On Tue, Mar 28, 2017 at 02:50:58PM -0400, John Ferlan wrote:
On 03/23/2017 11:26 AM, Ján Tomko wrote:
> @@ -1788,6 +1789,15 @@
> for hypervisor to decide.
> <span class="since">Since 2.1.0</span>
> </dd>
> + <dt><code>irqchip</code></dt>
> + <dd>Tune the in-kernel irqchip. Possible values for the
> + <code>mode</code> attribute are:
> + <code>on</code>, <code>split</code> and
<code>off</code>.
> + <code>split</code> is useful for using interrupt remapping
> + with the <a href="#elementsIommu">IOMMU device</a>.
> + The default is left for hypervisor to decide.
> + <span class="since">Since 3.3.0</span> (QEMU only)
> + </dd>
Did you somehow know this wouldn't be reviewed in time for 3.2.0? ;-)
Once I got to patch 4 & 7, I began to wonder if support there was
dependence upon the value being "split" or is "on" also acceptable.
It
would seem "off" and not present wouldn't allow intremap or caching to
work.
Looking at QEMU code, off also works but it seems the restriction is
temporary.
At the very least whatever it is that allows the -device intel-iommu
to
be present in QEMU 2.7 (if I read virQEMUCapsInitQMPMonitor correctly)
instead of ",iommu=on" would also seem to be a requirement. I know we
don't want to put versions there, but whatever it is that is required
should be listed (at least while it's still fresh in your mind).
The intremap option was introduced in QEMU 2.9, so libvirt cannot possibly
support it with 2.7's -machine iommu=on.
Jan