Re: [RFC PATCH 0/9] iommu/iommufd: Add hypervisor external attach

[Jacob Pan]

Hi Mostafa,

On Wed, 21 Oct 2026 10:55:36 +0000
Mostafa Saleh <smostafa@xxxxxxxxxx> wrote:

> On Fri, Sep 25, 2026 at 12:07:33PM -0700, Jacob Pan wrote:
> > Hi,
> > 
> > This RFC supports guest VM device assignment from a Linux host
> > using the Microsoft Hypervisor (MSHV). The goal is that a common
> > set of flows and infrastructure can be applied for Xen and other
> > type 1 hypervisors.
> > 
> > Background:
> > 
> > The IOMMUFD/VFIO stack used for device assignment has largely
> > evolved around the KVM/QEMU model, at least as the first enabling
> > target. In that model, the Linux host owns the host-side physical
> > IOMMU programming as well as the physical PCI devices.
> > 
> > Architecturally, MSHV and Xen differ from this KVM/QEMU ownership
> > model. At a high level:
> > 
> >   +------------------+
> >   | Guest VM         |
> >   | virtual IOMMU    |
> >   +--------+---------+
> >            | vIOMMU ops
> >            v
> >   +------------------+    external attach     +------------------+
> >   | MSHV             |<-----------------------| Linux host       |
> >   +--------+---------+                        +--------+---------+
> >            | programs                                  | owns
> >            v                                           v
> >   +------------------+                        +------------------+
> >   | HW IOMMU         |<---------------------->| PCI dev          |
> >   +------------------+          DMA           +------------------+
> > 
> > For MSHV, the specific ownership model is as follows:
> >   1. The Linux host OS owns the physical device and exposes it
> > through VFIO/IOMMUFD, but it does not directly own any of the IOMMU
> > hardware. 2. The device node (/dev/mshv) owns the VM partition file
> > used by userspace. 3. MSHV owns most of the physical IOMMU
> > programming, including stage-2 I/O page tables.
> >   4. The Linux MSHV root pvIOMMU driver bridges IOMMUFD/VFIO UAPIs
> > to hardware IOMMU programming through MSHV hypercalls.
> >   5. Guest virtual IOMMU emulation is done in MSHV rather than the
> > userspace VMM.
> >   6. Stage-2 IOPT may be shared between CPU and IOMMU. It is
> > managed by MSHV and abstracted from the Linux host, including TLB
> > synchronization.
> > 
> > These structural differences from the KVM/QEMU model create several
> > needs for IOMMUFD support:
> >   1. Move a device to a guest whose I/O page table is externally
> > managed. This patchset refers to that operation as external attach.
> > Unlike conventional IOMMU vendor drivers, where translation is
> > managed within the driver through paging domains, external attach
> > means that translation is managed outside the driver, such as by
> > the hypervisor. 2. Identify the guest-visible device across VMM,
> > host pvIOMMU driver, and hypervisor namespaces.
> >   3. Identify the guest partition in the host pvIOMMU driver for
> > device attach hypercalls.
> >   4. Tie the device-attachment lifetime to the hypervisor partition.
> >   5. Coordinate guest-memory pinning and mapping for DMA between
> > MSHV and the external IOMMU domain attachment.
> > 
> > External attach has a few caveats compared with KVM/QEMU, where
> > nesting is typically handled by the host IOMMU driver in one
> > transaction after the userspace VMM uses IOMMUFD to form nested
> > domains:
> > 
> >   1. MSHV creates a temporal split between host and guest
> > device-domain attach. The host first attaches S2 through a
> > hypercall; later, if virtual IOMMU is present, the guest
> >      allocates S1 and establishes nesting with MSHV directly
> > through a pvIOMMU or emulated vIOMMU.
> >   2. Guest-memory pinning before mapping is covered by the
> > MSHV/userspace VMM contract, not by the VFIO/IOMMUFD pin-and-map
> > DMA flow. 3. Device capabilities such as ATS and PASID need a
> > contract among MSHV, the userspace VMM, and the host pvIOMMU
> > driver. At the time of host external attach, only S2 is attached,
> > so the guest's later ATS/PASID usage is not yet known. For example,
> > whether to expose and enable PASID in the device could depend on
> > whether the userspace VMM creates a guest virtual IOMMU.
> > 
> > This RFC proposes an IOMMUFD external attach model for those cases.
> > The normal Linux-managed paging or nested domain is not the object
> > that ultimately owns the translation context. Instead, userspace
> > identifies a VM/domain object, binds a physical device to a
> > VM-visible device identity, and asks the host pvIOMMU driver to
> > attach the device to an externally managed address space.
> > 
> > The series makes that relationship explicit in IOMMUFD. The model
> > is anchored around a hypervisor vIOMMU object that represents the
> > per-VM partition abstraction. It introduces:
> > 
> >   1. IOMMU_DOMAIN_EXTERNAL for externally managed translation
> > domains. 2. IOMMU_VIOMMU_TYPE_HYPERVISOR for the file-backed VM
> > partition. 3. IOMMU_HWPT_EXTERNAL for user-visible external type
> > IOMMU domains.
> > 
> > It reuses the IOMMUFD vDEVICE object to carry the guest-VM-scoped
> > logical device ID visible to the hypervisor.
> >   
> 
> This is very similar to pKVM on arm64 also, where the IOMMU is managed
> by the hypervisor, and for device assignment, the host doesn't control
> it for the guest, but the hypervisor (in EL2), and there will be a
> channel between the guest and pKVM directly after the guest runs. 
> 
> The host only sets the virtual to physical topology (how many vIOMMUs,
> vSID -> SID mappings) and describes that for the hypervisor through
> hypercalls.
> In Android, we have an out of tree uAPI through KVM/VFIO bridge
> IOCTLs that proxies those hypercalls.
> 
> In that case, the host just creates a virtual IOMMU and the guest sees
> a paravirtual one which runs with single stage collapsed translation,
> although I don't see a difference on the host side if nested
> translation was used beyond an argument change to the hypercalls.
> 
By "argument change", do you mean that the host would pass the guest S1
context in the device-attach hypercall for nested translation? 

