Re: [patch 12/13] signal/timer/event fds v10 - eventfd wire up x86_64 arch ...

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 02 Apr 2007 15:46:31 -0700
Davide Libenzi <davidel@xmailserver.org> wrote:

> This patch wire the eventfd system call to the x86_64 architecture.
> 
> 
> 
> Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
> 
> 
> - Davide
> 
> 
> 
> Index: linux-2.6.21-rc5.fds/arch/x86_64/ia32/ia32entry.S
> ===================================================================
> --- linux-2.6.21-rc5.fds.orig/arch/x86_64/ia32/ia32entry.S	2007-04-02 15:06:40.000000000 -0700
> +++ linux-2.6.21-rc5.fds/arch/x86_64/ia32/ia32entry.S	2007-04-02 15:06:46.000000000 -0700
> @@ -721,4 +721,5 @@
>  	.quad sys_epoll_pwait
>  	.quad sys_signalfd		/* 320 */
>  	.quad sys_timerfd
> +	.quad sys_eventfd
>  ia32_syscall_end:		
> Index: linux-2.6.21-rc5.fds/include/asm-x86_64/unistd.h
> ===================================================================
> --- linux-2.6.21-rc5.fds.orig/include/asm-x86_64/unistd.h	2007-04-02 15:06:40.000000000 -0700
> +++ linux-2.6.21-rc5.fds/include/asm-x86_64/unistd.h	2007-04-02 15:06:46.000000000 -0700
> @@ -623,8 +623,10 @@
>  __SYSCALL(__NR_signalfd, sys_signalfd)
>  #define __NR_timerfd		281
>  __SYSCALL(__NR_timerfd, sys_timerfd)
> +#define __NR_eventfd		282
> +__SYSCALL(__NR_eventfd, sys_eventfd)
>  
> -#define __NR_syscall_max __NR_timerfd
> +#define __NR_syscall_max __NR_eventfd
>  
>  #ifndef __NO_STUBS
>  #define __ARCH_WANT_OLD_READDIR

General point: the first hunk here which wires up the 32-bit entry point
should arguably be part of the "wire up i386" patch, not the "wire up
x86_64" patch.

This is not a clear-cut thing.  The problem with the approach you've taken
here is that if we wire up i386 and not x86_64's i386 table (for example)
there are ways in which intervening patches can accidentally get i386 and
x86_64 emulation's syscall tables out of sync.  And, of course, 32-bit
applications will run on i386 but won't run on x86_64 or vice-versa.

So I think the best way to avoid all such problems is to always wire up
i386 and x86_64 (both 32- and 64-bit) in the same patch.