Re: [PATCH 01/12] xen/events: avoid race with raising an event in unmask_evtchn()

[David Vrabel <david.vrabel@citrix.com>]

On 20/03/13 11:00, Stefano Stabellini wrote:
> On Tue, 19 Mar 2013, David Vrabel wrote:
>> From: David Vrabel <david.vrabel@citrix.com>
>>
>> In unmask_evtchn(), when the mask bit is cleared after testing for
>> pending and the event becomes pending between the test and clear, then
>> the upcall will not become pending and the event may be lost or
>> delayed.
>>
>> Avoid this by always clearing the mask bit before checking for
>> pending.
>>
>> This fixes a regression introduced in 3.7 by
>> b5e579232d635b79a3da052964cb357ccda8d9ea (xen/events: fix
>> unmask_evtchn for PV on HVM guests) which reordered the clear mask and
>> check pending operations.
> 
> The race you are trying to fix is real, but the fix you are proposing
> breaks PV on HVM and ARM guests again.
> 
> From the description of b5e579232d635b79a3da052964cb357ccda8d9ea, it's
> clear that the reason to call EVTCHNOP_unmask is to trigger an event
> notification injection again.
> But if you clear the evtchn_mask bit *before* calling EVTCHNOP_unmask,
> EVTCHNOP_unmask won't reinject the event.
> From evtchn_unmask:
> 
>     if ( test_and_clear_bit(port, &shared_info(d, evtchn_mask)) &&
>          test_bit          (port, &shared_info(d, evtchn_pending)) &&
>          !test_and_set_bit (port / BITS_PER_EVTCHN_WORD(d),
>                             &vcpu_info(v, evtchn_pending_sel)) )
>     {
>         vcpu_mark_events_pending(v);
>     }
> 
> The first condition for reinjection would fail.

I missed this.  The only way I can think of fixing this is to set the
mask bit before call the unmask hypercall.

The FIFO-based ABI doesn't have this problem as it always tries to
relink the event whatever the previous state of the mask bit was.

David

>> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
>> Cc: stable@vger.kernel.org
>> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
>> ---
>>  drivers/xen/events.c |   10 +++++-----
>>  1 files changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/xen/events.c b/drivers/xen/events.c
>> index d17aa41..4bdd0a5 100644
>> --- a/drivers/xen/events.c
>> +++ b/drivers/xen/events.c
>> @@ -403,11 +403,13 @@ static void unmask_evtchn(int port)
>>  
>>  	if (unlikely((cpu != cpu_from_evtchn(port))))
>>  		do_hypercall = 1;
>> -	else
>> +	else {
>> +		sync_clear_bit(port, BM(&s->evtchn_mask[0]));
>>  		evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
>>  
>> -	if (unlikely(evtchn_pending && xen_hvm_domain()))
>> -		do_hypercall = 1;
>> +		if (unlikely(evtchn_pending && xen_hvm_domain()))
>> +			do_hypercall = 1;
>> +	}
>>  
>>  	/* Slow path (hypercall) if this is a non-local port or if this is
>>  	 * an hvm domain and an event is pending (hvm domains don't have
>> @@ -418,8 +420,6 @@ static void unmask_evtchn(int port)
>>  	} else {
>>  		struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
>>  
>> -		sync_clear_bit(port, BM(&s->evtchn_mask[0]));
>> -
>>  		/*
>>  		 * The following is basically the equivalent of
>>  		 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
>> -- 
>> 1.7.2.5
>>