Re: [Xen-devel] PML causing race condition during guest bootstorm and host crash on Broadwell cpu.

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 08.02.17 at 16:32, <JBeulich@xxxxxxxx> wrote:
>>>> On 07.02.17 at 18:26, <anshul.makkar@xxxxxxxxxx> wrote:
>> Facing a issue where bootstorm of guests leads to host crash. I debugged 
>> and found that that enabling PML  introduces a  race condition during 
>> guest teardown stage while disabling PML on a vcpu  and context switch 
>> happening for another vcpu.
>> 
>> Crash happens only on Broadwell processors as PML got introduced in this 
>> generation.
>> 
>> Here is my analysis:
>> 
>> Race condition:
>> 
>> vmcs.c vmx_vcpu_disable_pml (vcpu){ vmx_vmcs_enter() ; vm_write( 
>> disable_PML); vmx_vmcx_exit();)
>> 
>> In between I have a callpath from another pcpu executing context 
>> switch-> vmx_fpu_leave() and crash on vmwrite..
>> 
>>    if ( !(v->arch.hvm_vmx.host_cr0 & X86_CR0_TS) )
>> {
>>           v->arch.hvm_vmx.host_cr0 |= X86_CR0_TS;
>>           __vmwrite(HOST_CR0, v->arch.hvm_vmx.host_cr0);  //crash
>>       }
> 
> So that's after current has changed already, so it's effectively
> dealing with a foreign VMCS, but it doesn't use vmx_vmcs_enter().
> The locking done in vmx_vmcs_try_enter() / vmx_vmcs_exit(),
> however, assumes that any user of a VMCS either owns the lock
> or has current as the owner of the VMCS. Of course such a call
> also can't be added here, as a vcpu on the context-switch-from
> path can't vcpu_pause() itself.
> 
> That taken together with the bypassing of __context_switch()
> when the incoming vCPU is the idle one (which means that via
> context_saved() ->is_running will be cleared before running
> ->ctxt_switch_from()), the vcpu_pause() invocation in
> vmx_vmcs_try_enter() may not have to wait at all if the call
> happens between the clearing of ->is_running and the
> eventual invocation of vmx_ctxt_switch_from().
> 
> If the above makes sense (which I'm not sure at all), the
> question is whether using this_cpu(curr_vcpu) instead of
> current in the VMCS enter/exit functions would help.

This won't help, as it won't make vcpu_pause() wait (which is the
core of the problem).

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel