[U-Boot] [PATCH 1/3] armv7/fsl-ls102xa: Workaround for DDR erratum A008514
Affects: DDR Description: Memory controller performance is not optimal with default internal target queue register values. Impact: Memory controller performance is not optimal. Workaround: Write a value of 63b2_0002h to address: 157_020Ch.
Signed-off-by: Yuan Yao yao.yuan@freescale.com --- arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h | 2 +- board/freescale/ls1021aqds/ls1021aqds.c | 10 ++++++++++ board/freescale/ls1021atwr/ls1021atwr.c | 10 ++++++++++ 3 files changed, 21 insertions(+), 1 deletion(-)
diff --git a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h index d34044a..21bd65b 100644 --- a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h +++ b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h @@ -221,7 +221,7 @@ struct ccsr_scfg { u32 scfgrevcr; u32 coresrencr; u32 pex2pmrdsr; - u32 ddrc1cr; + u32 eddrtqcfg; u32 ddrc2cr; u32 ddrc3cr; u32 ddrc4cr; diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index d6ef6ba..52bffc8 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -299,6 +299,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/* + * Memory controller require a register write before being enabled. + * Affects: DDR + * Register: EDDRTQCFG + * Description: Memory controller performance is not optimal with + * default internal target queue register values. + * Workaround: Write a value of 63b2_0002h to address: 157_020Ch. + */ + out_be32(&scfg->eddrtqcfg, 0x63b20002); + + /* * Enable snoop requests and DVM message requests for * Slave insterface S4 (A7 core cluster) */ diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index b7458a9..c565e91 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -501,6 +501,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/* + * Memory controller require a register write before being enabled. + * Affects: DDR + * Register: EDDRTQCFG + * Description: Memory controller performance is not optimal with + * default internal target queue register values. + * Workaround: Write a value of 63b2_0002h to address: 157_020Ch. + */ + out_be32(&scfg->eddrtqcfg, 0x63b20002); + + /* * Enable snoop requests and DVM message requests for * Slave insterface S4 (A7 core cluster) */
EDDRTQCFG Registers are Integration Strap values which controls performance parameters for DDR Controller.
The bit 25 is used to disable priorities within DDR since DDR are connected backwards on silicon Rev2.0.
Signed-off-by: Yuan Yao yao.yuan@freescale.com --- board/freescale/ls1021aqds/ls1021aqds.c | 13 ++++++++++++- board/freescale/ls1021atwr/ls1021atwr.c | 13 ++++++++++++- 2 files changed, 24 insertions(+), 2 deletions(-)
diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index 52bffc8..1256ece 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -277,7 +277,7 @@ int board_early_init_f(void) { struct ccsr_scfg *scfg = (struct ccsr_scfg *)CONFIG_SYS_FSL_SCFG_ADDR; struct ccsr_cci400 *cci = (struct ccsr_cci400 *)CONFIG_SYS_CCI400_ADDR; - unsigned int major; + unsigned int major, reg;
#ifdef CONFIG_TSEC_ENET out_be32(&scfg->etsecdmamcr, SCFG_ETSECDMAMCR_LE_BD_FR); @@ -309,6 +309,17 @@ int board_early_init_f(void) out_be32(&scfg->eddrtqcfg, 0x63b20002);
/* + * EDDRTQCFG Registers are Integration Strap values which controls + * performance parameters for DDR Controller. + * The bit 25(0x40) is used for disable priorities within DDR. + * This is a workaround because of the DDR are connected backwards + * on Rev2.0. + */ + reg = in_be32(&scfg->eddrtqcfg); + reg |= 0x40; + out_be32(&scfg->eddrtqcfg, reg); + + /* * Enable snoop requests and DVM message requests for * Slave insterface S4 (A7 core cluster) */ diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index c565e91..53bea95 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -478,7 +478,7 @@ int board_early_init_f(void) { struct ccsr_scfg *scfg = (struct ccsr_scfg *)CONFIG_SYS_FSL_SCFG_ADDR; struct ccsr_cci400 *cci = (struct ccsr_cci400 *)CONFIG_SYS_CCI400_ADDR; - unsigned int major; + unsigned int major, reg;
#ifdef CONFIG_TSEC_ENET out_be32(&scfg->etsecdmamcr, SCFG_ETSECDMAMCR_LE_BD_FR); @@ -511,6 +511,17 @@ int board_early_init_f(void) out_be32(&scfg->eddrtqcfg, 0x63b20002);
/* + * EDDRTQCFG Registers are Integration Strap values which controls + * performance parameters for DDR Controller. + * The bit 25(0x40) is used for disable priorities within DDR. + * This is a workaround because of the DDR are connected backwards + * on Rev2.0. + */ + reg = in_be32(&scfg->eddrtqcfg); + reg |= 0x40; + out_be32(&scfg->eddrtqcfg, reg); + + /* * Enable snoop requests and DVM message requests for * Slave insterface S4 (A7 core cluster) */
Hi Yuan
On 14/08/15 02:54 AM, Yuan Yao wrote:
EDDRTQCFG Registers are Integration Strap values which controls performance parameters for DDR Controller.
