Re: [PATCH net-next] udp: Add tracepoint for udp_sendmsg()

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Breno Leitao wrote:
> Hello Willem,
> 
> On Wed, Apr 16, 2025 at 03:34:38PM -0400, Willem de Bruijn wrote:
> > Breno Leitao wrote:
> > > Add a lightweight tracepoint to monitor UDP send message operations,
> > > similar to the recently introduced tcp_sendmsg_locked() trace event in
> > > commit 0f08335ade712 ("trace: tcp: Add tracepoint for
> > > tcp_sendmsg_locked()")
> > > 
> > > This implementation uses DECLARE_TRACE instead of TRACE_EVENT to avoid
> > > creating extensive trace event infrastructure and exporting to tracefs,
> > > keeping it minimal and efficient.
> > > 
> > > Since this patch creates a rawtracepoint, it can be accessed using
> > > standard tracing tools like bpftrace:
> > > 
> > >     rawtracepoint:udp_sendmsg_tp {
> > >         ...
> > >     }
> > 
> > What does this enable beyond kfunc:udp_sendmsg?
> 
> A few things come to mind when evaluating the use of tracepoints.
> 
> One significant advantage is that tracepoints provide a stable API where
> programs can hook into, making it easier for users to interact with key
> functions.
> 
> However, kfunc/kprobes has some notable disadvantages. For instance,
> partial or total inlining can cause hooks to fail, which is not ideal
> and can lead to issues (mainly when we have partial inlines, and the
> hook works _sometimes_).

As Paolo explained, that is unlikely to happen in this case, as this
is a protocol specific callback function.
 
> In contrast, tracepoints create a more stable API for users to hook
> into, eliminating the need to patch the kernel with noinline function
> attributes. This solution may be ugly, but it is a common practice.
> (and this is my main goal with it, remove `noinline` from our internal
> kernel)
> 
> While tracepoints are not officially considered stable APIs, they tend
> to be "more stable" in practice due to their deliberate and strategic
> placement. As a result, they are less likely to get renamed or changed
> frequently.
> 
> Additionally, tracepoints offer performance benefits, being faster than
> both kfunc and kprobes. 

The performance argument is fair.

Perhaps we want to think this through more broadly for networking
tracepoints vs more flexible kprobes/kfuncs, rather than on a case
by case basis:

Where do we think the performance or functionality (if exposing
additional info, as for tcp_sendmsg) warrants the tracepoint?

I suspect that the use is predominantly for on-demand debugging,
where the performance cost (and latency impact) of measurement is
minor.
 
> For further discussion on this topic, please refer to same discussion in
> VFS:
> 
> https://lore.kernel.org/bpf/20250118033723.GV1977892@ZenIV/T/#m4c2fb2d904e839b34800daf8578dff0b9abd69a0
> 
> Thanks
> --breno