Re: [RFC PATCH 09/45] KVM: arm/arm64: vgic-new: Add GICv2 IRQ sync/flush

[Christoffer Dall <christoffer.dall@linaro.org>]

On Fri, Mar 25, 2016 at 02:04:32AM +0000, Andre Przywara wrote:
> From: Marc Zyngier <marc.zyngier@arm.com>
> 
> Implement the functionality for syncing IRQs between our emulation
> and the list registers, which represent the guest's view of IRQs.
> This is done in kvm_vgic_flush_hwstate and kvm_vgic_sync_hwstate,
> which gets called on guest entry and exit.
> 
> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
> Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
> Signed-off-by: Eric Auger <eric.auger@linaro.org>
> Signed-off-by: Andre Przywara <andre.przywara@arm.com>
> ---
>  include/kvm/vgic/vgic.h     |   4 +
>  virt/kvm/arm/vgic/vgic-v2.c | 161 ++++++++++++++++++++++++++++++++++
>  virt/kvm/arm/vgic/vgic.c    | 204 ++++++++++++++++++++++++++++++++++++++++++++
>  virt/kvm/arm/vgic/vgic.h    |   4 +
>  4 files changed, 373 insertions(+)
> 
> diff --git a/include/kvm/vgic/vgic.h b/include/kvm/vgic/vgic.h
> index f32b284..986f23f 100644
> --- a/include/kvm/vgic/vgic.h
> +++ b/include/kvm/vgic/vgic.h
> @@ -187,6 +187,10 @@ int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int intid,
>  #define vgic_valid_spi(k,i)	(((i) >= VGIC_NR_PRIVATE_IRQS) && \
>  			((i) < (k)->arch.vgic.nr_spis + VGIC_NR_PRIVATE_IRQS))
>  
> +bool kvm_vcpu_has_pending_irqs(struct kvm_vcpu *vcpu);
> +void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu);
> +void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu);
> +
>  /**
>   * kvm_vgic_get_max_vcpus - Get the maximum number of VCPUs allowed by HW
>   *
> diff --git a/virt/kvm/arm/vgic/vgic-v2.c b/virt/kvm/arm/vgic/vgic-v2.c
> index 0bf6f27..1cec423 100644
> --- a/virt/kvm/arm/vgic/vgic-v2.c
> +++ b/virt/kvm/arm/vgic/vgic-v2.c
> @@ -14,11 +14,172 @@
>   * along with this program.  If not, see <http://www.gnu.org/licenses/>.
>   */
>  
> +#include <linux/irqchip/arm-gic.h>
>  #include <linux/kvm.h>
>  #include <linux/kvm_host.h>
>  
>  #include "vgic.h"
>  
> +/*
> + * Call this function to convert a u64 value to an unsigned long * bitmask
> + * in a way that works on both 32-bit and 64-bit LE and BE platforms.
> + *
> + * Warning: Calling this function may modify *val.
> + */
> +static unsigned long *u64_to_bitmask(u64 *val)
> +{
> +#if defined(CONFIG_CPU_BIG_ENDIAN) && BITS_PER_LONG == 32
> +	*val = (*val >> 32) | (*val << 32);
> +#endif
> +	return (unsigned long *)val;
> +}
> +
> +void vgic_v2_process_maintenance(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_v2_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v2;
> +
> +	if (cpuif->vgic_misr & GICH_MISR_EOI) {
> +		u64 eisr = cpuif->vgic_eisr;
> +		unsigned long *eisr_bmap = u64_to_bitmask(&eisr);
> +		int lr;
> +
> +		for_each_set_bit(lr, eisr_bmap, vcpu->arch.vgic_cpu.nr_lr) {
> +			struct vgic_irq *irq;
> +			u32 intid = cpuif->vgic_lr[lr] & GICH_LR_VIRTUALID;
> +
> +			irq = vgic_get_irq(vcpu->kvm, vcpu, intid);
> +
> +			WARN_ON(irq->config == VGIC_CONFIG_EDGE);
> +			WARN_ON(cpuif->vgic_lr[lr] & GICH_LR_STATE);
> +
> +			kvm_notify_acked_irq(vcpu->kvm, 0,
> +					     intid - VGIC_NR_PRIVATE_IRQS);
> +
> +			cpuif->vgic_lr[lr] &= ~GICH_LR_STATE; /* Useful?? */
> +			cpuif->vgic_elrsr |= 1ULL << lr;
> +		}
> +	}
> +
> +	/* check and disable underflow maintenance IRQ */
> +	cpuif->vgic_hcr &= ~GICH_HCR_UIE;
> +
> +	/*
> +	 * In the next iterations of the vcpu loop, if we sync the
> +	 * vgic state after flushing it, but before entering the guest
> +	 * (this happens for pending signals and vmid rollovers), then
> +	 * make sure we don't pick up any old maintenance interrupts
> +	 * here.
