On Tue, Apr 28, 2020 at 01:47:24PM -0400, Michael S. Tsirkin wrote:
> On Tue, Apr 28, 2020 at 06:29:20PM +0100, Stefan Hajnoczi wrote:
> > On Wed, Apr 22, 2020 at 09:13:35PM -0700, elena.ufimtseva@oracle.com wrote:
> > > We will post separate patchsets for the following improvements for
> > > the experimental Qemu multi-process:
> > > - Live migration;
> > > - Asynchronous communication channel;
> > > - Libvirt support;
> > >
> > > We welcome all your ideas, concerns, and questions for this patchset.
> >
> > This patch series does two things:
> > 1. It introduces the remote device infrastructure.
> > 2. It creates the remote device program and the associated build changes
> > (makefiles, stubs, etc).
> >
> > There are many patches and it's likely that a bunch more revisions will
> > be necessary before this can be merged.
> >
> > I want to share an idea to reduce the scope and get patches merged more
> > quickly. It looks like the series can be reduced to 21 patches using
> > this approach.
> >
> > I suggest dropping the remote device program from this patch series (and
> > maybe never bringing it back). Instead, use the softmmu target for the
> > remote device.
> >
> > Why? Because the remote device program is just a QEMU that uses the
> > remote machine type and has no vCPUs:
> >
> > $ qemu-system-x86_64 -chardev id=char0,... \
> > -M remote,chardev=char0 \
> > -device lsi53c810 \
> > -drive if=none,id=drive0,file=vm.img,format=raw \
> > -device scsi-hd,drive=drive0
> >
> > This will use the remote machine type, interrupt controller, and PCI bus
> > that you have created.
> >
> > The remote machine type should default to no vCPUs and no memory
> > creation (the memory comes via the mpqemu link communications channel).
> >
> > At this point qemu-system-x86_64 contains a lot of code that you don't
> > want in the final remote device program. Let's ignore that for a
> > second.
> >
> > Now you can submit a 21-patch series containing just the remote device
> > infrastructure. This will be easier to merge.
> >
> > Returning to code size, the next step is to reduce the binary. QEMU has
> > a Kconfig-style system for optional features and dependencies. It's a
> > better approach than creating a separate make target because it
> > eliminates the duplication and mess in the makefiles.
> >
> > For example, you can disable TCG and KVM so that your binary has no
> > ability to execute guest code. Currently ./configure disallows this but
> > I've tried it and it works.
> >
> > You can add a new default-configs/ file that disables CONFIG_ISAPC,
> > CONFIG_I440FX, etc. When you compile QEMU most of hw/ will not be built
> > anymore. At this point you have a smaller binary that is still a
> > softmmu target so the makefiles are shared with the regular
> > qemu-system-x86_64.
> >
> > There will be some code for which there is no Kconfig option yet.
> > Further improvements can be made by adding Kconfig options for any code
> > that you wish to eliminate. Instead of writing makefile changes like
> > you did in this patch series you would be adding Kconfig options. The
> > nice thing is that this work isn't specific to the remote device program
> > - anyone can use the new Kconfig options to reduce the size of their
> > QEMU. So not only is it less messy than duplicating the makefiles,
> > but it also benefits everyone.
> >
> > The downside to doing this is that it will take a while to eliminate all
> > code that you don't want via Kconfig. However, your initial patch
> > series can be merged sooner and I think this direction is also cleaner.
> >
> > I hope I've explained the idea properly :). We can continue reviewing
> > the current series if you prefer, but I think it would be quicker to
> > drop the remote device program.
> >
> > Stefan
>
> Building QEMU twices just to get the remote is however not very
> attractive. So how about making remote a special target?
> Either remote-softmmu/ or if impossible x86_64-remote-softmmu/
Yes, that's a good idea. It needs to be the full x86_64-remote-softmmu
because hw/ code depends on the QEMU target :(.
To summarize the big advantage of this approach (besides reducing the
patch series): the existing makefile rules for softmmu will be used to
build the remote device program. No new main() and no new per-object
file makefile rules are needed.
Stefan