KVM: arm/arm64: Allow to use KVM without in-kernel irqchip

This patch set brings back functionality which was broken in v4.0. Unfortunately because of restrictions of such a hardware is is impossible to take advantage of virtual architected timer, therefore guest, running in such restricted mode, has to use some memory-mapped timer. But it is still better than nothing.

v2 => v3:
- Improved commit messages, added references to commits where the respective functionality was broken
- Explicitly specify that the solution currently affects only vGIC and has nothing to do with timer.
- Fixed code style according to previous notes
- Removed ARM64 save/restore patch introduced in v2 because it was already obsolete for linux-next
- Modify KVM_CAP_IRQFD handling in correct place

v1 => v2:
- Do not use defensive approach in patch 0001. Use correct conditions in callers instead
- Added ARM64-specific code, without which attempt to run a VM ends in a HYP crash because of unset vGIC save/restore function pointers

Pavel Fedin (3):
Fix NULL pointer dereference if KVM is used without in-kernel irqchip Detect vGIC presence at runtime
Make KVM_CAP_IRQFD dependent on KVM_CAP_IRQCHIP

 arch/arm/kvm/arm.c  | 19 ++++++++++++++++---
 virt/kvm/kvm_main.c |  5 +++--
 2 files changed, 19 insertions(+), 5 deletions(-)

2.4.4

Makes qemu working again with kernel-irqchip=off option, allowing to use GIC emulation in userspace.

Previously kvm_vgic_map_resources() used to include irqchip_in_kernel() check, and vgic_v2_map_resources() still has it, but now vm_ops are not initialized before kvm_vgic_create(). Therefore kvm_vgic_map_resources() call without vGIC instantiation attempts to call vm_ops.map_resources == NULL

The bug was introduced in b26e5fdac43c1b7c394502917e42e3b91f3aa1a3

---
 arch/arm/kvm/arm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 483b5d1..199a50a 100644
@@ -449,7 +449,7 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 	 * Map the VGIC hardware resources before running a vcpu the first
 	 * time on this VM.
 	 */
-	if (unlikely(!vgic_ready(kvm))) {
+	if (unlikely(irqchip_in_kernel(kvm) && !vgic_ready(kvm))) {
 		ret = kvm_vgic_map_resources(kvm);
 		if (ret)
 			return ret;

2.4.4

Before commit 662d9715840aef44dcb573b0f9fab9e8319c868a is was possible to compile the kernel without vGIC and vTimer support. Commit message says about possibility to detect vGIC support in runtine, but this has never been implemented.

This patch introdices runtime check, restoring the lost functionality. It again allows to use KVM on hardware without vGIC. Interrupt controller has to be emulated in userspace in this case.

-ENODEV return code from probe function means there's no GIC at all. -ENXIO happens when, for example, there is GIC node in the device tree, but it does not specify vGIC resources. Normally this means that vGIC hardware is defunct. Any other error code is still treated as full stop because it might mean some really serious problems.

This patch does not touch any virtual timer code, suggesting that timer hardware is actually in place. Normally on boards in question it is true, however since vGIC is missing, it is impossible to correctly utilize interrupts from the virtual timer. Since virtual timer handling is in active redevelopment now, handling in it userspace is out of scope at the moment. The guest is currently suggested to use some memory-mapped timer which can be emulated in userspace.

---
 arch/arm/kvm/arm.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

index 199a50a..1039161 100644
@@ -61,6 +61,8 @@ static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
 static u8 kvm_next_vmid;
 static DEFINE_SPINLOCK(kvm_vmid_lock);
 
+static bool vgic_present;
+
 static void kvm_arm_set_running_vcpu(struct kvm_vcpu *vcpu)
 {
 	BUG_ON(preemptible());

@@ -131,7 +133,8 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 	kvm->arch.vmid_gen = 0;
 
 	/* The maximum number of VCPUs is limited by the host's GIC model */
-	kvm->arch.max_vcpus = kvm_vgic_get_max_vcpus();
+	kvm->arch.max_vcpus = vgic_present ?
+				kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 
 	return ret;
 out_free_stage2_pgd:

@@ -171,6 +174,8 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 	int r;
 	switch (ext) {
 	case KVM_CAP_IRQCHIP:
+		r = vgic_present;
+		break;
 	case KVM_CAP_IOEVENTFD:
 	case KVM_CAP_DEVICE_CTRL:
 	case KVM_CAP_USER_MEMORY:

@@ -849,6 +854,8 @@ static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
 
 	switch (dev_id) {
 	case KVM_ARM_DEVICE_VGIC_V2:
+		if (!vgic_present)
+			return -ENXIO;
 		return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
 	default:
 		return -ENODEV;

@@ -863,6 +870,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 
 	switch (ioctl) {
 	case KVM_CREATE_IRQCHIP: {
+		if (!vgic_present)
+			return -ENXIO;
 		return kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
 	}
 	case KVM_ARM_SET_DEVICE_ADDR: {

@@ -1045,8 +1054,12 @@ static int init_hyp_mode(void)
 	 * Init HYP view of VGIC
 	 */
 	err = kvm_vgic_hyp_init();
-	if (err)
+	if (err == -ENODEV || err == -ENXIO)
+		vgic_present = false;
+	else if (err)
 		goto out_free_context;
+	else
+		vgic_present = true;
 
 	/*
 	 * Init HYP architected timer support

2.4.4

And that's a problem, see below.

