Re: [PATCH] KVM: x86: Fix potential preemption when get the current kvmclock timestamp

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/05/2017 14:15, Wanpeng Li wrote:
> 2017-05-11 20:10 GMT+08:00 Paolo Bonzini <pbonzini@redhat.com>:
>>
>>
>> On 11/05/2017 14:00, Wanpeng Li wrote:
>>> From: Wanpeng Li <wanpeng.li@hotmail.com>
>>>
>>>  BUG: using __this_cpu_read() in preemptible [00000000] code: qemu-system-x86/2809
>>>  caller is __this_cpu_preempt_check+0x13/0x20
>>>  CPU: 2 PID: 2809 Comm: qemu-system-x86 Not tainted 4.11.0+ #13
>>>  Call Trace:
>>>   dump_stack+0x99/0xce
>>>   check_preemption_disabled+0xf5/0x100
>>>   __this_cpu_preempt_check+0x13/0x20
>>>   get_kvmclock_ns+0x6f/0x110 [kvm]
>>>   get_time_ref_counter+0x5d/0x80 [kvm]
>>>   kvm_hv_process_stimers+0x2a1/0x8a0 [kvm]
>>>   ? kvm_hv_process_stimers+0x2a1/0x8a0 [kvm]
>>>   ? kvm_arch_vcpu_ioctl_run+0xac9/0x1ce0 [kvm]
>>>   kvm_arch_vcpu_ioctl_run+0x5bf/0x1ce0 [kvm]
>>>   kvm_vcpu_ioctl+0x384/0x7b0 [kvm]
>>>   ? kvm_vcpu_ioctl+0x384/0x7b0 [kvm]
>>>   ? __fget+0xf3/0x210
>>>   do_vfs_ioctl+0xa4/0x700
>>>   ? __fget+0x114/0x210
>>>   SyS_ioctl+0x79/0x90
>>>   entry_SYSCALL_64_fastpath+0x23/0xc2
>>>  RIP: 0033:0x7f9d164ed357
>>>  RSP: 002b:00007f9d0f6768f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
>>>  RAX: ffffffffffffffda RBX: ffffffffa64d53c3 RCX: 00007f9d164ed357
>>>  RDX: 0000000000000000 RSI: 000000000000ae80 RDI: 000000000000000d
>>>  RBP: ffffbb260856bf88 R08: 0000556b2a13eeb0 R09: 0000000000000000
>>>  R10: 00007f9d080000c8 R11: 0000000000000246 R12: 0000000000000000
>>>  R13: 00007f9d1853d000 R14: 0000000000000000 R15: 000000000000ae80
>>>   ? __this_cpu_preempt_check+0x13/0x20
>>>
>>> This can be reproduced by run kvm-unit-tests/hyperv_stimer.flat w/
>>> CONFIG_PREEMPT and CONFIG_DEBUG_PREEMPT enabled.
>>>
>>> Safe access to per-CPU data requires a couple of constraints, though: the
>>> thread working with the data cannot be preempted and it cannot be migrated
>>> while it manipulates per-CPU variables. If the thread is preempted, the
>>> thread that replaces it could try to work with the same variables; migration
>>> to another CPU could also cause confusion. However there is no preemption
>>> disable when reads host per-CPU tsc rate to calculate the current kvmclock
>>> timestamp.
>>>
>>> This patch fix it by holding pvclock_gtod_sync_lock lock when calculates
>>> pvclock's time scale in order to disable preemption for host per-CPU tsc
>>> rate read.
>>>
>>> Cc: Paolo Bonzini <pbonzini@redhat.com>
>>> Cc: Radim Krčmář <rkrcmar@redhat.com>
>>> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
>>> ---
>>>  arch/x86/kvm/x86.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>> index b54125b..8008d56 100644
>>> --- a/arch/x86/kvm/x86.c
>>> +++ b/arch/x86/kvm/x86.c
>>> @@ -1772,11 +1772,11 @@ u64 get_kvmclock_ns(struct kvm *kvm)
>>>
>>>       hv_clock.tsc_timestamp = ka->master_cycle_now;
>>>       hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
>>> -     spin_unlock(&ka->pvclock_gtod_sync_lock);
>>>
>>>       kvm_get_time_scale(NSEC_PER_SEC, __this_cpu_read(cpu_tsc_khz) * 1000LL,
>>>                          &hv_clock.tsc_shift,
>>>                          &hv_clock.tsc_to_system_mul);
>>> +     spin_unlock(&ka->pvclock_gtod_sync_lock);
>>>       return __pvclock_read_cycles(&hv_clock, rdtsc());
>>>  }
>>>
>>>
>>
>> This would not be enough for PREEMPT_RT.  You need to use
>> get_cpu/put_cpu (including __pvclock_read_cycles in the non-preemptable
>> section).
> 
> Actually the splat is for __this_cpu_read(cpu_tsc_khz), so I just protect it.

Yes, but spin_lock/spin_unlock would still be preemptable on PREEMPT_RT.

Also, rdtsc() should really be on the same CPU as __this_cpu_read.  We
know it's not really really necessary because the master clock is
active, but since we need a get_cpu/put_cpu pair, better be clean.

Paolo