Linux 4.8 introduces changes that really really require you to set
the APIC base (via either KVM_SET_SREGS or KVM_SET_MSRS) before the
KVM_SET_LAPIC ioctl, unless KVM_SET_LAPIC is immediately preceded by
KVM_GET_LAPIC. This is because in x2apic mode KVM now moves the APIC id
from bits 31-24 of the KVM_SET_LAPIC argument to bits 7-0 of the APIC_ID
register (0x20). If QEMU does not nudge the APIC out of x2apic at reset
time, the subsequent KVM_SET_LAPIC incorrectly shifts the APIC id right
and Linux sees (in bits 31-24) a zero id.
While QEMU has always done the right thing, the recent commit 78d6a05
("x86/lapic: Load LAPIC state at post_load", 2016-09-13) changed
that, causing a bug on newer kernels. Fix this by setting the APIC
base in kvm_apic_put. Another workaround could be to enable
KVM_X2APIC_API_USE_32BIT_IDS, which lets KVM not mangle the APIC ID.
Despite the "don't break userspace" mantra, I believe that this can be
considered a QEMU bug.
I checked kvmtool and it only calls KVM_SET_LAPIC after a KVM_GET_LAPIC,
which avoids the issue in the first place.
Dr. David Alan Gilbert (1):
kvm: apic: set APIC base as part of kvm_apic_put
Paolo Bonzini (1):
target-i386: introduce kvm_put_one_msr
v1->v2: send the right patches
hw/i386/kvm/apic.c | 2 ++
target-i386/kvm.c | 31 ++++++++++++++++++++++---------
target-i386/kvm_i386.h | 2 ++
3 files changed, 26 insertions(+), 9 deletions(-)
--
2.7.4
