Re: [PATCH] KVM: VMX: Update instruction length on intercepted BP

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 4425 bytes --]

Gleb Natapov wrote:
> On Sun, Feb 14, 2010 at 06:06:31PM +0100, Jan Kiszka wrote:
>> Gleb Natapov wrote:
>>> On Sun, Feb 14, 2010 at 05:37:39PM +0100, Jan Kiszka wrote:
>>>> Gleb Natapov wrote:
>>>>> On Sun, Feb 14, 2010 at 12:39:14PM +0100, Jan Kiszka wrote:
>>>>>> Gleb Natapov wrote:
>>>>>>> On Sun, Feb 14, 2010 at 11:47:58AM +0100, Jan Kiszka wrote:
>>>>>>>> Gleb Natapov wrote:
>>>>>>>>> On Sun, Feb 14, 2010 at 11:26:31AM +0100, Jan Kiszka wrote:
>>>>>>>>>> Gleb Natapov wrote:
>>>>>>>>>>> On Sat, Feb 13, 2010 at 10:31:12AM +0100, Jan Kiszka wrote:
>>>>>>>>>>>> From: Jan Kiszka <jan.kiszka@siemens.com>
>>>>>>>>>>>>
>>>>>>>>>>>> We intercept #BP while in guest debugging mode. As VM exists due to
>>>>>>>>>>>> intercepted exceptions do not necessarily come with valid
>>>>>>>>>>>> idt_vectoring, we have to update event_exit_inst_len explicitly in such
>>>>>>>>>>>> cases. At least in the absence of migration, this ensures that
>>>>>>>>>>>> re-injections of #BP will find and use the correct instruction length.
>>>>>>>>>>>>
>>>>>>>>>>> event_exit_inst_len is only used for event reinjection. Since event
>>>>>>>>>>> intercepted here will not be reinjected why updating event_exit_inst_len
>>>>>>>>>>> is needed here?
>>>>>>>>>> In guest debugging mode a #BP exception is always reported to user space
>>>>>>>>>> to find out what caused it. If it was the guest itself, the exception is
>>>>>>>>>> reinjected, on older kernels via KVM_SET_GUEST_DEBUG and since 2.6.33
>>>>>>>>>> via KVM_SET_VCPU_EVENTS (the latter requires some qemu patch that I will
>>>>>>>>>> post later).
>>>>>>>>>>
>>>>>>>>>> As we currently do not update event_exit_inst_len on #BP exits,
>>>>>>>>>> reinjecting fails unless event_exit_inst_len happens to be 1 from some
>>>>>>>>>> other exit.
>>>>>>>>>>
>>>>>>>>> Hmm, how does it work on SVM then where we do not have
>>>>>>>>> event_exit_inst_len so execution will resume on the same rip that caused
>>>>>>>>> #BP after event reinjection?
>>>>>>>>>
>>>>>>>> Maybe not at all. I don't think I've tested this scenario on amd so far.
>>>>>>>> Guess it needs some special handling in svm to move rip after the int3
>>>>>>>> when requesting to inject #BP.
>>>>>>>>
>>>>>>> This will work for VMX too, no? So may be we should design something
>>>>>>> that will work for both VMX and SVM before applying patches that make
>>>>>>> oly VMX work?
>>>>>> VMX used to work, so my patch is actually a regression fix. I bet this
>>>>>> was accidentally broken while cleaning up the interrupt handling of VMX.
>>>>>>
>>>>> VMX used to always reexecute instruction.
>>>> ...since 66fd3f7f90. And that was what broke this guest debugging corner
>>>> case.
>>>>
>>> I see. And I see why it worked, but it shouldn't have been working for
>>> SVM. I prefer to look for general solution here that works for SVM/VMX.
>> I don't see the need to emulate INT3 for the sake of unification. VMX
>> works today (with this patch), and SVM might work without further
>> efforts, at least on modern hosts:
>>
>> "Software interrupts cannot be properly injected if the processor does
>> not support the NextRIP field, indicated by EDX[3] = 1 as returned by
>> CPUID function 8000_000A. Hypervisor software should emulate the event
>> injection of software interrupts if NextRIP is not supported."
>>
>> (right below the paragraph I cited before)
>>
>> I assume, INT3 can be considered as software interrupt as well in this
>> context.
>>
> Lets check if SVM works. I can do that if you tell me how.

- Fire up some Linux guest with gdb installed
- Attach gdb to gdbstub of the VM
- Set a soft breakpoint in guest kernel, ideally where it does not
  immediately trigger, e.g. on sys_reboot (use grep sys_reboot
  /proc/kallsyms if you don't have symbols for the guest kernel)
- Start gdb /bin/true in the guest
- run

As gdb sets some automatic breakpoints, this already exercises the
reinjection of #BP.

> My concern is
> that if SVM doesn't work we will write another way to handle this
> situation that will work on SVM and VMX and then we will have code in
> VMX that is not needed, but we have to support it.

I understand your concerns, but I do not share them ATM. I bet we will
need some workaround for older SVM, but it will be some SVM-only thing.
Anyway, we can wait with this patch a few days until we know more.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 257 bytes --]