On Wed, Oct 23, 2024 at 11:43 PM Eric Dumazet [off-list ref] wrote:

On Wed, Oct 23, 2024 at 4:35 PM Yafang Shao [off-list ref] wrote:

quoted

On Wed, Oct 23, 2024 at 9:01 PM Eric Dumazet [off-list ref] wrote:

quoted

On Wed, Oct 23, 2024 at 2:33 PM Yafang Shao [off-list ref] wrote:

quoted

We previously hooked the tcp_drop_reason() function using BPF to monitor
TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC
11, tcp_drop_reason() is now inlined, preventing us from hooking into it.
To address this, it would be beneficial to introduce a dedicated tracepoint
for monitoring.

This patch would require changes in user space tracers.
I am surprised no one came up with a noinline variant.

__bpf_kfunc is using

#define __bpf_kfunc __used __retain noinline

I would rather not have include/trace/events/tcp.h becoming the
biggest file in TCP stack...

I’d prefer not to introduce a new tracepoint if we can easily hook it
with BPF. Does the following change look good to you?

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 092456b8f8af..ebea844cc974 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c

@@ -4720,7 +4720,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk,
        return res;
 }

-static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
+noinline static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
                            enum skb_drop_reason reason)
 {
        sk_drops_add(sk, skb);

I would suggest adding an explicit keyword, like the one we have for
noinline_for_stack, for documentation purposes.

noinline_for_tracing perhaps ?

Good suggestion! This approach eliminates the need to add comments for
each noinline.

--
Regards
Yafang