[Proposal] sh: Suspend in Ram on sh4

Commit Message

Francesco VIRLINZI March 10, 2009, 1:24 p.m. UTC

Hi all
I would propose an other version of suspend in ram.

In ST we have several constraints:
- Several SOCs with different Clock IPs
- No cache memory mapped (it was removed in recently version of ST40)
- No IPREF instruction..
- 29/32 bits supports (where in 32 bits we are managing the PMB in a 
different manner respect the original kernel.org code).

For these reason an hard-coded suspend assembly code isn't easy to use 
for us.

To address the issue we are using a mini interpreter able to execute 
macro instructions SOC based.
The assembly code preloads (with a jump sequence) in cache (both I$ and 
D$) the 'instruction-table'
and after that it executes the instruction-table.

This solution allow me to have no assembly code SOCs specific and
for each SOC I have only to produce the right "tables" for standby 
and/or mem-standby.

Moreover a 3rd table is supported to be able to produce or save data on 
the fly for the standby it-self (for example if we have different action 
on the same SOC but different Cut).

Let me know.
Regards
Francesco

P.S.: in attach an example on how I'm using the suspend core code.

Comments

Francesco VIRLINZI March 10, 2009, 1:27 p.m. UTC | #1

Sorry here the example on now to use.
Regards
 Francesco
>
>
> P.S.: in attach an example on how I'm using the suspend core code.

Magnus Damm March 11, 2009, 1:36 p.m. UTC | #2

Hi Francesco!

On Tue, Mar 10, 2009 at 10:24 PM, Francesco VIRLINZI
<[e-mail]> wrote:
> Hi all
> I would propose an other version of suspend in ram.
>
> In ST we have several constraints:
> - Several SOCs with different Clock IPs
> - No cache memory mapped (it was removed in recently version of ST40)
> - No IPREF instruction..
> - 29/32 bits supports (where in 32 bits we are managing the PMB in a
> different manner respect the original kernel.org code).

Just curious, how many different SoCs do you support with suspend?

> For these reason an hard-coded suspend assembly code isn't easy to use for
> us.
>
> To address the issue we are using a mini interpreter able to execute macro
> instructions SOC based.

Some week ago I was considering doing something similar, but then I
decided not to. The reason for this was simplicity. My suspend patch
only supports a few processors, so it's probably not fair to compare
it with your case. So please take this with a pinch of salt. And
correct me when I'm wrong.

The only code that needs special care is the self refresh code and
whatever runs when the system ram is put in self refresh. What happens
before and after can be written in C or assembly as usual. Having some
assembly code snippet that does this is very simple. The code for
suspend that I posted recently is relocatable and does exactly this.
It should be possible to load that into the instruction cache too
unless I'm mistaken.

I realize that the self-refresh mode switching code may vary with
processor type and maybe board type as well. And having special case
handling by ifdefs in the assembly snippet is not exacly clean.

> The assembly code preloads (with a jump sequence) in cache (both I$ and D$)
> the 'instruction-table'
> and after that it executes the instruction-table.

I tried to figure out how this cache population is working, but i
still don't understand. =) Something is magical with your JUMPER().
And you need both populate both instruction and data cache, right?

> This solution allow me to have no assembly code SOCs specific and
> for each SOC I have only to produce the right "tables" for standby and/or
> mem-standby.

I understand. This is pretty neat.

> Moreover a 3rd table is supported to be able to produce or save data on the
> fly for the standby it-self (for example if we have different action on the
> same SOC but different Cut).

How do you save data without writing to the self-refreshing ram?
Copy-back configuration with guaranteed space?

I like the idea with an interpreter. But I get the feeling that you're
using the interpreter to set all sorts of register configurations -
not only the ones that setup/restore self-refresh and sleep. Am I
wrong?

I'm not sure what the best way forward is. Are you planning on
submitting upstream support for some SoC? That would be nice so we
could share the same power management code. If not then it's maybe
best that I just fix up and resend my suspend patch and we can tie in
some interpreter whenever it's needed.

Thanks for your help so far!

Cheers,

/ magnus
--
To unsubscribe from this list: send the line "unsubscribe linux-sh" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Francesco VIRLINZI March 11, 2009, 2:31 p.m. UTC | #3

Hi again Magnus
> ...
> Just curious, how many different SoCs do you support with suspend?
Currently seven... ~11-12 for the end of the year...
> ...
>
>
> The only code that needs special care is the self refresh code and
> whatever runs when the system ram is put in self refresh. What happens
> before and after can be written in C or assembly as usual.
Agreed!
>> The assembly code preloads (with a jump sequence) in cache (both I$ and D$)
>> the 'instruction-table'
>> and after that it executes the instruction-table.
>
> I tried to figure out how this cache population is working, but i
> still don't understand. =) Something is magical with your JUMPER().
> And you need both populate both instruction and data cache, right?
Yes you are right!
As I said in the ST40 we removed the ipref instruction and we removed also
 the cache memory mapped therefore the only way to preload the code  in 
the ICache is executing it..

