[RFD] clear virtual machine memory when virtual machine is turned off

[RFD] clear virtual machine memory when virtual machine is turned off

[fei luo]

[-- Attachment #1: Type: text/plain, Size: 972 bytes --]

Hi,

When running the kvm virtual machine in Linux, because the virtual

machine may contain sensitive data, the user may not want these

data to remain in the memory after the virtual machine is turned off.


Although this part of memory will be cleared before being reused by

user-mode programs , But the sensitive data staying in the memory

for a long time will undoubtedly increase the risk of information leakage,

so I wonder whether it is possible to add a flag (like MAP_UNMAPZERO)

to the mmap(2) system call to indicate that the mapped memory needs

to be cleared zero when unmap called or when the program exits.


Of course, the page clear operation not only occurs when unmap called

or program exits, but also need to consider scenes such as page migration,

swap, balloon etc.


When reusing the page that has been cleared, there is no need to clear it

again, which also speeds up the memory allocation of user-mode programs.


Is this feature feasible?

[-- Attachment #2: Type: text/html, Size: 9391 bytes --]

Re: [RFD] clear virtual machine memory when virtual machine is turned off

[David Hildenbrand]

On 02.12.21 11:19, fei luo wrote:
> Hi,
> 
> When running the kvm virtual machine in Linux, because the virtual
> 
> machine may contain sensitive data, the user may not want these
> 
> data to remain in the memory after the virtual machine is turned off.
> 

Hi,

yes, just like if the VM is running.

> 
> Although this part of memory will be cleared before being reused by
> 
> user-mode programs , But the sensitive data staying in the memory
> 
> for a long time will undoubtedly increase the risk of information leakage,
> 
> so I wonder whether it is possible to add a flag (like MAP_UNMAPZERO)
> 
> to the mmap(2) system call to indicate that the mapped memory needs
> 
> to be cleared zero when unmap called or when the program exits.
> 

it's not immediately clear to me why data of user space program #1
should be more important than data of user space program #2 and why the
program should make that decision.

> 
> Of course, the page clear operation not only occurs when unmap called
> 
> or program exits, but also need to consider scenes such as page migration,
> 
> swap, balloon etc.

What about page migration (who clears the old memory location?),
swapping (who clears the swap space, also considering zram?), writeback
(who clears file storage)? Also, as you indicate, MADV_DONTNEED,
MADV_FREE, FALLOC_FL_PUNCH_HOLE would need care ...

To disable swapping you can use mlock(). To handle file storage ...
don't use files. You'd still have to handle any cases where physical
memory locations might be freed and land in the buddy, and for that we
do have ...

> 
> 
> When reusing the page that has been cleared, there is no need to clear it
> 
> again, which also speeds up the memory allocation of user-mode programs.
> 
> 
> Is this feature feasible?

"init_on_free=1" for the system as a whole, which might sounds like what
might tackle part of your use case.

-- 
Thanks,

David / dhildenb

Re: [RFD] clear virtual machine memory when virtual machine is turned off

[Mike Kravetz]

On 12/2/21 04:47, David Hildenbrand wrote:
> On 02.12.21 11:19, fei luo wrote:
>>
>> When reusing the page that has been cleared, there is no need to clear it
>>
>> again, which also speeds up the memory allocation of user-mode programs.
>>
>>
>> Is this feature feasible?
> 
> "init_on_free=1" for the system as a whole, which might sounds like what
> might tackle part of your use case.
> 

Certainly init_on_free will not make much difference if VMs are backed by
hugetlb pages.  We (Joao and myself) have thought about clearing hugetlb
pages from user space in an attempt speed up launching of VMs backed by
hugetlb pages.
-- 
Mike Kravetz

Re: [RFD] clear virtual machine memory when virtual machine is turned off

[David Hildenbrand]

On 02.12.21 18:27, Mike Kravetz wrote:
> On 12/2/21 04:47, David Hildenbrand wrote:
>> On 02.12.21 11:19, fei luo wrote:
>>>
>>> When reusing the page that has been cleared, there is no need to clear it
>>>
>>> again, which also speeds up the memory allocation of user-mode programs.
>>>
>>>
>>> Is this feature feasible?
>>
>> "init_on_free=1" for the system as a whole, which might sounds like what
>> might tackle part of your use case.
>>
> 
> Certainly init_on_free will not make much difference if VMs are backed by
> hugetlb pages.  We (Joao and myself) have thought about clearing hugetlb
> pages from user space in an attempt speed up launching of VMs backed by
> hugetlb pages.
> 

I remember that discussion. And I also recall a discussion regarding
extending init_on_free semantics to hugetlbfs.

