[PATCH v2 2/2] RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization

Anup Patel apatel at ventanamicro.com
Sun Jul 6 20:53:44 PDT 2025


Currently, the common AIA functions kvm_riscv_vcpu_aia_has_interrupts()
and kvm_riscv_aia_wakeon_hgei() lookup HGEI line using an array of VCPU
pointers before accessing HGEI[E|P] CSR which is slow and prone to race
conditions because there is a separate per-hart lock for the VCPU pointer
array and a separate per-VCPU rwlock for IMSIC VS-file (including HGEI
line) used by the VCPU. Due to these race conditions, it is observed
on QEMU RISC-V host that Guest VCPUs sleep in WFI and never wakeup even
with interrupt pending in the IMSIC VS-file because VCPUs were waiting
for HGEI wakeup on the wrong host CPU.

The IMSIC virtualization already keeps track of the HGEI line and the
associated IMSIC VS-file used by each VCPU so move the HGEI[E|P] CSR
access to IMSIC virtualization so that costly HGEI line lookup can be
avoided and likelihood of race-conditions when updating HGEI[E|P] CSR
is also reduced.

Reviewed-by: Atish Patra <atishp at rivosinc.com>
Tested-by: Atish Patra <atishp at rivosinc.com>
Tested-by: Heinrich Schuchardt <heinrich.schuchardt at canonical.com>
Fixes: 3385339296d1 ("RISC-V: KVM: Use IMSIC guest files when available")
Signed-off-by: Anup Patel <apatel at ventanamicro.com>
---
 arch/riscv/include/asm/kvm_aia.h |  4 ++-
 arch/riscv/kvm/aia.c             | 51 +++++---------------------------
 arch/riscv/kvm/aia_imsic.c       | 45 ++++++++++++++++++++++++++++
 arch/riscv/kvm/vcpu.c            |  2 --
 4 files changed, 55 insertions(+), 47 deletions(-)

diff --git a/arch/riscv/include/asm/kvm_aia.h b/arch/riscv/include/asm/kvm_aia.h
index 0a0f12496f00..b04ecdd1a860 100644
--- a/arch/riscv/include/asm/kvm_aia.h
+++ b/arch/riscv/include/asm/kvm_aia.h
@@ -87,6 +87,9 @@ DECLARE_STATIC_KEY_FALSE(kvm_riscv_aia_available);
 
 extern struct kvm_device_ops kvm_riscv_aia_device_ops;
 
+bool kvm_riscv_vcpu_aia_imsic_has_interrupt(struct kvm_vcpu *vcpu);
+void kvm_riscv_vcpu_aia_imsic_load(struct kvm_vcpu *vcpu, int cpu);
+void kvm_riscv_vcpu_aia_imsic_put(struct kvm_vcpu *vcpu);
 void kvm_riscv_vcpu_aia_imsic_release(struct kvm_vcpu *vcpu);
 int kvm_riscv_vcpu_aia_imsic_update(struct kvm_vcpu *vcpu);
 
@@ -161,7 +164,6 @@ void kvm_riscv_aia_destroy_vm(struct kvm *kvm);
 int kvm_riscv_aia_alloc_hgei(int cpu, struct kvm_vcpu *owner,
 			     void __iomem **hgei_va, phys_addr_t *hgei_pa);
 void kvm_riscv_aia_free_hgei(int cpu, int hgei);
-void kvm_riscv_aia_wakeon_hgei(struct kvm_vcpu *owner, bool enable);
 
 void kvm_riscv_aia_enable(void);
 void kvm_riscv_aia_disable(void);
diff --git a/arch/riscv/kvm/aia.c b/arch/riscv/kvm/aia.c
index 19afd1f23537..dad318185660 100644
--- a/arch/riscv/kvm/aia.c
+++ b/arch/riscv/kvm/aia.c
@@ -30,28 +30,6 @@ unsigned int kvm_riscv_aia_nr_hgei;
 unsigned int kvm_riscv_aia_max_ids;
 DEFINE_STATIC_KEY_FALSE(kvm_riscv_aia_available);
 
-static int aia_find_hgei(struct kvm_vcpu *owner)
-{
-	int i, hgei;
-	unsigned long flags;
-	struct aia_hgei_control *hgctrl = get_cpu_ptr(&aia_hgei);
-
-	raw_spin_lock_irqsave(&hgctrl->lock, flags);
-
-	hgei = -1;
-	for (i = 1; i <= kvm_riscv_aia_nr_hgei; i++) {
-		if (hgctrl->owners[i] == owner) {
-			hgei = i;
-			break;
-		}
-	}
-
-	raw_spin_unlock_irqrestore(&hgctrl->lock, flags);
-
-	put_cpu_ptr(&aia_hgei);
-	return hgei;
-}
-
 static inline unsigned long aia_hvictl_value(bool ext_irq_pending)
 {
 	unsigned long hvictl;
@@ -95,7 +73,6 @@ void kvm_riscv_vcpu_aia_sync_interrupts(struct kvm_vcpu *vcpu)
 
 bool kvm_riscv_vcpu_aia_has_interrupts(struct kvm_vcpu *vcpu, u64 mask)
 {
-	int hgei;
 	unsigned long seip;
 
 	if (!kvm_riscv_aia_available())
@@ -114,11 +91,7 @@ bool kvm_riscv_vcpu_aia_has_interrupts(struct kvm_vcpu *vcpu, u64 mask)
 	if (!kvm_riscv_aia_initialized(vcpu->kvm) || !seip)
 		return false;
 
