(mass filing with linux-libc-dev and libc6-dev)

/usr/include/asm/stat.h:        long            __unused[3];
/usr/include/linux/icmp.h:              __be16  __unused;
/usr/include/linux/sysctl.h:    unsigned long __unused[4];

These conflict with the traditional use of __unused by the BSDs, which
predates use of __unused as identifier by Linux and glibc headers.

For example, the Debian package libbsd-dev includes the following code:

#ifndef __unused
# ifdef __GNUC__
#  define __unused __attribute__((unused))
# else
#  define __unused
# endif
#endif

It however was ifdef’d out due to this problem. A personal package of
mine, mirmake, fails to build on more recent Debian systems because it
does _not_ ifdef it out, since otherwise, some BSD software cannot be
compiled (without extra patches).

Please file this bug with your respective upstreams *and* patch the
header files in question yourselves, to bridge over the response time
waiting for upstream to fix it.

Thanks!

Hi package maintainers,

maybe you can do such a thing during package installation?
This would tremendously help porting software (liberal in
what one accepts), such as NetBSD® makefs (ITP: #538171 –
for now I worked around the issue in it).


Joseph S. Myers dixit:

I do have a suggestion. Maybe we can get into an agreement still…

How about you use the following sed expression on installation?
This way you’d only have to change the Makefile.in in question.

s/[[:<:]]__unused[[:>:]]/&_/g

I do not know if your sed(1) does [[:<:]] and [[:>:]], ours does,
see http://www.mirbsd.org/man7/re_format – if not, something si-
milar would probably be possible.

As __unused* should never be used from userland, this would work
I think.


Comments?
//mirabilos

No, the only way I've seen them sold is for $40 with a free OpenBSD CD.
	-- Haroon Khalid and Steve Shockley in gmane.os.openbsd.misc

tags 522773 wontfix
severity 522773 wishlist
thanks

You need to get this changed upstream, we won't deviate from
upstream for this.

Cheers,
        Moritz

Moritz Muehlenhoff dixit:

You probably have better contact to upstream, would you be
so kind as to bring this forward with them?

Thanks in advance.


Sincerely,
//mirabilos

We don't have time for this. You should contact linux-kernel or
file a bug at http://bugzilla.kernel.org

Cheers,
        Moritz

It seems there are two traditions for attribute macros.

The GNU one produced things like:

#define __attribute_malloc__ __attribute__ ((__malloc__))
#define __attribute_pure__ __attribute__ ((__pure__))
#define __attribute_used__ __attribute__ ((__used__))
#define __attribute_noinline__ __attribute__ ((__noinline__))
#define __always_inline __inline __attribute__ ((__always_inline__))

The BSD one produced things like:

#define __malloc_like   __attribute__((__malloc__))
#define __pure          __attribute__((__pure__))
#define __used          __attribute__((__used__))
#define __noinline      __attribute__ ((__noinline__))
#define __always_inline __attribute__((__always_inline__))

When it comes to __attibute__((__unused__)), it's just a matter of
agreeing on whether to call it __attribute_unused__ or __unused.

The problem is that there's severe NIH to overcome.  As you can
see with the bug you submitted to Linux developers, they're generally
very hostile to portability.  The same applies to Glibc (aka Drepper).

Here are some possible ways out of this I can think of:

- Propose to BSD folks that they accept __attribute_* prefix and
  define those macros (in addition to __unused etc), then begin
  accepting patches that replace __unused with __attribute_unused.

- Propose to GCC folks that they define __attribute_* macros in
  <stddef.h> (they won't define __unused since this would break
  Linux and Glibc), then bring the same proposal to Clang folks.  If
  both accept, FreeBSD is much more likely to backport it to GCC 4.2.

- Work with standards bodies (POSIX?) so that they specify either
  definition (it doesn't matter to us which one, the offending
  definition will have to adapt).

What I would definitely NOT do is waste any time trying to persuade
Drepper to remove __unused from Glibc or Linux developers to remove
__unused from Linux.

Hi Robert,

Robert Millan wrote:

I don't agree.  It's perfectly fine for there to be multiple names
for the thing --- the task at hand is dividing up the __* namespace
between libbsd and libc.

Would it be possible to collect a near-exhaustive list of identifiers
like this (i.e., used in BSD application code and starting with __)
that could cause trouble?  Then we can propose sed-ing them away at
header installation time to the libc-ports maintainers (as Thorsten
suggested), and I wouldn't be surprised if such a patch is accepted.

