Messages in this thread | | | Date | Tue, 30 Oct 2012 17:43:19 +0000 | Subject | Re: [PATCH] scripts/recordmcount.pl: Support build with -ffunction-sections. | From | Will Newton <> |
| |
On Tue, Oct 30, 2012 at 4:05 PM, Steven Rostedt <rostedt@goodmis.org> wrote: > On Tue, 2012-10-30 at 15:35 +0000, Will Newton wrote: >> On Tue, Oct 30, 2012 at 3:05 PM, Steven Rostedt <rostedt@goodmis.org> wrote: >> > On Tue, 2012-10-30 at 14:51 +0000, Will Newton wrote: >> >> Scan any text section whose name begins with ".text." so >> >> we will find all the functions in a kernel built with >> >> -ffunction-sections. >> > >> > A couple of things. >> > >> > First, I'm very paranoid about a blanket "ok" on sections. We must >> > guarantee that all sections that starts with ".text" never is freed. And >> > if it is freed, that it must inform ftrace that it's about to free it >> > before it does. If we can not guarantee this, then we can't do it. >> >> I share that concern, however it looks like it should be safe. > > Can we confirm that it is safe. I'm not too confident with the words > "looks like" ;-)
Ok, well we are only concerned with C compiled code, otherwise there would be no calls to mcount? The only way I can think of to cause code to be emitted in a section of this type is to use the section attribute. A quick grep for __attribute__(__section__ or __section does not turn up any sections that are named .text.* on any architecture (freed or not).
Is there any other case I have missed? I also did a grep for .text. more generally and didn't see anything that looked problematic (although it is mainly .S files so not relevant).
>> >> > Second, most archs today do not use recordmcount.pl. They use the new >> > recordmcount.c file. >> >> Ok, thanks for letting me know. Is the plan to switch over to >> recordmcount.c exclusively at some point? > > When all the archs convert.
I'll rework my patch against recordmcount.c once I have it working.
| |