Hi Andre,
Thanks for looking into this.
On Thu, Apr 4, 2019 at 9:17 PM Andre Przywara <andre.przywara@xxxxxxx> wrote:
>
> On Thu, 4 Apr 2019 12:30:15 +0000
> Zenghui Yu <yuzenghui@xxxxxxxxxx> wrote:
>
> Hi,
>
> > Commit ad275b8bb1e6 ("KVM: arm/arm64: vgic-new: vgic_init: implement
> > vgic_init") had set irq->targets in kvm_vgic_vcpu_init(), regardless of
> > the GIC architecture (v2 or v3). When the number of vcpu reaches 32
> > (in v3), UBSAN will complain about it.
>
> The first part looks similar to this one:
> http://lists.infradead.org/pipermail/linux-arm-kernel/2019-March/637209.html
Yes. I have not noticed this, sorry.
>
> > ================================================================================
> > UBSAN: Undefined behaviour in arch/arm64/kvm/../../../virt/kvm/arm/vgic/vgic-init.c:223:21
> > shift exponent 32 is too large for 32-bit type 'unsigned int'
> > CPU: 13 PID: 833 Comm: CPU 32/KVM Kdump: loaded Not tainted 5.1.0-rc1+ #16
> > Hardware name: Huawei TaiShan 2280 /BC11SPCD, BIOS 1.58 10/24/2018
> > Call trace:
> > dump_backtrace+0x0/0x190
> > show_stack+0x24/0x30
> > dump_stack+0xc8/0x114
> > ubsan_epilogue+0x14/0x50
> > __ubsan_handle_shift_out_of_bounds+0x118/0x188
> > kvm_vgic_vcpu_init+0x1d4/0x200
> > kvm_arch_vcpu_init+0x3c/0x48
> > kvm_vcpu_init+0xa8/0x100
> > kvm_arch_vcpu_create+0x94/0x120
> > kvm_vm_ioctl+0x57c/0xe58
> > do_vfs_ioctl+0xc4/0x7f0
> > ksys_ioctl+0x8c/0xa0
> > __arm64_sys_ioctl+0x28/0x38
> > el0_svc_common+0xa0/0x190
> > el0_svc_handler+0x38/0x78
> > el0_svc+0x8/0xc
> > ================================================================================
> >
> > This patch Restricts setting irq->targets in GICv2, which only supports
> > a maximum of eight PEs, to keep UBSAN quiet. And since irq->mpidr will
> > only be used by SPI in GICv3, we decided to set mpidr to 0 for SGI and
> > PPI.
> >
> > Like commit ab2d5eb03dbb ("KVM: arm/arm64: vgic: Always initialize the
> > group of private IRQs"), we should also take the creating order of the
> > VGIC and VCPUs into consideration.
> >
> > Cc: Eric Auger <eric.auger@xxxxxxxxxx>
> > Cc: Marc Zyngier <marc.zyngier@xxxxxxx>
> > Cc: Andre Przywara <andre.przywara@xxxxxxx>
> > Cc: Christoffer Dall <christoffer.dall@xxxxxxx>
> > Signed-off-by: Zenghui Yu <yuzenghui@xxxxxxxxxx>
> > ---
> > virt/kvm/arm/vgic/vgic-init.c | 16 +++++++++++-----
> > 1 file changed, 11 insertions(+), 5 deletions(-)
> >
> > diff --git a/virt/kvm/arm/vgic/vgic-init.c b/virt/kvm/arm/vgic/vgic-init.c
> > index 3bdb31e..0cba92e 100644
> > --- a/virt/kvm/arm/vgic/vgic-init.c
> > +++ b/virt/kvm/arm/vgic/vgic-init.c
> > @@ -220,7 +220,6 @@ int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
> > irq->intid = i;
> > irq->vcpu = NULL;
> > irq->target_vcpu = vcpu;
> > - irq->targets = 1U << vcpu->vcpu_id;
> > kref_init(&irq->refcount);
> > if (vgic_irq_is_sgi(i)) {
> > /* SGIs */
> > @@ -231,10 +230,14 @@ int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
> > irq->config = VGIC_CONFIG_LEVEL;
> > }
> >
> > - if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
> > + if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
> > irq->group = 1;
> > - else
> > + irq->mpidr = 0;
> > + } else {
> > irq->group = 0;
> > + if (vcpu->vcpu_id < VGIC_V2_MAX_CPUS)
> > + irq->targets = 1U << vcpu->vcpu_id;
> > + }
> > }
> >
> > if (!irqchip_in_kernel(vcpu->kvm))
> > @@ -297,10 +300,13 @@ int vgic_init(struct kvm *kvm)
> >
> > for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
> > struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
> > - if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
> > + if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
> > irq->group = 1;
> > - else
> > + irq->mpidr = 0;
> > + } else {
> > irq->group = 0;
> > + irq->targets = 1U << vcpu->vcpu_id;
> > + }
>
> So why would you need this part? That does the same as above, doesn't it?
This idea comes from commit ab2d5eb03dbb. As Christoffer said, "we have no
enforced ordering of creating the VGIC and creating VCPUs". Without this
part, the VCPUs created before VGIC might still end up with the wrong
"mpidr (target)" set, since they don't know the actual vGIC model.
If we're using QEMU to boot a vGIC-v3 guest, we'll still find the incorrect
TARGET value from debugfs. That's QEMU will create and intialize all of the
VCPUs before VGIC.
A detailed explanation can be found at:
https://marc.info/?l=android-virt&m=154713085226516&w=2
thanks,
zenghui
>
> Cheers,
> Andre.
>
>
> > }
> > }
> >
>
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