Thanks for moving this forward.  Thoughts below about the alternatives
you mentioned.

Regards,
Jonathan

If I were in their shoes, I wouldn't be happy about such patches.  It
sounds like heavy patching without much immediate benefit, with no end
in sight (since glibc could start using another keyword the next day).

Likewise, in their shoes I wouldn't accept such patches.  The macros
are not in the C or POSIX standard and are easy to define in terms of
the __attribute__() feature, so they're not part of what a C compiler
is supposed to do.  Making each implementation of the standard headers
add these macros would hinder portability between implementations
(yes, there are more than two :)).

That sounds like an excellent idea!  Presumably the C working group
would be likely to consider standardizing __attribute__((__unused__))
if it is proposed, since that syntax is already widely implemented.
C++0x has its own attribute syntax which could make something like
[[unused]] possible.

2011/6/18 Jonathan Nieder <jrnieder@gmail.com>:

Providing multiple names is fine, the problem is which ones to use.

If BSD developers agree to provide both __attribute_unused and
__unused, and switch their users to __attribute_unused, problem is
solved.

If Drepper agrees to provide __unused, problem would be solved
but this won't happen.

Yes, just read the GNU and BSD versions of <sys/cdefs.h>...

... but I don't think this will fly.

I can see they wouldn't be excited about it, but they might also accept
them.  This is not a recurrent problem: it's irrelevant if glibc changes it
to another keyword (heck, there isn't even any equivalent to __unused
RIGHT NOW), what matters is that we'd have found a way out the name
collision.

Actually, I realized afer writing this that the point of these macros
is precisely
portability among different compilers.  Moving them to a GCC header is
out of the question (we might as well just use
__attribute__((__unused__)) then).

Well I have no experience working with standards bodies (and no spare
time), but if someone wants to take it, I'm almost certain this would
resolve our problem.

Robert Millan dixit:

You know that there are more than one BSD, but only one glibc,
IIRC Drepper isn’t even its maintainer any more. Try persuading
for example Theo de Raadt of anything which doesn’t have any
immediate technical merit… have fun ;-)

Honestly, when resolving this I’d go for “who has the older
rights”. Maybe look at how CSUR resolves different claims to
the same part of the Unicode PUA, or something like that.
Nevertheless, thanks on picking this up.

bye,
//mirabilos

According to <http://www.gnu.org/software/libc/> "Ulrich Drepper is
currently the foremost contributor and has overall responsibility for
maintenance and development."

Debian GNU/Linux is the older system; the Debian GNU/kFreeBSD
maintainers have to deal with compatibility issues.

Ben.

Ben Hutchings dixit:

The use of __undefined in the BSDs predates use of it by
both Linux and GNU. (But when using this argumentation
style, we’d probably better take this upstream… except
that upstream may not be helping…)

bye,
//mirabilos

We already asked you back in September 2009 to report this upstream.

If you don't think that will help, we can just close the bug, then.

Cheers,
        Moritz

2011/7/29 Moritz Mühlenhoff <jmm@inutil.org>:

He did, precisely in September 2009:

https://bugzilla.kernel.org/show_bug.cgi?id=14186

He still got no response though.

'__unused' is defined as a macro by some BSD-derived userland software.
We should avoid using it in the UAPI.

References: https://bugs.debian.org/522773
Reported-by: Thorsten Glaser <tg@mirbsd.de>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
---
Build-tested on x86 only.  I also grepped for uses of the supposedly
 __unused fields on other architectures and found nothing.

Ben.

 arch/alpha/include/uapi/asm/stat.h       | 2 +-
 arch/arm64/include/uapi/asm/ucontext.h   | 2 +-
 arch/ia64/include/uapi/asm/stat.h        | 2 +-
 arch/powerpc/include/uapi/asm/ucontext.h | 2 +-
 arch/s390/include/uapi/asm/stat.h        | 2 +-
 arch/s390/include/uapi/asm/ucontext.h    | 4 ++--
 arch/sparc/include/uapi/asm/fcntl.h      | 4 ++--
 arch/sparc/include/uapi/asm/stat.h       | 2 +-
 arch/x86/include/uapi/asm/stat.h         | 8 ++++----
 include/uapi/linux/icmp.h                | 2 +-
 include/uapi/linux/sysctl.h              | 2 +-
 11 files changed, 16 insertions(+), 16 deletions(-)

diff --git a/arch/alpha/include/uapi/asm/stat.h b/arch/alpha/include/uapi/asm/stat.h
index 07ad3e6b3f3e..f13899ad8d85 100644
--- a/arch/alpha/include/uapi/asm/stat.h
+++ b/arch/alpha/include/uapi/asm/stat.h
@@ -42,7 +42,7 @@ struct stat64 {
 	unsigned long	st_mtime_nsec;
 	unsigned long	st_ctime;
 	unsigned long   st_ctime_nsec;
-  	long		__unused[3];
+	long		__unused1[3];
 };

