[libvirt] [PATCH 2/2] drm/i915/gvt: export mdev device version to sysfs for Intel vGPU
Yan Zhao
yan.y.zhao at intel.com
Wed Apr 24 02:33:06 UTC 2019
On Tue, Apr 23, 2019 at 07:39:11PM +0800, Cornelia Huck wrote:
> On Fri, 19 Apr 2019 04:35:59 -0400
> Yan Zhao <yan.y.zhao at intel.com> wrote:
>
> > This feature implements the version attribute for Intel's vGPU mdev
> > devices.
> >
> > version attribute is rw. It is queried by userspace software like libvirt
> > to check whether two vGPUs are compatible for live migration.
> >
> > It consists of two parts: common part and vendor proprietary part.
> > common part: 32 bit. lower 16 bits is vendor id and higher 16 bits
> > identifies device type. e.g., for pci device, it is
> > "pci vendor id" | (VFIO_DEVICE_FLAGS_PCI << 16).
> > vendor proprietary part: this part is varied in length. vendor driver can
> > specify any string to identify a device.
> >
> > For Intel vGPU of gen8 and gen9, the vendor proprietary part currently
> > consists of 2 fields: "device id" + "mdev type".
> >
> > Reading from a vGPU's version attribute, a string is returned in below
> > format: 00028086-<device id>-<mdev type>. e.g.
> > 00028086-193b-i915-GVTg_V5_2.
> >
> > Writing a string to a vGPU's version attribute will trigger GVT to check
> > whether a vGPU identified by the written string is compatible with
> > current vGPU owning this version attribute. errno is returned if the two
> > vGPUs are incompatible. The length of written string is returned in
> > compatible case.
> >
> > For other platforms, and for GVT not supporting vGPU live migration
> > feature, errnos are returned when read/write of mdev devices' version
> > attributes.
> >
> > For old GVT versions where no version attributes exposed in sysfs, it is
> > regarded as not supporting vGPU live migration.
> >
> > For future platforms, besides the current 2 fields in vendor proprietary
> > part, more fields may be added to identify Intel vGPU well for live
> > migration purpose.
> >
> > Cc: Alex Williamson <alex.williamson at redhat.com>
> > Cc: Erik Skultety <eskultet at redhat.com>
> > Cc: "Dr. David Alan Gilbert" <dgilbert at redhat.com>
> > Cc: Cornelia Huck <cohuck at redhat.com>
> > Cc: "Tian, Kevin" <kevin.tian at intel.com>
> > Cc: Zhenyu Wang <zhenyuw at linux.intel.com>
> > Cc: "Wang, Zhi A" <zhi.a.wang at intel.com>
> > c: Neo Jia <cjia at nvidia.com>
> > Cc: Kirti Wankhede <kwankhede at nvidia.com>
> >
> > Signed-off-by: Yan Zhao <yan.y.zhao at intel.com>
> > ---
> > drivers/gpu/drm/i915/gvt/Makefile | 2 +-
> > drivers/gpu/drm/i915/gvt/device_version.c | 94 +++++++++++++++++++++++
> > drivers/gpu/drm/i915/gvt/gvt.c | 55 +++++++++++++
> > drivers/gpu/drm/i915/gvt/gvt.h | 6 ++
> > 4 files changed, 156 insertions(+), 1 deletion(-)
> > create mode 100644 drivers/gpu/drm/i915/gvt/device_version.c
> >
>
> (...)
>
> > +static bool is_compatible(const char *self, const char *remote)
> > +{
> > + if (strlen(remote) != strlen(self))
> > + return false;
> > +
> > + return (strncmp(self, remote, strlen(self))) ? false : true;
> > +}
> > +
> > +ssize_t intel_gvt_get_vfio_device_version_len(struct drm_i915_private *dev_priv)
> > +{
> > + if (!IS_GEN(dev_priv, 8) && !IS_GEN(dev_priv, 9))
> > + return -ENODEV;
> > +
> > + return PAGE_SIZE;
> > +}
> > +
> > +ssize_t intel_gvt_get_vfio_device_version(struct drm_i915_private *dev_priv,
> > + char *buf, const char *mdev_type)
> > +{
> > + int cnt = 0, ret = 0;
> > + const char *str = NULL;
> > +
> > + /* currently only gen8 & gen9 are supported */
> > + if (!IS_GEN(dev_priv, 8) && !IS_GEN(dev_priv, 9))
> > + return -ENODEV;
> > +
> > + /* first 32 bit common part specifying vendor id and it's a pci
> > + * device
> > + */
> > + cnt = snprintf(buf, GVT_DEVICE_VERSION_COMMON_LEN + 1,
> > + "%08x", GVT_VFIO_DEVICE_VENDOR_ID);
> > + buf += cnt;
> > + ret += cnt;
> > +
> > + /* vendor proprietary part: device id + mdev type */
> > + /* device id */
> > + cnt = snprintf(buf, GVT_DEVICE_VERSION_DEVICE_ID_LEN + 2,
> > + "-%04x", INTEL_DEVID(dev_priv));
> > + buf += cnt;
> > + ret += cnt;
> > +
> > + /* mdev type */
> > + str = mdev_type;
> > + cnt = snprintf(buf, strlen(str) + 3, "-%s\n", mdev_type);
> > + buf += cnt;
> > + ret += cnt;
> > +
> > + return ret;
> > +}
>
> Looking at this handling, it seems much easier to me to simply use a
> numeric value instead of a string: You don't have to build it via
> sprintf, there are generic functions for parsing a string input into a
> simple number, and you have more options for compatibility (e.g.
