The vmexit_ipi test can be used as a rough benchmark for IPI performance
since commit 8a8c1fc3b1f8 ("vmexit: measure IPI and EOI cost") added
reporting of the average number of cycles taken for IPI delivery. Avoid
exposing a PIT to the guest so that SVM AVIC is not inhibited and IPI
acceleration can be tested when available and enabled by the host.
Signed-off-by: Alejandro Jimenez <alejandro.j.jimenez@xxxxxxxxxx>
---
x86/unittests.cfg | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/x86/unittests.cfg b/x86/unittests.cfg
index 867a8ea2..70cdda72 100644
--- a/x86/unittests.cfg
+++ b/x86/unittests.cfg
@@ -81,10 +81,13 @@ file = vmexit.flat
extra_params = -append 'inl_from_pmtimer'
groups = vmexit
+# To allow IPIs to be accelerated by SVM AVIC when the feature is available and
+# enabled, do not create a Programmable Interval Timer (PIT, a.k.a 8254), since
+# such device will disable/inhibit AVIC if exposed to the guest.
[vmexit_ipi]
file = vmexit.flat
smp = 2
-extra_params = -append 'ipi'
+extra_params = -machine pit=off -append 'ipi'
groups = vmexit
[vmexit_ipi_halt]
base-commit: 00af1c849ced4e515f8659658d18652df2eb08fa
--
2.39.3
