Re: [PATCH] parisc: prefer _THIS_IP_ and _RET_IP_ statement expressions

From: Nick Desaulniers
Date: Tue Aug 07 2018 - 14:12:11 EST


On Fri, Aug 3, 2018 at 3:34 PM Helge Deller <deller@xxxxxx> wrote:
>
> On 03.08.2018 22:33, Nick Desaulniers wrote:
> > On Fri, Aug 3, 2018 at 12:09 PM John David Anglin <dave.anglin@xxxxxxxx> wrote:
> >>
> >> On 2018-08-03 2:11 PM, Nick Desaulniers wrote:
> >>> But the kernel uses the generic_THIS_IP_ *everywhere*, not parisc's
> >>> custom current_text_addr(). So if this did actually break unwinding,
> >>> you should have noticed by now.
> >> The unwind problem was noticed.
> >
> > So parisc is currently broken (doesn't unwind) due to the pervasive
> > use of _THIS_IP_ (generic C) throughout the kernel?
>
> I tested it on the 32bit kernel.
> The answer is: No. Unwinding works (with and without your patch).
>
> > If no, that implies this patch (generic C) causes no unwinding problems.
>
> correct.
>
> > If yes, that implies that the diff I posted later in this thread
> > (inline assembly) is preferable, and that parisc has bigger problems
> > (and probably needs to do rewrite the unwinding code to handle these
> > extra labels everywhere).
> >
> >> Patches were recently applied to gcc and binutils to try and fix it.
> >> The gcc patch moved
> >> branch tables to rodata so that the label at the head of the table
> >> wasn't in text.
> >>
> >> https://gcc.gnu.org/ml/gcc-patches/2018-07/msg01804.html
> >> https://sourceware.org/ml/binutils/2018-07/msg00474.html
> >>
> >> When I saw your suggested change, I realized there was another source of
> >> text labels
> >> that need linker relocations.
> >
> > Thank you for the links.
> >
> > On Fri, Aug 3, 2018 at 10:57 AM John David Anglin <dave.anglin@xxxxxxxx> wrote:
> >> The label breaks the unwind data, not the unwind code. So, localizing
> >> the use of
> >> current_text_addr() to the parisc unwind code doesn't help.
> >
> > Have you confirmed that applying my patch breaks *the ability to
> > unwind correctly*?
>
> I tested your patch (on 32bit).
> Your patch does not break anything.
>
> > It looks like return_address() is used in
> > ftrace_return_address(), so I assume you can boot a kernel with my
> > patch applied, and CONFIG_FTRACE=y, then run:
> >
> > $ sudo trace-cmd record -p function date
> > $ trace-cmd report | grep date- | less
> >
> > and see if the stacks aren't unwound or look messed up.
>
> I faced issues with trace-cmd, but calling ftracing functions manually worked.
>
> So, your patch is basically OK and doesn't break anything.
> But I agree with Dave that Andrew, that THIS_IP is ugly.

I don't disagree, and other maintainers have remarked on _THIS_IP_
being ugly, but renaming it en masse is a tree wide change, which I'm
trying to avoid at the moment.

It sounds like we have a working patch? Are there 64b parisc machines
to test on, or can this get merged?
--
Thanks,
~Nick Desaulniers