On Tue, Feb 18, 2020 at 05:28:49PM +0100, Erwan Velu wrote:
> On 14/02/2020 18:05, Sean Christopherson wrote:
> >This has come up before[*]. Using _once() doesn't fully solve the issue
> >when KVM is built as a module. The spam is more than likely a userspace
> >bug, i.e. userspace is probing KVM on every CPU.
>
> I made some progress on this.
>
>
> That's "/usr/bin/udevadm trigger --type=devices --action=add" the culprit.
>
> It does echo "add" in /sys/devices/system/cpu/cpu<x>/uevent
>
> For the each cpu, it does the 'add' which trigger the "disabled by bios"
> message from kvm_arch_init.
>
> Note that doing a "add" on the same processor will trigger the same message
> at every "add" event.
>
>
> So I tried the patch of using pr_err_once() instead of printk() and the
> behavior is fine : despite the number of "add" generated, there is a single
> line being printed out.
>
> Without the patch, every "add" generates the "disabled by bios" message.
That's a sort of unintentional side effect of KVM being split into two
modules, kvm and kvm_{intel,amd}. E.g. if userspace did 'rmmod kvm' on
failure of 'modprobe kvm_intel' then using _*once() would be ineffective.
> So the question is : do we want to handle the case where a possible bios
> missed the configuration of some cores ?
That's a question for AMD/SVM. Starting with kernel 5.6, Intel/VMX checks
for BIOS enabling on all CPUs.
That being said, checking for correct BIOS configuration on all CPUs is
orthogonal to this print statement issue. Probing kvm_intel on every CPU
doesn't do anything to address a misoncifgured BIOS, e.g. if VMX/SVM is
fully supported on CPU0 then additional probes of kvm_{intel,amd} are nops,
they don't actually check for support on other CPUs.
> If no, then the patch is fine and could be submitted. I don't see the need
> of printing this message at every call as it pollute the kernel log.
>
> If yes, then we need to keep a trace of the number of enabled/disabled cores
> so we can report a mismatch. As this message seems printed per cpu, that
> would kind of mean a global variable right ?
>
> What are your recommendations on this ?
Fix userspace to only do the "add" on one CPU.
Changing kvm_arch_init() to use pr_err_once() for the disabled_by_bios()
case "works", but it's effectively a hack to workaround a flawed userspace.