When the linux calls the 'sh4_suspend' the CPU begins a sequence of 
jumps "to ping" all the Icache line (but
 I'm not really executing code... I'm only jumping from a label to the 
next label...) until the
 end of the code where it jumps on 'sh4_really_suspend' where the CPU 
really executes the code (now in Icache).
>
> How do you save data without writing to the self-refreshing ram?
Both the 'instruction table' and the 'writable-table' are preloaded in 
Dcache.
> Copy-back configuration with guaranteed space?
>
> I like the idea with an interpreter. But I get the feeling that you're
> using the interpreter to set all sorts of register configurations -
> not only the ones that setup/restore self-refresh and sleep. Am I
> wrong?
No you aren't.
You are  right, I sets also some clocks but using the interpreter 
basically you can do what you want for example
 in a SOC the last thing I do to enter in standby isn't a sleep... it's 
a write in the ClockIP to turn-off the sh4_clk...
 the wakeup event routed to the ClockIP will ("automatically") turn-on 
again the sh4_clk.... therefore it's a kind
 of 'hard' sleep instruction.... (without sleep instruction)

Also in this case the code to manage the clock has to be in icache.
>
> I'm not sure what the best way forward is. Are you planning on
> submitting upstream support for some SoC?
Not so easy submit the per SOC code also because we use an older kernel 
than the original kernel.org (I'm on 2.6.23).
>  That would be nice so we
> could share the same power management code.
I'm already sharing this code for that reason.
Ciao
  Francesco

--
To unsubscribe from this list: send the line "unsubscribe linux-sh" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Magnus Damm March 19, 2009, 5:52 a.m. UTC | #4

On Wed, Mar 11, 2009 at 11:31 PM, Francesco VIRLINZI
<[e-mail]> wrote:
>> I'm not sure what the best way forward is. Are you planning on
>> submitting upstream support for some SoC?
>
> Not so easy submit the per SOC code also because we use an older kernel than
> the original kernel.org (I'm on 2.6.23).

I see. The joys of old kernels. =)

>>  That would be nice so we
>> could share the same power management code.
>
> I'm already sharing this code for that reason.

I appreciate your effort. You may have noted that we cherry pick
certain parts of the code that you post. Accepting everything would of
course be better, but there is some hesitation since we don't have the
full picture of your code base.

In my opinion, the best way to add new interfaces or change existing
ones is to also submit code that is using your interfaces. WIthout
such code it's difficult to understand the reason behind the change.

I wonder if it would be possible to take a single cpu and a single
board from your out of tree code base, and add support for that system
upstream. Then we can tie in whatever interfaces to that code. This
may take quite a bit of time though, but it's probably worth doing in
the long term.

Also, please try to submit your patches inline if possible. It makes
review much easier. I use sendpatchset for this purpose.

Thanks for your help so far!

/ magnus
--
To unsubscribe from this list: send the line "unsubscribe linux-sh" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

From dc7e67600f58bf04eae348c43854c1757f052f19 Mon Sep 17 00:00:00 2001
From: Francesco Virlinzi <[e-mail]>
Date: Mon, 9 Mar 2009 16:05:18 +0100
Subject: [PATCH] Suspend in Ram support on Sh4


Signed-off-by: Francesco Virlinzi <[e-mail]>
---
 arch/sh/Kconfig                       |    1 +
 arch/sh/include/cpu-sh4/suspend.h     |  260 ++++++++++++++++++++++++
 arch/sh/kernel/cpu/sh4/Makefile       |    4 +
 arch/sh/kernel/cpu/sh4/suspend-core.S |  350 +++++++++++++++++++++++++++++++++
 arch/sh/kernel/cpu/sh4/suspend.c      |  113 +++++++++++
 5 files changed, 728 insertions(+), 0 deletions(-)
 create mode 100644 arch/sh/include/cpu-sh4/suspend.h
 create mode 100644 arch/sh/kernel/cpu/sh4/suspend-core.S
 create mode 100644 arch/sh/kernel/cpu/sh4/suspend.c

diff --git a/arch/sh/Kconfig b/arch/sh/Kconfig
index a4c2c84..4b3504b 100644
--- a/arch/sh/Kconfig
+++ b/arch/sh/Kconfig
@@ -157,6 +157,7 @@  config CPU_SH4
 	bool
 	select CPU_HAS_INTEVT
 	select CPU_HAS_SR_RB
+	select ARCH_SUSPEND_POSSIBLE
 	select CPU_HAS_PTEA if !CPU_SH4A || CPU_SHX2
 	select CPU_HAS_FPU if !CPU_SH4AL_DSP
 
