Re: [PATCH] scripts: add script for translating stack dump function offsets

From: Josh Poimboeuf
Date: Fri Sep 16 2016 - 15:17:55 EST

On Fri, Sep 16, 2016 at 11:12:10AM -0700, Linus Torvalds wrote:
> On Fri, Sep 16, 2016 at 7:48 AM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > Here's something a lot faster than gdb, which also handles duplicate
> > symbols properly.
> Ack.
> Side note: I find addr2line almost completely useless in many cases
> not because of address space randomization, but because of how complex
> the inlining often is. I just had something where I decided to use
> addr2line and it just pointed me to the __read_once_size_nocheck()
> line in <linux/compiler.h>. That was not very useful.
> I ended up actually looking at the instructions *around* it, to find
> where that one instruction had been inlined from.
> So I'm wondering if this kind of helper script could be extended to
> have that "look around it" thing to help.

I think that issue is solved by addr2line's '--inline' option, which the
script uses:

$ scripts/faddr2line vmlinux show_stack_log_lvl+0x28

> Finally, I note that *if* you hit the "multiple copies of the same
> function name" issue, it might be a good idea to limit the function
> name copies by their size/offset. Also, shouldn't you filter the
> objdump for just functions Both the size and the function filtering
> should be possible with some additional awk script magic.
> For example, in my current kernel build, I have the following object
> names that are both functions and non-functions:
> event_function, irq_trigger, p_start, tbl_size, verbose, watchdog
> and I have 10 different versions of the function ("type_show()", with
> sizes ranging from 26 bytes to 166 bytes. So both the function offset
> filtering and the type filtering could definitely make a difference.

Yeah, good ideas. That would help reduce some of the false duplicates,
though they are quite rare. I'll see what I can do.