[U-Boot] [RFC] ARM: rmobile: create DT memory nodes for R8A7795 3.0 and newer
Uses an SMC call to the ATF to determine the memory configuration, then creates appropriate DT memory nodes.
Signed-off-by: Ulrich Hecht ulrich.hecht+renesas@gmail.com --- This is an attempt to avoid having to have a ton of different device trees for Renesas R8A7795 H3 SoCs from revision 3.0, which come in a number of different memory configurations.
The specific configuration is statically baked into the ARM Trusted Firmware binary, so this solution adds an SMC "SiP Service Call" to the ATF returning the memory configuration, and then has u-boot build appropriate memory nodes.
This has not been tested in the flesh successfully, as I only have a revision 1.0 board. I'd like to get some feedback on whether this approach is sound, though. Thanks.
Depends on "[RFC ATF] Add SMCCC_RENESAS_MEMCONF SMC call".
CU Uli
board/renesas/salvator-x/salvator-x.c | 51 +++++++++++++++++++++++++++++++++++ configs/r8a7795_salvator-x_defconfig | 2 ++ 2 files changed, 53 insertions(+)
diff --git a/board/renesas/salvator-x/salvator-x.c b/board/renesas/salvator-x/salvator-x.c index 651877c..307cb64 100644 --- a/board/renesas/salvator-x/salvator-x.c +++ b/board/renesas/salvator-x/salvator-x.c @@ -24,6 +24,7 @@ #include <asm/arch/sh_sdhi.h> #include <i2c.h> #include <mmc.h> +#include <linux/arm-smccc.h>
DECLARE_GLOBAL_DATA_PTR;
@@ -136,3 +137,53 @@ void reset_cpu(ulong addr) writel(RST_CODE, RST_CA57RESCNT); #endif } + +#define ARM_SMCCC_RENESAS_MEMCONF 0x82000000UL +int ft_board_setup(void *blob, bd_t *bd) +{ + if (rmobile_get_cpu_type() == RMOBILE_CPU_TYPE_R8A7795 && + rmobile_get_cpu_rev_integer() >= 3) { + u64 base[CONFIG_NR_DRAM_BANKS]; + u64 size[CONFIG_NR_DRAM_BANKS]; + struct arm_smccc_res res; + + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, + 0, 0, 0, 0, 0, 0, 0, &res); + + switch (res.a0) { + case 1: + base[0] = 0x048000000ULL; + size[0] = 0x038000000ULL; + base[1] = 0x500000000ULL; + size[1] = 0x040000000ULL; + base[2] = 0x600000000ULL; + size[2] = 0x040000000ULL; + base[3] = 0x700000000ULL; + size[3] = 0x040000000ULL; + fdt_fixup_memory_banks(blob, base, size, 4); + break; + case 2: + base[0] = 0x048000000ULL; + size[0] = 0x078000000ULL; + base[1] = 0x500000000ULL; + size[1] = 0x080000000ULL; + fdt_fixup_memory_banks(blob, base, size, 2); + break; + case 3: + base[0] = 0x048000000ULL; + size[0] = 0x078000000ULL; + base[1] = 0x500000000ULL; + size[1] = 0x080000000ULL; + base[2] = 0x600000000ULL; + size[2] = 0x080000000ULL; + base[3] = 0x700000000ULL; + size[3] = 0x080000000ULL; + fdt_fixup_memory_banks(blob, base, size, 4); + break; + default: + break; + } + } + + return 0; +} diff --git a/configs/r8a7795_salvator-x_defconfig b/configs/r8a7795_salvator-x_defconfig index fdfa41c..cd2bb59 100644 --- a/configs/r8a7795_salvator-x_defconfig +++ b/configs/r8a7795_salvator-x_defconfig @@ -61,3 +61,5 @@ CONFIG_USB_EHCI_GENERIC=y CONFIG_USB_STORAGE=y CONFIG_OF_LIBFDT_OVERLAY=y CONFIG_SMBIOS_MANUFACTURER="" +CONFIG_OF_BOARD_SETUP=y +CONFIG_ARM_SMCCC=y
Returns the memory configuration for Renesas R8A7795 (R-Car H3) SoCs, revision 3.0 and up.
Signed-off-by: Ulrich Hecht ulrich.hecht+renesas@gmail.com --- See "[RFC] ARM: rmobile: create DT memory nodes for R8A7795 3.0 and newer" for an explanation of this.
