Re: [PATCH v2 4/7] x86/vmx: add do_vmtrace_op

[Michał Leszczyński <michal.leszczynski@xxxxxxx>]

----- 22 cze 2020 o 18:16, Jan Beulich jbeulich@xxxxxxxx napisał(a):

> On 22.06.2020 18:02, Michał Leszczyński wrote:
>> ----- 22 cze 2020 o 17:22, Jan Beulich jbeulich@xxxxxxxx napisał(a):
>>> On 22.06.2020 16:35, Michał Leszczyński wrote:
>>>> ----- 22 cze 2020 o 15:25, Jan Beulich jbeulich@xxxxxxxx napisał(a):
>>>>> Is any of what you do in this switch() actually legitimate without
>>>>> hvm_set_vmtrace_pt_size() having got called for the guest? From
>>>>> remarks elsewhere I imply you expect the param that you currently
>>>>> use to be set upon domain creation time, but at the very least the
>>>>> potentially big buffer should imo not get allocated up front, but
>>>>> only when tracing is to actually be enabled.
>>>>
>>>> Wait... so you want to allocate these buffers in runtime?
>>>> Previously we were talking that there is too much runtime logic
>>>> and these enable/disable hypercalls should be stripped to absolute
>>>> minimum.
>>>
>>> Basic arrangements can be made at domain creation time. I don't
>>> think though that it would be a good use of memory if you
>>> allocated perhaps many gigabytes of memory just for possibly
>>> wanting to enable tracing on a guest.
>>>
>> 
>> From our previous conversations I thought that you want to have
>> as much logic moved to the domain creation as possible.
>> 
>> Thus, a parameter "vmtrace_pt_size" was introduced. By default it's
>> zero (= disabled), if you set it to a non-zero value, then trace buffers
>> of given size will be allocated for the domain and you have possibility
>> to use ipt_enable/ipt_disable at any moment.
>> 
>> This way the runtime logic is as thin as possible. I assume user knows
>> in advance whether he/she would want to use external monitoring with IPT
>> or not.
> 
> Andrew - I think you requested movement to domain_create(). Could
> you clarify if indeed you mean to also allocate the big buffers
> this early?
> 
>> It's also not "many gigabytes". In most use cases a buffer of 16/32/64 MB
>> would suffice, I think.
> 
> But that one such buffer per vCPU, isn't it? Plus these buffers
> need to be physically contiguous, which is an additional possibly
> severe constraint.

Yes. For my use case (VMI stuff) I estimate 16-64 MB per vCPU and for fuzzing
I think it would be even less.

And also yes - these buffers need to be physically contigous and aligned
because otherwise CPU would refuse to use them.


Best regards,
Michał Leszczyński
CERT Polska