PM / devfreq: Add initial imx support
Subsystems (5)
v2 cover PM / devfreq: Add initial imx support
This adds devfreq support for imx8mm, covering dynamic scaling of internal NOC and DDR Controller
Scaling for simple busses (NIC/NOC) is done through the clk framework while DRAM scaling is performed in firmware with an "imx-ddrc" wrapper for devfreq.
Changes since v1:
* bindings: Stop using "contains" for "compatible"
* bindings: Set "additionalProperties: false" and document missing stuff.
* Remove (c) from NXP copyright notice
* Fix various checkpatch issues
* Remove unused dram_alt_root clk from imx-ddrc
Link to v1: https://patchwork.kernel.org/project/linux-arm-kernel/list/?series=158773
Changes since RFC v3:
* Implement passive support and set NOC's parent to DDRC
* Drop scaling AHB/AXI for now (NOC/DDRC use most power anyway)
* Stop relying on clk_min_rate
* Split into two devreq drivers (and bindings) because the ddrc is
really a distinct piece of hardware.
* Perform DRAM frequency inside devfreq instead of clk, mostly due to
objections to earlier RFCs for imx8m-dram-clk.
* Fetch info about dram clk parents from firmware instead of
hardcoding in driver. This can more easily support additional rates.
* Link: https://patchwork.kernel.org/cover/11056779/
* Link: https://patchwork.kernel.org/patch/11049429/
Scaling buses can cause problems for devices with realtime bandwith requirements such as display, the intention is to use the interconnect framework to make DEV_PM_QOS_MIN_FREQUENCY to devfreq. There are separate patches for that:
* https://patchwork.kernel.org/cover/11104055/
* https://patchwork.kernel.org/cover/11078671/
Leonard Crestez (7):
clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
dt-bindings: devfreq: Add bindings for generic imx buses PM / devfreq: Add generic imx bus driver
dt-bindings: devfreq: Add bindings for imx ddr controller PM / devfreq: Add dynamic scaling for imx ddr controller PM / devfreq: imx-ddrc: Measure bandwidth with perf
arm64: dts: imx8mm: Add devfreq nodes
.../devicetree/bindings/devfreq/imx-ddrc.yaml | 60 ++ .../devicetree/bindings/devfreq/imx.yaml | 68 +++ arch/arm64/boot/dts/freescale/imx8mm.dtsi | 53 +- drivers/clk/imx/clk-imx8mm.c | 6 +- drivers/clk/imx/clk-imx8mn.c | 6 +- drivers/clk/imx/clk-imx8mq.c | 7 +- drivers/devfreq/Kconfig | 12 + drivers/devfreq/Makefile | 1 + drivers/devfreq/imx-ddrc.c | 514 ++++++++++++++++++ drivers/devfreq/imx-devfreq.c | 148 +++++ 10 files changed, 867 insertions(+), 8 deletions(-) create mode 100644 Documentation/devicetree/bindings/devfreq/imx-ddrc.yaml create mode 100644 Documentation/devicetree/bindings/devfreq/imx.yaml create mode 100644 drivers/devfreq/imx-ddrc.c create mode 100644 drivers/devfreq/imx-devfreq.c
Signature · 1 line
2.17.1
PATCH v2 1/7 clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
Dram frequency changes required modifying these clocks outside the control of clk framework. Mark them as CLK_GET_RATE_NOCACHE so that rates are always read back from registers.