CU Uli
include/services/arm_arch_svc.h | 1 + services/arm_arch_svc/arm_arch_svc_setup.c | 14 ++++++++++++++ 2 files changed, 15 insertions(+)
diff --git a/include/services/arm_arch_svc.h b/include/services/arm_arch_svc.h index 2961601..9132336 100644 --- a/include/services/arm_arch_svc.h +++ b/include/services/arm_arch_svc.h @@ -10,5 +10,6 @@ #define SMCCC_VERSION U(0x80000000) #define SMCCC_ARCH_FEATURES U(0x80000001) #define SMCCC_ARCH_WORKAROUND_1 U(0x80008000) +#define SMCCC_RENESAS_MEMCONF U(0x82000000)
#endif /* __ARM_ARCH_SVC_H__ */ diff --git a/services/arm_arch_svc/arm_arch_svc_setup.c b/services/arm_arch_svc/arm_arch_svc_setup.c index eedac86..f836ae8 100644 --- a/services/arm_arch_svc/arm_arch_svc_setup.c +++ b/services/arm_arch_svc/arm_arch_svc_setup.c @@ -56,6 +56,20 @@ uintptr_t arm_arch_svc_smc_handler(uint32_t smc_fid, */ SMC_RET0(handle); #endif + case SMCCC_RENESAS_MEMCONF: +#if (RCAR_DRAM_LPDDR4_MEMCONF == 0) + /* 4GB(1GBx4) */ + SMC_RET1(handle, 1); +#elif (RCAR_DRAM_LPDDR4_MEMCONF == 1) && (RCAR_DRAM_CHANNEL == 5) && \ + (RCAR_DRAM_SPLIT == 2) + /* 4GB(2GBx2 2ch split) */ + SMC_RET1(handle, 2); +#elif (RCAR_DRAM_LPDDR4_MEMCONF == 1) && (RCAR_DRAM_CHANNEL == 15) + /* 8GB(2GBx4: default) */ + SMC_RET1(handle, 3); +#else + SMC_RET1(handle, 0); +#endif /* RCAR_DRAM_LPDDR4_MEMCONF == 0 */ default: WARN("Unimplemented Arm Architecture Service Call: 0x%x \n", smc_fid);
On 06/15/2018 11:40 AM, Ulrich Hecht wrote:
Uses an SMC call to the ATF to determine the memory configuration, then creates appropriate DT memory nodes.
Signed-off-by: Ulrich Hecht ulrich.hecht+renesas@gmail.com
This is an attempt to avoid having to have a ton of different device trees for Renesas R8A7795 H3 SoCs from revision 3.0, which come in a number of different memory configurations.
The specific configuration is statically baked into the ARM Trusted Firmware binary, so this solution adds an SMC "SiP Service Call" to the ATF returning the memory configuration, and then has u-boot build appropriate memory nodes.
This has not been tested in the flesh successfully, as I only have a revision 1.0 board. I'd like to get some feedback on whether this approach is sound, though. Thanks.
Depends on "[RFC ATF] Add SMCCC_RENESAS_MEMCONF SMC call".
CU Uli
board/renesas/salvator-x/salvator-x.c | 51 +++++++++++++++++++++++++++++++++++ configs/r8a7795_salvator-x_defconfig | 2 ++ 2 files changed, 53 insertions(+)
diff --git a/board/renesas/salvator-x/salvator-x.c b/board/renesas/salvator-x/salvator-x.c index 651877c..307cb64 100644 --- a/board/renesas/salvator-x/salvator-x.c +++ b/board/renesas/salvator-x/salvator-x.c @@ -24,6 +24,7 @@ #include <asm/arch/sh_sdhi.h> #include <i2c.h> #include <mmc.h> +#include <linux/arm-smccc.h>
DECLARE_GLOBAL_DATA_PTR;
@@ -136,3 +137,53 @@ void reset_cpu(ulong addr) writel(RST_CODE, RST_CA57RESCNT); #endif }
+#define ARM_SMCCC_RENESAS_MEMCONF 0x82000000UL +int ft_board_setup(void *blob, bd_t *bd) +{
- if (rmobile_get_cpu_type() == RMOBILE_CPU_TYPE_R8A7795 &&
rmobile_get_cpu_rev_integer() >= 3) {u64 base[CONFIG_NR_DRAM_BANKS];u64 size[CONFIG_NR_DRAM_BANKS];struct arm_smccc_res res;arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);
Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;
Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
default:break;}- }
- return 0;
+} diff --git a/configs/r8a7795_salvator-x_defconfig b/configs/r8a7795_salvator-x_defconfig index fdfa41c..cd2bb59 100644 --- a/configs/r8a7795_salvator-x_defconfig +++ b/configs/r8a7795_salvator-x_defconfig @@ -61,3 +61,5 @@ CONFIG_USB_EHCI_GENERIC=y CONFIG_USB_STORAGE=y CONFIG_OF_LIBFDT_OVERLAY=y CONFIG_SMBIOS_MANUFACTURER="" +CONFIG_OF_BOARD_SETUP=y +CONFIG_ARM_SMCCC=y
Update the defconfig with make savedefconfig .