The bit 25 is used to disable priorities within DDR since DDR are connected backwards on silicon Rev2.0.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
board/freescale/ls1021aqds/ls1021aqds.c | 13 ++++++++++++- board/freescale/ls1021atwr/ls1021atwr.c | 13 ++++++++++++- 2 files changed, 24 insertions(+), 2 deletions(-)
diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index 52bffc8..1256ece 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -277,7 +277,7 @@ int board_early_init_f(void) { struct ccsr_scfg *scfg = (struct ccsr_scfg *)CONFIG_SYS_FSL_SCFG_ADDR; struct ccsr_cci400 *cci = (struct ccsr_cci400 *)CONFIG_SYS_CCI400_ADDR;
- unsigned int major;
unsigned int major, reg;
#ifdef CONFIG_TSEC_ENET out_be32(&scfg->etsecdmamcr, SCFG_ETSECDMAMCR_LE_BD_FR);
@@ -309,6 +309,17 @@ int board_early_init_f(void) out_be32(&scfg->eddrtqcfg, 0x63b20002);
/*
* EDDRTQCFG Registers are Integration Strap values which controls* performance parameters for DDR Controller.
Would this comment not be better placed in your patch 1/2. Please see my comments there.
* The bit 25(0x40) is used for disable priorities within DDR.* This is a workaround because of the DDR are connected backwards* on Rev2.0.*/- reg = in_be32(&scfg->eddrtqcfg);
- reg |= 0x40;
- out_be32(&scfg->eddrtqcfg, reg);
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index c565e91..53bea95 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -478,7 +478,7 @@ int board_early_init_f(void) { struct ccsr_scfg *scfg = (struct ccsr_scfg *)CONFIG_SYS_FSL_SCFG_ADDR; struct ccsr_cci400 *cci = (struct ccsr_cci400 *)CONFIG_SYS_CCI400_ADDR;
- unsigned int major;
unsigned int major, reg;
#ifdef CONFIG_TSEC_ENET out_be32(&scfg->etsecdmamcr, SCFG_ETSECDMAMCR_LE_BD_FR);
@@ -511,6 +511,17 @@ int board_early_init_f(void) out_be32(&scfg->eddrtqcfg, 0x63b20002);
/*
* EDDRTQCFG Registers are Integration Strap values which controls* performance parameters for DDR Controller.* The bit 25(0x40) is used for disable priorities within DDR.* This is a workaround because of the DDR are connected backwards* on Rev2.0.*/- reg = in_be32(&scfg->eddrtqcfg);
- reg |= 0x40;
- out_be32(&scfg->eddrtqcfg, reg);
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
Regards Sinan Akman
On 08/13/2015 11:54 PM, Yuan Yao wrote:
EDDRTQCFG Registers are Integration Strap values which controls performance parameters for DDR Controller.
The bit 25 is used to disable priorities within DDR since DDR are connected backwards on silicon Rev2.0.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
board/freescale/ls1021aqds/ls1021aqds.c | 13 ++++++++++++- board/freescale/ls1021atwr/ls1021atwr.c | 13 ++++++++++++- 2 files changed, 24 insertions(+), 2 deletions(-)
Yuan,
SoC erratum workaround shouldn't be put into board file.
York
For LS1021A Rev2.0 have already fixed the snoop silicon issue, So enable snoop requests and DVM message requests for all the slave insterfaces.