--- drivers/clk/imx/clk-imx8mm.c | 6 ++++-- drivers/clk/imx/clk-imx8mn.c | 6 ++++-- drivers/clk/imx/clk-imx8mq.c | 7 ++++--- 3 files changed, 12 insertions(+), 7 deletions(-)
drivers/clk/imx/clk-imx8mm.c +4 -2
@@ -526,12 +526,14 @@ static int imx8mm_clocks_probe(struct platform_device *pdev)/* IPG */ clks[IMX8MM_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MM_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 );/* IP */ - clks[IMX8MM_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 ); - clks[IMX8MM_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 ); + clks[IMX8MM_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MM_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE ); clks[IMX8MM_CLK_VPU_G1 ] =imx8m_clk_composite ("vpu_g1" , imx8mm_vpu_g1_sels, base +0xa100 ); clks[IMX8MM_CLK_VPU_G2 ] =imx8m_clk_composite ("vpu_g2" , imx8mm_vpu_g2_sels, base +0xa180 ); clks[IMX8MM_CLK_DISP_DTRC ] =imx8m_clk_composite ("disp_dtrc" , imx8mm_disp_dtrc_sels, base +0xa200 ); clks[IMX8MM_CLK_DISP_DC8000 ] =imx8m_clk_composite ("disp_dc8000" , imx8mm_disp_dc8000_sels, base +0xa280 ); clks[IMX8MM_CLK_PCIE1_CTRL ] =imx8m_clk_composite ("pcie1_ctrl" , imx8mm_pcie1_ctrl_sels, base +0xa300 );
drivers/clk/imx/clk-imx8mn.c +4 -2
@@ -529,12 +529,14 @@ static int imx8mn_clocks_probe(struct platform_device *pdev) clks[IMX8MN_CLK_AHB ] =imx8m_clk_composite_critical ("ahb" , imx8mn_ahb_sels, base +0x9000 ); clks[IMX8MN_CLK_AUDIO_AHB ] =imx8m_clk_composite ("audio_ahb" , imx8mn_audio_ahb_sels, base +0x9100 ); clks[IMX8MN_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MN_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 ); clks[IMX8MN_CLK_DRAM_CORE ] =imx_clk_mux2_flags ("dram_core_clk" , base +0x9800 ,24 ,1 , imx8mn_dram_core_sels,ARRAY_SIZE (imx8mn_dram_core_sels),CLK_IS_CRITICAL ); - clks[IMX8MN_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mn_dram_alt_sels, base +0xa000 ); - clks[IMX8MN_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mn_dram_apb_sels, base +0xa080 ); + clks[IMX8MN_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mn_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MN_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mn_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE ); clks[IMX8MN_CLK_DISP_PIXEL ] =imx8m_clk_composite ("disp_pixel" , imx8mn_disp_pixel_sels, base +0xa500 ); clks[IMX8MN_CLK_SAI2 ] =imx8m_clk_composite ("sai2" , imx8mn_sai2_sels, base +0xa600 ); clks[IMX8MN_CLK_SAI3 ] =imx8m_clk_composite ("sai3" , imx8mn_sai3_sels, base +0xa680 ); clks[IMX8MN_CLK_SAI5 ] =imx8m_clk_composite ("sai5" , imx8mn_sai5_sels, base +0xa780 ); clks[IMX8MN_CLK_SAI6 ] =imx8m_clk_composite ("sai6" , imx8mn_sai6_sels, base +0xa800 );
drivers/clk/imx/clk-imx8mq.c +4 -3
@@ -436,13 +436,14 @@ static int imx8mq_clocks_probe(struct platform_device *pdev) clks[IMX8MQ_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MQ_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 );/* IP */ clks[IMX8MQ_CLK_DRAM_CORE ] =imx_clk_mux2_flags ("dram_core_clk" , base +0x9800 ,24 ,1 , imx8mq_dram_core_sels,ARRAY_SIZE (imx8mq_dram_core_sels),CLK_IS_CRITICAL ); - - clks[IMX8MQ_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mq_dram_alt_sels, base +0xa000 ); - clks[IMX8MQ_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mq_dram_apb_sels, base +0xa080 ); + clks[IMX8MQ_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mq_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MQ_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mq_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE ); clks[IMX8MQ_CLK_VPU_G1 ] =imx8m_clk_composite ("vpu_g1" , imx8mq_vpu_g1_sels, base +0xa100 ); clks[IMX8MQ_CLK_VPU_G2 ] =imx8m_clk_composite ("vpu_g2" , imx8mq_vpu_g2_sels, base +0xa180 ); clks[IMX8MQ_CLK_DISP_DTRC ] =imx8m_clk_composite ("disp_dtrc" , imx8mq_disp_dtrc_sels, base +0xa200 ); clks[IMX8MQ_CLK_DISP_DC8000 ] =imx8m_clk_composite ("disp_dc8000" , imx8mq_disp_dc8000_sels, base +0xa280 ); clks[IMX8MQ_CLK_PCIE1_CTRL ] =imx8m_clk_composite ("pcie1_ctrl" , imx8mq_pcie1_ctrl_sels, base +0xa300 );
Signature · 1 line
2.17.1
PATCH v2 2/7 dt-bindings: devfreq: Add bindings for generic imx buses
In reply to Leonard CrestezAdd initial dt bindings for the interconnects inside i.MX chips. Multiple external IPs are involved but SOC integration means the software controllable interfaces are very similar.
