Re: [PATCH net-next 4/5] veth: introduce more xdp counters

[Toshiaki Makita <toshiaki.makita1@gmail.com>]

On 2020/03/24 23:36, Lorenzo Bianconi wrote:
> 
> [...]
> 
>>>> }
>>>
>>> actually I have a different idea..what about account tx stats on the peer rx
>>> queue as a result of XDP_TX or ndo_xdp_xmit and properly report this info in
>>> the ethool stats? In this way we do not have any locking issue and we still use
>>> the per-queue stats approach. Could you please take a look to the following patch?
>>
>> Thanks I think your idea is nice.
>>
>>> diff --git a/drivers/net/veth.c b/drivers/net/veth.c
>>> index b6505a6c7102..f2acd2ee6287 100644
>>> --- a/drivers/net/veth.c
>>> +++ b/drivers/net/veth.c
>>> @@ -92,17 +92,22 @@ struct veth_q_stat_desc {
>>>    static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
>>>    	{ "xdp_packets",	VETH_RQ_STAT(xdp_packets) },
>>>    	{ "xdp_bytes",		VETH_RQ_STAT(xdp_bytes) },
>>> -	{ "rx_drops",		VETH_RQ_STAT(rx_drops) },
>>> -	{ "rx_xdp_redirect",	VETH_RQ_STAT(xdp_redirect) },
>>> -	{ "rx_xdp_drops",	VETH_RQ_STAT(xdp_drops) },
>>> -	{ "rx_xdp_tx",		VETH_RQ_STAT(xdp_tx) },
>>> -	{ "rx_xdp_tx_errors",	VETH_RQ_STAT(xdp_tx_err) },
>>> -	{ "tx_xdp_xmit",	VETH_RQ_STAT(xdp_xmit) },
>>> -	{ "tx_xdp_xmit_errors",	VETH_RQ_STAT(xdp_xmit_err) },
>>> +	{ "drops",		VETH_RQ_STAT(rx_drops) },
>>> +	{ "xdp_redirect",	VETH_RQ_STAT(xdp_redirect) },
>>> +	{ "xdp_drops",		VETH_RQ_STAT(xdp_drops) },
>>>    };
>>>    #define VETH_RQ_STATS_LEN	ARRAY_SIZE(veth_rq_stats_desc)
>>> +static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
>>> +	{ "xdp_tx",		VETH_RQ_STAT(xdp_tx) },
>>> +	{ "xdp_tx_errors",	VETH_RQ_STAT(xdp_tx_err) },
>>
>> I'm wondering why xdp_tx is not accounted in rx_queue?
>> You can count xdp_tx/tx_error somewhere in rx path like veth_xdp_flush_bq().
>> xdp_redirect and xdp_drops are similar actions and in rx stats.
> 
> Hi,
> 
> thanks for the review :)
> 
> I moved the accounting of xdp_tx/tx_error in veth_xdp_xmit for two reason:
> 1- veth_xdp_tx in veth_xdp_rcv_one or veth_xdp_rcv_skb returns an error
>     for XDP_TX just if xdp_frame pointer is invalid but the packet can be
>     discarded in veth_xdp_xmit if the device is 'under-pressure' (and this can
>     be a problem since in XDP there are no queueing mechanisms so far)

Right, but you can track the discard in veth_xdp_flush_bq().

> 2- to be symmetric  with ndo_xdp_xmit

I thought consistency between drivers is more important. What about other drivers?