> +	 */
> +	cpuif->vgic_eisr = 0;
> +}
> +
> +void vgic_v2_set_underflow(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_v2_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v2;
> +
> +	cpuif->vgic_hcr |= GICH_HCR_UIE;
> +}
> +
> +/*
> + * transfer the content of the LRs back into the corresponding ap_list:
> + * - active bit is transferred as is
> + * - pending bit is
> + *   - transferred as is in case of edge sensitive IRQs
> + *   - set to the line-level (resample time) for level sensitive IRQs
> + */
> +void vgic_v2_fold_lr_state(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_v2_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v2;
> +	int lr;
> +
> +	for (lr = 0; lr < vcpu->arch.vgic_cpu.used_lrs; lr++) {
> +		u32 val = cpuif->vgic_lr[lr];
> +		u32 intid = val & GICH_LR_VIRTUALID;
> +		struct vgic_irq *irq;
> +
> +		irq = vgic_get_irq(vcpu->kvm, vcpu, intid);
> +
> +		spin_lock(&irq->irq_lock);
> +
> +		/* Always preserve the active bit */
> +		irq->active = !!(val & GICH_LR_ACTIVE_BIT);
> +
> +		/* Edge is the only case where we preserve the pending bit */
> +		if (irq->config == VGIC_CONFIG_EDGE &&
> +		    (val & GICH_LR_PENDING_BIT)) {
> +			irq->pending = true;
> +
> +			if (intid < VGIC_NR_SGIS) {
> +				u32 cpuid = val & GICH_LR_PHYSID_CPUID;
> +
> +				cpuid >>= GICH_LR_PHYSID_CPUID_SHIFT;
> +				irq->source |= (1 << cpuid);
> +			}
> +		}
> +
> +		/* Clear soft pending state when level IRQs have been acked */
> +		if (irq->config == VGIC_CONFIG_LEVEL &&
> +		    !(val & GICH_LR_PENDING_BIT)) {
> +			irq->soft_pending = false;
> +			irq->pending = irq->line_level;
> +		}
> +
> +		spin_unlock(&irq->irq_lock);
> +	}
> +}
> +
> +/*
> + * Populates the particular LR with the state of a given IRQ:
> + * - for an edge sensitive IRQ the pending state is reset in the struct
> + * - for a level sensitive IRQ the pending state value is unchanged;
> + *   it will be resampled on deactivation
> + *
> + * If irq is not NULL, the irq_lock must be hold already by the caller.
> + * If irq is NULL, the respective LR gets cleared.
> + */
> +void vgic_v2_populate_lr(struct kvm_vcpu *vcpu, struct vgic_irq *irq, int lr)
> +{
> +	u32 val;
> +
> +	if (!irq) {
> +		val = 0;
> +		goto out;
> +	}
> +
> +	val = irq->intid;
> +
> +	if (irq->pending) {
> +		val |= GICH_LR_PENDING_BIT;
> +
> +		if (irq->config == VGIC_CONFIG_EDGE)
> +			irq->pending = false;
> +
> +		if (irq->intid < VGIC_NR_SGIS) {
> +			u32 src = ffs(irq->source);
> +
> +			BUG_ON(!src);
> +			val |= (src - 1) << GICH_LR_PHYSID_CPUID_SHIFT;
> +			irq->source &= ~(1 << (src - 1));
> +			if (irq->source)
> +				irq->pending = true;
> +		}
> +	}
> +
> +	if (irq->active)
> +		val |= GICH_LR_ACTIVE_BIT;
> +
> +	if (irq->hw) {
> +		val |= GICH_LR_HW;
> +		val |= irq->hwintid << GICH_LR_PHYSID_CPUID_SHIFT;
> +	} else {
> +		if (irq->config == VGIC_CONFIG_LEVEL)
> +			val |= GICH_LR_EOI;
> +	}

shouldn't we start writing the priority here (and in the GICv3 version)?

(which has the fun consequence of having to compare priorities against
the virtual priority filter in PATCH 11).