 #endif
diff --git a/arch/arm64/include/uapi/asm/ucontext.h b/arch/arm64/include/uapi/asm/ucontext.h
index 791de8e89e35..37b343d2492e 100644
--- a/arch/arm64/include/uapi/asm/ucontext.h
+++ b/arch/arm64/include/uapi/asm/ucontext.h
@@ -24,7 +24,7 @@ struct ucontext {
 	stack_t		  uc_stack;
 	sigset_t	  uc_sigmask;
 	/* glibc uses a 1024-bit sigset_t */
-	__u8		  __unused[1024 / 8 - sizeof(sigset_t)];
+	__u8		  __unused1[1024 / 8 - sizeof(sigset_t)];
 	/* last for future expansion */
 	struct sigcontext uc_mcontext;
 };
diff --git a/arch/ia64/include/uapi/asm/stat.h b/arch/ia64/include/uapi/asm/stat.h
index 367bb90cdffa..aa1e4f942942 100644
--- a/arch/ia64/include/uapi/asm/stat.h
+++ b/arch/ia64/include/uapi/asm/stat.h
@@ -24,7 +24,7 @@ struct stat {
 	unsigned long	st_ctime_nsec;
 	unsigned long	st_blksize;
 	long		st_blocks;
-	unsigned long	__unused[3];
+	unsigned long	__unused1[3];
 };

 #define STAT_HAVE_NSEC 1
diff --git a/arch/powerpc/include/uapi/asm/ucontext.h b/arch/powerpc/include/uapi/asm/ucontext.h
index d9a4ddf0cc86..cb08c035e427 100644
--- a/arch/powerpc/include/uapi/asm/ucontext.h
+++ b/arch/powerpc/include/uapi/asm/ucontext.h
@@ -28,7 +28,7 @@ struct ucontext {
 	sigset_t	uc_sigmask;
 	/* glibc has 1024-bit signal masks, ours are 64-bit */
 #ifdef __powerpc64__
-	sigset_t	__unused[15];	/* Allow for uc_sigmask growth */
+	sigset_t	__unused1[15];	/* Allow for uc_sigmask growth */
 	struct sigcontext uc_mcontext;	/* last for extensibility */
 #else
 	int		uc_maskext[30];
diff --git a/arch/s390/include/uapi/asm/stat.h b/arch/s390/include/uapi/asm/stat.h
index b4ca97d91466..5e493c1aa1f6 100644
--- a/arch/s390/include/uapi/asm/stat.h
+++ b/arch/s390/include/uapi/asm/stat.h
@@ -93,7 +93,7 @@ struct stat {
 	unsigned long  st_ctime_nsec;
         unsigned long  st_blksize;
         long           st_blocks;
-        unsigned long  __unused[3];
+        unsigned long  __unused1[3];
 };

 #endif /* __s390x__ */
diff --git a/arch/s390/include/uapi/asm/ucontext.h b/arch/s390/include/uapi/asm/ucontext.h
index 64a69aa5dde0..59d915f9ac27 100644
--- a/arch/s390/include/uapi/asm/ucontext.h
+++ b/arch/s390/include/uapi/asm/ucontext.h
@@ -23,7 +23,7 @@ struct ucontext_extended {
 	_sigregs	  uc_mcontext;
 	sigset_t	  uc_sigmask;
 	/* Allow for uc_sigmask growth.  Glibc uses a 1024-bit sigset_t.  */
-	unsigned char	  __unused[128 - sizeof(sigset_t)];
+	unsigned char	  __unused1[128 - sizeof(sigset_t)];
 	_sigregs_ext	  uc_mcontext_ext;
 };

@@ -34,7 +34,7 @@ struct ucontext {
 	_sigregs          uc_mcontext;
 	sigset_t	  uc_sigmask;
 	/* Allow for uc_sigmask growth.  Glibc uses a 1024-bit sigset_t.  */
-	unsigned char	  __unused[128 - sizeof(sigset_t)];
+	unsigned char	  __unused1[128 - sizeof(sigset_t)];
 };

 #endif /* !_ASM_S390_UCONTEXT_H */
diff --git a/arch/sparc/include/uapi/asm/fcntl.h b/arch/sparc/include/uapi/asm/fcntl.h
index 7e8ace5bf760..9f5d4383170a 100644
--- a/arch/sparc/include/uapi/asm/fcntl.h
+++ b/arch/sparc/include/uapi/asm/fcntl.h
@@ -48,8 +48,8 @@
 #define F_WRLCK		2
 #define F_UNLCK		3

-#define __ARCH_FLOCK_PAD	short __unused;
-#define __ARCH_FLOCK64_PAD	short __unused;
+#define __ARCH_FLOCK_PAD	short __unused1;
+#define __ARCH_FLOCK64_PAD	short __unused1;

 #include <asm-generic/fcntl.h>

diff --git a/arch/sparc/include/uapi/asm/stat.h b/arch/sparc/include/uapi/asm/stat.h
index a232e9e1f4e5..0b4b4ae9a05d 100644
--- a/arch/sparc/include/uapi/asm/stat.h
+++ b/arch/sparc/include/uapi/asm/stat.h
@@ -43,7 +43,7 @@ struct stat64 {
 	unsigned long	st_mtime_nsec;
 	unsigned long	st_ctime;
 	unsigned long	st_ctime_nsec;
-	long		__unused[3];
+	long		__unused1[3];
 };

 #else
diff --git a/arch/x86/include/uapi/asm/stat.h b/arch/x86/include/uapi/asm/stat.h
index bc03eb5d6360..64ba4e324f57 100644
--- a/arch/x86/include/uapi/asm/stat.h
+++ b/arch/x86/include/uapi/asm/stat.h
@@ -99,15 +99,15 @@ struct stat {
 	__kernel_ulong_t	st_mtime_nsec;
 	__kernel_ulong_t	st_ctime;
 	__kernel_ulong_t	st_ctime_nsec;
-	__kernel_long_t		__unused[3];
+	__kernel_long_t		__unused1[3];
 };

 /* We don't need to memset the whole thing just to initialize the padding */
 #define INIT_STRUCT_STAT_PADDING(st) do {	\
 	st.__pad0 = 0;				\
-	st.__unused[0] = 0;			\
-	st.__unused[1] = 0;			\
-	st.__unused[2] = 0;			\
+	st.__unused1[0] = 0;			\
+	st.__unused1[1] = 0;			\
+	st.__unused1[2] = 0;			\
 } while (0)

 #endif
diff --git a/include/uapi/linux/icmp.h b/include/uapi/linux/icmp.h
index 16fff055f734..78d71b13f9e5 100644
--- a/include/uapi/linux/icmp.h
+++ b/include/uapi/linux/icmp.h
@@ -76,7 +76,7 @@ struct icmphdr {
 	} echo;
 	__be32	gateway;
 	struct {
-		__be16	__unused;
+		__be16	__unused1;
 		__be16	mtu;
 	} frag;
   } un;
diff --git a/include/uapi/linux/sysctl.h b/include/uapi/linux/sysctl.h
index 0956373b56db..177aa35a8ce8 100644
--- a/include/uapi/linux/sysctl.h
+++ b/include/uapi/linux/sysctl.h
@@ -40,7 +40,7 @@ struct __sysctl_args {
 	size_t __user *oldlenp;
 	void __user *newval;
 	size_t newlen;
-	unsigned long __unused[4];
+	unsigned long __unused1[4];
 };

 /* Define sysctl names first */

hi,

Given there was as well not reaction upstream AFAICS, as well to Ben's
posted patch I guess we can close this bug. Not doing so yet though.

Regards,
Salvatore

Another year has passed and it's now almost 10 years ago that the upstream bug
was closed (in a less then courteous way).
I just checked 'arch/alpha/include/uapi/asm/stat.h' in Linus' master branch
and the patch proposed by Ben, more then 6 years ago, hasn't been merged.

As I don't know the reason it wasn't closed last year, I won't do it, but
maybe it's time to finally close it?

Diederik de Haas dixit:

If the bug persists, it’s still a bug, so don’t close it.
Prod them occasionally.

I know that *some* __unused have been fixed; maybe the remaining
ones were an oversight?

bye,
//mirabilos