On 4/14/20 1:53 PM, Andrea Bolognani wrote:
This document describes the relationship between PCI addresses as
seen in the domain XML and by the guest OS, which is a topic that
people get confused by time and time again.
Signed-off-by: Andrea Bolognani <abologna@xxxxxxxxxx>
---
docs/formatdomain.html.in | 6 +-
docs/pci-addresses.rst | 184 ++++++++++++++++++++++++++++++++++++++
2 files changed, 189 insertions(+), 1 deletion(-)
create mode 100644 docs/pci-addresses.rst
diff --git a/docs/formatdomain.html.in b/docs/formatdomain.html.in
index 6f43976815..0077666862 100644
--- a/docs/formatdomain.html.in
+++ b/docs/formatdomain.html.in
@@ -4286,7 +4286,11 @@
element with no other attributes as an explicit request to
assign a PCI address for the device rather than some other
type of address that may also be appropriate for that same
- device (e.g. virtio-mmio).
+ device (e.g. virtio-mmio).<br/>
+ The relationship between the PCI addresses configured in the domain
+ XML and those seen by the guest OS can sometime seem confusing: a
+ separate document describes <a href="pci-addresses.html">how PCI
+ addresses work</a> in more detail.
</dd>
<dt><code>drive</code></dt>
<dd>Drive addresses have the following additional
diff --git a/docs/pci-addresses.rst b/docs/pci-addresses.rst
new file mode 100644
index 0000000000..96c6466899
--- /dev/null
+++ b/docs/pci-addresses.rst
@@ -0,0 +1,184 @@
+========================================
+PCI addresses in domain XML and guest OS
+========================================
+
+.. contents::
+
+When discussing PCI addresses, it's important to understand the the
+relationship between the addresses that can be seen in the domain XML
+and those that are visible inside the guest OS.
+
+
+Simple cases
+============
+
+When the PCI topology of the VM is very simple, the PCI addresses
+will usually match.
+
+For example, the domain XML snippet
+
+::
+
+ <controller type='pci' index='0' model='pcie-root'/>
+ <controller type='pci' index='1' model='pcie-root-port'>
+ <model name='pcie-root-port'/>
+ <target chassis='1' port='0x8'/>
+ <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x0'/>
+ </controller>
+ <interface type='network'>
+ <source network='default'/>
+ <model type='virtio'/>
+ <address type='pci' domain='0x0000' bus='0x01' slot='0x00' function='0x0'/>
+ </interface>
+
+will result in the PCI topology
+
+::
+
+ 0000:00:00.0 Host bridge: Intel Corporation 82G33/G31/P35/P31 Express DRAM Controller
+ 0000:00:01.0 PCI bridge: Red Hat, Inc. QEMU PCIe Root port
+ 0000:01:00.0 Ethernet controller: Red Hat, Inc. Virtio network device (rev 01)
+
+showing up in the guest OS.
+
+The PCI address of the ``virtio-net`` adapter, ``0000:01:00.0``, is
+the same in both cases, so there's no confusion.
+
+
+More complex cases
+==================
+
+In more complex cases, the PCI address visible in the domain XML will
+correlate to the one seen by the guest OS in a less obvious way.
+
+pcie-expander-bus
+-----------------
+
+This fairly uncommon device, which can be used with ``x86_64/q35``
+guests, will help illustrate one such scenario.
+
+For example, the domain XML snippet
+
+::
+
+ <controller type='pci' index='0' model='pcie-root'/>
+ <controller type='pci' index='1' model='pcie-expander-bus'>
+ <model name='pxb-pcie'/>
+ <target busNr='254'/>
+ <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x0'/>
+ </controller>
+ <controller type='pci' index='2' model='pcie-root-port'>
+ <model name='pcie-root-port'/>
+ <target chassis='2' port='0x0'/>
+ <address type='pci' domain='0x0000' bus='0x01' slot='0x00' function='0x0'/>
+ </controller>
+ <interface type='network'>
+ <source network='default'/>
+ <model type='virtio'/>
+ <address type='pci' domain='0x0000' bus='0x02' slot='0x00' function='0x0'/>
+ </interface>
+
+will result in the PCI topology
+
+::
+
+ 0000:00:00.0 Host bridge: Intel Corporation 82G33/G31/P35/P31 Express DRAM Controller
+ 0000:00:01.0 Host bridge: Red Hat, Inc. QEMU PCIe Expander bridge
+ 0000:fe:00.0 PCI bridge: Red Hat, Inc. QEMU PCIe Root port
+ 0000:ff:00.0 Ethernet controller: Red Hat, Inc. Virtio network device (rev 01)
+
+showing up in the guest OS.