> +
> +out:
> +	vcpu->arch.vgic_cpu.vgic_v2.vgic_lr[lr] = val;
> +}
> +
>  void vgic_v2_irq_change_affinity(struct kvm *kvm, u32 intid, u8 new_targets)
>  {
>  	struct vgic_dist *dist = &kvm->arch.vgic;
> diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
> index 29c753e..90a85bf 100644
> --- a/virt/kvm/arm/vgic/vgic.c
> +++ b/virt/kvm/arm/vgic/vgic.c
> @@ -273,3 +273,207 @@ int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int intid,
>  	vgic_update_irq_pending(kvm, vcpu, intid, level);
>  	return 0;
>  }
> +
> +/**
> + * vgic_prune_ap_list - Remove non-relevant interrupts from the list
> + *
> + * @vcpu: The VCPU pointer
> + *
> + * Go over the list of "interesting" interrupts, and prune those that we
> + * won't have to consider in the near future.
> + */
> +static void vgic_prune_ap_list(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
> +	struct vgic_irq *irq, *tmp;
> +
> +retry:
> +	spin_lock(&vgic_cpu->ap_list_lock);
> +
> +	list_for_each_entry_safe(irq, tmp, &vgic_cpu->ap_list_head, ap_list) {
> +		struct kvm_vcpu *target_vcpu, *vcpuA, *vcpuB;
> +
> +		spin_lock(&irq->irq_lock);
> +
> +		BUG_ON(vcpu != irq->vcpu);
> +
> +		target_vcpu = vgic_target_oracle(irq);
> +
> +		if (!target_vcpu) {
> +			/*
> +			 * We don't need to process this interrupt any
> +			 * further, move it off the list.
> +			 */
> +			list_del_init(&irq->ap_list);
> +			irq->vcpu = NULL;
> +			spin_unlock(&irq->irq_lock);
> +			continue;
> +		}
> +
> +		if (target_vcpu == vcpu) {
> +			/* We're on the right CPU */
> +			spin_unlock(&irq->irq_lock);
> +			continue;
> +		}
> +
> +		/* This interrupt looks like it has to be migrated. */
> +
> +		spin_unlock(&irq->irq_lock);
> +		spin_unlock(&vgic_cpu->ap_list_lock);
> +
> +		/*
> +		 * Ensure locking order by always locking the smallest
> +		 * ID first.
> +		 */
> +		if (vcpu->vcpu_id < target_vcpu->vcpu_id) {
> +			vcpuA = vcpu;
> +			vcpuB = target_vcpu;
> +		} else {
> +			vcpuA = target_vcpu;
> +			vcpuB = vcpu;
> +		}
> +
> +		spin_lock(&vcpuA->arch.vgic_cpu.ap_list_lock);
> +		spin_lock(&vcpuB->arch.vgic_cpu.ap_list_lock);
> +		spin_lock(&irq->irq_lock);
> +
> +		/*
> +		 * If the affinity has been preserved, move the
> +		 * interrupt around. Otherwise, it means things have
> +		 * changed while the interrupt was unlocked, and we
> +		 * need to replay this.
> +		 *
> +		 * In all cases, we cannot trust the list not to have
> +		 * changed, so we restart from the beginning.
> +		 */
> +		if (target_vcpu == vgic_target_oracle(irq)) {
> +			struct vgic_cpu *new_cpu = &target_vcpu->arch.vgic_cpu;
> +
> +			list_del_init(&irq->ap_list);
> +			irq->vcpu = target_vcpu;
> +			list_add_tail(&irq->ap_list, &new_cpu->ap_list_head);
> +		}
> +
> +		spin_unlock(&irq->irq_lock);
> +		spin_unlock(&vcpuB->arch.vgic_cpu.ap_list_lock);
> +		spin_unlock(&vcpuA->arch.vgic_cpu.ap_list_lock);
> +		goto retry;
> +	}
> +
> +	spin_unlock(&vgic_cpu->ap_list_lock);
> +}
> +
> +static inline void vgic_process_maintenance_interrupt(struct kvm_vcpu *vcpu)
> +{
> +	if (kvm_vgic_global_state.type == VGIC_V2)
> +		vgic_v2_process_maintenance(vcpu);
> +	else
> +		WARN(1, "GICv3 Not Implemented\n");
> +}
> +
> +static inline void vgic_fold_lr_state(struct kvm_vcpu *vcpu)
> +{
> +	if (kvm_vgic_global_state.type == VGIC_V2)
> +		vgic_v2_fold_lr_state(vcpu);
> +	else
> +		WARN(1, "GICv3 Not Implemented\n");
> +}
> +
> +/*
> + * Requires the ap_lock to be held.
> + * If irq is not NULL, requires the IRQ lock to be held as well.
