Re: [PATCH 0/4] __cpuinitconst and __devinitconst

[Sam Ravnborg <sam@ravnborg.org>]

On Fri, Jan 11, 2008 at 08:44:28PM +0100, Sam Ravnborg wrote:
> Hi Jan.
> 
> On Fri, Jan 11, 2008 at 08:55:29AM +0000, Jan Beulich wrote:
> > Since __cpuinitdata/__devinitdata don't allow const to be specified with
> > them (otherwise .init.data sections with and without the writeable attribute
> > will be generated by the compiler), and since __devinitdata except for
> > embedded systems evaluates to <empty> unconditionally and
> > __cpuinitdata at least in most production kernel configurations also
> > likely evaluates to <empty>, it seems appropriate to add an additional
> > attribute allowing the respective objects to end up in .rodata rather than
> > .data when not used at initialization time only.
> 
> How about a slightly diffrent approach...
> Consider
> __cpuinitconst => function is placed in section .init.const.text
> __devinitconst => function is placed in section .init.const.text
> 
> Then we in the linker scrip can distingush between the two
> and locate the sections as appropriate.
> 
> This will require some updates to modpost but are align with an
> idea I have had for a while.
> All of the following should have dedicated sections associated
> unconditionally:
> 
> __init
> __cpuinit
> __meminit
> __initdata
> __cpuinitdata
> __meminitdata
> 
> And then in the linker script we decide what to do with the section.
> In the built-in case we put them in the "to-be-discarded" section.
> In the module case we put them as today.
> 
> The primary tasks needed to accomplish this is:
> 1) Update all arch linker scripts (and some of them looks ugly)
> 2) Teach modpost about the new sections
> 
> If you following the suggestion above this is a simple step
> in this direction which would be good.

The following patch implment first step in this direction.
It is only an RFC as I have not touched anything else than
64 bit x86 for the arch specific parts.
But it should show what I tried to say above.

On top of x86.git mm-branch.

	Sam

diff --git a/arch/x86/kernel/vmlinux_64.lds.S b/arch/x86/kernel/vmlinux_64.lds.S
index ba8ea97..26c1d81 100644
--- a/arch/x86/kernel/vmlinux_64.lds.S
+++ b/arch/x86/kernel/vmlinux_64.lds.S
@@ -155,12 +155,16 @@ SECTIONS
   __init_begin = .;
   .init.text : AT(ADDR(.init.text) - LOAD_OFFSET) {
 	_sinittext = .;
+	INIT_TEXT
 	*(.init.text)
 	_einittext = .;
   }
-  __initdata_begin = .;
-  .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) { *(.init.data) }
-  __initdata_end = .;
+  .init.data : AT(ADDR(.init.data) - LOAD_OFFSET) {
+	__initdata_begin = .;
+	INIT_DATA
+	__initdata_end = .;
+   }
+
   . = ALIGN(16);
   __setup_start = .;
   .init.setup : AT(ADDR(.init.setup) - LOAD_OFFSET) { *(.init.setup) }
@@ -187,8 +191,12 @@ SECTIONS
   }
   /* .exit.text is discard at runtime, not link time, to deal with references
      from .altinstructions and .eh_frame */
-  .exit.text : AT(ADDR(.exit.text) - LOAD_OFFSET) { *(.exit.text) }
-  .exit.data : AT(ADDR(.exit.data) - LOAD_OFFSET) { *(.exit.data) }
+  .exit.text : AT(ADDR(.exit.text) - LOAD_OFFSET) {
+	EXIT_TEXT
+  }
+  .exit.data : AT(ADDR(.exit.data) - LOAD_OFFSET) {
+	EXIT_DATA
+  }
 
 /* vdso blob that is mapped into user space */
   vdso_start = . ;
diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
index 9f584cc..2f359d9 100644
--- a/include/asm-generic/vmlinux.lds.h
+++ b/include/asm-generic/vmlinux.lds.h
@@ -9,10 +9,48 @@
 /* Align . to a 8 byte boundary equals to maximum function alignment. */
 #define ALIGN_FUNCTION()  . = ALIGN(8)
 
+#ifdef CONFIG_HOTPLUG
+#define DEV_KEEP(sec)
+#define DEV_DISCARD(sec) *(.dev##sec)
+#else
+#define DEV_KEEP(sec)    *(.dev##sec)
+#define DEV_DISCARD(sec)
+#endif
+
+#ifdef CONFIG_HOTPLUG_CPU
+#define CPU_KEEP(sec)
+#define CPU_DISCARD(sec) *(.cpu##sec)
+#else
+#define CPU_KEEP(sec)    *(.cpu##sec)
+#define CPU_DISCARD(sec)
+#endif
+
+#if defined(CONFIG_MEMORY_HOTPLUG) || defined(CONFIG_ACPI_HOTPLUG_MEMORY) \
+        || defined(CONFIG_ACPI_HOTPLUG_MEMORY_MODULE)
+#define MEM_KEEP(sec)
+#define MEM_DISCARD(sec) *(.mem##sec)
+#else
+#define MEM_KEEP(sec)    *(.mem##sec)
+#define MEM_DISCARD(sec)
+#endif
+
+
 /* .data section */
 #define DATA_DATA							\
 	*(.data)							\
 	*(.data.init.refok)						\
+	DEV_KEEP(init.data)						\
+	DEV_KEEP(init.data.const)					\
+	DEV_KEEP(exit.data)						\
+	DEV_KEEP(exit.data.const)					\
+	CPU_KEEP(init.data)						\
+	CPU_KEEP(init.data.const)					\
+	CPU_KEEP(exit.data)						\
+	CPU_KEEP(exit.data.const)					\
+	MEM_KEEP(init.data)						\
+	MEM_KEEP(init.data.const)					\
+	MEM_KEEP(exit.data)						\
+	MEM_KEEP(exit.data.const)					\
 	. = ALIGN(8);							\
 	VMLINUX_SYMBOL(__start___markers) = .;				\
 	*(__markers)							\
@@ -159,7 +197,14 @@
 		ALIGN_FUNCTION();					\
 		*(.text)						\
 		*(.text.init.refok)					\
-		*(.exit.text.refok)
+		*(.exit.text.refok)					\
+	DEV_KEEP(init.text)						\
+	DEV_KEEP(exit.text)						\
+	CPU_KEEP(init.text)						\
+	CPU_KEEP(exit.text)						\
+	MEM_KEEP(init.text)						\
+	MEM_KEEP(exit.text)
+
 
