Re: [Qemu-devel] [PATCH RFC 0/5] Introduce "-object iothread-group"

[Fam Zheng <famz@redhat.com>]

On Tue, 07/11 15:15, Stefan Hajnoczi wrote:
> On Mon, Jul 10, 2017 at 03:20:22PM +0800, Fam Zheng wrote:
> > Last time we've looked at "-object iothread,spawns=N" but it was a bit abusive.
> > A dedicated "iothread-group" class is cleaner from the interface point of view.
> > This series does that.
> > 
> > It has the same set of poll parameters as the existing "iothread" object, plus
> > a "size" option to specify how many threads to start. Using iothread-group
> > doesn't require the user to explicitly create the contained IOThreads. The
> > IOThreads are created by the group object.
> > 
> > Internally, IOThreads share one AioContext.  This is to make it easier to adapt
> > this to the current data plane code (see the last patch). But it is an
> > implementation detail, and will change depending on the block layer multiqueue
> > needs.
> > 
> > TODO:
> > 
> > - qmp_query_iothread_groups, in addition to proper QOM @child property from
> >   IOThreadGroup to its IOThread instances.
> > - Add virtio-scsi.
> > - Variant of iothread_stop_all().
> > 
> > Fam Zheng (5):
> >   aio: Wrap poll parameters into AioContextPollParams
> >   iothread: Don't error on windows
> >   iothread: Extract iothread_start
> >   Introduce iothread-group
> >   virtio-blk: Add iothread-group property
> 
> From your TODO note above it looks like you plan to duplicate IOThread
> interfaces for IOThreadGroup?  This means existing query-iothreads users
> no longer see the full view of all IOThreads.
> 
> I think it would be cleaner to define and query IOThreads like they are
> today but change virtio-blk/virtio-scsi to accept a list of IOThreads.

That way the groups are formed passively and I'm not sure if it is better for
users/tools to manage in the long run. Consider this syntax:

    -object iothread,id=iot0 \
    -object iothread,id=iot1 \
    -object iothread,id=iot2 \
    -device virtio-blk-pci,id=vblk0,iothread.0=iot0,iothread.1=iot1 \
    -device virtio-blk-pci,id=vblk1,iothread.0=iot1,iothread.1=iot2

where vblk0 uses iot0 and iot1 and vblk1 uses iot1 and iot2. There is a
intersection between the two groups. IMO it is less clean compared to the
rule set by an explicit syntax:

    -object iothread-group,id=iotg0,size=4 \
    -object iothread-group,id=iotg1,size=4 \
    -device virtio-blk-pci,id=vblk0,iothread-group=iotg0 \
    -device virtio-blk-pci,id=vblk1,iothread-group=iotg1 \


Also I have not idea how easy it is to add a "list of links" qdev property. I
remember there was some related work in progress, but I've lost the pointers.

> That way existing management tool functionality can be used and the only
> tweak is that devices can now be added to multiple IOThreads.

Another way could be to still include any IOThreads created by IOThreadGroup in
"query-iothreads" output, and add a "group name" property so users know the
groupings.

> 
> It would be nice to express the group relationship in QOM instead of
> open coding a new group object.  The majority of the RFC code creates a
> child/link list relationship that QOM should support for any type, not
> just IOThread.

Sounds fine, but I'm not sure what exactly you have in mind (I think it is
extending QOM). Can you elaborate?

Fam