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut marek.vasut@gmail.com wrote:
arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
CU Uli
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut marek.vasut@gmail.com wrote:
arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.
Which means the board has a memory size of 0 and fails to boot ?
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut marek.vasut@gmail.com wrote:
arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.
Which means the board has a memory size of 0 and fails to boot ?
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
Hi Marek,
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut marek.vasut@gmail.com
wrote:
arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.>
Which means the board has a memory size of 0 and fails to boot ?
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
Hi Marek,
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut marek.vasut@gmail.com
wrote:
arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF,0, 0, 0, 0, 0, 0, 0, &res);Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.>
Which means the board has a memory size of 0 and fails to boot ?
switch (res.a0) {case 1:base[0] = 0x048000000ULL;size[0] = 0x038000000ULL;base[1] = 0x500000000ULL;size[1] = 0x040000000ULL;base[2] = 0x600000000ULL;size[2] = 0x040000000ULL;base[3] = 0x700000000ULL;size[3] = 0x040000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;case 2:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 2);break;case 3:base[0] = 0x048000000ULL;size[0] = 0x078000000ULL;base[1] = 0x500000000ULL;size[1] = 0x080000000ULL;base[2] = 0x600000000ULL;size[2] = 0x080000000ULL;base[3] = 0x700000000ULL;size[3] = 0x080000000ULL;fdt_fixup_memory_banks(blob, base, size, 4);break;Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
Hi Marek,
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut wrote:
> + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, > + 0, 0, 0, 0, 0, 0, 0, &res);
Will this call work on platforms without patched ATF ? (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.
Which means the board has a memory size of 0 and fails to boot ?
> + switch (res.a0) { > + case 1: > + base[0] = 0x048000000ULL; > + size[0] = 0x038000000ULL; > + base[1] = 0x500000000ULL; > + size[1] = 0x040000000ULL; > + base[2] = 0x600000000ULL; > + size[2] = 0x040000000ULL; > + base[3] = 0x700000000ULL; > + size[3] = 0x040000000ULL; > + fdt_fixup_memory_banks(blob, base, size, 4); > + break; > + case 2: > + base[0] = 0x048000000ULL; > + size[0] = 0x078000000ULL; > + base[1] = 0x500000000ULL; > + size[1] = 0x080000000ULL; > + fdt_fixup_memory_banks(blob, base, size, 2); > + break; > + case 3: > + base[0] = 0x048000000ULL; > + size[0] = 0x078000000ULL; > + base[1] = 0x500000000ULL; > + size[1] = 0x080000000ULL; > + base[2] = 0x600000000ULL; > + size[2] = 0x080000000ULL; > + base[3] = 0x700000000ULL; > + size[3] = 0x080000000ULL; > + fdt_fixup_memory_banks(blob, base, size, 4); > + break;
Obvious design question is -- since you're adding new SMC call anyway, can't the call just return the memory layout table itself, so that it won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
Hi Marek,
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote:
On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut wrote: >> + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, >> + 0, 0, 0, 0, 0, 0, 0, &res); > > Will this call work on platforms without patched ATF ? > (I think not, don't you need to handle return value?)
I have not actually tested that, but if I understand the ATF code correctly, unimplemented calls return SMC_UNK (0xffffffff), which should be handled by the default case (NOP) below.
Which means the board has a memory size of 0 and fails to boot ?