Toshiaki Makita

> 
>>
>>> +	{ "xdp_xmit",		VETH_RQ_STAT(xdp_xmit) },
>>> +	{ "xdp_xmit_errors",	VETH_RQ_STAT(xdp_xmit_err) },
>>> +};
>>> +
>>> +#define VETH_TQ_STATS_LEN	ARRAY_SIZE(veth_tq_stats_desc)
>>> +
>>>    static struct {
>>>    	const char string[ETH_GSTRING_LEN];
>>>    } ethtool_stats_keys[] = {
>>> @@ -142,6 +147,14 @@ static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
>>>    				p += ETH_GSTRING_LEN;
>>>    			}
>>>    		}
>>> +		for (i = 0; i < dev->real_num_tx_queues; i++) {
>>> +			for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
>>> +				snprintf(p, ETH_GSTRING_LEN,
>>> +					 "tx_queue_%u_%.18s",
>>> +					 i, veth_tq_stats_desc[j].desc);
>>> +				p += ETH_GSTRING_LEN;
>>> +			}
>>> +		}
>>>    		break;
>>>    	}
>>>    }
>>> @@ -151,7 +164,8 @@ static int veth_get_sset_count(struct net_device *dev, int sset)
>>>    	switch (sset) {
>>>    	case ETH_SS_STATS:
>>>    		return ARRAY_SIZE(ethtool_stats_keys) +
>>> -		       VETH_RQ_STATS_LEN * dev->real_num_rx_queues;
>>> +		       VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
>>> +		       VETH_TQ_STATS_LEN * dev->real_num_tx_queues;
>>>    	default:
>>>    		return -EOPNOTSUPP;
>>>    	}
>>> @@ -160,7 +174,7 @@ static int veth_get_sset_count(struct net_device *dev, int sset)
>>>    static void veth_get_ethtool_stats(struct net_device *dev,
>>>    		struct ethtool_stats *stats, u64 *data)
>>>    {
>>> -	struct veth_priv *priv = netdev_priv(dev);
>>> +	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
>>>    	struct net_device *peer = rtnl_dereference(priv->peer);
>>>    	int i, j, idx;
>>> @@ -181,6 +195,26 @@ static void veth_get_ethtool_stats(struct net_device *dev,
>>>    		} while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
>>>    		idx += VETH_RQ_STATS_LEN;
>>>    	}
>>> +
>>> +	if (!peer)
>>> +		return;
>>> +
>>> +	rcv_priv = netdev_priv(peer);
>>> +	for (i = 0; i < peer->real_num_rx_queues; i++) {
>>> +		const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
>>> +		const void *stats_base = (void *)&rq_stats->vs;
>>> +		unsigned int start, tx_idx;
>>> +		size_t offset;
>>> +
>>> +		tx_idx = (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
>>
>> I'm a bit concerned that this can fold multiple rx queue counters into one
>> tx counter. But I cannot think of a better idea provided that we want to
>> align XDP stats between drivers. So I'm OK with this.
> 
> Since peer->real_num_rx_queues can be greater than dev->real_num_tx_queues,
> right? IIUC the only guarantee we have is:
> 
> peer->real_num_tx_queues < dev->real_num_rx_queues
> 
> If you are fine with that approach, I will post a patch before net-next
> closure.
> 
> Regards,
> Lorenzo
> 
> 
>>
>> Toshiaki Makita
>>
>>> +		do {
>>> +			start = u64_stats_fetch_begin_irq(&rq_stats->syncp);
>>> +			for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
>>> +				offset = veth_tq_stats_desc[j].offset;
>>> +				data[tx_idx + idx + j] += *(u64 *)(stats_base + offset);
>>> +			}
>>> +		} while (u64_stats_fetch_retry_irq(&rq_stats->syncp, start));
>>> +	}
>>>    }
>>>    static const struct ethtool_ops veth_ethtool_ops = {
>>> @@ -340,8 +374,7 @@ static void veth_get_stats64(struct net_device *dev,
>>>    	tot->tx_packets = packets;
>>>    	veth_stats_rx(&rx, dev);
>>> -	tot->tx_dropped += rx.xdp_xmit_err + rx.xdp_tx_err;
>>> -	tot->rx_dropped = rx.rx_drops;
>>> +	tot->rx_dropped = rx.rx_drops + rx.xdp_tx_err + rx.xdp_xmit_err;
>>>    	tot->rx_bytes = rx.xdp_bytes;
>>>    	tot->rx_packets = rx.xdp_packets;
>>> @@ -353,7 +386,7 @@ static void veth_get_stats64(struct net_device *dev,
>>>    		tot->rx_packets += packets;
>>>    		veth_stats_rx(&rx, peer);
>>> -		tot->rx_dropped += rx.xdp_xmit_err + rx.xdp_tx_err;
>>> +		tot->tx_dropped += rx.xdp_xmit_err + rx.xdp_tx_err;
>>>    		tot->tx_bytes += rx.xdp_bytes;
>>>    		tot->tx_packets += rx.xdp_packets;
>>>    	}
>>> @@ -394,9 +427,9 @@ static int veth_xdp_xmit(struct net_device *dev, int n,
>>>    			 u32 flags, bool ndo_xmit)
>>>    {
>>>    	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
>>> -	unsigned int qidx, max_len;
>>>    	struct net_device *rcv;
>>>    	int i, ret, drops = n;
>>> +	unsigned int max_len;
>>>    	struct veth_rq *rq;
>>>    	rcu_read_lock();
>>> @@ -414,8 +447,7 @@ static int veth_xdp_xmit(struct net_device *dev, int n,
>>>    	}
>>>    	rcv_priv = netdev_priv(rcv);
>>> -	qidx = veth_select_rxq(rcv);
>>> -	rq = &rcv_priv->rq[qidx];
>>> +	rq = &rcv_priv->rq[veth_select_rxq(rcv)];
>>>    	/* Non-NULL xdp_prog ensures that xdp_ring is initialized on receive
>>>    	 * side. This means an XDP program is loaded on the peer and the peer
>>>    	 * device is up.
>>> @@ -446,7 +478,6 @@ static int veth_xdp_xmit(struct net_device *dev, int n,
>>>    	ret = n - drops;
>>>    drop:
>>> -	rq = &priv->rq[qidx];
>>>    	u64_stats_update_begin(&rq->stats.syncp);
>>>    	if (ndo_xmit) {
>>>    		rq->stats.vs.xdp_xmit += n - drops;
>>>
>>>>
>>>> Thoughts?
>>>>
>>>> Toshiaki Makita