mmc: sdhci-esdhc: Add SDHCI_QUIRK_32BIT_DMA_ADDR
Subsystems (4)
PATCH 1/4 mmc: sdhci-esdhc: Add SDHCI_QUIRK_32BIT_DMA_ADDR
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of mmc_test:
11: Badly aligned write
12: Badly aligned read
13: Badly aligned multi-block write
14: Badly aligned multi-block read
--- drivers/mmc/host/sdhci-esdhc.h | 1 + 1 file changed, 1 insertion(+)
drivers/mmc/host/sdhci-esdhc.h +1 -0
@@ -19,6 +19,7 @@*/ #define ESDHC_DEFAULT_QUIRKS (SDHCI_QUIRK_FORCE_BLK_SZ_2048 | \ + SDHCI_QUIRK_32BIT_DMA_ADDR | \ SDHCI_QUIRK_NO_BUSY_IRQ | \ SDHCI_QUIRK_DATA_TIMEOUT_USES_SDCLK | \ SDHCI_QUIRK_PIO_NEEDS_DELAY | \
Signature · 1 line
2.7.4
PATCH 2/4 mmc: sdhci-esdhc-imx: Fix DAT line software reset
On i.MX25, the eSDHC DAT line software reset (SYSCTL.RSTD) unexpectedly clears at least the data transfer width (PROCTL.DTW), which then results in data CRC errors. This behavior is not documented, but it has actually been observed. Consequently, the DAT line software resets triggered by sdhci.c in case of errors caused unrecoverable errors.
Fix this by making sure that the DAT line software reset does not alter the Host Control register. This behavior being undocumented, it may also be present on other i.MX SoCs, so apply this fix for the whole i.MX family.
--- drivers/mmc/host/sdhci-esdhc-imx.c | 59 ++++++++++++++++++++++++-------------- 1 file changed, 38 insertions(+), 21 deletions(-)
drivers/mmc/host/sdhci-esdhc-imx.c +38 -21
@@ -579,7 +579,7 @@ static void esdhc_writeb_le(struct sdhci_host *host, u8 val, int reg) {struct sdhci_pltfm_host *pltfm_host =sdhci_priv (host);struct pltfm_imx_data *imx_data =sdhci_pltfm_priv (pltfm_host); -u32 new_val; +u32 new_val =0 ;u32 mask;switch (reg) {
@@ -610,29 +610,46 @@ static void esdhc_writeb_le(struct sdhci_host *host, u8 val, int reg)esdhc_clrset_le (host, mask, new_val, reg);return ; +case SDHCI_SOFTWARE_RESET : +if (val &SDHCI_RESET_DATA ) + new_val =readl (host->ioaddr +SDHCI_HOST_CONTROL ); +break ; }esdhc_clrset_le (host,0xff , val, reg); -/* -* The esdhc has a design violation to SDHC spec which tells -* that software reset should not affect card detection circuit. -* But esdhc clears its SYSCTL register bits [0..2] during the -* software reset. This will stop those clocks that card detection -* circuit relies on. To work around it, we turn the clocks on back -* to keep card detection circuit functional. -*/ -if ((reg ==SDHCI_SOFTWARE_RESET ) && (val &1 )) { -esdhc_clrset_le (host,0x7 ,0x7 ,ESDHC_SYSTEM_CONTROL ); -/* -* The reset on usdhc fails to clear MIX_CTRL register. -* Do it manually here. -*/ -if (esdhc_is_usdhc (imx_data)) { -/* the tuning bits should be kept during reset */ - new_val =readl (host->ioaddr +ESDHC_MIX_CTRL ); -writel (new_val &ESDHC_MIX_CTRL_TUNING_MASK , - host->ioaddr +ESDHC_MIX_CTRL ); - imx_data->is_ddr =0 ; +if (reg ==SDHCI_SOFTWARE_RESET ) { +if (val &SDHCI_RESET_ALL ) { +/* +* The esdhc has a design violation to SDHC spec which +* tells that software reset should not affect card +* detection circuit. But esdhc clears its SYSCTL +* register bits [0..2] during the software reset. This +* will stop those clocks that card detection circuit +* relies on. To work around it, we turn the clocks on +* back to keep card detection circuit functional. +*/ +esdhc_clrset_le (host,0x7 ,0x7 ,ESDHC_SYSTEM_CONTROL ); +/* +* The reset on usdhc fails to clear MIX_CTRL register. +* Do it manually here. +*/ +if (esdhc_is_usdhc (imx_data)) { +/* +* the tuning bits should be kept during reset +*/ + new_val =readl (host->ioaddr +ESDHC_MIX_CTRL ); +writel (new_val &ESDHC_MIX_CTRL_TUNING_MASK , + host->ioaddr +ESDHC_MIX_CTRL ); + imx_data->is_ddr =0 ; + } + }else if (val &SDHCI_RESET_DATA ) { +/* +* The eSDHC DAT line software reset clears at least the +* data transfer width on i.MX25, so make sure that the +* Host Control register is unaffected. +*/ +esdhc_clrset_le (host,0xff , new_val, +SDHCI_HOST_CONTROL ); } } }
Signature · 1 line
2.7.4
Re: [PATCH 2/4] mmc: sdhci-esdhc-imx: Fix DAT line software reset
Quoted 6 lines · Benoît Thébaudeau
On i.MX25, the eSDHC DAT line software reset (SYSCTL.RSTD) unexpectedly clears at least the data transfer width (PROCTL.DTW), which then results in data CRC errors. This behavior is not documented, but it has actually been observed. Consequently, the DAT line software resets triggered by
sdhci.cin case of errors caused unrecoverable errors.
