Re: [PATCH v2] KVM: X86: Fix softlockup when get the current kvmclock timestamp

[Radim KrÄmÃÅ]

2017-11-06 04:17-0800, Wanpeng Li:
> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> 
>  watchdog: BUG: soft lockup - CPU#6 stuck for 22s! [qemu-system-x86:10185]
>  CPU: 6 PID: 10185 Comm: qemu-system-x86 Tainted: G           OE   4.14.0-rc4+ #4
>  RIP: 0010:kvm_get_time_scale+0x4e/0xa0 [kvm]
>  Call Trace:
>   ? get_kvmclock_ns+0xa3/0x140 [kvm]
>   get_time_ref_counter+0x5a/0x80 [kvm]
>   kvm_hv_process_stimers+0x120/0x5f0 [kvm]
>   ? kvm_hv_process_stimers+0x120/0x5f0 [kvm]
>   ? preempt_schedule+0x27/0x30
>   ? ___preempt_schedule+0x16/0x18
>   kvm_arch_vcpu_ioctl_run+0x4b4/0x1690 [kvm]
>   ? kvm_arch_vcpu_load+0x47/0x230 [kvm]
>   kvm_vcpu_ioctl+0x33a/0x620 [kvm]
>   ? kvm_vcpu_ioctl+0x33a/0x620 [kvm]
>   ? kvm_vm_ioctl_check_extension_generic+0x3b/0x40 [kvm]
>   ? kvm_dev_ioctl+0x279/0x6c0 [kvm]
>   do_vfs_ioctl+0xa1/0x5d0
>   ? __fget+0x73/0xa0
>   SyS_ioctl+0x79/0x90
>   entry_SYSCALL_64_fastpath+0x1e/0xa9
> 
> This can be reproduced when running kvm-unit-tests/hyperv_stimer.flat and 
> cpu-hotplug stress simultaneously. __this_cpu_read(cpu_tsc_khz) returns 0 
> (set in kvmclock_cpu_down_prep()) when the pCPU is unhotplug which results 
> in kvm_get_time_scale() gets into an infinite loop.
> 
> This patch fixes it by skipping to fill the hv_clock when the pCPU is offline.
> 
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> ---
> v1 -> v2:
>  * avoid infinite loop
> 
>  arch/x86/kvm/x86.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 03869eb..d2507c6 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -1259,6 +1259,9 @@ static void kvm_get_time_scale(uint64_t scaled_hz, uint64_t base_hz,
>  	uint64_t tps64;
>  	uint32_t tps32;
>  
> +	if (unlikely(base_hz == 0))
> +		return;

This is a sensible thing to do and will prevent the loop, but KVM will
still have a minor bug:  get_kvmclock_ns() passes uninitialized stack
values with the expectation that kvm_get_time_scale() will set them, but
returning here would result in __pvclock_read_cycles() with random data
and inject timer interrupts early (if not worse).

I think it would be best if kvm_get_time_scale() wasn't executing when
cpu_tsc_khz is 0, by clearing cpu_tsc_khz later and setting earlier;
do you see any problems with moving the CPUHP_AP_X86_KVM_CLK_ONLINE
before CPUHP_AP_ONLINE?

Thanks.