Re: [Xen-devel] [PATCH v1] x86/mm: Supresses vm_events caused by page-walks

[Tamas K Lengyel <tamas@xxxxxxxxxxxxx>]

On Mon, Oct 30, 2017 at 11:19 AM, Razvan Cojocaru
<rcojocaru@xxxxxxxxxxxxxxx> wrote:
> On 10/30/2017 07:07 PM, Tamas K Lengyel wrote:
>> On Mon, Oct 30, 2017 at 11:01 AM, Razvan Cojocaru
>> <rcojocaru@xxxxxxxxxxxxxxx> wrote:
>>> On 10/30/2017 06:39 PM, Tamas K Lengyel wrote:
>>>> On Mon, Oct 30, 2017 at 10:24 AM, Razvan Cojocaru
>>>> <rcojocaru@xxxxxxxxxxxxxxx> wrote:
>>>>> On 30.10.2017 18:01, Tamas K Lengyel wrote:
>>>>>> On Mon, Oct 30, 2017 at 4:32 AM, Alexandru Isaila
>>>>>> <aisaila@xxxxxxxxxxxxxxx> wrote:
>>>>>>> This patch is adding a way to enable/disable nested pagefault
>>>>>>> events. It introduces the xc_monitor_nested_pagefault function
>>>>>>> and adds the nested_pagefault_disabled in the monitor structure.
>>>>>>> This is needed by the introspection so it will only get gla
>>>>>>> faults and not get spammed with other faults.
>>>>>>> In p2m_set_ad_bits the v->arch.sse_pg_dirty.eip and
>>>>>>> v->arch.sse_pg_dirty.gla are used to mark that this is the
>>>>>>> second time a fault occurs and the dirty bit is set.
>>>>>>
>>>>>> Could you describe under what conditions do you get these other faults?
>>>>>
>>>>> Hey Tamas, the whole story is at page 8 of this document:
>>>>>
>>>>> https://www.researchgate.net/publication/281835515_Proposed_Processor_Extensions_for_Significant_Speedup_of_Hypervisor_Memory_Introspection
>>>>
>>>> Hi Razvan,
>>>> thanks but I'm not sure that doc addresses my question. You
>>>> effectively filter out npfec_kind_in_gpt and npfec_kind_unknown in
>>>> this patch. The first, npfec_kind_in_gpt should only happen if you
>>>> have restricted access to the gpt with ept and the processor couldn't
>>>> walk the table. But if you don't want to get events of these types
>>>> then why not simply not restrict access the gpt to begin with? And as
>>>> for npfec_kind_unknown, I don't think that gets generated under any
>>>> situation. So hence my question, what is your setup that makes this
>>>> patch necessary?
>>>
>>> On the npfec_kind_unknown case, indeed, we were wondering when that
>>> might possibly occur when discussing this patch - it's probably reserved
>>> for the future?
>>>
>>> On why our introspection engine decides to restrict access to those
>>> specific pages, I am not intimate with its inner workings, and not sure
>>> how much could be disclosed here in any case. Is it not a worthwhile
>>> (and otherwise harmless) tool to be able to switch A/D bits-triggered
>>> EPT faults anyway, for introspection purposes?
>>
>> It changes the default behavior of mem_access events so I just wanted
>> to get some background on when that is really required. Technically
>> there is no reason why we couldn't do that filtering in Xen. I think
>> it might be better to flip the filter the other way though so the
>> default behavior remains as is (ie. change the option to enable
>> filtering instead of enabling monitoring).
>
> Wait, it shouldn't change the default behaviour at all. If nobody calls
> that function, all the EPT event kinds should be sent out - the new
> monitor flag is a "disable" flag for non-GLA event (the so-called
> "nested page fault" events).

Oh yea you are right, I completely overlooked that it is named
"nested_pagefault_disabled" =) Maybe a comment in the domctl header
would be warranted to note that this is enabled by default when
mem_access is used.

Tamas

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