This is initially only for imx8mm but add an "fsl,imx8m-nic" fallback similar to exynos-bus.
--- .../devicetree/bindings/devfreq/imx.yaml | 68 +++++++++++++++++++ 1 file changed, 68 insertions(+) create mode 100644 Documentation/devicetree/bindings/devfreq/imx.yaml
Documentation/devicetree/bindings/devfreq/imx.yaml new file +68 -0
@@ -0,0 +1,68 @@ +# SPDX-License-Identifier: GPL-2.0 +%YAML 1.2 +--- +$id: http://devicetree.org/schemas/devfreq/imx.yaml# +$schema: http://devicetree.org/meta-schemas/core.yaml# + +title: Generic i.MX bus frequency device + +maintainers: + - Leonard Crestez <[email protected]> + +description: | + The i.MX SoC family has multiple buses for which clock frequency (and sometimes + voltage) can be adjusted. + + Some of those buses expose register areas mentioned in the memory maps as GPV + ("Global Programmers View") but not all. Access to this area might be denied for + normal world. + + The buses are based on externally licensed IPs such as ARM NIC-301 and Arteris + FlexNOC but DT bindings are specific to the integration of these bus + interconnect IPs into imx SOCs. + +properties: + compatible: + oneOf: + - items: + - enum: + - fsl,imx8mn-nic + - fsl,imx8mm-nic + - fsl,imx8mq-nic + - const: fsl,imx8m-nic + - items: + - enum: + - fsl,imx8mn-noc + - fsl,imx8mm-noc + - fsl,imx8mq-noc + - const: fsl,imx8m-noc + + reg: + maxItems: 1 + + clocks: + maxItems: 1 + + operating-points-v2: true + + devfreq: + description: | + Phandle to another devfreq device to match OPPs with by using the + passive governor. + $ref: "/schemas/types.yaml#/definitions/phandle" + +required: + - compatible + - clocks + +additionalProperties: false + +examples: + - | + #include <dt-bindings/clock/imx8mm-clock.h> + noc: noc@32700000 { + compatible = "fsl,imx8mm-noc", "fsl,imx8m-noc"; + reg = <0x32700000 0x100000>; + clocks = <&clk IMX8MM_CLK_NOC>; + operating-points-v2 = <&noc_opp_table>; + };
Signature · 1 line
2.17.1
PATCH v2 3/7 PM / devfreq: Add generic imx bus driver
In reply to Leonard CrestezAdd initial support for dynamic frequency switching on pieces of the imx interconnect fabric.
All this driver actually does is set a clk rate based on an opp table.
No attempt is made to map registers or anything clever.
