Re: [PATCH v2 08/29] LoongArch: KVM: Implement vcpu handle exit interface

[Paolo Bonzini <pbonzini@redhat.com>]

On 2/20/23 07:57, Tianrui Zhao wrote:
> + * Return value is in the form (errcode<<2 | RESUME_FLAG_HOST | RESUME_FLAG_NV)

As far as I can see, RESUME_FLAG_NV does not exist anymore and this is 
just copied from arch/mips?

You can keep RESUME_HOST/RESUME_GUEST for the individual functions, but 
here please make it just "1" for resume guest, and "<= 0" for resume 
host.  This is easy enough to check from assembly and removes the srai by 2.

> +static int _kvm_handle_exit(struct kvm_run *run, struct kvm_vcpu *vcpu)
> +{
> +	unsigned long exst = vcpu->arch.host_estat;
> +	u32 intr = exst & 0x1fff; /* ignore NMI */
> +	u32 exccode = (exst & CSR_ESTAT_EXC) >> CSR_ESTAT_EXC_SHIFT;
> +	u32 __user *opc = (u32 __user *) vcpu->arch.pc;
> +	int ret = RESUME_GUEST, cpu;
> +
> +	vcpu->mode = OUTSIDE_GUEST_MODE;
> +
> +	/* Set a default exit reason */
> +	run->exit_reason = KVM_EXIT_UNKNOWN;
> +	run->ready_for_interrupt_injection = 1;
> +
> +	/*
> +	 * Set the appropriate status bits based on host CPU features,
> +	 * before we hit the scheduler
> +	 */

Stale comment?

> +	local_irq_enable();

Please add guest_state_exit_irqoff() here.

> +	kvm_debug("%s: exst: %lx, PC: %p, kvm_run: %p, kvm_vcpu: %p\n",
> +			__func__, exst, opc, run, vcpu);

Please add the information to the kvm_exit tracepoint (thus also 
removing variables such as "exst" or "opc" from this function) instead 
of calling kvm_debug().

> +	trace_kvm_exit(vcpu, exccode);
> +	if (exccode) {
> +		ret = _kvm_handle_fault(vcpu, exccode);
> +	} else {
> +		WARN(!intr, "suspicious vm exiting");
> +		++vcpu->stat.int_exits;
> +
> +		if (need_resched())
> +			cond_resched();

This "if" is not necessary because there is already a cond_resched() below.

> +		ret = RESUME_GUEST;

This "ret" is not necessary because "ret" is already initialized to 
RESUME_GUEST above, you can either remove it or remove the initializer.

> +	}
> +
> +	cond_resched();
> +	local_irq_disable();

At this point, ret is either RESUME_GUEST or RESUME_HOST.  So, the "if"s 
below are either all taken or all not taken, and most of this code:

	kvm_acquire_timer(vcpu);
	_kvm_deliver_intr(vcpu);

	if (signal_pending(current)) {
		run->exit_reason = KVM_EXIT_INTR;
		ret = (-EINTR << 2) | RESUME_HOST;
		++vcpu->stat.signal_exits;
		// no need for a tracepoint here
		// trace_kvm_exit(vcpu, KVM_TRACE_EXIT_SIGNAL);
	}

	trace_kvm_reenter(vcpu);

	/*
	 * Make sure the read of VCPU requests in vcpu_reenter()
	 * callback is not reordered ahead of the write to vcpu->mode,
	 * or we could miss a TLB flush request while the requester sees
	 * the VCPU as outside of guest mode and not needing an IPI.
	 */
	smp_store_mb(vcpu->mode, IN_GUEST_MODE);

	cpu = smp_processor_id();
	_kvm_check_requests(vcpu, cpu);
	_kvm_check_vmid(vcpu, cpu);
	vcpu->arch.host_eentry = csr_read64(LOONGARCH_CSR_EENTRY);