diff --git a/arch/sh/include/cpu-sh4/suspend.h b/arch/sh/include/cpu-sh4/suspend.h
new file mode 100644
index 0000000..4d7dfb7
--- /dev/null
+++ b/arch/sh/include/cpu-sh4/suspend.h
@@ -0,0 +1,260 @@ 
+/*
+ * -------------------------------------------------------------------------
+ * Copyright (C) 2008  STMicroelectronics
+ * Copyright (C) 2009  STMicroelectronics
+ * Author: Francesco M. Virlinzi  <[e-mail]>
+ *
+ * May be copied or modified under the terms of the GNU General Public
+ * License V.2 ONLY.  See linux/COPYING for more information.
+ *
+ * ------------------------------------------------------------------------- */
+#ifndef __suspend_sh4_h__
+#define __suspend_sh4_h__
+
+#define BASE_DATA		(0x0)
+	/* to identify the ClockGenA registers */
+#define BASE_CLK		(0x1)
+	/* to identify the ClockGenB registers */
+#define BASE_CLKB		(0x2)
+	/* to identify the Sysconf registers */
+#define BASE_SYS		(0x3)
+
+
+#define OP_END			(0*4)	/* no more data in the table */
+#define OP_END_NO_SLEEP		(1*4)
+#define OP_SOURCE		(2*4)
+#define OP_LOAD			(3*4)	/* load  @(offset, Reg_idx) */
+#define OP_ILOAD_SRC0		(4*4)	/* load_imm (from itable) on r1 */
+#define OP_ILOAD_SRC1		(5*4)	/* load_imm (from itable) on r3 */
+#define OP_ILOAD_SRC2		(6*4)	/* load_imm (from itable) on r4 */
+#define OP_ILOAD_DEST		(7*4)	/* load_imm (from table) on r2 */
+
+#define OP_STORE		(8*4)	/* store @(offset, Reg_idx) */
+#define OP_OR			(9*4)
+#define OP_AND			(10*4)
+#define OP_NOT			(11*4)
+/* WHILE_EQ (idx, offset, mask, value)
+ * wait until the mask bits is equal to value
+ */
+#define OP_WHILE_EQ		(12*4)
+/* WHILE_NEQ (idx, offset, mask, value)
+ * wait until the mask bits isn't equal to value
+ */
+#define OP_WHILE_NEQ		(13*4)
+
+#define OP_DELAY		(14*4)	/* A loop delay */
+
+#define OP_LOAD_SRC0		(15*4)	/* Load SRC_0 from resources */
+#define OP_LOAD_SRC1		(16*4)	/* Load SRC_1 from  */
+#define OP_LOAD_SRC2		(17*4)	/* Load SRC_2 from table */
+#define _OPCODE_TABLE_SIZE_	3
+
+#ifndef __ASSEMBLY__
+
+struct sh4_suspend_t {
+	unsigned long *iobase;   /* the external iomemory resource 		*/
+	unsigned long l_p_j;
+	unsigned long wrt_tbl; /* the writeable table address			*/
+	unsigned long wrt_size; /* the writeable table size in dcache line!	*/
+	unsigned long stby_tbl;	/* the standby instruction table address	*/
+	unsigned long stby_size;/* the standby instruction table size in dcache line*/
+	unsigned long mem_tbl;	/* the mem instruction table address		*/
+	unsigned long mem_size;	/* the mem instruction table size in dcache line*/
+	int (*evt_to_irq)(unsigned long evt); /* translate the INTEVT code
+					       * to the irq number */
+	struct platform_suspend_ops ops;
+};
+
+int sh_register_suspend(struct sh4_suspend_t *pdata);
+
+/* Operations */
+#define _OR()					OP_OR
+#define _AND()					OP_AND
+#define _NOT()					OP_NOT
