Re: [Qemu-devel] [PATCH] virtio-pci: fix migration for pci bus master

[Greg Kurz <gkurz@linux.vnet.ibm.com>]

On Mon, 6 Oct 2014 19:26:21 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:
> On Mon, Oct 06, 2014 at 04:51:35PM +0200, Greg Kurz wrote:
> > On Wed, 17 Sep 2014 20:39:25 +0300
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > On Wed, Sep 17, 2014 at 07:21:09PM +0200, Greg Kurz wrote:
> > > > On Sun, 14 Sep 2014 21:30:36 +0300
> > > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > 
> > > > > Current support for bus master (clearing OK bit)
> > > > > together with the need to support guests which do not
> > > > > enable PCI bus mastering, leads to extra state in
> > > > > VIRTIO_PCI_FLAG_BUS_MASTER_BUG bit, which isn't robust
> > > > > in case of cross-version migration for the case when
> > > > > guests use the device before setting DRIVER_OK.
> > > > > 
> > > > > Rip out VIRTIO_PCI_FLAG_BUS_MASTER_BUG and implement a simpler
> > > > > work-around: treat clearing of PCI_COMMAND as a virtio reset.  Old
> > > > > guests never touch this bit so they will work.
> > > > > 
> > > > > As reset clears device status, DRIVER and MASTER bits are
> > > > > now in sync, so we can fix up cross-version migration simply
> > > > > by synchronising them, without need to detect a buggy guest
> > > > > explicitly.
> > > > > 
> > > > > Drop tracking VIRTIO_PCI_FLAG_BUS_MASTER_BUG completely.
> > > > > 
> > > > > As reset makes the device quiescent, in the future we'll be able to drop
> > > > > checking OK bit in a bunch of places.
> > > > > 
> > > > > Cc: Jason Wang <jasowang@redhat.com>
> > > > > Cc: Greg Kurz <gkurz@linux.vnet.ibm.com>
> > > > > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > > > ---
> > > > 
> > > > Hi Michael,
> > > > 
> > > > I am not quite sure how to test this patch with my pseries based setup...
> > > > Migrating from qemu-2.1 to qemu-master ?
> > > > 
> > > > Cheers.
> > > > 
> > > > --
> > > > Greg
> > > 
> > > Exactly. And back! Pls don't forget to specify the 2.1 machine type.
> > > Thanks!
> > > 
> > 
> > Michael,
> > 
> > Nikunj and I had started to investigate the pseries breakage: the QEMU
> > originated reset brought by this patch clears the vq and breaks SLOF.
> > This isn't a surprise since reset should always come from the driver,
> > not the device.
> > 
> > Since commit 45363e46aeebfc99753389649eac7c7fc22bfe52 has reverted this
> > patch, QEMU works again for pseries and virtio. :)
> > 
> > So back to the initial issue, I've tried to migrate a pseries-2.1 guest running
> > rhel65, from QEMU v2.1.2 to QEMU master, back and forth, several times and it
> > always succeeded... what symptom this patch was expected to fix ?
> > 
> > Cheers.
> > 
> > --
> > Greg
> > 
> > > > >  hw/virtio/virtio-pci.c | 39 ++++++++++++++++++++-------------------
> > > > >  1 file changed, 20 insertions(+), 19 deletions(-)
> > > > > 
> > > > > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > > > > index a827cd4..f560814 100644
> > > > > --- a/hw/virtio/virtio-pci.c
> > > > > +++ b/hw/virtio/virtio-pci.c
> > > > > @@ -86,9 +86,6 @@
> > > > >   * 12 is historical, and due to x86 page size. */
> > > > >  #define VIRTIO_PCI_QUEUE_ADDR_SHIFT    12
> > > > > 
> > > > > -/* Flags track per-device state like workarounds for quirks in older guests. */
> > > > > -#define VIRTIO_PCI_FLAG_BUS_MASTER_BUG  (1 << 0)
> > > > > -
> > > > >  static void virtio_pci_bus_new(VirtioBusState *bus, size_t bus_size,
> > > > >                                 VirtIOPCIProxy *dev);
> > > > > 
> > > > > @@ -323,14 +320,6 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
> > > > >                                       proxy->pci_dev.config[PCI_COMMAND] |
> > > > >                                       PCI_COMMAND_MASTER, 1);
> > > > >          }
> > > > > -
> > > > > -        /* Linux before 2.6.34 sets the device as OK without enabling
> > > > > -           the PCI device bus master bit. In this case we need to disable
