Re: [PATCH] virtio-net: failover: add missing remove_migration_state_change_notifier()

[Philippe Mathieu-Daudé]

On 4/27/21 4:02 PM, Dr. David Alan Gilbert wrote:
> * Laurent Vivier (lvivier@redhat.com) wrote:
>> In the failover case configuration, virtio_net_device_realize() uses an
>> add_migration_state_change_notifier() to add a state notifier, but this
>> notifier is not removed by the unrealize function when the virtio-net
>> card is unplugged.
>>
>> If the card is unplugged and a migration is started, the notifier is
>> called and as it is not valid anymore QEMU crashes.
>>
>> This patch fixes the problem by adding the
>> remove_migration_state_change_notifier() in virtio_net_device_unrealize().
>>
>> The problem can be reproduced with:
>>
>>   $ qemu-system-x86_64 -enable-kvm -m 1g -M q35 \
>>     -device pcie-root-port,slot=4,id=root1 \
>>     -device pcie-root-port,slot=5,id=root2 \
>>     -device 
>> virtio-net-pci,id=net1,mac=52:54:00:6f:55:cc,failover=on,bus=root1 \
>>     -monitor stdio disk.qcow2
>>   (qemu) device_del net1
>>   (qemu) migrate "exec:gzip -c > STATEFILE.gz"
>>
>>   Thread 1 "qemu-system-x86" received signal SIGSEGV, Segmentation fault.
>>   0x0000000000000000 in ?? ()
>>   (gdb) bt
>>   #0  0x0000000000000000 in  ()
>>   #1  0x0000555555d726d7 in notifier_list_notify (...)
>>       at .../util/notify.c:39
>>   #2  0x0000555555842c1a in migrate_fd_connect (...)
>>       at .../migration/migration.c:3975
>>   #3  0x0000555555950f7d in migration_channel_connect (...)
>>       error@entry=0x0) at .../migration/channel.c:107
>>   #4  0x0000555555910922 in exec_start_outgoing_migration (...)
>>       at .../migration/exec.c:42
>>
>> Reported-by: Igor Mammedov <imammedo@redhat.com>
>> Signed-off-by: Laurent Vivier <lvivier@redhat.com>
> 
> Yep, I think that's OK.

IIUC HMP device_del ends calling device_finalize(), and per Igor's
explanation of qdev transition states:
https://www.mail-archive.com/qemu-devel@nongnu.org/msg661853.html
I'd expect this to be done in instance_init/instance_finalize,
but since 'failover' is a virtio_net_properties, the callbacks
has to be registered in the realize() function, thus unregistered
in unrealize().

Mentioning it is in realize/unrealize due to 'failover' being a property:
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>

> Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> 
>> ---
>>  hw/net/virtio-net.c | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
>> index 66b9ff451185..914051feb75b 100644
>> --- a/hw/net/virtio-net.c
>> +++ b/hw/net/virtio-net.c
>> @@ -3373,6 +3373,7 @@ static void virtio_net_device_unrealize(DeviceState 
>> *dev)
>>  
>>      if (n->failover) {
>>          device_listener_unregister(&n->primary_listener);
>> +        remove_migration_state_change_notifier(&n->migration_state);
>>      }
>>  
>>      max_queues = n->multiqueue ? n->max_queues : 1;
>> -- 
>> 2.30.2
>>