| Message ID | 1500525154-5200-2-git-send-email-anup.patel@broadcom.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Jul 20, 2017 at 10:02:33AM +0530, Anup Patel wrote: > Not allowing No-IOMMU mode for devices already having > iommu_ops on their bus is very conservative. > > We now have IOMMU (such as ARM SMMU) which can bypass > transcations when IOMMU is not configured for a given > device. In addition, it is not necessary to have all > devices on bus to be upstream to an IOMMU on that bus. How does the SMMU know to bypass in these cases? As I explained before, the driver-specific command line option is the wrong way to go about arranging this. Will
On Thu, 20 Jul 2017 10:02:33 +0530 Anup Patel <anup.patel@broadcom.com> wrote: > Not allowing No-IOMMU mode for devices already having > iommu_ops on their bus is very conservative. > > We now have IOMMU (such as ARM SMMU) which can bypass > transcations when IOMMU is not configured for a given > device. In addition, it is not necessary to have all > devices on bus to be upstream to an IOMMU on that bus. > > Due above reasons, having iommu_present() check for > VFIO No-IOMMU mode is not appropriate. > > Signed-off-by: Anup Patel <anup.patel@broadcom.com> > --- > drivers/vfio/vfio.c | 7 +++---- > 1 file changed, 3 insertions(+), 4 deletions(-) > > diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c > index 330d505..9d90de7 100644 > --- a/drivers/vfio/vfio.c > +++ b/drivers/vfio/vfio.c > @@ -124,11 +124,10 @@ struct iommu_group *vfio_iommu_group_get(struct device *dev) > #ifdef CONFIG_VFIO_NOIOMMU > /* > * With noiommu enabled, an IOMMU group will be created for a device > - * that doesn't already have one and doesn't have an iommu_ops on their > - * bus. We set iommudata simply to be able to identify these groups > - * as special use and for reclamation later. > + * that doesn't already have one. We set iommudata simply to be able > + * to identify these groups as special use and for reclamation later. > */ > - if (group || !noiommu || iommu_present(dev->bus)) > + if (group || !noiommu) > return group; > > group = iommu_group_alloc(); Nak, no-iommu is intentionally conservative. How do we tell the difference between the iommu_ops available to us with a fake group vs a proper group created by the iommu if we ignore iommu_present()? If an iommu exists, use it. Thanks, Alex
diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c index 330d505..9d90de7 100644 --- a/drivers/vfio/vfio.c +++ b/drivers/vfio/vfio.c @@ -124,11 +124,10 @@ struct iommu_group *vfio_iommu_group_get(struct device *dev) #ifdef CONFIG_VFIO_NOIOMMU /* * With noiommu enabled, an IOMMU group will be created for a device - * that doesn't already have one and doesn't have an iommu_ops on their - * bus. We set iommudata simply to be able to identify these groups - * as special use and for reclamation later. + * that doesn't already have one. We set iommudata simply to be able + * to identify these groups as special use and for reclamation later. */ - if (group || !noiommu || iommu_present(dev->bus)) + if (group || !noiommu) return group; group = iommu_group_alloc();
Not allowing No-IOMMU mode for devices already having iommu_ops on their bus is very conservative. We now have IOMMU (such as ARM SMMU) which can bypass transcations when IOMMU is not configured for a given device. In addition, it is not necessary to have all devices on bus to be upstream to an IOMMU on that bus. Due above reasons, having iommu_present() check for VFIO No-IOMMU mode is not appropriate. Signed-off-by: Anup Patel <anup.patel@broadcom.com> --- drivers/vfio/vfio.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)