Gleb Natapov wrote on 2012-11-26:
> On Mon, Nov 26, 2012 at 03:51:04AM +0000, Zhang, Yang Z wrote:
>> Gleb Natapov wrote on 2012-11-25:
>>> On Wed, Nov 21, 2012 at 04:09:39PM +0800, Yang Zhang wrote:
>>>> Posted Interrupt allows vAPICV interrupts to inject into guest directly
>>>> without any vmexit.
>>>>
>>>> - When delivering a interrupt to guest, if target vcpu is running,
>>>> update Posted-interrupt requests bitmap and send a notification event
>>>> to the vcpu. Then the vcpu will handle this interrupt automatically,
>>>> without any software involvemnt.
>>> Looks like you allocating one irq vector per vcpu per pcpu and then
>>> migrate it or reallocate when vcpu move from one pcpu to another.
>>> This is not scalable and migrating irq migration slows things down.
>>> What's wrong with allocating one global vector for posted interrupt
>>> during vmx initialization and use it for all vcpus?
>>
>> Consider the following situation:
>> If vcpu A is running when notification event which belong to vcpu B is arrived,
> since the vector match the vcpu A's notification vector, then this event
> will be consumed by vcpu A(even it do nothing) and the interrupt cannot
> be handled in time. The exact same situation is possible with your code.
> vcpu B can be migrated from pcpu and vcpu A will take its place and will
> be assigned the same vector as vcpu B. But I fail to see why is this a
No, the on bit will be set to prevent notification event when vcpu B start migration. And it only free the vector before it going to run in another pcpu.
> problem. vcpu A will ignore PI since pir will be empty and vcpu B should
> detect new event during next vmentry.
Yes, but the next vmentry may happen long time later and interrupt cannot be serviced until next vmentry. In current way, it will cause vmexit and re-schedule the vcpu B.
>>
>>>
>>>> + if (!cfg) {
>>>> + free_irq_at(irq, NULL);
>>>> + return 0;
>>>> + }
>>>> +
>>>> + raw_spin_lock_irqsave(&vector_lock, flags);
>>>> + if (!__assign_irq_vector(irq, cfg, mask))
>>>> + ret = irq;
>>>> + raw_spin_unlock_irqrestore(&vector_lock, flags);
>>>> +
>>>> + if (ret) {
>>>> + irq_set_chip_data(irq, cfg);
>>>> + irq_clear_status_flags(irq, IRQ_NOREQUEST);
>>>> + } else {
>>>> + free_irq_at(irq, cfg);
>>>> + }
>>>> + return ret;
>>>> +}
>>>
>>> This function is mostly cut&paste of create_irq_nr().
>>
>> Yes, this function allow to allocate vector from specified cpu.
>>
> Does not justify code duplication.
ok. will change it in next version.
>>>>
>>>> if (kvm_x86_ops->has_virtual_interrupt_delivery(vcpu))
>>>> apic->vid_enabled = true;
>>>> +
>>>> + if (kvm_x86_ops->has_posted_interrupt(vcpu))
>>>> + apic->pi_enabled = true;
>>>> +
>>> This is global state, no need per apic variable.
Even all vcpus use the same setting, but according to SDM, apicv really is a per apic variable.
Anyway, if you think we should not put it here, where is the best place?
>>>> @@ -1555,6 +1611,11 @@ static void vmx_vcpu_load(struct kvm_vcpu
> *vcpu,
>>> int cpu)
>>>> struct desc_ptr *gdt = &__get_cpu_var(host_gdt);
>>>> unsigned long sysenter_esp;
>>>> + if (enable_apicv_pi && to_vmx(vcpu)->pi)
>>>> + pi_set_on(to_vmx(vcpu)->pi);
>>>> +
>>> Why?
>>
>> Here means the vcpu start migration. So we should prevent the
>> notification event until migration end.