Quoted 6 lines
Fix this by making sure that the DAT line software reset does not alter the Host Control register. This behavior being undocumented, it may also be present on other i.MX SoCs, so apply this fix for the whole i.MX family.
Signed-off-by Benoît Thébaudeau <[email protected]>
I would expect to see Acks from other sdhci-esdhc-imx users. Nevertheless, for sdhci:
Quoted 85 lines
--- drivers/mmc/host/sdhci-esdhc-imx.c | 59 ++++++++++++++++++++++++-------------- 1 file changed, 38 insertions(+), 21 deletions(-)drivers/mmc/host/sdhci-esdhc-imx.c +38
-21@@ -579,7 +579,7 @@ static void esdhc_writeb_le(struct sdhci_host *host, u8 val, int reg) {struct sdhci_pltfm_host *pltfm_host =sdhci_priv (host);struct pltfm_imx_data *imx_data =sdhci_pltfm_priv (pltfm_host); -u32 new_val; +u32 new_val =0 ;u32 mask;switch (reg) {@@ -610,29 +610,46 @@ static void esdhc_writeb_le(struct sdhci_host *host, u8 val, int reg)esdhc_clrset_le (host, mask, new_val, reg);return ; +case SDHCI_SOFTWARE_RESET : +if (val &SDHCI_RESET_DATA ) + new_val =readl (host->ioaddr +SDHCI_HOST_CONTROL ); +break ; }esdhc_clrset_le (host,0xff , val, reg); -/* -* The esdhc has a design violation to SDHC spec which tells -* that software reset should not affect card detection circuit. -* But esdhc clears its SYSCTL register bits [0..2] during the -* software reset. This will stop those clocks that card detection -* circuit relies on. To work around it, we turn the clocks on back -* to keep card detection circuit functional. -*/ -if ((reg ==SDHCI_SOFTWARE_RESET ) && (val &1 )) { -esdhc_clrset_le (host,0x7 ,0x7 ,ESDHC_SYSTEM_CONTROL ); -/* -* The reset on usdhc fails to clear MIX_CTRL register. -* Do it manually here. -*/ -if (esdhc_is_usdhc (imx_data)) { -/* the tuning bits should be kept during reset */ - new_val =readl (host->ioaddr +ESDHC_MIX_CTRL ); -writel (new_val &ESDHC_MIX_CTRL_TUNING_MASK , - host->ioaddr +ESDHC_MIX_CTRL ); - imx_data->is_ddr =0 ; +if (reg ==SDHCI_SOFTWARE_RESET ) { +if (val &SDHCI_RESET_ALL ) { +/* +* The esdhc has a design violation to SDHC spec which +* tells that software reset should not affect card +* detection circuit. But esdhc clears its SYSCTL +* register bits [0..2] during the software reset. This +* will stop those clocks that card detection circuit +* relies on. To work around it, we turn the clocks on +* back to keep card detection circuit functional. +*/ +esdhc_clrset_le (host,0x7 ,0x7 ,ESDHC_SYSTEM_CONTROL ); +/* +* The reset on usdhc fails to clear MIX_CTRL register. +* Do it manually here. +*/ +if (esdhc_is_usdhc (imx_data)) { +/* +* the tuning bits should be kept during reset +*/ + new_val =readl (host->ioaddr +ESDHC_MIX_CTRL ); +writel (new_val &ESDHC_MIX_CTRL_TUNING_MASK , + host->ioaddr +ESDHC_MIX_CTRL ); + imx_data->is_ddr =0 ; + } + }else if (val &SDHCI_RESET_DATA ) { +/* +* The eSDHC DAT line software reset clears at least the +* data transfer width on i.MX25, so make sure that the +* Host Control register is unaffected. +*/ +esdhc_clrset_le (host,0xff , new_val, +SDHCI_HOST_CONTROL ); } } }
Re: [PATCH 2/4] mmc: sdhci-esdhc-imx: Fix DAT line software reset
In reply to Benoît ThébaudeauQuoted 6 lines · Benoît Thébaudeau
On i.MX25, the eSDHC DAT line software reset (SYSCTL.RSTD) unexpectedly clears at least the data transfer width (PROCTL.DTW), which then results in data CRC errors. This behavior is not documented, but it has actually been observed. Consequently, the DAT line software resets triggered by
sdhci.cin case of errors caused unrecoverable errors.
