Re: [BUG] vhost-vdpa: qemu-system-s390x crashes with second virtio-net-ccw device

[Cornelia Huck <cohuck@redhat.com>]

On Fri, 24 Jul 2020 11:17:57 -0400
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Fri, Jul 24, 2020 at 04:56:27PM +0200, Cornelia Huck wrote:
> > On Fri, 24 Jul 2020 09:30:58 -0400
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> >   
> > > On Fri, Jul 24, 2020 at 03:27:18PM +0200, Cornelia Huck wrote:  
> > > > When I start qemu with a second virtio-net-ccw device (i.e. adding
> > > > -device virtio-net-ccw in addition to the autogenerated device), I get
> > > > a segfault. gdb points to
> > > > 
> > > > #0  0x000055d6ab52681d in virtio_net_get_config (vdev=<optimized out>, 
> > > >     config=0x55d6ad9e3f80 "RT") at /home/cohuck/git/qemu/hw/net/virtio-net.c:146
> > > > 146	    if (nc->peer->info->type == NET_CLIENT_DRIVER_VHOST_VDPA) {
> > > > 
> > > > (backtrace doesn't go further)  
> > 
> > The core was incomplete, but running under gdb directly shows that it
> > is just a bog-standard config space access (first for that device).
> > 
> > The cause of the crash is that nc->peer is not set... no idea how that
> > can happen, not that familiar with that part of QEMU. (Should the code
> > check, or is that really something that should not happen?)
> > 
> > What I don't understand is why it is set correctly for the first,
> > autogenerated virtio-net-ccw device, but not for the second one, and
> > why virtio-net-pci doesn't show these problems. The only difference
> > between -ccw and -pci that comes to my mind here is that config space
> > accesses for ccw are done via an asynchronous operation, so timing
> > might be different.  
> 
> Hopefully Jason has an idea. Could you post a full command line
> please? Do you need a working guest to trigger this? Does this trigger
> on an x86 host?

Yes, it does trigger with tcg-on-x86 as well. I've been using

s390x-softmmu/qemu-system-s390x -M s390-ccw-virtio,accel=tcg -cpu qemu,zpci=on 
-m 1024 -nographic -device virtio-scsi-ccw,id=scsi0,devno=fe.0.0001 
-drive file=/path/to/image,format=qcow2,if=none,id=drive-scsi0-0-0-0 
-device scsi-hd,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=drive-scsi0-0-0-0,id=scsi0-0-0-0,bootindex=1 
-device virtio-net-ccw

It seems it needs the guest actually doing something with the nics; I
cannot reproduce the crash if I use the old advent calendar moon buggy
image and just add a virtio-net-ccw device.

(I don't think it's a problem with my local build, as I see the problem
both on my laptop and on an LPAR.)

> 
> > > > 
> > > > Starting qemu with no additional "-device virtio-net-ccw" (i.e., only
> > > > the autogenerated virtio-net-ccw device is present) works. Specifying
> > > > several "-device virtio-net-pci" works as well.
> > > > 
> > > > Things break with 1e0a84ea49b6 ("vhost-vdpa: introduce vhost-vdpa net
> > > > client"), 38140cc4d971 ("vhost_net: introduce set_config & get_config")
> > > > works (in-between state does not compile).    
> > > 
> > > Ouch. I didn't test all in-between states :(
> > > But I wish we had a 0-day instrastructure like kernel has,
> > > that catches things like that.  
> > 
> > Yep, that would be useful... so patchew only builds the complete series?
> >   
> > >   
> > > > This is reproducible with tcg as well. Same problem both with
> > > > --enable-vhost-vdpa and --disable-vhost-vdpa.
> > > > 
> > > > Have not yet tried to figure out what might be special with
> > > > virtio-ccw... anyone have an idea?
> > > > 
> > > > [This should probably be considered a blocker?]    
> > 
> > I think so, as it makes s390x unusable with more that one
> > virtio-net-ccw device, and I don't even see a workaround.  
>