On Wed, Aug 18, 2021, Maciej S. Szmigiero wrote:
> On 18.08.2021 13:49, Paul Menzel wrote:
> > In scripts, running
> >
> > modprobe kvm_amd 2>/dev/null
> > modprobe kvm_intel 2>/dev/null
> >
> > to ensure the modules are loaded causes Linux to log errors.
> >
> > $ dmesg --level=err
> > [ 0.641747] [Firmware Bug]: TSC_DEADLINE disabled due to Errata; please update microcode to version: 0x3a (or later)
> > [ 40.196868] kvm: already loaded the other module
> > [ 40.219857] kvm: already loaded the other module
> > [ 55.501362] kvm [1177]: vcpu0, guest rIP: 0xffffffff96e5b644 disabled perfctr wrmsr: 0xc2 data 0xffff
> > [ 56.397974] kvm [1418]: vcpu0, guest rIP: 0xffffffff81046158 disabled perfctr wrmsr: 0xc1 data 0xabcd
> > [1007981.827781] kvm: already loaded the other module
> > [1008000.394089] kvm: already loaded the other module
> > [1008030.706999] kvm: already loaded the other module
> > [1020396.054470] kvm: already loaded the other module
> > [1020405.614774] kvm: already loaded the other module
> > [1020410.140069] kvm: already loaded the other module
> > [1020704.049231] kvm: already loaded the other module
> >
> > As one of the two KVM modules is already loaded, KVM is functioning, and
> > their is no error condition. Therefore, demote the log message level to
> > informational.
Hrm, but there is an error condition. Userspace explicitly requested something
and KVM couldn't satisfy the request.
KVM is also going to complain at level=err one way or another, e.g. if a script
probes kvm_amd before kvm_intel on an Intel CPU it's going to get "kvm: no hardware
support", so this isn't truly fixing the problem. Is the issue perhaps that this
particular message isn't ratelimited?
It's also easy for the script to grep /proc/cpuinfo, so it's hard to feel too
bad about the kludgy message, e.g. look for a specific vendor, 'vmx' or 'svm', etc...
if [[ -z $kvm ]]; then
grep vendor_id "/proc/cpuinfo" | grep -q AuthenticAMD
if [[ $? -eq 0 ]]; then
kvm=kvm_amd
else
kvm=kvm_intel
fi
fi
> Shouldn't this return ENODEV when loading one of these modules instead
> as there is no hardware that supports both VMX and SVM?
Probably not, as KVM would effectively be speculating, e.g. someone could load an
out-of-tree variant of kvm_{intel,amd}. Maybe instead of switching to ENODEV,
reword the comment, make it ratelimited, and shove it down? That way the message
and -EEXIST fires iff the vendor module actually has some chance of being loaded.
>From 3528e66bd5107d5ac4f6a6ae50503cf64446866a Mon Sep 17 00:00:00 2001
From: Sean Christopherson <seanjc@xxxxxxxxxx>
Date: Wed, 18 Aug 2021 15:17:43 -0700
Subject: [PATCH] KVM: x86: Tweak handling and message when vendor module is
already loaded
Reword KVM's error message if a vendor module is already loaded to state
exactly that instead of assuming "the other" module is loaded, ratelimit
said message to match the other errors, and move the check down below the
basic functionality checks so that attempting to load an unsupported
module provides the same result regardless of whether or not a supported
vendor module is already loaded.
Reported-by: Paul Menzel <pmenzel@xxxxxxxxxxxxx>
Cc: Maciej S. Szmigiero <mail@xxxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
---
arch/x86/kvm/x86.c | 12 ++++++------
1 file changed, 6 insertions(+), 6 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index fdc0c18339fb..15bd4bd3c81d 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8357,12 +8357,6 @@ int kvm_arch_init(void *opaque)
struct kvm_x86_init_ops *ops = opaque;
int r;
- if (kvm_x86_ops.hardware_enable) {
- printk(KERN_ERR "kvm: already loaded the other module\n");
- r = -EEXIST;
- goto out;
- }
-
if (!ops->cpu_has_kvm_support()) {
pr_err_ratelimited("kvm: no hardware support\n");
r = -EOPNOTSUPP;
@@ -8374,6 +8368,12 @@ int kvm_arch_init(void *opaque)
goto out;
}
+ if (kvm_x86_ops.hardware_enable) {
+ pr_err_ratelimited("kvm: already loaded a vendor module\n");
+ r = -EEXIST;
+ goto out;
+ }
+
/*
* KVM explicitly assumes that the guest has an FPU and
* FXSAVE/FXRSTOR. For example, the KVM_GET_FPU explicitly casts the
--
2.33.0.rc2.250.ged5fa647cd-goog