Signed-off-by: Yuan Yao yao.yuan@freescale.com --- board/freescale/ls1021aqds/ls1021aqds.c | 8 +++++++- board/freescale/ls1021atwr/ls1021atwr.c | 8 +++++++- 2 files changed, 14 insertions(+), 2 deletions(-)
diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index 1256ece..2f604f5 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -321,8 +321,14 @@ int board_early_init_f(void)
/* * Enable snoop requests and DVM message requests for - * Slave insterface S4 (A7 core cluster) + * All the slave insterfaces. */ + out_le32(&cci->slave[0].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); + out_le32(&cci->slave[1].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); + out_le32(&cci->slave[2].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); out_le32(&cci->slave[4].snoop_ctrl, CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN);
diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index 53bea95..ac6de7a 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -523,8 +523,14 @@ int board_early_init_f(void)
/* * Enable snoop requests and DVM message requests for - * Slave insterface S4 (A7 core cluster) + * All the slave insterfaces. */ + out_le32(&cci->slave[0].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); + out_le32(&cci->slave[1].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); + out_le32(&cci->slave[2].snoop_ctrl, + CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN); out_le32(&cci->slave[4].snoop_ctrl, CCI400_DVM_MESSAGE_REQ_EN | CCI400_SNOOP_REQ_EN);
On 08/13/2015 11:54 PM, Yuan Yao wrote:
For LS1021A Rev2.0 have already fixed the snoop silicon issue, So enable snoop requests and DVM message requests for all the slave insterfaces.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
board/freescale/ls1021aqds/ls1021aqds.c | 8 +++++++- board/freescale/ls1021atwr/ls1021atwr.c | 8 +++++++- 2 files changed, 14 insertions(+), 2 deletions(-)
Yuan,
Do not put SoC specific code into board file, please.
York
Hi Yuan
On 14/08/15 02:54 AM, Yuan Yao wrote:
Affects: DDR Description: Memory controller performance is not optimal with default internal target queue register values. Impact: Memory controller performance is not optimal. Workaround: Write a value of 63b2_0002h to address: 157_020Ch.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h | 2 +- board/freescale/ls1021aqds/ls1021aqds.c | 10 ++++++++++ board/freescale/ls1021atwr/ls1021atwr.c | 10 ++++++++++ 3 files changed, 21 insertions(+), 1 deletion(-)
diff --git a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h index d34044a..21bd65b 100644 --- a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h +++ b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h @@ -221,7 +221,7 @@ struct ccsr_scfg { u32 scfgrevcr; u32 coresrencr; u32 pex2pmrdsr;
- u32 ddrc1cr;
- u32 eddrtqcfg;
Is this a simple name change on the register, that ddrc1cr is not needed/used anywhere else ?
u32 ddrc2cr; u32 ddrc3cr; u32 ddrc4cr; diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index d6ef6ba..52bffc8 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -299,6 +299,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/*
* Memory controller require a register write before being enabled.* Affects: DDR* Register: EDDRTQCFG* Description: Memory controller performance is not optimal with* default internal target queue register values.* Workaround: Write a value of 63b2_0002h to address: 157_020Ch.*/- out_be32(&scfg->eddrtqcfg, 0x63b20002);
Can you perhaps bring the comment from your 2/3 patch that explains EDDRTQCFG Registers are so that the reader sees it before its first usage.
Also, is it possible to explain here what that magic value 0x63b20002 corresponds to and perhaps use a #define that explains its function ?
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index b7458a9..c565e91 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -501,6 +501,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/*
* Memory controller require a register write before being enabled.* Affects: DDR* Register: EDDRTQCFG* Description: Memory controller performance is not optimal with* default internal target queue register values.* Workaround: Write a value of 63b2_0002h to address: 157_020Ch.*/- out_be32(&scfg->eddrtqcfg, 0x63b20002);
Same here.
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
Regards Sinan Akman
Hi Sinan,
Thanks for your review. Please see my comments.
Best Regards, Yuan Yao
-----Original Message----- From: Sinan Akman [mailto:sinan@writeme.com] Sent: Saturday, August 15, 2015 12:28 AM To: Yuan Yao-B46683; Sun York-R58495; Wang Huan-B18965 Cc: u-boot@lists.denx.de Subject: Re: [U-Boot] [PATCH 1/3] armv7/fsl-ls102xa: Workaround for DDR erratum A008514
Hi Yuan
On 14/08/15 02:54 AM, Yuan Yao wrote:
Affects: DDR Description: Memory controller performance is not optimal with default internal target queue register values. Impact: Memory controller performance is not optimal. Workaround: Write a value of 63b2_0002h to address: 157_020Ch.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h | 2 +- board/freescale/ls1021aqds/ls1021aqds.c | 10 ++++++++++ board/freescale/ls1021atwr/ls1021atwr.c | 10 ++++++++++ 3 files changed, 21 insertions(+), 1 deletion(-)
diff --git a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h index d34044a..21bd65b 100644 --- a/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h +++ b/arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h @@ -221,7 +221,7 @@ struct ccsr_scfg { u32 scfgrevcr; u32 coresrencr; u32 pex2pmrdsr;
- u32 ddrc1cr;
- u32 eddrtqcfg;
Is this a simple name change on the register, that ddrc1cr is not needed/used anywhere else ?
