[PATCH] KVM: SVM: Initialize prev_ga_tag before use

[PATCH] KVM: SVM: Initialize prev_ga_tag before use

[Suravee Suthikulpanit]

The function amd_ir_set_vcpu_affinity makes use of the parameter struct
amd_iommu_pi_data.prev_ga_tag to determine if it should delete struct
amd_iommu_pi_data from a list when not running in AVIC mode.

However, prev_ga_tag is initialized only when AVIC is enabled. The non-zero
uninitialized value can cause unintended code path, which ends up making
use of the struct vcpu_svm.ir_list and ir_list_lock without being
initialized (since they are intended only for the AVIC case).

This triggers NULL pointer dereference bug in the function vm_ir_list_del
with the following call trace:

    svm_update_pi_irte+0x3c2/0x550 [kvm_amd]
    ? proc_create_single_data+0x41/0x50
    kvm_arch_irq_bypass_add_producer+0x40/0x60 [kvm]
    __connect+0x5f/0xb0 [irqbypass]
    irq_bypass_register_producer+0xf8/0x120 [irqbypass]
    vfio_msi_set_vector_signal+0x1de/0x2d0 [vfio_pci]
    vfio_msi_set_block+0x77/0xe0 [vfio_pci]
    vfio_pci_set_msi_trigger+0x25c/0x2f0 [vfio_pci]
    vfio_pci_set_irqs_ioctl+0x88/0xb0 [vfio_pci]
    vfio_pci_ioctl+0x2ea/0xed0 [vfio_pci]
    ? alloc_file_pseudo+0xa5/0x100
    vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
    ? vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
    __x64_sys_ioctl+0x96/0xd0
    do_syscall_64+0x37/0x80
    entry_SYSCALL_64_after_hwframe+0x44/0xa9

Therefore, initialize prev_ga_tag to zero before use. This should be safe
because ga_tag value 0 is invalid (see function avic_vm_init).

Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into avic_init_vcpu()")
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
---
 arch/x86/kvm/svm/avic.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index ac830cd50830..381d22daa4ac 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -868,6 +868,7 @@ int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
 			 * - Tell IOMMU to use legacy mode for this interrupt.
 			 * - Retrieve ga_tag of prior interrupt remapping data.
 			 */
+			pi.prev_ga_tag = 0;
 			pi.is_guest_mode = false;
 			ret = irq_set_vcpu_affinity(host_irq, &pi);
 
-- 
2.17.1

Re: [PATCH] KVM: SVM: Initialize prev_ga_tag before use

[Suravee Suthikulpanit]

Paolo,

Are there any issues or concerns about this patch?

Thank you,
Suravee

On 10/4/20 6:27 AM, Suravee Suthikulpanit wrote:
> The function amd_ir_set_vcpu_affinity makes use of the parameter struct
> amd_iommu_pi_data.prev_ga_tag to determine if it should delete struct
> amd_iommu_pi_data from a list when not running in AVIC mode.
> 
> However, prev_ga_tag is initialized only when AVIC is enabled. The non-zero
> uninitialized value can cause unintended code path, which ends up making
> use of the struct vcpu_svm.ir_list and ir_list_lock without being
> initialized (since they are intended only for the AVIC case).
> 
> This triggers NULL pointer dereference bug in the function vm_ir_list_del
> with the following call trace:
> 
>      svm_update_pi_irte+0x3c2/0x550 [kvm_amd]
>      ? proc_create_single_data+0x41/0x50
>      kvm_arch_irq_bypass_add_producer+0x40/0x60 [kvm]
>      __connect+0x5f/0xb0 [irqbypass]
>      irq_bypass_register_producer+0xf8/0x120 [irqbypass]
>      vfio_msi_set_vector_signal+0x1de/0x2d0 [vfio_pci]
>      vfio_msi_set_block+0x77/0xe0 [vfio_pci]
>      vfio_pci_set_msi_trigger+0x25c/0x2f0 [vfio_pci]
>      vfio_pci_set_irqs_ioctl+0x88/0xb0 [vfio_pci]
>      vfio_pci_ioctl+0x2ea/0xed0 [vfio_pci]
>      ? alloc_file_pseudo+0xa5/0x100
>      vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
>      ? vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
>      __x64_sys_ioctl+0x96/0xd0
>      do_syscall_64+0x37/0x80
>      entry_SYSCALL_64_after_hwframe+0x44/0xa9
> 
> Therefore, initialize prev_ga_tag to zero before use. This should be safe
> because ga_tag value 0 is invalid (see function avic_vm_init).
> 
> Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into avic_init_vcpu()")
> Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
> ---
>   arch/x86/kvm/svm/avic.c | 1 +
>   1 file changed, 1 insertion(+)
> 
> diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> index ac830cd50830..381d22daa4ac 100644
> --- a/arch/x86/kvm/svm/avic.c
> +++ b/arch/x86/kvm/svm/avic.c
> @@ -868,6 +868,7 @@ int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
>   			 * - Tell IOMMU to use legacy mode for this interrupt.
>   			 * - Retrieve ga_tag of prior interrupt remapping data.
>   			 */
> +			pi.prev_ga_tag = 0;
>   			pi.is_guest_mode = false;
>   			ret = irq_set_vcpu_affinity(host_irq, &pi);
>   
> 

