Re: [PATCH v4 04/17] iommufd: Document overview of iommufd

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

On 11/8/22 01:48, Jason Gunthorpe wrote:
> From: Kevin Tian <kevin.tian@xxxxxxxxx>
>
> Add iommufd into the documentation tree, and supply initial documentation.
> Much of this is linked from code comments by kdoc.
>
> Tested-by: Nicolin Chen <nicolinc@xxxxxxxxxx>
> Signed-off-by: Kevin Tian <kevin.tian@xxxxxxxxx>
> Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
> ---
>  Documentation/userspace-api/index.rst   |   1 +
>  Documentation/userspace-api/iommufd.rst | 222 ++++++++++++++++++++++++
>  2 files changed, 223 insertions(+)
>  create mode 100644 Documentation/userspace-api/iommufd.rst
>
> diff --git a/Documentation/userspace-api/index.rst b/Documentation/userspace-api/index.rst
> index c78da9ce0ec44e..f16337bdb8520f 100644
> --- a/Documentation/userspace-api/index.rst
> +++ b/Documentation/userspace-api/index.rst
> @@ -25,6 +25,7 @@ place where this information is gathered.
>     ebpf/index
>     ioctl/index
>     iommu
> +   iommufd
>     media/index
>     netlink/index
>     sysfs-platform_profile
> diff --git a/Documentation/userspace-api/iommufd.rst b/Documentation/userspace-api/iommufd.rst
> new file mode 100644
> index 00000000000000..64a135f3055adc
> --- /dev/null
> +++ b/Documentation/userspace-api/iommufd.rst
> @@ -0,0 +1,222 @@
> +.. SPDX-License-Identifier: GPL-2.0+
> +
> +=======
> +IOMMUFD
> +=======
> +
> +:Author: Jason Gunthorpe
> +:Author: Kevin Tian
> +
> +Overview
> +========
> +
> +IOMMUFD is the user API to control the IOMMU subsystem as it relates to managing
> +IO page tables from userspace using file descriptors. It intends to be general
> +and consumable by any driver that wants to expose DMA to userspace. These
> +drivers are eventually expected to deprecate any internal IOMMU logic if exists
they may already/historically implement (eg. vfio_iommu_type1.c)?
> +(e.g. vfio_iommu_type1.c).
> +
> +At minimum iommufd provides universal support of managing I/O address spaces and
> +I/O page tables for all IOMMUs, with room in the design to add non-generic
> +features to cater to specific hardware functionality.
> +
> +In this context the capital letter (IOMMUFD) refers to the subsystem while the
> +small letter (iommufd) refers to the file descriptors created via /dev/iommu for
> +use by userspace.
> +
> +Key Concepts
> +============
> +
> +User Visible Objects
> +--------------------
> +
> +Following IOMMUFD objects are exposed to userspace:
> +
> +- IOMMUFD_OBJ_IOAS, representing an I/O address space (IOAS), allowing map/unmap
> +  of user space memory into ranges of I/O Virtual Address (IOVA).
> +
> +  The IOAS is a functional replacement for the VFIO container, and like the VFIO
> +  container it copies an IOVA map to a list of iommu_domains held within it.
> +
> +- IOMMUFD_OBJ_DEVICE, representing a device that is bound to iommufd by an
> +  external driver.
> +
> +- IOMMUFD_OBJ_HW_PAGETABLE, representing an actual hardware I/O page table
> +  (i.e. a single struct iommu_domain) managed by the iommu driver.
> +
> +  The IOAS has a list of HW_PAGETABLES that share the same IOVA mapping and
> +  it will synchronize its mapping with each member HW_PAGETABLE.
> +
> +All user-visible objects are destroyed via the IOMMU_DESTROY uAPI.
> +
> +The diagram below shows relationship between user-visible objects and kernel
> +datastructures (external to iommufd), with numbers referred to operations
> +creating the objects and links::
> +
> +  _________________________________________________________
