> -----Original Message-----
> From: Vladimir Oltean <olteanv@xxxxxxxxx>
> Sent: 2021年4月27日 2:40
> To: Richard Cochran <richardcochran@xxxxxxxxx>
> Cc: Y.b. Lu <yangbo.lu@xxxxxxx>; netdev@xxxxxxxxxxxxxxx; Vladimir Oltean
> <vladimir.oltean@xxxxxxx>; David S . Miller <davem@xxxxxxxxxxxxx>; Jakub
> Kicinski <kuba@xxxxxxxxxx>; Jonathan Corbet <corbet@xxxxxxx>; Kurt
> Kanzenbach <kurt@xxxxxxxxxxxxx>; Andrew Lunn <andrew@xxxxxxx>; Vivien
> Didelot <vivien.didelot@xxxxxxxxx>; Florian Fainelli <f.fainelli@xxxxxxxxx>;
> Claudiu Manoil <claudiu.manoil@xxxxxxx>; Alexandre Belloni
> <alexandre.belloni@xxxxxxxxxxx>; UNGLinuxDriver@xxxxxxxxxxxxx;
> linux-doc@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [net-next, v2, 3/7] net: dsa: free skb->cb usage in core driver
>
> On Mon, Apr 26, 2021 at 06:38:46AM -0700, Richard Cochran wrote:
> > On Mon, Apr 26, 2021 at 05:37:58PM +0800, Yangbo Lu wrote:
> > > @@ -624,7 +623,7 @@ static netdev_tx_t dsa_slave_xmit(struct sk_buff
> > > *skb, struct net_device *dev)
> > >
> > > dev_sw_netstats_tx_add(dev, 1, skb->len);
> > >
> > > - DSA_SKB_CB(skb)->clone = NULL;
> > > + memset(skb->cb, 0, 48);
> >
> > Replace hard coded 48 with sizeof() please.
>
> You mean just a trivial change like this, right?
>
> memset(skb->cb, 0, sizeof(skb->cb));
>
> And not what I had suggested in v1, which would have looked something like
> this:
>
> -----------------------------[cut here]-----------------------------
> diff --git a/include/net/dsa.h b/include/net/dsa.h index
> e1a2610a0e06..c75b249e846f 100644
> --- a/include/net/dsa.h
> +++ b/include/net/dsa.h
> @@ -92,6 +92,7 @@ struct dsa_device_ops {
> */
> bool (*filter)(const struct sk_buff *skb, struct net_device *dev);
> unsigned int overhead;
> + unsigned int skb_cb_size;
> const char *name;
> enum dsa_tag_protocol proto;
> /* Some tagging protocols either mangle or shift the destination MAC diff
> --git a/net/dsa/slave.c b/net/dsa/slave.c index 2033d8bac23d..2230596b48b7
> 100644
> --- a/net/dsa/slave.c
> +++ b/net/dsa/slave.c
> @@ -610,11 +610,14 @@ static int dsa_realloc_skb(struct sk_buff *skb, struct
> net_device *dev) static netdev_tx_t dsa_slave_xmit(struct sk_buff *skb,
> struct net_device *dev) {
> struct dsa_slave_priv *p = netdev_priv(dev);
> + const struct dsa_device_ops *tag_ops;
> struct sk_buff *nskb;
>
> dev_sw_netstats_tx_add(dev, 1, skb->len);
>
> - memset(skb->cb, 0, 48);
> + tag_ops = p->dp->cpu_dp->tag_ops;
> + if (tag_ops->skb_cb_size)
> + memset(skb->cb, 0, tag_ops->skb_cb_size);
>
> /* Handle tx timestamp if any */
> dsa_skb_tx_timestamp(p, skb);
> diff --git a/net/dsa/tag_sja1105.c b/net/dsa/tag_sja1105.c index
> 50496013cdb7..1b337fa104dc 100644
> --- a/net/dsa/tag_sja1105.c
> +++ b/net/dsa/tag_sja1105.c
> @@ -365,6 +365,7 @@ static const struct dsa_device_ops
> sja1105_netdev_ops = {
> .overhead = VLAN_HLEN,
> .flow_dissect = sja1105_flow_dissect,
> .promisc_on_master = true,
> + .skb_cb_size = sizeof(struct sja1105_skb_cb),
> };
>
> MODULE_LICENSE("GPL v2");
> -----------------------------[cut here]-----------------------------
>
> I wanted to see how badly impacted would the performance be, so I created
> an IPv4 forwarding setup on the NXP LS1021A-TSN board (gianfar +
> sja1105):
>
> #!/bin/bash
>
> ETH0=swp3
> ETH1=swp2
>
> systemctl stop ptp4l # runs a BPF classifier on every packet systemctl stop
> phc2sys
>
> echo 1 > /proc/sys/net/ipv4/ip_forward
> ip addr flush $ETH0 && ip addr add 192.168.100.1/24 dev $ETH0 && ip link
> set $ETH0 up ip addr flush $ETH1 && ip addr add 192.168.200.1/24 dev $ETH1
> && ip link set $ETH1 up
>
> arp -s 192.168.100.2 00:04:9f:06:00:09 dev $ETH0 arp -s 192.168.200.2
> 00:04:9f:06:00:0a dev $ETH1
>
> ethtool --config-nfc eth2 flow-type ether dst 00:1f:7b:63:01:d4 m ff:ff:ff:ff:ff:ff
> action 0
>
> and I got the following results on 1 CPU, 64B UDP packets (yes, I know the
> baseline results suck, I haven't investigated why that is, but nonetheless, it
> should still be relevant as far as comparative results
> go):
>
> Unpatched net-next:
> proto 17: 65695 pkt/s
> proto 17: 65725 pkt/s
> proto 17: 65732 pkt/s
> proto 17: 65720 pkt/s
> proto 17: 65695 pkt/s
> proto 17: 65725 pkt/s
> proto 17: 65732 pkt/s
> proto 17: 65720 pkt/s
>
>
> After patch 1:
> proto 17: 72679 pkt/s
> proto 17: 72677 pkt/s
> proto 17: 72669 pkt/s
> proto 17: 72707 pkt/s
> proto 17: 72696 pkt/s
> proto 17: 72699 pkt/s
>
> After patch 2:
> proto 17: 72292 pkt/s
> proto 17: 72425 pkt/s
> proto 17: 72485 pkt/s
> proto 17: 72478 pkt/s
>
> After patch 4 (as 3 doesn't build):
> proto 17: 72437 pkt/s
> proto 17: 72510 pkt/s
> proto 17: 72479 pkt/s
> proto 17: 72499 pkt/s
> proto 17: 72497 pkt/s
> proto 17: 72427 pkt/s
>
> With the change I pasted above:
> proto 17: 71891 pkt/s
> proto 17: 71810 pkt/s
> proto 17: 71850 pkt/s
> proto 17: 71826 pkt/s
> proto 17: 71798 pkt/s
> proto 17: 71786 pkt/s
> proto 17: 71814 pkt/s
> proto 17: 71814 pkt/s
> proto 17: 72010 pkt/s
>
> So basically, not only are we better off just zero-initializing the complete
> skb->cb instead of looking up the tagger's skb_cb_size, but zero-initializing the
> skb->cb isn't even all that bad. Yangbo's change is an overall win anyway, all
> things considered. So just change the memset as Richard suggested, make sure
> all patches compile, and we should be good to go.
Ah... I had thought 48bytes memset was acceptable for now by you.
I actually didn't consider the performance affected by this. I just did this because I believed the direction was right:)
That's really good testing. Thank you very much, Vladimir.
With the data, we can feel free to use the changes.
