Re: virtio: why no full reset on virtio_set_status 0 ?

[Michael S. Tsirkin]

On Thu, Jul 28, 2022 at 09:43:56AM +0200, Claudio Fontana wrote:
> On 7/28/22 03:27, Jason Wang wrote:
> > On Wed, Jul 27, 2022 at 11:32 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >>
> >> On Wed, Jul 27, 2022 at 12:51:31PM +0200, Claudio Fontana wrote:
> >>> Hi Michael and all,
> >>>
> >>> I have started researching a qemu / ovs / dpdk bug:
> >>>
> >>> https://inbox.dpdk.org/dev/322122fb-619d-96f6-5c3e-9eabdbf3819a@redhat.com/T/
> >>>
> >>> that seems to be affecting multiple parties in the telco space,
> >>>
> >>> and during this process I noticed that qemu/hw/virtio/virtio.c does not 
> >>> do a full virtio reset
> >>> in virtio_set_status, when receiving a status value of 0.
> >>>
> >>> It seems it has always been this way, so I am clearly missing / 
> >>> forgetting something basic,
> >>>
> >>> I checked the virtio spec at https://docs.oasis-open.org/
> >>>
> >>> and from:
> >>>
> >>> "
> >>> 4.1.4.3 Common configuration structure layout
> >>>
> >>> device_status
> >>> The driver writes the device status here (see 2.1). Writing 0 into this 
> >>> field resets the device.
> >>>
> >>> "
> >>>
> >>> and
> >>>
> >>> "
> >>> 2.4.1 Device Requirements: Device Reset
> >>> A device MUST reinitialize device status to 0 after receiving a reset.
> >>> "
> >>>
> >>> I would conclude that in virtio.c::virtio_set_status we should 
> >>> unconditionally do a full virtio_reset.
> >>>
> >>> Instead, we have just the check:
> >>>
> >>>     if ((vdev->status & VIRTIO_CONFIG_S_DRIVER_OK) !=
> >>>         (val & VIRTIO_CONFIG_S_DRIVER_OK)) {
> >>>         virtio_set_started(vdev, val & VIRTIO_CONFIG_S_DRIVER_OK);
> >>>     }
> >>>
> >>> which just sets the started field,
> >>>
> >>> and then we have the call to the virtio device class set_status 
> >>> (virtio_net...),
> >>> but the VirtioDevice is not fully reset, as per the virtio_reset() call 
> >>> we are missing:
> >>>
> >>> "
> >>>     vdev->start_on_kick = false;
> >>>     vdev->started = false;
> >>>     vdev->broken = false;
> >>>     vdev->guest_features = 0;
> >>>     vdev->queue_sel = 0;
> >>>     vdev->status = 0;
> >>>     vdev->disabled = false;
> >>>     qatomic_set(&vdev->isr, 0);
> >>>     vdev->config_vector = VIRTIO_NO_VECTOR;
> >>>     virtio_notify_vector(vdev, vdev->config_vector);
> >>>
> >>>     for(i = 0; i < VIRTIO_QUEUE_MAX; i++) {
> >>>         ... initialize vdev->vq[i] ...
> >>>     }
> >>> "
> >>>
> >>> Doing a full reset seems to fix the problem for me, so I can send 
> >>> tentative patches if necessary,
> >>> but what am I missing here?
> >>>
> >>> Thanks,
> >>>
> >>> Claudio
> >>>
> >>> --
> >>> Claudio Fontana
> >>> Engineering Manager Virtualization, SUSE Labs Core
> >>>
> >>> SUSE Software Solutions Italy Srl
> >>
> >>
> >> So for example for pci:
> >>
> >>     case VIRTIO_PCI_STATUS:
> >>
> >>
> >>         ....
> >>
> >>         if (vdev->status == 0) {
> >>             virtio_pci_reset(DEVICE(proxy));
> >>         }
> >>
> >> which I suspect is a bug because:
> >>
> >> static void virtio_pci_reset(DeviceState *qdev)
> >> {
> >>     VirtIOPCIProxy *proxy = VIRTIO_PCI(qdev);
> >>     VirtioBusState *bus = VIRTIO_BUS(&proxy->bus);
> >>     PCIDevice *dev = PCI_DEVICE(qdev);
> >>     int i;
> >>
> >>     virtio_bus_reset(bus);
> > 
> > Note that we do virtio_reset() here.
> 
> 
> Yes, thank you, I completely overlooked it, I noticed this in Michael's 
> response as well.
> 
> However we end up with multiple calls to k->set_status, one from the 
> virtio_set_status call,
> and one from the virtio_bus_reset(), which is probably something we don't 
> want.
> 
> All in all it is not clear what the meaning of virtio_set_status is supposed 
> to be I think,
> and I wonder what the assumptions are among all the callers.
> If it is supposed to be an implementation of the virtio standard field as 
> described, I think we should do the reset right then and there,
> but maybe the true meaning of the function is another one I couldn't 
> understand, since _some_ of the cases are processes there.
> 
> And there is a question about ordering:
> 
> in virtio_pci we end up calling virtio_set_status(0), which gets us 
> k->set_status(vdev, 0), which lands in virtio_net_set_status(0) and 
> virtio_net_vhost_status,
> which causes a vhost_net_stop().
> Should we instead land in virtio_net_reset() first, by doing a virtio reset 
> earlier when detecting a 0 value from the driver?


Well we want to first stop the backend and only then reset our
local state. Seems to make sense ...


> in the scenario I am looking at (with vhost-user, ovs/dpdk, and a guest 
> testpmd application),
> the guest application goes away without any chance to signal (kill -9), then 
> gets immediately restarted and does a write of 0 to status, while qemu and 
> ovs still hold the state for the device.
> 
> As QEMU lands in vhost_net_stop(), it seems to cause a chain of events that 
> crash ovs which is trying to read an rx burst from the queue,

Not sure I got this part.

> while QEMU is left hanging waiting forever for a response to 
> VHOST_USER_GET_VRING_BASE issued as a result of vhost_net_stop.

Ineteresting why doesn't socket close after ovs crash cause the read to fail.

> Just saying, I am having more success with the second ordering, but I am 
> still studying, don't have the full picture yet.
> 
> Thanks,
> 
> Claudio
> 
> > 
> >>     msix_unuse_all_vectors(&proxy->pci_dev);
> >>
> >>     for (i = 0; i < VIRTIO_QUEUE_MAX; i++) {
> >>         proxy->vqs[i].enabled = 0;
> >>         proxy->vqs[i].num = 0;
> >>         proxy->vqs[i].desc[0] = proxy->vqs[i].desc[1] = 0;
> >>         proxy->vqs[i].avail[0] = proxy->vqs[i].avail[1] = 0;
> >>         proxy->vqs[i].used[0] = proxy->vqs[i].used[1] = 0;
> >>     }
> >>
> >>
> >> so far so good
> >>
> >>     if (pci_is_express(dev)) {
> >>         pcie_cap_deverr_reset(dev);
> >>         pcie_cap_lnkctl_reset(dev);
> >>
> >>         pci_set_word(dev->config + dev->exp.pm_cap + PCI_PM_CTRL, 0);
> >>     }
> >>
> >> this part is wrong I think, it got here by mistake since the same
> >> function is used for bus level reset.
> >>
> >> Jason, Marcel, any input?
> > 
> > Yes, I think we don't need PCI stuff here. We do virtio reset not pci.
> > 
> > Thanks
> > 
> >>
> >> --
> >> MST
> >>
> > 
> >