-- 
Thanks,

David / dhildenb

Re: [RFD] clear virtual machine memory when virtual machine is turned off

[fei luo]

David Hildenbrand <david@redhat.com> 于2021年12月2日周四 20:47写道：
>
> >
> > Although this part of memory will be cleared before being reused by
> >
> > user-mode programs , But the sensitive data staying in the memory
> >
> > for a long time will undoubtedly increase the risk of information leakage,
> >
> > so I wonder whether it is possible to add a flag (like MAP_UNMAPZERO)
> >
> > to the mmap(2) system call to indicate that the mapped memory needs
> >
> > to be cleared zero when unmap called or when the program exits.
> >
>
> it's not immediately clear to me why data of user space program #1
> should be more important than data of user space program #2 and why the
> program should make that decision.
>

What I mean here is that by adding a flag to the mmap(2) system call to
indicate that this memory will be cleared after munmap() called or
the process exits, and no sensitive data will be left in the system memory,
so as to avoid information leakage after the process exits.And the task
of clearing memory needs to be done in the kernel

> >
> > Of course, the page clear operation not only occurs when unmap called
> >
> > or program exits, but also need to consider scenes such as page migration,
> >
> > swap, balloon etc.
>
> What about page migration (who clears the old memory location?),
> swapping (who clears the swap space, also considering zram?), writeback
> (who clears file storage)? Also, as you indicate, MADV_DONTNEED,
> MADV_FREE, FALLOC_FL_PUNCH_HOLE would need care ...
>
> To disable swapping you can use mlock(). To handle file storage ...
> don't use files. You'd still have to handle any cases where physical
> memory locations might be freed and land in the buddy, and for that we
> do have ...
>

Yes, this feature needs to consider when page migration, the content
of the old page needs to be cleared, and the swap space needs to
be cleared before swap. Of course, for security reasons, swap can be
prohibited. Here I just listed some of the changes involved, not all
aspects. This feature is mainly aimed at clearing the memory of
the virtual machine after shutdown, so it is more aimed at anonymous
mapping and huge page mapping

> >
> >
> > When reusing the page that has been cleared, there is no need to clear it
> >
> > again, which also speeds up the memory allocation of user-mode programs.
> >
> >
> > Is this feature feasible?
>
> "init_on_free=1" for the system as a whole, which might sounds like what
> might tackle part of your use case.
>

This feature is mainly to prevent the used memory information from leaking,
not to clear the memory before use.

--
Thanks

Re: [RFD] clear virtual machine memory when virtual machine is turned off

[David Hildenbrand]

> Yes, this feature needs to consider when page migration, the content
> of the old page needs to be cleared, and the swap space needs to
> be cleared before swap. Of course, for security reasons, swap can be
> prohibited. Here I just listed some of the changes involved, not all
> aspects. This feature is mainly aimed at clearing the memory of
> the virtual machine after shutdown, so it is more aimed at anonymous
> mapping and huge page mapping
> 
>>>
>>>
>>> When reusing the page that has been cleared, there is no need to clear it
>>>
>>> again, which also speeds up the memory allocation of user-mode programs.
>>>
>>>
>>> Is this feature feasible?
>>
>> "init_on_free=1" for the system as a whole, which might sounds like what
>> might tackle part of your use case.
>>
> 
> This feature is mainly to prevent the used memory information from leaking,
> not to clear the memory before use.

That's the whole purpose of init_on_free -- maybe you should give that a
second look.

I don't think MAP_UNMAPZERO is what we want.

-- 
Thanks,

David / dhildenb

[RFD] Clear virtual machine memory when virtual machine is turned off

[fei luo]

Hi,

When running the kvm virtual machine in Linux, because the virtual
machine may contain sensitive data, the user may not want this
data to remain in the memory after the virtual machine is turned off.

Although this part of memory will be cleared before being reused by
user-mode programs , But the sensitive data staying in the memory
for a long time will undoubtedly increase the risk of information leakage,
So I wonder whether it is possible to add a flag (like MAP_UNMAPZERO)
to the mmap(2) system call to indicate that the mapped memory needs
to be cleared zero when munmap(2) called or when the program exits.

Of course, the page clear operation not only occurs when munmap(2)
called or program exits, but also needs to consider scenes such as
page migration, swap, balloon etc.

When reusing the page that has been cleared, there is no need to clear it
again, which also speeds up the memory allocation of user-mode programs.

Is this feature feasible?