[PATCH] amd64_edac: fix size decoding error on K8

[PATCH] amd64_edac: fix size decoding error on K8

[Niklas Söderlund]

Use a lookup table to calculate the size of a chip select. The old
method of calculating size return erroneous values for some memory
configurations. The lookup table is transcribed from "DRAM CS
Address Mapping Register" [1], pages 88-89.

[1] BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
    Opteron Processors - Revision 3.30 - doc #26094 - February 2006
    http://support.amd.com/us/Processor_TechDocs/26094.pdf

Signed-off-by: Niklas Söderlund <niklas.soderlund@ericsson.com>
---
 drivers/edac/amd64_edac.c |   10 +++++-----
 1 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
index 9a8bebc..396ea76 100644
--- a/drivers/edac/amd64_edac.c
+++ b/drivers/edac/amd64_edac.c
@@ -1108,6 +1108,10 @@ static int ddr2_cs_size(unsigned i, bool dct_width)
 static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
 				  unsigned cs_mode)
 {
+	static const int rev_de_lookup[] = { 32, 64, 128, 128, 256, 512, 256,
+		512, 1024, 1024, 2048 };
+
+
 	u32 dclr = dct ? pvt->dclr1 : pvt->dclr0;
 
 	if (pvt->ext_model >= K8_REV_F) {
@@ -1116,11 +1120,7 @@ static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
 	}
 	else if (pvt->ext_model >= K8_REV_D) {
 		WARN_ON(cs_mode > 10);
-
-		if (cs_mode == 3 || cs_mode == 8)
-			return 32 << (cs_mode - 1);
-		else
-			return 32 << cs_mode;
+		return rev_de_lookup[cs_mode];
 	}
 	else {
 		WARN_ON(cs_mode > 6);
-- 
1.7.7.1

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Borislav Petkov]

On Wed, Nov 09, 2011 at 03:32:58PM +0100, Niklas Söderlund wrote:
> Use a lookup table to calculate the size of a chip select. The old
> method of calculating size return erroneous values for some memory
> configurations.

Can you elaborate more on those erroneous values, please?

Thanks.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Niklas Söderlund]

On 11/09/2011 03:42 PM, Borislav Petkov wrote:
> On Wed, Nov 09, 2011 at 03:32:58PM +0100, Niklas Söderlund wrote:
>> Use a lookup table to calculate the size of a chip select. The old
>> method of calculating size return erroneous values for some memory
>> configurations.
>
> Can you elaborate more on those erroneous values, please?
>
> Thanks.
>

The decoding table from the documentation:

cs_mode		size (mb)
0		32
1		64
2		128
3		128
4		256
5		512
6		256
7		512
8		1024
9		1024
10		2048

The original code:

if (cs_mode == 3 || cs_mode == 8)
	return 32 << (cs_mode - 1);
else
	return 32 << cs_mode;

The code tries to compensate for the irregularity in the decoding table 
but fails for cs_mode greater or equal to 4. It fails because:

- It do not consider the cumulating effect of reoccurring sizes for all 
values of cs_mode. It only tries for the special cases and it only 
succeeds in one.

- It do not consider the out of order values of cs_mode values 4-7: 256, 
512, 256, 512.

Best Regards
// Niklas

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Borislav Petkov]

Hi Niklas,

On Wed, Nov 09, 2011 at 04:13:37PM +0100, Niklas Söderlund wrote:
> The decoding table from the documentation:
> 
> cs_mode		size (mb)
> 0		32
> 1		64
> 2		128
> 3		128
> 4		256
> 5		512
> 6		256
> 7		512
> 8		1024
> 9		1024
> 10		2048
> 
> The original code:
> 
> if (cs_mode == 3 || cs_mode == 8)
> 	return 32 << (cs_mode - 1);
> else
> 	return 32 << cs_mode;
> 
> The code tries to compensate for the irregularity in the decoding
> table but fails for cs_mode greater or equal to 4. It fails because:
> 
> - It do not consider the cumulating effect of reoccurring sizes for
> all values of cs_mode. It only tries for the special cases and it
> only succeeds in one.
> 
> - It do not consider the out of order values of cs_mode values 4-7:
> 256, 512, 256, 512.

Agreed, thanks for reporting this. However, I removed the static tables
at the time because I didn't want to carry unnecessary weight around for
no reason (tables were static back then). And I know, they're small but
dead weight like that tends to accumulate.

Your approach is ok with creating the lookup table on the stack but how
about the following instead - it is much smaller and it is only code,
no need for stack memory at all. It should give you the proper mapping,
IINM.

