This patch adds test that ensure the boundary conditions for the various

constants introduced in the previous patches is met.  No code is generated.
 fs/eventfd.c      |    4 ++++

 fs/eventpoll.c    |    3 +++

 fs/inotify_user.c |    4 ++++

 fs/signalfd.c     |    4 ++++

 fs/timerfd.c      |    4 ++++

 net/socket.c      |    7 +++++++

 6 files changed, 26 insertions(+)
Signed-off-by: Ulrich Drepper <drepper@redhat.com>
diff --git a/fs/eventfd.c b/fs/eventfd.c

index 3ed4466..08bf558 100644

--- a/fs/eventfd.c

+++ b/fs/eventfd.c

@@ -203,6 +203,10 @@ asmlinkage long sys_eventfd2(unsigned int count, int flags)

 	int fd;

 	struct eventfd_ctx *ctx;
+	/* Check the EFD_* constants for consistency.  */

+	BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);

+	BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);

+

 	if (flags & ~(EFD_CLOEXEC | EFD_NONBLOCK))

 		return -EINVAL;
diff --git a/fs/eventpoll.c b/fs/eventpoll.c

index 3fd4014..2fdad42 100644

--- a/fs/eventpoll.c

+++ b/fs/eventpoll.c

@@ -1051,6 +1051,9 @@ asmlinkage long sys_epoll_create2(int size, int flags)

 	int error, fd = -1;

 	struct eventpoll *ep;
+	/* Check the EPOLL_* constant for consistency.  */

+	BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);

+

 	if (flags & ~EPOLL_CLOEXEC)

 		return -EINVAL;
diff --git a/fs/inotify_user.c b/fs/inotify_user.c

index dc7e1f6..fe79c25 100644

--- a/fs/inotify_user.c

+++ b/fs/inotify_user.c

@@ -574,6 +574,10 @@ asmlinkage long sys_inotify_init1(int flags)

 	struct file *filp;

 	int fd, ret;
+	/* Check the IN_* constants for consistency.  */

+	BUILD_BUG_ON(IN_CLOEXEC != O_CLOEXEC);

+	BUILD_BUG_ON(IN_NONBLOCK != O_NONBLOCK);

+

 	if (flags & ~(IN_CLOEXEC | IN_NONBLOCK))

 		return -EINVAL;
diff --git a/fs/signalfd.c b/fs/signalfd.c

index 5441a4b..9c39bc7 100644

--- a/fs/signalfd.c

+++ b/fs/signalfd.c

@@ -211,6 +211,10 @@ asmlinkage long sys_signalfd4(int ufd, sigset_t __user *user_mask,

 	sigset_t sigmask;

 	struct signalfd_ctx *ctx;
+	/* Check the SFD_* cons...
[ message continues ]

On alpha the
	BUILD_BUG_ON(SOCK_NONBLOCK & SOCK_TYPE_MASK);
seems to have gone away, but now
	BUILD_BUG_ON(SOCK_NONBLOCK != O_NONBLOCK);
is triggering.
--

-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1
After the addition of arch-specific SOCK_NONBLOCK values this line must

go.  I thought one of my patches already did that.
- --

➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.4.9 (GNU/Linux)
iEYEARECAAYFAkg21SoACgkQ2ijCOnn/RHQG2wCeNS8o/cVxuOYXRbCnwFTg1rcx

590AoIvJBnwNP9O26kIHNY3ExVTA2M2a

=5P4P

-----END PGP SIGNATURE-----

--

The fifth assertion triggers with alpha allmodconfig.