-	hgei = aia_find_hgei(vcpu);
-	if (hgei > 0)
-		return !!(ncsr_read(CSR_HGEIP) & BIT(hgei));
-
-	return false;
+	return kvm_riscv_vcpu_aia_imsic_has_interrupt(vcpu);
 }
 
 void kvm_riscv_vcpu_aia_update_hvip(struct kvm_vcpu *vcpu)
@@ -164,6 +137,9 @@ void kvm_riscv_vcpu_aia_load(struct kvm_vcpu *vcpu, int cpu)
 		csr_write(CSR_HVIPRIO2H, csr->hviprio2h);
 #endif
 	}
+
+	if (kvm_riscv_aia_initialized(vcpu->kvm))
+		kvm_riscv_vcpu_aia_imsic_load(vcpu, cpu);
 }
 
 void kvm_riscv_vcpu_aia_put(struct kvm_vcpu *vcpu)
@@ -174,6 +150,9 @@ void kvm_riscv_vcpu_aia_put(struct kvm_vcpu *vcpu)
 	if (!kvm_riscv_aia_available())
 		return;
 
+	if (kvm_riscv_aia_initialized(vcpu->kvm))
+		kvm_riscv_vcpu_aia_imsic_put(vcpu);
+
 	if (kvm_riscv_nacl_available()) {
 		nsh = nacl_shmem();
 		csr->vsiselect = nacl_csr_read(nsh, CSR_VSISELECT);
@@ -472,22 +451,6 @@ void kvm_riscv_aia_free_hgei(int cpu, int hgei)
 	raw_spin_unlock_irqrestore(&hgctrl->lock, flags);
 }
 
-void kvm_riscv_aia_wakeon_hgei(struct kvm_vcpu *owner, bool enable)
-{
-	int hgei;
-
-	if (!kvm_riscv_aia_available())
-		return;
-
-	hgei = aia_find_hgei(owner);
-	if (hgei > 0) {
-		if (enable)
-			csr_set(CSR_HGEIE, BIT(hgei));
-		else
-			csr_clear(CSR_HGEIE, BIT(hgei));
-	}
-}
-
 static irqreturn_t hgei_interrupt(int irq, void *dev_id)
 {
 	int i;
diff --git a/arch/riscv/kvm/aia_imsic.c b/arch/riscv/kvm/aia_imsic.c
index ea1a36836d9c..fda0346f0ea1 100644
--- a/arch/riscv/kvm/aia_imsic.c
+++ b/arch/riscv/kvm/aia_imsic.c
@@ -677,6 +677,48 @@ static void imsic_swfile_update(struct kvm_vcpu *vcpu,
 	imsic_swfile_extirq_update(vcpu);
 }
 
+bool kvm_riscv_vcpu_aia_imsic_has_interrupt(struct kvm_vcpu *vcpu)
+{
+	struct imsic *imsic = vcpu->arch.aia_context.imsic_state;
+	unsigned long flags;
+	bool ret = false;
+
+	/*
+	 * The IMSIC SW-file directly injects interrupt via hvip so
+	 * only check for interrupt when IMSIC VS-file is being used.
+	 */
+
+	read_lock_irqsave(&imsic->vsfile_lock, flags);
+	if (imsic->vsfile_cpu > -1)
+		ret = !!(csr_read(CSR_HGEIP) & BIT(imsic->vsfile_hgei));
+	read_unlock_irqrestore(&imsic->vsfile_lock, flags);
+
+	return ret;
+}
+
+void kvm_riscv_vcpu_aia_imsic_load(struct kvm_vcpu *vcpu, int cpu)
+{
+	/*
+	 * No need to explicitly clear HGEIE CSR bits because the
+	 * hgei interrupt handler (aka hgei_interrupt()) will always
+	 * clear it for us.
+	 */
+}
+
+void kvm_riscv_vcpu_aia_imsic_put(struct kvm_vcpu *vcpu)
+{
+	struct imsic *imsic = vcpu->arch.aia_context.imsic_state;
+	unsigned long flags;
+
+	if (!kvm_vcpu_is_blocking(vcpu))
+		return;
+
+	read_lock_irqsave(&imsic->vsfile_lock, flags);
+	if (imsic->vsfile_cpu > -1)
+		csr_set(CSR_HGEIE, BIT(imsic->vsfile_hgei));
+	read_unlock_irqrestore(&imsic->vsfile_lock, flags);
+}
+
 void kvm_riscv_vcpu_aia_imsic_release(struct kvm_vcpu *vcpu)
 {
 	unsigned long flags;
@@ -781,6 +823,9 @@ int kvm_riscv_vcpu_aia_imsic_update(struct kvm_vcpu *vcpu)
 	 * producers to the new IMSIC VS-file.
 	 */
 
+	/* Ensure HGEIE CSR bit is zero before using the new IMSIC VS-file */
+	csr_clear(CSR_HGEIE, BIT(new_vsfile_hgei));
+
 	/* Zero-out new IMSIC VS-file */
 	imsic_vsfile_local_clear(new_vsfile_hgei, imsic->nr_hw_eix);
 
diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
index fe028b4274df..b26bf35a0a19 100644
--- a/arch/riscv/kvm/vcpu.c
+++ b/arch/riscv/kvm/vcpu.c
@@ -211,12 +211,10 @@ int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 
 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 {
-	kvm_riscv_aia_wakeon_hgei(vcpu, true);
 }
 
 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 {
-	kvm_riscv_aia_wakeon_hgei(vcpu, false);
 }
 
 int kvm_arch_vcpu_runnable(struct kvm_vcpu *vcpu)
-- 
2.43.0




More information about the kvm-riscv mailing list