First, something that is not clear to me: how do you deal with a guest writing 1 to VSIP.SSIP? I think that could lead to lost interrupts if you have the following sequence 1) guest writes 1 to VSIP.SSIP 2) guest leaves VS-mode 3) host syncs VSIP 4) user mode triggers interrupt 5) host reenters guest 6) host moves irqs_pending to VSIP and clears VSIP.SSIP in the process Perhaps irqs_pending needs to be split in two fields, irqs_pending and irqs_pending_mask, and then you can do this: /* * irqs_pending and irqs_pending_mask have multiple-producer/single- * consumer semantics; therefore bits can be set in the mask without * a lock, but clearing the bits requires vcpu_lock. Furthermore, * consumers should never write to irqs_pending, and should not * use bits of irqs_pending that weren't 1 in the mask. */ int kvm_riscv_vcpu_set_interrupt(struct kvm_vcpu *vcpu, unsigned int irq) { ... set_bit(irq, &vcpu->arch.irqs_pending); smp_mb__before_atomic(); set_bit(irq, &vcpu->arch.irqs_pending_mask); kvm_vcpu_kick(vcpu); } int kvm_riscv_vcpu_unset_interrupt(struct kvm_vcpu *vcpu, unsigned int irq) { ... clear_bit(irq, &vcpu->arch.irqs_pending); smp_mb__before_atomic(); set_bit(irq, &vcpu->arch.irqs_pending_mask); } static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu) { ... WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0); } and kvm_riscv_vcpu_flush_interrupts can leave aside VSIP bits that aren't in vcpu->arch.irqs_pending_mask: if (atomic_read(&vcpu->arch.irqs_pending_mask)) { u32 mask, val; mask = xchg_acquire(&vcpu->arch.irqs_pending_mask, 0); val = READ_ONCE(vcpu->arch.irqs_pending) & mask; vcpu->arch.guest_csr.vsip &= ~mask; vcpu->arch.guest_csr.vsip |= val; csr_write(CSR_VSIP, vsip); } Also, the getter of CSR_VSIP should call kvm_riscv_vcpu_flush_interrupts, while the setter should clear irqs_pending_mask. On 29/07/19 13:56, Anup Patel wrote: > + kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); > + kvm_vcpu_kick(vcpu); The request is not needed as long as kvm_riscv_vcpu_flush_interrupts is called *after* smp_store_mb(vcpu->mode, IN_GUEST_MODE) in kvm_arch_vcpu_ioctl_run. This is the "request-less vCPU kick" pattern in Documentation/virtual/kvm/vcpu-requests.rst. The smp_store_mb then orders the write of IN_GUEST_MODE before the read of irqs_pending (or irqs_pending_mask in my proposal above); in the producers, there is a dual memory barrier in kvm_vcpu_exiting_guest_mode(), ordering the write of irqs_pending(_mask) before the read of vcpu->mode. Similar to other VS* CSRs, I'd rather have a ONE_REG interface for VSIE and VSIP from the beginning as well. Note that the VSIP setter would clear irqs_pending_mask, while the getter would call kvm_riscv_vcpu_flush_interrupts before reading. It's up to userspace to ensure that no interrupt injections happen between the calls to the getter and the setter. Paolo > + csr_write(CSR_VSIP, vcpu->arch.irqs_pending); > + vcpu->arch.guest_csr.vsip = vcpu->arch.irqs_pending; > + }