--
diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
index c9eee6d33e9a..3702f20d784e 100644
--- a/drivers/edac/amd64_edac.c
+++ b/drivers/edac/amd64_edac.c
@@ -1132,12 +1132,12 @@ static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
 		return ddr2_cs_size(cs_mode, dclr & WIDTH_128);
 	}
 	else if (pvt->ext_model >= K8_REV_D) {
+		unsigned diff;
 		WARN_ON(cs_mode > 10);
 
-		if (cs_mode == 3 || cs_mode == 8)
-			return 32 << (cs_mode - 1);
-		else
-			return 32 << cs_mode;
+		diff = cs_mode / 3 + (unsigned)(cs_mode > 5);
+
+		return 32 << (cs_mode - diff);
 	}
 	else {
 		WARN_ON(cs_mode > 6);
--

Thanks.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Tony Luck]

On Wed, Nov 9, 2011 at 12:35 PM, Borislav Petkov <bp@amd64.org> wrote:
> +               diff = cs_mode / 3 + (unsigned)(cs_mode > 5);
> +
> +               return 32 << (cs_mode - diff);

Code is correct - but looks like an entry for an obfuscated C
competition. Perhaps
it deserves a comment!

-Tony

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Nils Carlson]

Hi Borislav,

On Nov 9, 2011, at 9:35 PM, Borislav Petkov wrote:

> Hi Niklas,
> 
> On Wed, Nov 09, 2011 at 04:13:37PM +0100, Niklas Söderlund wrote:
>> The decoding table from the documentation:
>> 

>> <snip>

>> - It do not consider the out of order values of cs_mode values 4-7:
>> 256, 512, 256, 512.
> 
> Agreed, thanks for reporting this. However, I removed the static tables
> at the time because I didn't want to carry unnecessary weight around for
> no reason (tables were static back then). And I know, they're small but
> dead weight like that tends to accumulate.
> 
> Your approach is ok with creating the lookup table on the stack but how
> about the following instead - it is much smaller and it is only code,
> no need for stack memory at all. It should give you the proper mapping,
> IINM.
> 

I have to say that I much prefer Niklas approach, the code below is difficult
to understand compared to the table, and seeing as the data sheet contains
the table it's much easier for a reader to understand the correspondance.

Cheers,
Nils Carlson

> --
> diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
> index c9eee6d33e9a..3702f20d784e 100644
> --- a/drivers/edac/amd64_edac.c
> +++ b/drivers/edac/amd64_edac.c
> @@ -1132,12 +1132,12 @@ static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
> 		return ddr2_cs_size(cs_mode, dclr & WIDTH_128);
> 	}
> 	else if (pvt->ext_model >= K8_REV_D) {
> +		unsigned diff;
> 		WARN_ON(cs_mode > 10);
> 
> -		if (cs_mode == 3 || cs_mode == 8)
> -			return 32 << (cs_mode - 1);
> -		else
> -			return 32 << cs_mode;
> +		diff = cs_mode / 3 + (unsigned)(cs_mode > 5);
> +
> +		return 32 << (cs_mode - diff);
> 	}
> 	else {
> 		WARN_ON(cs_mode > 6);
> --
> 
> Thanks.
> 
> -- 
> Regards/Gruss,
> Boris.
> 
> Advanced Micro Devices GmbH
> Einsteinring 24, 85609 Dornach
> GM: Alberto Bozzo
> Reg: Dornach, Landkreis Muenchen
> HRB Nr. 43632 WEEE Registernr: 129 19551
> --
> To unsubscribe from this list: send the line "unsubscribe linux-edac" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH] amd64_edac: fix size decoding error on K8

[Borislav Petkov]

On Wed, Nov 09, 2011 at 12:50:42PM -0800, Tony Luck wrote:
> On Wed, Nov 9, 2011 at 12:35 PM, Borislav Petkov <bp@amd64.org> wrote:
> > +               diff = cs_mode / 3 + (unsigned)(cs_mode > 5);
> > +
> > +               return 32 << (cs_mode - diff);
> 
> Code is correct - but looks like an entry for an obfuscated C
> competition.

I know, right?! Especially casting the boolean to an unsigned int. Yuck!
But it is compact the stupid thing.

> Perhaps it deserves a comment!

Sure it does, will do.

Thanks.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551

[PATCH] amd64_edac: Fix K8 revD and later chip select sizes

[Borislav Petkov]

Fix DRAM chip select sizes calculation for K8, revisions D and E.

