New VT-D spec available

New VT-D spec available

[Nakajima, Jun]

An updated specification (Rev 1.0 update) of Intel(r) virtualization
technology for Directed I/O is now available on Intel web site: 

http://download.intel.com/technology/computing/vptech/Intel(r)_VT_for_Di
rect_IO.pdf 

The specification update includes DMA remapping capability, Interrupt
remapping capability, queue management, invalidation information and
enhancements.

We are thinking how this should be enabled for KVM as well especially
when assigning real devices to guests. Suggestions are welcome.

Jun
---
Intel Open Source Technology Center

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: New VT-D spec available

[Avi Kivity]

Nakajima, Jun wrote:
> An updated specification (Rev 1.0 update) of Intel(r) virtualization
> technology for Directed I/O is now available on Intel web site: 
>
> http://download.intel.com/technology/computing/vptech/Intel(r)_VT_for_Di
> rect_IO.pdf 
>
> The specification update includes DMA remapping capability, Interrupt
> remapping capability, queue management, invalidation information and
> enhancements.
>
> We are thinking how this should be enabled for KVM as well especially
> when assigning real devices to guests. Suggestions are welcome.
>   

I've had a quick look-see. There are two general approaches which I 
think make sense:

- tightly coupled kvm integration, where kvm is extended with a 
root-only 'assign pci dev x to this vm'.  As the page table format is 
the same as used by ept (I'm guessing... haven't actually read the ept 
spec), this should be quite easy to do on once npt/ept support is 
integrated.

- general interface for manipulating VT-d, with the ability to use it 
with kvm.  With this, it is possible to assign pci devices to ordinary 
userspace processes, some of which could be hosting virtual machines.  
This may fit will with the UIO patchset.

Unless I misread things, VT-d does not honor the U/S bit on the page 
table entries.  That makes the general approach fairly difficult.  I 
would recommend starting with the tightly-coupled, less generic approach 
first, and generalizing it if/when there is demand for it.

-- 
error compiling committee.c: too many arguments to function


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: New VT-D spec available

[Nakajima, Jun]

Avi Kivity wrote:
> Nakajima, Jun wrote:
> > An updated specification (Rev 1.0 update) of Intel(r) virtualization
> > technology for Directed I/O is now available on Intel web site:
> > 
> >
http://download.intel.com/technology/computing/vptech/Intel(r)_VT_for_Di
> > rect_IO.pdf 
> > 
> > The specification update includes DMA remapping capability,
Interrupt
> > remapping capability, queue management, invalidation information and
> > enhancements. 
> > 
> > We are thinking how this should be enabled for KVM as well
especially
> > when assigning real devices to guests. Suggestions are welcome.
> > 
> 
> I've had a quick look-see. There are two general approaches which I
> think make sense:
> 
> - tightly coupled kvm integration, where kvm is extended with a
> root-only 'assign pci dev x to this vm'.  As the page table format is
> the same as used by ept (I'm guessing... haven't actually read the ept
> spec), this should be quite easy to do on once npt/ept support is
> integrated.

I agree that it's straightforward to handle those page tables. The
question is interrupt remapping, i.e. how to inject particular external
interupts to the vm that is reposible for the interrupts without
impacting the kernel code. We are considering the irq chip infrastrure.

> 
> - general interface for manipulating VT-d, with the ability to use it
> with kvm.  With this, it is possible to assign pci devices to ordinary
> userspace processes, some of which could be hosting virtual machines.
> This may fit will with the UIO patchset.
> 
> Unless I misread things, VT-d does not honor the U/S bit on the page
> table entries.  That makes the general approach fairly difficult.  I
> would recommend starting with the tightly-coupled, less generic
approach
> first, and generalizing it if/when there is demand for it.

I agree that we should start with the tightly-coupled.

Jun
---
Intel Open Source Technology Center

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: New VT-D spec available

[Avi Kivity]

Nakajima, Jun wrote:
> I agree that it's straightforward to handle those page tables. The
> question is interrupt remapping, i.e. how to inject particular external
> interupts to the vm that is reposible for the interrupts without
> impacting the kernel code. We are considering the irq chip infrastrure.
>
>   

So Windows guests would be handled by the regular apic emulation,
without direct injection?



-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: New VT-D spec available

[Nakajima, Jun]

Avi Kivity wrote:
> Nakajima, Jun wrote:
> > I agree that it's straightforward to handle those page tables. The
> > question is interrupt remapping, i.e. how to inject particular
external
> > interupts to the vm that is reposible for the interrupts without
> > impacting the kernel code. We are considering the irq chip
infrastrure.
> > 
> > 
> 
> So Windows guests would be handled by the regular apic emulation,
> without direct injection?

Direct injection is almost impossible to handle because such guests can
be scheduled out. And they cannot get other irrelevant interrutps. So
(at least part of) such APIC emulation should be in the kernel.

Jun
---
Intel Open Source Technology Center

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/