On 11.07.2012 14:03, Jiri Denemark wrote:
On Wed, Jul 11, 2012 at 09:00:05 +0200, Hendrik Schwartke wrote:
On 10.07.2012 21:57, Jiri Denemark wrote:
I know I'm late in this vendor_id stuff but it hasn't been released yet so I'm
not too late. I assume the reason for introducing it is to be able to lie to a
guest. Please, correct me, if this is not the case.
Well, the reason is to specify a vendor_id which is different to the
actual host vendor id
diff --git a/docs/formatdomain.html.in b/docs/formatdomain.html.in
index 94c555f..b6e0d5d 100644
--- a/docs/formatdomain.html.in
+++ b/docs/formatdomain.html.in
@@ -779,7 +779,11 @@
in which case an attempt to start a domain requesting an unsupported
CPU model will fail. Supported values for<code>fallback</code>
attribute are:<code>allow</code> (this is the default), and
-<code>forbid</code>.</dd>
+<code>forbid</code>. The optional<code>vendor_id</code> attribute
+ (<span class="since">Since 0.9.14</span>) can be used to set the
+ vendor id seen by the guest. It must be exactly 12 characters long.
+ If not set the vendor id of the host is used. Typical possible
+ values are "AuthenticAMD" and "GenuineIntel".</dd>
<dt><code>vendor</code></dt>
<dd><span class="since">Since 0.8.3</span> the content of the
This is wrong (unless your previous patch explicitly modified the code to
behave like this). If vendor_id is not set, a guest should see model's default
vendor. If a guest is configured with, e.g., SandyBridge model, its vendor
will be "GenuineIntel". If a guest uses Opteron_G3, it will see "AuthenticAMD"
vendor. All this regardless on the vendor of the host CPU as longs as the host
CPU is capable enough to support all features of a given guest CPU.
I have tested this (with the qemu driver) and that's not correct. I'm
sure you have to set the vendor attribute in the qemu command line to
set the vendor id. So "kvm -cpu core2duo -cdrom
debian-6.0.5-i386-netinst.iso" on an amd machine results in vendor id =
AuthenticAMD (just as with -cpu 486)
Oh, you are right. I tried with Opteron_G1 on Sandy Bridge host and indeed the
guest sees it as Opteron made by Intel. I think this is a QEMU bug as its CPU
configuration says:
[cpudef]
name = "Opteron_G1"
level = "5"
vendor = "AuthenticAMD"
...
No, it's not a bug it's a feature ;). There are a few instructions that
are implemented differently on intel and amd cpus (e.g.SYSENTER and
SYSCALL). So in the past it was important for the guest to know the
correct cpu vendor and therefor it wasn't possible to fake it. But since
a while these few instructions can be emulated by qemu, so that it's
know unproblematic to change it. (See
http://developer.amd.com/assets/CrossVendorMigration.pdf)
I'm not sure but I'm supposing that the old behavior of qemu to set all
cpu properties with the exception of the vendor id is still in use to
avoid compatibility issues. If you want to specify a guest cpu
*including* the vendor id you have to set the additional vendor attribute.
Anyway, to be honest, I'm not a big fan of the new vendor_id attribute.
Currently we have
<cpu ...>
<model vendor_id="bleblablebla">Model</model>
...
</cpu>
to force "bleblablebla" vendor ID on the guest CPU and
<cpu ...>
<model>Model</model>
<vendor>Intel</vendor>
...
</cpu>
to make sure the guest will be run only on a host with Intel CPU.
I think it would be much better to reuse the already existing<vendor>
element. We could perhaps add new force attribute for vendor element. Thus,
These are two different things. One is the vendor_id the guest should
see and the other the vendor id the host should have to start the
domain. So if the CPUID instruction in the guest should return
"NOT-INTEL" and the host cpu should be an intel cpu a force bit is not
enough. Furthermore it would be confusing to use only one attribute to
specify the vendor of the host and the guest cpu.
Yes, I know there are two different things here. However, having<vendor>
element specifying host's vendor and vendor_id attribute for model element
specifying guest's vendor feels a bit odd. Let's see what other thinks about
this matter.
Jirka
--
libvir-list mailing list
libvir-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/libvir-list