Virtualization - Apr 2021 - [PATCH net-next v3 3/8] virtio-net: xsk zero copy xmit setup

? 2021/4/7 ??3:27, Xuan Zhuo ??:
> On Tue, 6 Apr 2021 12:27:14 +0800, Jason Wang <jasowang at
redhat.com> wrote:
>> ? 2021/3/31 ??3:11, Xuan Zhuo ??:
>>> xsk is a high-performance packet receiving and sending technology.
>>>
>>> This patch implements the binding and unbinding operations of xsk
and
>>> the virtio-net queue for xsk zero copy xmit.
>>>
>>> The xsk zero copy xmit depends on tx napi. So if tx napi is not
opened,
>>> an error will be reported. And the entire operation is under the
>>> protection of rtnl_lock
>>>
>>> Signed-off-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com>
>>> Reviewed-by: Dust Li <dust.li at linux.alibaba.com>
>>> ---
>>>    drivers/net/virtio_net.c | 66
++++++++++++++++++++++++++++++++++++++++
>>>    1 file changed, 66 insertions(+)
>>>
>>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
>>> index bb4ea9dbc16b..4e25408a2b37 100644
>>> --- a/drivers/net/virtio_net.c
>>> +++ b/drivers/net/virtio_net.c
>>> @@ -22,6 +22,7 @@
>>>    #include <net/route.h>
>>>    #include <net/xdp.h>
>>>    #include <net/net_failover.h>
>>> +#include <net/xdp_sock_drv.h>
>>>
>>>    static int napi_weight = NAPI_POLL_WEIGHT;
>>>    module_param(napi_weight, int, 0444);
>>> @@ -133,6 +134,11 @@ struct send_queue {
>>>    	struct virtnet_sq_stats stats;
>>>
>>>    	struct napi_struct napi;
>>> +
>>> +	struct {
>>> +		/* xsk pool */
>>> +		struct xsk_buff_pool __rcu *pool;
>>> +	} xsk;
>>>    };
>>>
>>>    /* Internal representation of a receive virtqueue */
>>> @@ -2526,11 +2532,71 @@ static int virtnet_xdp_set(struct
net_device *dev, struct bpf_prog *prog,
>>>    	return err;
>>>    }
>>>
>>> +static int virtnet_xsk_pool_enable(struct net_device *dev,
>>> +				   struct xsk_buff_pool *pool,
>>> +				   u16 qid)
>>> +{
>>> +	struct virtnet_info *vi = netdev_priv(dev);
>>> +	struct send_queue *sq;
>>> +	int ret = -EBUSY;
>>
>> I'd rather move this under the check of xsk.pool.
>>
>>
>>> +
>>> +	if (qid >= vi->curr_queue_pairs)
>>> +		return -EINVAL;
>>> +
>>> +	sq = &vi->sq[qid];
>>> +
>>> +	/* xsk zerocopy depend on the tx napi */
>>
>> Need more comments to explain why tx NAPI is required here.
>>
>> And what's more important, tx NAPI could be enabled/disable via
ethtool,
>> what if the NAPI is disabled after xsk is enabled?
>>
> If napi_tx is turned off, then the xmit will be affected.

Please document what kind of effect that prevents xsk from working.

> Maybe I
> should restrict that tx NAPI be disable via ethtool after xsk is enabled.

It can work.

>
>>> +	if (!sq->napi.weight)
>>> +		return -EPERM;
>>> +
>>> +	rcu_read_lock();
>>> +	if (rcu_dereference(sq->xsk.pool))
>>> +		goto end;
>>
>> Under what condition can we reach here?
> When the user tries to bind repeatedly.

Ok, I am a little suprised that it was not checked by xsk_bind().


>
>>
>>> +
>>> +	/* Here is already protected by rtnl_lock, so rcu_assign_pointer
is
>>> +	 * safe.
>>> +	 */
>>> +	rcu_assign_pointer(sq->xsk.pool, pool);
>>> +	ret = 0;
>>> +end:
>>> +	rcu_read_unlock();
>>> +
>>> +	return ret;
>>> +}
>>> +
>>> +static int virtnet_xsk_pool_disable(struct net_device *dev, u16
qid)
>>> +{
>>> +	struct virtnet_info *vi = netdev_priv(dev);
>>> +	struct send_queue *sq;
>>> +
>>> +	if (qid >= vi->curr_queue_pairs)
>>> +		return -EINVAL;
>>> +
>>> +	sq = &vi->sq[qid];
>>> +
>>> +	/* Here is already protected by rtnl_lock, so rcu_assign_pointer
is
>>> +	 * safe.
>>> +	 */
>>> +	rcu_assign_pointer(sq->xsk.pool, NULL);
>>> +
>>> +	synchronize_rcu(); /* Sync with the XSK wakeup and with NAPI. */
>>
>> Since rtnl is held here, I guess it's better to use
synchornize_net().
>>
>>
>>> +
>>> +	return 0;
>>> +}
>>> +
>>>    static int virtnet_xdp(struct net_device *dev, struct netdev_bpf
*xdp)
>>>    {
>>>    	switch (xdp->command) {
>>>    	case XDP_SETUP_PROG:
>>>    		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
>>> +	case XDP_SETUP_XSK_POOL:
>>> +		/* virtio net not use dma before call vring api */
>>> +		xdp->xsk.check_dma = false;
>>
>> I think it's better not open code things like this. How about
introduce
>> new parameters in xp_assign_dev()?
> xp_assign_dev is called by the user, we should let xsk perceive that the
current
> dev does not directly use dma. Is it possible to use dev->priv_flags? I
only
> know that this is the case with virtio-net!!

Ok, then it should be fine, but we need a better name other than 
"check_dma". Maybe "use_dma_addr" instead.

Thanks

>
> Thanks very much.
>
>>
>>> +		if (xdp->xsk.pool)
>>> +			return virtnet_xsk_pool_enable(dev, xdp->xsk.pool,
>>> +						       xdp->xsk.queue_id);
>>> +		else
>>> +			return virtnet_xsk_pool_disable(dev, xdp->xsk.queue_id);
>>>    	default:
>>>    		return -EINVAL;
>>>    	}
>>
>> Thanks
>>