Re: [musl] SA_RESTORER for rv64?

[enh <enh@google.com>]

On Sun, Feb 5, 2023 at 3:54 PM Rich Felker <dalias@libc.org> wrote:
>
> On Fri, Feb 03, 2023 at 10:44:56AM -0800, enh wrote:
> > oops, never actually sent the patch. attached...
> >
> > On Thu, Nov 10, 2022 at 9:31 AM Khem Raj <raj.khem@gmail.com> wrote:
> > >
> > > On Thu, Nov 10, 2022 at 9:19 AM Rich Felker <dalias@libc.org> wrote:
> > > >
> > > > On Thu, Nov 10, 2022 at 07:44:23AM -0800, enh wrote:
> > > > > arch/riscv64/bits/signal.h has contained a definition for SA_RESTORER since
> > > > > the initial commit, but i think that's just copy & paste from whichever
> > > > > architecture the rv64 headers were based on? the linux kernel itself
> > > > > doesn't have SA_RESTORER for rv64, unless i'm missing something?
> > > >
> > > > I suspect this is just a mistake. Have you seen any ill effects from
> > > > it? If riscv folks can confirm it's wrong, I'll remove it.
> > >
> > > Yeah I think it should be removed. Perhaps mips is in same boat.
> > >
> > > >
> > > > Rich
>
> > From 6413de6d9f785c98e5bc0cf40be947f1169d2fd7 Mon Sep 17 00:00:00 2001
> > From: Elliott Hughes <enh@google.com>
> > Date: Fri, 3 Feb 2023 10:42:55 -0800
> > Subject: [PATCH] risc-v does not have SA_RESTORER.
> >
> > The kernel's include/uapi/asm-generic/signal-defs.h explicitly calls
> > this out as obsolete. New architectures like risc-v do not define it.
> > ---
> >  arch/riscv64/bits/signal.h | 1 -
> >  1 file changed, 1 deletion(-)
> >
> > diff --git a/arch/riscv64/bits/signal.h b/arch/riscv64/bits/signal.h
> > index 287367db..fd6157a3 100644
> > --- a/arch/riscv64/bits/signal.h
> > +++ b/arch/riscv64/bits/signal.h
> > @@ -76,7 +76,6 @@ typedef struct __ucontext
> >  #define SA_RESTART   0x10000000
> >  #define SA_NODEFER   0x40000000
> >  #define SA_RESETHAND 0x80000000
> > -#define SA_RESTORER  0x04000000
> >
> >  #endif
> >
> > --
> > 2.39.1.519.gcb327c4b5f-goog
> >
>
> I don't think this patch works as-is, since musl unconditionally uses
> SA_RESTORER. We probably need to make that conditional on its
> presence, and it looks like there's also a wrong-struct-layout issue
> on archs where it's absent...

yeah, bionic just uses the kernel uapi headers directly, and they look
like this:

struct sigaction {
  __sighandler_t sa_handler;
  unsigned long sa_flags;
#ifdef SA_RESTORER
  __sigrestore_t sa_restorer;
#endif
  sigset_t sa_mask;
};


> Rich