--

-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1
This means the simple approach won't work for alpha.
The question is how to solve it?  I suggest the following:
- - define SOCK_NONBLOCK to some value > 16 in include/asm-alpha/socket.h
- - add #ifndef SOCK_NONBLOCK around the definition in include/linux/net.h
- - in sys_socket, sys_socketpair, sys_paccept
   int fflags = flags;

   if (SOCK_NONBLOCK != O_NONBLOCK && (flags & SOCK_NONBLOCK)) {

     fflags &= ~SOCK_NONBLOCK;

     fflags |= O_NONBLOCK;

   }
  and use fflags instead of flags in the places where we pass it on.
This has no costs on platforms other than alpha.
Shall I sent a patch?
- --

➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.4.9 (GNU/Linux)

Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org
iEYEARECAAYFAkgpEPoACgkQ2ijCOnn/RHQ38gCgyVOI/urGeoIr7CFIfs8C7OyR

I80AniPGGWyxZYUQg0pQlySQHqX/UYlx

=5fA2

-----END PGP SIGNATURE-----

--

From: Ulrich Drepper <drepper@redhat.com>
You'll need to handle MIPS and PARISC as well, as they have

similar conflicts.

--

m68k too..

--

-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1
I cannot build it here but looking at the headers:
- - include/asm-m68k/fcntl.h does not define O_NONBLOCK; and therefore
- - the definition from include/asm-generic/fcntl.h is used which is 04000
Is there any magic I miss?
- --

➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.4.7 (GNU/Linux)
iD8DBQFIKSLi2ijCOnn/RHQRAvItAJ45n2h20UasryYucg18pd9ERWUpbACglsvR

ftad99qXQ20UwezP0DjbssA=

=orzm

-----END PGP SIGNATURE-----

--

Different bug, I think.
net/socket.c: In function 'sys_paccept':

net/socket.c:1543: error: implicit declaration of function 'set_restore_sigmask'
--

-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1
Right, that's the signal mask stuff.  Should be fixed by the other patch

I sent.
- --

➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.4.7 (GNU/Linux)
iD8DBQFIKSpa2ijCOnn/RHQRAmPvAJ90D3ZHrIWs1ZE5K9XVEdZBh8vcdQCePcaw

1Hnyz58x9oizIXe6qiS82xU=

=o78L

-----END PGP SIGNATURE-----

--

mips allmodconfig is unhappy still.
net/socket.c: In function `sys_socket':

net/socket.c:1225: error: `SOCK_CLOEXEC' undeclared (first use in this function)

net/socket.c:1225: error: (Each undeclared identifier is reported only once

net/socket.c:1225: error: for each function it appears in.)

net/socket.c:1226: error: `SOCK_NONBLOCK' undeclared (first use in this function)

net/socket.c:1227: error: `SOCK_TYPE_MASK' undeclared (first use in this function)

net/socket.c: In function `sys_socketpair':

net/socket.c:1268: error: `SOCK_TYPE_MASK' undeclared (first use in this function)

net/socket.c:1269: error: `SOCK_CLOEXEC' undeclared (first use in this function)

net/socket.c:1269: error: `SOCK_NONBLOCK' undeclared (first use in this function)

net/socket.c: In function `do_accept':

net/socket.c:1438: error: `SOCK_CLOEXEC' undeclared (first use in this function)

net/socket.c:1438: error: `SOCK_NONBLOCK' undeclared (first use in this function)

--

-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1
Is MIPS really a problem?  The value is 0x80.  128 protocol types should

be enough.  Even if not, there is no reason why the numbers should be

consecutive.  It's easy enough to skip over a bit.  I.e., after procotol

0x7f the next one would be 0x100.
And PA seems to have gotten O_NONBLOCK wrong.  SPARC does it right AFAICS.
- --

➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.4.7 (GNU/Linux)
iD8DBQFIKRdJ2ijCOnn/RHQRAovKAKDFz2ZJ0E2XYVu5Qn8fME+ieZZopwCgmH4D

ldfvNrhK8LdvjZzNPViILV0=

=Z9Dj

-----END PGP SIGNATURE-----

--

From: Ulrich Drepper <drepper@redhat.com>
I agree.
But that value is compiled into every parisc binary out there,

limiting what we can do about it in the short term.

--

Sounds reasonable.

--

From: Andrew Morton <akpm@linux-foundation.org>
Unfortunately, MIPS and PARISC look like they will as well :-/

--