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

[Wanpeng Li]

2017-11-06 18:21 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
> On 06/11/2017 11:06, Wanpeng Li wrote:
>> 2017-11-06 17:29 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>>> On 06/11/2017 01:55, Wanpeng Li wrote:
>>>> This can be reproduced when running kvm-unit-tests/hyperv_stimer.flat and
>>>> cpu-hotplug stress simultaneously. kvm_get_time_scale() takes too long which
>>>> results in softlockup.
>>>
>>> Apart from the pr_debug, kvm_get_time_scale should take less than a
>>> microsecond.  The patch is fine, but can you confirm that pr_debug is
>>> the culprit?
>>
>> I can still encounter softlockup after removing the pr_debug.
>
> Is kvm_get_time_scale getting into an infinite loop then?  That would be
> the actual bug.

I think so, kvm_get_time_scale almost occupy 100% cpu utilization.

         :      static void kvm_get_time_scale(uint64_t scaled_hz,
uint64_t base_hz,
         :                                     s8 *pshift, u32 *pmultiplier)
         :      {
    8.97 :        21a34:       test   %rdx,%rbx
    0.00 :        21a37:       jne    21a86 <kvm_get_time_scale+0xb6>
         :              uint64_t tps64;
         :              uint32_t tps32;
         :
         :              tps64 = base_hz;
         :              scaled64 = scaled_hz;
         :              while (tps64 > scaled64*2 || tps64 &
0xffffffff00000000ULL) {
   28.01 :        21a39:       test   %r13d,%r13d
         :                      tps64 >>= 1;
    0.00 :        21a3c:       js     21a86 <kvm_get_time_scale+0xb6>
         :                      shift--;
    6.16 :        21a3e:       add    %r13d,%r13d
         :              uint64_t tps64;
         :              uint32_t tps32;
         :
         :              tps64 = base_hz;
         :              scaled64 = scaled_hz;
         :              while (tps64 > scaled64*2 || tps64 &
0xffffffff00000000ULL) {
    7.22 :        21a41:       mov    %r13d,%eax
   17.86 :        21a44:       add    $0x1,%r14d
    0.00 :        21a48:       cmp    %rax,%rbx
         :                      tps64 >>= 1;
         :                      shift--;
         :              }
         :
         :              tps32 = (uint32_t)tps64;
         :              while (tps32 <= scaled64 || scaled64 &
0xffffffff00000000ULL) {
   31.78 :        21a4b:       jae    21a34 <kvm_get_time_scale+0x64>
    0.00 :        21a4d:       mov    %r15,%rdi
         :              while (tps64 > scaled64*2 || tps64 &
0xffffffff00000000ULL) {
         :                      tps64 >>= 1;
         :                      shift--;
         :              }

Regards,
Wanpeng Li