Re: [Qemu-devel] [PATCH v4 06/10] vfio: add check host bus reset is support or not

[Alex Williamson]

On Mon, 21 Mar 2016 18:08:42 +0800
Cao jin <address@hidden> wrote:

> From: Chen Fan <address@hidden>
> 
> when boot up a VM that assigning vfio devices with aer enabled, we
> must check the vfio device whether support host bus reset. because
> when one error occur. OS driver always recover the device by do a
> bus reset, in order to recover the vfio device, qemu must able to do
> a host bus reset to recover the device to default status. and for all
> affected devices by the bus reset. we must check them whether all
> are assigned to the VM and on the same virtual bus. meanwhile, for
> simply done, the devices which don't affected by the host bus reset
> are not allowed to assign to the same virtual bus.

Rewording/expansion suggestion:

  When assigning a vfio device with AER enabled, we must check whether
  the device supports a host bus reset (ie. hot reset) as this may be
  used by the guest OS in order to recover the device from an AER
  error.  QEMU must therefore have the ability to perform a physical
  host bus reset using the existing vfio APIs in response to a virtual
  bus reset in the VM.  A physical bus reset affects all of the devices
  on the host bus, therefore we place a few simplifying configuration
  restriction on the VM:

   - All physical devices affected by a bus reset must be assigned to
     the VM with AER enabled on each and be configured on the same
     virtual bus in the VM.

   - No devices unaffected by the bus reset, be they physical, emulated,
     or paravirtual may be configured on the same virtual bus as a
     device supporting AER signaling through vfio.

  In other words users wishing to enable AER on a multifunction device
  need to assign all functions of the device to the same virtual bus
  and enable AER support for each device.  The easiest way to
  accomplish this is to identity map the physical functions to virtual
  functions with multifunction enabled on the virtual device.

> 
> Signed-off-by: Chen Fan <address@hidden>
> ---
>  hw/vfio/pci.c | 205 
> +++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
>  hw/vfio/pci.h |   1 +
>  2 files changed, 205 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> index 8842b7f..dce3b6d 100644
> --- a/hw/vfio/pci.c
> +++ b/hw/vfio/pci.c
> @@ -38,6 +38,10 @@
>  static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
>  static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
>  
> +#define HOST_CMP_FUNC_MASK       (1 << 0)
> +static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name,
> +                                uint8_t mask);
> +

This would of course go back to a _slot version of the function as
you've had in previous versions.

