[PATCH] introduce __init_exit function annotation

[PATCH] introduce __init_exit function annotation

[Domen Puncer]

Introduce __init_exit, which is useful ie. for drivers that call
cleanup functions when they fail in __init functions.


Signed-off-by: Domen Puncer <domen.puncer@telargo.com>
---
 include/linux/init.h |    2 ++
 1 file changed, 2 insertions(+)

Index: work-powerpc.git/include/linux/init.h
===================================================================
--- work-powerpc.git.orig/include/linux/init.h
+++ work-powerpc.git/include/linux/init.h
@@ -60,8 +60,10 @@
 
 #ifdef MODULE
 #define __exit		__attribute__ ((__section__(".exit.text")))
+#define __init_exit
 #else
 #define __exit		__attribute_used__ __attribute__ ((__section__(".exit.text")))
+#define __init_exit	__init
 #endif
 
 /* For assembly routines */

Re: [PATCH] introduce __init_exit function annotation

[Adrian Bunk]

On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> Introduce __init_exit, which is useful ie. for drivers that call
> cleanup functions when they fail in __init functions.
> 
> 
> Signed-off-by: Domen Puncer <domen.puncer@telargo.com>
> ---
>  include/linux/init.h |    2 ++
>  1 file changed, 2 insertions(+)
> 
> Index: work-powerpc.git/include/linux/init.h
> ===================================================================
> --- work-powerpc.git.orig/include/linux/init.h
> +++ work-powerpc.git/include/linux/init.h
> @@ -60,8 +60,10 @@
>  
>  #ifdef MODULE
>  #define __exit		__attribute__ ((__section__(".exit.text")))
> +#define __init_exit
>  #else
>  #define __exit		__attribute_used__ __attribute__ ((__section__(".exit.text")))
> +#define __init_exit	__init
>  #endif
>  
>  /* For assembly routines */

This doesn't work on architectures like i386 where __exit code is
discarded at runtime.

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed

Re: [PATCH] introduce __init_exit function annotation

[Domen Puncer]

On 17/07/07 10:31 +0200, Adrian Bunk wrote:
> On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > Introduce __init_exit, which is useful ie. for drivers that call
> > cleanup functions when they fail in __init functions.
> > 
> > 
> > Signed-off-by: Domen Puncer <domen.puncer@telargo.com>
> > ---
> >  include/linux/init.h |    2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > Index: work-powerpc.git/include/linux/init.h
> > ===================================================================
> > --- work-powerpc.git.orig/include/linux/init.h
> > +++ work-powerpc.git/include/linux/init.h
> > @@ -60,8 +60,10 @@
> >  
> >  #ifdef MODULE
> >  #define __exit		__attribute__ ((__section__(".exit.text")))
> > +#define __init_exit
> >  #else
> >  #define __exit		__attribute_used__ __attribute__ ((__section__(".exit.text")))
> > +#define __init_exit	__init
> >  #endif
> >  
> >  /* For assembly routines */
> 
> This doesn't work on architectures like i386 where __exit code is
> discarded at runtime.

If it's a module, then it shouldn't be discarded until unload anyway.
If it's in-kernel, then it'll be discarded as __init stuff.
I don't see a problem?

BTW. can you point me to reasoning for discarding __exit at runtime?


	Domen

> 
> cu
> Adrian
> 
> -- 
> 
>        "Is there not promise of rain?" Ling Tan asked suddenly out
>         of the darkness. There had been need of rain for many days.
>        "Only a promise," Lao Er said.
>                                        Pearl S. Buck - Dragon Seed

Re: [PATCH] introduce __init_exit function annotation

[Sam Ravnborg]

On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> Introduce __init_exit, which is useful ie. for drivers that call
> cleanup functions when they fail in __init functions.

This is wrong.
On arm (just one example of several) the __exit section are discarded
at buildtime so any reference from __init to __exit will cause the
linker to error out.

The real solution is only to declare functiones used solely during
exit as __exit.
The whole point of using __exit is to tell that this can be safely
dropped when built-in because it is not used then.

	Sam

Re: [PATCH] introduce __init_exit function annotation

[Takashi Iwai]

At Tue, 17 Jul 2007 15:02:30 +0200,
Sam Ravnborg wrote:
> 
> On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > Introduce __init_exit, which is useful ie. for drivers that call
> > cleanup functions when they fail in __init functions.
> 
> This is wrong.
> On arm (just one example of several) the __exit section are discarded
> at buildtime so any reference from __init to __exit will cause the
> linker to error out.

Hmm, from what I see, it adds __init to the function.  There is no
reference to __exit.