+#define _DELAY()				OP_DELAY
+#define _WHILE_NEQ()				OP_WHILE_NEQ
+#define _WHILE_EQ()				OP_WHILE_EQ
+#define _LOAD()					OP_LOAD
+#define _STORE()				OP_STORE
+/*
+ * N.S.: DATA_LOAD and DATA_STORE work directly on DEST reg.
+ *       To load something on SCR0, SRC1 and SRC2 Use
+ *       following instructions
+ */
+#define _LOAD_SRC0()				OP_LOAD_SRC0
+#define _LOAD_SRC1()				OP_LOAD_SRC1
+#define _LOAD_SRC2()				OP_LOAD_SRC2
+
+#define _END()					OP_END
+#define _END_NO_SLEEP()				OP_END_NO_SLEEP
+
+#define DATA_SOURCE(idx)					\
+	OP_SOURCE, BASE_DATA, (4*(idx))
+
+#define RAW_SOURCE(orig, reg_offset)				\
+	OP_SOURCE, (4*(orig)), (reg_offset)
+
+#define SYS_SOURCE(reg_offset)					\
+	OP_SOURCE, BASE_SYS, (reg_offset)
+
+#define CLK_SOURCE(reg_offset)					\
+	OP_SOURCE, BASE_CLK, (reg_offset)
+
+#define DATA_LOAD(idx)				DATA_SOURCE(idx), _LOAD()
+#define DATA_STORE(idx)				DATA_SOURCE(idx), _STORE()
+
+/* a raw load */
+#define RAW_LOAD(base, reg_offset)		RAW_SOURCE(base, reg_offset), _LOAD()
+
+#define SYS_LOAD(reg_offset)			RAW_LOAD(BASE_SYS, reg_offset)
+#define CLK_LOAD(reg_offset)			RAW_LOAD(BASE_CLK, reg_offset)
+#define CLKB_LOAD(reg_offset)			RAW_LOAD(BASE_CLKB, reg_offset)
+
+/* A raw store          */
+#define RAW_STORE(base, reg_offset)		RAW_SOURCE(base, reg_offset), _STORE()
+#define SYS_STORE(reg_offset)			RAW_STORE(BASE_SYS, reg_offset)
+#define CLK_STORE(reg_offset)			RAW_STORE(BASE_CLK, reg_offset)
+#define CLKB_STORE(reg_offset)			RAW_STORE(BASE_CLKB, reg_offset)
+
+#define IMMEDIATE_SRC0(value)			OP_ILOAD_SRC0, (value)
+#define IMMEDIATE_SRC1(value)			OP_ILOAD_SRC1, (value)
+#define IMMEDIATE_SRC2(value)			OP_ILOAD_SRC2, (value)
+#define IMMEDIATE_DEST(value)			OP_ILOAD_DEST, (value)
+
+/* Set Or-ing the bits in the register */
+#define RAW_OR_LONG(orig, reg_offset, or_bits)	\
+	RAW_SOURCE(orig, reg_offset),		\
+	 _LOAD(),				\
+	IMMEDIATE_SRC0(or_bits),		\
+	_OR(),					\
+	RAW_SOURCE(orig, reg_offset),		\
+	_STORE()
+
+#define SYS_OR_LONG(reg_offset, or_bits)			\
+	RAW_OR_LONG(BASE_SYS, reg_offset, or_bits)
+
+#define CLK_OR_LONG(reg_offset, or_bits)			\
+	RAW_OR_LONG(BASE_CLK, reg_offset, or_bits)
+
+#define CLKB_OR_LONG(reg_offset, or_bits)			\
+	RAW_OR_LONG(BASE_CLKB, reg_offset, or_bits)
+
+
+#define DATA_OR_LONG(idx_mem, idx_mask)			\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_LOAD(),					\
+	DATA_SOURCE(idx_mask),				\
+	_LOAD_SRC0(),					\
+	_OR(),						\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_STORE()
+
+#define DATA_AND_LONG(idx_mem, idx_mask)		\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_LOAD(),					\
+	DATA_SOURCE(idx_mask),				\
+	_LOAD_SRC0(),					\
+	_AND(),						\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_STORE()
+
+#define DATA_AND_NOT_LONG(idx_mem, idx_mask)		\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_LOAD(),					\
