diff mbox

[v2] virtio-pci: enable bus master for old guests

Message ID 20140911180039.2c3df0d2@bahia.local
State New
Headers show

Commit Message

Greg Kurz Sept. 11, 2014, 4 p.m. UTC
On Wed, 10 Sep 2014 20:30:47 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> commit cc943c36faa192cd4b32af8fe5edb31894017d35
>     pci: Use bus master address space for delivering MSI/MSI-X messages
> breaks virtio-net for rhel6.[56] x86 guests because they don't
> enable bus mastering for virtio PCI devices
> 
> Old guests forgot to enable bus mastering, enable it
> automatically when driver discovers device.
> 
> Cc: qemu-stable@nongnu.org
> Reported-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> ---
> 
> OK, this should have better luck.

Unfortunately not... it is even worse than V1 as
both x86 and ppc64 guests fail. I've added some
debug messages in virtio_ioport_write() and
virtio_write_config() to track:
- VIRTIO_PCI_STATUS changes
- bus master disablement

This is what I get for x86 with a virtio-net device:

virtio_write_config: PCI_COMMAND = 0x3
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x7
virtio_ioport_write: detected bus master bug ==> v1 patch would have
                                                 enabled BM here

and for ppc64 when booting from a virtio-blk disk:

virtio_write_config: PCI_COMMAND = 0x3
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM

I think we should:
- detect the "old driver" when reaching DRIVER as it works for
  both architectures
- always re-enable BM when an "old driver" disables it

That is basically your v1 plus its follow-up patch:

https://lists.nongnu.org/archive/html/qemu-devel/2014-09/msg01918.html

plus:


I could verify it fixes the issues. I will send a patch shortly.


Cheers.

--
Greg

> This also makes it possible to simplify code,
> will do that in a follow-up patch.
> 
> 
>  hw/virtio/virtio-pci.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> index ddb5da1..a29d94f 100644
> --- a/hw/virtio/virtio-pci.c
> +++ b/hw/virtio/virtio-pci.c
> @@ -303,6 +303,14 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
>              virtio_pci_stop_ioeventfd(proxy);
>          }
> 
> +        /* Linux before 2.6.34 uses the device without enabling
> +           the PCI device bus master bit. As a work-around, enable it
> +           automatically when driver detects the device. */
> +        if (val == VIRTIO_CONFIG_S_ACKNOWLEDGE) {
> +            memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,
> +                                      true);
> +        }
> +
>          virtio_set_status(vdev, val & 0xFF);
> 
>          if (val & VIRTIO_CONFIG_S_DRIVER_OK) {

Comments

Michael S. Tsirkin Sept. 11, 2014, 4:49 p.m. UTC | #1
On Thu, Sep 11, 2014 at 06:00:39PM +0200, Greg Kurz wrote:
> On Wed, 10 Sep 2014 20:30:47 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > commit cc943c36faa192cd4b32af8fe5edb31894017d35
> >     pci: Use bus master address space for delivering MSI/MSI-X messages
> > breaks virtio-net for rhel6.[56] x86 guests because they don't
> > enable bus mastering for virtio PCI devices
> > 
> > Old guests forgot to enable bus mastering, enable it
> > automatically when driver discovers device.
> > 
> > Cc: qemu-stable@nongnu.org
> > Reported-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
> > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > ---
> > 
> > OK, this should have better luck.
> 
> Unfortunately not... it is even worse than V1 as
> both x86 and ppc64 guests fail. I've added some
> debug messages in virtio_ioport_write() and
> virtio_write_config() to track:
> - VIRTIO_PCI_STATUS changes
> - bus master disablement
> 
> This is what I get for x86 with a virtio-net device:
> 
> virtio_write_config: PCI_COMMAND = 0x3
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
> virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM

Weird. What code in guest causes this second write?
Any idea?

> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x7
> virtio_ioport_write: detected bus master bug ==> v1 patch would have
>                                                  enabled BM here
> 
> and for ppc64 when booting from a virtio-blk disk:
> 
> virtio_write_config: PCI_COMMAND = 0x3
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
> virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM
> 
> I think we should:
> - detect the "old driver" when reaching DRIVER as it works for
>   both architectures

Basically take v2 and replace ACKNOWLEDGE with DRIVER.
Yea, maybe.
But I would like to understand what is going on here.