Reported-by: Niklas Söderlund <niklas.soderlund@ericsson.com
Link: http://lkml.kernel.org/r/1320849178-23340-1-git-send-email-niklas.soderlund@ericsson.com
Signed-off-by: Borislav Petkov <borislav.petkov@amd.com>
---

@Niklas: would you please verify this patch fixes your issue?

Thanks.

 drivers/edac/amd64_edac.c |   32 ++++++++++++++++++++++++++++----
 1 files changed, 28 insertions(+), 4 deletions(-)

diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
index c9eee6d..6a83d49 100644
--- a/drivers/edac/amd64_edac.c
+++ b/drivers/edac/amd64_edac.c
@@ -1132,12 +1132,36 @@ static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
 		return ddr2_cs_size(cs_mode, dclr & WIDTH_128);
 	}
 	else if (pvt->ext_model >= K8_REV_D) {
+		unsigned diff;
 		WARN_ON(cs_mode > 10);
 
-		if (cs_mode == 3 || cs_mode == 8)
-			return 32 << (cs_mode - 1);
-		else
-			return 32 << cs_mode;
+		/*
+		 * the below calculation, besides trying to win an obfuscated C
+		 * contest, maps cs_mode values to DIMM chip select sizes. The
+		 * mappings are:
+		 *
+		 * cs_mode	CS size (mb)
+		 * =======	============
+		 * 0		32
+		 * 1		64
+		 * 2		128
+		 * 3		128
+		 * 4		256
+		 * 5		512
+		 * 6		256
+		 * 7		512
+		 * 8		1024
+		 * 9		1024
+		 * 10		2048
+		 *
+		 * Basically, it calculates a value with which to shift the
+		 * smallest CS size of 32MB.
+		 *
+		 * ddr[23]_cs_size have a similar purpose.
+		 */
+		diff = cs_mode/3 + (unsigned)(cs_mode > 5);
+
+		return 32 << (cs_mode - diff);
 	}
 	else {
 		WARN_ON(cs_mode > 6);
-- 
1.7.8.rc0

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551

Re: [PATCH] amd64_edac: Fix K8 revD and later chip select sizes

[Niklas Söderlund]

On Mon 14 Nov 2011 05:54:59 PM CET, Borislav Petkov wrote:
>
> Fix DRAM chip select sizes calculation for K8, revisions D and E.
>
> Reported-by: Niklas Söderlund<niklas.soderlund@ericsson.com
> Link: http://lkml.kernel.org/r/1320849178-23340-1-git-send-email-niklas.soderlund@ericsson.com
> Signed-off-by: Borislav Petkov<borislav.petkov@amd.com>
> ---
>
> @Niklas: would you please verify this patch fixes your issue?
>
> Thanks.
>
>   drivers/edac/amd64_edac.c |   32 ++++++++++++++++++++++++++++----
>   1 files changed, 28 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
> index c9eee6d..6a83d49 100644
> --- a/drivers/edac/amd64_edac.c
> +++ b/drivers/edac/amd64_edac.c
> @@ -1132,12 +1132,36 @@ static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, u8 dct,
>   		return ddr2_cs_size(cs_mode, dclr&  WIDTH_128);
>   	}
>   	else if (pvt->ext_model>= K8_REV_D) {
> +		unsigned diff;
>   		WARN_ON(cs_mode>  10);
>
> -		if (cs_mode == 3 || cs_mode == 8)
> -			return 32<<  (cs_mode - 1);
> -		else
> -			return 32<<  cs_mode;
> +		/*
> +		 * the below calculation, besides trying to win an obfuscated C
> +		 * contest, maps cs_mode values to DIMM chip select sizes. The
> +		 * mappings are:
> +		 *
> +		 * cs_mode	CS size (mb)
> +		 * =======	============
> +		 * 0		32
> +		 * 1		64
> +		 * 2		128
> +		 * 3		128
> +		 * 4		256
> +		 * 5		512
> +		 * 6		256
> +		 * 7		512
> +		 * 8		1024
> +		 * 9		1024
> +		 * 10		2048
> +		 *
> +		 * Basically, it calculates a value with which to shift the
> +		 * smallest CS size of 32MB.
> +		 *
> +		 * ddr[23]_cs_size have a similar purpose.
> +		 */
> +		diff = cs_mode/3 + (unsigned)(cs_mode>  5);
> +
> +		return 32<<  (cs_mode - diff);
>   	}
>   	else {
>   		WARN_ON(cs_mode>  6);

Hi Borislav,

Yes the patch fixes my problem, but it truly is obfuscated C.

Thanks
// Niklas