Quoted 6 lines
Fix this by making sure that the DAT line software reset does not alter the Host Control register. This behavior being undocumented, it may also be present on other i.MX SoCs, so apply this fix for the whole i.MX family.
Signed-off-by Benoît Thébaudeau <[email protected]>
PATCH 3/4 mmc: sdhci-esdhc-imx: Allow all supported prescaler values
In reply to Benoît ThébaudeauOn i.MX, SYSCTL.SDCLKFS may always be set to 0 in order to make the SD clock frequency prescaler divide by 1 in SDR mode, even with the eSDHC. The previous minimum prescaler value of 2 in SDR mode with the eSDHC was a code remnant from PowerPC, which actually has this limitation on earlier revisions.
In DDR mode, the prescaler can divide by up to 512.
The maximum SD clock frequency in High Speed mode is 50 MHz. On i.MX25, this change makes it possible to get 48 MHz from the USB PLL (240 MHz / 5 / 1) instead of only 40 MHz from the USB PLL (240 MHz / 3 / 2) or 33.25 MHz from the AHB clock (133 MHz / 2 / 2).
--- drivers/mmc/host/sdhci-esdhc-imx.c | 18 +++++++----------- 1 file changed, 7 insertions(+), 11 deletions(-)
drivers/mmc/host/sdhci-esdhc-imx.c +7 -11
@@ -674,7 +674,8 @@ static inline void esdhc_pltfm_set_clock(struct sdhci_host *host,struct sdhci_pltfm_host *pltfm_host =sdhci_priv (host);struct pltfm_imx_data *imx_data =sdhci_pltfm_priv (pltfm_host);unsigned int host_clock = pltfm_host->clock; -int pre_div =2 ; +int ddr_pre_div = imx_data->is_ddr ?2 :1 ; +int pre_div =1 ;int div =1 ;u32 temp, val;
@@ -689,28 +690,23 @@ static inline void esdhc_pltfm_set_clock(struct sdhci_host *host,return ; } -if (esdhc_is_usdhc (imx_data) && !imx_data->is_ddr) - pre_div =1 ; - temp =sdhci_readl (host,ESDHC_SYSTEM_CONTROL ); temp &= ~(ESDHC_CLOCK_IPGEN |ESDHC_CLOCK_HCKEN |ESDHC_CLOCK_PEREN |ESDHC_CLOCK_MASK );sdhci_writel (host, temp,ESDHC_SYSTEM_CONTROL ); -while (host_clock / pre_div /16 > clock && pre_div <256 ) +while (host_clock / (16 * pre_div * ddr_pre_div) > clock && + pre_div <256 ) pre_div *=2 ; -while (host_clock / pre_div / div > clock && div <16 ) +while (host_clock / (div * pre_div * ddr_pre_div) > clock && div <16 ) div++; - host->mmc->actual_clock = host_clock / pre_div / div; + host->mmc->actual_clock = host_clock / (div * pre_div * ddr_pre_div);dev_dbg (mmc_dev (host->mmc),"desired SD clock: %d, actual: %d\n" , clock, host->mmc->actual_clock); -if (imx_data->is_ddr) - pre_div >>=2 ; -else - pre_div >>=1 ; + pre_div >>=1 ; div--; temp =sdhci_readl (host,ESDHC_SYSTEM_CONTROL );
Signature · 1 line
2.7.4
Re: [PATCH 3/4] mmc: sdhci-esdhc-imx: Allow all supported prescaler values
Quoted 6 lines · Benoît Thébaudeau
On i.MX, SYSCTL.SDCLKFS may always be set to 0 in order to make the SD clock frequency prescaler divide by 1 in SDR mode, even with the eSDHC. The previous minimum prescaler value of 2 in SDR mode with the eSDHC was a code remnant from PowerPC, which actually has this limitation on earlier revisions.