--- drivers/devfreq/Kconfig | 12 +++ drivers/devfreq/Makefile | 1 + drivers/devfreq/imx-devfreq.c | 148 ++++++++++++++++++++++++++++++++++ 3 files changed, 161 insertions(+) create mode 100644 drivers/devfreq/imx-devfreq.c
drivers/devfreq/Kconfig +12 -0
@@ -90,10 +90,22 @@ config ARM_EXYNOS_BUS_DEVFREQ Each memory bus group could contain many memoby bus block. It reads PPMU counters of memory controllers by using DEVFREQ-event device and adjusts the operating frequencies and voltages with OPP support. This does not yet operate with optimal voltages. +config ARM_IMX_DEVFREQ + tristate "i.MX DEVFREQ Driver" + depends on ARCH_MXC || COMPILE_TEST + select DEVFREQ_GOV_PASSIVE + select DEVFREQ_GOV_SIMPLE_ONDEMAND + select DEVFREQ_GOV_USERSPACE + select PM_OPP + help + This adds the DEVFREQ driver for the i.MX family of SoCs. + It allows adjusting frequencies for DDRC (DDR Controller) and various + NICs and NOCs which form the SOC interconnect fabric + config ARM_TEGRA_DEVFREQ tristate "NVIDIA Tegra30/114/124/210 DEVFREQ Driver" depends on ARCH_TEGRA_3x_SOC || ARCH_TEGRA_114_SOC || \ ARCH_TEGRA_132_SOC || ARCH_TEGRA_124_SOC || \ ARCH_TEGRA_210_SOC || \
drivers/devfreq/Makefile +1 -0
@@ -7,10 +7,11 @@ obj-$(CONFIG_DEVFREQ_GOV_POWERSAVE) += governor_powersave.o obj-$(CONFIG_DEVFREQ_GOV_USERSPACE) += governor_userspace.o obj-$(CONFIG_DEVFREQ_GOV_PASSIVE) += governor_passive.o # DEVFREQ Drivers obj-$(CONFIG_ARM_EXYNOS_BUS_DEVFREQ) += exynos-bus.o +obj-$(CONFIG_ARM_IMX_DEVFREQ) += imx-devfreq.o obj-$(CONFIG_ARM_RK3399_DMC_DEVFREQ) += rk3399_dmc.o obj-$(CONFIG_ARM_TEGRA_DEVFREQ) += tegra30-devfreq.o obj-$(CONFIG_ARM_TEGRA20_DEVFREQ) += tegra20-devfreq.o # DEVFREQ Event Drivers
drivers/devfreq/imx-devfreq.c new file +148 -0
148 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH v2 4/7 dt-bindings: devfreq: Add bindings for imx ddr controller
In reply to Leonard CrestezAdd devicetree bindings for the i.MX DDR Controller on imx8m series chips. It supports dynamic frequency switching between multiple data rates and this is exposed to Linux via the devfreq subsystem.
--- .../devicetree/bindings/devfreq/imx-ddrc.yaml | 60 +++++++++++++++++++ 1 file changed, 60 insertions(+) create mode 100644 Documentation/devicetree/bindings/devfreq/imx-ddrc.yaml
Documentation/devicetree/bindings/devfreq/imx-ddrc.yaml new file +60 -0
@@ -0,0 +1,60 @@ +# SPDX-License-Identifier: GPL-2.0 +%YAML 1.2 +--- +$id: http://devicetree.org/schemas/devfreq/imx-devfreq.yaml# +$schema: http://devicetree.org/meta-schemas/core.yaml# + +title: i.MX DDR Controller + +maintainers: + - Leonard Crestez <[email protected]> + +properties: + compatible: + items: + - enum: + - fsl,imx8mn-ddrc + - fsl,imx8mm-ddrc + - fsl,imx8mq-ddrc + - const: fsl,imx8m-ddrc + + reg: + maxItems: 1 + + clocks: + maxItems: 4 + + clock-names: + items: + - const: dram_core + - const: dram_pll + - const: dram_alt + - const: dram_apb + + operating-points-v2: true + + devfreq-events: + description: Phandle of PMU node + $ref: "/schemas/types.yaml#/definitions/phandle" + +required: + - reg + - compatible + - clocks + - clock-names + +additionalProperties: false + +examples: + - | + #include <dt-bindings/clock/imx8mm-clock.h> + ddrc: dram-controller@3d400000 { + compatible = "fsl,imx8mm-ddrc", "fsl,imx8m-ddrc"; + reg = <0x3d400000 0x400000>; + clock-names = "dram_core", "dram_pll", "dram_alt", "dram_apb"; + clocks = <&clk IMX8MM_CLK_DRAM_CORE>, + <&clk IMX8MM_DRAM_PLL>, + <&clk IMX8MM_CLK_DRAM_ALT>, + <&clk IMX8MM_CLK_DRAM_APB>; + operating-points-v2 = <&ddrc_opp_table>; + };
Signature · 1 line
2.17.1
PATCH v2 5/7 PM / devfreq: Add dynamic scaling for imx ddr controller
In reply to Leonard CrestezAdd driver for dynamic scaling the DDR Controller on imx8m chips. Actual frequency switching is implemented in TF-A, this driver just wraps the SMC calls and updates the clk tree.
