Re: [PATCH v3 25/25] KVM: TDX: Fix list_add corruption during vcpu_load()

[Yan Zhao <yan.y.zhao@intel.com>]

On Thu, Oct 16, 2025 at 05:32:43PM -0700, Sean Christopherson wrote:
> From: Yan Zhao <yan.y.zhao@intel.com>
> 
> During vCPU creation, a vCPU may be destroyed immediately after
> kvm_arch_vcpu_create() (e.g., due to vCPU id confiliction). However, the
> vcpu_load() inside kvm_arch_vcpu_create() may have associate the vCPU to
> pCPU via "list_add(&tdx->cpu_list, &per_cpu(associated_tdvcpus, cpu))"
> before invoking tdx_vcpu_free().
> 
> Though there's no need to invoke tdh_vp_flush() on the vCPU, failing to
> dissociate the vCPU from pCPU (i.e., "list_del(&to_tdx(vcpu)->cpu_list)")
> will cause list corruption of the per-pCPU list associated_tdvcpus.
> 
> Then, a later list_add() during vcpu_load() would detect list corruption
> and print calltrace as shown below.
> 
> Dissociate a vCPU from its associated pCPU in tdx_vcpu_free() for the vCPUs
> destroyed immediately after creation which must be in
> VCPU_TD_STATE_UNINITIALIZED state.
> 
> kernel BUG at lib/list_debug.c:29!
> Oops: invalid opcode: 0000 [#2] SMP NOPTI
> RIP: 0010:__list_add_valid_or_report+0x82/0xd0
> 
> Call Trace:
>  <TASK>
>  tdx_vcpu_load+0xa8/0x120
>  vt_vcpu_load+0x25/0x30
>  kvm_arch_vcpu_load+0x81/0x300
>  vcpu_load+0x55/0x90
>  kvm_arch_vcpu_create+0x24f/0x330
>  kvm_vm_ioctl_create_vcpu+0x1b1/0x53
>  kvm_vm_ioctl+0xc2/0xa60
>   __x64_sys_ioctl+0x9a/0xf0
>  x64_sys_call+0x10ee/0x20d0
>  do_syscall_64+0xc3/0x470
>  entry_SYSCALL_64_after_hwframe+0x77/0x7f
> 
> Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
> Signed-off-by: Sean Christopherson <seanjc@google.com>
> ---
>  arch/x86/kvm/vmx/tdx.c | 43 +++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 38 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/x86/kvm/vmx/tdx.c b/arch/x86/kvm/vmx/tdx.c
> index d6541b08423f..daec88d4b88d 100644
> --- a/arch/x86/kvm/vmx/tdx.c
> +++ b/arch/x86/kvm/vmx/tdx.c
> @@ -833,19 +833,52 @@ void tdx_vcpu_put(struct kvm_vcpu *vcpu)
>  	tdx_prepare_switch_to_host(vcpu);
>  }
>  
> +/*
> + * Life cycles for a TD and a vCPU:
> + * 1. KVM_CREATE_VM ioctl.
> + *    TD state is TD_STATE_UNINITIALIZED.
> + *    hkid is not assigned at this stage.
> + * 2. KVM_TDX_INIT_VM ioctl.
> + *    TD transitions to TD_STATE_INITIALIZED.
> + *    hkid is assigned after this stage.
> + * 3. KVM_CREATE_VCPU ioctl. (only when TD is TD_STATE_INITIALIZED).
> + *    3.1 tdx_vcpu_create() transitions vCPU state to VCPU_TD_STATE_UNINITIALIZED.
> + *    3.2 vcpu_load() and vcpu_put() in kvm_arch_vcpu_create().
> + *    3.3 (conditional) if any error encountered after kvm_arch_vcpu_create()
> + *        kvm_arch_vcpu_destroy() --> tdx_vcpu_free().
> + * 4. KVM_TDX_INIT_VCPU ioctl.
> + *    tdx_vcpu_init() transitions vCPU state to VCPU_TD_STATE_INITIALIZED.
> + *    vCPU control structures are allocated at this stage.
> + * 5. kvm_destroy_vm().
> + *    5.1 tdx_mmu_release_hkid(): (1) tdh_vp_flush(), disassociates all vCPUs.
> + *                                (2) puts hkid to !assigned state.
> + *    5.2 kvm_destroy_vcpus() --> tdx_vcpu_free():
> + *        transitions vCPU to VCPU_TD_STATE_UNINITIALIZED state.
> + *    5.3 tdx_vm_destroy()
> + *        transitions TD to TD_STATE_UNINITIALIZED state.
> + *
> + * tdx_vcpu_free() can be invoked only at 3.3 or 5.2.
> + * - If at 3.3, hkid is still assigned, but the vCPU must be in
> + *   VCPU_TD_STATE_UNINITIALIZED state.
> + * - if at 5.2, hkid must be !assigned and all vCPUs must be in
> + *   VCPU_TD_STATE_INITIALIZED state and have been dissociated.
> + */
>  void tdx_vcpu_free(struct kvm_vcpu *vcpu)
>  {
>  	struct kvm_tdx *kvm_tdx = to_kvm_tdx(vcpu->kvm);
>  	struct vcpu_tdx *tdx = to_tdx(vcpu);
>  	int i;
>  
> +	if (vcpu->cpu != -1) {
> +		KVM_BUG_ON(tdx->state == VCPU_TD_STATE_INITIALIZED, vcpu->kvm);
> +		tdx_disassociate_vp(vcpu);
Sorry, I should use "tdx_flush_vp_on_cpu(vcpu);" rather than invoking
tdx_disassociate_vp() directly.

This is to ensure that list_del() in tdx_disassociate_vp() runs on the physical
cpu that owns the list and to which the vcpu is associated with in the previous
tdx_vcpu_load().

> +		return;
> +	}
> +
>  	/*
>  	 * It is not possible to reclaim pages while hkid is assigned. It might
> -	 * be assigned if:
> -	 * 1. the TD VM is being destroyed but freeing hkid failed, in which
> -	 * case the pages are leaked
> -	 * 2. TD VCPU creation failed and this on the error path, in which case
> -	 * there is nothing to do anyway
> +	 * be assigned if the TD VM is being destroyed but freeing hkid failed,
> +	 * in which case the pages are leaked.
>  	 */
>  	if (is_hkid_assigned(kvm_tdx))
>  		return;
> -- 
> 2.51.0.858.gf9c4a03a3a-goog
>