Quoted 8 lines
In DDR mode, the prescaler can divide by up to 512.
The maximum SD clock frequency in High Speed mode is 50 MHz. On i.MX25, this change makes it possible to get 48 MHz from the USB PLL (240 MHz / 5 / 1) instead of only 40 MHz from the USB PLL (240 MHz / 3 / 2) or 33.25 MHz from the AHB clock (133 MHz / 2 / 2).
Signed-off-by Benoît Thébaudeau <[email protected]>
I would expect to see Acks from other sdhci-esdhc-imx users. Nevertheless, for sdhci:
Quoted 53 lines
--- drivers/mmc/host/sdhci-esdhc-imx.c | 18 +++++++----------- 1 file changed, 7 insertions(+), 11 deletions(-)drivers/mmc/host/sdhci-esdhc-imx.c +7
-11@@ -674,7 +674,8 @@ static inline void esdhc_pltfm_set_clock(struct sdhci_host *host,struct sdhci_pltfm_host *pltfm_host =sdhci_priv (host);struct pltfm_imx_data *imx_data =sdhci_pltfm_priv (pltfm_host);unsigned int host_clock = pltfm_host->clock; -int pre_div =2 ; +int ddr_pre_div = imx_data->is_ddr ?2 :1 ; +int pre_div =1 ;int div =1 ;u32 temp, val;@@ -689,28 +690,23 @@ static inline void esdhc_pltfm_set_clock(struct sdhci_host *host,return ; } -if (esdhc_is_usdhc (imx_data) && !imx_data->is_ddr) - pre_div =1 ; - temp =sdhci_readl (host,ESDHC_SYSTEM_CONTROL ); temp &= ~(ESDHC_CLOCK_IPGEN |ESDHC_CLOCK_HCKEN |ESDHC_CLOCK_PEREN |ESDHC_CLOCK_MASK );sdhci_writel (host, temp,ESDHC_SYSTEM_CONTROL ); -while (host_clock / pre_div /16 > clock && pre_div <256 ) +while (host_clock / (16 * pre_div * ddr_pre_div) > clock && + pre_div <256 ) pre_div *=2 ; -while (host_clock / pre_div / div > clock && div <16 ) +while (host_clock / (div * pre_div * ddr_pre_div) > clock && div <16 ) div++; - host->mmc->actual_clock = host_clock / pre_div / div; + host->mmc->actual_clock = host_clock / (div * pre_div * ddr_pre_div);dev_dbg (mmc_dev (host->mmc),"desired SD clock: %d, actual: %d\n" , clock, host->mmc->actual_clock); -if (imx_data->is_ddr) - pre_div >>=2 ; -else - pre_div >>=1 ; + pre_div >>=1 ; div--; temp =sdhci_readl (host,ESDHC_SYSTEM_CONTROL );
Re: [PATCH 3/4] mmc: sdhci-esdhc-imx: Allow all supported prescaler values
In reply to Benoît ThébaudeauQuoted 6 lines · Benoît Thébaudeau
On i.MX, SYSCTL.SDCLKFS may always be set to 0 in order to make the SD clock frequency prescaler divide by 1 in SDR mode, even with the eSDHC. The previous minimum prescaler value of 2 in SDR mode with the eSDHC was a code remnant from PowerPC, which actually has this limitation on earlier revisions.
Quoted 8 lines
In DDR mode, the prescaler can divide by up to 512.
The maximum SD clock frequency in High Speed mode is 50 MHz. On i.MX25, this change makes it possible to get 48 MHz from the USB PLL (240 MHz / 5 / 1) instead of only 40 MHz from the USB PLL (240 MHz / 3 / 2) or 33.25 MHz from the AHB clock (133 MHz / 2 / 2).