Takashi

Re: [PATCH] introduce __init_exit function annotation

[Sam Ravnborg]

On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> At Tue, 17 Jul 2007 15:02:30 +0200,
> Sam Ravnborg wrote:
> > 
> > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > Introduce __init_exit, which is useful ie. for drivers that call
> > > cleanup functions when they fail in __init functions.
> > 
> > This is wrong.
> > On arm (just one example of several) the __exit section are discarded
> > at buildtime so any reference from __init to __exit will cause the
> > linker to error out.
> 
> Hmm, from what I see, it adds __init to the function.  There is no
> reference to __exit.

The cleanup functions are marked __exit in the referenced case.

	Sam

Re: [PATCH] introduce __init_exit function annotation

[Takashi Iwai]

At Tue, 17 Jul 2007 17:14:32 +0200,
Sam Ravnborg wrote:
> 
> On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > At Tue, 17 Jul 2007 15:02:30 +0200,
> > Sam Ravnborg wrote:
> > > 
> > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > cleanup functions when they fail in __init functions.
> > > 
> > > This is wrong.
> > > On arm (just one example of several) the __exit section are discarded
> > > at buildtime so any reference from __init to __exit will cause the
> > > linker to error out.
> > 
> > Hmm, from what I see, it adds __init to the function.  There is no
> > reference to __exit.
> 
> The cleanup functions are marked __exit in the referenced case.

My understanding is that it's the very purpose of this patch --
change the mark from __exit to __init_exit for such clean-up
functions.


Takashi

Re: [PATCH] introduce __init_exit function annotation

[Sam Ravnborg]

On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> At Tue, 17 Jul 2007 17:14:32 +0200,
> Sam Ravnborg wrote:
> > 
> > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > Sam Ravnborg wrote:
> > > > 
> > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > cleanup functions when they fail in __init functions.
> > > > 
> > > > This is wrong.
> > > > On arm (just one example of several) the __exit section are discarded
> > > > at buildtime so any reference from __init to __exit will cause the
> > > > linker to error out.
> > > 
> > > Hmm, from what I see, it adds __init to the function.  There is no
> > > reference to __exit.
> > 
> > The cleanup functions are marked __exit in the referenced case.
> 
> My understanding is that it's the very purpose of this patch --
> change the mark from __exit to __init_exit for such clean-up
> functions.

And that is wrong.
See following example:

static void __init foo_init()
{
	if (error)
		foo_exit();
}

static void __exit foo_exit()
{
}

If foo_init is annotated with __init_exit then in the build-in case it
become __init and there is a reference to a non existing function because
functions marked __exit are discarded during link or run-time (depending on arch).

If foo_exit() are marked __init_exit then it becomes __init in the non-module case
which seems coorrect. If this is the intention of the patch then it should
be OK but then this intention should be spelled out.

	Sam

Re: [PATCH] introduce __init_exit function annotation

[Takashi Iwai]

At Tue, 17 Jul 2007 17:32:36 +0200,
Sam Ravnborg wrote:
> 
> On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> > At Tue, 17 Jul 2007 17:14:32 +0200,
> > Sam Ravnborg wrote:
> > > 
> > > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > > Sam Ravnborg wrote:
> > > > > 
> > > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > > cleanup functions when they fail in __init functions.
> > > > > 
> > > > > This is wrong.
> > > > > On arm (just one example of several) the __exit section are discarded
> > > > > at buildtime so any reference from __init to __exit will cause the
> > > > > linker to error out.
> > > > 
> > > > Hmm, from what I see, it adds __init to the function.  There is no
> > > > reference to __exit.
> > > 
> > > The cleanup functions are marked __exit in the referenced case.
> > 
> > My understanding is that it's the very purpose of this patch --
> > change the mark from __exit to __init_exit for such clean-up
> > functions.
> 
> And that is wrong.

You misunderstood.  What I meant is the case like this:

static void __init_exit cleanup()
{
	...
}

static void __init foo_init()
{
	if (error)
		cleanup();
}

static void __exit foo_exit()
{
	cleanup();
}

Currently, there is no proper way to mark cleanup().  Neither __init,
__exit, __devinit nor __devexit can be used there.


Takashi

Re: [PATCH] introduce __init_exit function annotation

[Sam Ravnborg]