> + |                         iommufd                         |
> + |       [1]                                               |
> + |  _________________                                      |
> + | |                 |                                     |
> + | |                 |                                     |
> + | |                 |                                     |
> + | |                 |                                     |
> + | |                 |                                     |
> + | |                 |                                     |
> + | |                 |        [3]                 [2]      |
> + | |                 |    ____________         __________  |
> + | |      IOAS       |<--|            |<------|          | |
> + | |                 |   |HW_PAGETABLE|       |  DEVICE  | |
> + | |                 |   |____________|       |__________| |
> + | |                 |         |                   |       |
> + | |                 |         |                   |       |
> + | |                 |         |                   |       |
> + | |                 |         |                   |       |
> + | |                 |         |                   |       |
> + | |_________________|         |                   |       |
> + |         |                   |                   |       |
> + |_________|___________________|___________________|_______|
> +           |                   |                   |
> +           |              _____v______      _______v_____
> +           | PFN storage |            |    |             |
> +           |------------>|iommu_domain|    |struct device|
> +                         |____________|    |_____________|
> +
> +1. IOMMUFD_OBJ_IOAS is created via the IOMMU_IOAS_ALLOC uAPI. An iommufd can
> +   hold multiple IOAS objects. IOAS is the most generic object and does not
> +   expose interfaces that are specific to single IOMMU drivers. All operations
> +   on the IOAS must operate equally on each of the iommu_domains inside of it.
> +
> +2. IOMMUFD_OBJ_DEVICE is created when an external driver calls the IOMMUFD kAPI
> +   to bind a device to an iommufd. The driver is expected to implement proper a
s/proper/properly?
> +   set of ioctls to allow userspace to initiate the binding operation.
> +   Successful completion of this operation establishes the desired DMA ownership
> +   over the device. The driver must also set the driver_managed_dma flag and
> +   must not touch the device until this operation succeeds.
> +
> +3. IOMMUFD_OBJ_HW_PAGETABLE is created when an external driver calls the IOMMUFD
> +   kAPI to attach a bound device to an IOAS. Similarly the external driver uAPI
> +   allows userspace to initiate the attaching operation. If a compatible
> +   pagetable already exists then it is reused for the attachment. Otherwise a
> +   new pagetable object and iommu_domain is created. Successful completion of
> +   this operation sets up the linkages among IOAS, device and iommu_domain. Once
> +   this completes the device could do DMA.
> +
> +   Every iommu_domain inside the IOAS is also represented to userspace as a
> +   HW_PAGETABLE object.
> +
> +   .. note::
> +
> +      Future IOMMUFD updates will provide an API to create and manipulate the
> +      HW_PAGETABLE directly.
> +
> +A device can only bind to an iommufd due to DMA ownership claim and attach to at
> +most one IOAS object (no support of PASID yet).
> +
> +Currently only PCI device is allowed to use IOMMUFD.
is it still true? device_bind() now takes a struct device *

