Skip to contentSkip to conversation

KVM: arm64: vgic: Don't hold config_lock while unregistering redistributors

PATCH Applied Marc Zyngier 3 participants 3 messages Aug 19, 2024

Subsystems (3)
lore
Share & tools
lore

Marc Zyngier

author maintainer
#0 lore raw

PATCH 1/1 KVM: arm64: vgic: Don't hold config_lock while unregistering redistributors

We recently moved the teardown of the vgic part of a vcpu inside a critical section guarded by the config_lock. This teardown phase involves calling into kvm_io_bus_unregister_dev(), which takes the kvm->srcu lock.

However, this violates the established order where kvm->srcu is taken on a memory fault (such as an MMIO access), possibly followed by taking the config_lock if the GIC emulation requires mutual exclusion from the other vcpus.

It therefore results in a bad lockdep splat, as reported by Zenghui.

Fix this by moving the call to kvm_io_bus_unregister_dev() outside of the config_lock critical section. At this stage, there shouln't be any need to hold the config_lock.

As an additional bonus, document the ordering between kvm->slots_lock, kvm->srcu and kvm->arch.config_lock so that I cannot pretend I didn't know about those anymore.

Fixes 9eb18136af9f ("KVM: arm64: vgic: Hold config_lock while tearing down a CPU interface")
---
 arch/arm64/kvm/vgic/vgic-init.c | 9 ++++++---
 arch/arm64/kvm/vgic/vgic.c      | 5 +++++
 2 files changed, 11 insertions(+), 3 deletions(-)
arch/arm64/kvm/vgic/vgic-init.c +6 -3
index 41feb858ff9a..e7c53e8af3d1 100644
@@ -417,10 +417,8 @@ static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
 	kfree(vgic_cpu->private_irqs);
 	vgic_cpu->private_irqs = NULL;
 
-	if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
-		vgic_unregister_redist_iodev(vcpu);
+	if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
 		vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
-	}
 }
 
 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
@@ -448,6 +446,11 @@ void kvm_vgic_destroy(struct kvm *kvm)
 	kvm_vgic_dist_destroy(kvm);
 
 	mutex_unlock(&kvm->arch.config_lock);
+
+	if (kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
+		kvm_for_each_vcpu(i, vcpu, kvm)
+			vgic_unregister_redist_iodev(vcpu);
+
 	mutex_unlock(&kvm->slots_lock);
 }
 
arch/arm64/kvm/vgic/vgic.c +5 -0
index 2caa64415ff3..f50274fd5581 100644
@@ -36,6 +36,11 @@ struct vgic_global kvm_vgic_global_state __ro_after_init = {
  * we have to disable IRQs before taking this lock and everything lower
  * than it.
  *
+ * The config_lock has additional ordering requirements:
+ * kvm->slots_lock
+ *   kvm->srcu
+ *     kvm->arch.config_lock
+ *
  * If you need to take multiple locks, always take the upper lock first,
  * then the lower ones, e.g. first take the its_lock, then the irq_lock.
  * If you are already holding a lock and need to take a higher one, you
Signature · 1 line
2.39.2
reviewer Reviewed-by Tested-by on v1 1/1
#1 lore raw
Quoted 16 lines · Marc Zyngier

We recently moved the teardown of the vgic part of a vcpu inside a critical section guarded by the config_lock. This teardown phase involves calling into kvm_io_bus_unregister_dev(), which takes the kvm->srcu lock.

However, this violates the established order where kvm->srcu is taken on a memory fault (such as an MMIO access), possibly followed by taking the config_lock if the GIC emulation requires mutual exclusion from the other vcpus.

It therefore results in a bad lockdep splat, as reported by Zenghui.

Fix this by moving the call to kvm_io_bus_unregister_dev() outside of the config_lock critical section. At this stage, there shouln't be any need to hold the config_lock.

Quoted 7 lines

As an additional bonus, document the ordering between kvm->slots_lock, kvm->srcu and kvm->arch.config_lock so that I cannot pretend I didn't know about those anymore.

Fixes 9eb18136af9f ("KVM: arm64: vgic: Hold config_lock while tearing down a CPU interface")

Thanks,
Zenghui

Applied
#2 lore raw
In reply to Marc Zyngier
Quoted 11 lines · Marc Zyngier

We recently moved the teardown of the vgic part of a vcpu inside a critical section guarded by the config_lock. This teardown phase involves calling into kvm_io_bus_unregister_dev(), which takes the kvm->srcu lock.

However, this violates the established order where kvm->srcu is taken on a memory fault (such as an MMIO access), possibly followed by taking the config_lock if the GIC emulation requires mutual exclusion from the other vcpus.

[...]

Tested this w/ kvm-unit-tests, selftests, and a few VMs on a lockdep kernel.

Applied to kvmarm/fixes, thanks!

[1/1] KVM: arm64: vgic: Don't hold config_lock while unregistering redistributors

      https://git.kernel.org/kvmarm/kvmarm/c/f616506754d3

--
Best,
Oliver