+	DATA_SOURCE(idx_mask),				\
+	_LOAD_SRC1(),					\
+	_NOT(),						\
+	_AND(),						\
+	DATA_SOURCE(idx_mem),				\
+	_LOAD_SRC0(),					\
+	_STORE()
+
+/* Set And-ing the bits in the register */
+#define RAW_AND_LONG(orig, reg_offset, and_bits)	\
+	RAW_SOURCE(orig, reg_offset),			\
+	_LOAD(), /* dest = @(iomem) */			\
+	IMMEDIATE_SRC0(and_bits),			\
+	_AND(),	 /* dest &= src0 */			\
+	RAW_SOURCE(orig, reg_offset),			\
+	_STORE() /* @(iomem) = dest */
+
+#define SYS_AND_LONG(reg_offset, and_bits)			\
+		RAW_AND_LONG(BASE_SYS, reg_offset, and_bits)
+
+#define CLK_AND_LONG(reg_offset, and_bits)			\
+		RAW_AND_LONG(BASE_CLK, reg_offset, and_bits)
+
+#define CLKB_AND_LONG(reg_offset, and_bits)			\
+		RAW_AND_LONG(BASE_CLKB, reg_offset, and_bits)
+
+/* Standard Poke */
+#define RAW_POKE(base, reg_offset, value)		\
+	IMMEDIATE_DEST(value),				\
+	RAW_SOURCE(base, reg_offset),			\
+	_STORE()
+
+#define SYS_POKE(reg_offset, value)			\
+	RAW_POKE(BASE_SYS, reg_offset, value)
+
+#define CLK_POKE(reg_offset, value)			\
+	RAW_POKE(BASE_CLK, reg_offset, value)
+
+#define CLKB_POKE(reg_offset, value)			\
+	RAW_POKE(BASE_CLKB, reg_offset, value)
+
+/* While operation */
+#define RAW_WHILE_EQ(orig, offset, mask, value)		\
+	RAW_SOURCE(orig, offset),			\
+	IMMEDIATE_SRC1(mask),				\
+	IMMEDIATE_SRC2(value),				\
+	_WHILE_EQ()
+
+#define RAW_WHILE_NEQ(orig, offset, mask, value)	\
+	RAW_SOURCE(orig, offset),			\
+	IMMEDIATE_SRC1(mask),				\
+	IMMEDIATE_SRC2(value),				\
+	_WHILE_NEQ()
+
+#define DATA_WHILE_EQ(idx_iomem, idx_mask, idx_value)	\
+	DATA_SOURCE(idx_value),				\
+	_LOAD_SRC2(),					\
+	DATA_SOURCE(idx_mask),				\
+	_LOAD_SRC1(),					\
+	DATA_SOURCE(idx_iomem),				\
+	_LOAD_SRC0(),					\
+	_WHILE_EQ()
+
+#define DATA_WHILE_NEQ(idx_iomem, idx_mask, idx_value)	\
+	DATA_SOURCE(idx_value),				\
+	_LOAD_SRC2(),					\
+	DATA_SOURCE(idx_mask),				\
+	_LOAD_SRC1(),					\
+	DATA_SOURCE(idx_iomem),				\
+	_LOAD_SRC0(),					\
+	_WHILE_NEQ()
+
+#define SYS_WHILE_EQ(offset, mask, value)		\
+	RAW_WHILE_EQ(BASE_SYS, offset, mask, value)
+
+#define CLK_WHILE_EQ(offset, mask, value)		\
+	RAW_WHILE_EQ(BASE_CLK, offset, mask, value)
+
+#define CLKB_WHILE_EQ(offset, mask, value)		\
+	RAW_WHILE_EQ(BASE_CLKB, offset, mask, value)
+
+#define SYS_WHILE_NEQ(offset, mask, value)		\
+	RAW_WHILE_NEQ(BASE_SYS, offset, mask, value)
+
+#define CLK_WHILE_NEQ(offset, mask, value)		\
+	RAW_WHILE_NEQ(BASE_CLK, offset, mask, value)
+
+#define CLKB_WHILE_NEQ(offset, mask, value)		\
+	RAW_WHILE_NEQ(BASE_CLKB, offset, mask, value)
+
+#endif
+#endif
diff --git a/arch/sh/kernel/cpu/sh4/Makefile b/arch/sh/kernel/cpu/sh4/Makefile
index d608557..381a331 100644
--- a/arch/sh/kernel/cpu/sh4/Makefile
+++ b/arch/sh/kernel/cpu/sh4/Makefile
@@ -27,3 +27,7 @@  endif
 clock-$(CONFIG_CPU_SUBTYPE_SH4_202)	+= clock-sh4-202.o
 
