Re: [for-next][PATCH 10/31] scripts: ftrace - move the sort-processing in ftrace_init

[Nathan Chancellor]

On Sat, Jan 15, 2022 at 10:59:20PM -0500, Steven Rostedt wrote:
> On Sat, 15 Jan 2022 13:36:04 -0700
> Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
> 
> > Hi Steven and Yinan,
> > 
> > On Tue, Jan 11, 2022 at 12:30:41PM -0500, Steven Rostedt wrote:
> > > From: Yinan Liu <yinan@xxxxxxxxxxxxxxxxx>
> > > 
> > > When the kernel starts, the initialization of ftrace takes
> > > up a portion of the time (approximately 6~8ms) to sort mcount
> > > addresses. We can save this time by moving mcount-sorting to
> > > compile time.
> > > 
> > > Link: https://lkml.kernel.org/r/20211212113358.34208-2-yinan@xxxxxxxxxxxxxxxxx
> > > 
> > > Signed-off-by: Yinan Liu <yinan@xxxxxxxxxxxxxxxxx>
> > > Reported-by: kernel test robot <lkp@xxxxxxxxx>
> > > Reported-by: kernel test robot <oliver.sang@xxxxxxxxx>
> > > Signed-off-by: Steven Rostedt <rostedt@xxxxxxxxxxx>  
> > 
> > This change as commit 72b3942a173c ("scripts: ftrace - move the
> > sort-processing in ftrace_init") in -next causes a bunch of warnings at
> > the beginning of the build when using clang as the host compiler:
> > 
> 
> 
> > 
> > Should mcount_sort_thread be zero initialized or is there something else
> > going on here? I am currently hunting down a bunch of other regressions
> > so apologies for just the report rather than a patch to fix it.
> 
> Can this really happen? We have:

The way the code is written now, yes.

>         if (pthread_create(&mcount_sort_thread, NULL, &sort_mcount_loc, &mstruct)) {
>                 fprintf(stderr,
>                         "pthread_create mcount_sort_thread failed '%s': %s\n",
>                         strerror(errno), fname);
>                 goto out;
>         }
> [..]
> 
>         if (mcount_sort_thread) {
>                 void *retval = NULL;
>                 /* wait for mcount sort done */
>                 rc = pthread_join(mcount_sort_thread, &retval);
>                 if (rc) {
>                         fprintf(stderr,
>                                 "pthread_join failed '%s': %s\n",
>                                 strerror(errno), fname);
>                 } else if (retval) {
>                         rc = -1;
>                         fprintf(stderr,
>                                 "failed to sort mcount '%s': %s\n",
>                                 (char *)retval, fname);
>                 }
>         }
> 
> Shouldn't the pthread_create() initialize it? And I'm not even sure if we
> need that if statement?
> 
> Or is there a path to get there without pthread_create() initializing it?

Yes. If the if statment right above the pthread_create() call triggers,
we jump to the out label, which hits the if (mcount_sort_thread), and
mcount_sort_thread won't be initialized.

	if (!mstruct.init_data_sec || !_start_mcount_loc || !_stop_mcount_loc) {
		fprintf(stderr,
			"incomplete mcount's sort in file: %s\n",
			fname);
		goto out;
	}

	if (pthread_create(&mcount_sort_thread, ...)) {
...

out:
...
	if (mcount_sort_thread) {

If I am misunderstanding something, please let me know.

Cheers,
Nathan