[Question-kvm] Can hva_to_pfn_fast be executed in interrupt context?

[Question-kvm] Can hva_to_pfn_fast be executed in interrupt context?

[Bharath Vedartham]

Hi all,

I was looking at the function hva_to_pfn_fast(in virt/kvm/kvm_main) which is 
executed in an atomic context(even in non-atomic context, since
hva_to_pfn_fast is much faster than hva_to_pfn_slow).

My question is can this be executed in an interrupt context? 

The motivation for this question is that in an interrupt context, we cannot
assume "current" to be the task_struct of the process of interest.
__get_user_pages_fast assume current->mm when walking the process page
tables. 

So if this function hva_to_pfn_fast can be executed in an
interrupt context, it would not be safe to retrive the pfn with
__get_user_pages_fast. 

Thoughts on this?

Thank you
Bharath

Re: [Question-kvm] Can hva_to_pfn_fast be executed in interrupt context?

[Paolo Bonzini]

On 13/08/19 21:14, Bharath Vedartham wrote:
> Hi all,
> 
> I was looking at the function hva_to_pfn_fast(in virt/kvm/kvm_main) which is 
> executed in an atomic context(even in non-atomic context, since
> hva_to_pfn_fast is much faster than hva_to_pfn_slow).
> 
> My question is can this be executed in an interrupt context? 

No, it cannot for the reason you mention below.

Paolo

> The motivation for this question is that in an interrupt context, we cannot
> assume "current" to be the task_struct of the process of interest.
> __get_user_pages_fast assume current->mm when walking the process page
> tables. 
> 
> So if this function hva_to_pfn_fast can be executed in an
> interrupt context, it would not be safe to retrive the pfn with
> __get_user_pages_fast. 
> 
> Thoughts on this?
> 
> Thank you
> Bharath
> 

Re: [Question-kvm] Can hva_to_pfn_fast be executed in interrupt context?

[Bharath Vedartham]

On Tue, Aug 13, 2019 at 10:17:09PM +0200, Paolo Bonzini wrote:
> On 13/08/19 21:14, Bharath Vedartham wrote:
> > Hi all,
> > 
> > I was looking at the function hva_to_pfn_fast(in virt/kvm/kvm_main) which is 
> > executed in an atomic context(even in non-atomic context, since
> > hva_to_pfn_fast is much faster than hva_to_pfn_slow).
> > 
> > My question is can this be executed in an interrupt context? 
> 
> No, it cannot for the reason you mention below.
> 
> Paolo
hmm.. Well I expected the answer to be kvm specific. 
Because I observed a similar use-case for a driver (sgi-gru) where 
we want to retrive the physical address of a virtual address. This was
done in atomic and non-atomic context similar to hva_to_pfn_fast and
hva_to_pfn_slow. __get_user_pages_fast(for atomic case) 
would not work as the driver could execute in interrupt context.

The driver manually walked the page tables to handle this issue.

Since kvm is a widely used piece of code, I asked this question to know
how kvm handled this issue. 

Thank you for your time.

Thank you
Bharath
> > The motivation for this question is that in an interrupt context, we cannot
> > assume "current" to be the task_struct of the process of interest.
> > __get_user_pages_fast assume current->mm when walking the process page
> > tables. 
> > 
> > So if this function hva_to_pfn_fast can be executed in an
> > interrupt context, it would not be safe to retrive the pfn with
> > __get_user_pages_fast. 
> > 
> > Thoughts on this?
> > 
> > Thank you
> > Bharath
> > 
> 

Re: [Question-kvm] Can hva_to_pfn_fast be executed in interrupt context?

[Bharath Vedartham]

On Thu, Aug 15, 2019 at 08:26:43PM +0200, Paolo Bonzini wrote:
> Oh, I see. Sorry I didn't understand the question. In the case of KVM,
> there's simply no code that runs in interrupt context and needs to use
> virtual addresses.
> 
> In fact, there's no code that runs in interrupt context at all. The only
> code that deals with host interrupts in a virtualization host is in VFIO,
> but all it needs to do is signal an eventfd.
> 
> Paolo
Great, answers my question. Thank you for your time.

Thank you
Bharath
> 
> Il gio 15 ago 2019, 19:18 Bharath Vedartham <linux.bhar@gmail.com> ha
> scritto:
> 
> > On Tue, Aug 13, 2019 at 10:17:09PM +0200, Paolo Bonzini wrote:
> > > On 13/08/19 21:14, Bharath Vedartham wrote:
> > > > Hi all,
> > > >
> > > > I was looking at the function hva_to_pfn_fast(in virt/kvm/kvm_main)
> > which is
> > > > executed in an atomic context(even in non-atomic context, since
> > > > hva_to_pfn_fast is much faster than hva_to_pfn_slow).
> > > >
> > > > My question is can this be executed in an interrupt context?
> > >
> > > No, it cannot for the reason you mention below.
> > >
> > > Paolo
> > hmm.. Well I expected the answer to be kvm specific.
> > Because I observed a similar use-case for a driver (sgi-gru) where
> > we want to retrive the physical address of a virtual address. This was
> > done in atomic and non-atomic context similar to hva_to_pfn_fast and
> > hva_to_pfn_slow. __get_user_pages_fast(for atomic case)
> > would not work as the driver could execute in interrupt context.
> >
> > The driver manually walked the page tables to handle this issue.
> >
> > Since kvm is a widely used piece of code, I asked this question to know
> > how kvm handled this issue.
> >
> > Thank you for your time.
> >
> > Thank you
> > Bharath
> > > > The motivation for this question is that in an interrupt context, we
> > cannot
> > > > assume "current" to be the task_struct of the process of interest.
> > > > __get_user_pages_fast assume current->mm when walking the process page
> > > > tables.
> > > >
> > > > So if this function hva_to_pfn_fast can be executed in an
> > > > interrupt context, it would not be safe to retrive the pfn with
> > > > __get_user_pages_fast.
> > > >
> > > > Thoughts on this?
> > > >
> > > > Thank you
> > > > Bharath
> > > >
> > >
> >