In [PATCH v4 12/17] iommufd: Add kAPI toward external drivers for
physical devices "PCI" is used at several places
but shouldn't it be removed now?

> +
> +Kernel Datastructure
> +--------------------
> +
> +User visible objects are backed by following datastructures:
> +
> +- iommufd_ioas for IOMMUFD_OBJ_IOAS.
> +- iommufd_device for IOMMUFD_OBJ_DEVICE.
> +- iommufd_hw_pagetable for IOMMUFD_OBJ_HW_PAGETABLE.
> +
> +Several terminologies when looking at these datastructures:
> +
> +- Automatic domain - refers to an iommu domain created automatically when
> +  attaching a device to an IOAS object. This is compatible to the semantics of
> +  VFIO type1.
> +
> +- Manual domain - refers to an iommu domain designated by the user as the
> +  target pagetable to be attached to by a device. Though currently there are
> +  no uAPIs to directly create such domain, the datastructure and algorithms
> +  are ready for handling that use case.
> +
> +- In-kernel user - refers to something like a VFIO mdev that is using the
> +  IOMMUFD access interface to access the IOAS. This starts by creating an
> +  iommufd_access object that is similar to the domain binding a physical device
> +  would do. The access object will then allow converting IOVA ranges into struct
> +  page * lists, or doing direct read/write to an IOVA.
> +
> +iommufd_ioas serves as the metadata datastructure to manage how IOVA ranges are
> +mapped to memory pages, composed of:
> +
> +- struct io_pagetable holding the IOVA map
> +- struct iopt_areas representing populated portions of IOVA
> +- struct iopt_pages representing the storage of PFNs
> +- struct iommu_domain representing the IO page table in the IOMMU
> +- struct iopt_pages_access representing in-kernel users of PFNs
> +- struct xarray pinned_pfns holding a list of pages pinned by in-kernel users
> +
> +Each iopt_pages represents a logical linear array of full PFNs. The PFNs are
> +ultimately derived from userspave VAs via an mm_struct. Once they have been
> +pinned the PFN is stored in IOPTEs of an iommu_domain or inside the pinned_pages
s/is/are
> +xarray if they have been pinned through an iommufd_access.
> +
> +PFN have to be copied between all combinations of storage locations, depending
> +on what domains are present and what kinds of in-kernel "software access" users
> +exists. The mechanism ensures that a page is pinned only once.
> +
> +An io_pagetable is composed of iopt_areas pointing at iopt_pages, along with a
> +list of iommu_domains that mirror the IOVA to PFN map.
> +
> +Multiple io_pagetable-s, through their iopt_area-s, can share a single
> +iopt_pages which avoids multi-pinning and double accounting of page
> +consumption.
> +
> +iommufd_ioas is sharable between subsystems, e.g. VFIO and VDPA, as long as
> +devices managed by different subsystems are bound to a same iommufd.
> +
> +IOMMUFD User API
> +================
> +
> +.. kernel-doc:: include/uapi/linux/iommufd.h
> +
> +IOMMUFD Kernel API
> +==================
> +
> +The IOMMUFD kAPI is device-centric with group-related tricks managed behind the
> +scene. This allows the external drivers calling such kAPI to implement a simple
> +device-centric uAPI for connecting its device to an iommufd, instead of
> +explicitly imposing the group semantics in its uAPI as VFIO does.
> +
> +.. kernel-doc:: drivers/iommu/iommufd/device.c
> +   :export:
> +
> +VFIO and IOMMUFD
> +----------------
> +
> +Connecting a VFIO device to iommufd can be done in two ways.
> +
> +First is a VFIO compatible way by directly implementing the /dev/vfio/vfio
> +container IOCTLs by mapping them into io_pagetable operations. Doing so allows
> +the use of iommufd in legacy VFIO applications by symlinking /dev/vfio/vfio to
> +/dev/iommufd or extending VFIO to SET_CONTAINER using an iommufd instead of a
> +container fd.
> +
> +The second approach directly extends VFIO to support a new set of device-centric
> +user API based on aforementioned IOMMUFD kernel API. It requires userspace
> +change but better matches the IOMMUFD API semantics and easier to support new
> +iommufd features when comparing it to the first approach.
> +
> +Currently both approaches are still work-in-progress.
> +
> +There are still a few gaps to be resolved to catch up with VFIO type1, as
> +documented in iommufd_vfio_check_extension().
> +
> +Future TODOs
> +============
> +
> +Currently IOMMUFD supports only kernel-managed I/O page table, similar to VFIO
> +type1. New features on the radar include:
> +
> + - Binding iommu_domain's to PASID/SSID
> + - Userspace page tables, for ARM, x86 and S390
> + - Kernel bypass'd invalidation of user page tables
> + - Re-use of the KVM page table in the IOMMU
> + - Dirty page tracking in the IOMMU
> + - Runtime Increase/Decrease of IOPTE size
> + - PRI support with faults resolved in userspace
Thanks

Eric




[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux