| Message ID | 20240613092359.847145-2-eric.auger@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | VIRTIO-IOMMU/VFIO: Fix host iommu geometry handling for hotplugged devices | expand |
On 6/13/24 11:20 AM, Eric Auger wrote: > Store the agent device (VFIO or VDPA) in the host IOMMU device. > This will allow easy access to some of its resources. > > Signed-off-by: Eric Auger <eric.auger@redhat.com> > Reviewed-by: Zhenzhong Duan <zhenzhong.duan@intel.com> > --- > include/sysemu/host_iommu_device.h | 1 + > hw/vfio/container.c | 1 + > hw/vfio/iommufd.c | 2 ++ > 3 files changed, 4 insertions(+) > > diff --git a/include/sysemu/host_iommu_device.h b/include/sysemu/host_iommu_device.h > index a57873958b..3e5f058e7b 100644 > --- a/include/sysemu/host_iommu_device.h > +++ b/include/sysemu/host_iommu_device.h > @@ -34,6 +34,7 @@ struct HostIOMMUDevice { > Object parent_obj; > > char *name; > + void *agent; /* pointer to agent device, ie. VFIO or VDPA device */ > HostIOMMUDeviceCaps caps; > }; > > diff --git a/hw/vfio/container.c b/hw/vfio/container.c > index 26e6f7fb4f..b728b978a2 100644 > --- a/hw/vfio/container.c > +++ b/hw/vfio/container.c > @@ -1145,6 +1145,7 @@ static bool hiod_legacy_vfio_realize(HostIOMMUDevice *hiod, void *opaque, > > hiod->name = g_strdup(vdev->name); > hiod->caps.aw_bits = vfio_device_get_aw_bits(vdev); > + hiod->agent = opaque; > > return true; > } > diff --git a/hw/vfio/iommufd.c b/hw/vfio/iommufd.c > index 409ed3dcc9..dbdae1adbb 100644 > --- a/hw/vfio/iommufd.c > +++ b/hw/vfio/iommufd.c > @@ -631,6 +631,8 @@ static bool hiod_iommufd_vfio_realize(HostIOMMUDevice *hiod, void *opaque, > struct iommu_hw_info_vtd vtd; > } data; > > + hiod->agent = opaque; > + This opaque pointer could be assigned in vfio_attach_device(). Talking of which, why are we passing a 'VFIODevice *' parameter to HostIOMMUDeviceClass::realize ? I don't see a good reason I think a 'VFIOContainerBase *' would be more appropriate since 'HostIOMMUDevice' represents a device on the host which is common to all VFIO devices. In that case, HostIOMMUDevice::agent wouldn't need to be opaque anymore. It could simply be a 'VFIOContainerBase *' and hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the 'iova_ranges' from the 'VFIOContainerBase *' directly. This means some rework : * vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead. * HostIOMMUDevice::name would be removed. This is just for error messages. * hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be. That said, I think we need the QOMification changes first. Thanks, C. > if (!iommufd_backend_get_device_info(vdev->iommufd, vdev->devid, > &type, &data, sizeof(data), errp)) { > return false;
Hi Cédric, On 6/14/24 11:13, Cédric Le Goater wrote: > On 6/13/24 11:20 AM, Eric Auger wrote: >> Store the agent device (VFIO or VDPA) in the host IOMMU device. >> This will allow easy access to some of its resources. >> >> Signed-off-by: Eric Auger <eric.auger@redhat.com> >> Reviewed-by: Zhenzhong Duan <zhenzhong.duan@intel.com> >> --- >> include/sysemu/host_iommu_device.h | 1 + >> hw/vfio/container.c | 1 + >> hw/vfio/iommufd.c | 2 ++ >> 3 files changed, 4 insertions(+) >> >> diff --git a/include/sysemu/host_iommu_device.h >> b/include/sysemu/host_iommu_device.h >> index a57873958b..3e5f058e7b 100644 >> --- a/include/sysemu/host_iommu_device.h >> +++ b/include/sysemu/host_iommu_device.h >> @@ -34,6 +34,7 @@ struct HostIOMMUDevice { >> Object parent_obj; >> char *name; >> + void *agent; /* pointer to agent device, ie. VFIO or VDPA device */ >> HostIOMMUDeviceCaps caps; >> }; >> diff --git a/hw/vfio/container.c b/hw/vfio/container.c >> index 26e6f7fb4f..b728b978a2 100644 >> --- a/hw/vfio/container.c >> +++ b/hw/vfio/container.c >> @@ -1145,6 +1145,7 @@ static bool >> hiod_legacy_vfio_realize(HostIOMMUDevice *hiod, void *opaque, >> hiod->name = g_strdup(vdev->name); >> hiod->caps.aw_bits = vfio_device_get_aw_bits(vdev); >> + hiod->agent = opaque; >> return true; >> } >> diff --git a/hw/vfio/iommufd.c b/hw/vfio/iommufd.c >> index 409ed3dcc9..dbdae1adbb 100644 >> --- a/hw/vfio/iommufd.c >> +++ b/hw/vfio/iommufd.c >> @@ -631,6 +631,8 @@ static bool >> hiod_iommufd_vfio_realize(HostIOMMUDevice *hiod, void *opaque, >> struct iommu_hw_info_vtd vtd; >> } data; >> + hiod->agent = opaque; >> + > > This opaque pointer could be assigned in vfio_attach_device(). > > Talking of which, why are we passing a 'VFIODevice *' parameter to > HostIOMMUDeviceClass::realize ? I don't see a good reason > > I think a 'VFIOContainerBase *' would be more appropriate since > 'HostIOMMUDevice' represents a device on the host which is common > to all VFIO devices. > > In that case, HostIOMMUDevice::agent wouldn't need to be opaque > anymore. It could simply be a 'VFIOContainerBase *' and > hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the > 'iova_ranges' from the 'VFIOContainerBase *' directly. > > This means some rework : > > * vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead. > * HostIOMMUDevice::name would be removed. This is just for error > messages. > * hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be. > > That said, I think we need the QOMification changes first. OK I need to review your series first. At the moment I have just addressed Zhenzhong's comment in v4, just sent. Thanks Eric > > Thanks, > > C. > > > > >> if (!iommufd_backend_get_device_info(vdev->iommufd, vdev->devid, >> &type, &data, >> sizeof(data), errp)) { >> return false; >
>> Talking of which, why are we passing a 'VFIODevice *' parameter to >> HostIOMMUDeviceClass::realize ? I don't see a good reason >> >> I think a 'VFIOContainerBase *' would be more appropriate since >> 'HostIOMMUDevice' represents a device on the host which is common >> to all VFIO devices. >> >> In that case, HostIOMMUDevice::agent wouldn't need to be opaque >> anymore. It could simply be a 'VFIOContainerBase *' and >> hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the >> 'iova_ranges' from the 'VFIOContainerBase *' directly. >> >> This means some rework : >> >> * vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead. >> * HostIOMMUDevice::name would be removed. This is just for error >> messages. >> * hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be. >> >> That said, I think we need the QOMification changes first. > > OK I need to review your series first. At the moment I have just > addressed Zhenzhong's comment in v4, just sent. Yep. Just take a look at mine. If both of you agree with above proposal, I can care of it and resend all 3. It's a small change. Thanks, C.
Hi Cédric, >-----Original Message----- >From: Cédric Le Goater <clg@redhat.com> >Sent: Friday, June 14, 2024 6:05 PM >To: eric.auger@redhat.com; eric.auger.pro@gmail.com; qemu- >devel@nongnu.org; qemu-arm@nongnu.org; mst@redhat.com; jean- >philippe@linaro.org; peter.maydell@linaro.org; yanghliu@redhat.com; Duan, >Zhenzhong <zhenzhong.duan@intel.com> >Cc: alex.williamson@redhat.com; jasowang@redhat.com; >pbonzini@redhat.com; berrange@redhat.com >Subject: Re: [PATCH v3 1/7] HostIOMMUDevice: Store the VFIO/VDPA agent > > >>> Talking of which, why are we passing a 'VFIODevice *' parameter to >>> HostIOMMUDeviceClass::realize ? I don't see a good reason >>> >>> I think a 'VFIOContainerBase *' would be more appropriate since >>> 'HostIOMMUDevice' represents a device on the host which is common >>> to all VFIO devices. >>> >>> In that case, HostIOMMUDevice::agent wouldn't need to be opaque >>> anymore. It could simply be a 'VFIOContainerBase *' and >>> hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the >>> 'iova_ranges' from the 'VFIOContainerBase *' directly. >>> >>> This means some rework : >>> >>> * vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead. >>> * HostIOMMUDevice::name would be removed. This is just for error >>> messages. >>> * hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be. >>> >>> That said, I think we need the QOMification