On Fri, May 28, 2010 at 03:25:34AM -0700, Boris Derzhavets wrote:
> What is an advantage of PV on HVM ?
Pure HVM guests using the Qemu emulated disk/network devices are slow.
PV-on-HVM drivers make disk- and network IO fast for HVM guests.
> Kernel 2.6.34 with Stefano's patches may be built i believe only on Linux
> HVM DomU.
Exactly. They're meant for an upstream kernel, running as Xen HVM guest.
> At the same time any recent Linux ( >=24 or >=26) supports PV guest
> install ( it's in
> mainline for a while).
> What i am missing here ?
HVM guests might be faster for some workloads compared to PV guests.
Kernel compilation could be one example..ie. workloads spawning a lot
of new processes all the time.
> --- On Mon, 5/24/10, Stefano Stabellini <Stefano.Stabellini@xxxxxxxxxxxxx>
> From: Stefano Stabellini <Stefano.Stabellini@xxxxxxxxxxxxx>
> Subject: [Xen-devel] [PATCH 0 of 12] PV on HVM Xen
> To: "linux-kernel@xxxxxxxxxxxxxxx" <linux-kernel@xxxxxxxxxxxxxxx>
> Cc: "Stefano Stabellini" <Stefano.Stabellini@xxxxxxxxxxxxx>, "Jeremy
> Fitzhardinge" <jeremy@xxxxxxxx>, "xen-devel@xxxxxxxxxxxxxxxxxxx"
> <xen-devel@xxxxxxxxxxxxxxxxxxx>, "Don Dutile" <ddutile@xxxxxxxxxx>,
> "Sheng Yang" <sheng@xxxxxxxxxxxxxxx>
> Date: Monday, May 24, 2010, 2:25 PM
> Hi all,
> this is another update of the PV on HVM Xen series that addresses
> Jeremy's comments.
> The platform_pci hooks have been removed, suspend/resume for HVM
> domains is now much more similar to the PV case and shares the same
> do_suspend function.
> Alloc_xen_mmio_hook has been removed has well, now the memory allocation
> the grant table is done by the xen platform pci driver directly.
> The per_cpu xen_vcpu variable is set by a cpu_notifier function so that
> secondary vcpus have the variable set correctly no matter what the xen
> features are on the host.
> The kernel command line option xen_unplug has been renamed to
> xen_emul_unplug and the code that makes use of it has been moved to a
> separate file (arch/x86/xen/platform-pci-unplug.c).
> Xen_unplug_emulated_devices is now able to detect if blkfront, netfront
> and the Xen platform PCI driver have been compiled, and set the default
> value of xen_emul_unplug accordingly.
> The patch "Initialize xenbus device structs with ENODEV as
> default" has been removed from the series and it will be sent
> Finally the comments on most of the patches have been improved.
> The series is based on 2.6.34 and supports Xen PV frontends running
> in a HVM domain, including netfront, blkfront and the VIRQ_TIMER.
> In order to be able to use VIRQ_TIMER and to improve performances you
> need a patch to Xen to implement the vector callback mechanism
> for event channel delivery.
> A git tree is also available here:
> branch name 2.6.34-pvhvm-v2.
> Xen-devel mailing list
> Visible links
> 1. file:///mc/compose?to=Xen-devel@xxxxxxxxxxxxxxxxxxx
> 2. http://lists.xensource.com/xen-devel
> Xen-devel mailing list
Xen-devel mailing list