diff mbox

[v3,03/13] pm: at91: Workaround DDRSDRC self-refresh bug with LPDDR1 memories.

Message ID 1422338006-3371-1-git-send-email-wenyou.yang@atmel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Wenyou Yang Jan. 27, 2015, 5:53 a.m. UTC
From: Peter Rosin <peda@axentia.se>

The DDRSDR controller fails miserably to put LPDDR1 memories in
self-refresh. Force the controller to think it has DDR2 memories
during the self-refresh period, as the DDR2 self-refresh spec is
equivalent to LPDDR1, and is correctly implemented in the
controller.

Assume that the second controller has the same fault, but that is
untested.

Signed-off-by: Peter Rosin <peda@axentia.se>
Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>
---
 arch/arm/mach-at91/pm_slowclock.S  |   43 +++++++++++++++++++++++++++++++-----
 include/soc/at91/at91sam9_ddrsdr.h |    2 +-
 2 files changed, 39 insertions(+), 6 deletions(-)

Comments

Sergei Shtylyov Jan. 27, 2015, 10:24 a.m. UTC | #1
Hello.

On 1/27/2015 8:53 AM, Wenyou Yang wrote:

> From: Peter Rosin <peda@axentia.se>

> The DDRSDR controller fails miserably to put LPDDR1 memories in
> self-refresh. Force the controller to think it has DDR2 memories
> during the self-refresh period, as the DDR2 self-refresh spec is
> equivalent to LPDDR1, and is correctly implemented in the
> controller.

> Assume that the second controller has the same fault, but that is
> untested.

> Signed-off-by: Peter Rosin <peda@axentia.se>
> Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>
> ---
>   arch/arm/mach-at91/pm_slowclock.S  |   43 +++++++++++++++++++++++++++++++-----
>   include/soc/at91/at91sam9_ddrsdr.h |    2 +-
>   2 files changed, 39 insertions(+), 6 deletions(-)

> diff --git a/arch/arm/mach-at91/pm_slowclock.S b/arch/arm/mach-at91/pm_slowclock.S
> index e2bfaf5..1155217 100644
> --- a/arch/arm/mach-at91/pm_slowclock.S
> +++ b/arch/arm/mach-at91/pm_slowclock.S
[...]
> @@ -108,14 +118,26 @@ ddr_sr_enable:
>
>   	/* figure out if we use the second ram controller */
>   	cmp	ramc1, #0
> -	ldrne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> -	strne	tmp2, .saved_sam9_lpr1
> -	bicne	tmp2, #AT91_DDRSDRC_LPCB
> -	orrne	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
> +	beq	ddr_no_2nd_ctrl
> +
> +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> +	str	tmp2, .saved_sam9_mdr1
> +	bic	tmp2, tmp2, #~AT91_DDRSDRC_MD
> +	cmp	tmp2, #AT91_DDRSDRC_MD_LOW_POWER_DDR
> +	ldreq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> +	biceq	tmp2, tmp2, #AT91_DDRSDRC_MD

    Didn't you forget ~? Either that, or ~ above is not needed, I think.

> +	orreq	tmp2, tmp2, #AT91_DDRSDRC_MD_DDR2
> +	streq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> +
> +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> +	str	tmp2, .saved_sam9_lpr1
> +	bic	tmp2, #AT91_DDRSDRC_LPCB

    Didn't you forget ~? And isn't it 3-operand instruction (as seen in the 
above code)?

> +	orr	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH

    Only 2 operands?

[...]

WBR, Sergei
Peter Rosin Jan. 27, 2015, 12:14 p.m. UTC | #2
Sergei Shtylyov wrote:
> Hello.

Hi!