+
+This time the addresses don't match: this is because the ``busNr``
+property for the ``pcie-expander-bus`` controller causes it to show
+up as bus 254 (``0xfe`` in hexadecimal) instead of bus 1 as one might
+expect based on its ``index`` property.
+
+How can the domain XML shown above work at all, then? Surely the
+``pcie-root-port`` controller and the ``virtio-net`` adapter should
+use ``bus=0xfe`` and ``bus=0xff`` respectively for the configuration
+to be accepted by libvirt?
+
+As it turns out, that's not the case. The reason for this is that
+QEMU, and consequently libvirt, uses the ``bus`` property of a
+device's PCI address only to match it with the PCI controller that
+has the same ``index`` property, and not to set the actual PCI
+address, which is decided by the guest OS.
+
+So, by looking at the XML snippet above, we can see that the
+``virtio-net`` adapter plugs into the ``pcie-root-port`` controller,
+which plugs into the ``pcie-expander-bus`` controller, which plugs
+into ``pcie-root``: the guest OS sees the same topology, but assigns
+different PCI addresses to some of its component.
+
+The takeaway is that the *relationship* between controllers are the
+very same whether you look at the domain XML or at the guest OS, but
+the *actual PCI addresses* are not guaranteed to match and in fact,
+except for the very simplest cases, they usually will not.
and it doesn't necessarily take a pcie-expander-bus to make the
numbering appear "off". It really is 100% up to the guest OS what bus
number is given to each bus that it discovers, and there's nothing the
host can do about it. This is similar to the target device name for
block devices - you can put <target dev='sde'/> in the config for a disk
as much as you want and it's not going to make any difference - the
guest OS will name it whatever it feels like naming it, and you'd better
damn well like it! :-)
Sorry, that was a digression. What you've said is fine.
+
+spapr-pci-host-bridge
+---------------------
+
+This device, which is unique to ``ppc64/pseries`` guests, will help
+illustrate another scenario.
+
+For example, the domain XML snippet
+
+::
+
+ <controller type='pci' index='0' model='pci-root'>
+ <model name='spapr-pci-host-bridge'/>
+ <target index='0'/>
+ </controller>
+ <controller type='pci' index='1' model='pci-root'>
+ <model name='spapr-pci-host-bridge'/>
+ <target index='1'/>
+ </controller>
+ <interface type='network'>
+ <source network='default'/>
+ <model type='virtio'/>
+ <address type='pci' domain='0x0000' bus='0x01' slot='0x01' function='0x0'/>
+ </interface>
+
+will result in the PCI topology
+
+::
+
+ 0001:00:01.0 Ethernet controller: Red Hat, Inc. Virtio network device
+
+showing up in the guest OS. Note that the two
+``spapr-pci-host-bridge`` controllers are not listed.
+
+This time, in addition to the bus not matching just like in the
+previous example, the interesting part is that the domain doesn't
+match either: this is because each ``spapr-pci-host-bridge``
+controller creates a separate PCI domain.
+
+Once again, while the PCI addresses seen in the domain XML and those
+seen by the guest OS do not match, the relationships between the
+various devices are preserved.
+
+
+Device assignment
+=================
+
+When using VFIO to assign host devices to a guest, an additional
+caveat to keep in mind that the guest OS will base its decisions upon
+the *target address* rather than the *source address*.
+
+For example, the domain XML snippet
+
+::
+
+ <hostdev mode='subsystem' type='pci' managed='yes'>
+ <driver name='vfio'/>
+ <source>
+ <address domain='0x0001' bus='0x08' slot='0x00' function='0x0'/>
+ </source>
+ <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x0'/>
+ </hostdev>
+
+will result in the device showing up as ``0000:00:01.0`` in the
+guest OS rather than as ``0001:08:00.1``.
"... which is the address of the device *on the host*."
or something like that.
Reviewed-by: Laine Stump <laine@xxxxxxxxxx>
+
+Of course, all the rules and behaviors described above still apply.