 obj-y	+= $(clock-y)
+
+ifdef CONFIG_SUSPEND
+obj-y					+= suspend.o suspend-core.o
+endif
diff --git a/arch/sh/kernel/cpu/sh4/suspend-core.S b/arch/sh/kernel/cpu/sh4/suspend-core.S
new file mode 100644
index 0000000..79fa5f3
--- /dev/null
+++ b/arch/sh/kernel/cpu/sh4/suspend-core.S
@@ -0,0 +1,350 @@ 
+/*
+ * -------------------------------------------------------------------------
+ * <linux_root>/arch/sh/kernel/cpu/sh4/suspend-core.S
+ * -------------------------------------------------------------------------
+ * Copyright (C) 2008  STMicroelectronics
+ * Copyright (C) 2009  STMicroelectronics
+ * Author: Francesco M. Virlinzi  <[e-mail]>
+ *
+ * May be copied or modified under the terms of the GNU General Public
+ * License V.2 ONLY.  See linux/COPYING for more information.
+ *
+ * ------------------------------------------------------------------------- */
+
+#include <cpu-sh4/suspend.h>
+#include <cpu-sh4/cpu/mmu_context.h>
+/*
+ * Some register are dedicated for special purpose
+ */
+#define IOREGS_BASE		r14
+#define ITABLE_ADDRESS		r13
+#define DTABLE_ADDRESS		r12
+#define DELAY_REG		r11
+
+#define OFFSET_IOBASE		0x0
+#define OFFSET_LPJ		0x4
+#define OFFSET_DTABLE		0x8
+#define OFFSET_DTABLE_SIZE	0xc
+#define REG_INSTR		r5
+#define REG_INSTR_END		r6
+
+
+#define JUMPER()		bra 201f;	\
+				 nop;		\
+			200:	bra 200f;	\
+				 nop;	;	\
+			201:
+
+#undef ENTRY
+#define ENTRY(name, align)	\
+  .balign align;		\
+  .globl name;			\
+  name:
+
+.text
+ENTRY(sh4_suspend, 32)		! to be icache aligned
+	bra 200f		! start the jump sequence
+	 nop
+sh4_really_suspend:
+	mov.l   r14, @-r15
+	mov.l   r13, @-r15
+	mov.l   r12, @-r15
+	mov.l   r11, @-r15
+	mov.l   r10, @-r15
+	mov.l   r9,  @-r15
+	mov.l   r8,  @-r15
+	sts.l	pr,  @-r15	! save the pr (we can call other function)
+	stc.l	sr,  @-r15
+	stc	vbr, r0
+
+	JUMPER()
+
+	mov.l	r0,  @-r15	! save the original vbr on the stack
+
+	mov.l	@(OFFSET_IOBASE, r4), IOREGS_BASE	! save ioregs address
+	mov.l	@(OFFSET_LPJ, r4),    DELAY_REG
+	mov.l	@(OFFSET_DTABLE, r4), DTABLE_ADDRESS
+	mov	REG_INSTR,  	      ITABLE_ADDRESS	! the instruction table!
+
+/*
+ *	runs the suspend iteration tables
+ */
+	bsr	do_decode
+	 nop
+
+	cmp/eq  #1, r0		! check if we have to sleep or not
+	bt	__resume	! it depends if we complete the table
+				! with END or END_NO_SLEEP
+
+	mova	vbr_base_suspend, r0	! this mova isn't a problem
+					! because vbr_base_suspend is
+					! 4 bytes alligned
+	ldc	r0, vbr			! install the wakeup_interrupt
+	mov	#0x3c, r1
+
+	JUMPER()
+
+	shll2	r1
+	not	r1,   r1
+	stc	sr,   r0
+	and	r1,   r0
+	ldc	r0,   sr		! enable the interrups
+
+	sleep				! SLEEP!!!
+
+/*
+ *	runs the resume instruction tables
+ */
+__resume:
+	nop
+	bsr     do_decode
+	 nop
+
+	mov.l	@r15+, r0
+	ldc	r0,    vbr		! Restore the original vbr
+	mov.l	@r15+, r0		! Original sr (on interrupts disabled)
+
+	JUMPER()
+
+	lds.l	@r15+, pr
+	mov.l   @r15+, r8
+	mov.l   @r15+, r9
+	mov.l   @r15+, r10
+	mov.l   @r15+, r11
+	mov.l   @r15+, r12
+	mov.l   @r15+, r13
+	mov.l	@r15+, r14
+	mov.l   1f,  r1
+	mov.l   @r1, r1			! who waked up
+	ldc	r0, sr			! Restore the original sr
+	rts
+	 mov	r1, r0			! who waked up
+	JUMPER()
+
+.balign 4
+1:			.long	INTEVT
+
+
+.balign       	1024,	0,	1024
+vbr_base_suspend:
+	.long   0
+.balign         1024,   0,      1024
+
+	.long 0
+.balign         512,	0,	512
+wakeup_interrupt:
+	JUMPER()
+	!	Disable the interrupts in the ssr
+	!	and returns to the context (asap)....
+	stc	ssr,   r0
+	or	#0xf0, r0
+	ldc	r0, ssr		! to avoit recursive irq...
+				! this means the context will be resumed
+				! with interrupt disabled!!!
+/*
+ * Here we could have a problem (a sleep with interrupt disables!!!)
+ * It could happen if we detect an interrupt between
