Re: [PATCH v3] virtio-net: prevent offloads reset on migration

[Michael S. Tsirkin]

On Thu, Oct 10, 2019 at 04:58:09PM +0100, Dr. David Alan Gilbert wrote:
> * Michael S. Tsirkin (address@hidden) wrote:
> > On Thu, Oct 10, 2019 at 02:11:15PM +0200, Mikhail Sennikovsky wrote:
> > > Currently offloads disabled by guest via the 
> > > VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
> > > command are not preserved on VM migration.
> > > Instead all offloads reported by guest features (via 
> > > VIRTIO_PCI_GUEST_FEATURES)
> > > get enabled.
> > > What happens is: first the VirtIONet::curr_guest_offloads gets restored
> > > and offloads are getting set correctly:
> > > 
> > >  #0  qemu_set_offload (nc=0x555556a11400, csum=1, tso4=0, tso6=0, ecn=0, 
> > > ufo=0) at net/net.c:474
> > >  #1  virtio_net_apply_guest_offloads (n=0x555557701ca0) at 
> > > hw/net/virtio-net.c:720
> > >  #2  virtio_net_post_load_device (opaque=0x555557701ca0, version_id=11) 
> > > at hw/net/virtio-net.c:2334
> > >  #3  vmstate_load_state (f=0x5555569dc010, vmsd=0x555556577c80 
> > > <vmstate_virtio_net_device>, opaque=0x555557701ca0, version_id=11)
> > >      at migration/vmstate.c:168
> > >  #4  virtio_load (vdev=0x555557701ca0, f=0x5555569dc010, version_id=11) 
> > > at hw/virtio/virtio.c:2197
> > >  #5  virtio_device_get (f=0x5555569dc010, opaque=0x555557701ca0, size=0, 
> > > field=0x55555668cd00 <__compound_literal.5>) at hw/virtio/virtio.c:2036
> > >  #6  vmstate_load_state (f=0x5555569dc010, vmsd=0x555556577ce0 
> > > <vmstate_virtio_net>, opaque=0x555557701ca0, version_id=11) at 
> > > migration/vmstate.c:143
> > >  #7  vmstate_load (f=0x5555569dc010, se=0x5555578189e0) at 
> > > migration/savevm.c:829
> > >  #8  qemu_loadvm_section_start_full (f=0x5555569dc010, 
> > > mis=0x5555569eee20) at migration/savevm.c:2211
> > >  #9  qemu_loadvm_state_main (f=0x5555569dc010, mis=0x5555569eee20) at 
> > > migration/savevm.c:2395
> > >  #10 qemu_loadvm_state (f=0x5555569dc010) at migration/savevm.c:2467
> > >  #11 process_incoming_migration_co (opaque=0x0) at 
> > > migration/migration.c:449
> > > 
> > > However later on the features are getting restored, and offloads get 
> > > reset to
> > > everything supported by features:
> > > 
> > >  #0  qemu_set_offload (nc=0x555556a11400, csum=1, tso4=1, tso6=1, ecn=0, 
> > > ufo=0) at net/net.c:474
> > >  #1  virtio_net_apply_guest_offloads (n=0x555557701ca0) at 
> > > hw/net/virtio-net.c:720
> > >  #2  virtio_net_set_features (vdev=0x555557701ca0, features=5104441767) 
> > > at hw/net/virtio-net.c:773
> > >  #3  virtio_set_features_nocheck (vdev=0x555557701ca0, val=5104441767) at 
> > > hw/virtio/virtio.c:2052
> > >  #4  virtio_load (vdev=0x555557701ca0, f=0x5555569dc010, version_id=11) 
> > > at hw/virtio/virtio.c:2220
> > >  #5  virtio_device_get (f=0x5555569dc010, opaque=0x555557701ca0, size=0, 
> > > field=0x55555668cd00 <__compound_literal.5>) at hw/virtio/virtio.c:2036
> > >  #6  vmstate_load_state (f=0x5555569dc010, vmsd=0x555556577ce0 
> > > <vmstate_virtio_net>, opaque=0x555557701ca0, version_id=11) at 
> > > migration/vmstate.c:143
> > >  #7  vmstate_load (f=0x5555569dc010, se=0x5555578189e0) at 
> > > migration/savevm.c:829
> > >  #8  qemu_loadvm_section_start_full (f=0x5555569dc010, 
> > > mis=0x5555569eee20) at migration/savevm.c:2211
> > >  #9  qemu_loadvm_state_main (f=0x5555569dc010, mis=0x5555569eee20) at 
> > > migration/savevm.c:2395
> > >  #10 qemu_loadvm_state (f=0x5555569dc010) at migration/savevm.c:2467
> > >  #11 process_incoming_migration_co (opaque=0x0) at 
> > > migration/migration.c:449
> > > 
> > > Fix this by making the virtio_net_set_features not reset
> > > the curr_guest_offloads in case the VM runstate is RUN_STATE_INMIGRATE
> > > 
> > > Signed-off-by: Mikhail Sennikovsky <address@hidden>
> > 
> > I thought more about it and I really think the problem here is
> > different: we really should first set features and then apply offloads.
> > 
> > It is simply dumb that we are calling device specific
> > code before initializing generic virtio.
> > 
> > The logical order of initialization should be:
> > 
> > init base class
> > init derived class
> > 
> > 
> > instead we have it all spread out :(
> 
> Yeh; we've hit problems there before.
> 
> > 
> > What to do about it?
> > Could we maybe teach vmstate to record post load hooks
> > it needs to run *somewhere*?
> > Then we could load device specific state and record post load
> > hooks, then load generic state, run generic post load,
> > then run device specific post load hooks.
> > 
> > Dave, what do you think?
> 
> I worry about changing all existing virtio devices; who knows
> what existing ordering constraints those devices might have on their
> post_load's - for example one field in the vmstate might rely on
> another post_load having already run and set a flag; that's not
> that unusual.
> 
> I can imagine having a Notifier* chain; you initialise it at the start
> of virtio_load, provide a virtio_load_post_notifier_add(...) and any
> post_load that wants delaying can add itself and then you call it
> at the right point in virtio_load; but I'd be too scared to change
> the ordering of all post_loads of all devices.  You might be
> OK to do that for the top level post_load of a virtio device,
> because there are only a few to check.
> 
> Dave


Okay so let's just add a new hook for that.
I'll send a patch that does this shortly.
Mikhail, I'll send a couple of untested patches,
please try them - either ack or rework in any way you like.