Fixes live-migrations & snapshot-rollback of VMs with a restricted
CPU type (e.g., qemu64) from our 5.15 based kernel (default Proxmox
VE 7.4) to the 6.2 (and future newer) of Proxmox VE 8.0.
Previous to (upstream kernel) commit ad856280ddea ("x86/kvm/fpu: Limit
guest user_xfeatures to supported bits of XCR0") the PKRU bit of the
host could leak into the state from the guest, which caused trouble
when migrating between hosts with different CPUs, i.e., where the
source supported it but the target did not, causing a general
protection fault when the guest tried to use a pkru related
instruction after the migration.
But the fix, while welcome, caused a temporary out-of-sync state when
migrating such a VM from a kernel without the fix to a kernel with
the fix, as it threw of KVM when the CPUID of the guest and most of
the state doesn't report XSAVE and thus any xfeatures, but PKRU and
the related state is set as enabled, causing the vCPU to spin at 100%
without any progress forever.
The fix could be at two sites, either in QEMU or in the kernel, I
choose the kernel as we have all the info there for a targeted
heuristic so that we don't have to adapt QEMU and qemu-server, the
latter even on both sides.
Still, a short summary of the possible fixes and short drawbacks:
* on QEMU-side either
- clear the PKRU state in the migration saved state would be rather
complicated to implement as the vCPU is initialised way before we
have the saved xfeature state available to check what we'd need
to do, plus the user-space only gets a memory blob from ioctl
KVM_GET_XSAVE2 that it passes to KVM_SET_XSAVE ioctl, there are
no ABI guarantees, and while the struct seem stable for 5.15 to
6.5-rc1, that doesn't has to be for future kernels, so off the
table.
- enforce that the CPUID reports PKU support even if it normally
wouldn't. While this works (tested by hard-coding it as POC) it
is a) not really nice and b) needs some interaction from
qemu-server to enable this flag as otherwise we have no good info
to decide when it's OK to do this, which means we need to adapt
both PVE 7 and 8's qemu-server and also pve-qemu, workable but
not optimal
* on Kernel/KVM-side we can hook into the set XSAVE ioctl specific to
the KVM subsystem, which already reduces chance of regression for
all other places. There we have access to the union/struct
definitions of the saved state and thus can savely cast to that.
We also got access to the vCPU's CPUID capabilities, meaning we can
check if the XCR0 (first XSAVE Control Register) reports
that it support the PKRU feature, and if it does *NOT* but the
saved xfeatures register from XSAVE *DOES* report it, we can safely
assume that this combination is due to an migration from an older,
leaky kernel – and clear the bit in the xfeature register before
restoring it to the guest vCPU KVM state, avoiding the confusing
situation that made the vCPU spin at 100%.
This should be safe to do, as the guest vCPU CPUID never reported
support for the PKRU feature, and it's also a relatively niche and
newish feature.
If it gains us something we can drop this patch a bit in the future
Proxmox VE 9 major release, but we should ensure that VMs that where
started before PVE 8 cannot be directly live-migrated to the release
that includes that change; so we should rather only drop it if the
maintenance burden is high.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