IMHO, I don't see any change needed for the host-side attach between hw
nested and collapsed single stage, if guest and host attach are
split. I think it is up to pKVM or MSHV to handle guest virtual iommu
mapping for both cases.

> My goal is to converge to IOMMUFD and I believe this series is the way
> forward, By relying on the vDEVICE virtual toplogy uAPI and the
> external attach introduced here.
> 
> I will try to prototype something based on that to prove that it
> works. However that work has a dependency on the host SMMUv3 support
> [1] which needs to land first.
> 
> [1] https://lore.kernel.org/all/20261015200903.2735565-1-smostafa@xxxxxxxxxx/
> 
Same here, our pviommu root driver must land first.
https://lore.kernel.org/linux-iommu/20260929223622.778617-5-mrathor@xxxxxxxxxxxxxxxxxxx/

My next step is to rebase on Aneesh's v7 such that we have a converged
iommufd viommu layer, i will remove mshv specific helper registration
in this series.
https://lore.kernel.org/linux-iommu/20261008055955.4014342-1-aneesh.kumar@xxxxxxxxxx/

> Thanks,
> Mostafa
> 
> 
> > Roughly, the VMM flow is as follows:
> >   1. iommufd = open("/dev/iommu")
> >   2. vfio_fd = open("/dev/vfio/devices/vfioN")
> >   3. dev_id = bind_vfio_device(vfio_fd, iommufd)
> >   4. create_mshv_partition(&mshv_fd, &vm_fd)
> >      // Creates the guest vIOMMU and configures guest-memory
> > pinning policy. 5. alloc_hypervisor_viommu(iommufd, &viommu_id,
> > dev_id, vm_fd) // Uses the hypervisor vIOMMU type.
> >   6. alloc_vdevice(iommufd, viommu_id, dev_id, vm_device_id,
> > &vdevice_id) 7. alloc_external_hwpt(iommufd, dev_id, viommu_id,
> > &external_hwpt_id) 8. attach_external_hwpt(vfio_fd,
> > external_hwpt_id)
> > 
> > Step 4 may be completed before steps 1-3, but it must precede step
> > 5 because the partition file is required to allocate the hypervisor
> > vIOMMU.
> > 
> > The attach_external_hwpt() operation enters the MSHV root pvIOMMU
> > driver's external-HWPT device_attach path:  
> >   -> resolve partition ID from hypervisor vIOMMU
> >   -> resolve logical device ID from vDEVICE
> >   -> invoke the partition prepare_attach callback to force pinning
> > if policy requires  
> >   -> MSHV attach-device hypercall  
> > 
> > Object lifetime:
> >   - The hypervisor vIOMMU takes a reference on the MSHV partition
> > file and holds it until the vIOMMU is destroyed. Closing the
> > userspace copy of the partition fd therefore does not destroy the
> > partition while the vIOMMU still exists.
> >   - Each vDEVICE and external HWPT holds a reference on its parent
> > vIOMMU. The vIOMMU cannot be destroyed until these child objects
> > are removed.
> >   - An attached device holds a reference on the external HWPT, so
> > userspace cannot destroy the HWPT while the device is attached.
> >   - Closing the IOMMUFD fd destroys the object graph from the
> > leaves upward. Once the child objects are gone, destroying the
> > vIOMMU drops its partition file reference and allows the MSHV