> > > > > -           some safety checks. */
> > > > > -        if ((val & VIRTIO_CONFIG_S_DRIVER_OK) &&
> > > > > -            !(proxy->pci_dev.config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
> > > > > -            proxy->flags |= VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
> > > > > -        }
> > > > >          break;
> > > > >      case VIRTIO_MSI_CONFIG_VECTOR:
> > > > >          msix_vector_unuse(&proxy->pci_dev, vdev->config_vector);
> > > > > @@ -480,13 +469,18 @@ static void virtio_write_config(PCIDevice *pci_dev, uint32_t address,
> > > > >      VirtIOPCIProxy *proxy = DO_UPCAST(VirtIOPCIProxy, pci_dev, pci_dev);
> > > > >      VirtIODevice *vdev = virtio_bus_get_device(&proxy->bus);
> > > > > 
> > > > > +    uint8_t cmd = proxy->pci_dev.config[PCI_COMMAND];
> > > > > +
> > > > >      pci_default_write_config(pci_dev, address, val, len);
> > > > > 
> > > > >      if (range_covers_byte(address, len, PCI_COMMAND) &&
> > > > >          !(pci_dev->config[PCI_COMMAND] & PCI_COMMAND_MASTER) &&
> > > > > -        !(proxy->flags & VIRTIO_PCI_FLAG_BUS_MASTER_BUG)) {
> > > > > +        (cmd & PCI_COMMAND_MASTER)) {
> > > > > +        /* Bus driver disables bus mastering - make it act
> > > > > +         * as a kind of reset to render the device quiescent. */
> > > > >          virtio_pci_stop_ioeventfd(proxy);
> > > > > -        virtio_set_status(vdev, vdev->status & ~VIRTIO_CONFIG_S_DRIVER_OK);
> > > > > +        virtio_reset(vdev);
> > > > > +        msix_unuse_all_vectors(&proxy->pci_dev);
> > > > >      }
> > > > >  }
> > > > > 
> > > > > @@ -895,11 +889,19 @@ static void virtio_pci_vmstate_change(DeviceState *d, bool running)
> > > > >      VirtIODevice *vdev = virtio_bus_get_device(&proxy->bus);
> > > > > 
> > > > >      if (running) {
> > > > > -        /* Try to find out if the guest has bus master disabled, but is
> > > > > -           in ready state. Then we have a buggy guest OS. */
> > > > > -        if ((vdev->status & VIRTIO_CONFIG_S_DRIVER_OK) &&
> > > > > -            !(proxy->pci_dev.config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
> > > > > -            proxy->flags |= VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
> > > > > +        /* Linux before 2.6.34 drives the device without enabling
> > > > > +           the PCI device bus master bit. Enable it automatically
> > > > > +           for the guest. This is a PCI spec violation but so is
> > > > > +           initiating DMA with bus master bit clear.
> > > > > +           Note: this only makes a difference when migrating
> > > > > +           across QEMU versions from an old QEMU, as for new QEMU
> > > > > +           bus master and driver bits are always in sync.
> > > > > +           TODO: consider enabling conditionally for compat machine types. */
> > > > > +        if (vdev->status & (VIRTIO_CONFIG_S_ACKNOWLEDGE |
> > > > > +                            VIRTIO_CONFIG_S_DRIVER)) {
> > > > > +            pci_default_write_config(&proxy->pci_dev, PCI_COMMAND,
> > > > > +                                     proxy->pci_dev.config[PCI_COMMAND] |
> > > > > +                                     PCI_COMMAND_MASTER, 1);
> > > > >          }
> > > > >          virtio_pci_start_ioeventfd(proxy);
> > > > >      } else {
> > > > > @@ -1040,7 +1042,6 @@ static void virtio_pci_reset(DeviceState *qdev)
> > > > >      virtio_pci_stop_ioeventfd(proxy);
> > > > >      virtio_bus_reset(bus);
> > > > >      msix_unuse_all_vectors(&proxy->pci_dev);
> > > > > -    proxy->flags &= ~VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
> > > > >  }
> > > > > 
> > > > >  static Property virtio_pci_properties[] = {
> > > 
> 
> 
> It's a race condition: you have to catch it migrating
> in the window before DRIVER_OK is set.
> 

Ok I see... when DRIVER and MASTER bits are not in sync. I'll patch QEMU to
trigger a migration in this window then.

Thanks.

--
Greg