On 11/6/19 10:14 AM, Sean Christopherson wrote:
> On Wed, Nov 06, 2019 at 08:36:38AM -0500, Nitesh Narayan Lal wrote:
>> In IOAPIC fixed delivery mode instead of flushing the scan
>> requests to all vCPUs, we should only send the requests to
>> vCPUs specified within the destination field.
>>
>> This patch introduces kvm_get_dest_vcpus_mask() API which
>> retrieves an array of target vCPUs by using
>> kvm_apic_map_get_dest_lapic() and then based on the
>> vcpus_idx, it sets the bit in a bitmap. However, if the above
>> fails kvm_get_dest_vcpus_mask() finds the target vCPUs by
>> traversing all available vCPUs. Followed by setting the
>> bits in the bitmap.
>>
>> If we had different vCPUs in the previous request for the
>> same redirection table entry then bits corresponding to
>> these vCPUs are also set. This to done to keep
>> ioapic_handled_vectors synchronized.
>>
>> This bitmap is then eventually passed on to
>> kvm_make_vcpus_request_mask() to generate a masked request
>> only for the target vCPUs.
>>
>> This would enable us to reduce the latency overhead on isolated
>> vCPUs caused by the IPI to process due to KVM_REQ_IOAPIC_SCAN.
>>
>> Suggested-by: Marcelo Tosatti <mtosatti@redhat.com>
>> Signed-off-by: Nitesh Narayan Lal <nitesh@redhat.com>
>> ---
>>  arch/x86/include/asm/kvm_host.h |  2 ++
>>  arch/x86/kvm/ioapic.c           | 33 ++++++++++++++++++++++++++++--
>>  arch/x86/kvm/lapic.c            | 45 +++++++++++++++++++++++++++++++++++++++++
>>  arch/x86/kvm/lapic.h            |  3 +++
>>  arch/x86/kvm/x86.c              |  6 ++++++
>>  include/linux/kvm_host.h        |  2 ++
>>  virt/kvm/kvm_main.c             | 14 +++++++++++++
>>  7 files changed, 103 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
>> index 24d6598..b2aca6d 100644
>> --- a/arch/x86/include/asm/kvm_host.h
>> +++ b/arch/x86/include/asm/kvm_host.h
>> @@ -1571,6 +1571,8 @@ int kvm_pv_send_ipi(struct kvm *kvm, unsigned long ipi_bitmap_low,
>>  
>>  void kvm_make_mclock_inprogress_request(struct kvm *kvm);
>>  void kvm_make_scan_ioapic_request(struct kvm *kvm);
>> +void kvm_make_scan_ioapic_request_mask(struct kvm *kvm,
>> +				       unsigned long *vcpu_bitmap);
>>  
>>  void kvm_arch_async_page_not_present(struct kvm_vcpu *vcpu,
>>  				     struct kvm_async_pf *work);
>> diff --git a/arch/x86/kvm/ioapic.c b/arch/x86/kvm/ioapic.c
>> index d859ae8..c8d0a83 100644
>> --- a/arch/x86/kvm/ioapic.c
>> +++ b/arch/x86/kvm/ioapic.c
>> @@ -271,8 +271,9 @@ static void ioapic_write_indirect(struct kvm_ioapic *ioapic, u32 val)
>>  {
>>  	unsigned index;
>>  	bool mask_before, mask_after;
>> -	int old_remote_irr, old_delivery_status;
>>  	union kvm_ioapic_redirect_entry *e;
>> +	unsigned long vcpu_bitmap;
>> +	int old_remote_irr, old_delivery_status, old_dest_id, old_dest_mode;
>>  
>>  	switch (ioapic->ioregsel) {
>>  	case IOAPIC_REG_VERSION:
>> @@ -296,6 +297,8 @@ static void ioapic_write_indirect(struct kvm_ioapic *ioapic, u32 val)
>>  		/* Preserve read-only fields */
>>  		old_remote_irr = e->fields.remote_irr;
>>  		old_delivery_status = e->fields.delivery_status;
>> +		old_dest_id = e->fields.dest_id;
>> +		old_dest_mode = e->fields.dest_mode;
>>  		if (ioapic->ioregsel & 1) {
>>  			e->bits &= 0xffffffff;