Signed-off-by Benoît Thébaudeau <[email protected]>
PATCH 4/4 mmc: sdhci-esdhc-imx: Remove the ENGcm07207 workaround
In reply to Benoît ThébaudeauThe SDHCI_QUIRK_NO_MULTIBLOCK quirk was used as a workaround for the ENGcm07207 erratum. However, it caused excruciatingly slow SD transfers (300 kB/s on average), and this erratum actually does not imply that multiple-block transfers are not supported, so this was overkill.
The suggested workaround for this erratum is to set SYSCTL.RSTA, but the simple DAT line software reset (which resets the DMA circuit among others) triggered by sdhci_finish_data() in case of errors seems to be sufficient. Indeed, generating errors in a controlled manner on i.MX25 using the FEVT register right in the middle of read data transfers without this quirk shows that nothing is written to the buffer by the eSDHC past CMD12, and no extra Auto CMD12 is sent with AC12EN set, so the data transfers on AHB are properly aborted. For write data transfers, neither extra data nor extra Auto CMD12 is sent, as expected. Moreover, after intensive stress tests on i.MX25, removing SDHCI_QUIRK_NO_MULTIBLOCK seems to be safe.
--- drivers/mmc/host/sdhci-esdhc-imx.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)
drivers/mmc/host/sdhci-esdhc-imx.c +3 -13
@@ -115,11 +115,6 @@*/ #define ESDHC_FLAG_MULTIBLK_NO_INT BIT(1)/* -* The flag enables the workaround for ESDHC errata ENGcm07207 which -* affects i.MX25 and i.MX35. -*/ -#define ESDHC_FLAG_ENGCM07207 BIT(2) -/* * The flag tells that the ESDHC controller is an USDHC block that is * integrated on the i.MX6 series. */
@@ -149,11 +144,11 @@ struct esdhc_soc_data { };static struct esdhc_soc_data esdhc_imx25_data = { - .flags =ESDHC_FLAG_ENGCM07207 , + .flags =ESDHC_FLAG_ERR004536 , };static struct esdhc_soc_data esdhc_imx35_data = { - .flags =ESDHC_FLAG_ENGCM07207 , + .flags =ESDHC_FLAG_ERR004536 , };static struct esdhc_soc_data esdhc_imx51_data = {
@@ -286,7 +281,7 @@ static u32 esdhc_readl_le(struct sdhci_host *host, int reg)* ADMA2 capability of esdhc, but this bit is messed up on * some SOCs (e.g. on MX25, MX35 this bit is set, but they * don't actually support ADMA2). So set the BROKEN_ADMA -* uirk on MX25/35 platforms. +* quirk on MX25/35 platforms. */ if (val &SDHCI_CAN_DO_ADMA1 ) {
@@ -1278,11 +1273,6 @@ static int sdhci_esdhc_imx_probe(struct platform_device *pdev)if (IS_ERR (imx_data->pins_default))dev_warn (mmc_dev (host->mmc),"could not get default state\n" ); -if (imx_data->socdata->flags &ESDHC_FLAG_ENGCM07207 ) -/* Fix errata ENGcm07207 present on i.MX25 and i.MX35 */ - host->quirks |=SDHCI_QUIRK_NO_MULTIBLOCK - |SDHCI_QUIRK_BROKEN_ADMA ; -if (esdhc_is_usdhc (imx_data)) { host->quirks2 |=SDHCI_QUIRK2_PRESET_VALUE_BROKEN ; host->mmc->caps |=MMC_CAP_1_8V_DDR ;
Signature · 1 line
2.7.4
Re: [PATCH 4/4] mmc: sdhci-esdhc-imx: Remove the ENGcm07207 workaround
Quoted 17 lines · Benoît Thébaudeau
The
SDHCI_QUIRK_NO_MULTIBLOCKquirk was used as a workaround for the ENGcm07207 erratum. However, it caused excruciatingly slow SD transfers (300 kB/s on average), and this erratum actually does not imply that multiple-block transfers are not supported, so this was overkill.The suggested workaround for this erratum is to set SYSCTL.RSTA, but the simple DAT line software reset (which resets the DMA circuit among others) triggered by
sdhci_finish_data()in case of errors seems to be sufficient. Indeed, generating errors in a controlled manner on i.MX25 using the FEVT register right in the middle of read data transfers without this quirk shows that nothing is written to the buffer by the eSDHC past CMD12, and no extra Auto CMD12 is sent with AC12EN set, so the data transfers on AHB are properly aborted. For write data transfers, neither extra data nor extra Auto CMD12 is sent, as expected. Moreover, after intensive stress tests on i.MX25, removingSDHCI_QUIRK_NO_MULTIBLOCKseems to be safe.