On Tue, Jul 17, 2007 at 05:40:15PM +0200, Takashi Iwai wrote:
> At Tue, 17 Jul 2007 17:32:36 +0200,
> Sam Ravnborg wrote:
> > 
> > On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> > > At Tue, 17 Jul 2007 17:14:32 +0200,
> > > Sam Ravnborg wrote:
> > > > 
> > > > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > > > Sam Ravnborg wrote:
> > > > > > 
> > > > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > > > cleanup functions when they fail in __init functions.
> > > > > > 
> > > > > > This is wrong.
> > > > > > On arm (just one example of several) the __exit section are discarded
> > > > > > at buildtime so any reference from __init to __exit will cause the
> > > > > > linker to error out.
> > > > > 
> > > > > Hmm, from what I see, it adds __init to the function.  There is no
> > > > > reference to __exit.
> > > > 
> > > > The cleanup functions are marked __exit in the referenced case.
> > > 
> > > My understanding is that it's the very purpose of this patch --
> > > change the mark from __exit to __init_exit for such clean-up
> > > functions.
> > 
> > And that is wrong.
> 
> You misunderstood.  What I meant is the case like this:
> 
> static void __init_exit cleanup()
> {
> 	...
> }
> 
> static void __init foo_init()
> {
> 	if (error)
> 		cleanup();
> }
> 
> static void __exit foo_exit()
> {
> 	cleanup();
> }
> 
> Currently, there is no proper way to mark cleanup().  Neither __init,
> __exit, __devinit nor __devexit can be used there.

Then you get the annotation sorted out so cleanup() get discarded in the
built-in case. But you leave no room for automated tools to detect this.

If this is really necessary (and I daught) then a specific section should be
dedicated for this usage.

We have lot of issues with current __init/__exit, __devinit/__devexit, __cpuint/__cpuexit
and introducing more of the kind does not help it.
So even if it saves a few bytes in some odd cases the added complaxity is IMHO not worth it.

	Sam

Re: [PATCH] introduce __init_exit function annotation

[Takashi Iwai]

At Tue, 17 Jul 2007 18:48:46 +0200,
Sam Ravnborg wrote:
> 
> On Tue, Jul 17, 2007 at 05:40:15PM +0200, Takashi Iwai wrote:
> > At Tue, 17 Jul 2007 17:32:36 +0200,
> > Sam Ravnborg wrote:
> > > 
> > > On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> > > > At Tue, 17 Jul 2007 17:14:32 +0200,
> > > > Sam Ravnborg wrote:
> > > > > 
> > > > > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > > > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > > > > Sam Ravnborg wrote:
> > > > > > > 
> > > > > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > > > > cleanup functions when they fail in __init functions.
> > > > > > > 
> > > > > > > This is wrong.
> > > > > > > On arm (just one example of several) the __exit section are discarded
> > > > > > > at buildtime so any reference from __init to __exit will cause the
> > > > > > > linker to error out.
> > > > > > 
> > > > > > Hmm, from what I see, it adds __init to the function.  There is no
> > > > > > reference to __exit.
> > > > > 
> > > > > The cleanup functions are marked __exit in the referenced case.
> > > > 
> > > > My understanding is that it's the very purpose of this patch --
> > > > change the mark from __exit to __init_exit for such clean-up
> > > > functions.
> > > 
> > > And that is wrong.
> > 
> > You misunderstood.  What I meant is the case like this:
> > 
> > static void __init_exit cleanup()
> > {
> > 	...
> > }
> > 
> > static void __init foo_init()
> > {
> > 	if (error)
> > 		cleanup();
> > }
> > 
> > static void __exit foo_exit()
> > {
> > 	cleanup();
> > }
> > 
> > Currently, there is no proper way to mark cleanup().  Neither __init,
> > __exit, __devinit nor __devexit can be used there.
> 
> Then you get the annotation sorted out so cleanup() get discarded in the
> built-in case. But you leave no room for automated tools to detect this.
> 
> If this is really necessary (and I daught) then a specific section should be
> dedicated for this usage.
> 
> We have lot of issues with current __init/__exit, __devinit/__devexit, __cpuint/__cpuexit
> and introducing more of the kind does not help it.
> So even if it saves a few bytes in some odd cases the added complaxity is IMHO not worth it.

Well, I don't think it's a few bytes and not so odd, but I agree that
this solution isn't the best way.  And, I now remember that this won't
work anyway, too.  Calling __init from __exit also causes error...


BTW, this reminds me why we have to add annotations for each
subisdiary function manually.  A tool to parse the code statically and
give the proper annotations/hints would be really nice.


Takashi

Re: [PATCH] introduce __init_exit function annotation

[Domen Puncer]

On 17/07/07 17:40 +0200, Takashi Iwai wrote:
> At Tue, 17 Jul 2007 17:32:36 +0200,
> Sam Ravnborg wrote:
> > 
> > On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> > > At Tue, 17 Jul 2007 17:14:32 +0200,
> > > Sam Ravnborg wrote:
> > > > 
> > > > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > > > Sam Ravnborg wrote:
> > > > > > 
> > > > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > > > cleanup functions when they fail in __init functions.
> > > > > > 
> > > > > > This is wrong.
> > > > > > On arm (just one example of several) the __exit section are discarded
> > > > > > at buildtime so any reference from __init to __exit will cause the
> > > > > > linker to error out.
> > > > > 
> > > > > Hmm, from what I see, it adds __init to the function.  There is no
> > > > > reference to __exit.
> > > > 
> > > > The cleanup functions are marked __exit in the referenced case.
> > > 
> > > My understanding is that it's the very purpose of this patch --
> > > change the mark from __exit to __init_exit for such clean-up
> > > functions.
> > 
> > And that is wrong.
> 
> You misunderstood.  What I meant is the case like this:
> 
> static void __init_exit cleanup()
> {
> 	...
> }
> 
> static void __init foo_init()
> {
> 	if (error)
> 		cleanup();
> }
> 
> static void __exit foo_exit()
> {
> 	cleanup();
> }

Uh, yes, this, or just __init_exit foo_exit() as in Sam's example.
It seemed obvious to me, sorry.

