docs: add info about factors affecting CPU compatibility

While libvirt solves most of the problem of ensuring compatibility, when
there is incompatibility it can be hard for users to track down the
cause. Everything knows to check the physical CPU model, but there are a
surprisingly large number of other factors influencing compatibility.

Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>

diff --git a/docs/drvqemu.rst b/docs/drvqemu.rst
index c174bdb4cf0a6c09fd029c5b55d8169b0b69567c..c9f24322687fa2fe88d3c2a3155c814dc1ed9271 100644 (file)
--- a/docs/drvqemu.rst
+++ b/docs/drvqemu.rst
@@ -416,6 +416,89 @@ following command should be run as root, prior to starting libvirtd
 libvirt will then place each virtual machine in a cgroup at
 ``/dev/cgroup/libvirt/qemu/$VMNAME/``
 
+
+Live migration compatibility
+----------------------------
+
+Many factors can affect the ability to live migrate a guest between a pair
+of hosts. It is critical that when QEMU is started on the destination host,
+the exposed guest machine ABI matches what was exposed by the existing QEMU
+process on the source host. To facilitate this, when libvirt receives a
+guest configuration document, it will attempt to expand any features that
+were not specified, to ensure a stable guest machine ABI. Mostly this involves
+adding address information to all devices, and adding controllers to attach
+the devices to.
+
+Certain features that affect the guest ABI, however, may only be known at the
+time the guest is started and can be influenced by features of the host OS
+and its hardware. This means that even if the guest XML configuration is the
+same, it may still be impossible to migrate the guest between two hosts.
+
+Migration CPU model compatibility
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The most common problems with migration compatibility surround the use of
+the guest CPU ``host-model`` or ``host-passthrough`` modes. Both of these
+modes attempt to expose the full host CPU featureset to the guest. The
+``host-model`` mode attempts to expose as many features as possible while
+retaining the ability to accurately check compatibility between hosts prior
+to migration running. The ``host-passthrough`` mode attempts to expose the
+host CPU as precisely as possible, but with the cost that it is not possible
+for libvirt to check compatibility prior to migration.
+
+If using ``host-model`` the target host hardware and software deployment
+must expose a superset of the features of the source host CPU. If using
+``host-passthrough`` the target host CPU and software deployment must
+always expose a superset of the fetures, however, it is further strongly
+recommended that the source and destination hosts be identical in every
+way.
+
+In both cases, there are a number of factors that will influence the CPU
+features available to the guest
+
+* **Physical CPU model** - the core constraint on what features are available.
+  Check ``/proc/cpuinfo`` for CPU model name.
+* **Firmware revision (BIOS/UEFI/etc)** - firmware updates may bundle microcode
+  updates which arbitrarily add or remove CPU features, typically in response
+  to new hardware vulnerabilities. Check ``dmidecode`` for details on ``x86``
+  and ``aarch64`` platforms for firmware version, and ``/proc/cpuinfo`` for
+  associated microcode version (if not updated by the OS).
+* **Firmware settings** - certain firmware settings can affect accessibility of
+  features. For example, turning on/off SMT/HT not only affects the number
+  of logical CPUs available to the OS, but can indirectly influence other
+  factors such as the number of performance counters available for use. Check
+  the firmware specific configuration interface.
+* **Host kernel version** - the host kernel software version may have a
+  need to block certain physical CPU features from use in the guest. It can
+  also emulate certain features that may not exist in the silicon, for example,
+  x2apic. Check ``uname -r`` output for kernel version.
+* **Host kernel settings** - the kernel command line options can be used to
+  block certain physical CPU features from use in the guest, for example,
+  ``tsx=off``, ``l1tf=...`` or ``nosmt``. Check ``/proc/cmdline`` and
+  ``/etc/modprobe.d/*.conf``.
+* **microcode update version** - while the firmware will load the initial
+  microcode in to the CPU, the OS may ship packages providing newer microcode
+  updates since these can be deployed on a more timely manner than firmware
+  updates. These updates can arbitrarily load add or remove CPU features.
+  Check ``/proc/cpuinfo`` for microcode version.
+* **QEMU version** - even when the kernel supports exposing a CPU feature to
+  the guest, an update in the QEMU emulator version will be required to unlock
+  its usage with a guest, except with ``host-passthrough``. Check the output
+  of ``$QEMU -version``.
+* **libvirt version** - even when the kernel and QEMU support exposing a CPU
+  feature to the guest, an update in the libvirt version will be required to
+  unlock its usage with a guest, except with ``host-passthrough``. Check
+  ``virsh version``.
+* **Nested virtualization** - due to the limitations of nested virtualization,
+  a L1 nested host may not be able to expose the same featureset as a bare
+  metal host, even if everything else is the same.
+
+The ``virsh capabilities`` output will provide information on the high level
+CPU model, its features, microcode version. Most of the time this will provide
+enough information to know whether the CPUs of two hosts will be compatible.
+If there are unexpected differences though, checking the above list of
+influencing factors can reveal where the difference arises from.
+
 Import and export of libvirt domain XML configs
 -----------------------------------------------