--- drivers/devfreq/Makefile | 2 +- drivers/devfreq/imx-ddrc.c | 374 +++++++++++++++++++++++++++++++++++++ 2 files changed, 375 insertions(+), 1 deletion(-) create mode 100644 drivers/devfreq/imx-ddrc.c
drivers/devfreq/Makefile +1 -1
@@ -7,11 +7,11 @@ obj-$(CONFIG_DEVFREQ_GOV_POWERSAVE) += governor_powersave.o obj-$(CONFIG_DEVFREQ_GOV_USERSPACE) += governor_userspace.o obj-$(CONFIG_DEVFREQ_GOV_PASSIVE) += governor_passive.o # DEVFREQ Drivers obj-$(CONFIG_ARM_EXYNOS_BUS_DEVFREQ) += exynos-bus.o -obj-$(CONFIG_ARM_IMX_DEVFREQ) += imx-devfreq.o +obj-$(CONFIG_ARM_IMX_DEVFREQ) += imx-devfreq.o imx-ddrc.o obj-$(CONFIG_ARM_RK3399_DMC_DEVFREQ) += rk3399_dmc.o obj-$(CONFIG_ARM_TEGRA_DEVFREQ) += tegra30-devfreq.o obj-$(CONFIG_ARM_TEGRA20_DEVFREQ) += tegra20-devfreq.o # DEVFREQ Event Drivers
drivers/devfreq/imx-ddrc.c new file +374 -0
374 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH v2 6/7 PM / devfreq: imx-ddrc: Measure bandwidth with perf
In reply to Leonard CrestezThe imx8m ddrc has a performance monitoring block attached which can be used to measure bandwidth usage and automatically adjust frequency.
There is already a perf driver for that block so instead of implementing a devfreq events driver use the in-kernel perf API to implement get_dev_status directly.
--- drivers/devfreq/imx-ddrc.c | 146 ++++++++++++++++++++++++++++++++++++- 1 file changed, 143 insertions(+), 3 deletions(-)
drivers/devfreq/imx-ddrc.c +143 -3
@@ -11,10 +11,13 @@#include <linux/pm_opp.h> #include <linux/clk.h> #include <linux/clk-provider.h> #include <linux/arm-smccc.h> +#include <asm/perf_event.h> +#include <linux/perf_event.h> +#define IMX_SIP_DDR_DVFS 0xc2000004/* Values starting from 0 switch to specific frequency */ #define IMX_SIP_DDR_FREQ_SET_HIGH 0x00
@@ -80,10 +83,22 @@ struct imx_ddrc {struct clk *dram_alt;struct clk *dram_apb;int freq_count;struct imx_ddrc_freq freq_table[IMX_DDRC_MAX_FREQ_COUNT ]; + +/* For measuring load with perf events: */ +struct platform_device *pmu_pdev; +struct pmu *pmu; + +struct perf_event_attr rd_event_attr; +struct perf_event_attr wr_event_attr; +struct perf_event *rd_event; +struct perf_event *wr_event; + +u64 last_rd_val, last_rd_ena, last_rd_run; +u64 last_wr_val, last_wr_ena, last_wr_run; };static struct imx_ddrc_freq *imx_ddrc_find_freq (struct imx_ddrc *priv,unsigned long rate) {
@@ -228,19 +243,127 @@ static int imx_ddrc_get_cur_freq(struct device *dev, unsigned long *freq)return 0 ; }static int imx_ddrc_get_dev_status (struct device *dev, -struct devfreq_dev_status *stat) +struct devfreq_dev_status *stat) {struct imx_ddrc *priv =dev_get_drvdata (dev); - stat->busy_time =0 ; - stat->total_time =0 ; stat->current_frequency =clk_get_rate (priv->dram_core); +if (priv->rd_event && priv->wr_event) { +u64 rd_delta, rd_val, rd_ena, rd_run; +u64 wr_delta, wr_val, wr_ena, wr_run; + + rd_val =perf_event_read_value (priv->rd_event, + &rd_ena, &rd_run); + wr_val =perf_event_read_value (priv->wr_event, + &wr_ena, &wr_run); + + rd_delta = (rd_val - priv->last_rd_val) * + (rd_ena - priv->last_rd_ena) / + (rd_run - priv->last_rd_run); + priv->last_rd_val = rd_val; + priv->last_rd_ena = rd_ena; + priv->last_rd_run = rd_run; + wr_delta = (wr_val - priv->last_wr_val) * + (wr_ena - priv->last_wr_ena) / + (wr_run - priv->last_wr_run); + priv->last_wr_val = wr_val; + priv->last_wr_ena = wr_ena; + priv->last_wr_run = wr_run; + +/* magic numbers, possibly wrong */ + stat->busy_time =4 * (rd_delta + wr_delta); + stat->total_time = stat->current_frequency; + }else { + stat->busy_time =0 ; + stat->total_time =0 ; + } + +return 0 ; +} + +static int imx_ddrc_perf_disable (struct imx_ddrc *priv) +{ +/* release and set to NULL */ +if (!IS_ERR_OR_NULL (priv->rd_event)) +perf_event_release_kernel (priv->rd_event); +if (!IS_ERR_OR_NULL (priv->wr_event)) +perf_event_release_kernel (priv->wr_event); + priv->rd_event =NULL ; + priv->wr_event =NULL ; + +return 0 ; +} + +static int imx_ddrc_perf_enable (struct imx_ddrc *priv) +{ +int ret; + + priv->rd_event_attr.size =sizeof (priv->rd_event_attr); + priv->rd_event_attr.type = priv->pmu->type; + priv->rd_event_attr.config =0x2a ; + + priv->rd_event =perf_event_create_kernel_counter ( + &priv->rd_event_attr,0 ,NULL ,NULL ,NULL ); +if (IS_ERR (priv->rd_event)) { + ret =PTR_ERR (priv->rd_event); + goto err; + } + + priv->wr_event_attr.size =sizeof (priv->wr_event_attr); + priv->wr_event_attr.type = priv->pmu->type; + priv->wr_event_attr.config =0x2b ; + + priv->wr_event =perf_event_create_kernel_counter ( + &priv->wr_event_attr,0 ,NULL ,NULL ,NULL ); +if (IS_ERR (priv->wr_event)) { + ret =PTR_ERR (priv->wr_event); + goto err; + } +return 0 ; + +err: +imx_ddrc_perf_disable (priv); +return ret; +} + +static int imx_ddrc_init_events (struct device *dev, +struct device_node *events_node) +{ +struct imx_ddrc *priv =dev_get_drvdata (dev); +struct device_driver *driver; + +/* +* We need pmu->type for perf_event_attr but there is no API for +* mapping device_node to pmu. Fetch private data for imx-ddr-pmu and +* cast that to a struct pmu instead. +*/ + priv->pmu_pdev =of_find_device_by_node (events_node); +if (!priv->pmu_pdev) +return -EPROBE_DEFER ; + driver = priv->pmu_pdev->dev.driver; +if (!driver) +return -EPROBE_DEFER ; +if (strcmp (driver->name,"imx-ddr-pmu" )) { +dev_warn (dev,"devfreq-events node %pOF has unexpected driver %s\n" , + events_node, driver->name); +return -ENODEV ; + } + + priv->pmu =platform_get_drvdata (priv->pmu_pdev); +if (!priv->pmu) { +dev_err (dev,"devfreq-events device missing private data\n" ); +return -EINVAL ; + } + +dev_dbg (dev,"events from pmu %s\n" , priv->pmu->name); + +return imx_ddrc_perf_enable (priv); }static int imx_ddrc_init_freq_info (struct device *dev) {struct imx_ddrc *priv =dev_get_drvdata (dev);
@@ -288,10 +411,15 @@ static int imx_ddrc_init_freq_info(struct device *dev)return 0 ; }static void imx_ddrc_exit (struct device *dev) { +struct imx_ddrc *priv =dev_get_drvdata (dev); + +imx_ddrc_perf_disable (priv); +platform_device_put (priv->pmu_pdev); +dev_pm_opp_of_remove_table (dev); }static int imx_ddrc_probe (struct platform_device *pdev) {