>> + switch (res.a0) { >> + case 1: >> + base[0] = 0x048000000ULL; >> + size[0] = 0x038000000ULL; >> + base[1] = 0x500000000ULL; >> + size[1] = 0x040000000ULL; >> + base[2] = 0x600000000ULL; >> + size[2] = 0x040000000ULL; >> + base[3] = 0x700000000ULL; >> + size[3] = 0x040000000ULL; >> + fdt_fixup_memory_banks(blob, base, size, 4); >> + break; >> + case 2: >> + base[0] = 0x048000000ULL; >> + size[0] = 0x078000000ULL; >> + base[1] = 0x500000000ULL; >> + size[1] = 0x080000000ULL; >> + fdt_fixup_memory_banks(blob, base, size, 2); >> + break; >> + case 3: >> + base[0] = 0x048000000ULL; >> + size[0] = 0x078000000ULL; >> + base[1] = 0x500000000ULL; >> + size[1] = 0x080000000ULL; >> + base[2] = 0x600000000ULL; >> + size[2] = 0x080000000ULL; >> + base[3] = 0x700000000ULL; >> + size[3] = 0x080000000ULL; >> + fdt_fixup_memory_banks(blob, base, size, 4); >> + break; > > Obvious design question is -- since you're adding new SMC call anyway, > can't the call just return the memory layout table itself, so that it > won't be duplicated both in U-Boot and ATF ?
My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
Hi Marek,
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote:
On 06/15/2018 12:37 PM, Ulrich Hecht wrote: > On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut wrote: >>> + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, >>> + 0, 0, 0, 0, 0, 0, 0, &res); >> >> Will this call work on platforms without patched ATF ? >> (I think not, don't you need to handle return value?) > > I have not actually tested that, but if I understand the ATF code > correctly, unimplemented calls return > SMC_UNK (0xffffffff), which should be handled by the default case > (NOP) > below.
Which means the board has a memory size of 0 and fails to boot ?
>>> + switch (res.a0) { >>> + case 1: >>> + base[0] = 0x048000000ULL; >>> + size[0] = 0x038000000ULL; >>> + base[1] = 0x500000000ULL; >>> + size[1] = 0x040000000ULL; >>> + base[2] = 0x600000000ULL; >>> + size[2] = 0x040000000ULL; >>> + base[3] = 0x700000000ULL; >>> + size[3] = 0x040000000ULL; >>> + fdt_fixup_memory_banks(blob, base, size, 4); >>> + break; >>> + case 2: >>> + base[0] = 0x048000000ULL; >>> + size[0] = 0x078000000ULL; >>> + base[1] = 0x500000000ULL; >>> + size[1] = 0x080000000ULL; >>> + fdt_fixup_memory_banks(blob, base, size, 2); >>> + break; >>> + case 3: >>> + base[0] = 0x048000000ULL; >>> + size[0] = 0x078000000ULL; >>> + base[1] = 0x500000000ULL; >>> + size[1] = 0x080000000ULL; >>> + base[2] = 0x600000000ULL; >>> + size[2] = 0x080000000ULL; >>> + base[3] = 0x700000000ULL; >>> + size[3] = 0x080000000ULL; >>> + fdt_fixup_memory_banks(blob, base, size, 4); >>> + break; >> >> Obvious design question is -- since you're adding new SMC call >> anyway, >> can't the call just return the memory layout table itself, so that >> it >> won't be duplicated both in U-Boot and ATF ? > > My gut feeling was to go with the smallest interface possible.
But this doesn't scale. The API here uses some ad-hoc constants to identify memory layout tables which have to be encoded both in ATF and U-Boot, both of which must be kept in sync.
The ATF already has those memory layout tables, it's only a matter of passing them to U-Boot. If you do just that, the ad-hoc constants and encoding of tables into U-Boot goes away and in fact simplifies the design.
