KVM: arm64: vgic-v3: Retire all pending LPIs on vcpu destroy
authorZenghui Yu <yuzenghui@huawei.com>
Tue, 14 Apr 2020 03:03:47 +0000 (11:03 +0800)
committerMarc Zyngier <maz@kernel.org>
Thu, 23 Apr 2020 15:26:56 +0000 (16:26 +0100)
It's likely that the vcpu fails to handle all virtual interrupts if
userspace decides to destroy it, leaving the pending ones stay in the
ap_list. If the un-handled one is a LPI, its vgic_irq structure will
be eventually leaked because of an extra refcount increment in
vgic_queue_irq_unlock().

This was detected by kmemleak on almost every guest destroy, the
backtrace is as follows:

unreferenced object 0xffff80725aed5500 (size 128):
comm "CPU 5/KVM", pid 40711, jiffies 4298024754 (age 166366.512s)
hex dump (first 32 bytes):
00 00 00 00 00 00 00 00 08 01 a9 73 6d 80 ff ff ...........sm...
c8 61 ee a9 00 20 ff ff 28 1e 55 81 6c 80 ff ff .a... ..(.U.l...
backtrace:
[<000000004bcaa122>] kmem_cache_alloc_trace+0x2dc/0x418
[<0000000069c7dabb>] vgic_add_lpi+0x88/0x418
[<00000000bfefd5c5>] vgic_its_cmd_handle_mapi+0x4dc/0x588
[<00000000cf993975>] vgic_its_process_commands.part.5+0x484/0x1198
[<000000004bd3f8e3>] vgic_its_process_commands+0x50/0x80
[<00000000b9a65b2b>] vgic_mmio_write_its_cwriter+0xac/0x108
[<0000000009641ebb>] dispatch_mmio_write+0xd0/0x188
[<000000008f79d288>] __kvm_io_bus_write+0x134/0x240
[<00000000882f39ac>] kvm_io_bus_write+0xe0/0x150
[<0000000078197602>] io_mem_abort+0x484/0x7b8
[<0000000060954e3c>] kvm_handle_guest_abort+0x4cc/0xa58
[<00000000e0d0cd65>] handle_exit+0x24c/0x770
[<00000000b44a7fad>] kvm_arch_vcpu_ioctl_run+0x460/0x1988
[<0000000025fb897c>] kvm_vcpu_ioctl+0x4f8/0xee0
[<000000003271e317>] do_vfs_ioctl+0x160/0xcd8
[<00000000e7f39607>] ksys_ioctl+0x98/0xd8

Fix it by retiring all pending LPIs in the ap_list on the destroy path.

p.s. I can also reproduce it on a normal guest shutdown. It is because
userspace still send LPIs to vcpu (through KVM_SIGNAL_MSI ioctl) while
the guest is being shutdown and unable to handle it. A little strange
though and haven't dig further...

Reviewed-by: James Morse <james.morse@arm.com>
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
[maz: moved the distributor deallocation down to avoid an UAF splat]
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200414030349.625-2-yuzenghui@huawei.com
virt/kvm/arm/vgic/vgic-init.c

index a963b9d..30dbec9 100644 (file)
@@ -348,6 +348,12 @@ void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
 {
        struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 
+       /*
+        * Retire all pending LPIs on this vcpu anyway as we're
+        * going to destroy it.
+        */
+       vgic_flush_pending_lpis(vcpu);
+
        INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
 }
 
@@ -359,10 +365,10 @@ static void __kvm_vgic_destroy(struct kvm *kvm)
 
        vgic_debug_destroy(kvm);
 
-       kvm_vgic_dist_destroy(kvm);
-
        kvm_for_each_vcpu(i, vcpu, kvm)
                kvm_vgic_vcpu_destroy(vcpu);
+
+       kvm_vgic_dist_destroy(kvm);
 }
 
 void kvm_vgic_destroy(struct kvm *kvm)