[PATCH 1/1] RISC-V: KVM: Avoid lock inversion in SBI_EXT_HSM_HART_START

[Anup Patel]

On Tue, Mar 26, 2024 at 3:41 PM Yong-Xuan Wang <yongxuan.wang at sifive.com> wrote:
>
> Documentation/virt/kvm/locking.rst advises that kvm->lock should be
> acquired outside vcpu->mutex and kvm->srcu. However, when KVM/RISC-V
> handling SBI_EXT_HSM_HART_START, the lock ordering is vcpu->mutex,
> kvm->srcu then kvm->lock.
>
> Although the lockdep checking no longer complains about this after commit
> f0f44752f5f6 ("rcu: Annotate SRCU's update-side lockdep dependencies"),
> it's necessary to replace kvm->lock with a new dedicated lock to ensure
> only one hart can execute the SBI_EXT_HSM_HART_START call for the target
> hart simultaneously.
>
> Signed-off-by: Yong-Xuan Wang <yongxuan.wang at sifive.com>
> ---
>  arch/riscv/include/asm/kvm_host.h | 1 +
>  arch/riscv/kvm/vcpu.c             | 1 +
>  arch/riscv/kvm/vcpu_sbi_hsm.c     | 5 ++---
>  3 files changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> index 484d04a92fa6..537099413344 100644
> --- a/arch/riscv/include/asm/kvm_host.h
> +++ b/arch/riscv/include/asm/kvm_host.h
> @@ -254,6 +254,7 @@ struct kvm_vcpu_arch {
>
>         /* VCPU power-off state */
>         bool power_off;
> +       struct mutex hsm_start_lock;

Instead of a mutex hsm_start_lock, let's introduce spinlock mp_state_lock
which needs to be taken whenever power_off is accessed. Also, we should
rename "power_off" to "mp_state" with two possible values.

>
>         /* Don't run the VCPU (blocked) */
>         bool pause;
> diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> index b5ca9f2e98ac..4d89b5b5afbf 100644
> --- a/arch/riscv/kvm/vcpu.c
> +++ b/arch/riscv/kvm/vcpu.c
> @@ -119,6 +119,7 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
>         spin_lock_init(&vcpu->arch.hfence_lock);
>
>         /* Setup reset state of shadow SSTATUS and HSTATUS CSRs */
> +       mutex_init(&vcpu->arch.hsm_start_lock);
>         cntx = &vcpu->arch.guest_reset_context;
>         cntx->sstatus = SR_SPP | SR_SPIE;
>         cntx->hstatus = 0;
> diff --git a/arch/riscv/kvm/vcpu_sbi_hsm.c b/arch/riscv/kvm/vcpu_sbi_hsm.c
> index 7dca0e9381d9..b528f6e880ae 100644
> --- a/arch/riscv/kvm/vcpu_sbi_hsm.c
> +++ b/arch/riscv/kvm/vcpu_sbi_hsm.c
> @@ -71,14 +71,13 @@ static int kvm_sbi_ext_hsm_handler(struct kvm_vcpu *vcpu, struct kvm_run *run,
>  {
>         int ret = 0;
>         struct kvm_cpu_context *cp = &vcpu->arch.guest_context;
> -       struct kvm *kvm = vcpu->kvm;
>         unsigned long funcid = cp->a6;
>
>         switch (funcid) {
>         case SBI_EXT_HSM_HART_START:
> -               mutex_lock(&kvm->lock);
> +               mutex_lock(&vcpu->arch.hsm_start_lock);
>                 ret = kvm_sbi_hsm_vcpu_start(vcpu);
> -               mutex_unlock(&kvm->lock);
> +               mutex_unlock(&vcpu->arch.hsm_start_lock);

The use of kvm->lock over here was also protecting
simultaneous updates to VCPU reset context. It's better
to introduce a separate lock for protecting VCPU reset
context access.

>                 break;
>         case SBI_EXT_HSM_HART_STOP:
>                 ret = kvm_sbi_hsm_vcpu_stop(vcpu);
> --
> 2.17.1
>

I think this patch can be broken down into two patches:
1) Patch replacing VCPU "power_off" with "enum mp_state"
    and introducing "mp_state_lock"
2) Patch introducing VCPU reset context lock

Regards,
Anup