Yet, I have to wonder if ATF doesn't already contain some sort of standard SMC call to get memory topology. It surprises me that it wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Hi Laurent,
On Tue, Jun 19, 2018 at 11:15 AM, Laurent Pinchart laurent.pinchart@ideasonboard.com wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
Hi Marek,
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote: > On 06/15/2018 12:37 PM, Ulrich Hecht wrote: >> On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut wrote: >>>> + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, >>>> + 0, 0, 0, 0, 0, 0, 0, &res); >>> >>> Will this call work on platforms without patched ATF ? >>> (I think not, don't you need to handle return value?) >> >> I have not actually tested that, but if I understand the ATF code >> correctly, unimplemented calls return >> SMC_UNK (0xffffffff), which should be handled by the default case >> (NOP) >> below. > > Which means the board has a memory size of 0 and fails to boot ? > >>>> + switch (res.a0) { >>>> + case 1: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x038000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x040000000ULL; >>>> + base[2] = 0x600000000ULL; >>>> + size[2] = 0x040000000ULL; >>>> + base[3] = 0x700000000ULL; >>>> + size[3] = 0x040000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 4); >>>> + break; >>>> + case 2: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x078000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x080000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 2); >>>> + break; >>>> + case 3: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x078000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x080000000ULL; >>>> + base[2] = 0x600000000ULL; >>>> + size[2] = 0x080000000ULL; >>>> + base[3] = 0x700000000ULL; >>>> + size[3] = 0x080000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 4); >>>> + break; >>> >>> Obvious design question is -- since you're adding new SMC call >>> anyway, >>> can't the call just return the memory layout table itself, so that >>> it >>> won't be duplicated both in U-Boot and ATF ? >> >> My gut feeling was to go with the smallest interface possible. > > But this doesn't scale. The API here uses some ad-hoc constants to > identify memory layout tables which have to be encoded both in ATF and > U-Boot, both of which must be kept in sync. > > The ATF already has those memory layout tables, it's only a matter of > passing them to U-Boot. If you do just that, the ad-hoc constants and > encoding of tables into U-Boot goes away and in fact simplifies the > design. > > Yet, I have to wonder if ATF doesn't already contain some sort of > standard SMC call to get memory topology. It surprises me that it > wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Thanks for the clarification. I wonder how much of the split between ATF and "the rest" can be adjusted. It smells like just a software architecture policy, my gut feeling is that LIFEC can be programmed to adjust the assignment between secure side and "regular". At least it can do so for a wide range of bus masters. However if this is the case for FCNL remains to be seen.
As a side note, for FCNL I personally would prefer a more dynamic solution where we manage physically contiguous ranges from Linux during run time instead of relying of statically setup windows.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
-- Regards,
Laurent Pinchart
Hi Magnus,
On Tuesday, 19 June 2018 08:43:31 EEST Magnus Damm wrote:
On Tue, Jun 19, 2018 at 11:15 AM, Laurent Pinchart wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote: > On 06/15/2018 01:43 PM, Marek Vasut wrote:
[snip]
>> Yet, I have to wonder if ATF doesn't already contain some sort of >> standard SMC call to get memory topology. It surprises me that it >> wouldn't. > > In fact, Laurent (CCed) was solving some similar issue with lossy > decomp and I think this involved some passing of memory layout > information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Thanks for the clarification. I wonder how much of the split between ATF and "the rest" can be adjusted. It smells like just a software architecture policy, my gut feeling is that LIFEC can be programmed to adjust the assignment between secure side and "regular". At least it can do so for a wide range of bus masters. However if this is the case for FCNL remains to be seen.
As a side note, for FCNL I personally would prefer a more dynamic solution where we manage physically contiguous ranges from Linux during run time instead of relying of statically setup windows.