> On 1/27/2015 8:53 AM, Wenyou Yang wrote:
> 
> > From: Peter Rosin <peda@axentia.se>
> 
> > The DDRSDR controller fails miserably to put LPDDR1 memories in
> > self-refresh. Force the controller to think it has DDR2 memories
> > during the self-refresh period, as the DDR2 self-refresh spec is
> > equivalent to LPDDR1, and is correctly implemented in the controller.
> 
> > Assume that the second controller has the same fault, but that is
> > untested.
> 
> > Signed-off-by: Peter Rosin <peda@axentia.se>
> > Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>
> > ---
> >   arch/arm/mach-at91/pm_slowclock.S  |   43
> +++++++++++++++++++++++++++++++-----
> >   include/soc/at91/at91sam9_ddrsdr.h |    2 +-
> >   2 files changed, 39 insertions(+), 6 deletions(-)
> 
> > diff --git a/arch/arm/mach-at91/pm_slowclock.S
> > b/arch/arm/mach-at91/pm_slowclock.S
> > index e2bfaf5..1155217 100644
> > --- a/arch/arm/mach-at91/pm_slowclock.S
> > +++ b/arch/arm/mach-at91/pm_slowclock.S
> [...]
> > @@ -108,14 +118,26 @@ ddr_sr_enable:
> >
> >   	/* figure out if we use the second ram controller */
> >   	cmp	ramc1, #0
> > -	ldrne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> > -	strne	tmp2, .saved_sam9_lpr1
> > -	bicne	tmp2, #AT91_DDRSDRC_LPCB
> > -	orrne	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
> > +	beq	ddr_no_2nd_ctrl
> > +
> > +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > +	str	tmp2, .saved_sam9_mdr1
> > +	bic	tmp2, tmp2, #~AT91_DDRSDRC_MD
> > +	cmp	tmp2, #AT91_DDRSDRC_MD_LOW_POWER_DDR
> > +	ldreq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > +	biceq	tmp2, tmp2, #AT91_DDRSDRC_MD
> 
>     Didn't you forget ~? Either that, or ~ above is not needed, I think.

The code is correct, the first bic with ~ clears bits not in the relevant
field in order to compare if LPDDR mode is active. The second bic(eq)
w/o ~ clears the field, to make way for the bits in the below orreq
when actually changing the register content into DDR2 mode.

> > +	orreq	tmp2, tmp2, #AT91_DDRSDRC_MD_DDR2
> > +	streq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > +
> > +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> > +	str	tmp2, .saved_sam9_lpr1
> > +	bic	tmp2, #AT91_DDRSDRC_LPCB
> 
>     Didn't you forget ~? And isn't it 3-operand instruction (as seen in the above
> code)?

The logic for the LPR register is from the old code, the "only" thing
I did to it was changing the instruction sequence to not have the
???ne form, i.e. ldrne, strne, bicne, orrne became ldr, str, bic, orr
with a jump around it instead. So, the original code also had a two
argument bic(ne), which indeed is strange, and I don't know why
there is no warning from the assembler. Since there is no warning,
my guess is that the assembler somehow mends it? Or does the
patch actually break the second controller? It would be a surprise
it the assembler handles 2-operand bicne differently from a
2-operand bic, no?

> > +	orr	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
> 
>     Only 2 operands?

Same argument as above. I didn't touch it (sort of...)

Should I update the patch and fix this collateral 2-operand problem as
well? To me, it feels like a separate patch, no?