changes first. >> >> OK I need to review your series first. At the moment I have just >> addressed Zhenzhong's comment in v4, just sent. > >Yep. Just take a look at mine. If both of you agree with above >proposal, I can care of it and resend all 3. It's a small change. I would suggest using opaque pointer and VFIODevice for two reasons, 1. in nesting series vIOMMU needs to attach/detaching hwpt which is VFIODevice operations. See https://github.com/yiliu1765/qemu/commit/3ca559d35adc9840555e361a56708af4c6338b3d 2. If we plan to support VDPA Device in future, the opaque pointer can also point to an VDPADevice structure. Thanks Zhenzhong
On 6/17/24 3:25 AM, Duan, Zhenzhong wrote: > Hi Cédric, > >> -----Original Message----- >> From: Cédric Le Goater <clg@redhat.com> >> Sent: Friday, June 14, 2024 6:05 PM >> To: eric.auger@redhat.com; eric.auger.pro@gmail.com; qemu- >> devel@nongnu.org; qemu-arm@nongnu.org; mst@redhat.com; jean- >> philippe@linaro.org; peter.maydell@linaro.org; yanghliu@redhat.com; Duan, >> Zhenzhong <zhenzhong.duan@intel.com> >> Cc: alex.williamson@redhat.com; jasowang@redhat.com; >> pbonzini@redhat.com; berrange@redhat.com >> Subject: Re: [PATCH v3 1/7] HostIOMMUDevice: Store the VFIO/VDPA agent >> >> >>>> Talking of which, why are we passing a 'VFIODevice *' parameter to >>>> HostIOMMUDeviceClass::realize ? I don't see a good reason >>>> >>>> I think a 'VFIOContainerBase *' would be more appropriate since >>>> 'HostIOMMUDevice' represents a device on the host which is common >>>> to all VFIO devices. >>>> >>>> In that case, HostIOMMUDevice::agent wouldn't need to be opaque >>>> anymore. It could simply be a 'VFIOContainerBase *' and >>>> hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the >>>> 'iova_ranges' from the 'VFIOContainerBase *' directly. >>>> >>>> This means some rework : >>>> >>>> * vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead. >>>> * HostIOMMUDevice::name would be removed. This is just for error >>>> messages. >>>> * hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be. >>>> >>>> That said, I think we need the QOMification changes first. >>> >>> OK I need to review your series first. At the moment I have just >>> addressed Zhenzhong's comment in v4, just sent. >> >> Yep. Just take a look at mine. If both of you agree with above >> proposal, I can care of it and resend all 3. It's a small change. > > I would suggest using opaque pointer and VFIODevice for two reasons, > 1. in nesting series vIOMMU needs to attach/detaching hwpt which is VFIODevice operations. > See https://github.com/yiliu1765/qemu/commit/3ca559d35adc9840555e361a56708af4c6338b3d OK. I realized later on that we needed the device id anyhow. > > 2. If we plan to support VDPA Device in future, the opaque pointer can also point > to an VDPADevice structure. Thanks, C.
diff --git a/include/sysemu/host_iommu_device.h b/include/sysemu/host_iommu_device.h index a57873958b..3e5f058e7b 100644 --- a/include/sysemu/host_iommu_device.h +++ b/include/sysemu/host_iommu_device.h @@ -34,6 +34,7 @@ struct HostIOMMUDevice { Object parent_obj; char *name; + void *agent; /* pointer to agent device, ie. VFIO or VDPA device */ HostIOMMUDeviceCaps caps; }; diff --git a/hw/vfio/container.c b/hw/vfio/container.c index 26e6f7fb4f..b728b978a2 100644 --- a/hw/vfio/container.c +++ b/hw/vfio/container.c @@ -1145,6 +1145,7 @@ static bool hiod_legacy_vfio_realize(HostIOMMUDevice *hiod, void *opaque, hiod->name = g_strdup(vdev->name); hiod->caps.aw_bits = vfio_device_get_aw_bits(vdev); + hiod->agent = opaque; return true; } diff --git a/hw/vfio/iommufd.c b/hw/vfio/iommufd.c index 409ed3dcc9..dbdae1adbb 100644 --- a/hw/vfio/iommufd.c +++ b/hw/vfio/iommufd.c @@ -631,6 +631,8 @@ static bool hiod_iommufd_vfio_realize(HostIOMMUDevice *hiod, void *opaque, struct iommu_hw_info_vtd vtd; } data; + hiod->agent = opaque; + if (!iommufd_backend_get_device_info(vdev->iommufd, vdev->devid, &type, &data, sizeof(data), errp)) { return false;