[Yuan Yao] Yes, there is no ddrc1cr in ls102xa, the register on 0x157020c is named eddrtqcfg.
u32 ddrc2cr; u32 ddrc3cr; u32 ddrc4cr; diff --git a/board/freescale/ls1021aqds/ls1021aqds.c b/board/freescale/ls1021aqds/ls1021aqds.c index d6ef6ba..52bffc8 100644 --- a/board/freescale/ls1021aqds/ls1021aqds.c +++ b/board/freescale/ls1021aqds/ls1021aqds.c @@ -299,6 +299,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/*
* Memory controller require a register write before being enabled.* Affects: DDR* Register: EDDRTQCFG* Description: Memory controller performance is not optimal with* default internal target queue register values.* Workaround: Write a value of 63b2_0002h to address: 157_020Ch.*/- out_be32(&scfg->eddrtqcfg, 0x63b20002);
Can you perhaps bring the comment from your 2/3 patch that explains EDDRTQCFG Registers are so that the reader sees it before its first usage.
Also, is it possible to explain here what that magic value 0x63b20002 corresponds to and perhaps use a #define that explains its function ?
[Yuan Yao] Sorry for here, I also don't sure what's the means of the magic value about 0x63b20002. Also I'm not suggest to use the #define. Because the 0x63b20002 is just a silicon erratum workaround. I just get the information for the document for silicon erratum. There is no more information for software. The register eddrtqcfg is a reserve register for software. It should not open for developer as silicon team said.
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
diff --git a/board/freescale/ls1021atwr/ls1021atwr.c b/board/freescale/ls1021atwr/ls1021atwr.c index b7458a9..c565e91 100644 --- a/board/freescale/ls1021atwr/ls1021atwr.c +++ b/board/freescale/ls1021atwr/ls1021atwr.c @@ -501,6 +501,16 @@ int board_early_init_f(void) out_be32(&scfg->endiancr, SCFG_ENDIANCR_LE);
/*
* Memory controller require a register write before being enabled.* Affects: DDR* Register: EDDRTQCFG* Description: Memory controller performance is not optimal with* default internal target queue register values.* Workaround: Write a value of 63b2_0002h to address: 157_020Ch.*/- out_be32(&scfg->eddrtqcfg, 0x63b20002);
Same here.
[Yuan Yao] The same here.
- /*
*/
- Enable snoop requests and DVM message requests for
- Slave insterface S4 (A7 core cluster)
Regards Sinan Akman
On 08/13/2015 11:54 PM, Yuan Yao wrote:
Affects: DDR Description: Memory controller performance is not optimal with default internal target queue register values. Impact: Memory controller performance is not optimal. Workaround: Write a value of 63b2_0002h to address: 157_020Ch.
Signed-off-by: Yuan Yao yao.yuan@freescale.com
arch/arm/include/asm/arch-ls102xa/immap_ls102xa.h | 2 +- board/freescale/ls1021aqds/ls1021aqds.c | 10 ++++++++++ board/freescale/ls1021atwr/ls1021atwr.c | 10 ++++++++++ 3 files changed, 21 insertions(+), 1 deletion(-)
Yuan,
This is a wrong place to implement this workaround. It is a SoC erratum, not a board erratum. This erratum applies to both LS1021A and LS2085A. The workaround for LS2085A is already implemented in drivers/ddr/fsl/fsl_ddr_gen4.c. LS1021A supports both DDR3 and DDR4 modes. I started a discussion (CC'ed you) with the designer to clarify if both modes are impacted. If yes, you can add the workaround in drivers/ddr/fsl/arm_ddr_gen3.c, and make sure to add proper address offset in drivers/ddr/fsl/fsl_ddr_gen4.c.
York
participants (4)
-
Sinan Akman -
Yao Yuan -
York Sun -
Yuan Yao