sys/mount.h vs fcntl.h

[Florian Weimer]

* DJ Delorie:

> diff --git a/sysdeps/unix/sysv/linux/sys/mount.h b/sysdeps/unix/sysv/linux/sys/mount.h
> index 457d8dcff2..d52c0121a3 100644
> --- a/sysdeps/unix/sysv/linux/sys/mount.h
> +++ b/sysdeps/unix/sysv/linux/sys/mount.h
> @@ -21,7 +21,6 @@
>  #ifndef _SYS_MOUNT_H
>  #define _SYS_MOUNT_H	1
>  
> -#include <fcntl.h>
>  #include <features.h>
>  #include <stdint.h>
>  #include <stddef.h>
> @@ -266,7 +265,13 @@ enum fsconfig_command
>  
>  /* open_tree flags.  */
>  #define OPEN_TREE_CLONE    1         /* Clone the target tree and attach the clone */
> -#define OPEN_TREE_CLOEXEC  O_CLOEXEC /* Close the file on execve() */
> +/* The kernel version of fcntl.h defines this as O_CLOEXEC but that is
> +   not a valid flag according to the manual, so use our private
> +   version instead to avoid namespace pollution and the fcntl.h
> +   requirement.  */
> +#include <bits/cloexec.h>
> +#undef  OPEN_TREE_CLOEXEC
> +#define OPEN_TREE_CLOEXEC  __O_CLOEXEC /* Close the file on execve() */

I suggest this instead:

#include <bits/cloexec.h>
#define O_CLOEXEC __O_CLOEXEC
#define OPEN_TREE_CLOEXEC O_CLOEXEC

So it's identical to the kernel macro definition and no #undef is
needed.

> diff --git a/sysdeps/unix/sysv/linux/tst-fcntl-cloexec.c b/sysdeps/unix/sysv/linux/tst-fcntl-cloexec.c
> new file mode 100644
> index 0000000000..442fa9f7d5
> --- /dev/null
> +++ b/sysdeps/unix/sysv/linux/tst-fcntl-cloexec.c
> @@ -0,0 +1,47 @@

> +/* Both fcntl.h and sys/mount.h include APIs that need O_CLOEXEC.
> +
> +   So, there's an arch-specific header for this constant, just for
> +   this purpose.  We test that it is defined correctly here, and that
> +   fcntl.h can use it.  */

You can include both <bits/cloexec.h> and the kernel header
<linux/fcntl.h> and use static_assert.

Or write a tst-fcntl-consts.py test that checks all constants.  This
would completely break the build if there's an inconsistency.

Thanks,
Florian