Re: [PATCH] kallsyms: ignore local block labels generated by compiler

[Changbin Du]

On Fri, Jan 28, 2022 at 10:59:55AM -0800, Nick Desaulniers wrote:
> On Fri, Jan 28, 2022 at 7:21 AM Changbin Du <changbin.du@xxxxxxxxx> wrote:
> >
> > On Fri, Jan 28, 2022 at 06:57:46PM +0800, Changbin Du wrote:
> > > The llvm compiler can generate lots of local block labels and they might
> > > overlap with C defined symbols. So let's ignore such local labels.
> > >
> > > Before this change, dumpstack shows a local symbol for epc:
> > > [    0.040341][    T0] Hardware name: riscv-virtio,qemu (DT)
> > > [    0.040376][    T0] epc : .LBB6_14+0x22/0x6a
> > > [    0.040452][    T0]  ra : restore_all+0x12/0x6e
> > >
> > > After this change, the C defined symbol is shown which is expected:
> > > [    0.035795][    T0] Hardware name: riscv-virtio,qemu (DT)
> > > [    0.036332][    T0] epc : trace_hardirqs_on+0x54/0x13c
> > > [    0.036567][    T0]  ra : restore_all+0x12/0x6e
> > >
> > > Signed-off-by: Changbin Du <changbin.du@xxxxxxxxx>
> > > ---
> > >  scripts/kallsyms.c | 1 +
> > >  1 file changed, 1 insertion(+)
> > >
> > > diff --git a/scripts/kallsyms.c b/scripts/kallsyms.c
> > > index 54ad86d13784..5f4be9d72a32 100644
> > > --- a/scripts/kallsyms.c
> > > +++ b/scripts/kallsyms.c
> > > @@ -108,6 +108,7 @@ static bool is_ignored_symbol(const char *name, char type)
> > >       /* Symbol names that begin with the following are ignored.*/
> > >       static const char * const ignored_prefixes[] = {
> > >               "$",                    /* local symbols for ARM, MIPS, etc. */
> > > +             ".LBB",                 /* local block labels generated by compiler */
> > I aslo found many symbols like '.Ltmpxxx', '.L__unnamed_xx'. So should we just
> > ignore all symbols prefixed by '.L'?
> 
> .L prefixes are "local labels" meaning they don't get an entry in the
> symbol table.  The trade off is that there is no such convention
> between compiler generated local label naming conventions, vs hand
> written ones.  If we omit all local labels, then it might be
> surprising that a handwritten local label disappeared from this list.
> I think though that we should omit all local labels, and if someone
> needs a label to appear in their dumpstack, then they should simply
> drop the .L prefix on their handwritten label.
>
Agree. I checked vmlinux for riscv and x86, seeing no handwritten function
labels prefixed by '.L'. (just some data definition symbols). It should be
safte to simply ignore all these symbols.

> >
> > >               ".LASANPC",             /* s390 kasan local symbols */
> > >               "__crc_",               /* modversions */
> > >               "__efistub_",           /* arm64 EFI stub namespace */
> > > --
> > > 2.32.0
> > >
> >
> > --
> > Cheers,
> > Changbin Du
> 
> 
> 
> -- 
> Thanks,
> ~Nick Desaulniers

-- 
Cheers,
Changbin Du