Re: [PATCH] objtool: Fix STACK_FRAME_NON_STANDARD reloc type

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Apr 29, 2022 at 03:56:36PM -0700, Josh Poimboeuf wrote:
> On Fri, Apr 29, 2022 at 02:00:44PM +0200, Peter Zijlstra wrote:
> > On Fri, Apr 29, 2022 at 11:20:24AM +0200, Peter Zijlstra wrote:
> > > 
> > > STACK_FRAME_NON_STANDARD results in inconsistent relocation types
> > > depending on .c or .S usage:
> > > 
> > >   Relocation section '.rela.discard.func_stack_frame_non_standard' at offset 0x3c01090 contains 5 entries:
> > >   Offset             Info             Type               Symbol's Value  Symbol's Name + Addend
> > >   0000000000000000  00020c2200000002 R_X86_64_PC32          0000000000047b40 do_suspend_lowlevel + 0
> > >   0000000000000008  0002461e00000001 R_X86_64_64            00000000000480a0 machine_real_restart + 0
> > >   0000000000000010  0000001400000001 R_X86_64_64            0000000000000000 .rodata + b3d4
> > >   0000000000000018  0002444600000002 R_X86_64_PC32          00000000000678a0 __efi64_thunk + 0
> > >   0000000000000020  0002659d00000001 R_X86_64_64            0000000000113160 __crash_kexec + 0
> > 
> > So that weird .rodata entry is optprobe_template_func.
> > 
> > It being in .rodata also means it's not validated and there is no ORC
> > data generated, is that all intentional? The changelog for:
> > 
> >   877b145f0f47 ("x86/kprobes: Move trampoline code into RODATA")
> > 
> > doesn't really say anything useful about any of that :/
> > 
> > I also don't see any kprobe/optprobe hooks in unwind.h, so what happens
> > if we hit an optprobe?
> 
> Same as for any other generated code, the unwinder will try to fall back
> to frame pointers, and if that doesn't work, the unwind stops.
> 
> That commit didn't change anything since it was already not being
> directly executed anyway, but rather used to generate code on the fly.
> 
> And before that commit it was being ignored by ORC anyway, thanks to
> STACK_FRAME_NON_STANDARD.  Which can now be removed since this code is
> now data and objtool will no longer try to understand it.

Right; but I suppose I'm wondering if we should fix this. It seems a
rather sub-optimal state of affairs.