Re: [PATCH net-next] net: dev: introduce netdev_drop_inc()

February 9, 2022 11:53 AM, "Jakub Kicinski" [off-list ref] wrote:

On Wed, 09 Feb 2022 02:20:07 +0000 yajun.deng@linux.dev wrote:

quoted

February 9, 2022 8:27 AM, "Jakub Kicinski" [off-list ref] wrote:

On Tue, 8 Feb 2022 14:43:18 +0800 Yajun Deng wrote:

We will use 'sudo perf record -g -a -e skb:kfree_skb' command to trace
the dropped packets when dropped increase in the output of ifconfig.
But there are two cases, one is only called kfree_skb(), another is
increasing the dropped and called kfree_skb(). The latter is what
we need. So we need to separate these two cases.

From the other side, the dropped packet came from the core network and
the driver, we also need to separate these two cases.

Add netdev_drop_inc() and add a tracepoint for the core network dropped
packets. use 'sudo perf record -g -a -e net:netdev_drop' and 'sudo perf
script' will recored the dropped packets by the core network.

Signed-off-by: Yajun Deng <redacted>

Have you seen the work that's being done around kfree_skb_reason()?

Yes, I saw it. The focus of kfree_skb_reason() is trace kfree_skb() and the reason,
but the focus of this patch only traces this case of the dropped packet.

I don't want to trace all kfree_skb(), but I just want to trace the dropped packet.

This command 'sudo perf record -g -a -e skb:kfree_skb' would trace all kfree_skb(),
kfree_skb() would drowned out the case of dropped packets when the samples were too large.

IIRC perf support filters, I think with -f? We can't add a tracepoint
for every combination of attributes.

Yes, we can use a command like this: " sudo perf record -g -a -e skb:kfree_skb --filter 'protocol == 0x0800' ",
However, only the filter is defined in kfree_skb tracepoint are available.

The purpose of this patch is record {rx_dropped, tx_dropped, rx_nohandler} in struct net_device, to distinguish 
with struct net_device_stats. 

We don't have any tracepoint records {rx_dropped, tx_dropped, rx_nohandler} in struct net_device now. 
Can we add {rx_dropped, tx_dropped, rx_nohandler} in kfree_skb tracepoint?  like this:

        TP_STRUCT__entry(
                __field(void *,         skbaddr)
                __field(void *,         location)
                __field(unsigned short, protocol)
                __field(enum skb_drop_reason,   reason)
                __field(unsigned long,  rx_dropped)
                __field(unsigned long,  tx_dropped)
                __field(unsigned long,  rx_nohandler)

        ),

        TP_fast_assign(
                __entry->skbaddr = skb;
                __entry->location = location;
                __entry->protocol = ntohs(skb->protocol);
                __entry->reason = reason;
                __entry->rx_dropped   = (unsigned long)atomic_long_read(&skb->dev->rx_dropped);
                __entry->tx_dropped   = (unsigned long)atomic_long_read(&skb->dev->tx_dropped);
                __entry->rx_nohandler = (unsigned long)atomic_long_read(&skb->dev->rx_nohandler);
        ),

If so, we can record this but not add a tracepoint.