest become very slow and the log files of
>>>>>>>> the
>>>>>>>> host are filled with messages (and qemu uses 100% of the CPU). On a
>>>>>>>> smaller machine like a  PowerMac G5, the machine becomes simply
>>>>>>>> unusable.
>>>>>>> It's probably more related to your verbosity level of kernel messages.
>>>>>>> If you pass loglevel=0 (or quiet) to you kernel cmdline you won't get
>>>>>>> the messages printed to serial which is what's slowing you down.
>>>>>>>
>>>>>>> The other problem sounds pretty severe, but the only thing your patch
>>>>>>> does any different from the current code flow would be the patch below.
>>>>>>> Or did I miss anything?
>>>>>>>
>>>>>>> diff --git a/arch/powerpc/kvm/emulate.c b/arch/powerpc/kvm/emulate.c
>>>>>>> index 5cc2e7a..4672bc2 100644
>>>>>>> --- a/arch/powerpc/kvm/emulate.c
>>>>>>> +++ b/arch/powerpc/kvm/emulate.c
>>>>>>> @@ -302,7 +302,11 @@ int kvmppc_emulate_instruction(struct kvm_run
>>>>>>> *run,
>>>>>>> struct kvm_vcpu *vcpu)
>>>>>>>                          advance = 0;
>>>>>>>                          printk(KERN_ERR "Couldn't emulate instruction
>>>>>>> 0x%08x "
>>>>>>>                                 "(op %d xop %d)\n", inst, get_op(inst),
>>>>>>> get_xop(inst));
>>>>>>> +#ifdef CONFIG_PPC_BOOK3S
>>>>>>> +                       kvmppc_core_queue_program(vcpu, SRR1_PROGILL);
>>>>>>> +#else
>>>>>>>                          kvmppc_core_queue_program(vcpu, 0);
>>>>>>> +#endif
>>>>>>>                  }
>>>>>>>          }
>>>>>>>
>>>>> Do you want I send an updated patch with your changes?
>>>>
>>>> Well, you reported the issue and narrowed it down, so feel free to send
>>>> it under your name :). I merely simplified your patch a bit.
>>>
>>> Well, while I was trying to update the patch, I've re-tested this... and
>>> it fails. I don't know what I'm doing bad now or what I did bad before
>>> but it seems it doesn't work. :(
>>>
>>> Thomas, could try the patch from Alex?
>>
>> The patch from Alex also does not work for me.
> 
> I've now had a closer look at the code, and I think the endless loop is
> caused by the fact that we try to inject a program interrupt twice here.
> 
> In kvmppc_handle_exit_pr(), the code flow looks like this:
> 
> 	case BOOK3S_INTERRUPT_H_EMUL_ASSIST:
> 	{
> 		...
> 		er = kvmppc_emulate_instruction(run, vcpu);
> 		switch (er) {
> 			...
> 		case EMULATE_FAIL:
> 			printk(KERN_CRIT "%s: emulation at %lx failed (%08x)\n",
> 			       __func__, kvmppc_get_pc(vcpu), last_inst);
> 			kvmppc_core_queue_program(vcpu, flags);
> 			printk(KERN_CRIT "%s: pc = %lx \n",
> 			       __func__, kvmppc_get_pc(vcpu));
> 			r = RESUME_GUEST;
> 			break;
> 		...
> 	}
> 
> But when you look at the end of kvmppc_emulate_instruction(), you
> can see that the interrupt has also already been injected there:
> 
> 	if (emulated == EMULATE_FAIL) {
> 		emulated = vcpu->kvm->arch.kvm_ops->emulate_op(run, vcpu, inst,
> 							       &advance);
> 		if (emulated == EMULATE_AGAIN) {
> 			advance = 0;
> 		} else if (emulated == EMULATE_FAIL) {
> 			advance = 0;
> 			printk(KERN_ERR "Couldn't emulate instruction 0x%08x "
> 			       "(op %d xop %d)\n", inst, get_op(inst), get_xop(inst));
> 			kvmppc_core_queue_program(vcpu, 0);
> 		}
> 	}
> 
> Injecting the program interrupt twice of course destroys the
> return address in SRR0, causing this strange behavior.
> If I comment out the kvmppc_core_queue_program() in
> kvmppc_emulate_instruction(), the endless loop is gone.

I've now also checked the other callers of kvmppc_emulate_instruction(),
and all are already trying to inject an interrupt on their own, so
removing the kvmppc_core_queue_program() from
kvmppc_emulate_instruction() sounds like the right solution for me.
I'll submit a corresponding patch...

 Thomas