Re: [PATCH v3 2/2] KVM: x86: fix underflow in TSC deadline calculation

From: Paolo Bonzini
Date: Thu Jun 30 2016 - 06:22:00 EST




On 30/06/2016 02:52, Wanpeng Li wrote:
> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
>
> INFO: rcu_sched detected stalls on CPUs/tasks:
> 1-...: (11800 GPs behind) idle=45d/140000000000000/0 softirq=0/0 fqs=21663
> (detected by 0, t=65016 jiffies, g=11500, c=11499, q=719)
> Task dump for CPU 1:
> qemu-system-x86 R running task 0 3529 3525 0x00080808
> ffff8802021791a0 ffff880212895040 0000000000000001 00007f1c2c00db40
> ffff8801dd20fcd3 ffffc90002b98000 ffff8801dd20fc88 ffff8801dd20fcf8
> 0000000000000286 ffff8801dd2ac538 ffff8801dd20fcc0 ffffffffc06949c9
> Call Trace:
> ? kvm_write_guest_cached+0xb9/0x160 [kvm]
> ? __delay+0xf/0x20
> ? wait_lapic_expire+0x14a/0x200 [kvm]
> ? kvm_arch_vcpu_ioctl_run+0xcbe/0x1b00 [kvm]
> ? kvm_arch_vcpu_ioctl_run+0xe34/0x1b00 [kvm]
> ? kvm_vcpu_ioctl+0x2d3/0x7c0 [kvm]
> ? __fget+0x5/0x210
> ? do_vfs_ioctl+0x96/0x6a0
> ? __fget_light+0x2a/0x90
> ? SyS_ioctl+0x79/0x90
> ? do_syscall_64+0x7c/0x1e0
> ? entry_SYSCALL64_slow_path+0x25/0x25
>
> This can be reproduced readily by running a full dynticks guest(since hrtimer
> in guest is heavily used) w/ lapic_timer_advance disabled.
>
> If fail to program hardware preemption timer, we will fallback to hrtimer based
> method, however, a previous programmed preemption timer miss to cancel in this
> scenario which results in one hardware preemption timer and one hrtimer emulated
> tsc deadline timer run simultaneously. So sometimes the target guest deadline
> tsc is earlier than guest tsc, which leads to the computation in vmx_set_hv_timer
> can underflow and cause delta_tsc to be set a huge value, then host soft lockup
> as above.
>
> This patch fix it by cancelling the previous programmed preemption timer if there
> is once we failed to program the new preemption timer and fallback to hrtimer
> based method.
>
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
> Cc: Yunhong Jiang <yunhong.jiang@xxxxxxxxx>
> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> ---
> v2 -> v3:
> * depends on start_hv_tscdeadline caller to start sw_timer
> * move lapic_timer.pending check in start_hv_tscdeadline

Thanks, this looks good. There are only a couple aesthetic changes
I'd like to make:

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index b44c5879b9b6..22a6474af220 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -1384,7 +1384,7 @@ static bool start_hv_tscdeadline(struct kvm_lapic *apic)
}
trace_kvm_hv_timer_state(apic->vcpu->vcpu_id,
apic->lapic_timer.hv_timer_in_use);
- return kvm_lapic_hv_timer_in_use(apic->vcpu);
+ return apic->lapic_timer.hv_timer_in_use;
}

void kvm_lapic_switch_to_hv_timer(struct kvm_vcpu *vcpu)
@@ -1460,9 +1460,10 @@ static void start_apic_timer(struct kvm_lapic *apic)
apic->lapic_timer.period,
ktime_to_ns(ktime_add_ns(now,
apic->lapic_timer.period)));
- } else if (apic_lvtt_tscdeadline(apic))
+ } else if (apic_lvtt_tscdeadline(apic)) {
if (!(kvm_x86_ops->set_hv_timer && start_hv_tscdeadline(apic)))
start_sw_tscdeadline(apic);
+ }
}

static void apic_manage_nmi_watchdog(struct kvm_lapic *apic, u32 lvt0_val)


Paolo