+ * the enabled irq and the sleep!!!
+ * Restoring the (raw) spc we will go to execute a sleep with the
+ * interrupt disabled !!!!
+ * To avoid that in any case we will return on the resume_address
+ * label
+ */
+	mov.l	resume_address, r0
+	ldc	r0, spc
+	rte
+	 nop
+
+200:
+/*
+ *	load the instruction datas
+ */
+	mov.l   resume_address,	r0
+	mov	REG_INSTR, r0		/* start address I-table */
+	mov	REG_INSTR_END, r1	/* I-table size */
+	tst	r1, r1
+2:
+	mov.l   @r0, r2			/* Load the I-tables in cache */
+	add	#32, r0
+        bf/s	2b
+         dt	r1
+/*
+ *      load the writeable datas
+ */
+	mov.l	@(OFFSET_DTABLE, r4), r0
+	mov.l	@(OFFSET_DTABLE_SIZE, r4),   r1
+	tst	r1, r1
+2:
+	mov.l   @r0, r2			/* Load the d-tables in cache */
+	add	#32, r0
+	bf/s	2b
+	 dt	r1
+	bra	200f
+	 nop
+
+.balign 4
+resume_address:		.long __resume
+
+#define SRC0		r1
+#define SRC1		r2
+#define SRC2		r3
+#define DEST		r4
+#define TMP		r5
+
+.balign 2
+	JUMPER()
+ENTRY(do_decode, 2)
+	mov.l	@ITABLE_ADDRESS+, r0	! opcode
+	mov.l	s_jmp_table_address, TMP
+	mov.l	@(r0, TMP), TMP
+	jmp	@TMP
+	 nop
+
+l_end:	! OP_END
+	rts				! Return point
+	 mov	#0, r0			! r0 = 0 to say return and sleep
+
+	JUMPER()
+
+l_end_no_sleep:	! OP_END_NO_SLEEP
+	rts				! Return point
+	 mov	#1, r0			! r0 = 1 to say return and Don't sleep
+
+l_source: ! OP_SOURCE
+	mov.l	@ITABLE_ADDRESS+, r0	! load the source reg base
+	mov.l	@(r0, IOREGS_BASE), TMP	! load ioreg in r5
+	mov.l	@ITABLE_ADDRESS+, SRC0	! load the offset
+	bra     do_decode
+	 add	TMP, SRC0		! r2 = the iomem address of source
+
+	JUMPER()
+
+	/* Load a @SRC0 in Dest*/
+l_load: ! #OP_LOAD
+	bra	do_decode
+	 mov.l	@SRC0, DEST		! load the value
+
+	/* Load a value from table in SRC0 */
+l_iload_scr0: ! OP_ILOAD_SRC0
+	bra	do_decode
+	 mov.l	@ITABLE_ADDRESS+, SRC0	! the value is in SRC0 !!!
+
+	/* Load a value from table in SRC1 */
+l_iload_src1: ! OP_ILOAD_SRC1
+	bra	do_decode
+	 mov.l	@ITABLE_ADDRESS+, SRC1	! the value is in SRC1 !!!
+
+	/* Load a value from table in SRC2 */
+l_iload_src2: ! OP_ILOAD_SRC2
+	bra	do_decode
+	 mov.l @ITABLE_ADDRESS+, SRC2	! the value is in SRC2 !!!
+
+	/* Load a value from table in the DEST */
+l_iload_dest: ! OP_ILOAD_DEST
+	bra	do_decode
+	 mov.l @ITABLE_ADDRESS+, DEST
+
+	JUMPER()
+
+	/* Store DEST value in @SRC0 */
+l_store: ! OP_STORE
+	bra	do_decode
+	 mov.l DEST, @(0,SRC0)		! store the value
+
+	/* Or operation: DEST |= SRC0 */
+l_or:	! OP_OR
+	bra	do_decode
+	 or	SRC0, DEST
+
+	/* And operation: DEST &= SRC0 */
+l_and:	! OP_AND
+	bra	do_decode
+	 and	SRC0, DEST
+
+	/* Not operation: SRC0 = ~SRC1*/
+	/* It's a bit dirty that the NOT operation works on SRC1 instead of DEST or SRC0*/
+l_not:	! OP_NOT
+	bra	do_decode
+	 not	SRC1, SRC0
+
+	JUMPER()
+
+	/* While bits equal to value. This operation assumes:
+		- SRC0: the iomemory address
+		- SRC1: the bitmask
+		- SRC2: the result
+	*/
+l_while_eq: !	OP_WHILE_EQ
+	mov.l	@SRC0, TMP
+2:	and     SRC1, TMP
+	cmp/eq	SRC2, TMP			! (@SRC0 and SRC1) ?!= SRC2)
+	bt/s	2b
+	 mov.l   @SRC0, TMP
+	bra	do_decode
+	 nop
+
+	JUMPER()
+	/* While bits not equal to value. This operation assumes:
+		   - SRC0: the iomemory address
+		   - SRC1: the bitmask
+		   - SRC2: the result
+	*/
+l_while_neq: ! OP_WHILE_NEQ
+	mov.l	@SRC0, TMP
+2:	and	SRC1, TMP
+	cmp/eq  SRC2, TMP		! (@SRC0 and SRC1) ?== SRC2)
+	bf/s	2b
+	 mov.l	@SRC0, TMP
+	bra	do_decode
+	 nop
+
+	JUMPER()
+
+	/* Delay operation */
+l_delay: ! OP_DELAY
+	mov     DELAY_REG, TMP
+	tst	TMP, TMP
+2:
+	bf/s   2b
+	 dt	TMP
+	bra	do_decode
+	 nop
+
+	/*  SCR0 = @SRC0 */
+l_load_src0: ! OP_LOAD_SRC0
+	mov.l	@SRC0, SRC0
+	bra	do_decode
+	 nop
+
+	JUMPER()
+
+l_load_src1: ! OP_LOAD_SRC1	=> SRC1 = @SRC0
+	mov.l  @SRC0, SRC1