Quoted 1 line
Signed-off-by Benoît Thébaudeau <[email protected]>
Aside from one comment below...
I would expect to see Acks from other sdhci-esdhc-imx users. Nevertheless, for sdhci:
Quoted 35 lines
--- drivers/mmc/host/sdhci-esdhc-imx.c | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-)drivers/mmc/host/sdhci-esdhc-imx.c +2
-7@@ -115,11 +115,6 @@*/ #define ESDHC_FLAG_MULTIBLK_NO_INT BIT(1)/* -* The flag enables the workaround for ESDHC errata ENGcm07207 which -* affects i.MX25 and i.MX35. -*/ -#define ESDHC_FLAG_ENGCM07207 BIT(2) -/* * The flag tells that the ESDHC controller is an USDHC block that is * integrated on the i.MX6 series. */ @@ -149,11 +144,11 @@ struct esdhc_soc_data { };static struct esdhc_soc_data esdhc_imx25_data = { - .flags =ESDHC_FLAG_ENGCM07207 , + .flags =ESDHC_FLAG_ERR004536 , };static struct esdhc_soc_data esdhc_imx35_data = { - .flags =ESDHC_FLAG_ENGCM07207 , + .flags =ESDHC_FLAG_ERR004536 , };static struct esdhc_soc_data esdhc_imx51_data = {
drivers/mmc/host/sdhci-esdhc-imx.c static u32 esdhc_readl_le(struct sdhci_host *host, int reg) · line 284
Quoted 5 lines
drivers/mmc/host/sdhci-esdhc-imx.c +1
-1@@ -286,7 +281,7 @@ static u32 esdhc_readl_le(struct sdhci_host *host, int reg)* ADMA2 capability of esdhc, but this bit is messed up on * some SOCs (e.g. on MX25, MX35 this bit is set, but they * don't actually support ADMA2). So set the BROKEN_ADMA -* uirk on MX25/35 platforms. +* quirk on MX25/35 platforms.
Please do spelling fixes as a separate patch. Note there are several spelling errors in comments in drivers/mmc/host/sdhci-esdhc-imx.c so you could easily do a few of them to make a separate patch more worthwhile.
Quoted 16 lines
drivers/mmc/host/sdhci-esdhc-imx.c +0
-5*/ if (val &SDHCI_CAN_DO_ADMA1 ) {@@ -1278,11 +1273,6 @@ static int sdhci_esdhc_imx_probe(struct platform_device *pdev)if (IS_ERR (imx_data->pins_default))dev_warn (mmc_dev (host->mmc),"could not get default state\n" ); -if (imx_data->socdata->flags &ESDHC_FLAG_ENGCM07207 ) -/* Fix errata ENGcm07207 present on i.MX25 and i.MX35 */ - host->quirks |=SDHCI_QUIRK_NO_MULTIBLOCK - |SDHCI_QUIRK_BROKEN_ADMA ; -if (esdhc_is_usdhc (imx_data)) { host->quirks2 |=SDHCI_QUIRK2_PRESET_VALUE_BROKEN ; host->mmc->caps |=MMC_CAP_1_8V_DDR ;
Re: [PATCH 4/4] mmc: sdhci-esdhc-imx: Remove the ENGcm07207 workaround
Quoted 18 lines · Adrian Hunter
Quoted 17 lines · Benoît Thébaudeau
The
SDHCI_QUIRK_NO_MULTIBLOCKquirk was used as a workaround for the ENGcm07207 erratum. However, it caused excruciatingly slow SD transfers (300 kB/s on average), and this erratum actually does not imply that multiple-block transfers are not supported, so this was overkill.The suggested workaround for this erratum is to set SYSCTL.RSTA, but the simple DAT line software reset (which resets the DMA circuit among others) triggered by
sdhci_finish_data()in case of errors seems to be sufficient. Indeed, generating errors in a controlled manner on i.MX25 using the FEVT register right in the middle of read data transfers without this quirk shows that nothing is written to the buffer by the eSDHC past CMD12, and no extra Auto CMD12 is sent with AC12EN set, so the data transfers on AHB are properly aborted. For write data transfers, neither extra data nor extra Auto CMD12 is sent, as expected. Moreover, after intensive stress tests on i.MX25, removingSDHCI_QUIRK_NO_MULTIBLOCKseems to be safe.