> [...]
> 
> WBR, Sergei

Cheers,
Peter
Peter Rosin Jan. 27, 2015, 9:55 p.m. UTC | #3
I wrote:
> Sergei Shtylyov wrote:
> > On 1/27/2015 8:53 AM, Wenyou Yang wrote:
> >
> > > From: Peter Rosin <peda@axentia.se>
> >
> > > The DDRSDR controller fails miserably to put LPDDR1 memories in
> > > self-refresh. Force the controller to think it has DDR2 memories
> > > during the self-refresh period, as the DDR2 self-refresh spec is
> > > equivalent to LPDDR1, and is correctly implemented in the controller.
> >
> > > Assume that the second controller has the same fault, but that is
> > > untested.
> >
> > > Signed-off-by: Peter Rosin <peda@axentia.se>
> > > Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>
> > > ---
> > >   arch/arm/mach-at91/pm_slowclock.S  |   43
> > +++++++++++++++++++++++++++++++-----
> > >   include/soc/at91/at91sam9_ddrsdr.h |    2 +-
> > >   2 files changed, 39 insertions(+), 6 deletions(-)
> >
> > > diff --git a/arch/arm/mach-at91/pm_slowclock.S
> > > b/arch/arm/mach-at91/pm_slowclock.S
> > > index e2bfaf5..1155217 100644
> > > --- a/arch/arm/mach-at91/pm_slowclock.S
> > > +++ b/arch/arm/mach-at91/pm_slowclock.S
> > [...]
> > > @@ -108,14 +118,26 @@ ddr_sr_enable:
> > >
> > >   	/* figure out if we use the second ram controller */
> > >   	cmp	ramc1, #0
> > > -	ldrne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> > > -	strne	tmp2, .saved_sam9_lpr1
> > > -	bicne	tmp2, #AT91_DDRSDRC_LPCB
> > > -	orrne	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
> > > +	beq	ddr_no_2nd_ctrl
> > > +
> > > +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > > +	str	tmp2, .saved_sam9_mdr1
> > > +	bic	tmp2, tmp2, #~AT91_DDRSDRC_MD
> > > +	cmp	tmp2, #AT91_DDRSDRC_MD_LOW_POWER_DDR
> > > +	ldreq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > > +	biceq	tmp2, tmp2, #AT91_DDRSDRC_MD
> >
> >     Didn't you forget ~? Either that, or ~ above is not needed, I think.
> 
> The code is correct, the first bic with ~ clears bits not in the relevant
> field in order to compare if LPDDR mode is active. The second bic(eq)
> w/o ~ clears the field, to make way for the bits in the below orreq
> when actually changing the register content into DDR2 mode.
> 
> > > +	orreq	tmp2, tmp2, #AT91_DDRSDRC_MD_DDR2
> > > +	streq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
> > > +
> > > +	ldr	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
> > > +	str	tmp2, .saved_sam9_lpr1
> > > +	bic	tmp2, #AT91_DDRSDRC_LPCB
> >
> >     Didn't you forget ~? And isn't it 3-operand instruction (as seen in the above
> > code)?
> 
> The logic for the LPR register is from the old code, the "only" thing
> I did to it was changing the instruction sequence to not have the
> ???ne form, i.e. ldrne, strne, bicne, orrne became ldr, str, bic, orr
> with a jump around it instead. So, the original code also had a two
> argument bic(ne), which indeed is strange, and I don't know why
> there is no warning from the assembler. Since there is no warning,
> my guess is that the assembler somehow mends it? Or does the
> patch actually break the second controller? It would be a surprise
> it the assembler handles 2-operand bicne differently from a

s/it the/if the/

> 2-operand bic, no?
> 
> > > +	orr	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
> >
> >     Only 2 operands?
> 
> Same argument as above. I didn't touch it (sort of...)
> 
> Should I update the patch and fix this collateral 2-operand problem as
> well? To me, it feels like a separate patch, no?

I have now checked the assembler output, and apparently it mends the
input, just as I thought. That might be a fluke, of course, or it might be a
deliberate shorthand when the destination register is the same as the
following operand? But I also note that there are more instances of this
2 vs. 3 argument syntax, and I suggest that they are all fixed in one go,
if it is determined that they need fixing. I am obviously not an authority
when it comes to arm assembler syntax, so someone else will have to
advise...

Cheers,
Peter
Russell King - ARM Linux Jan. 27, 2015, 10:43 p.m. UTC | #4
On Tue, Jan 27, 2015 at 09:55:52PM +0000, Peter Rosin wrote:
> I have now checked the assembler output, and apparently it mends the
> input, just as I thought. That might be a fluke, of course, or it might be a
> deliberate shorthand when the destination register is the same as the
> following operand?

I believe it is because we have the assembler in unified asm mode, where
such things are legal (due to Thumb2.)  That means these two are
legal:

	orr	r2, #const
	orr	r2, r2, #const

and equivalent in this context.
diff mbox

Patch

diff --git a/arch/arm/mach-at91/pm_slowclock.S b/arch/arm/mach-at91/pm_slowclock.S
index e2bfaf5..1155217 100644
--- a/arch/arm/mach-at91/pm_slowclock.S
+++ b/arch/arm/mach-at91/pm_slowclock.S
@@ -100,6 +100,16 @@  ddr_sr_enable:
 	cmp	memctrl, #AT91_MEMCTRL_DDRSDR
 	bne	sdr_sr_enable
 