	M.

Jazz is not dead. It just smells funny...

Hello! Thank you for quick response.

Christoffer disliked it in v1, so i thought a bit and changed it. Ok, will change it back.

But, this seems to be already done, isn't it?
According to http://lxr.free-electrons.com/source/arch/arm/kvm/arm.c#L439:
--- cut ---
459 /*

460          * Enable the arch timers only if we have an in-kernel VGIC
461          * and it has been properly initialized, since we cannot handle
462          * interrupts from the virtual timer with a userspace gic.
463          */

464 if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
465 kvm_timer_enable(kvm);
--- cut ---

Without kvm->arch.timer.enabled set to 1 by kvm_timer_enable() VM context save/restore code will
not actually touch timer registers. Therefore the host part of the code will not do anything.
As to guest itself, only userspace can stop it from accessing timer registers. My experimental qemu
does this by removing generic timer node from guest's device tree. Virtual timer access simply cannot be trapped, otherwise there would be no problem at all. But, OK, even if the guest programs timer, we will just see "Unexpected IRQ 27" on the console, and the guest will not work, so it's not terribly fatal.

You know, i actually looked at it before posting v3. I tried to omit kvm_timer_hyp_init() call too,
and got lots of crashes because:
1. kvm_timer_init() is called unconditionally
2. qemu does some initialization of timer registers unconditionally using ioctl, and they end up in kvm_arm_timer_set_reg()
Both of these points end up in kvm_phys_timer_read() which dereferences timecounter == NULL. Well, i could make kvm_phys_timer_read() just returning 0 in this case, but this could mis-trigger
kvm_timer_should_fire() in some circumstances. I would have to patch it too... At this point i decided to stop because the result perhaps does not worth the effort and amount of patching.

While writing this message i was walking through this code once again, and... I have a suggestion.
Actually, if we are really paranoid, we could be afraid of kvm_vgic_inject_irq() being called, which would do some weird things without vGIC. It is possible to add a check for kvm->arch.timer.enabled in kvm_timer_sync_hwstate() and kvm_timer_flush_hwstate(). If the timer is disabled those functions will simply return doing nothing. This would guarantee that interrupt injection is never attempted.

What do you think?

And some more. Actually, it is possible to emulate generic timer in userspace, just not the virtual
one. IIRC access to physical timer can be trapped. So, if we modify guest's device tree by removing virtual timer IRQ, the guest will fall back to physical timer. And this will be caught by the hypervisor. After this all we have to do is to add corresponding exit code which would allow the userspace to emulate missing CP15 (or system in case of ARM64) registers. So, this timer issue is not grave, just i postpone implementing it until GIC issues are settled down.

Kind regards,
Pavel Fedin
Expert Engineer
Samsung Electronics Research center Russia

Right, I failed to remember that one. Sorry. It should be safe then. Hopefully.

[...]

This is completely Linux-specific, unfortunately. And it relies on userpace to expose a modified DT, so you need to be able to report back to userspace that you can't deal with the virtual timer.

Which brings me to the next point: how do you tell userspace that your timers are non-functional?

Thanks,

	M.

Jazz is not dead. It just smells funny...

Hello!

Yes. But better than nothing.

Easy. If KVM_CAP_IRQCHIP == 0, then we apparently don't have vGIC, and since we know that vGIC and
vTimer are paired, we know that there is no vTimer too.

Kind regards,
Pavel Fedin
Expert Engineer
Samsung Electronics Research center Russia

Now at least ARM is able to determine whether the machine has virtualization support for irqchip or not at runtime. Obviously, irqfd requires irqchip.

---
This has to be verified with PowerPC and S390 maintainers. I've got an impression that KVM_CAP_IRQCHIP is forgotten on these architectures (their KConfig selects HAVE_KVM_IRQCHIP but KVM_CAP_IRQCHIP is not mentioned anywhere in the respective code).

---
 virt/kvm/kvm_main.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 9097741..7b49618 100644
@@ -2478,12 +2478,13 @@ static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
 #ifdef CONFIG_HAVE_KVM_MSI
 	case KVM_CAP_SIGNAL_MSI:
 #endif
+	case KVM_CAP_CHECK_EXTENSION_VM:
+		return 1;
 #ifdef CONFIG_HAVE_KVM_IRQFD
 	case KVM_CAP_IRQFD:
 	case KVM_CAP_IRQFD_RESAMPLE:
+		return kvm_vm_ioctl_check_extension(kvm, KVM_CAP_IRQCHIP);
 #endif
-	case KVM_CAP_CHECK_EXTENSION_VM:
-		return 1;
 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 	case KVM_CAP_IRQ_ROUTING:
 		return KVM_MAX_IRQ_ROUTES;

2.4.4