The current logic keeps all secondary VCPUs powered off on their
first run in kvm_arch_vcpu_postcreate(), relying on the boot VCPU 
to wake them up by sbi call. This is correct for a fresh VM start,
where VCPUs begin execution at the bootaddress (0x80000000).

However, this behavior is not suitable for VCPUs that are being
restored from a state (e.g., during migration resume or snapshot
load). These VCPUs have a saved program counter (sepc). Forcing
them to wait for a wake-up from the boot VCPU, which may not
happen or may happen incorrectly, leaves them in a stuck state
when using Qemu to migration if smp is larger than one.

So check a cold start and a warm resumption by the value of the 
guest's sepc register. If the VCPU is running for the first time 
*and* its sepc is not the hardware boot address, it indicates a 
resumed vCPU that must be powered on immediately to continue 
execution from its saved context.

Signed-off-by: Jinyu Tang <tjytimi@163.com>
Tested-by: Tianshun Sun <stsmail163@163.com>
---
 arch/riscv/kvm/vcpu.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
index 3ebcfffaa..86aeba886 100644
--- a/arch/riscv/kvm/vcpu.c
+++ b/arch/riscv/kvm/vcpu.c
@@ -867,8 +867,16 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 	struct kvm_cpu_trap trap;
 	struct kvm_run *run = vcpu->run;
 
-	if (!vcpu->arch.ran_atleast_once)
+	if (!vcpu->arch.ran_atleast_once) {
 		kvm_riscv_vcpu_setup_config(vcpu);
+		/*
+		 * For VCPUs that are resuming (e.g., from migration)
+		 * and not starting from the boot address, explicitly
+		 * power them on.
+		 */
+		if (vcpu->arch.guest_context.sepc != 0x80000000)
+			kvm_riscv_vcpu_power_on(vcpu);
+	}
 
 	/* Mark this VCPU ran at least once */
 	vcpu->arch.ran_atleast_once = true;
-- 
2.43.0

2025-09-15T20:23:34+08:00, Jinyu Tang <tjytimi@163.com>:
> The current logic keeps all secondary VCPUs powered off on their
> first run in kvm_arch_vcpu_postcreate(), relying on the boot VCPU 
> to wake them up by sbi call. This is correct for a fresh VM start,
> where VCPUs begin execution at the bootaddress (0x80000000).
>
> However, this behavior is not suitable for VCPUs that are being
> restored from a state (e.g., during migration resume or snapshot
> load). These VCPUs have a saved program counter (sepc). Forcing
> them to wait for a wake-up from the boot VCPU, which may not
> happen or may happen incorrectly, leaves them in a stuck state
> when using Qemu to migration if smp is larger than one.
>
> So check a cold start and a warm resumption by the value of the 
> guest's sepc register. If the VCPU is running for the first time 
> *and* its sepc is not the hardware boot address, it indicates a 
> resumed vCPU that must be powered on immediately to continue 
> execution from its saved context.
>
> Signed-off-by: Jinyu Tang <tjytimi@163.com>
> Tested-by: Tianshun Sun <stsmail163@163.com>
> ---

I don't like this approach.  Userspace controls the state of the VM, and
KVM shouldn't randomly change the state that userspace wants.

> diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> @@ -867,8 +867,16 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
>  	struct kvm_cpu_trap trap;
>  	struct kvm_run *run = vcpu->run;
>  
> -	if (!vcpu->arch.ran_atleast_once)
> +	if (!vcpu->arch.ran_atleast_once) {
>  		kvm_riscv_vcpu_setup_config(vcpu);
> +		/*
> +		 * For VCPUs that are resuming (e.g., from migration)
> +		 * and not starting from the boot address, explicitly
> +		 * power them on.
> +		 */
> +		if (vcpu->arch.guest_context.sepc != 0x80000000)

Offlined VCPUs are not guaranteed to have sepc == 0x80000000, so this
patch would incorrectly wake them up.
(Depending on vcpu->arch.ran_atleast_once is flaky at best as well.)

Please try to fix userspace instead,

Thanks.

On Mon, Sep 15, 2025 at 04:19:21PM +0200, Radim Krčmář wrote:
> 2025-09-15T20:23:34+08:00, Jinyu Tang <tjytimi@163.com>:
> > The current logic keeps all secondary VCPUs powered off on their
> > first run in kvm_arch_vcpu_postcreate(), relying on the boot VCPU 
> > to wake them up by sbi call. This is correct for a fresh VM start,
> > where VCPUs begin execution at the bootaddress (0x80000000).
> >
> > However, this behavior is not suitable for VCPUs that are being
> > restored from a state (e.g., during migration resume or snapshot
> > load). These VCPUs have a saved program counter (sepc). Forcing
> > them to wait for a wake-up from the boot VCPU, which may not
> > happen or may happen incorrectly, leaves them in a stuck state
> > when using Qemu to migration if smp is larger than one.
> >
> > So check a cold start and a warm resumption by the value of the 
> > guest's sepc register. If the VCPU is running for the first time 
> > *and* its sepc is not the hardware boot address, it indicates a 
> > resumed vCPU that must be powered on immediately to continue 
> > execution from its saved context.
> >
> > Signed-off-by: Jinyu Tang <tjytimi@163.com>
> > Tested-by: Tianshun Sun <stsmail163@163.com>
> > ---
> 
> I don't like this approach.  Userspace controls the state of the VM, and
> KVM shouldn't randomly change the state that userspace wants.
> 
> > diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> > @@ -867,8 +867,16 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
> >  	struct kvm_cpu_trap trap;
> >  	struct kvm_run *run = vcpu->run;
> >  
> > -	if (!vcpu->arch.ran_atleast_once)
> > +	if (!vcpu->arch.ran_atleast_once) {
> >  		kvm_riscv_vcpu_setup_config(vcpu);
> > +		/*
> > +		 * For VCPUs that are resuming (e.g., from migration)
> > +		 * and not starting from the boot address, explicitly
> > +		 * power them on.
> > +		 */
> > +		if (vcpu->arch.guest_context.sepc != 0x80000000)
> 
> Offlined VCPUs are not guaranteed to have sepc == 0x80000000, so this
> patch would incorrectly wake them up.
> (Depending on vcpu->arch.ran_atleast_once is flaky at best as well.)
> 
> Please try to fix userspace instead,

Yes, and maybe it's already fixed
https://lore.kernel.org/all/20250915070811.3422578-1-xb@ultrarisc.com/

Thanks,
drew