> + * If irq is NULL, the list register gets cleared.
> + */
> +static inline void vgic_populate_lr(struct kvm_vcpu *vcpu,
> +				    struct vgic_irq *irq, int lr)
> +{
> +	if (kvm_vgic_global_state.type == VGIC_V2)
> +		vgic_v2_populate_lr(vcpu, irq, lr);
> +	else
> +		WARN(1, "GICv3 Not Implemented\n");
> +}
> +
> +static inline void vgic_set_underflow(struct kvm_vcpu *vcpu)
> +{
> +	if (kvm_vgic_global_state.type == VGIC_V2)
> +		vgic_v2_set_underflow(vcpu);
> +	else
> +		WARN(1, "GICv3 Not Implemented\n");
> +}
> +
> +static int compute_ap_list_depth(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
> +	struct vgic_irq *irq;
> +	int count = 0;
> +
> +	list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
> +		spin_lock(&irq->irq_lock);
> +		/* GICv2 SGIs can count for more than one... */
> +		if (irq->intid < VGIC_NR_SGIS && irq->source)
> +			count += hweight8(irq->source);
> +		else
> +			count++;
> +		spin_unlock(&irq->irq_lock);
> +	}
> +	return count;
> +}
> +
> +/* requires the vcpu ap_lock to be held */
> +static void vgic_populate_lrs(struct kvm_vcpu *vcpu)
> +{
> +	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
> +	u32 model = vcpu->kvm->arch.vgic.vgic_model;
> +	struct vgic_irq *irq;
> +	int count = 0;
> +
> +	if (compute_ap_list_depth(vcpu) > vcpu->arch.vgic_cpu.nr_lr) {
> +		vgic_set_underflow(vcpu);
> +		vgic_sort_ap_list(vcpu);
> +	}
> +
> +	list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
> +		spin_lock(&irq->irq_lock);
> +
> +		if (unlikely(vgic_target_oracle(irq) != vcpu))
> +			goto next;
> +
> +		/*
> +		 * If we get an SGI with multiple sources, try to get
> +		 * them in all at once.
> +		 */
> +		if (model == KVM_DEV_TYPE_ARM_VGIC_V2 &&
> +		    irq->intid < VGIC_NR_SGIS) {
> +			while (irq->source && count < vcpu->arch.vgic_cpu.nr_lr)
> +				vgic_populate_lr(vcpu, irq, count++);
> +		} else {
> +			vgic_populate_lr(vcpu, irq, count++);
> +		}
> +
> +next:
> +		spin_unlock(&irq->irq_lock);
> +
> +		if (count == vcpu->arch.vgic_cpu.nr_lr)
> +			break;
> +	}
> +
> +	vcpu->arch.vgic_cpu.used_lrs = count;
> +
> +	/* Nuke remaining LRs */
> +	for ( ; count < vcpu->arch.vgic_cpu.nr_lr; count++)
> +		vgic_populate_lr(vcpu, NULL, count);
> +}
> +
> +void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
> +{
> +	vgic_process_maintenance_interrupt(vcpu);
> +	vgic_fold_lr_state(vcpu);
> +	vgic_prune_ap_list(vcpu);
> +}
> +
> +void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
> +{
> +	spin_lock(&vcpu->arch.vgic_cpu.ap_list_lock);
> +	vgic_populate_lrs(vcpu);
> +	spin_unlock(&vcpu->arch.vgic_cpu.ap_list_lock);
> +}
> diff --git a/virt/kvm/arm/vgic/vgic.h b/virt/kvm/arm/vgic/vgic.h
> index b2faf00..95ef3cf 100644
> --- a/virt/kvm/arm/vgic/vgic.h
> +++ b/virt/kvm/arm/vgic/vgic.h
> @@ -21,5 +21,9 @@ struct vgic_irq *vgic_get_irq(struct kvm *kvm, struct kvm_vcpu *vcpu,
>  bool vgic_queue_irq(struct kvm *kvm, struct vgic_irq *irq);
>  
>  void vgic_v2_irq_change_affinity(struct kvm *kvm, u32 intid, u8 target);
> +void vgic_v2_process_maintenance(struct kvm_vcpu *vcpu);
> +void vgic_v2_fold_lr_state(struct kvm_vcpu *vcpu);
> +void vgic_v2_populate_lr(struct kvm_vcpu *vcpu, struct vgic_irq *irq, int lr);
> +void vgic_v2_set_underflow(struct kvm_vcpu *vcpu);
>  
>  #endif
> -- 
> 2.7.3
>