On Wed, Aug 21, 2024 at 04:32:46PM +0100, Mark Rutland wrote:
> On Wed, Aug 21, 2024 at 07:05:39AM +0900, Masami Hiramatsu wrote:
> > On Tue, 20 Aug 2024 08:10:42 -0700
> > Sami Tolvanen <samitolva...@google.com> wrote:
> >
> > > On Tue, Aug 20, 2024 at 3:48 AM Mark Rutland <mark.rutl...@arm.com> wrote:
> > > >
> > > > On Tue, Aug 20, 2024 at 10:03:30AM +0900, Masami Hiramatsu wrote:
> > > > > On Mon, 19 Aug 2024 12:02:44 -0400
> > > > > Steven Rostedt <rost...@goodmis.org> wrote:
> > > > >
> > > > > > On Tue, 20 Aug 2024 00:56:49 +0900
> > > > > > Masami Hiramatsu (Google) <mhira...@kernel.org> wrote:
> > > > > > >
> > > > > > > >
> > > > > > > > We may need to add "noinline" or something to make sure those
> > > > > > > > functions
> > > > > > > > don't get inlined for LTO.
> > > > > > >
> > > > > > > Yeah, we need such option at least for function call test.
> > > > > >
> > > > > > Could you add the noinline, and if it fixes the issue send a patch?
> > > > >
> > > > > I found the target function already has "noinline". I tried to add
> > > > > noinline
> > > > > to the testing function (callsite), but it also did not work.
> > > > > I think "noinline" is for the compiler, but LTO is done by the linker.
> > > >
> > > > If LTO is breaking noinline, then that has much larger implications for
> > > > noinstr code and similar, and means that LTO is unsound...
> > >
> > > The noinline attribute is preserved in LLVM IR, so it should continue
> > > to work with LTO. Which function are we talking about here? Are you
> > > sure the function was inlined instead of being dropped completely?
> > > Does marking the function __used help?
> >
> > We are talking about trace_selftest_startup_dynamic_tracing() in
> > kernel/trace/trace_selftest.c. The callee is func() which is actually
> > DYN_FTRACE_TEST_NAME() in kernel/trace/trace_selftest_dynamic.c.
> > That function passed as pointer (but the compiler can embed it by constant
> > propagation.)
>
> Ah, so IIUC the function isn't being inlined; the call is being
> optimized away becase callee() has no side-effects.
>
> That can happen without LTO if the caller is in the same compilation
> unit, and I have worked around that in the past by adding a barrier()
> into the callee.
FWIW, that was in samples/ftrace/ftrace-ops.c, where tracee_relevant() and
tracee_irrelevant() have the barrier():
| /*
| * Marked as noinline to ensure that an out-of-line traceable copy is
| * generated by the compiler.
| *
| * The barrier() ensures the compiler won't elide calls by determining there
| * are no side-effects.
| */
| static noinline void tracee_relevant(void)
| {
| barrier();
| }
... so we already have precedent for that in tracing code.
Mark.
