Re: [PATCH 4/7] hw/scsi/virtio-scsi: Use VIRTIO_SCSI_COMMON() macro

[Kevin Wolf]

Am 31.10.2023 um 14:48 hat Richard W.M. Jones geschrieben:
> On Tue, Oct 31, 2023 at 02:17:56PM +0100, Kevin Wolf wrote:
> > Am 17.10.2023 um 16:01 hat Philippe Mathieu-Daudé geschrieben:
> > > Access QOM parent with the proper QOM VIRTIO_SCSI_COMMON() macro.
> > > 
> > > Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> > > ---
> > >  hw/scsi/virtio-scsi.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c
> > > index 45b95ea070..fa53f0902c 100644
> > > --- a/hw/scsi/virtio-scsi.c
> > > +++ b/hw/scsi/virtio-scsi.c
> > > @@ -761,7 +761,7 @@ static void virtio_scsi_fail_cmd_req(VirtIOSCSIReq 
> > > *req)
> > >  
> > >  static int virtio_scsi_handle_cmd_req_prepare(VirtIOSCSI *s, 
> > > VirtIOSCSIReq *req)
> > >  {
> > > -    VirtIOSCSICommon *vs = &s->parent_obj;
> > > +    VirtIOSCSICommon *vs = VIRTIO_SCSI_COMMON(s);
> > >      SCSIDevice *d;
> > >      int rc;
> > 
> > Why is a dynamic cast more "proper" than a static type-safe access, even
> > more so in a hot I/O path?
> > 
> > Rich Jones posted a flamegraph the other day that surprised me because
> > object_class_dynamic_class_assert() and object_dynamic_cast_assert()
> > were shown to be a big part of scsi_req_new(). In the overall
> > performance, it's probably dwarved by other issues, but unnecessary
> > little things can add up, too.
> 
> I think Kevin is referring to one of these flamegraphs:
> 
>   http://oirase.annexia.org/tmp/2023-kvm-build-on-device.svg
>   http://oirase.annexia.org/tmp/2023-kvm-build.svg
> 
> Here's a zoom showing scsi_req_new (hopefully this URL is stable ...):
> 
>   
> http://oirase.annexia.org/tmp/2023-kvm-build-on-device.svg?s=scsi_req_new&x=512.9&y=501
> 
> Note that qemu has been compiled with QOM cast debug.  This is the
> default for Fedora (not RHEL) because we'd like to get early detection
> of bugs from Fedora users.

Indeed, object_dynamic_cast_assert() probably wouldn't show up any more
if you disabled QOM cast debugging. But we enable it by default, so
that's an explicit setting users have to make, and as you say, at least
Fedora doesn't.

The other one, object_class_dynamic_cast_assert() isn't even affected
by the debug setting because it does more than just asserting a plain
cast. Working around it isn't as trivial either, but I did think about
caching it in the object to avoid a string based class lookup for every
single request.

Looking at it again, strangely CONFIG_QOM_CAST_DEBUG enables a cache in
object_class_dynamic_cast_assert() that isn't used otherwise. Why don't
we want this optimisation without the debug flag?

Kevin