So would I, but whether we can be allowed to access those registers from Linux isn't my decision :-/
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Hi Laurent,
On Tue, Jun 19, 2018 at 4:15 AM Laurent Pinchart laurent.pinchart@ideasonboard.com wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote:
On 06/15/2018 01:43 PM, Marek Vasut wrote: > On 06/15/2018 12:37 PM, Ulrich Hecht wrote: >> On Fri, Jun 15, 2018 at 12:09 PM, Marek Vasut wrote: >>>> + arm_smccc_smc(ARM_SMCCC_RENESAS_MEMCONF, >>>> + 0, 0, 0, 0, 0, 0, 0, &res); >>> >>> Will this call work on platforms without patched ATF ? >>> (I think not, don't you need to handle return value?) >> >> I have not actually tested that, but if I understand the ATF code >> correctly, unimplemented calls return >> SMC_UNK (0xffffffff), which should be handled by the default case >> (NOP) >> below. > > Which means the board has a memory size of 0 and fails to boot ? > >>>> + switch (res.a0) { >>>> + case 1: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x038000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x040000000ULL; >>>> + base[2] = 0x600000000ULL; >>>> + size[2] = 0x040000000ULL; >>>> + base[3] = 0x700000000ULL; >>>> + size[3] = 0x040000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 4); >>>> + break; >>>> + case 2: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x078000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x080000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 2); >>>> + break; >>>> + case 3: >>>> + base[0] = 0x048000000ULL; >>>> + size[0] = 0x078000000ULL; >>>> + base[1] = 0x500000000ULL; >>>> + size[1] = 0x080000000ULL; >>>> + base[2] = 0x600000000ULL; >>>> + size[2] = 0x080000000ULL; >>>> + base[3] = 0x700000000ULL; >>>> + size[3] = 0x080000000ULL; >>>> + fdt_fixup_memory_banks(blob, base, size, 4); >>>> + break; >>> >>> Obvious design question is -- since you're adding new SMC call >>> anyway, >>> can't the call just return the memory layout table itself, so that >>> it >>> won't be duplicated both in U-Boot and ATF ? >> >> My gut feeling was to go with the smallest interface possible. > > But this doesn't scale. The API here uses some ad-hoc constants to > identify memory layout tables which have to be encoded both in ATF and > U-Boot, both of which must be kept in sync. > > The ATF already has those memory layout tables, it's only a matter of > passing them to U-Boot. If you do just that, the ad-hoc constants and > encoding of tables into U-Boot goes away and in fact simplifies the > design. > > Yet, I have to wonder if ATF doesn't already contain some sort of > standard SMC call to get memory topology. It surprises me that it > wouldn't.
In fact, Laurent (CCed) was solving some similar issue with lossy decomp and I think this involved some passing of memory layout information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Pass a "table"? Or an FDT containing the /memory nodes? The latter allows for easier future extension.
Gr{oetje,eeting}s,
Geert
Hi Geert,
On Tuesday, 19 June 2018 09:58:59 EEST Geert Uytterhoeven wrote:
On Tue, Jun 19, 2018 at 4:15 AM Laurent Pinchart wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote:
On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote: > On 06/15/2018 01:43 PM, Marek Vasut wrote:
[snip]
>>>> Obvious design question is -- since you're adding new SMC call >>>> anyway, can't the call just return the memory layout table >>>> itself, so that it won't be duplicated both in U-Boot and ATF ? >>> >>> My gut feeling was to go with the smallest interface possible. >> >> But this doesn't scale. The API here uses some ad-hoc constants to >> identify memory layout tables which have to be encoded both in ATF >> and U-Boot, both of which must be kept in sync. >> >> The ATF already has those memory layout tables, it's only a matter >> of passing them to U-Boot. If you do just that, the ad-hoc >> constants and encoding of tables into U-Boot goes away and in fact >> simplifies the design. >> >> Yet, I have to wonder if ATF doesn't already contain some sort of >> standard SMC call to get memory topology. It surprises me that it >> wouldn't. > > In fact, Laurent (CCed) was solving some similar issue with lossy > decomp and I think this involved some passing of memory layout > information from ATF to U-Boot too, or am I mistaken ?
That's correct, ATF stores information about the memory layout at a fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Pass a "table"? Or an FDT containing the /memory nodes? The latter allows for easier future extension.
ATF passes a table to U-Boot, and U-Boot then updates the FDT accordingly before starting Linux.
Hi Laurent,
On Tue, Jun 19, 2018 at 9:10 AM Laurent Pinchart laurent.pinchart@ideasonboard.com wrote:
On Tuesday, 19 June 2018 09:58:59 EEST Geert Uytterhoeven wrote:
On Tue, Jun 19, 2018 at 4:15 AM Laurent Pinchart wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote: > On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote: >> On 06/15/2018 01:43 PM, Marek Vasut wrote:
[snip]
>>>>> Obvious design question is -- since you're adding new SMC call >>>>> anyway, can't the call just return the memory layout table >>>>> itself, so that it won't be duplicated both in U-Boot and ATF ? >>>> >>>> My gut feeling was to go with the smallest interface possible. >>> >>> But this doesn't scale. The API here uses some ad-hoc constants to >>> identify memory layout tables which have to be encoded both in ATF >>> and U-Boot, both of which must be kept in sync. >>> >>> The ATF already has those memory layout tables, it's only a matter >>> of passing them to U-Boot. If you do just that, the ad-hoc >>> constants and encoding of tables into U-Boot goes away and in fact >>> simplifies the design. >>> >>> Yet, I have to wonder if ATF doesn't already contain some sort of >>> standard SMC call to get memory topology. It surprises me that it >>> wouldn't. >> >> In fact, Laurent (CCed) was solving some similar issue with lossy >> decomp and I think this involved some passing of memory layout >> information from ATF to U-Boot too, or am I mistaken ? > > That's correct, ATF stores information about the memory layout at a > fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Pass a "table"? Or an FDT containing the /memory nodes? The latter allows for easier future extension.
