[RFC PATCH net-next] net: use a dedicated tracepoint for kfree_skb_list()

[RFC PATCH net-next] net: use a dedicated tracepoint for kfree_skb_list()

[Davide Caratti]

kfree_skb_list() calls kfree_skb(), thus triggering as many dropwatch
events as the number of skbs in the list. This can disturb the analysis
of packet drops, e.g. with fragmented echo requests generated by ICMP
sockets, or with regular SCTP packets: when consume_skb() frees them,
the kernel's drop monitor may wrongly account for several packet drops:

 consume skb()
   skb_release_data()
     kfree_skb_list()
       kfree_skb() <-- false dropwatch event

don't call kfree_skb() when freeing a skb list, use a dedicated
tracepoint instead. By printing "cur" and "next", it also becomes
possible to reconstruct the skb list from its members.

Signed-off-by: Davide Caratti <dcaratti@redhat.com>
---
 include/trace/events/skb.h | 19 +++++++++++++++++++
 net/core/skbuff.c          |  6 +++++-
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/include/trace/events/skb.h b/include/trace/events/skb.h
index 9e92f22eb086..b16e3544bbbe 100644
--- a/include/trace/events/skb.h
+++ b/include/trace/events/skb.h
@@ -51,6 +51,25 @@ TRACE_EVENT(consume_skb,
 	TP_printk("skbaddr=%p", __entry->skbaddr)
 );
 
+TRACE_EVENT(kfree_skb_list,
+
+	TP_PROTO(struct sk_buff *cur, struct sk_buff *next),
+
+	TP_ARGS(cur, next),
+
+	TP_STRUCT__entry(
+		__field(	void *,	cur_addr	)
+		__field(	void *,	next_addr	)
+	),
+
+	TP_fast_assign(
+		__entry->cur_addr = cur;
+		__entry->next_addr = next;
+	),
+
+	TP_printk("cur=%p next=%p", __entry->cur_addr, __entry->next_addr)
+);
+
 TRACE_EVENT(skb_copy_datagram_iovec,
 
 	TP_PROTO(const struct sk_buff *skb, int len),
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 1ba8f0163744..7ed6bfc5dfd0 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -702,7 +702,11 @@ void kfree_skb_list(struct sk_buff *segs)
 	while (segs) {
 		struct sk_buff *next = segs->next;
 
-		kfree_skb(segs);
+		if (!skb_unref(segs))
+			continue;
+
+		trace_kfree_skb_list(segs, next);
+		__kfree_skb(segs);
 		segs = next;
 	}
 }
-- 
2.26.2

Re: [RFC PATCH net-next] net: use a dedicated tracepoint for kfree_skb_list()

[Marcelo Ricardo Leitner]

On Fri, Oct 23, 2020 at 10:52:14PM +0200, Davide Caratti wrote:
> kfree_skb_list() calls kfree_skb(), thus triggering as many dropwatch
> events as the number of skbs in the list. This can disturb the analysis
> of packet drops, e.g. with fragmented echo requests generated by ICMP
> sockets, or with regular SCTP packets: when consume_skb() frees them,
> the kernel's drop monitor may wrongly account for several packet drops:
> 
>  consume skb()
>    skb_release_data()
>      kfree_skb_list()
>        kfree_skb() <-- false dropwatch event

Seems the problem lies with skb_release_data() calling
kfree_skb_list() while it should have been a, say, consume_skb_list(),
and not generate further kfree_skb calls.

Maybe a bool parameter on skb_release_data to signal that it should
call consume_skb_list (which doesn't exist) instead?

> 
> don't call kfree_skb() when freeing a skb list, use a dedicated
> tracepoint instead. By printing "cur" and "next", it also becomes
> possible to reconstruct the skb list from its members.

I like the new probe alone. It helps to have more visibility on drops
such as those from __dev_xmit_skb() and how they happen.

But as a solution to the problem stated, seems it can be confusing.
Say one is debugging a tx drop issue. AFAICT one would have to watch
both probe points anyway, as the drop could be on a layer below than
SCTP. So I'm not seeing how it helps much, other than possibly causing
drop_watch to miss drops (by not listening to the new trace point).

  Marcelo