>>
> You check for IN_GUEST_MODE while sending notification. Why is this not
For interrupt from emulated device, it enough. But VT-d device doesn't know the vcpu is migrating, so set the on bit to prevent the notification event when target vcpu is migrating.
> enough? Also why vmx_vcpu_load() call means that vcpu start migration?
I think the follow check can ensure the vcpu is in migration, am I wrong?
if (vmx->loaded_vmcs->cpu != cpu)
{
if (enable_apicv_pi && to_vmx(vcpu)->pi)
pi_set_on(to_vmx(vcpu)->pi);
}
>>>> + kvm_make_request(KVM_REQ_POSTED_INTR, vcpu);
>>>> +
>>>> kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu); local_irq_disable();
>>>> list_add(&vmx->loaded_vmcs->loaded_vmcss_on_cpu_link, @@ -1582,6
>>>> +1643,8 @@ static void vmx_vcpu_put(struct kvm_vcpu *vcpu)
>>>> vcpu->cpu = -1; kvm_cpu_vmxoff(); }
>>>> + if (enable_apicv_pi && to_vmx(vcpu)->pi)
>>>> + pi_set_on(to_vmx(vcpu)->pi);
>>> Why?
>>
>> When vcpu schedule out, no need to send notification event to it, just set the
> PIR and wakeup it is enough.
> Same as above. When vcpu is scheduled out it will no be in IN_GUEST_MODE
Right.
> mode. Also in this case we probably should set bit directly in IRR and leave
> PIR alone.
>From the view of hypervisor, IRR and PIR are same. For each vmentry, if PI is enabled, the IRR equal to (IRR | PIR). So there is no difference to set IRR or PIR if target vcpu is not running.
>
>>
>>>> }
>>>>
>>>> static void vmx_fpu_activate(struct kvm_vcpu *vcpu)
>>>> @@ -2451,12 +2514,6 @@ static __init int setup_vmcs_config(struct
>>> vmcs_config *vmcs_conf)
>>>> u32 _vmexit_control = 0;
>>>> u32 _vmentry_control = 0;
>>>> - min = PIN_BASED_EXT_INTR_MASK | PIN_BASED_NMI_EXITING;
>>>> - opt = PIN_BASED_VIRTUAL_NMIS;
>>>> - if (adjust_vmx_controls(min, opt, MSR_IA32_VMX_PINBASED_CTLS,
>>>> - &_pin_based_exec_control) < 0)
>>>> - return -EIO;
>>>> -
>>>> min = CPU_BASED_HLT_EXITING |
>>>> #ifdef CONFIG_X86_64
>>>> CPU_BASED_CR8_LOAD_EXITING |
>>>> @@ -2531,6 +2588,17 @@ static __init int setup_vmcs_config(struct
>>> vmcs_config *vmcs_conf)
>>>> &_vmexit_control) < 0)
>>>> return -EIO;
>>>> + min = PIN_BASED_EXT_INTR_MASK | PIN_BASED_NMI_EXITING;
>>>> + opt = PIN_BASED_VIRTUAL_NMIS | PIN_BASED_POSTED_INTR;
>>>> + if (adjust_vmx_controls(min, opt, MSR_IA32_VMX_PINBASED_CTLS,
>>>> + &_pin_based_exec_control) < 0)
>>>> + return -EIO;
>>>> +
>>>> + if (!(_cpu_based_2nd_exec_control &
>>>> + SECONDARY_EXEC_VIRTUAL_INTR_DELIVERY) ||
>>>> + !(_vmexit_control & VM_EXIT_ACK_INTR_ON_EXIT))
>>>> + _pin_based_exec_control &= ~PIN_BASED_POSTED_INTR;
>>>> +
>>>> min = 0; opt = VM_ENTRY_LOAD_IA32_PAT; if
>>>> (adjust_vmx_controls(min, opt, MSR_IA32_VMX_ENTRY_CTLS, @@ -2715,6
>>>> +2783,9 @@ static __init int hardware_setup(void) if
>>>> (!cpu_has_vmx_virtual_intr_delivery()) enable_apicv_vid = 0;
>>>> + if (!cpu_has_vmx_posted_intr() || !x2apic_enabled())
>>> In nested guest x2apic may be enabled without irq remapping. Check for
>>> irq remapping here.
