[RFC] __initdata strings

[RFC] __initdata strings

[Oleg Nesterov]

Hello.

This patch is not intended for inclusion, just for illustration.

__init functions leaves strings (mainly printk's arguments) in
.data section. It make sense to move them in .init.data.

Is there anyone else who would consider this useful?

Oleg.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>

diff -rup 2.6.9-rc3-clean/include/linux/init.h 2.6.9-rc3-i_str/include/linux/init.h
--- 2.6.9-rc3-clean/include/linux/init.h	Sun Oct 10 11:48:46 2004
+++ 2.6.9-rc3-i_str/include/linux/init.h	Sun Oct 10 11:55:07 2004
@@ -61,6 +61,31 @@
 /*
  * Used for initialization calls..
  */
+
+#define I_STRING(str)	\
+({						\
+	static char data[] __initdata = (str);	\
+	data;					\
+})
+
+#if	defined(__GNUC__) && (__GNUC__ >= 3)
+#define CK_FMTSTR(expr)	do if (0) { expr; } while (0)
+#else
+#define CK_FMTSTR(expr)	do ; while (0)
+#endif
+
+#define i_printk(fmt, args...)	\
+({						\
+	CK_FMTSTR(printk(fmt , ##args));	\
+	printk(I_STRING(fmt) , ##args);		\
+})
+
+#define i_panic(fmt, args...)	\
+({						\
+	CK_FMTSTR(panic(fmt , ##args));		\
+	panic(I_STRING(fmt) , ##args);		\
+})
+
 typedef int (*initcall_t)(void);
 typedef void (*exitcall_t)(void);
 
diff -rup 2.6.9-rc3-clean/init/do_mounts.c 2.6.9-rc3-i_str/init/do_mounts.c
--- 2.6.9-rc3-clean/init/do_mounts.c	Sun Oct 10 11:48:57 2004
+++ 2.6.9-rc3-i_str/init/do_mounts.c	Sat Oct  9 18:18:30 2004
@@ -265,10 +265,10 @@ static int __init do_mount_root(char *na
 
 	sys_chdir("/root");
 	ROOT_DEV = current->fs->pwdmnt->mnt_sb->s_dev;
-	printk("VFS: Mounted root (%s filesystem)%s.\n",
+	i_printk("VFS: Mounted root (%s filesystem)%s.\n",
 	       current->fs->pwdmnt->mnt_sb->s_type->name,
 	       current->fs->pwdmnt->mnt_sb->s_flags & MS_RDONLY ? 
-	       " readonly" : "");
+	       I_STRING(" readonly") : "");
 	return 0;
 }
 
@@ -296,13 +296,13 @@ retry:
 		 * and bad superblock on root device.
 		 */
 		__bdevname(ROOT_DEV, b);
-		printk("VFS: Cannot open root device \"%s\" or %s\n",
+		i_printk("VFS: Cannot open root device \"%s\" or %s\n",
 				root_device_name, b);
-		printk("Please append a correct \"root=\" boot option\n");
+		i_printk("Please append a correct \"root=\" boot option\n");
 
-		panic("VFS: Unable to mount root fs on %s", b);
+		i_panic("VFS: Unable to mount root fs on %s", b);
 	}
-	panic("VFS: Unable to mount root fs on %s", __bdevname(ROOT_DEV, b));
+	i_panic("VFS: Unable to mount root fs on %s", __bdevname(ROOT_DEV, b));
 out:
 	putname(fs_names);
 }
@@ -336,7 +336,7 @@ void __init change_floppy(char *fmt, ...
 		sys_ioctl(fd, FDEJECT, 0);
 		sys_close(fd);
 	}
-	printk(KERN_NOTICE "VFS: Insert %s and press ENTER\n", buf);
+	i_printk(KERN_NOTICE "VFS: Insert %s and press ENTER\n", buf);
 	fd = sys_open("/dev/console", O_RDWR, 0);
 	if (fd >= 0) {
 		sys_ioctl(fd, TCGETS, (long)&termios);
@@ -357,7 +357,7 @@ void __init mount_root(void)
 		if (mount_nfs_root())
 			return;
 
-		printk(KERN_ERR "VFS: Unable to mount root fs via NFS, trying floppy.\n");
+		i_printk(KERN_ERR "VFS: Unable to mount root fs via NFS, trying floppy.\n");
 		ROOT_DEV = Root_FD0;
 	}
 #endif

Re: [RFC] __initdata strings

[Andi Kleen]

Oleg Nesterov <oleg@tv-sign.ru> writes:

> Hello.
>
> This patch is not intended for inclusion, just for illustration.
>
> __init functions leaves strings (mainly printk's arguments) in
> .data section. It make sense to move them in .init.data.
>
> Is there anyone else who would consider this useful?

There is a more generic way to do this with gcc extensions. Something like
(uncompiled/untested)

#define __i(x) ({ static char __str[] __initdata = x; __str; })

But I'm not sure the few bytes saved are worth the code uglification. 
Probably not. likely/unlikely is already bad enough.

-Andi

Re: [RFC] __initdata strings

[Oleg Nesterov]

Andi Kleen wrote:
> 
> There is a more generic way to do this with gcc extensions. Something like
> 
> #define __i(x) ({ static char __str[] __initdata = x; __str; })

I can't see any difference with:

#define I_STRING(str)  \
({                                             \
       static char data[] __initdata = (str);  \
       data;                                   \
})

> But I'm not sure the few bytes saved are worth the code uglification.

Probably you are right, but i think it would be few kilobytes,
there are so many __init functions.

Oleg.

Re: [RFC] __initdata strings

[Matt Mackall]

On Sun, Oct 10, 2004 at 07:28:29PM +0400, Oleg Nesterov wrote:
> Hello.
> 
> This patch is not intended for inclusion, just for illustration.
> 
> __init functions leaves strings (mainly printk's arguments) in
> .data section. It make sense to move them in .init.data.

Probably better to do this with something like objcopy?

-- 
Mathematics is the supreme nostalgia of our time.

Re: [RFC] __initdata strings

[Arnaldo Carvalho de Melo]

Oleg Nesterov wrote:
> Andi Kleen wrote:
> 
>>There is a more generic way to do this with gcc extensions. Something like

Or with a pre-processor, using a sparse based tool such as this one I've
written a long time ago:

http://www.kernel.org/pub/linux/kernel/people/acme/sparse/initstr.c

That uses what Andi described, but does this automatically if initstr is
made part of the building process.

>>
>>#define __i(x) ({ static char __str[] __initdata = x; __str; })

Regards,

- Arnaldo

Re: [RFC] __initdata strings

[Arnaldo Carvalho de Melo]

Matt Mackall wrote:
> On Sun, Oct 10, 2004 at 07:28:29PM +0400, Oleg Nesterov wrote:
> 
>>Hello.
>>
>>This patch is not intended for inclusion, just for illustration.
>>
>>__init functions leaves strings (mainly printk's arguments) in
>>.data section. It make sense to move them in .init.data.
> 
> 
> Probably better to do this with something like objcopy?
> 

Yes, this is another way of doing that, but the kernel has to be
prepared to get such treatment, think about register functions
that only save a pointer to strings passed from __init functions...

Ah, nothing related to this specific way of doing what is intended,
any scheme that moves strings in __init functions to .data.init has
to deal with this.

- Arnaldo