> > partition to be released.
> > 
> > Thus, the external-attach lifetime is anchored directly by IOMMUFD
> > object references and the partition file. It does not require a
> > separate VFIO-hypervisor bridge, such as kvm-vfio: the MSHV root
> > IOMMU driver's detach operation directly tears down the
> > hypervisor-visible external-attach state.
> > 
> > The root pvIOMMU driver changes in this RFC address the following
> > concerns received from the older version[1]:
> >   1. The target partition comes from the fd-backed hypervisor
> > vIOMMU created by userspace, not from the thread issuing the attach.
> >   2. External attach is represented explicitly, avoiding
> > paging-domain hacks. 3. Guest memory is pinned exactly once. In the
> > older model, the same memory could be pinned once by /dev/mshv and
> > again by the VFIO DMA map ioctl.
> > 
> > This RFC builds on the MSHV PCI passthrough/root pvIOMMU base work
> > [2]. The mergeable MSHV driver changes should be added on top of a
> > future version of that base root pvIOMMU driver.
> > 
> > Feedback would be especially useful on:
> > 
> >   1. Whether IOMMU_DOMAIN_EXTERNAL is the right IOMMU core
> > representation for an externally managed translation (Xen, perhaps
> > TSM [4]). 2. Whether IOMMU_VIOMMU_TYPE_HYPERVISOR/VDEVICE can be
> > common to MSHV, Xen, and similar type-1 hypervisor flows, and
> > whether it can cover the semantics in CoCo TDISP that would
> > otherwise require tsm_bind() [3]. 3. Whether a file-backed
> > VM/domain reference is the right way to anchor the external attach
> > lifetime. 4. Whether vDEVICE virt_id is the right generic source
> > for the hypervisor-assigned logical device ID.
> > 
> > At LPC 2026, I will be presenting this work and have open
> > discussions. https://lpc.events/event/20/contributions/2520/
> > 
> > Future work:
> > 1. ATS/PASID policy integration with the root pvIOMMU driver.
> > ATS/PASID policy is intentionally left as a separate contract
> > issue. I do not think the root pvIOMMU driver should blindly enable
> > or disable PASID/ATS on a device without knowing the guest
> > configuration or receiving acknowledgment from MSHV, which owns the
> > device-TLB flush for S2 changes.
> > 
> > 2. Live update support
> > The current LUO proposal preserves DMA mappings through
> > IOMMUFD-owned state. External attach requires a different
> > preservation model because the S2 IOPT is managed by MSHV rather
> > than IOMMUFD.
> > 
> > 3. vDEVICE lifetime during external attachment
> > An attached external HWPT does not currently retain the vDEVICE
> > used by MSHV to resolve the guest logical device ID. We can either
> > let the core retain the vDEVICE identity for external attach, or
> > provide an opt-in mechanism to the vendor IOMMU drivers as needed.
> > I am not sure if all type 1 hypervisors require a vDEVICE identity.
> > 
> > 4. Revokable dma-buf support for P2PDMA
> > Now that MSHV, instead of IOMMUFD, manages the S2 IOPT, we need to
