> From: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> > Sent: Wednesday, January 13, 2021 10:50 AM > > Hi Jean, > > On 1/12/21 5:16 PM, Jean-Philippe Brucker wrote: > > Hi Baolu, > > > > On Tue, Jan 12, 2021 at 12:31:23PM +0800, Lu Baolu wrote: > >> Hi Jean, > >> > >> On 1/8/21 10:52 PM, Jean-Philippe Brucker wrote: > >>> Some devices manage I/O Page Faults (IOPF) themselves instead of > relying > >>> on PCIe PRI or Arm SMMU stall. Allow their drivers to enable SVA without > >>> mandating IOMMU-managed IOPF. The other device drivers now need to > first > >>> enable IOMMU_DEV_FEAT_IOPF before enabling > IOMMU_DEV_FEAT_SVA. > >>> > >>> Signed-off-by: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx> > >>> --- > >>> Cc: Arnd Bergmann <arnd@xxxxxxxx> > >>> Cc: David Woodhouse <dwmw2@xxxxxxxxxxxxx> > >>> Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> > >>> Cc: Joerg Roedel <joro@xxxxxxxxxx> > >>> Cc: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> > >>> Cc: Will Deacon <will@xxxxxxxxxx> > >>> Cc: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx> > >>> Cc: Zhou Wang <wangzhou1@xxxxxxxxxxxxx> > >>> --- > >>> include/linux/iommu.h | 20 +++++++++++++++++--- > >>> 1 file changed, 17 insertions(+), 3 deletions(-) > >>> > >>> diff --git a/include/linux/iommu.h b/include/linux/iommu.h > >>> index 583c734b2e87..701b2eeb0dc5 100644 > >>> --- a/include/linux/iommu.h > >>> +++ b/include/linux/iommu.h > >>> @@ -156,10 +156,24 @@ struct iommu_resv_region { > >>> enum iommu_resv_type type; > >>> }; > >>> -/* Per device IOMMU features */ > >>> +/** > >>> + * enum iommu_dev_features - Per device IOMMU features > >>> + * @IOMMU_DEV_FEAT_AUX: Auxiliary domain feature > >>> + * @IOMMU_DEV_FEAT_SVA: Shared Virtual Addresses > >>> + * @IOMMU_DEV_FEAT_IOPF: I/O Page Faults such as PRI or Stall. > Generally using > >>> + * %IOMMU_DEV_FEAT_SVA > requires %IOMMU_DEV_FEAT_IOPF, but > >>> + * some devices manage I/O Page Faults themselves > instead > >>> + * of relying on the IOMMU. When supported, this > feature > >>> + * must be enabled before and disabled after > >>> + * %IOMMU_DEV_FEAT_SVA. > >> > >> Is this only for SVA? We may see more scenarios of using IOPF. For > >> example, when passing through devices to user level, the user's pages > >> could be managed dynamically instead of being allocated and pinned > >> statically. > > > > Hm, isn't that precisely what SVA does? I don't understand the > > difference. That said FEAT_IOPF doesn't have to be only for SVA. It could > > later be used as a prerequisite some another feature. For special cases > > device drivers can always use the iommu_register_device_fault_handler() > > API and handle faults themselves. > > From the perspective of IOMMU, there is a little difference between > these two. For SVA, the page table is from CPU side, so IOMMU only needs > to call handle_mm_fault(); For above pass-through case, the page table > is from IOMMU side, so the device driver (probably VFIO) needs to > register a fault handler and call iommu_map/unmap() to serve the page > faults. > > If we think about the nested mode (or dual-stage translation), it's more > complicated since the kernel (probably VFIO) handles the second level > page faults, while the first level page faults need to be delivered to > user-level guest. Obviously, this hasn't been fully implemented in any > IOMMU driver. > Thinking more the confusion might come from the fact that we mixed hardware capability with software capability. IOMMU_FEAT describes the hardware capability. When FEAT_IOPF is set, it purely means whatever page faults that are enabled by the software are routed through the IOMMU. Nothing more. Then the software (IOMMU drivers) may choose to support only limited faulting scenarios and then evolve to support more complex usages gradually. For example, the intel-iommu driver only supports 1st-level fault (thus SVA) for now, while FEAT_IOPF as a separate feature may give the impression that 2nd-level faults are also allowed. From this angle once we start to separate page fault from SVA, we may also need a way to report the software capability (e.g. a set of faulting categories) and also extend iommu_register_device_fault_handler to allow specifying which category is enabled respectively. The example categories could be: - IOPF_BIND, for page tables which are bound/linked to the IOMMU. Apply to bare metal SVA and guest SVA case; - IOPF_MAP, for page tables which are managed through explicit IOMMU map interfaces. Apply to removing VFIO pinning restriction; Both categories can be enabled together in nested translation, with additional information provided to differentiate them in fault information. Using paging/staging level doesn't make much sense as it's IOMMU driver's internal knowledge, e.g. VT-d driver plans to use 1st level for GPA if no nesting and then turn to 2nd level when nesting is enabled. Thanks Kevin