VMX: configure posted interrupt descriptor when assigning device
cover VMX: configure posted interrupt descriptor when assigning device
Configuration of the posted interrupt descriptor is incorrect when devices are hotplugged to the guest (and vcpus are halted).
See patch 2 for details.
PATCH 1/2 KVM: x86: add start_assignment hook to kvm_x86_ops
Add a start_assignment hook to kvm_x86_ops, which is called when kvm_arch_start_assignment is done.
The hook is required to update the wakeup vector of a sleeping vCPU when a device is assigned to the guest.
arch/x86/include/asm/kvm_host.h +1 -0
@@ -1322,6 +1322,7 @@ struct kvm_x86_ops {int (*update_pi_irte)(struct kvm *kvm,unsigned int host_irq,uint32_t guest_irq,bool set); +void (*start_assignment)(struct kvm *kvm,int device_count);void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);bool (*dy_apicv_has_pending_interrupt)(struct kvm_vcpu *vcpu);
arch/x86/kvm/svm/svm.c +1 -0
@@ -4601,6 +4601,7 @@ static struct kvm_x86_ops svm_x86_ops __ .deliver_posted_interrupt = svm_deliver_avic_intr, .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt, .update_pi_irte = svm_update_pi_irte, + .start_assignment = NULL, .setup_mce = svm_setup_mce, .smi_allowed = svm_smi_allowed,
arch/x86/kvm/vmx/vmx.c +1 -0
@@ -7732,6 +7732,7 @@ static struct kvm_x86_ops vmx_x86_ops __ .nested_ops = &vmx_nested_ops, .update_pi_irte = pi_update_irte, + .start_assignment =NULL , #ifdef CONFIG_X86_64 .set_hv_timer = vmx_set_hv_timer,
arch/x86/kvm/x86.c +5 -1
@@ -11295,7 +11295,11 @@ bool kvm_arch_can_dequeue_async_page_prevoid kvm_arch_start_assignment (struct kvm *kvm) { -atomic_inc (&kvm->arch.assigned_device_count); +int ret; + + ret =atomic_inc_return (&kvm->arch.assigned_device_count); +if (kvm_x86_ops.start_assignment) + kvm_x86_ops.start_assignment (kvm, ret); }EXPORT_SYMBOL_GPL (kvm_arch_start_assignment);
Sean Christopherson
maintainerRe: [patch 1/2] KVM: x86: add start_assignment hook to kvm_x86_ops
Quoted 5 lines · Marcelo Tosatti
Index:
kvm/arch/x86/kvm/x86.c
===================================================================
arch/x86/kvm/x86.c bool kvm_arch_can_dequeue_async_page_pre · line 11302
Quoted 9 lines
arch/x86/kvm/x86.c +5
-1@@ -11295,7 +11295,11 @@ bool kvm_arch_can_dequeue_async_page_prevoid kvm_arch_start_assignment (struct kvm *kvm) { -atomic_inc (&kvm->arch.assigned_device_count); +int ret; + + ret =atomic_inc_return (&kvm->arch.assigned_device_count); +if (kvm_x86_ops.start_assignment) + kvm_x86_ops.start_assignment (kvm, ret);
This can be a static_call(), just needs an entry in kvm-x86-ops.h.
Quoted 5 lines
arch/x86/kvm/x86.c +0
-0}EXPORT_SYMBOL_GPL (kvm_arch_start_assignment);
PATCH 2/2 KVM: VMX: update vcpu posted-interrupt descriptor when assigning device
In reply to Marcelo TosattiFor VMX, when a vcpu enters HLT emulation, pi_post_block will:
1) Add vcpu to per-cpu list of blocked vcpus.
2) Program the posted-interrupt descriptor "notification vector" to POSTED_INTR_WAKEUP_VECTOR
With interrupt remapping, an interrupt will set the PIR bit for the vector programmed for the device on the CPU, test-and-set the ON bit on the posted interrupt descriptor, and if the ON bit is clear generate an interrupt for the notification vector.
