On Mon, 2021-01-04 at 16:27 +0000, Marc Zyngier wrote:
> On 2021-01-04 16:22, Qian Cai wrote:
> > On Mon, 2021-01-04 at 16:08 +0000, Marc Zyngier wrote:
> > > On 2021-01-04 15:47, Qian Cai wrote:
> > > > On Thu, 2020-12-10 at 08:30 +0000, Marc Zyngier wrote:
> > > > > We reset the guest's view of PMCR_EL0 unconditionally, based on
> > > > > the host's view of this register. It is however legal for an
> > > > > imnplementation not to provide any PMU, resulting in an UNDEF.
> > > > >
> > > > > The obvious fix is to skip the reset of this shadow register
> > > > > when no PMU is available, sidestepping the issue entirely.
> > > > > If no PMU is available, the guest is not able to request
> > > > > a virtual PMU anyway, so not doing nothing is the right thing
> > > > > to do!
> > > > >
> > > > > It is unlikely that this bug can hit any HW implementation
> > > > > though, as they all provide a PMU. It has been found using nested
> > > > > virt with the host KVM not implementing the PMU itself.
> > > > >
> > > > > Fixes: ab9468340d2bc ("arm64: KVM: Add access handler for PMCR
> > > > > register")
> > > > > Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx>
> > > >
> > > > Reverting this commit on the top of today's linux-next fixed a qemu-kvm
> > > > coredump
> > > > issue on TX2 while starting a guest.
> > > >
> > > > - host kernel .config:
> > > > https://cailca.coding.net/public/linux/mm/git/files/master/arm64.config
> > > >
> > > > # /usr/libexec/qemu-kvm -name ubuntu-20.04-server-cloudimg -cpu host
> > > > -smp 2 -m 2g
> > > > -drive
> > > > if=none,format=qcow2,file=./ubuntu-20.04-server-cloudimg.qcow2,id=hd
> > > > -device virtio-scsi -device scsi-hd,drive=hd -cdrom
> > > > ./ubuntu-20.04-server-cloudimg.iso
> > > > -bios /usr/share/AAVMF/AAVMF_CODE.fd -M gic-version=host -nographic
> > > > -nic user,model=virtio,hostfwd=tcp::2222-:22
> > > >
> > > > qemu-kvm: /builddir/build/BUILD/qemu-4.2.0/target/arm/helper.c:1812:
> > > > pmevcntr_rawwrite: Assertion `counter < pmu_num_counters(env)' failed.
> > >
> > > You don't have KVM_ARM_PMU selected in your config, so QEMU cannot
> > > access the PMU registers, and no counters are exposed.
> >
> > Well, isn't it the rule that don't break the userspace? qemu works fine
> > with
> > KVM_ARM_PMU=n until this commit.
>
> No, it doesn't "work fine". It gets random data that potentially makes
> no sense,
> depending on the HW this runs on.
>
> Now, userspace tells you that your kernel is misconfigured. I see it as
> an improvement.
Marc, do you suggest that CONFIG_KVM=y should select KVM_ARM_PMU=y then?
Otherwise, this is rather difficult for users to figure out and a core dump with
an implicit error message from qemu is not that helpful.