ATF passes a table to U-Boot, and U-Boot then updates the FDT accordingly before starting Linux.
I know it passes a table. A table is very easy to parse, but complicates future extension.
Gr{oetje,eeting}s,
Geert
On 06/19/2018 09:11 AM, Laurent Pinchart wrote:
Hi Geert,
On Tuesday, 19 June 2018 09:58:59 EEST Geert Uytterhoeven wrote:
On Tue, Jun 19, 2018 at 4:15 AM Laurent Pinchart wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote:
On 06/16/2018 01:21 AM, Laurent Pinchart wrote: > On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote: >> On 06/15/2018 01:43 PM, Marek Vasut wrote:
[snip]
>>>>> Obvious design question is -- since you're adding new SMC call >>>>> anyway, can't the call just return the memory layout table >>>>> itself, so that it won't be duplicated both in U-Boot and ATF ? >>>> >>>> My gut feeling was to go with the smallest interface possible. >>> >>> But this doesn't scale. The API here uses some ad-hoc constants to >>> identify memory layout tables which have to be encoded both in ATF >>> and U-Boot, both of which must be kept in sync. >>> >>> The ATF already has those memory layout tables, it's only a matter >>> of passing them to U-Boot. If you do just that, the ad-hoc >>> constants and encoding of tables into U-Boot goes away and in fact >>> simplifies the design. >>> >>> Yet, I have to wonder if ATF doesn't already contain some sort of >>> standard SMC call to get memory topology. It surprises me that it >>> wouldn't. >> >> In fact, Laurent (CCed) was solving some similar issue with lossy >> decomp and I think this involved some passing of memory layout >> information from ATF to U-Boot too, or am I mistaken ? > > That's correct, ATF stores information about the memory layout at a > fixed address in system memory, and U-Boot can read it.
Well, that sounds good ! Maybe we can avoid adding SMC call altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Pass a "table"? Or an FDT containing the /memory nodes? The latter allows for easier future extension.
ATF passes a table to U-Boot, and U-Boot then updates the FDT accordingly before starting Linux.
At this point, U-Boot does not parse any such table. It could be some fork does that, but mainline does not. But that code could (and probably should) be added.
I don't think ATF has FDT support, but I might be wrong. And if ATF can pass DT fragment or some simple DT, that'd be neat.
Hi Marek, All,
I am looking forward for the solution/approach you are going to reach in this thread. Currently, integrating Renesas Yocto v3.7.0 release [1], I can see that the increase of SoC/SiP versions supported in this release increases the number of ATF/U-boot build variants dramatically.
Renesas U-boot starts to rely not only on defconfig, but also on build-time flags passed via Yocto recipes (see usage of get_uboot_config_opt in [2,3]). This doesn't look right (I don't see U-Boot KCFLAGS being adjusted too often in mainline). Latest Renesas ATF also adds a bunch of new make flags, all being related to SiP DRAM configuration, which means it's very easy to pick the wrong setting at build-time and then deal with obscure issues afterwards (hopefully the target just doesn't boot) :(
FWIW, with regard to FDT support in ATF, it seems to be there starting with commit [4] and there are already users/callers of libfdt APIs in the tree [5]. Something I am wondering about is why not getting the memory configuration in ATF at runtime? Haven't DRAM vendors thought this information (number/size of DDR banks) would be useful and should be exposed by HW and accessible by SW? Currently Renesas tells their customers that the correct SiP-specific DDR configuration has to be hardcoded during build process. If anybody has an explanation why runtime DDR detection is not possible in ATF, could you please share?
Best regards, Eugeniu.