> - always re-enable BM when an "old driver" disables it
> 
> That is basically your v1 plus its follow-up patch:
> 
> https://lists.nongnu.org/archive/html/qemu-devel/2014-09/msg01918.html
> 
> plus:
> 
> --- a/hw/virtio/virtio-pci.c
> +++ b/hw/virtio/virtio-pci.c
> @@ -317,7 +317,7 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
>          /* 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) &&
> +        if ((val & VIRTIO_CONFIG_S_DRIVER) &&
>              !(proxy->pci_dev.config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
>              proxy->flags |= VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
>              memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,
> 
> I could verify it fixes the issues. I will send a patch shortly.
> 
> 
> Cheers.
> 
> --
> Greg
> 
> > This also makes it possible to simplify code,
> > will do that in a follow-up patch.
> > 
> > 
> >  hw/virtio/virtio-pci.c | 8 ++++++++
> >  1 file changed, 8 insertions(+)
> > 
> > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > index ddb5da1..a29d94f 100644
> > --- a/hw/virtio/virtio-pci.c
> > +++ b/hw/virtio/virtio-pci.c
> > @@ -303,6 +303,14 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
> >              virtio_pci_stop_ioeventfd(proxy);
> >          }
> > 
> > +        /* Linux before 2.6.34 uses the device without enabling
> > +           the PCI device bus master bit. As a work-around, enable it
> > +           automatically when driver detects the device. */
> > +        if (val == VIRTIO_CONFIG_S_ACKNOWLEDGE) {
> > +            memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,
> > +                                      true);
> > +        }
> > +
> >          virtio_set_status(vdev, val & 0xFF);
> > 
> >          if (val & VIRTIO_CONFIG_S_DRIVER_OK) {
Michael S. Tsirkin Sept. 11, 2014, 5:52 p.m. UTC | #2
On Thu, Sep 11, 2014 at 06:00:39PM +0200, Greg Kurz wrote:
> On Wed, 10 Sep 2014 20:30:47 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > commit cc943c36faa192cd4b32af8fe5edb31894017d35
> >     pci: Use bus master address space for delivering MSI/MSI-X messages
> > breaks virtio-net for rhel6.[56] x86 guests because they don't
> > enable bus mastering for virtio PCI devices
> > 
> > Old guests forgot to enable bus mastering, enable it
> > automatically when driver discovers device.
> > 
> > Cc: qemu-stable@nongnu.org
> > Reported-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
> > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > ---
> > 
> > OK, this should have better luck.
> 
> Unfortunately not... it is even worse than V1 as
> both x86 and ppc64 guests fail. I've added some
> debug messages in virtio_ioport_write() and
> virtio_write_config() to track:
> - VIRTIO_PCI_STATUS changes
> - bus master disablement
> 
> This is what I get for x86 with a virtio-net device:
> 
> virtio_write_config: PCI_COMMAND = 0x3
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
> virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x7
> virtio_ioport_write: detected bus master bug ==> v1 patch would have
>                                                  enabled BM here
> 
> and for ppc64 when booting from a virtio-blk disk:
> 
> virtio_write_config: PCI_COMMAND = 0x3
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x0
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x1 ==> v2 patch enables BM
> virtio_ioport_write: VIRTIO_PCI_STATUS = 0x3
> virtio_write_config: PCI_COMMAND = 0x3       ==> guest disables BM


So guest sets DRIVER then overwrites PCI_COMMAND.  Weird.
Any idea which guest code does this?

> I think we should:
> - detect the "old driver" when reaching DRIVER as it works for
>   both architectures
> - always re-enable BM when an "old driver" disables it
> 
> That is basically your v1 plus its follow-up patch:
> 
> https://lists.nongnu.org/archive/html/qemu-devel/2014-09/msg01918.html
> 
> plus:
> 
> --- a/hw/virtio/virtio-pci.c
> +++ b/hw/virtio/virtio-pci.c
> @@ -317,7 +317,7 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
>          /* 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) &&
> +        if ((val & VIRTIO_CONFIG_S_DRIVER) &&
>              !(proxy->pci_dev.config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
>              proxy->flags |= VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
>              memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,
> 
> I could verify it fixes the issues. I will send a patch shortly.
> 
> 
> Cheers.
> 
> --
> Greg
> 
> > This also makes it possible to simplify code,
> > will do that in a follow-up patch.
> > 
> > 
> >  hw/virtio/virtio-pci.c | 8 ++++++++
> >  1 file changed, 8 insertions(+)
> > 
> > diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> > index ddb5da1..a29d94f 100644
> > --- a/hw/virtio/virtio-pci.c
> > +++ b/hw/virtio/virtio-pci.c
> > @@ -303,6 +303,14 @@ static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
> >              virtio_pci_stop_ioeventfd(proxy);
> >          }
> > 
> > +        /* Linux before 2.6.34 uses the device without enabling
> > +           the PCI device bus master bit. As a work-around, enable it
> > +           automatically when driver detects the device. */
> > +        if (val == VIRTIO_CONFIG_S_ACKNOWLEDGE) {
> > +            memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,
> > +                                      true);
> > +        }
> > +
> >          virtio_set_status(vdev, val & 0xFF);
> > 
> >          if (val & VIRTIO_CONFIG_S_DRIVER_OK) {
diff mbox

Patch

--- a/hw/virtio/virtio-pci.c
+++ b/hw/virtio/virtio-pci.c
@@ -317,7 +317,7 @@  static void virtio_ioport_write(void *opaque, uint32_t addr, uint32_t val)
         /* 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) &&
+        if ((val & VIRTIO_CONFIG_S_DRIVER) &&
             !(proxy->pci_dev.config[PCI_COMMAND] & PCI_COMMAND_MASTER)) {
             proxy->flags |= VIRTIO_PCI_FLAG_BUS_MASTER_BUG;
             memory_region_set_enabled(&proxy->pci_dev.bus_master_enable_region,