>>  			e->bits |= (u64) val << 32;
>> @@ -321,7 +324,33 @@ static void ioapic_write_indirect(struct kvm_ioapic *ioapic, u32 val)
>>  		if (e->fields.trig_mode == IOAPIC_LEVEL_TRIG
>>  		    && ioapic->irr & (1 << index))
>>  			ioapic_service(ioapic, index, false);
>> -		kvm_make_scan_ioapic_request(ioapic->kvm);
>> +		if (e->fields.delivery_mode == APIC_DM_FIXED) {
>> +			struct kvm_lapic_irq irq;
>> +
>> +			irq.shorthand = 0;
>> +			irq.vector = e->fields.vector;
>> +			irq.delivery_mode = e->fields.delivery_mode << 8;
>> +			irq.dest_id = e->fields.dest_id;
>> +			irq.dest_mode = e->fields.dest_mode;
>> +			kvm_get_dest_vcpus_mask(ioapic->kvm, &irq,
>> +						&vcpu_bitmap);
>> +			if (old_dest_mode != e->fields.dest_mode ||
>> +			    old_dest_id != e->fields.dest_id) {
>> +				/*
>> +				 * Update vcpu_bitmap with vcpus specified in
>> +				 * the previous request as well. This is done to
>> +				 * keep ioapic_handled_vectors synchronized.
>> +				 */
>> +				irq.dest_id = old_dest_id;
>> +				irq.dest_mode = old_dest_mode;
>> +				kvm_get_dest_vcpus_mask(ioapic->kvm, &irq,
>> +							&vcpu_bitmap);
>> +			}
>> +			kvm_make_scan_ioapic_request_mask(ioapic->kvm,
>> +							  &vcpu_bitmap);
>> +		} else {
>> +			kvm_make_scan_ioapic_request(ioapic->kvm);
>> +		}
>>  		break;
>>  	}
>>  }
>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>> index b29d00b..90869c4 100644
>> --- a/arch/x86/kvm/lapic.c
>> +++ b/arch/x86/kvm/lapic.c
>> @@ -1124,6 +1124,51 @@ static int __apic_accept_irq(struct kvm_lapic *apic, int delivery_mode,
>>  	return result;
>>  }
>>  
>> +/*
>> + * This routine identifies the destination vcpus mask meant to receive the
>> + * IOAPIC interrupts. It either uses kvm_apic_map_get_dest_lapic() to find
>> + * out the destination vcpus array and set the bitmap or it traverses to
>> + * each available vcpu to identify the same.
>> + */
>> +void kvm_get_dest_vcpus_mask(struct kvm *kvm, struct kvm_lapic_irq *irq,
>> +			     unsigned long *vcpu_bitmap)
>> +{
>> +	struct kvm_lapic **dest_vcpu = NULL;
>> +	struct kvm_lapic *src = NULL;
>> +	struct kvm_apic_map *map;
>> +	struct kvm_vcpu *vcpu;
>> +	unsigned long bitmap;
>> +	int i, vcpus_idx;
>> +	bool ret;
>> +
>> +	rcu_read_lock();
>> +	map = rcu_dereference(kvm->arch.apic_map);
>> +
>> +	ret = kvm_apic_map_get_dest_lapic(kvm, &src, irq, map, &dest_vcpu,
>> +					  &bitmap);
>> +	if (ret) {
>> +		for_each_set_bit(i, &bitmap, 16) {
>> +			if (!dest_vcpu[i])
>> +				continue;
>> +			vcpus_idx = dest_vcpu[i]->vcpu->vcpus_idx;
>> +			__set_bit(vcpus_idx, vcpu_bitmap);
>> +		}
>> +	} else {
>> +		kvm_for_each_vcpu(i, vcpu, kvm) {
>> +			if (!kvm_apic_present(vcpu))
>> +				continue;
>> +			if (!kvm_apic_match_dest(vcpu, NULL,
>> +						 irq->delivery_mode,
>> +						 irq->dest_id,
>> +						 irq->dest_mode))
>> +				continue;
>> +			vcpus_idx = dest_vcpu[i]->vcpu->vcpus_idx;
> This can't possibly be correct.  AFAICT, dest_vcpu is guaranteed to be
> *NULL* when kvm_apic_map_get_dest_lapic() returns false, and on top of
> that I'm pretty sure it's not intended to be indexed by the vcpu index.
>
> But vcpus_idx isn't needed here, you already have @vcpu, and even that
> is superfluous as @i itself is the vcpu index.