>  /*
>   * Disabling BAR mmaping can be slow, but toggling it around INTx can
>   * also be a huge overhead.  We try to get the best of both worlds by
> @@ -1877,6 +1881,185 @@ static int vfio_add_std_cap(VFIOPCIDevice *vdev, 
> uint8_t pos)
>      return 0;
>  }
>  
> +static void vfio_check_hot_bus_reset(VFIOPCIDevice *vdev, Error **errp)
> +{
> +    PCIBus *bus = vdev->pdev.bus;
> +    struct vfio_pci_hot_reset_info *info = NULL;
> +    struct vfio_pci_dependent_device *devices;
> +    VFIOGroup *group;
> +    int ret, i, devfn;
> +
> +    ret = vfio_get_hot_reset_info(vdev, &info);
> +    if (ret) {
> +        error_setg(errp, "vfio: Cannot enable AER for device %s,"
> +                   " device does not support hot reset.",
> +                   vdev->vbasedev.name);
> +        return;
> +    }
> +
> +    /* List all affected devices by bus reset */
> +    devices = &info->devices[0];
> +
> +    /* Verify that we have all the groups required */
> +    for (i = 0; i < info->count; i++) {
> +        PCIHostDeviceAddress host;
> +        VFIOPCIDevice *tmp;
> +        VFIODevice *vbasedev_iter;
> +        bool found = false;
> +
> +        host.domain = devices[i].segment;
> +        host.bus = devices[i].bus;
> +        host.slot = PCI_SLOT(devices[i].devfn);
> +        host.function = PCI_FUNC(devices[i].devfn);
> +
> +        /* Skip the current device */
> +        if (vfio_pci_host_match(&host, vdev->vbasedev.name, 0)) {
> +            continue;
> +        }
> +
> +        /* Ensure we own the group of the affected device */
> +        QLIST_FOREACH(group, &vfio_group_list, next) {
> +            if (group->groupid == devices[i].group_id) {
> +                break;
> +            }
> +        }
> +
> +        if (!group) {
> +            error_setg(errp, "vfio: Cannot enable AER for device %s, "
> +                       "depends on group %d which is not owned.",
> +                       vdev->vbasedev.name, devices[i].group_id);
> +            goto out;
> +        }
> +
> +        /* Ensure affected devices for reset on the same bus */
> +        QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
> +            if (vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
> +                continue;
> +            }
> +            tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
> +            if (vfio_pci_host_match(&host, tmp->vbasedev.name, 0)) {
> +                /*
> +                 * AER errors may be broadcast to all functions of a multi-
> +                 * function endpoint.  If any of those sibling functions are
> +                 * also assigned, they need to have AER enabled or else an
> +                 * error may continue to cause a vm_stop condition.  IOW,
> +                 * AER setup of this function would be pointless.
> +                 */
> +                if (vfio_pci_host_match(&host, vdev->vbasedev.name,
> +                                        HOST_CMP_FUNC_MASK) &&
> +                    !(tmp->features & VFIO_FEATURE_ENABLE_AER)) {
> +                    error_setg(errp, "vfio: Cannot enable AER for device %s, 
> on same slot"
> +                               " the dependent device %s which does not 
> enable AER.",
> +                               vdev->vbasedev.name, tmp->vbasedev.name);
> +                    goto out;
> +                }
> +
> +                if (tmp->pdev.bus != bus) {
> +                    error_setg(errp, "vfio: Cannot enable AER for device %s, 
> "
> +                               "the dependent device %s is not on the same 
> bus",
> +                               vdev->vbasedev.name, tmp->vbasedev.name);
> +                    goto out;
> +                }
> +                found = true;
> +                break;
> +            }
> +        }
> +
> +        /* Ensure all affected devices assigned to VM */
> +        if (!found) {
> +            error_setg(errp, "vfio: Cannot enable AER for device %s, "
> +                       "the dependent device %04x:%02x:%02x.%x "
> +                       "is not assigned to VM.",
> +                       vdev->vbasedev.name, host.domain, host.bus,
> +                       host.slot, host.function);
> +            goto out;
> +        }
> +    }
> +
> +    /*
> +     * To make things simply, functions must are combined
> +     * in the same way as on the host. no other irrelative
> +     * functions on the virtual same bus was required.
> +     */

Suggestion:

  The above code verified that all devices affected by a bus reset
  exist on the same bus in the VM.  To further simplify, we also
  require that there are no additional devices beyond those existing on
  the VM bus.

> +    for (devfn = 0; devfn < 8; devfn++) {

What about ARI?  Don't we potentially need to walk up to 255 here if
the parent bridge device has ARI enabled?  Should we always walk up
through 255 to prevent users from using non-zero slot numbers for other
devices?

> +        VFIOPCIDevice *tmp;
> +        PCIDevice *dev;
> +        bool found = false;
> +
> +        dev = pci_find_device(bus, pci_bus_num(bus),
> +                  PCI_DEVFN(PCI_SLOT(vdev->pdev.devfn), devfn));
> +
> +        if (!dev) {
> +            continue;
> +        }
> +
> +        if (!object_dynamic_cast(OBJECT(dev), "vfio-pci")) {
> +            error_setg(errp, "vfio: Cannot enable AER for device %s, "
> +                       "only support all dependent devices on the same bus, "
> +                       "%s is not one of the dependent devices.",
> +                       vdev->vbasedev.name, dev->name);

"vfio: Cannot enable AER for device %s, device %s cannot be configured
on the same virtual bus"

Is dev->name sufficiently descriptive?  Perhaps print the conflicting
PCI address as well.

> +            goto out;
> +        }
> +
> +        tmp = DO_UPCAST(VFIOPCIDevice, pdev, dev);
> +        for (i = 0; i < info->count; i++) {
> +            PCIHostDeviceAddress host;
> +
> +            host.domain = devices[i].segment;
> +            host.bus = devices[i].bus;
> +            host.slot = PCI_SLOT(devices[i].devfn);
> +            host.function = PCI_FUNC(devices[i].devfn);
> +
> +            if (vfio_pci_host_match(&host, tmp->vbasedev.name, 0)) {
> +                found = true;
> +                break;
> +            }
> +        }
> +
> +        if (!found) {
> +            error_setg(errp, "vfio: Cannot enable AER for device %s, "
> +                       "only support all dependent devices on the same bus, "
> +                       "%s is not one of the dependent devices.",
> +                       vdev->vbasedev.name, tmp->vbasedev.name);

Update this comment similar to above too.

