[U-Boot] [PATCH] ddr: altera: Add ECC DRAM scrubbing support for Stratix 10

From: Tien Fong Chee tien.fong.chee@intel.com
The SDRAM must first be rewritten by zeroes if ECC is used to initialize the ECC metadata. Make the CPU overwrite the DRAM with zeroes in such a case. This scrubbing implementation turns the caches on temporarily, then overwrites the whole RAM with zeroes, flushes the caches and turns them off again. This provides satisfactory performance.
Signed-off-by: Tien Fong Chee tien.fong.chee@intel.com --- drivers/ddr/altera/sdram_s10.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-)
diff --git a/drivers/ddr/altera/sdram_s10.c b/drivers/ddr/altera/sdram_s10.c index 48f4f47..cce261f 100644 --- a/drivers/ddr/altera/sdram_s10.c +++ b/drivers/ddr/altera/sdram_s10.c @@ -8,6 +8,7 @@ #include <errno.h> #include <div64.h> #include <asm/io.h> +#include <linux/sizes.h> #include <wait_bit.h> #include <asm/arch/firewall_s10.h> #include <asm/arch/sdram_s10.h> @@ -134,6 +135,47 @@ static int poll_hmc_clock_status(void) SYSMGR_HMC_CLK_STATUS_MSK, true, 1000, false); }
+/* Initialize SDRAM ECC bits to avoid false DBE */ +static void sdram_init_ecc_bits(unsigned long long size) +{ + /* 1GB per chunk */ + unsigned long long size_byte = SZ_1G; + unsigned long long remaining_size; + unsigned long long dst_addr = 0x8000; + unsigned int start = get_timer(0); + + icache_enable(); + + memset(0, 0, dst_addr); + gd->arch.tlb_addr = 0x4000; + gd->arch.tlb_size = PGTABLE_SIZE; + + dcache_enable(); + + remaining_size = size - dst_addr; + printf("DDRCAL: Scrubbing ECC RAM (%d MiB).\n", (u32)(size >> 20)); + + while (remaining_size) { + if (remaining_size <= size_byte) { + memset((void *)dst_addr, 0, remaining_size); + break; + } else { + memset((void *)dst_addr, 0, size_byte); + dst_addr += size_byte; + } + + WATCHDOG_RESET(); + remaining_size -= size_byte; + } + + flush_dcache_all(); + printf("DDRCAL: Scrubbing ECC RAM done.\n"); + dcache_disable(); + + printf("SDRAM-ECC: Initialized success with %d ms\n", + (unsigned)get_timer(start)); +} + /** * sdram_mmr_init_full() - Function to initialize SDRAM MMR * @@ -351,6 +393,8 @@ int sdram_mmr_init_full(unsigned int unused) setbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL2, (DDR_HMC_ECCCTL2_RMW_EN_SET_MSK | DDR_HMC_ECCCTL2_AWB_EN_SET_MSK)); + + sdram_init_ecc_bits(gd->ram_size); } else { clrbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL1, (DDR_HMC_ECCCTL_AWB_CNT_RST_SET_MSK |

On 07/23/2018 10:20 AM, tien.fong.chee@intel.com wrote:
From: Tien Fong Chee tien.fong.chee@intel.com
The SDRAM must first be rewritten by zeroes if ECC is used to initialize the ECC metadata. Make the CPU overwrite the DRAM with zeroes in such a case. This scrubbing implementation turns the caches on temporarily, then overwrites the whole RAM with zeroes, flushes the caches and turns them off again. This provides satisfactory performance.
Signed-off-by: Tien Fong Chee tien.fong.chee@intel.com
drivers/ddr/altera/sdram_s10.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-)
diff --git a/drivers/ddr/altera/sdram_s10.c b/drivers/ddr/altera/sdram_s10.c index 48f4f47..cce261f 100644 --- a/drivers/ddr/altera/sdram_s10.c +++ b/drivers/ddr/altera/sdram_s10.c @@ -8,6 +8,7 @@ #include <errno.h> #include <div64.h> #include <asm/io.h> +#include <linux/sizes.h> #include <wait_bit.h> #include <asm/arch/firewall_s10.h> #include <asm/arch/sdram_s10.h> @@ -134,6 +135,47 @@ static int poll_hmc_clock_status(void) SYSMGR_HMC_CLK_STATUS_MSK, true, 1000, false); }
+/* Initialize SDRAM ECC bits to avoid false DBE */ +static void sdram_init_ecc_bits(unsigned long long size) +{
- /* 1GB per chunk */
- unsigned long long size_byte = SZ_1G;
- unsigned long long remaining_size;
- unsigned long long dst_addr = 0x8000;
- unsigned int start = get_timer(0);
- icache_enable();
- memset(0, 0, dst_addr);
- gd->arch.tlb_addr = 0x4000;
- gd->arch.tlb_size = PGTABLE_SIZE;
Are you sure this is valid on arm64 ? It looks like something copies from arria10.