> > reconsider how revokable dma-bufs are handled. Perhaps MSHV can be
> > the importer of revokable dma-bufs for P2PDMA and unmap them when
> > the dma-buf is revoked.
> > 
> > [1] https://lore.kernel.org/linux-iommu/20260731223427.2554388-1-mrathor@xxxxxxxxxxxxxxxxxxx/
> > [2] https://lore.kernel.org/all/20260902235609.GG2890729@xxxxxxxx/
> > [3] https://lore.kernel.org/all/20260921225028.4007330-1-mrathor@xxxxxxxxxxxxxxxxxxx/T/#t
> > [4] https://lore.kernel.org/linux-iommu/yq5a33uxka2g.fsf@xxxxxxxxxx/
> > 
> > Patch groups:
> > 
> > IOMMUFD core and selftest:
> > 
> >   1. iommu: Introduce external attach domain type
> >   2. iommufd: Introduce hypervisor vIOMMU type
> >   3. iommufd: Add external HWPT support
> >   4. iommufd/selftest: Add hypervisor external attach backend
> > 
> > MSHV:
> > 
> >   5. mshv: Add partition file identity helper
> >   6. mshv: Add prepare callback for external device attach
> > 
> > MSHV root pIOMMU driver:
> > 
> >   7. iommu/hyperv: Split root IOMMU declarations
> >   8. iommu/hyperv: Add fd-backed vIOMMU support
> >   9. iommu/hyperv: Add IOMMUFD external domains
> > 
> > 
> > Thanks,
> > Jacob
> > 
> > Jacob Pan (9):
> >   iommu: Introduce external attach domain type
> >   iommufd: Introduce hypervisor vIOMMU type
> >   iommufd: Add external HWPT support
> >   iommufd/selftest: Add hypervisor external attach backend
> >   mshv: Add partition file identity helper
> >   mshv: Add prepare callback for external device attach
> >   iommu/hyperv: Split root IOMMU declarations
> >   iommu/hyperv: Add fd-backed vIOMMU support
> >   iommu/hyperv: Add IOMMUFD external domains
> > 
> >  drivers/hv/hv_common.c                  |  91 ++++++++++++++++
> >  drivers/hv/mshv_regions.c               |   6 ++
> >  drivers/hv/mshv_root.h                  |   1 +
> >  drivers/hv/mshv_root_main.c             |  71 ++++++++++++-
> >  drivers/iommu/amd/iommufd.c             |  10 +-
> >  drivers/iommu/hyperv/Makefile           |   2 +-
> >  drivers/iommu/hyperv/hv-iommu-iommufd.c | 135
> > ++++++++++++++++++++++++ drivers/iommu/hyperv/hv-iommu-root.c    |
> > 102 ++++++++++++++++-- drivers/iommu/hyperv/hv-iommu.h         |
> > 40 +++++++ drivers/iommu/iommu.c                   |   2 +
> >  drivers/iommu/iommufd/device.c          |   9 +-
> >  drivers/iommu/iommufd/hw_pagetable.c    |  88 +++++++++++++--
> >  drivers/iommu/iommufd/iommufd_private.h |   7 ++
> >  drivers/iommu/iommufd/main.c            |   4 +
> >  drivers/iommu/iommufd/selftest.c        | 120 +++++++++++++++++++--
> >  drivers/iommu/iommufd/viommu.c          |  37 ++++---
> >  include/asm-generic/mshyperv.h          |  42 ++++++++
> >  include/linux/iommu.h                   |  17 ++-
> >  include/linux/iommufd.h                 |  11 ++
> >  include/uapi/linux/iommufd.h            |  42 +++++++-
> >  20 files changed, 790 insertions(+), 47 deletions(-)
> >  create mode 100644 drivers/iommu/hyperv/hv-iommu-iommufd.c
> >  create mode 100644 drivers/iommu/hyperv/hv-iommu.h
> > 
> > -- 
> > 2.43.0
> > 
> >