[1] https://github.com/renesas-rcar/meta-renesas/releases/tag/Renesas-Yocto-v3.7... [2] https://github.com/renesas-rcar/meta-renesas/blob/Renesas-Yocto-v3.7.0/meta-... [3] https://github.com/renesas-rcar/meta-renesas/blob/Renesas-Yocto-v3.7.0/meta-... [4] https://github.com/ARM-software/arm-trusted-firmware/commit/91176bc6363d ("Import libfdt v1.4.1") [5] git grep -El "fdt_|fdtw_" -- ":!*lib/libfdt/*" "*.c" common/fdt_wrappers.c plat/arm/common/arm_dyn_cfg_helpers.c plat/arm/css/sgi/sgi_image_load.c plat/qemu/dt.c plat/qemu/qemu_bl2_setup.c On Wed, Jun 20, 2018 at 6:58 AM Marek Vasut marek.vasut@gmail.com wrote:
On 06/19/2018 09:11 AM, Laurent Pinchart wrote:
Hi Geert,
On Tuesday, 19 June 2018 09:58:59 EEST Geert Uytterhoeven wrote:
On Tue, Jun 19, 2018 at 4:15 AM Laurent Pinchart wrote:
On Sunday, 17 June 2018 03:08:02 EEST Marek Vasut wrote:
On 06/16/2018 05:44 PM, Laurent Pinchart wrote:
On Saturday, 16 June 2018 02:42:30 EEST Marek Vasut wrote: > On 06/16/2018 01:21 AM, Laurent Pinchart wrote: >> On Friday, 15 June 2018 15:00:31 EEST Marek Vasut wrote: >>> On 06/15/2018 01:43 PM, Marek Vasut wrote:
[snip]
>>>>>> Obvious design question is -- since you're adding new SMC call >>>>>> anyway, can't the call just return the memory layout table >>>>>> itself, so that it won't be duplicated both in U-Boot and ATF ? >>>>> >>>>> My gut feeling was to go with the smallest interface possible. >>>> >>>> But this doesn't scale. The API here uses some ad-hoc constants to >>>> identify memory layout tables which have to be encoded both in ATF >>>> and U-Boot, both of which must be kept in sync. >>>> >>>> The ATF already has those memory layout tables, it's only a matter >>>> of passing them to U-Boot. If you do just that, the ad-hoc >>>> constants and encoding of tables into U-Boot goes away and in fact >>>> simplifies the design. >>>> >>>> Yet, I have to wonder if ATF doesn't already contain some sort of >>>> standard SMC call to get memory topology. It surprises me that it >>>> wouldn't. >>> >>> In fact, Laurent (CCed) was solving some similar issue with lossy >>> decomp and I think this involved some passing of memory layout >>> information from ATF to U-Boot too, or am I mistaken ? >> >> That's correct, ATF stores information about the memory layout at a >> fixed address in system memory, and U-Boot can read it. > > Well, that sounds good ! Maybe we can avoid adding SMC call > altogether then? :)
I'd prefer that, yes.
Let's not duplicate the mechanism used to pass FCNL information from ATF to U- Boot, but instead create a data table format that can store all the information we need, in an easily extensible way.
To see how the mechanism is implemented for FCNL, search for 47FD7000 in the Renesas ATF sources (git://github.com/renesas-rcar/arm-trusted-firmware.git).
For everyone involved, can you explain what FCNL is ? ;-)
FCNL is Frame Compression Near Lossless. It's a way to reduce memory bandwidth by transparent compression and decompression of video frames. Compression is handled by an IP core called FCP, and decompression is handled by the DRAM controller. ATF programs the DRAM controller with ranges of memory addresses that will be dynamically decompressed. The registers containing those ranges are accessible in secure mode only, so neither U-Boot nor Linux can read them. That's why ATF has to pass the information to U-Boot, in order to add the ranges as reserved memory in DT.
Any yes, I agree this sounds good. I had a discussion on the U-Boot IRC about passing the memory configuration around and the result is basically the same -- pass a table from ATF to U-Boot. If there's already something, great.
Pass a "table"? Or an FDT containing the /memory nodes? The latter allows for easier future extension.
ATF passes a table to U-Boot, and U-Boot then updates the FDT accordingly before starting Linux.
At this point, U-Boot does not parse any such table. It could be some fork does that, but mainline does not. But that code could (and probably should) be added.
I don't think ATF has FDT support, but I might be wrong. And if ATF can pass DT fragment or some simple DT, that'd be neat.
-- Best regards, Marek Vasut _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot
participants (6)
-
Eugeniu Rosca -
Geert Uytterhoeven -
Laurent Pinchart -
Magnus Damm -
Marek Vasut -
Ulrich Hecht