Re: [GIT PULL] tracing: Updates for 5.10

[Masami Hiramatsu]

On Thu, 15 Oct 2020 22:21:39 -0400
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Thu, 15 Oct 2020 18:54:34 -0700
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Thu, Oct 15, 2020 at 10:53 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > >
> > > Updates for tracing and bootconfig:  
> > 
> > Hmm. I haven't verified that this came from you, but it seems likely..
> > Once again my clang build shows something that I don't see in my
> > allmodconfig gcc build:
> > 
> >    WARNING: modpost: vmlinux.o(.text+0x1e5b06): Section mismatch in
> > reference from the function __trace_early_add_events() to the function
> > .init.text:__trace_early_add_new_event()
> >    The function __trace_early_add_events() references
> >    the function __init __trace_early_add_new_event().
> >    This is often because __trace_early_add_events lacks a __init
> >    annotation or the annotation of __trace_early_add_new_event is wrong.
> > 
> > Hmm?
> > 
> >                Linus
> 
> I see the issue, and I wonder if it has to do with optimization, for gcc
> not to warn.

I also couldn't reproduce it even with CONFIG_DEBUG_SECTION_MISMATCH=y.
It seems that the __trace_early_add_new_event() is inlined in 
__trace_early_add_events(). 

$ eu-readelf -w kernel/trace/trace_events.o
...
 [ 1af20]    subprogram           abbrev: 53
             name                 (strp) "__trace_early_add_new_event"
             decl_file            (data1) trace_events.c (1)
             decl_line            (data2) 2502
             decl_column          (data1) 1
             prototyped           (flag_present) yes
             type                 (ref4) [    cc]
             inline               (data1) inlined (1)
$ gcc -v
...
gcc version 9.3.0 (Ubuntu 9.3.0-17ubuntu1~20.04) 

> 
> The issue is that we have:
> 
> trace_array_create() that can be called at any time. And it has:
> 
> 	if (trace_instance_dir) {
> 		ret = trace_array_create_dir(tr);
> 		if (ret)
> 			goto out_free_tr;
> 	} else
> 		__trace_early_add_events(tr);
> 
> 
> Where trace_instance_dir gets set at boot up, and thus the else statement
> will never get called after that.
> 
> The __trace_early_add_events() then calls __trace_early_add_new_events()
> which is __init.
> 
> I don't know how gcc didn't trigger this and clang does.

> 
> I'll have to think about how to untangle this. Is there some kind of
> annotation that makes it show that a path can only be called at boot up and
> not later?

What happen if we use Peter's static_call() and update it after boot up? 
Or, we might need to break apart the trace_array_create() and restruct
it as __init trace_array_early_create() and trace_array_create().

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>