This way, the target CPU wakes upon a device interrupt and wakes up the target vcpu.
Problem is that pi_post_block only programs the notification vector if kvm_arch_has_assigned_device() is true. Its possible for the following to happen:
1) vcpu V HLTs on pcpu P, kvm_arch_has_assigned_device is false, notification vector is not programmed
2) device is assigned to VM
3) device interrupts vcpu V, sets ON bit (notification vector not programmed, so pcpu P remains in idle)
4) vcpu 0 IPIs vcpu V (in guest), but since pi descriptor ON bit is set, kvm_vcpu_kick is skipped
5) vcpu 0 busy spins on vcpu V's response for several seconds, until RCU watchdog NMIs all vCPUs.
To fix this, use the start_assignment kvm_x86_ops callback to program the notification vector when assigned device count changes from 0 to 1.
arch/x86/kvm/vmx/posted_intr.c +70 -10
@@ -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 ); +} + +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)) +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 ) +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 (); + 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 (); + } +}/* * pi_update_irte - set IRTE for Posted-Interrupts
arch/x86/kvm/vmx/posted_intr.h +1 -0
@@ -95,5 +95,6 @@ void __init pi_init_cpu(int cpu);bool pi_has_pending_interrupt (struct kvm_vcpu *vcpu);int pi_update_irte (struct kvm *kvm,unsigned int host_irq,uint32_t guest_irq,bool set); +void vmx_pi_start_assignment (struct kvm *kvm,int device_count);#endif /* __KVM_X86_VMX_POSTED_INTR_H */
arch/x86/kvm/vmx/vmx.h +3 -0
@@ -336,6 +336,9 @@ struct vcpu_vmx {DECLARE_BITMAP (read,MAX_POSSIBLE_PASSTHROUGH_MSRS );DECLARE_BITMAP (write,MAX_POSSIBLE_PASSTHROUGH_MSRS ); } shadow_msr_intercept; + +/* true if vcpu is between pre_block and post_block */ +bool in_blocked_section; };enum ept_pointers_status {
Re: [patch 2/2] KVM: VMX: update vcpu posted-interrupt descriptor when assigning device
Quoted 29 lines · Marcelo Tosatti
For VMX, when a vcpu enters HLT emulation,
pi_post_blockwill:1) Add vcpu to per-cpu list of blocked vcpus.
2) Program the posted-interrupt descriptor "notification vector" to
POSTED_INTR_WAKEUP_VECTORWith interrupt remapping, an interrupt will set the PIR bit for the vector programmed for the device on the CPU, test-and-set the ON bit on the posted interrupt descriptor, and if the ON bit is clear generate an interrupt for the notification vector.
This way, the target CPU wakes upon a device interrupt and wakes up the target vcpu.
Problem is that
pi_post_blockonly programs the notification vector ifkvm_arch_has_assigned_device()is true. Its possible for the following to happen:1) vcpu V HLTs on pcpu P,
kvm_arch_has_assigned_deviceis false, notification vector is not programmed
2) device is assigned to VM
3) device interrupts vcpu V, sets ON bit (notification vector not programmed, so pcpu P remains in idle)
4) vcpu 0 IPIs vcpu V (in guest), but since pi descriptor ON bit is set,kvm_vcpu_kickis skipped
5) vcpu 0 busy spins on vcpu V's response for several seconds, until RCU watchdog NMIs all vCPUs.
Quoted 5 lines
To fix this, use the
start_assignmentkvm_x86_opscallback to program the notification vector when assigned device count changes from 0 to 1.Reported-by Pei Zhang <[email protected]>Signed-off-by Marcelo Tosatti <[email protected]>
Argh, missing setting vmx_pi_start_assignment, will resend v2.