>>
>> There are no posted interrupt available in nested case. We don't need
>> to check IR here.
>>
> One day emulation will be added. If pre-request for PI is IR check
> for IR.
> BTW why IR is needed for PI. To deliver assigned devices interrupts
> directly into a guest sure, but why is it required for delivering
> interrupts from emulated devices or IPIs?
Posted Interrupt support is Xeon only and these platforms will have x2APIC. So, Linux will enable x2APIC on these platforms. So we only want to enable PI when x2apic is enabled and IR is required for x2apic.
>>>
>>>> + enable_apicv_pi = 0;
>>>> +
>>>> if (nested) nested_vmx_setup_ctls_msrs(); @@ -3881,6 +3952,93 @@
>>>> static void ept_set_mmio_spte_mask(void)
>>>> kvm_mmu_set_mmio_spte_mask(0xffull << 49 | 0x6ull); }
>>>> +irqreturn_t pi_handler(int irq, void *data)
>>>> +{
>>>> + struct vcpu_vmx *vmx = data;
>>>> +
>>>> + kvm_make_request(KVM_REQ_EVENT, &vmx->vcpu);
>>>> + kvm_vcpu_kick(&vmx->vcpu);
>>>> +
>>>> + return IRQ_HANDLED;
>>>> +}
>>>> +
>>>> +static int vmx_has_posted_interrupt(struct kvm_vcpu *vcpu)
>>>> +{
>>>> + return irqchip_in_kernel(vcpu->kvm) && enable_apicv_pi;
>>>> +}
>>>> +
>>>> +static void vmx_pi_migrate(struct kvm_vcpu *vcpu)
>>>> +{
>>>> + int ret = 0;
>>>> + struct vcpu_vmx *vmx = to_vmx(vcpu);
>>>> +
>>>> + if (!enable_apicv_pi)
>>>> + return ;
>>>> +
>>>> + preempt_disable();
>>>> + local_irq_disable();
>>>> + if (!vmx->irq) {
>>>> + ret = arch_pi_alloc_irq(vmx);
>>>> + if (ret < 0) {
>>>> + vmx->irq = -1;
>>>> + goto out;
>>>> + }
>>>> + vmx->irq = ret;
>>>> +
>>>> + ret = request_irq(vmx->irq, pi_handler, IRQF_NO_THREAD,
>>>> + "Posted Interrupt", vmx);
>>>> + if (ret) {
>>>> + vmx->irq = -1;
>>>> + goto out;
>>>> + }
>>>> +
>>>> + ret = arch_pi_get_vector(vmx->irq);
>>>> + } else
>>>> + ret = arch_pi_migrate(vmx->irq, smp_processor_id());
>>>> +
>>>> + if (ret < 0) {
>>>> + vmx->irq = -1;
>>>> + goto out;
>>>> + } else {
>>>> + vmx->vector = ret;
>>>> + vmcs_write16(POSTED_INTR_NV, vmx->vector);
>>>> + pi_clear_on(vmx->pi);
>>>> + }
>>>> +out:
>>>> + local_irq_enable();
>>>> + preempt_enable();
>>>> + return ;
>>>> +}
>>>> +
>>>> +static int vmx_send_nv(struct kvm_vcpu *vcpu,
>>>> + int vector)
>>>> +{
>>>> + struct vcpu_vmx *vmx = to_vmx(vcpu);
>>>> +
>>>> + if (unlikely(vmx->irq == -1))
>>>> + return 0;
>>>> +
>>>> + if (vcpu->cpu == smp_processor_id()) {
>>>> + pi_set_on(vmx->pi);
>>> Why? You clear this bit anyway in vmx_update_irq() during guest entry.