Yeah, its a bug. I will correct it.

>
> It's probably worth manually testing this path by forcing @ret to false,
> I'm guessing it's not being hit in normal operation.

Sure, good idea.

>
>> +			__set_bit(vcpus_idx, vcpu_bitmap);
>> +		}
>> +	}
>> +	rcu_read_unlock();
>> +}
>> +
>>  int kvm_apic_compare_prio(struct kvm_vcpu *vcpu1, struct kvm_vcpu *vcpu2)
>>  {
>>  	return vcpu1->arch.apic_arb_prio - vcpu2->arch.apic_arb_prio;
>> diff --git a/arch/x86/kvm/lapic.h b/arch/x86/kvm/lapic.h
>> index 1f501485..49b0c6c 100644
>> --- a/arch/x86/kvm/lapic.h
>> +++ b/arch/x86/kvm/lapic.h
>> @@ -226,6 +226,9 @@ static inline int kvm_lapic_latched_init(struct kvm_vcpu *vcpu)
>>  
>>  void kvm_wait_lapic_expire(struct kvm_vcpu *vcpu);
>>  
>> +void kvm_get_dest_vcpus_mask(struct kvm *kvm, struct kvm_lapic_irq *irq,
>> +			     unsigned long *vcpu_bitmap);
>> +
>>  bool kvm_intr_is_single_vcpu_fast(struct kvm *kvm, struct kvm_lapic_irq *irq,
>>  			struct kvm_vcpu **dest_vcpu);
>>  int kvm_vector_to_index(u32 vector, u32 dest_vcpus,
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index ff395f8..ee6945f 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -7838,6 +7838,12 @@ static void process_smi(struct kvm_vcpu *vcpu)
>>  	kvm_make_request(KVM_REQ_EVENT, vcpu);
>>  }
>>  
>> +void kvm_make_scan_ioapic_request_mask(struct kvm *kvm,
>> +				       unsigned long *vcpu_bitmap)
>> +{
>> +	kvm_make_cpus_request_mask(kvm, KVM_REQ_SCAN_IOAPIC, vcpu_bitmap);
>> +}
>> +
>>  void kvm_make_scan_ioapic_request(struct kvm *kvm)
>>  {
>>  	kvm_make_all_cpus_request(kvm, KVM_REQ_SCAN_IOAPIC);
>> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
>> index 31c4fde..2f69eae 100644
>> --- a/include/linux/kvm_host.h
>> +++ b/include/linux/kvm_host.h
>> @@ -786,6 +786,8 @@ int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
>>  bool kvm_make_vcpus_request_mask(struct kvm *kvm, unsigned int req,
>>  				 unsigned long *vcpu_bitmap, cpumask_var_t tmp);
>>  bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req);
>> +bool kvm_make_cpus_request_mask(struct kvm *kvm, unsigned int req,
>> +				unsigned long *vcpu_bitmap);
>>  
>>  long kvm_arch_dev_ioctl(struct file *filp,
>>  			unsigned int ioctl, unsigned long arg);
>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> index 24ab711..9e85df8 100644
>> --- a/virt/kvm/kvm_main.c
>> +++ b/virt/kvm/kvm_main.c
>> @@ -242,6 +242,20 @@ bool kvm_make_vcpus_request_mask(struct kvm *kvm, unsigned int req,
>>  	return called;
>>  }
>>  
>> +bool kvm_make_cpus_request_mask(struct kvm *kvm, unsigned int req,
>> +				unsigned long *vcpu_bitmap)
>> +{
>> +	cpumask_var_t cpus;
>> +	bool called;
>> +
>> +	zalloc_cpumask_var(&cpus, GFP_ATOMIC);
>> +
>> +	called = kvm_make_vcpus_request_mask(kvm, req, vcpu_bitmap, cpus);
>> +
>> +	free_cpumask_var(cpus);
>> +	return called;
> kvm_make_all_cpus_request() should call this new function, the code is
> identical except for its declared vcpu_bitmap.  

I was also skeptical about this particular change.
The other way I thought of doing this was by modifying the definition of
kvm_make_all_cpus_request(). But I was not sure if that is a good idea as it
would have required more changes in the existing code.

>> +}
>> +
>>  bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req)
>>  {
>>  	cpumask_var_t cpus;
>> -- 
>> 1.8.3.1
>>
-- 
Thanks
Nitesh