On 14/02/18 18:22, Andrew Cooper wrote:
> On 14/02/18 16:10, Alexandru Stefan ISAILA wrote:
>> On Lu, 2018-02-12 at 15:54 +0000, Andrew Cooper wrote:
>>> On 12/02/18 15:08, Alexandru Isaila wrote:
>>>> @@ -2619,14 +2634,31 @@ void svm_vmexit_handler(struct
>>>> cpu_user_regs *regs)
>>>> case VMEXIT_EXCEPTION_BP:
>>>> - if ( !v->domain->debugger_attached )
>>>> - goto unexpected_exit_type;
>>>> - /* AMD Vol2, 15.11: INT3, INTO, BOUND intercepts do not
>>>> update RIP. */
>>>> - if ( (inst_len = __get_instruction_length(v, INSTR_INT3))
>>>> == 0 )
>>>> + inst_len = __get_instruction_length(v, INSTR_INT3);
>>> There are multiple ways of ending up with this vmexit, and INT3 is
>>> the only way.
>>> The old code was somewhat broken (but only in the case that a
>>> was attached), but now with this introspection hook active,
>>> `0xcd 0x03` will end up crashing the domain because of a length
>>> looking for 0xcc.
>>> You need to inspect EXITINTINFO to work out what went on here, and
>>> distinguish INT3 from INT $3.
>>> Can I suggest that you run this unit test
>>> http://xenbits.xen.org/docs/xtf/test-swint-emulation.html under debug
>>> introspection an check that you get all expected events? Every time
>>> touch this code, we seem to break it :(
>> I've tested on Intel and AMD and I only get events on int3. Further
>> more, I don't think there is any way to use the vmcb->exitintinfo
>> because all the fields are 0 on the time of VMEXIT_EXCEPTION_BP. Did I
>> understand the test scenario correctly?
> Quite possibly, but now I'm even more confused. I'll have a quick play.
Ok - after some investigation, executing `int $3` triggers VMEXIT_SWINT,
with the vector in EXITINFO1, as opposed to triggering VMEXIT_EXCP3,
except that we don't have INTERCEPT_SWINT active, so it completes
Therefore, in your patch, we do expect only ever to find an int3
triggering VMEXIT_EXCEPTION_BP. Sorry for the noise.
However, do you mind rebasing the remainder of your series onto
staging? It doesn't apply cleanly any more.
Xen-devel mailing list