>> Here means the target vcpu already in vmx non-root mode. Then it will
> consume the interrupt on next vm entry and we don't need to send the
> notification event from other cpu, just update PIR is enough.
> I understand why you avoid sending PI IPI here, but you do not update
> pir in this case either. You only set "on" bit here and set vector directly
> in IRR in __apic_accept_irq() since vmx_send_nv() returns 0 in this case.
> Interrupt is delivered from IRR on the next entry.
As I mentioned, IRR is basically same as PIR.
>>
>>>> + return 0; + } + + pi_set_pir(vector, vmx->pi); + if
>>>> (!pi_test_and_set_on(vmx->pi) && (vcpu->mode == IN_GUEST_MODE)) {
>>>> + apic->send_IPI_mask(get_cpu_mask(vcpu->cpu), vmx->vector);
>>>> + return 1; + } + return 0; +} + +static void free_pi(struct
>>>> vcpu_vmx *vmx) +{ + if (enable_apicv_pi) { + kfree(vmx->pi);
>>>> + arch_pi_free_irq(vmx->irq, vmx); + } +} +
>>>> /*
>>>> * Sets up the vmcs for emulated real mode.
>>>> */
>>>> @@ -3890,6 +4048,7 @@ static int vmx_vcpu_setup(struct vcpu_vmx *vmx)
>>>> unsigned long a;
>>>> #endif
>>>> int i;
>>>> + u32 pin_based_exec_ctrl = vmcs_config.pin_based_exec_ctrl;
>>>>
>>>> /* I/O */ vmcs_write64(IO_BITMAP_A, __pa(vmx_io_bitmap_a)); @@
>>>> -3901,8 +4060,10 @@ static int vmx_vcpu_setup(struct vcpu_vmx *vmx)
>>>> vmcs_write64(VMCS_LINK_POINTER, -1ull); /* 22.3.1.5 */
>>>>
>>>> /* Control */
>>>> - vmcs_write32(PIN_BASED_VM_EXEC_CONTROL,
>>>> - vmcs_config.pin_based_exec_ctrl); + if (!enable_apicv_pi)
>>>> + pin_based_exec_ctrl &= ~PIN_BASED_POSTED_INTR; +
>>>> + vmcs_write32(PIN_BASED_VM_EXEC_CONTROL, pin_based_exec_ctrl);
>>>>
>>>> vmcs_write32(CPU_BASED_VM_EXEC_CONTROL,
>>> vmx_exec_control(vmx));
>>>>
>>>> @@ -3920,6 +4081,12 @@ static int vmx_vcpu_setup(struct vcpu_vmx
> *vmx)
>>>> vmcs_write16(GUEST_INTR_STATUS, 0);
>>>> }
>>>> + if (enable_apicv_pi) {
>>>> + vmx->pi = kmalloc(sizeof(struct pi_desc),
>>>> + GFP_KERNEL | __GFP_ZERO);
>>>> + vmcs_write64(POSTED_INTR_DESC_ADDR, __pa((vmx->pi)));
>>>> + }
>>>> +
>>>> if (ple_gap) { vmcs_write32(PLE_GAP, ple_gap);
>>>> vmcs_write32(PLE_WINDOW, ple_window); @@ -6161,6 +6328,11 @@
>>>> static void vmx_update_irq(struct kvm_vcpu *vcpu) if
>>>> (!enable_apicv_vid) return ;
>>>> + if (enable_apicv_pi) {
>>>> + kvm_apic_update_irr(vcpu, (unsigned int *)vmx->pi->pir);
>>>> + pi_clear_on(vmx->pi);
>>> Why do you do that? Isn't VMX process posted interrupts on vmentry if
>>> "on" bit is set?
> Can you answer this question?
No, vmentry do nothing for PI. Posted interrupt only happens when an unmasked external interrupt arrived and the target vcpu is running. Beyond that, cpu follow the old way.
>> Best regards,
>> Yang
>>
>
> --
> Gleb.
Best regards,
Yang
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