	/*
	 * If FPU are enabled (i.e. the guest's FPU context
	 * is live), restore FCSR0.
	 */
	if (_kvm_guest_has_fpu(&vcpu->arch) &&
		read_csr_euen() & (CSR_EUEN_FPEN)) {
		kvm_restore_fcsr(&vcpu->arch.fpu);
	}

(all except for the "if (signal_pending(current))" and the final "if") 
is pretty much duplicated with kvm_arch_vcpu_ioctl_run(); the remaining 
code can also be done from kvm_arch_vcpu_ioctl_run(), the cost is small. 
  Please move it to a separate function, for example:

int kvm_pre_enter_guest(struct kvm_vcpu *vcpu)
{
	if (signal_pending(current)) {
		run->exit_reason = KVM_EXIT_INTR;
		++vcpu->stat.signal_exits;
		return -EINTR;
	}

	kvm_acquire_timer(vcpu);
	_kvm_deliver_intr(vcpu);

	...

	if (_kvm_guest_has_fpu(&vcpu->arch) &&
		read_csr_euen() & (CSR_EUEN_FPEN)) {
		kvm_restore_fcsr(&vcpu->arch.fpu);
	}
	return 1;
}

Call it from _kvm_handle_exit():

	if (ret == RESUME_HOST)
		return 0;

	r = kvm_pre_enter_guest(vcpu);
	if (r > 0) {
		trace_kvm_reenter(vcpu);
		guest_state_enter_irqoff();
	}

	return r;

and from kvm_arch_vcpu_ioctl_run():

	local_irq_disable();
	guest_timing_enter_irqoff();
	r = kvm_pre_enter_guest(vcpu);
	if (r > 0) {
		trace_kvm_enter(vcpu);
		/*
		 * This should actually not be a function pointer, but
		 * just for clarity */
		 */
		guest_state_enter_irqoff();
		r = vcpu->arch.vcpu_run(run, vcpu);
		/* guest_state_exit_irqoff() already done.  */
		trace_kvm_out(vcpu);
	}
	guest_timing_exit_irqoff();
	local_irq_enable();

out:
	kvm_sigset_deactivate(vcpu);

	vcpu_put(vcpu);
	return r;

Paolo

> +	if (ret == RESUME_GUEST)
> +		kvm_acquire_timer(vcpu);
> +
> +	if (!(ret & RESUME_HOST)) {
> +		_kvm_deliver_intr(vcpu);
> +		/* Only check for signals if not already exiting to userspace */
> +		if (signal_pending(current)) {
> +			run->exit_reason = KVM_EXIT_INTR;
> +			ret = (-EINTR << 2) | RESUME_HOST;
> +			++vcpu->stat.signal_exits;
> +			trace_kvm_exit(vcpu, KVM_TRACE_EXIT_SIGNAL);
> +		}
> +	}
> +
> +	if (ret == RESUME_GUEST) {
> +		trace_kvm_reenter(vcpu);
> +
> +		/*
> +		 * Make sure the read of VCPU requests in vcpu_reenter()
> +		 * callback is not reordered ahead of the write to vcpu->mode,
> +		 * or we could miss a TLB flush request while the requester sees
> +		 * the VCPU as outside of guest mode and not needing an IPI.
> +		 */
> +		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
> +
> +		cpu = smp_processor_id();
> +		_kvm_check_requests(vcpu, cpu);
> +		_kvm_check_vmid(vcpu, cpu);
> +		vcpu->arch.host_eentry = csr_read64(LOONGARCH_CSR_EENTRY);
> +
> +		/*
> +		 * If FPU are enabled (i.e. the guest's FPU context
> +		 * is live), restore FCSR0.
> +		 */
> +		if (_kvm_guest_has_fpu(&vcpu->arch) &&
> +			read_csr_euen() & (CSR_EUEN_FPEN)) {
> +			kvm_restore_fcsr(&vcpu->arch.fpu);
> +		}
> +	}
> +
> +	return ret;
> +}
> +
>   int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
>   {
>   	int i;