Quoted 6 lines
Quoted 1 line
Signed-off-by Benoît Thébaudeau <[email protected]>Aside from one comment below...
I would expect to see Acks from other sdhci-esdhc-imx users. Nevertheless, for sdhci:
Yes, I would also appreciate some ack/tested by, from the corresponding sdhci variant users of this series.
However, to allow it to get some results from linux-next, I have queued up this series for next (amending $subject patch according to the comment from Adrian). Thanks!
Quoted 4 lines
Acked-by Adrian Hunter <[email protected]>
[...]
Kind regards
Uffe
Re: [PATCH 4/4] mmc: sdhci-esdhc-imx: Remove the ENGcm07207 workaround
Quoted 26 lines · Ulf Hansson
Quoted 18 lines · Adrian Hunter
Quoted 17 lines · Benoît Thébaudeau
The
SDHCI_QUIRK_NO_MULTIBLOCKquirk was used as a workaround for the ENGcm07207 erratum. However, it caused excruciatingly slow SD transfers (300 kB/s on average), and this erratum actually does not imply that multiple-block transfers are not supported, so this was overkill.The suggested workaround for this erratum is to set SYSCTL.RSTA, but the simple DAT line software reset (which resets the DMA circuit among others) triggered by
sdhci_finish_data()in case of errors seems to be sufficient. Indeed, generating errors in a controlled manner on i.MX25 using the FEVT register right in the middle of read data transfers without this quirk shows that nothing is written to the buffer by the eSDHC past CMD12, and no extra Auto CMD12 is sent with AC12EN set, so the data transfers on AHB are properly aborted. For write data transfers, neither extra data nor extra Auto CMD12 is sent, as expected. Moreover, after intensive stress tests on i.MX25, removingSDHCI_QUIRK_NO_MULTIBLOCKseems to be safe.Quoted 6 lines
Quoted 1 line
Signed-off-by Benoît Thébaudeau <[email protected]>Aside from one comment below...
I would expect to see Acks from other sdhci-esdhc-imx users. Nevertheless, for sdhci:
Quoted 6 lines
Yes, I would also appreciate some ack/tested by, from the corresponding sdhci variant users of this series.
However, to allow it to get some results from linux-next, I have queued up this series for next (amending $subject patch according to the comment from Adrian). Thanks!
Please note that I had superseded this series with a v2 following Adrian's comment.
Quoted 6 lines
Quoted 4 lines
Acked-by Adrian Hunter <[email protected]>[...]
Best regards,
Benoît
Re: [PATCH 4/4] mmc: sdhci-esdhc-imx: Remove the ENGcm07207 workaround
In reply to Benoît ThébaudeauQuoted 17 lines · Benoît Thébaudeau
The
SDHCI_QUIRK_NO_MULTIBLOCKquirk was used as a workaround for the ENGcm07207 erratum. However, it caused excruciatingly slow SD transfers (300 kB/s on average), and this erratum actually does not imply that multiple-block transfers are not supported, so this was overkill.The suggested workaround for this erratum is to set SYSCTL.RSTA, but the simple DAT line software reset (which resets the DMA circuit among others) triggered by
sdhci_finish_data()in case of errors seems to be sufficient. Indeed, generating errors in a controlled manner on i.MX25 using the FEVT register right in the middle of read data transfers without this quirk shows that nothing is written to the buffer by the eSDHC past CMD12, and no extra Auto CMD12 is sent with AC12EN set, so the data transfers on AHB are properly aborted. For write data transfers, neither extra data nor extra Auto CMD12 is sent, as expected. Moreover, after intensive stress tests on i.MX25, removingSDHCI_QUIRK_NO_MULTIBLOCKseems to be safe.
Quoted 1 line
Signed-off-by Benoît Thébaudeau <[email protected]>
Quoted 9 lines · Benoît Thébaudeau
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of
mmc_test:11: Badly aligned write 12: Badly aligned read 13: Badly aligned multi-block write 14: Badly aligned multi-block readSigned-off-by Benoît Thébaudeau <[email protected]>
Is this the right description? I thought that SDHCI_QUIRK_32BIT_DMA_ADDR was for devices that cannot address high memory above 0xffffffff, rather than requiring a specific alignment.
If this is indeed an address range problem rather than an alignment problem, are you sure it is the SD controller that is wrong here, rather than having a 64-bit DMA capable SDHCI connected to a 32-bit parent bus? In the latter case, the dma-ranges property in the parent bus should limit the addressing, not the device.
Arnd
Quoted 15 lines · Arnd Bergmann
Quoted 9 lines · Benoît Thébaudeau
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of
mmc_test:11: Badly aligned write 12: Badly aligned read 13: Badly aligned multi-block write 14: Badly aligned multi-block readSigned-off-by Benoît Thébaudeau <[email protected]>Is this the right description? I thought that
SDHCI_QUIRK_32BIT_DMA_ADDRwas for devices that cannot address high memory above0xffffffff, rather than requiring a specific alignment.
Quoted 5 lines
If this is indeed an address range problem rather than an alignment problem, are you sure it is the SD controller that is wrong here, rather than having a 64-bit DMA capable SDHCI connected to a 32-bit parent bus? In the latter case, the dma-ranges property in the parent bus should limit the addressing, not the device.
No, this is the right description. This quirk really is about alignment, and not about address range. See:
drivers/mmc/host/sdhci.h:
Quoted 1 line
---
/* Controller can only DMA from 32-bit aligned addresses */
#define SDHCI_QUIRK_32BIT_DMA_ADDR (1<<7)
<---
drivers/mmc/host/sdhci.c @ sdhci_prepare_data():
Quoted 1 line
---
offset_mask = 0;
[...]
if (host->quirks & SDHCI_QUIRK_32BIT_DMA_ADDR)
offset_mask = 3;
[...]
if (sg->offset & offset_mask) {
DBG("Reverting to PIO because of bad alignment\n");
host->flags &= ~SDHCI_REQ_USE_DMA;
break;
}<---
Benoît
Quoted 31 lines · Benoît Thébaudeau
Quoted 15 lines · Arnd Bergmann
Quoted 9 lines · Benoît Thébaudeau
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of
mmc_test:11: Badly aligned write 12: Badly aligned read 13: Badly aligned multi-block write 14: Badly aligned multi-block readSigned-off-by Benoît Thébaudeau <[email protected]>Is this the right description? I thought that
SDHCI_QUIRK_32BIT_DMA_ADDRwas for devices that cannot address high memory above0xffffffff, rather than requiring a specific alignment.Quoted 5 lines
If this is indeed an address range problem rather than an alignment problem, are you sure it is the SD controller that is wrong here, rather than having a 64-bit DMA capable SDHCI connected to a 32-bit parent bus? In the latter case, the dma-ranges property in the parent bus should limit the addressing, not the device.
No, this is the right description. This quirk really is about alignment, and not about address range. See:
drivers/mmc/host/sdhci.h:Quoted 1 line
---
/* Controller can only DMA from 32-bit aligned addresses */
#defineSDHCI_QUIRK_32BIT_DMA_ADDR(1<<7)
<---
Quoted 6 lines
drivers/mmc/host/sdhci.c@sdhci_prepare_data():Quoted 1 line
---
offset_mask = 0; [...] if (host->quirks & SDHCI_QUIRK_32BIT_DMA_ADDR) offset_mask = 3;
Ok, thanks for the clarification. I guess I should have checked this myself first.
Arnd
Quoted 9 lines · Benoît Thébaudeau
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of
mmc_test:
11: Badly aligned write
12: Badly aligned read
13: Badly aligned multi-block write
14: Badly aligned multi-block readSigned-off-by Benoît Thébaudeau <[email protected]>
I would expect to see Acks from other sdhci-esdhc users. Nevertheless for sdhci:
Quoted 17 lines
--- drivers/mmc/host/sdhci-esdhc.h | 1 + 1 file changed, 1 insertion(+)drivers/mmc/host/sdhci-esdhc.h +1
-0@@ -19,6 +19,7 @@*/ #define ESDHC_DEFAULT_QUIRKS (SDHCI_QUIRK_FORCE_BLK_SZ_2048 | \ + SDHCI_QUIRK_32BIT_DMA_ADDR | \ SDHCI_QUIRK_NO_BUSY_IRQ | \ SDHCI_QUIRK_DATA_TIMEOUT_USES_SDCLK | \ SDHCI_QUIRK_PIO_NEEDS_DELAY | \
Quoted 9 lines · Benoît Thébaudeau
The eSDHC can only DMA from 32-bit-aligned addresses.
This fixes the following test cases of
mmc_test:11: Badly aligned write 12: Badly aligned read 13: Badly aligned multi-block write 14: Badly aligned multi-block readSigned-off-by Benoît Thébaudeau <[email protected]>