[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20190212183222.GN23599@brightrain.aerifal.cx>
Date: Tue, 12 Feb 2019 13:32:22 -0500
From: Rich Felker <dalias@...c.org>
To: musl@...ts.openwall.com
Subject: Re: [PATCH] powerpc64: use a type for mcontext_t regs field
On Tue, Feb 12, 2019 at 01:24:09PM -0500, David Edelsohn wrote:
> On Tue, Feb 12, 2019 at 1:21 PM Rich Felker <dalias@...c.org> wrote:
> >
> > On Tue, Feb 12, 2019 at 01:17:34PM -0500, David Edelsohn wrote:
> > > On Tue, Feb 12, 2019 at 12:23 PM Rich Felker <dalias@...c.org> wrote:
> > > >
> > > > On Tue, Feb 12, 2019 at 12:05:17PM -0500, David Edelsohn wrote:
> > > > > On Tue, Feb 12, 2019 at 11:18 AM Rich Felker <dalias@...c.org> wrote:
> > > > > >
> > > > > > On Tue, Feb 12, 2019 at 09:35:22AM -0600, A. Wilcox wrote:
> > > > > > > GCC Go dereferences `regs` for `nip`. Without this change, compilation
> > > > > > > fails with the following message:
> > > > > > >
> > > > > > > .../../../libgo/runtime/go-signal.c: In function ‘getSiginfo’:
> > > > > > > .../../../libgo/runtime/go-signal.c:225:56: warning: dereferencing ‘void *’ pointer
> > > > > > > ret.sigpc = ((ucontext_t*)(context))->uc_mcontext.regs->nip;
> > > > > > > ^~
> > > > > > > .../../../libgo/runtime/go-signal.c:225:56: error: request for member ‘nip’ in something not a structure or union
> > > > > > > ---
> > > > > > > arch/powerpc64/bits/signal.h | 4 +++-
> > > > > > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > > > > > >
> > > > > > > diff --git a/arch/powerpc64/bits/signal.h b/arch/powerpc64/bits/signal.h
> > > > > > > index 34693a68..6736c69a 100644
> > > > > > > --- a/arch/powerpc64/bits/signal.h
> > > > > > > +++ b/arch/powerpc64/bits/signal.h
> > > > > > > @@ -8,6 +8,8 @@
> > > > > > >
> > > > > > > #if defined(_GNU_SOURCE) || defined(_BSD_SOURCE)
> > > > > > >
> > > > > > > +#include <bits/user.h>
> > > > > > > +
> > > > > >
> > > > > > This is almost surely not right. You can never use a bits/* header
> > > > > > except from the corresponding public header that includes it. They do
> > > > > > not have multiple-inclusion guards and are not designed to be usable
> > > > > > independently. And including sys/user.h is also almost surely wrong
> > > > > > since it's a problematic header you don't want getting included
> > > > > > (conflicts with linux/*.h stuff).
> > > > > >
> > > > > > > typedef unsigned long greg_t, gregset_t[48];
> > > > > > >
> > > > > > > typedef struct {
> > > > > > > @@ -29,7 +31,7 @@ typedef struct sigcontext {
> > > > > > > int _pad0;
> > > > > > > unsigned long handler;
> > > > > > > unsigned long oldmask;
> > > > > > > - void *regs;
> > > > > > > + struct pt_regs *regs;
> > > > > > > gregset_t gp_regs;
> > > > > > > fpregset_t fp_regs;
> > > > > > > vrregset_t *v_regs;
> > > > > > > --
> > > > > > > 2.19.2
> > > > > >
> > > > > > Do you know if there's a reason Go is trying to use this regs member
> > > > > > rather than gp_regs? I think that's the real issue here.
> > > > >
> > > > > Apparently GCCGo runtime has dependencies on GLibc and has not been
> > > > > ported to Musl Libc. The GCCGo runtime is trying to obtain the PC for
> > > > > a signal.
> > > > >
> > > > > #ifdef __PPC__
> > > > > #ifdef __linux__
> > > > > ret.sigpc = ((ucontext_t*)(context))->uc_mcontext.regs->nip;
> > > > > #endif
> > > > > #ifdef _AIX
> > > > > ret.sigpc = ((ucontext_t*)(context))->uc_mcontext.jmp_context.iar;
> > > > > #endif
> > > > > #endif
> > > > >
> > > > > The program counter is not part of gp_regs.
> > > >
> > > > Yes it is, as musl already depends on having access to the program
> > > > counter for thread cancellation. See arch/powerpc64/pthread_arch.h:
> > > >
> > > > // the kernel calls the ip "nip", it's the first saved value after the 32
> > > > // GPRs.
> > > > #define MC_PC gp_regs[32]
> > >
> > > Apparently Ian Taylor now is aware of this and investigating a
> > > solution -- possibly another ifdef in the above for PPC Musl Libc.
> >
> > This is not a public #define that you could #ifdef on. It's part of
> > musl thread implementation internals and only present in the musl
> > source tree, not anywhere installed. I just showed it to demonstrate
> > that the NIP register (as well as 15 other registers that are just not
> > the 32 numbered GPRs) is part of the gp_regs array.
>
> I meant that it appears this should be fixed in GCC Go runtime --
> somehow -- not in Musl Libc referring to nip. GCC Go apparently will
> need another case for Musl Libc and somehow know to reference
> gp_regs[32], or some other way to know about the internals of Musl
> Libc and access the program counter.
This is not knowledge of musl internals. Both accessing a pt_regs
struct pointed to by the regs member, and accessing gp_regs[32], are
part of the kernel ABI for powerpc64 mcontext_t. musl simply does not
expose the struct pt_regs type here because there's no safe way to get
a definition that doesn't clash with kernel headers, so I suggested
using the alternate way that *also works on glibc*. I can't see any
reason to make this change conditional on musl.
I think it would also be acceptable to change the void * to struct
pt_regs * in musl's mcontext_t, but then it would be a pointer to
incomplete type unless you also include sys/user.h. Generally, user.h
is an awful header you want to avoid unless you have no choice,
though, so I think it's highly preferable for gccgo *not* to use it,
even if we do make this change, and instead use the gp_regs NIP slot
that avoids touching anything from user.h.
Rich
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