- dcache_enable();
- remaining_size = size - dst_addr;
- printf("DDRCAL: Scrubbing ECC RAM (%d MiB).\n", (u32)(size >> 20));
- while (remaining_size) {
if (remaining_size <= size_byte) {
memset((void *)dst_addr, 0, remaining_size);
break;
} else {
memset((void *)dst_addr, 0, size_byte);
dst_addr += size_byte;
}
WATCHDOG_RESET();
remaining_size -= size_byte;
- }
How long does this take ?
- flush_dcache_all();
- printf("DDRCAL: Scrubbing ECC RAM done.\n");
- dcache_disable();
- printf("SDRAM-ECC: Initialized success with %d ms\n",
- (unsigned)get_timer(start));
+}
/**
- sdram_mmr_init_full() - Function to initialize SDRAM MMR
@@ -351,6 +393,8 @@ int sdram_mmr_init_full(unsigned int unused) setbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL2, (DDR_HMC_ECCCTL2_RMW_EN_SET_MSK | DDR_HMC_ECCCTL2_AWB_EN_SET_MSK));
} else { clrbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL1, (DDR_HMC_ECCCTL_AWB_CNT_RST_SET_MSK |sdram_init_ecc_bits(gd->ram_size);

On Mon, 2018-07-23 at 11:46 +0200, Marek Vasut wrote:
On 07/23/2018 10:20 AM, tien.fong.chee@intel.com wrote:
From: Tien Fong Chee tien.fong.chee@intel.com
The SDRAM must first be rewritten by zeroes if ECC is used to initialize the ECC metadata. Make the CPU overwrite the DRAM with zeroes in such a case. This scrubbing implementation turns the caches on temporarily, then overwrites the whole RAM with zeroes, flushes the caches and turns them off again. This provides satisfactory performance.
Signed-off-by: Tien Fong Chee tien.fong.chee@intel.com
drivers/ddr/altera/sdram_s10.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-)
diff --git a/drivers/ddr/altera/sdram_s10.c b/drivers/ddr/altera/sdram_s10.c index 48f4f47..cce261f 100644 --- a/drivers/ddr/altera/sdram_s10.c +++ b/drivers/ddr/altera/sdram_s10.c @@ -8,6 +8,7 @@ #include <errno.h> #include <div64.h> #include <asm/io.h> +#include <linux/sizes.h> #include <wait_bit.h> #include <asm/arch/firewall_s10.h> #include <asm/arch/sdram_s10.h> @@ -134,6 +135,47 @@ static int poll_hmc_clock_status(void) SYSMGR_HMC_CLK_STATUS_MSK, true, 1000, false); } +/* Initialize SDRAM ECC bits to avoid false DBE */ +static void sdram_init_ecc_bits(unsigned long long size) +{
- /* 1GB per chunk */
- unsigned long long size_byte = SZ_1G;
- unsigned long long remaining_size;
- unsigned long long dst_addr = 0x8000;
- unsigned int start = get_timer(0);
- icache_enable();
- memset(0, 0, dst_addr);
- gd->arch.tlb_addr = 0x4000;
- gd->arch.tlb_size = PGTABLE_SIZE;
Are you sure this is valid on arm64 ? It looks like something copies from arria10.
The cache on/off is copied from your implementation on Arria 10. Yes, i have tested it, it is working on Stratix 10 board.
- dcache_enable();
- remaining_size = size - dst_addr;
- printf("DDRCAL: Scrubbing ECC RAM (%d MiB).\n", (u32)(size
20));
- while (remaining_size) {
if (remaining_size <= size_byte) {
memset((void *)dst_addr, 0,
remaining_size);
break;
} else {
memset((void *)dst_addr, 0, size_byte);
dst_addr += size_byte;
}
WATCHDOG_RESET();
remaining_size -= size_byte;
- }
How long does this take ?
1359ms for 2GB. But I have no idea why Arria 10 board can't achieve the same result. Could you try again on your Arria 10 ES board?