+	bra	do_decode
+	 nop
+
+l_load_src2: ! OP_LOAD_SRC2	=> SRC2 = @SRC0
+	mov.l  @SRC0, SRC2
+	bra	do_decode
+	 nop
+
+200:	! Preload the jump table
+	mov.l	s_jmp_table_address, r1
+	mov	#_OPCODE_TABLE_SIZE_, r0
+	cmp/eq	#0, r0
+load_jtable:
+	mov.l	@r1, r2
+	add	#32, r1
+	bf/s	load_jtable
+	 dt	r0
+
+	bra sh4_really_suspend		! Now we jump on sh4_really_suspend
+	 nop				! to really suspend (and resume... ;-)
+
+.balign 32
+s_jmp_table:
+.long l_end, l_end_no_sleep, l_source, l_load, l_iload_scr0, l_iload_src1, l_iload_src2, l_iload_dest 
+.long l_store, l_or, l_and, l_not, l_while_eq, l_while_neq, l_delay, l_load_src0
+.long l_load_src1, l_load_src2   
+s_jmp_table_address:
+.long s_jmp_table
diff --git a/arch/sh/kernel/cpu/sh4/suspend.c b/arch/sh/kernel/cpu/sh4/suspend.c
new file mode 100644
index 0000000..754de56
--- /dev/null
+++ b/arch/sh/kernel/cpu/sh4/suspend.c
@@ -0,0 +1,113 @@ 
+/*
+ * -------------------------------------------------------------------------
+ * <linux_root>/arch/sh/kernel/suspend.c
+ * -------------------------------------------------------------------------
+ * Copyright (C) 2008  STMicroelectronics
+ * Copyright (C) 2009  STMicroelectronics
+ * Author: Francesco M. Virlinzi  <[e-mail]>
+ *
+ * May be copied or modified under the terms of the GNU General Public
+ * License V.2 ONLY.  See linux/COPYING for more information.
+ *
+ * ------------------------------------------------------------------------- */
+
+#include <linux/init.h>
+#include <linux/suspend.h>
+#include <linux/errno.h>
+#include <linux/time.h>
+#include <linux/delay.h>
+#include <linux/irqflags.h>
+#include <linux/kobject.h>
+#include <linux/stat.h>
+#include <linux/clk.h>
+#include <linux/hardirq.h>
+#include <linux/jiffies.h>
+#include <asm/system.h>
+#include <asm/io.h>
+#include <asm-generic/bug.h>
+#include <cpu-sh4/suspend.h>
+#undef  dbg_print
+
+#ifdef CONFIG_PM_DEBUG
+#define dbg_print(fmt, args...)		\
+		printk(KERN_DEBUG "%s: " fmt, __FUNCTION__ , ## args)
+#else
+#define dbg_print(fmt, args...)
+#endif
+
+
+static struct sh4_suspend_t *sh4_data;
+
+unsigned long sh4_suspend(struct sh4_suspend_t *pdata,
+	unsigned long instr_tbl, unsigned long instr_tbl_end);
+
+static inline unsigned long _10_ms_lpj(void)
+{
+	static struct clk *sh4_clk;
+
+	if (!sh4_clk)
+		sh4_clk = clk_get(NULL, "sh4_clk");
+
+	return clk_get_rate(sh4_clk) / (100 * 2);
+}
+
+static int sh4_suspend_enter(suspend_state_t state)
+{
+	unsigned long flags;
+	unsigned long instr_tbl, instr_tbl_end;
+	unsigned long wokenup_by;
+
+	sh4_data->l_p_j = _10_ms_lpj();
+
+	/* Must wait for serial buffers to clear */
+	mdelay(500);
+
+	local_irq_save(flags);
+
+	/* sets the right instruction table */
+	if (state == PM_SUSPEND_STANDBY) {
+		instr_tbl     = sh4_data->stby_tbl;
+		instr_tbl_end = sh4_data->stby_size;
+	} else {
+		instr_tbl     = sh4_data->mem_tbl;
+		instr_tbl_end = sh4_data->mem_size;
+	}
+
+	BUG_ON(in_irq());
+
+	wokenup_by = sh4_suspend(sh4_data, instr_tbl, instr_tbl_end);
+
+/*
+ *  without the evt_to_irq function the INTEVT is returned
+ */
+	if (sh4_data->evt_to_irq)
+		wokenup_by = sh4_data->evt_to_irq(wokenup_by);
+
+	BUG_ON(in_irq());
+
+	local_irq_restore(flags);
+
+	printk(KERN_INFO "sh4 woken up by: 0x%x\n", wokenup_by);
+
+	return 0;
+}
+
+static int sh4_suspend_valid_both(suspend_state_t state)
+{
+	return 1;
+}
+
+int sh_register_suspend(struct sh4_suspend_t *pdata)
+{
+	sh4_data = pdata;
+	sh4_data->ops.enter = sh4_suspend_enter;
+	if (sh4_data->stby_tbl)
+		sh4_data->ops.valid = sh4_suspend_valid_both;
+	else
+		sh4_data->ops.valid = suspend_valid_only_mem;
+	suspend_set_ops(&sh4_data->ops);
+
+	printk(KERN_INFO "sh4 suspend support registered\n");
+
+	return 0;
+}
-- 
1.5.6.6