+	/* LPDDR1 --> force DDR2 mode during self-refresh */
+	ldr	tmp1, [sdramc, #AT91_DDRSDRC_MDR]
+	str	tmp1, .saved_sam9_mdr
+	bic	tmp1, tmp1, #~AT91_DDRSDRC_MD
+	cmp	tmp1, #AT91_DDRSDRC_MD_LOW_POWER_DDR
+	ldreq	tmp1, [sdramc, #AT91_DDRSDRC_MDR]
+	biceq	tmp1, tmp1, #AT91_DDRSDRC_MD
+	orreq	tmp1, tmp1, #AT91_DDRSDRC_MD_DDR2
+	streq	tmp1, [sdramc, #AT91_DDRSDRC_MDR]
+
 	/* prepare for DDRAM self-refresh mode */
 	ldr	tmp1, [sdramc, #AT91_DDRSDRC_LPR]
 	str	tmp1, .saved_sam9_lpr
@@ -108,14 +118,26 @@  ddr_sr_enable:
 
 	/* figure out if we use the second ram controller */
 	cmp	ramc1, #0
-	ldrne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
-	strne	tmp2, .saved_sam9_lpr1
-	bicne	tmp2, #AT91_DDRSDRC_LPCB
-	orrne	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
+	beq	ddr_no_2nd_ctrl
+
+	ldr	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
+	str	tmp2, .saved_sam9_mdr1
+	bic	tmp2, tmp2, #~AT91_DDRSDRC_MD
+	cmp	tmp2, #AT91_DDRSDRC_MD_LOW_POWER_DDR
+	ldreq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
+	biceq	tmp2, tmp2, #AT91_DDRSDRC_MD
+	orreq	tmp2, tmp2, #AT91_DDRSDRC_MD_DDR2
+	streq	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
+
+	ldr	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
+	str	tmp2, .saved_sam9_lpr1
+	bic	tmp2, #AT91_DDRSDRC_LPCB
+	orr	tmp2, #AT91_DDRSDRC_LPCB_SELF_REFRESH
 
 	/* Enable DDRAM self-refresh mode */
+	str	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
+ddr_no_2nd_ctrl:
 	str	tmp1, [sdramc, #AT91_DDRSDRC_LPR]
-	strne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
 
 	b	sdr_sr_done
 
@@ -227,12 +249,17 @@  sdr_sr_done:
 	 */
 	cmp	memctrl, #AT91_MEMCTRL_DDRSDR
 	bne	sdr_en_restore
+	/* Restore MDR in case of LPDDR1 */
+	ldr	tmp1, .saved_sam9_mdr
+	str	tmp1, [sdramc, #AT91_DDRSDRC_MDR]
 	/* Restore LPR on AT91 with DDRAM */
 	ldr	tmp1, .saved_sam9_lpr
 	str	tmp1, [sdramc, #AT91_DDRSDRC_LPR]
 
 	/* if we use the second ram controller */
 	cmp	ramc1, #0
+	ldrne	tmp2, .saved_sam9_mdr1
+	strne	tmp2, [ramc1, #AT91_DDRSDRC_MDR]
 	ldrne	tmp2, .saved_sam9_lpr1
 	strne	tmp2, [ramc1, #AT91_DDRSDRC_LPR]
 
@@ -263,5 +290,11 @@  ram_restored:
 .saved_sam9_lpr1:
 	.word 0
 
+.saved_sam9_mdr:
+	.word 0
+
+.saved_sam9_mdr1:
+	.word 0
+
 ENTRY(at91_slow_clock_sz)
 	.word .-at91_slow_clock
diff --git a/include/soc/at91/at91sam9_ddrsdr.h b/include/soc/at91/at91sam9_ddrsdr.h
index 0210797..dc10c52 100644
--- a/include/soc/at91/at91sam9_ddrsdr.h
+++ b/include/soc/at91/at91sam9_ddrsdr.h
@@ -92,7 +92,7 @@ 
 #define		AT91_DDRSDRC_UPD_MR	(3 << 20)	 /* Update load mode register and extended mode register */
 
 #define AT91_DDRSDRC_MDR	0x20	/* Memory Device Register */
-#define		AT91_DDRSDRC_MD		(3 << 0)		/* Memory Device Type */
+#define		AT91_DDRSDRC_MD		(7 << 0)	/* Memory Device Type */
 #define			AT91_DDRSDRC_MD_SDR		0
 #define			AT91_DDRSDRC_MD_LOW_POWER_SDR	1
 #define			AT91_DDRSDRC_MD_LOW_POWER_DDR	3