Re: [Qemu-devel] [PATCH v4] Add machine parameter qemu-kvm-migration for live migrate compatibility with qemu-kvm

From: Markus Armbruster <armbru@redhat.com>
To: Alex Bligh <alex@alex.org.uk>
Cc: "Ryan Harper" <ryan.harper@canonical.com>,
	"Serge Hallyn" <serge.hallyn@canonical.com>,
	"Michael S. Tsirkin" <mst@redhat.com>,
	Libvirt <libvir-list@redhat.com>,
	"Serge Hallyn" <serge.hallyn@ubuntu.com>,
	qemu-devel@nongnu.org, "Alexander Graf" <agraf@suse.de>,
	"quintela@redhat.com" <quintela@redhat.com>,
	"Cole Robinson" <crobinso@redhat.com>,
	"Amit Shah" <amit.shah@redhat.com>,
	"Bruce Rogers" <brogers@suse.com>,
	"Andreas Färber" <afaerber@suse.de>,
	"Serge E. Hallyn" <serge@hallyn.com>
Subject: Re: [Qemu-devel] [PATCH v4] Add machine parameter qemu-kvm-migration for live migrate compatibility with qemu-kvm
Date: Wed, 24 Sep 2014 10:05:19 +0200	[thread overview]
Message-ID: <87a95p78ao.fsf@blackfin.pond.sub.org> (raw)
In-Reply-To: <1411414496-46245-1-git-send-email-alex@alex.org.uk> (Alex Bligh's message of "Mon, 22 Sep 2014 20:34:55 +0100")

Alex Bligh <alex@alex.org.uk> writes:

> This patch series adds inbound migrate capability from qemu-kvm version
> 1.0. The main ideas are those set out in Cole Robinson's patch here:
> http://pkgs.fedoraproject.org/cgit/qemu.git/tree/0001-Fix-migration-from-qemu-kvm.patch?h=f20
> however, rather than patching statically (and breaking inbound
> migration on existing machine types), I have added a new machine
> parameter (qemu-kvm-migration) which when turned on affects the pc-1.0
> machine type. Usage:
>         -machine pc-1.0,qemu-kvm-migration=on

Forgive me if this has been discussed already: why not simply a separate
machine type "pc-1.0-qemu-kvm"?

> Three aproaches are taken:
>
> * cirrus-vga.vgamem_mb defaults to 16 rather than 8. In	order to
>   keep -global cirrus-vga.vgamem_mb working even with
>   qemu-kvm-migration=on, this is monkey-patched	 into the default
>   value			 of the MachineState structure's  compat_props list.

This part fires only for pc-1.0, because it's in
pc_early_init_pci_1_0().

> * In hw/timer/i8254_common.c, the VMSTATE_UINT32_TEST macro
>   is used to test the version for the irq_disable flags,
>   allowing version 3 or more, or version 2 for an inbound
>   migrate from qemu-kvm (only).
>
> * In hw/acpi/piix4.c, qemu-kvm incorrectly uses version 2 for
>   a version 3 structure, causing acpi_load_old to be used.
>   acpi_load_old detects this situation based on the machine type
>   and restarts the attempt to load the vmstate using a
>   customised VMStateDescription. The above cleaner approach is
>   unavailable here.

These parts apply to all machine types, don't they?

> The above monkey-patching must be done between the selection of
> the MachineClass and the processing of the machine parameters
> (on the		 one hand) and the processing of   the compat_props list
> and the		 globals   on the command line. To do this	    I have added
> an earlyinit function to MachineState and QEMUMachine.
>
> I developed this on qemu 2.0 but have forward ported it (trivially)
> to master. My testing has been on a VM live-migrated-to-file from
> Ubuntu Precise qemu-kvm 1.0.
>
> I have given this a moderate degree of testing but it could do
> with more.
>
> Note that certain hardware devices (including QXL) will not
> migrate properly due to a fundamental difference in their internal
> state between versions.
>
> Also note that (as expected) migration from qemu-2.x to qemu-1.0
> will not work, even if the machine types are the same.
>
> Changes since v1:
> * Do not use a machine type, use a machine parameter.

Okay, it has been discussed already.  I'd appreciate a brief recap all
the same.