[PATCH v5] virtio: optimize stats counters performance

[lihuisong (C)]

在 2024/8/2 0:03, Morten Brørup 写道:
> Optimized the performance of updating the virtio statistics counters by
> reducing the number of branches.
>
> Ordered the packet size comparisons according to the probability with
> typical internet traffic mix.
>
> Signed-off-by: Morten Brørup <mb at smartsharesystems.com>
> ---
> v5:
> * Do not inline the function. (Stephen)
> v4:
> * Consider multicast/broadcast packets unlikely.
> v3:
> * Eliminated a local variable.
> * Note: Substituted sizeof(uint32_t)*4 by 32UL, using unsigned long type
>    to keep optimal offsetting in generated assembler output.
> * Removed unnecessary curly braces.
> v2:
> * Fixed checkpatch warning about line length.
> ---
>   drivers/net/virtio/virtio_rxtx.c | 39 ++++++++++++--------------------
>   drivers/net/virtio/virtio_rxtx.h |  4 ++--
>   2 files changed, 16 insertions(+), 27 deletions(-)
>
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index f69b9453a2..b67f063b31 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -82,37 +82,26 @@ vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
>   }
>   
>   void
> -virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
> +virtio_update_packet_stats(struct virtnet_stats *const stats,
> +		const struct rte_mbuf *const mbuf)
The two const is also for performace?  Is there gain?
>   {
>   	uint32_t s = mbuf->pkt_len;
> -	struct rte_ether_addr *ea;
> +	const struct rte_ether_addr *const ea =
> +			rte_pktmbuf_mtod(mbuf, const struct rte_ether_addr *);
>   
>   	stats->bytes += s;
>   
> -	if (s == 64) {
> -		stats->size_bins[1]++;
> -	} else if (s > 64 && s < 1024) {
> -		uint32_t bin;
> -
> -		/* count zeros, and offset into correct bin */
> -		bin = (sizeof(s) * 8) - rte_clz32(s) - 5;
> -		stats->size_bins[bin]++;
> -	} else {
> -		if (s < 64)
> -			stats->size_bins[0]++;
> -		else if (s < 1519)
> -			stats->size_bins[6]++;
> -		else
> -			stats->size_bins[7]++;
> -	}
> +	if (s >= 1024)
> +		stats->size_bins[6 + (s > 1518)]++;
> +	else if (s <= 64)
> +		stats->size_bins[s >> 6]++;
> +	else
> +		stats->size_bins[32UL - rte_clz32(s) - 5]++;
>   
> -	ea = rte_pktmbuf_mtod(mbuf, struct rte_ether_addr *);
> -	if (rte_is_multicast_ether_addr(ea)) {
> -		if (rte_is_broadcast_ether_addr(ea))
> -			stats->broadcast++;
> -		else
> -			stats->multicast++;
> -	}
> +	RTE_BUILD_BUG_ON(offsetof(struct virtnet_stats, broadcast) !=
> +			offsetof(struct virtnet_stats, multicast) + sizeof(uint64_t));
> +	if (unlikely(rte_is_multicast_ether_addr(ea)))
> +		(&stats->multicast)[rte_is_broadcast_ether_addr(ea)]++;
The "rte_is_broadcast_ether_addr(ea) " will be calculated twice if 
packet is mulitcast.
How about coding like:
-->
is_mulitcast = rte_is_multicast_ether_addr(ea);
if (unlikely(is_mulitcast))
(&stats->multicast)[rte_is_broadcast_ether_addr(ea)]++;
>   }
>   
>   static inline void
> diff --git a/drivers/net/virtio/virtio_rxtx.h b/drivers/net/virtio/virtio_rxtx.h
> index afc4b74534..68034c914b 100644
> --- a/drivers/net/virtio/virtio_rxtx.h
> +++ b/drivers/net/virtio/virtio_rxtx.h
> @@ -35,7 +35,7 @@ struct virtnet_tx {
>   };
>   
>   int virtio_rxq_vec_setup(struct virtnet_rx *rxvq);
> -void virtio_update_packet_stats(struct virtnet_stats *stats,
> -				struct rte_mbuf *mbuf);
> +void virtio_update_packet_stats(struct virtnet_stats *const stats,
> +		const struct rte_mbuf *const mbuf);
>   
>   #endif /* _VIRTIO_RXTX_H_ */