Re: [RFC PATCH 4/5] tracing: Remove conditional locking from __DO_TRACE()

From: Przemek Kitszel
Date: Mon Nov 25 2024 - 10:35:58 EST


On 11/25/24 15:26, Peter Zijlstra wrote:
On Mon, Nov 25, 2024 at 09:18:18AM -0500, Mathieu Desnoyers wrote:
On 2024-11-23 12:38, Linus Torvalds wrote:

I tried the following alteration to the code, which triggers an
unexpected compiler warning on master, but not on v6.12. I suspect
this is something worth discussing:

static inline void trace_##name(proto) \
{ \
if (static_branch_unlikely(&__tracepoint_##name.key)) { \
if (cond) \
scoped_guard(preempt_notrace) \
__DO_TRACE_CALL(name, TP_ARGS(args)); \

So coding style would like braces here for it being multi-line. As
opposed to C that only mandates it for multi-statement. And then the
problem doesn't occur.

} \
if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) { \
WARN_ONCE(!rcu_is_watching(), \
"RCU not watching for tracepoint"); \
} \
}


I suspect this is caused by the "else" at the end of the __scoped_guard() macro:

#define __scoped_guard(_name, _label, args...) \
for (CLASS(_name, scope)(args); \
__guard_ptr(_name)(&scope) || !__is_cond_ptr(_name); \
({ goto _label; })) \
if (0) { \
_label: \
break; \
} else

#define scoped_guard(_name, args...) \
__scoped_guard(_name, __UNIQUE_ID(label), args)

AFAIU this is a new warning introduced by

commit fcc22ac5baf ("cleanup: Adjust scoped_guard() macros to avoid potential warning")

Yeah,.. So strictly speaking the code is fine, but the various compilers
don't like it when that else dangles :/

At one point I had a version that did:
if (0)
label: ;
else
for (....)

but it is goto-jumping back in the code
https://lore.kernel.org/netdev/20241001145718.8962-1-przemyslaw.kitszel@xxxxxxxxx/#t

I could switch to it again to reduce noise like this problem, but such
change would be to essentially allow bad formatting