Re: [Qemu-devel] [RFC 6/6] device-crash-test: Multi-device device_add test

[Dr. David Alan Gilbert]

* Thomas Huth (address@hidden) wrote:
> On 27.09.2017 01:07, Eduardo Habkost wrote:
> > When running device_add tests, test all devices in a single run
> > instead of restarting QEMU every time.
> > 
> > There's a plug_all testcase argument that can be used to make the
> > test code plug all devices at once.  I'm adding this mode because
> > there's a crash that was detected while testing the script
> > without any device_del command, but the crash goes away if we
> > device_del every device immediately:
> > 
> >   $ ../scripts/device-crash-test ./x86_64-softmmu/qemu-system-x86_64 
> > --quick -t method=device_add  --strict -t plug_all=1
> >   INFO: running test case: binary=./x86_64-softmmu/qemu-system-x86_64 
> > plug_all=1 accel=kvm machine=pc-0.12 device=* method=device_add
> >   WARNING: qemu received signal -6: ./x86_64-softmmu/qemu-system-x86_64 
> > -chardev socket,id=mon,path=/var/tmp/qemu-23578-monitor.sock -mon 
> > chardev=mon,mode=control -display none -vga none -S -machine 
> > pc-0.12,accel=kvm
> >   ERROR: result: binary=./x86_64-softmmu/qemu-system-x86_64 plug_all=1 
> > accel=kvm machine=pc-0.12 device=* method=device_add
> >   ERROR: cmdline: ./x86_64-softmmu/qemu-system-x86_64 -S -machine 
> > pc-0.12,accel=kvm
> >   ERROR: log: audio: Could not init `oss' audio driver
> >   ERROR: log: qemu-system-x86_64: .../qemu/migration/savevm.c:721: 
> > vmstate_register_with_alias_id: Assertion `!se->compat || se->instance_id 
> > == 0' failed.
> >   ERROR: last device_add device: usb-net
> 
> FWIW, I've seen similar crashes while working on my HMP device_add
> tester, when I skipped the device_del and plugged-in the devices a
> couple of times instead. I think this happens because some devices are
> doing a vmstate_register() in their instance_init() or realize()
> function, instead of using dc->vmsd as they should. However, with the
> git master branch as of today (31bc1d8481af414cfa2857f905e), I am
> currently not able anymore to reproduce the problem with the HMP
> device_add tester, so one of the recent "set user_creatable = false"
> patches might have fixed the issue for me already...

I think I've seen this problem when two things end up trying to register
with the same migration name;  I'll be honest I can't remember the
details, but I think the choice is either between having a full bus
name, e.g.

    pci-xx:yy.z:aa:bb:.c:...... then similar for the USB chain

or
    mydevice  + number

where number is the instance=id.  So to hit this I think
it's typically a case of messing up that long path and it no
longer being unique.
Printing the path in vmstate_register_with_alias_id after the
pstrcat might help.
I did add some sanity checking in there for a case I'd
previously hit where the name just got truncated, but I don't
think you should be able to hit that any more.

Dave

>  Thomas
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK