On Thu, May 06, 2021, Marcelo Tosatti wrote:
> Index: kvm/arch/x86/kvm/vmx/posted_intr.c
> ===================================================================
> --- kvm.orig/arch/x86/kvm/vmx/posted_intr.c
> +++ kvm/arch/x86/kvm/vmx/posted_intr.c
> @@ -114,7 +114,7 @@ static void __pi_post_block(struct kvm_v
> } while (cmpxchg64(&pi_desc->control, old.control,
> new.control) != old.control);
>
> - if (!WARN_ON_ONCE(vcpu->pre_pcpu == -1)) {
> + if (vcpu->pre_pcpu != -1) {
> spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> list_del(&vcpu->blocked_vcpu_list);
> spin_unlock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> @@ -135,20 +135,13 @@ static void __pi_post_block(struct kvm_v
> * this case, return 1, otherwise, return 0.
> *
> */
> -int pi_pre_block(struct kvm_vcpu *vcpu)
> +static int __pi_pre_block(struct kvm_vcpu *vcpu)
> {
> unsigned int dest;
> struct pi_desc old, new;
> struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
>
> - if (!kvm_arch_has_assigned_device(vcpu->kvm) ||
> - !irq_remapping_cap(IRQ_POSTING_CAP) ||
> - !kvm_vcpu_apicv_active(vcpu))
> - return 0;
> -
> - WARN_ON(irqs_disabled());
> - local_irq_disable();
> - if (!WARN_ON_ONCE(vcpu->pre_pcpu != -1)) {
> + if (vcpu->pre_pcpu == -1) {
> vcpu->pre_pcpu = vcpu->cpu;
> spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> list_add_tail(&vcpu->blocked_vcpu_list,
> @@ -188,12 +181,33 @@ int pi_pre_block(struct kvm_vcpu *vcpu)
> if (pi_test_on(pi_desc) == 1)
> __pi_post_block(vcpu);
>
> + return (vcpu->pre_pcpu == -1);
Nothing checks the return of __pi_pre_block(), this can be dropped and the
helper can be a void return.
> +}
> +
> +int pi_pre_block(struct kvm_vcpu *vcpu)
> +{
> + struct vcpu_vmx *vmx = to_vmx(vcpu);
> +
> + vmx->in_blocked_section = true;
> +
> + if (!kvm_arch_has_assigned_device(vcpu->kvm) ||
> + !irq_remapping_cap(IRQ_POSTING_CAP) ||
> + !kvm_vcpu_apicv_active(vcpu))
Opportunistically fix the indentation?
> + return 0;
> +
> + WARN_ON(irqs_disabled());
> + local_irq_disable();
> + __pi_pre_block(vcpu);
> local_irq_enable();
> +
> return (vcpu->pre_pcpu == -1);
> }
>
> void pi_post_block(struct kvm_vcpu *vcpu)
> {
> + struct vcpu_vmx *vmx = to_vmx(vcpu);
> +
> + vmx->in_blocked_section = false;
> if (vcpu->pre_pcpu == -1)
> return;
>
> @@ -236,6 +250,52 @@ bool pi_has_pending_interrupt(struct kvm
> (pi_test_sn(pi_desc) && !pi_is_pir_empty(pi_desc));
> }
>
> +static void pi_update_wakeup_vector(void *data)
> +{
> + struct vcpu_vmx *vmx;
> + struct kvm_vcpu *vcpu = data;
> +
> + vmx = to_vmx(vcpu);
> +
> + /* race with pi_post_block ? */
> + if (vcpu->pre_pcpu != -1)
This seems wrong. The funky code in __pi_pre_block() regarding pre_cpu muddies
the waters, but I don't think it's safe to call __pi_pre_block() from a pCPU
other than the pCPU that is associated with the vCPU.
If the vCPU is migrated after vmx_pi_start_assignment() grabs vcpu->cpu but
before the IPI arrives (to run pi_update_wakeup_vector()), then it's possible
that a different pCPU could be running __pi_pre_block() concurrently with this
code. If that happens, both pcPUs could see "vcpu->pre_cpu == -1" and corrupt
the list due to a double list_add_tail.
The existing code is unnecessarily confusing, but unless I'm missing something,
it's guaranteed to call pi_pre_block() from the pCPU that is associated with the
pCPU, i.e. arguably it could/should use this_cpu_ptr(). Because the existing
code grabs vcpu->cpu with IRQs disabled and is called only from KVM_RUN,
vcpu->cpu is guaranteed to match the current pCPU since vcpu->cpu will be set to
the current pCPU when the vCPU is scheduled in.
Assuming my analysis is correct (definitely not guaranteed), I'm struggling to
come up with an elegant solution. But, do we need an elegant solution? E.g.
can the start_assignment() hook simply kick all vCPUs with APICv active?
> + return;
> +
> + if (!vmx->in_blocked_section)
> + return;
> +
> + __pi_pre_block(vcpu);
> +}
> +
> +void vmx_pi_start_assignment(struct kvm *kvm, int device_count)
> +{
> + struct kvm_vcpu *vcpu;
> + int i;
> +
> + if (!irq_remapping_cap(IRQ_POSTING_CAP))
> + return;
> +
> + /* only care about first device assignment */
> + if (device_count != 1)
> + return;
> +
> + /* Update wakeup vector and add vcpu to blocked_vcpu_list */
> + kvm_for_each_vcpu(i, vcpu, kvm) {
> + struct vcpu_vmx *vmx = to_vmx(vcpu);
> + int pcpu;
> +
> + if (!kvm_vcpu_apicv_active(vcpu))
> + continue;
> +
> + preempt_disable();
Any reason not to do "cpu = get_cpu()"? Might make sense to do that outside of
the for-loop, too.
> + pcpu = vcpu->cpu;
> + if (vmx->in_blocked_section && vcpu->pre_pcpu == -1 &&
> + pcpu != -1 && pcpu != smp_processor_id())
> + smp_call_function_single(pcpu, pi_update_wakeup_vector,
> + vcpu, 1);
> + preempt_enable();
> + }
> +}
