[libvirt] [RFC] 5-level paging Support
Hi all, Now both qemu and kvm support 5-level paging. We can start qemu with a "cpu,+la57" to set 57-bit vitrual address space. So VM can be aware that it need to enable 5-level paging. We can also set another "cpu,phys-bits=52" to set the VM physical address space. Actually, VM can still turn on 5 level paging even without "phys-bits=52", yet this means the guest physical address width are limited, meaning less practical benefits. In to support 5-level paging, I suggest to add two attribute for the domain cpu element in libvirt. <cpu la57='yes', phys-bits='52'\> Here we need to be able to set phys-bits directly, because it is potentially migration sensitive. If la57='no', libvirt will ignore phys-bits, no matter whether hardware support la57. If la57='yes', libvirt will probe the host capability, and will throw error if hardware does support la57. If la57='yes', even phys-bits less than 52, libvirt will also pass it to qemu though less practical benefits. [info] [1] [Qemu-devel] [PATCH]x86: implement la57 paging mode <https://lists.gnu.org/archive/html/qemu-devel/2016-12/msg02096.html> [2] [Qemu-devel] [PATCH v4 2/5] x86: Allow physical address bits to be set <https://lists.gnu.org/archive/html/qemu-devel/2016-07/msg01950.html> BR Shaohe Feng
On Wed, Dec 13, 2017 at 08:43:38PM +0800, Feng, Shaohe wrote:
Hi all,
Now both qemu and kvm support 5-level paging.
We can start qemu with a "cpu,+la57" to set 57-bit vitrual address space.
So VM can be aware that it need to enable 5-level paging.
We can also set another "cpu,phys-bits=52" to set the VM physical address space.
Actually, VM can still turn on 5 level paging even without "phys-bits=52", yet this means
the guest physical address width are limited, meaning less practical benefits.
In to support 5-level paging, I suggest to add two attribute for the domain cpu element in libvirt.
<cpu la57='yes', phys-bits='52'\>
This doesn't make sense - 'la57' is just a CPU feature like any other CPU feature reported in /proc/cpuinfo. IOW, we should just have <cpu> .... <feature name="la57" policy="require"/> </cpu>
If la57='yes', libvirt will probe the host capability, and will throw error if hardware does support la57.
I'm not seeing why we need to special case la57 handling, when we can just use standard CPUID checks we already have
If la57='yes', even phys-bits less than 52, libvirt will also pass it to qemu though less practical benefits.
I would suggest separating la57 from phys-bits, as they're actually completely independant features. They merely happen to be commonly used at the same time. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On 2017年12月13日 21:42, Daniel P. Berrange wrote:
On Wed, Dec 13, 2017 at 08:43:38PM +0800, Feng, Shaohe wrote:
Hi all,
Now both qemu and kvm support 5-level paging.
We can start qemu with a "cpu,+la57" to set 57-bit vitrual address space.
So VM can be aware that it need to enable 5-level paging.
We can also set another "cpu,phys-bits=52" to set the VM physical address space.
Actually, VM can still turn on 5 level paging even without "phys-bits=52", yet this means
the guest physical address width are limited, meaning less practical benefits.
In to support 5-level paging, I suggest to add two attribute for the domain cpu element in libvirt.
<cpu la57='yes', phys-bits='52'\> This doesn't make sense - 'la57' is just a CPU feature like any other CPU feature reported in /proc/cpuinfo. IOW, we should just have
<cpu> .... <feature name="la57" policy="require"/> </cpu>
If la57='yes', libvirt will probe the host capability, and will throw error if hardware does support la57. I'm not seeing why we need to special case la57 handling, when we can just use standard CPUID checks we already have
If la57='yes', even phys-bits less than 52, libvirt will also pass it to qemu though less practical benefits. I would suggest separating la57 from phys-bits, as they're actually completely independant features. They merely happen to be commonly used at the same time. agree, keep them independant.
And still keep phys-bits as attribute of cpu element? <cpu phys-bits='52'> .... <feature name="la57" policy="require"/> </cpu>
Regards, Daniel
BR Shaohe Feng
participants (2)
-
Daniel P. Berrange -
Feng, Shaohe