PATCH v2 2/2 KVM: VMX: update vcpu posted-interrupt descriptor when assigning device
In reply to Marcelo TosattiFor VMX, when a vcpu enters HLT emulation, pi_post_block will:
1) Add vcpu to per-cpu list of blocked vcpus.
2) Program the posted-interrupt descriptor "notification vector" to POSTED_INTR_WAKEUP_VECTOR
With interrupt remapping, an interrupt will set the PIR bit for the vector programmed for the device on the CPU, test-and-set the ON bit on the posted interrupt descriptor, and if the ON bit is clear generate an interrupt for the notification vector.
This way, the target CPU wakes upon a device interrupt and wakes up the target vcpu.
Problem is that pi_post_block only programs the notification vector if kvm_arch_has_assigned_device() is true. Its possible for the following to happen:
1) vcpu V HLTs on pcpu P, kvm_arch_has_assigned_device is false, notification vector is not programmed
2) device is assigned to VM
3) device interrupts vcpu V, sets ON bit (notification vector not programmed, so pcpu P remains in idle)
4) vcpu 0 IPIs vcpu V (in guest), but since pi descriptor ON bit is set, kvm_vcpu_kick is skipped
5) vcpu 0 busy spins on vcpu V's response for several seconds, until RCU watchdog NMIs all vCPUs.
To fix this, use the start_assignment kvm_x86_ops callback to program the notification vector when assigned device count changes from 0 to 1.
---
v2: add vmx_pi_start_assignment to vmx's kvm_x86_ops
arch/x86/kvm/vmx/posted_intr.c +70 -10
@@ -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 ); +} + +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)) +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 ) +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 (); + 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 (); + } +}/* * pi_update_irte - set IRTE for Posted-Interrupts
arch/x86/kvm/vmx/posted_intr.h +1 -0
@@ -95,5 +95,6 @@ void __init pi_init_cpu(int cpu);bool pi_has_pending_interrupt (struct kvm_vcpu *vcpu);int pi_update_irte (struct kvm *kvm,unsigned int host_irq,uint32_t guest_irq,bool set); +void vmx_pi_start_assignment (struct kvm *kvm,int device_count);#endif /* __KVM_X86_VMX_POSTED_INTR_H */
arch/x86/kvm/vmx/vmx.h +3 -0
@@ -336,6 +336,9 @@ struct vcpu_vmx {DECLARE_BITMAP (read,MAX_POSSIBLE_PASSTHROUGH_MSRS );DECLARE_BITMAP (write,MAX_POSSIBLE_PASSTHROUGH_MSRS ); } shadow_msr_intercept; + +/* true if vcpu is between pre_block and post_block */ +bool in_blocked_section; };enum ept_pointers_status {
arch/x86/kvm/vmx/vmx.c +1 -1
@@ -7732,7 +7732,7 @@ static struct kvm_x86_ops vmx_x86_ops __ .nested_ops = &vmx_nested_ops, .update_pi_irte = pi_update_irte, - .start_assignment =NULL , + .start_assignment = vmx_pi_start_assignment, #ifdef CONFIG_X86_64 .set_hv_timer = vmx_set_hv_timer,
Sean Christopherson
maintainerRe: [patch 2/2 V2] KVM: VMX: update vcpu posted-interrupt descriptor when assigning device
Quoted 37 lines · Marcelo Tosatti
Index:
kvm/arch/x86/kvm/vmx/posted_intr.c
===================================================================arch/x86/kvm/vmx/posted_intr.c +3
-10@@ -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,
arch/x86/kvm/vmx/posted_intr.c int pi_pre_block(struct kvm_vcpu *vcpu) · line 184
Quoted 4 lines
arch/x86/kvm/vmx/posted_intr.c +1
-0@@ -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.
arch/x86/kvm/vmx/posted_intr.c int pi_pre_block(struct kvm_vcpu *vcpu) · line 195
Quoted 11 lines
arch/x86/kvm/vmx/posted_intr.c +11
-0+} + +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?
