kvm_ioapic_eoi_inject_work() re-delivers a throttled level-triggered interrupt via kvm_irq_delivery_to_apic(), which walks kvm->arch.apic_map and dereferences the destination vCPU's APIC. The work is cancelled only in kvm_ioapic_destroy(), which runs after kvm_destroy_vcpus() has freed the vCPUs and their APICs. kvm_free_lapic() does not rebuild apic_map, so the map is left with dangling pointers, and a work item that fires during that window reads freed memory: BUG: KASAN: slab-use-after-free in __kvm_irq_delivery_to_apic_fast (arch/x86/kvm/lapic.c:1248) Read of size 8 by task kworker/3:1 Workqueue: events kvm_ioapic_eoi_inject_work __kvm_irq_delivery_to_apic_fast (arch/x86/kvm/lapic.c:1248) __kvm_irq_delivery_to_apic (arch/x86/kvm/lapic.c:1343) ioapic_service (arch/x86/kvm/ioapic.c:492) kvm_ioapic_eoi_inject_work (arch/x86/kvm/ioapic.c:525) process_one_work Freed by task 153: kvm_arch_vcpu_destroy (arch/x86/kvm/x86.c:12871) kvm_destroy_vcpus (virt/kvm/kvm_main.c:489) kvm_arch_destroy_vm (arch/x86/kvm/x86.c:13402) kvm_destroy_vm (virt/kvm/kvm_main.c:1302) kvm_vm_release (virt/kvm/kvm_main.c:1363) A guest arms the work by EOIing a level-triggered pin 10000 times in a row, so the window is reachable from guest ring 0 whenever its VM is torn down soon after. Destroy the in-kernel PIC and IOAPIC in kvm_arch_pre_destroy_vm(), before vCPUs are freed, so the eoi_inject work is cancelled while the target APICs are still valid. This also unregisters the PIC/IOAPIC MMIO devices while the KVM buses still exist; kvm_destroy_vm() tears the buses down right after kvm_free_irq_routing() and before kvm_arch_destroy_vm(), so the previous kvm_io_bus_unregister_dev() in kvm_ioapic_destroy() was a no-op. Fixes: 184564efae4d ("kvm: ioapic: conditionally delay irq delivery duringeoi broadcast") Link: https://lore.kernel.org/all/88ba60ad32ba851426a3f6590b0e402210991b4a.e33b58ce.0008.4e1c.aa62.c1024b242cbf@bytedance.com/ Suggested-by: Kai Huang Reported-by: Zhong Wang Reported-by: Xuanqing Shi Signed-off-by: Weiming Shi --- v1: https://lore.kernel.org/all/20260705045450.1325048-2-bestswngs@gmail.com/ v2: - Per Kai's suggestion, instead of adding a kvm_ioapic_pre_destroy() helper that only cancels the eoi_inject work, move kvm_pic_destroy()/kvm_ioapic_destroy() as a whole into kvm_arch_pre_destroy_vm(). This also fixes the stale kvm_io_bus_unregister_dev() Kai pointed out. arch/x86/kvm/x86.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index fd1c4a36b593..5925da351a9a 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -13370,10 +13370,14 @@ void kvm_arch_pre_destroy_vm(struct kvm *kvm) * Stop all background workers and kthreads before destroying vCPUs, as * iterating over vCPUs in a different task while vCPUs are being freed * is unsafe, i.e. will lead to use-after-free. The PIT also needs to - * be stopped before IRQ routing is freed. + * be stopped before IRQ routing is freed. The PIC and IOAPIC need to + * be destroyed here too, for the same reason, and because they must + * be destroyed before the KVM buses are torn down. */ #ifdef CONFIG_KVM_IOAPIC kvm_free_pit(kvm); + kvm_pic_destroy(kvm); + kvm_ioapic_destroy(kvm); #endif kvm_mmu_pre_destroy_vm(kvm); @@ -13400,10 +13404,6 @@ void kvm_arch_destroy_vm(struct kvm *kvm) perf_release_mediated_pmu(); kvm_destroy_vcpus(kvm); kvm_free_msr_filter(srcu_dereference_check(kvm->arch.msr_filter, &kvm->srcu, 1)); -#ifdef CONFIG_KVM_IOAPIC - kvm_pic_destroy(kvm); - kvm_ioapic_destroy(kvm); -#endif kvfree(rcu_dereference_check(kvm->arch.apic_map, 1)); kfree(srcu_dereference_check(kvm->arch.pmu_event_filter, &kvm->srcu, 1)); kvm_mmu_uninit_vm(kvm); -- 2.54.0