The struct vcpu_svm.ir_list and ir_list_lock are being accessed even when
AVIC is not enabled, while current code only initialize the list and
the lock only when AVIC is enabled. This ended up trigger 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, move the initialziation code before checking for AVIC enabled
so that it is always excuted.
Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into avic_init_vcpu()")
Signed-off-by: Suravee Suthikulpanit <[email protected]>
---
arch/x86/kvm/svm/avic.c | 2 --
arch/x86/kvm/svm/svm.c | 3 +++
2 files changed, 3 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index ac830cd50830..1ccf13783785 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -572,8 +572,6 @@ int avic_init_vcpu(struct vcpu_svm *svm)
if (ret)
return ret;
- INIT_LIST_HEAD(&svm->ir_list);
- spin_lock_init(&svm->ir_list_lock);
svm->dfr_reg = APIC_DFR_FLAT;
return ret;
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index c44f3e9140d5..714d791fe5a5 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -1225,6 +1225,9 @@ static int svm_create_vcpu(struct kvm_vcpu *vcpu)
svm_init_osvw(vcpu);
vcpu->arch.microcode_version = 0x01000065;
+ INIT_LIST_HEAD(&svm->ir_list);
+ spin_lock_init(&svm->ir_list_lock);
+
return 0;
free_page4:
--
2.17.1
Hi,
Are there any issues or concerns about this patch?
Thank you,
Suravee
On 9/22/20 3:44 PM, Suravee Suthikulpanit wrote:
> The struct vcpu_svm.ir_list and ir_list_lock are being accessed even when
> AVIC is not enabled, while current code only initialize the list and
> the lock only when AVIC is enabled. This ended up trigger 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, move the initialziation code before checking for AVIC enabled
> so that it is always excuted.
>
> Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into avic_init_vcpu()")
> Signed-off-by: Suravee Suthikulpanit <[email protected]>
> ---
> arch/x86/kvm/svm/avic.c | 2 --
> arch/x86/kvm/svm/svm.c | 3 +++
> 2 files changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
> index ac830cd50830..1ccf13783785 100644
> --- a/arch/x86/kvm/svm/avic.c
> +++ b/arch/x86/kvm/svm/avic.c
> @@ -572,8 +572,6 @@ int avic_init_vcpu(struct vcpu_svm *svm)
> if (ret)
> return ret;
>
> - INIT_LIST_HEAD(&svm->ir_list);
> - spin_lock_init(&svm->ir_list_lock);
> svm->dfr_reg = APIC_DFR_FLAT;
>
> return ret;
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index c44f3e9140d5..714d791fe5a5 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -1225,6 +1225,9 @@ static int svm_create_vcpu(struct kvm_vcpu *vcpu)
> svm_init_osvw(vcpu);
> vcpu->arch.microcode_version = 0x01000065;
>
> + INIT_LIST_HEAD(&svm->ir_list);
> + spin_lock_init(&svm->ir_list_lock);
> +
> return 0;
>
> free_page4:
>
On 28/09/20 07:53, Suravee Suthikulpanit wrote:
> Hi,
>
> Are there any issues or concerns about this patch?
Yes, sorry I haven't replied yet. Looks like Linus is doing an -rc8 so
there's plenty of time to have it in 5.9.
The thing I'm wondering is, why is svm_update_pi_irte doing anything if
you don't have AVIC enabled? In other word, this might not be the root
cause of the bug. You always get to the "else" branch of the loop of
course, and I'm not sure how irq_set_vcpu_affinity returns something
with pi.prev_ga_tag set.
Thanks,
Paolo
> Thank you,
> Suravee
>
> On 9/22/20 3:44 PM, Suravee Suthikulpanit wrote:
>> The struct vcpu_svm.ir_list and ir_list_lock are being accessed even when
>> AVIC is not enabled, while current code only initialize the list and
>> the lock only when AVIC is enabled. This ended up trigger 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, move the initialziation code before checking for AVIC enabled
>> so that it is always excuted.
>>
>> Fixes: dfa20099e26e ("KVM: SVM: Refactor AVIC vcpu initialization into
>> avic_init_vcpu()")
>> Signed-off-by: Suravee Suthikulpanit <[email protected]>
>> ---
>> arch/x86/kvm/svm/avic.c | 2 --
>> arch/x86/kvm/svm/svm.c | 3 +++
>> 2 files changed, 3 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
>> index ac830cd50830..1ccf13783785 100644
>> --- a/arch/x86/kvm/svm/avic.c
>> +++ b/arch/x86/kvm/svm/avic.c
>> @@ -572,8 +572,6 @@ int avic_init_vcpu(struct vcpu_svm *svm)
>> if (ret)
>> return ret;
>> - INIT_LIST_HEAD(&svm->ir_list);
>> - spin_lock_init(&svm->ir_list_lock);
>> svm->dfr_reg = APIC_DFR_FLAT;
>> return ret;
>> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
>> index c44f3e9140d5..714d791fe5a5 100644
>> --- a/arch/x86/kvm/svm/svm.c
>> +++ b/arch/x86/kvm/svm/svm.c
>> @@ -1225,6 +1225,9 @@ static int svm_create_vcpu(struct kvm_vcpu *vcpu)
>> svm_init_osvw(vcpu);
>> vcpu->arch.microcode_version = 0x01000065;
>> + INIT_LIST_HEAD(&svm->ir_list);
>> + spin_lock_init(&svm->ir_list_lock);
>> +
>> return 0;
>> free_page4:
>>
>
Paolo,
On 9/28/20 3:01 PM, Paolo Bonzini wrote:
> On 28/09/20 07:53, Suravee Suthikulpanit wrote:
>> Hi,
>>
>> Are there any issues or concerns about this patch?
>
> Yes, sorry I haven't replied yet. Looks like Linus is doing an -rc8 so
> there's plenty of time to have it in 5.9.
>
> The thing I'm wondering is, why is svm_update_pi_irte doing anything if
> you don't have AVIC enabled? In other word, this might not be the root
> cause of the bug. You always get to the "else" branch of the loop of
> course, and I'm not sure how irq_set_vcpu_affinity returns something
> with pi.prev_ga_tag set.
You are right. pi_prev_ga_tag needs to be initialized before used
(in case AVIC is not enabled). I have already sent out another patch
to properly fix the issue instead with subject
(KVM: SVM: Initialize prev_ga_tag before use).
Thanks,
Suravee