Quoted 21 lines
arch/x86/kvm/vmx/posted_intr.c +9
-0+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)); }
arch/x86/kvm/vmx/posted_intr.c bool pi_has_pending_interrupt(struct kvm · line 261
Quoted 10 lines
arch/x86/kvm/vmx/posted_intr.c +9
-0+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?
Quoted 19 lines
arch/x86/kvm/vmx/posted_intr.c +19
-0+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 ;
arch/x86/kvm/vmx/posted_intr.c bool pi_has_pending_interrupt(struct kvm · line 290
Quoted 10 lines
arch/x86/kvm/vmx/posted_intr.c +10
-0+ +/* 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.
Quoted 8 lines
arch/x86/kvm/vmx/posted_intr.c +8
-0+ 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 (); + } +}
Re: [patch 2/2 V2] KVM: VMX: update vcpu posted-interrupt descriptor when assigning device
Hi Sean,
Quoted 41 lines · Sean Christopherson
Quoted 40 lines · Marcelo Tosatti
Index:
kvm/arch/x86/kvm/vmx/posted_intr.c
===================================================================arch/x86/kvm/vmx/posted_intr.c +3
-10@@ -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);
Quoted 4 lines
arch/x86/kvm/vmx/posted_intr.c int pi_pre_block(struct kvm_vcpu *vcpu) · line 184
Quoted 1 line
arch/x86/kvm/vmx/posted_intr.c +1
-0+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.
Done.
Quoted 5 lines
Quoted 5 lines
arch/x86/kvm/vmx/posted_intr.c +5
-0+} + +int pi_pre_block (struct kvm_vcpu *vcpu) +{ +struct vcpu_vmx *vmx =to_vmx (vcpu);
Quoted 8 lines
arch/x86/kvm/vmx/posted_intr.c int pi_pre_block(struct kvm_vcpu *vcpu) · line 195
Quoted 6 lines
arch/x86/kvm/vmx/posted_intr.c +6
-0+ + 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?
Done.
Quoted 22 lines
Quoted 22 lines
arch/x86/kvm/vmx/posted_intr.c +9
-0+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)); }
Quoted 13 lines
arch/x86/kvm/vmx/posted_intr.c bool pi_has_pending_interrupt(struct kvm · line 261
Quoted 9 lines
arch/x86/kvm/vmx/posted_intr.c +9
-0+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()regardingpre_cpumuddies 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.
From Intel's manual:
"29.6 POSTED-INTERRUPT PROCESSING
...
Use of the posted-interrupt descriptor differs from that of other data structures that are referenced by pointers in a VMCS. There is a general requirement that software ensure that each such data structure is modified only when no logical processor with a current VMCS that references it is in VMX non-root operation. That requirement does not apply to the posted-interrupt descriptor. There is a requirement, however, that such modifications be done using locked read-modify-write instructions."
Quoted 5 lines
If the vCPU is migrated after
vmx_pi_start_assignment()grabsvcpu->cpubut before the IPI arrives (to runpi_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 doublelist_add_tail.
Good point.
Quoted 3 lines
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 usethis_cpu_ptr().
Well problem is it might not exit kvm_vcpu_block(). However that can be fixed.
Quoted 33 lines
Because the existing
code grabsvcpu->cpuwith IRQs disabled and is called only fromKVM_RUN,vcpu->cpuis guaranteed to match the current pCPU sincevcpu->cpuwill 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?Quoted 24 lines
arch/x86/kvm/vmx/posted_intr.c +24
-0+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;
Quoted 8 lines
arch/x86/kvm/vmx/posted_intr.c bool pi_has_pending_interrupt(struct kvm · line 290
Quoted 5 lines
arch/x86/kvm/vmx/posted_intr.c +5
-0+ +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.
kvm_vcpu_kick seems cleaner, just need to add another arch hook to allow kvm_vcpu_block() to return.
Thanks for the review! Will resend after testing.