docs/specs: add riscv-iommu

Add a simple guideline to use the existing RISC-V IOMMU support we just
added.

This doc will be updated once we add the riscv-iommu-sys device.

Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Message-ID: <20241016204038.649340-13-dbarboza@ventanamicro.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>

3 files changed, 104 insertions, 0 deletions

diff --git a/docs/specs/index.rst b/docs/specs/index.rst
index 6495ed5ed9..ff5a1f03da 100644
--- a/docs/specs/index.rst
+++ b/docs/specs/index.rst
@@ -36,3 +36,4 @@ guest hardware that is specific to QEMU.
    vmgenid
    rapl-msr
    rocker
+   riscv-iommu
diff --git a/docs/specs/riscv-iommu.rst b/docs/specs/riscv-iommu.rst
new file mode 100644
index 0000000000..463f4cffb6
--- /dev/null
+++ b/docs/specs/riscv-iommu.rst
@@ -0,0 +1,90 @@
+.. _riscv-iommu:
+
+RISC-V IOMMU support for RISC-V machines
+========================================
+
+QEMU implements a RISC-V IOMMU emulation based on the RISC-V IOMMU spec
+version 1.0 `iommu1.0`_.
+
+The emulation includes a PCI reference device, riscv-iommu-pci, that QEMU
+RISC-V boards can use.  The 'virt' RISC-V machine is compatible with this
+device.
+
+riscv-iommu-pci reference device
+--------------------------------
+
+This device implements the RISC-V IOMMU emulation as recommended by the section
+"Integrating an IOMMU as a PCIe device" of `iommu1.0`_: a PCI device with base
+class 08h, sub-class 06h and programming interface 00h.
+
+As a reference device it doesn't implement anything outside of the specification,
+so it uses a generic default PCI ID given by QEMU: 1b36:0014.
+
+To include the device in the 'virt' machine:
+
+.. code-block:: bash
+
+  $ qemu-system-riscv64 -M virt -device riscv-iommu-pci,[optional_pci_opts] (...)
+
+This will add a RISC-V IOMMU PCI device in the board following any additional
+PCI parameters (like PCI bus address).  The behavior of the RISC-V IOMMU is
+defined by the spec but its operation is OS dependent.
+
+As of this writing the existing Linux kernel support `linux-v8`_, not yet merged,
+does not have support for features like VFIO passthrough.  The IOMMU emulation
+was tested using a public Ventana Micro Systems kernel repository in
+`ventana-linux`_.  This kernel is based on `linux-v8`_ with additional patches that
+enable features like KVM VFIO passthrough with irqbypass.  Until the kernel support
+is feature complete feel free to use the kernel available in the Ventana Micro Systems
+mirror.
+
+The current Linux kernel support will use the IOMMU device to create IOMMU groups
+with any eligible cards available in the system, regardless of factors such as the
+order in which the devices are added in the command line.
+
+This means that these command lines are equivalent as far as the current
+IOMMU kernel driver behaves:
+
+.. code-block:: bash
+
+  $ qemu-system-riscv64 \
+        -M virt,aia=aplic-imsic,aia-guests=5 \
+        -device riscv-iommu-pci,addr=1.0,vendor-id=0x1efd,device-id=0xedf1 \
+        -device e1000e,netdev=net1 -netdev user,id=net1,net=192.168.0.0/24 \
+        -device e1000e,netdev=net2 -netdev user,id=net2,net=192.168.200.0/24 \
+        (...)
+
+  $ qemu-system-riscv64 \
+        -M virt,aia=aplic-imsic,aia-guests=5 \
+        -device e1000e,netdev=net1 -netdev user,id=net1,net=192.168.0.0/24 \
+        -device e1000e,netdev=net2 -netdev user,id=net2,net=192.168.200.0/24 \
+        -device riscv-iommu-pci,addr=1.0,vendor-id=0x1efd,device-id=0xedf1 \
+        (...)
+
+Both will create iommu groups for the two e1000e cards.
+
+Another thing to notice on `linux-v8`_ and `ventana-linux`_ is that the kernel driver
+considers an IOMMU identified as a Rivos device, i.e. it uses Rivos vendor ID.  To
+use the riscv-iommu-pci device with the existing kernel support we need to emulate
+a Rivos PCI IOMMU by setting 'vendor-id' and 'device-id':
+
+.. code-block:: bash
+
+  $ qemu-system-riscv64 -M virt	\
+     -device riscv-iommu-pci,vendor-id=0x1efd,device-id=0xedf1 (...)
+
+Several options are available to control the capabilities of the device, namely:
+
+- "bus": the bus that the IOMMU device uses
+- "ioatc-limit": size of the Address Translation Cache (default to 2Mb)
+- "intremap": enable/disable MSI support
+- "ats": enable ATS support
+- "off" (Out-of-reset translation mode: 'on' for DMA disabled, 'off' for 'BARE' (passthrough))
+- "s-stage": enable s-stage support
+- "g-stage": enable g-stage support
+
+.. _iommu1.0: https://github.com/riscv-non-isa/riscv-iommu/releases/download/v1.0/riscv-iommu.pdf
+
+.. _linux-v8: https://lore.kernel.org/linux-riscv/cover.1718388908.git.tjeznach@rivosinc.com/
+
+.. _ventana-linux: https://github.com/ventanamicro/linux/tree/dev-upstream
diff --git a/docs/system/riscv/virt.rst b/docs/system/riscv/virt.rst
index 9a06f95a34..8e9a2e4dda 100644
--- a/docs/system/riscv/virt.rst
+++ b/docs/system/riscv/virt.rst
@@ -84,6 +84,19 @@ none``, as in
 
 Firmware images used for pflash must be exactly 32 MiB in size.
 
+riscv-iommu support
+-------------------
+
+The board has support for the riscv-iommu-pci device by using the following
+command line:
+
+.. code-block:: bash
+
+  $ qemu-system-riscv64 -M virt -device riscv-iommu-pci (...)
+
+Refer to :ref:`riscv-iommu` for more information on how the RISC-V IOMMU support
+works.
+
 Machine-specific options
 ------------------------