- flush_dcache_all();
- printf("DDRCAL: Scrubbing ECC RAM done.\n");
- dcache_disable();
- printf("SDRAM-ECC: Initialized success with %d ms\n",
- (unsigned)get_timer(start));
+}
/** * sdram_mmr_init_full() - Function to initialize SDRAM MMR * @@ -351,6 +393,8 @@ int sdram_mmr_init_full(unsigned int unused) setbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL2, (DDR_HMC_ECCCTL2_RMW_EN_SET_MSK | DDR_HMC_ECCCTL2_AWB_EN_SET_MSK));
sdram_init_ecc_bits(gd->ram_size);
} else { clrbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL1, (DDR_HMC_ECCCTL_AWB_CNT_RST_SET_MSK |

On 07/24/2018 10:10 AM, Chee, Tien Fong wrote:
On Mon, 2018-07-23 at 11:46 +0200, Marek Vasut wrote:
On 07/23/2018 10:20 AM, tien.fong.chee@intel.com wrote:
From: Tien Fong Chee tien.fong.chee@intel.com
The SDRAM must first be rewritten by zeroes if ECC is used to initialize the ECC metadata. Make the CPU overwrite the DRAM with zeroes in such a case. This scrubbing implementation turns the caches on temporarily, then overwrites the whole RAM with zeroes, flushes the caches and turns them off again. This provides satisfactory performance.
Signed-off-by: Tien Fong Chee tien.fong.chee@intel.com
drivers/ddr/altera/sdram_s10.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-)
diff --git a/drivers/ddr/altera/sdram_s10.c b/drivers/ddr/altera/sdram_s10.c index 48f4f47..cce261f 100644 --- a/drivers/ddr/altera/sdram_s10.c +++ b/drivers/ddr/altera/sdram_s10.c @@ -8,6 +8,7 @@ #include <errno.h> #include <div64.h> #include <asm/io.h> +#include <linux/sizes.h> #include <wait_bit.h> #include <asm/arch/firewall_s10.h> #include <asm/arch/sdram_s10.h> @@ -134,6 +135,47 @@ static int poll_hmc_clock_status(void) SYSMGR_HMC_CLK_STATUS_MSK, true, 1000, false); } +/* Initialize SDRAM ECC bits to avoid false DBE */ +static void sdram_init_ecc_bits(unsigned long long size) +{
- /* 1GB per chunk */
- unsigned long long size_byte = SZ_1G;
- unsigned long long remaining_size;
- unsigned long long dst_addr = 0x8000;
- unsigned int start = get_timer(0);
- icache_enable();
- memset(0, 0, dst_addr);
- gd->arch.tlb_addr = 0x4000;
- gd->arch.tlb_size = PGTABLE_SIZE;
Are you sure this is valid on arm64 ? It looks like something copies from arria10.
The cache on/off is copied from your implementation on Arria 10. Yes, i have tested it, it is working on Stratix 10 board.
Right, except S10 is arm64, A10 is arm32, which is why I wonder whether careless copying is enough.
- dcache_enable();
- remaining_size = size - dst_addr;
- printf("DDRCAL: Scrubbing ECC RAM (%d MiB).\n", (u32)(size
20));
- while (remaining_size) {
if (remaining_size <= size_byte) {
memset((void *)dst_addr, 0,
remaining_size);
break;
} else {
memset((void *)dst_addr, 0, size_byte);
dst_addr += size_byte;
}
WATCHDOG_RESET();
remaining_size -= size_byte;
- }
How long does this take ?
1359ms for 2GB.
So why do you need this watchdog reset hack ?
But I have no idea why Arria 10 board can't achieve the same result. Could you try again on your Arria 10 ES board?
There's nothing to try, the scrubbing works fine on A10.
- flush_dcache_all();
- printf("DDRCAL: Scrubbing ECC RAM done.\n");
- dcache_disable();
- printf("SDRAM-ECC: Initialized success with %d ms\n",
- (unsigned)get_timer(start));
+}
/** * sdram_mmr_init_full() - Function to initialize SDRAM MMR * @@ -351,6 +393,8 @@ int sdram_mmr_init_full(unsigned int unused) setbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL2, (DDR_HMC_ECCCTL2_RMW_EN_SET_MSK | DDR_HMC_ECCCTL2_AWB_EN_SET_MSK));
sdram_init_ecc_bits(gd->ram_size);
} else { clrbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL1, (DDR_HMC_ECCCTL_AWB_CNT_RST_SET_MSK |

On Tue, 2018-07-24 at 12:52 +0200, Marek Vasut wrote:
On 07/24/2018 10:10 AM, Chee, Tien Fong wrote:
On Mon, 2018-07-23 at 11:46 +0200, Marek Vasut wrote:
On 07/23/2018 10:20 AM, tien.fong.chee@intel.com wrote:
From: Tien Fong Chee tien.fong.chee@intel.com
The SDRAM must first be rewritten by zeroes if ECC is used to initialize the ECC metadata. Make the CPU overwrite the DRAM with zeroes in such a case. This scrubbing implementation turns the caches on temporarily, then overwrites the whole RAM with zeroes, flushes the caches and turns them off again. This provides satisfactory performance.
Signed-off-by: Tien Fong Chee tien.fong.chee@intel.com
drivers/ddr/altera/sdram_s10.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-)
diff --git a/drivers/ddr/altera/sdram_s10.c b/drivers/ddr/altera/sdram_s10.c index 48f4f47..cce261f 100644 --- a/drivers/ddr/altera/sdram_s10.c +++ b/drivers/ddr/altera/sdram_s10.c @@ -8,6 +8,7 @@ #include <errno.h> #include <div64.h> #include <asm/io.h> +#include <linux/sizes.h> #include <wait_bit.h> #include <asm/arch/firewall_s10.h> #include <asm/arch/sdram_s10.h> @@ -134,6 +135,47 @@ static int poll_hmc_clock_status(void) SYSMGR_HMC_CLK_STATUS_MSK, true, 1000, false); } +/* Initialize SDRAM ECC bits to avoid false DBE */ +static void sdram_init_ecc_bits(unsigned long long size) +{
- /* 1GB per chunk */
- unsigned long long size_byte = SZ_1G;
- unsigned long long remaining_size;
- unsigned long long dst_addr = 0x8000;
- unsigned int start = get_timer(0);
- icache_enable();
- memset(0, 0, dst_addr);
- gd->arch.tlb_addr = 0x4000;
- gd->arch.tlb_size = PGTABLE_SIZE;
Are you sure this is valid on arm64 ? It looks like something copies from arria10.
The cache on/off is copied from your implementation on Arria 10. Yes, i have tested it, it is working on Stratix 10 board.
Right, except S10 is arm64, A10 is arm32, which is why I wonder whether careless copying is enough.
The PGTABLE_SIZE is getting from the function get_page_table_size() which is specific for armv8 architecture. I'm also testing it on board to ensure sufficient RAM for the page table.
- dcache_enable();
- remaining_size = size - dst_addr;
- printf("DDRCAL: Scrubbing ECC RAM (%d MiB).\n",
(u32)(size
20));
- while (remaining_size) {
if (remaining_size <= size_byte) {
memset((void *)dst_addr, 0,
remaining_size);
break;
} else {
memset((void *)dst_addr, 0,
size_byte);
dst_addr += size_byte;
}
WATCHDOG_RESET();
remaining_size -= size_byte;
- }
How long does this take ?
1359ms for 2GB.
So why do you need this watchdog reset hack ?
Yeah, for size scalable. When size growing up, it would take longer to scrub. Without watchdog reset on time, it might be causing system to reset.
But I have no idea why Arria 10 board can't achieve the same result. Could you try again on your Arria 10 ES board?
There's nothing to try, the scrubbing works fine on A10.
Then, that really weird it took so long to finish scrubbing.
- flush_dcache_all();
- printf("DDRCAL: Scrubbing ECC RAM done.\n");
- dcache_disable();
- printf("SDRAM-ECC: Initialized success with %d ms\n",
- (unsigned)get_timer(start));
+}
/** * sdram_mmr_init_full() - Function to initialize SDRAM MMR * @@ -351,6 +393,8 @@ int sdram_mmr_init_full(unsigned int unused) setbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL2, (DDR_HMC_ECCCTL2_RMW_EN_SET_MSK | DDR_HMC_ECCCTL2_AWB_EN_SET_MSK)) ;
sdram_init_ecc_bits(gd->ram_size);
} else { clrbits_le32(SOCFPGA_SDR_ADDRESS + ECCCTRL1, (DDR_HMC_ECCCTL_AWB_CNT_RST_SET_M SK |
participants (3)
-
Chee, Tien Fong
-
Marek Vasut
-
tien.fong.chee@intel.com