[dpdk-dev] [PATCH RFC v3 1/3] vhost: use SMP barriers instead of compiler ones.

[Ilya Maximets]

CC to list.

On 18.03.2016 15:19, Ilya Maximets wrote:
> 
> 
> On 18.03.2016 14:42, Ilya Maximets wrote:
>> On 18.03.2016 14:30, Xie, Huawei wrote:
>>> On 3/18/2016 7:00 PM, Ilya Maximets wrote:
>>>> On 18.03.2016 13:47, Xie, Huawei wrote:
>>>>> On 3/18/2016 6:39 PM, Ilya Maximets wrote:
>>>>>> On 18.03.2016 13:27, Xie, Huawei wrote:
>>>>>>> On 3/18/2016 6:23 PM, Ilya Maximets wrote:
>>>>>>>> On 18.03.2016 13:08, Xie, Huawei wrote:
>>>>>>>>> On 2/24/2016 7:47 PM, Ilya Maximets wrote:
>>>>>>>>>>  		 * Wait until it's our turn to add our buffer
>>>>>>>>>> @@ -979,7 +979,7 @@ rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
>>>>>>>>>>  		entry_success++;
>>>>>>>>>>  	}
>>>>>>>>>>  
>>>>>>>>>> -	rte_compiler_barrier();
>>>>>>>>>> +	rte_smp_rmb();
>>>>>>>>> smp_rmb()?
>>>>>>>> There is no such function 'smp_rmb' in DPDK.
>>>>>>>> But:
>>>>>>>> .../arch/arm/rte_atomic.h:#define rte_smp_rmb() rte_rmb()
>>>>>>>> .../arch/ppc_64/rte_atomic.h:#define rte_smp_rmb() rte_compiler_barrier()
>>>>>>>> .../arch/tile/rte_atomic.h:#define rte_smp_rmb() rte_compiler_barrier()
>>>>>>>> .../arch/x86/rte_atomic.h:#define rte_smp_rmb() rte_compiler_barrier()
>>>>>>> I mean shoudn't be rte_smp_wmb()?
>>>>>> No. Here we need to be sure that copying of data from descriptor to
>>>>>> our local mbuf completed before 'vq->used->idx += entry_success;'.
>>>>>>
>>>>>> Read memory barrier will help us with it.
>>>>>>
>>>>>> In other places write barriers used because copying performed in
>>>>>> opposite direction.
>>>>> What about the udpate to the used ring?
>>>> Next line is the only place where this vq->used->idx accessed.
>>>> So, there must be no issues.
>>>
>>> The update to the used ring entries, happened before the update to the
>>> used ring index.
>>
>> Oh. Sorry. In that case we need full barrier here because we need reads and
>> writes both completed before updating of used->idx:
>> 	rte_smp_mb();
> 
> Hmmm.. But as I see, rte_smp_mb is a real mm_fence on x86. May be the pair
> of barriers will be better here:
> 	rte_smp_rmb();
> 	rte_smp_wmb();
> 
> It is normal because next increment uses read and write. So, we don't need to
> synchronize reads with writes here.
> 
>>
>>>
>>>>
>>>>>>>>>>  	vq->used->idx += entry_success;
>>>>>>>>>>  	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
>>>>>>>>>>  			sizeof(vq->used->idx));
>>>>>>>>>> -- 2.5.0
>>>>>>>
>>>>>
>>>>>
>>>
>>>
>>>