> +            goto out;
> +        }
> +    }
> +
> +out:
> +    g_free(info);
> +    return;
> +}
> +
> +static void vfio_aer_check_host_bus_reset(Error **errp)
> +{
> +    VFIOGroup *group;
> +    VFIODevice *vbasedev;
> +    VFIOPCIDevice *vdev;
> +    Error *local_err = NULL;
> +
> +    /* Check All vfio-pci devices if have bus reset capability */
> +    QLIST_FOREACH(group, &vfio_group_list, next) {
> +        QLIST_FOREACH(vbasedev, &group->device_list, next) {
> +            if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
> +                continue;
> +            }
> +            vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
> +            if (vdev->features & VFIO_FEATURE_ENABLE_AER) {
> +                if (!pci_get_function_0(&vdev->pdev)) {
> +                    continue;
> +                }

Is this function_0 test really necessary?  It seems like it's only
relevant to the hotplug case, in which case it should be included in
the next patch, but even there I don't see a path where we'll get here
when function 0 is not populated.  Is this perhaps leftover from
previous versions?  It actually seems like it's opening a gap where
function 0 could be a non-vfio-pci device (or non-existent) and bypass
our validation check here.

> +                vfio_check_hot_bus_reset(vdev, &local_err);
> +                if (local_err) {
> +                    error_propagate(errp, local_err);
> +                    return;
> +                }
> +            }
> +        }
> +    }
> +
> +    return;
> +}
> +
>  static int vfio_setup_aer(VFIOPCIDevice *vdev, uint8_t cap_ver,
>                            int pos, uint16_t size)
>  {
> @@ -2060,7 +2243,6 @@ static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
>      vfio_intx_enable(vdev);
>  }
>  
> -#define HOST_CMP_FUNC_MASK       (1 << 0)
>  static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name,
>                                  uint8_t mask)
>  {
> @@ -2587,6 +2769,22 @@ static void vfio_unregister_req_notifier(VFIOPCIDevice 
> *vdev)
>      vdev->req_enabled = false;
>  }
>  
> +static void vfio_pci_machine_done_notify(Notifier *notifier, void *unused)
> +{
> +    Error *local_err = NULL;
> +
> +    vfio_aer_check_host_bus_reset(&local_err);
> +    if (local_err) {
> +        fprintf(stderr, "%s\n", error_get_pretty(local_err));
> +        error_free(local_err);
> +        exit(1);
> +    }
> +}
> +
> +static Notifier machine_notifier = {
> +    .notify = vfio_pci_machine_done_notify,
> +};
> +
>  static int vfio_initfn(PCIDevice *pdev)
>  {
>      VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
> @@ -2932,6 +3130,11 @@ static const TypeInfo vfio_pci_dev_info = {
>  static void register_vfio_pci_dev_type(void)
>  {
>      type_register_static(&vfio_pci_dev_info);
> +    /*
> +     * Register notifier when machine init is done, since we need
> +     * check the configration manner after all vfio device are inited.
> +     */

Suggestion:

  The AER configuration may depend on multiple devices, so we cannot
  validate consistency after each device is initialized.  We can only
  depend on function initialization order (function 0 last) for hotplug
  devices, therefore a machine-init-done notifier is used to validate
  the configuration after all cold-plug devices are processed.

> +    qemu_add_machine_init_done_notifier(&machine_notifier);
>  }
>  
>  type_init(register_vfio_pci_dev_type)
> diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
> index 7b3924e..db7c6d5 100644
> --- a/hw/vfio/pci.h
> +++ b/hw/vfio/pci.h
> @@ -15,6 +15,7 @@
>  #include "qemu-common.h"
>  #include "exec/memory.h"
>  #include "hw/pci/pci.h"
> +#include "hw/pci/pci_bus.h"
>  #include "hw/pci/pci_bridge.h"
>  #include "hw/vfio/vfio-common.h"
>  #include "qemu/event_notifier.h"