> "version must be between m and n" instead of an exact match).
>
> If you still need to encode the device id here, you should be able to
> easily do something like (device_id << 16) | migration_version -- do
> you think that could work?
>
hi Cornelia,
using string is based on the consideration that we want to make this
version string a thing that can distinguish a mdev, so we incoportate
vendor id, device id to identify parent device first, then mdev type to
describe the mdev based on the parent device. And that's only for gen8 and
gen9. For future platforms, we may incorpate more information, e.g. besides
vendor id and device id, different device revision number, or even a value
in a register on the run may be needed to identify a parent device.
I think it's cleaner than a numeric version between m and n, because in that
case we have to maintain what m's configration is and what n's is.
whenever a mdev type is added (like changing a resolution type in mdev
type) a new version is generated. it's too complicated.
That's why we use current way: version describe parent device + mdev type
elaborately, then vendor driver checks compatibility according to this
information.
Do you think it's all right?
> > +
> > +ssize_t intel_gvt_check_vfio_device_version(struct drm_i915_private *dev_priv,
> > + const char *self, const char *remote)
> > +{
> > +
> > + /* currently only gen8 & gen9 are supported */
> > + if (!IS_GEN(dev_priv, 8) && !IS_GEN(dev_priv, 9))
> > + return -ENODEV;
> > +
> > + if (!is_compatible(self, remote))
> > + return -EINVAL;
>
> I think the meaning of the error codes should really be standardized
> across vendor drivers, if we need a value for "this device does not
> support migration at all". (Your choices look reasonable for that.)
>
Agree. thank you. I'll keep error codes consistently in future.
> > +
> > + return 0;
> > +}
> > diff --git a/drivers/gpu/drm/i915/gvt/gvt.c b/drivers/gpu/drm/i915/gvt/gvt.c
> > index 43f4242062dd..e720465b93d8 100644
> > --- a/drivers/gpu/drm/i915/gvt/gvt.c
> > +++ b/drivers/gpu/drm/i915/gvt/gvt.c
> > @@ -105,14 +105,69 @@ static ssize_t description_show(struct kobject *kobj, struct device *dev,
> > type->weight);
> > }
> >
> > +static ssize_t version_show(struct kobject *kobj, struct device *dev,
> > + char *buf)
> > +{
> > +#ifdef GVT_MIGRATION_VERSION
> > + struct drm_i915_private *i915 = kdev_to_i915(dev);
> > + const char *mdev_type = kobject_name(kobj);
> > +
> > + return intel_gvt_get_vfio_device_version(i915, buf, mdev_type);
> > +#else
> > + /* do not support live migration */
> > + return -EINVAL;
>
> ...but this looks inconsistent. I would expect -ENODEV here, same as
> for non-gen{8,9}.
Right, case "not suppporting live migration" should return -ENODEV.
Thanks:)
> Or simply do not create the attribute at all in that case?
That's also a good choice :)
> > +#endif
> > +}
> > +
> > +static ssize_t version_store(struct kobject *kobj, struct device *dev,
> > + const char *buf, size_t count)
> > +{
> > +#ifdef GVT_MIGRATION_VERSION
> > + char *remote = NULL, *self = NULL;
> > + int len, ret = 0;
> > + struct drm_i915_private *i915 = kdev_to_i915(dev);
> > + const char *mdev_type = kobject_name(kobj);
> > +
> > + len = intel_gvt_get_vfio_device_version_len(i915);
> > + if (len < 0)
> > + return len;
> > +
> > + self = kmalloc(len, GFP_KERNEL);
> > + if (!self)
> > + return -ENOMEM;
> > +
> > + ret = intel_gvt_get_vfio_device_version(i915, self, mdev_type);
> > + if (ret < 0)
> > + goto out;
> > +
> > + remote = kstrndup(buf, count, GFP_KERNEL);
> > + if (!remote) {
> > + ret = -ENOMEM;
> > + goto out;
> > + }
> > +
> > + ret = intel_gvt_check_vfio_device_version(i915, self, remote);
> > +
> > +out:
> > + kfree(self);
> > + kfree(remote);
> > + return (ret < 0 ? ret : count);
> > +#else
> > + /* do not support live migration */
> > + return -EINVAL;
> > +#endif
> > +}
> > +
> > static MDEV_TYPE_ATTR_RO(available_instances);
> > static MDEV_TYPE_ATTR_RO(device_api);
> > static MDEV_TYPE_ATTR_RO(description);
> > +static MDEV_TYPE_ATTR_RW(version);
> >
> > static struct attribute *gvt_type_attrs[] = {
> > &mdev_type_attr_available_instances.attr,
> > &mdev_type_attr_device_api.attr,
> > &mdev_type_attr_description.attr,
> > + &mdev_type_attr_version.attr,
> > NULL,
> > };
> >
> (...)
> _______________________________________________
> intel-gvt-dev mailing list
> intel-gvt-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gvt-dev
More information about the libvir-list
mailing list