 /* sched.text is aling to function alignment to secure we have same
  * address even at second ld pass when generating System.map */
@@ -255,6 +300,36 @@
   	*(.initcall7.init)						\
   	*(.initcall7s.init)
 
+#define INIT_DATA							\
+	*(.init.data)							\
+	DEV_DISCARD(init.data)						\
+	DEV_DISCARD(init.data.const)					\
+	CPU_DISCARD(init.data)						\
+	CPU_DISCARD(init.data.const)					\
+	MEM_DISCARD(init.data)						\
+	MEM_DISCARD(init.data.const)
+
+#define INIT_TEXT							\
+	*(.init.text)							\
+	DEV_DISCARD(init.text)						\
+	CPU_DISCARD(init.text)						\
+	MEM_DISCARD(init.text)
+
+#define EXIT_DATA							\
+	*(.exit.data)							\
+	DEV_DISCARD(exit.data)						\
+	DEV_DISCARD(exit.data.const)					\
+	CPU_DISCARD(exit.data)						\
+	CPU_DISCARD(exit.data.const)					\
+	MEM_DISCARD(exit.data)						\
+	MEM_DISCARD(exit.data.const)
+
+#define EXIT_TEXT							\
+	*(.exit.text)							\
+	DEV_DISCARD(exit.text)						\
+	CPU_DISCARD(exit.text)						\
+	MEM_DISCARD(exit.text)
+
 #define PERCPU(align)							\
 	. = ALIGN(align);						\
 	__per_cpu_start = .;						\
diff --git a/include/linux/init.h b/include/linux/init.h
index 86e7e94..b72897d 100644
--- a/include/linux/init.h
+++ b/include/linux/init.h
@@ -254,42 +254,26 @@ void __init parse_early_param(void);
 #define __initdata_or_module __initdata
 #endif /*CONFIG_MODULES*/
 
-#ifdef CONFIG_HOTPLUG
-#define __devinit
-#define __devinitdata
-#define __devexit
-#define __devexitdata
-#else
-#define __devinit __init
-#define __devinitdata __initdata
-#define __devexit __exit
-#define __devexitdata __exitdata
-#endif
-
-#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PM_CPUINIT)
-#define __cpuinit
-#define __cpuinitdata
-#define __cpuexit
-#define __cpuexitdata
-#else
-#define __cpuinit	__init
-#define __cpuinitdata __initdata
-#define __cpuexit __exit
-#define __cpuexitdata	__exitdata
-#endif
-
-#if defined(CONFIG_MEMORY_HOTPLUG) || defined(CONFIG_ACPI_HOTPLUG_MEMORY) \
-	|| defined(CONFIG_ACPI_HOTPLUG_MEMORY_MODULE)
-#define __meminit
-#define __meminitdata
-#define __memexit
-#define __memexitdata
-#else
-#define __meminit	__init
-#define __meminitdata __initdata
-#define __memexit __exit
-#define __memexitdata	__exitdata
-#endif
+#define __devinit        __attribute__ ((__section__ (".devinit.text")))
+#define __devinitdata    __attribute__ ((__section__ (".devinit.data")))
+#define __devinitcdata   __attribute__ ((__section__ (".devinit.data.const")))
+#define __devexit        __attribute__ ((__section__ (".devexit.text")))
+#define __devexitdata    __attribute__ ((__section__ (".devexit.data")))
+#define __devexitcdata   __attribute__ ((__section__ (".devexit.data.const")))
+
+#define __cpuinit        __attribute__ ((__section__ (".cpuinit.text")))
+#define __cpuinitdata    __attribute__ ((__section__ (".cpuinit.data")))
+#define __cpuinitcdata   __attribute__ ((__section__ (".cpuinit.data.const")))
+#define __cpuexit        __attribute__ ((__section__ (".cpuexit.text")))
+#define __cpuexitdata    __attribute__ ((__section__ (".cpuexit.data")))
+#define __cpuexitcdata   __attribute__ ((__section__ (".cpuexit.data.const")))
+
+#define __meminit        __attribute__ ((__section__ (".meminit.text")))
+#define __meminitdata    __attribute__ ((__section__ (".meminit.data")))
+#define __meminitcdata   __attribute__ ((__section__ (".meminit.data.const")))
+#define __memexit        __attribute__ ((__section__ (".memexit.text")))
+#define __memexitdata    __attribute__ ((__section__ (".memexit.data")))
+#define __memexitcdata   __attribute__ ((__section__ (".memexit.data.const")))
 
 /* Functions marked as __devexit may be discarded at kernel link time, depending
    on config options.  Newer versions of binutils detect references from