Re: [PATCH] KVM: SVM: Initialize prev_ga_tag before use

[Paolo Bonzini]

On 04/10/20 01:27, Suravee Suthikulpanit wrote:
> The function amd_ir_set_vcpu_affinity makes use of the parameter struct
> amd_iommu_pi_data.prev_ga_tag to determine if it should delete struct
> amd_iommu_pi_data from a list when not running in AVIC mode.
> 
> However, prev_ga_tag is initialized only when AVIC is enabled. The non-zero
> uninitialized value can cause unintended code path, which ends up making
> use of the struct vcpu_svm.ir_list and ir_list_lock without being
> initialized (since they are intended only for the AVIC case).
> 
> This triggers NULL pointer dereference bug in the function vm_ir_list_del
> with the following call trace:
> 
>     svm_update_pi_irte+0x3c2/0x550 [kvm_amd]
>     ? proc_create_single_data+0x41/0x50
>     kvm_arch_irq_bypass_add_producer+0x40/0x60 [kvm]
>     __connect+0x5f/0xb0 [irqbypass]
>     irq_bypass_register_producer+0xf8/0x120 [irqbypass]
>     vfio_msi_set_vector_signal+0x1de/0x2d0 [vfio_pci]
>     vfio_msi_set_block+0x77/0xe0 [vfio_pci]
>     vfio_pci_set_msi_trigger+0x25c/0x2f0 [vfio_pci]
>     vfio_pci_set_irqs_ioctl+0x88/0xb0 [vfio_pci]
>     vfio_pci_ioctl+0x2ea/0xed0 [vfio_pci]
>     ? alloc_file_pseudo+0xa5/0x100
>     vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
>     ? vfio_device_fops_unl_ioctl+0x26/0x30 [vfio]
>     __x64_sys_ioctl+0x96/0xd0
>     do_syscall_64+0x37/0x80
>     entry_SYSCALL_64_after_hwframe+0x44/0xa9
> 
> Therefore, initialize prev_ga_tag to zero before use. This should be safe
> because ga_tag value 0 is invalid (see function avic_vm_init).
> 
> Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into avic_init_vcpu()")
> Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
> ---
>  arch/x86/kvm/svm/avic.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> index ac830cd50830..381d22daa4ac 100644
> --- a/arch/x86/kvm/svm/avic.c
> +++ b/arch/x86/kvm/svm/avic.c
> @@ -868,6 +868,7 @@ int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
>  			 * - Tell IOMMU to use legacy mode for this interrupt.
>  			 * - Retrieve ga_tag of prior interrupt remapping data.
>  			 */
> +			pi.prev_ga_tag = 0;
>  			pi.is_guest_mode = false;
>  			ret = irq_set_vcpu_affinity(host_irq, &pi);
>  
> 

Queued (with Cc: stable), thanks.

Paolo