> 
> Currently, there is no proper way to mark cleanup().  Neither __init,
> __exit, __devinit nor __devexit can be used there.
> 
> 
> Takashi

Re: [PATCH] introduce __init_exit function annotation

[Domen Puncer]

On 17/07/07 19:02 +0200, Takashi Iwai wrote:
> At Tue, 17 Jul 2007 18:48:46 +0200,
> Sam Ravnborg wrote:
> > 
> > On Tue, Jul 17, 2007 at 05:40:15PM +0200, Takashi Iwai wrote:
> > > At Tue, 17 Jul 2007 17:32:36 +0200,
> > > Sam Ravnborg wrote:
> > > > 
> > > > On Tue, Jul 17, 2007 at 05:16:13PM +0200, Takashi Iwai wrote:
> > > > > At Tue, 17 Jul 2007 17:14:32 +0200,
> > > > > Sam Ravnborg wrote:
> > > > > > 
> > > > > > On Tue, Jul 17, 2007 at 04:52:12PM +0200, Takashi Iwai wrote:
> > > > > > > At Tue, 17 Jul 2007 15:02:30 +0200,
> > > > > > > Sam Ravnborg wrote:
> > > > > > > > 
> > > > > > > > On Tue, Jul 17, 2007 at 10:02:48AM +0200, Domen Puncer wrote:
> > > > > > > > > Introduce __init_exit, which is useful ie. for drivers that call
> > > > > > > > > cleanup functions when they fail in __init functions.
> > > > > > > > 
> > > > > > > > This is wrong.
> > > > > > > > On arm (just one example of several) the __exit section are discarded
> > > > > > > > at buildtime so any reference from __init to __exit will cause the
> > > > > > > > linker to error out.
> > > > > > > 
> > > > > > > Hmm, from what I see, it adds __init to the function.  There is no
> > > > > > > reference to __exit.
> > > > > > 
> > > > > > The cleanup functions are marked __exit in the referenced case.
> > > > > 
> > > > > My understanding is that it's the very purpose of this patch --
> > > > > change the mark from __exit to __init_exit for such clean-up
> > > > > functions.
> > > > 
> > > > And that is wrong.
> > > 
> > > You misunderstood.  What I meant is the case like this:
> > > 
> > > static void __init_exit cleanup()
> > > {
> > > 	...
> > > }
> > > 
> > > static void __init foo_init()
> > > {
> > > 	if (error)
> > > 		cleanup();
> > > }
> > > 
> > > static void __exit foo_exit()
> > > {
> > > 	cleanup();
> > > }
> > > 
> > > Currently, there is no proper way to mark cleanup().  Neither __init,
> > > __exit, __devinit nor __devexit can be used there.
> > 
> > Then you get the annotation sorted out so cleanup() get discarded in the
> > built-in case. But you leave no room for automated tools to detect this.
> > 
> > If this is really necessary (and I daught) then a specific section should be
> > dedicated for this usage.
> > 
> > We have lot of issues with current __init/__exit, __devinit/__devexit, __cpuint/__cpuexit
> > and introducing more of the kind does not help it.
> > So even if it saves a few bytes in some odd cases the added complaxity is IMHO not worth it.
> 
> Well, I don't think it's a few bytes and not so odd, but I agree that
> this solution isn't the best way.  And, I now remember that this won't
> work anyway, too.  Calling __init from __exit also causes error...

I made this patch because I saw __init calling __exit in yet another
driver (gianfar). Guess I'll just send the old way fix, and remove __exit.


As for calling __init_exit from __exit:
1 - in kernel, there's no __exit => no problem
2 - module, __init_exit is a no-op => no problem

the code in question again:
>  #ifdef MODULE
>  #define __exit               __attribute__ ((__section__(".exit.text")))
> +#define __init_exit
>  #else
>  #define __exit               __attribute_used__ __attribute__ ((__section__(".exit.text")))
> +#define __init_exit  __init
>  #endif

Or maybe it's the name that is confuzing, but it makes sense to me:
__init_exit - you can call it from __init or __exit.
__init_or_exit?


	Domen