@@ -337,10 +465,20 @@ static int imx_ddrc_probe(struct platform_device *pdev) priv->profile.get_dev_status = imx_ddrc_get_dev_status; priv->profile.exit = imx_ddrc_exit; priv->profile.get_cur_freq = imx_ddrc_get_cur_freq; priv->profile.initial_freq =clk_get_rate (priv->dram_core); +/* Handle devfreq-events */ + events_node =of_parse_phandle (dev->of_node,"devfreq-events" ,0 ); +if (events_node) { + ret =imx_ddrc_init_events (dev, events_node); +of_node_put (events_node); +if (ret) + goto err; + gov =DEVFREQ_GOV_SIMPLE_ONDEMAND ; + } + priv->devfreq =devm_devfreq_add_device (dev, &priv->profile, gov,NULL );if (IS_ERR (priv->devfreq)) { ret =PTR_ERR (priv->devfreq);dev_err (dev,"failed to add devfreq device: %d\n" , ret);
@@ -348,10 +486,12 @@ static int imx_ddrc_probe(struct platform_device *pdev) }return 0 ; err: +imx_ddrc_perf_disable (priv); +platform_device_put (priv->pmu_pdev);dev_pm_opp_of_remove_table (dev);return ret; }static const struct of_device_id imx_ddrc_of_match[] = {
Signature · 1 line
2.17.1
PATCH v2 7/7 arm64: dts: imx8mm: Add devfreq nodes
In reply to Leonard CrestezAdd initial support for bus scaling on imx8m, starting with noc and ddrc because they're the biggest power hogs.
Add a devfreq-event link to the PMU in order to support on-demand scaling of ddrc based on measured dram bandwith usage. Make ddrc a parent of the NOC because all traffic to ddrc goes through
Support for proactive scaling via interconnect and support for scaling additional NICs will come later. The high-performance bus masters which need that (display, vpu, gpu) are not yet enabled in upstream anyway.
--- arch/arm64/boot/dts/freescale/imx8mm.dtsi | 53 ++++++++++++++++++++++- 1 file changed, 52 insertions(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mm.dtsi +52 -1
@@ -139,10 +139,38 @@ clock-latency-ns = <150000>; opp-suspend; }; }; + ddrc_opp_table: ddrc-opp-table { + compatible = "operating-points-v2"; + + opp-25M { + opp-hz = /bits/ 64 <25000000>; + }; + opp-100M { + opp-hz = /bits/ 64 <100000000>; + }; + opp-750M { + opp-hz = /bits/ 64 <750000000>; + }; + }; + + noc_opp_table: noc-opp-table { + compatible = "operating-points-v2"; + + opp-150M { + opp-hz = /bits/ 64 <150000000>; + }; + opp-375M { + opp-hz = /bits/ 64 <375000000>; + }; + opp-750M { + opp-hz = /bits/ 64 <750000000>; + }; + }; + memory@40000000 { device_type = "memory"; reg = <0x0 0x40000000 0 0x80000000>; };
@@ -773,10 +801,18 @@ status = "disabled"; }; }; + noc: noc@32700000 { + compatible = "fsl,imx8mm-noc", "fsl,imx8m-noc"; + reg = <0x32700000 0x100000>; + clocks = <&clk IMX8MM_CLK_NOC>; + devfreq = <&ddrc>; + operating-points-v2 = <&noc_opp_table>; + }; + aips4: bus@32c00000 { compatible = "fsl,aips-bus", "simple-bus"; #address-cells = <1>; #size-cells = <1>; ranges = <0x32c00000 0x32c00000 0x400000>;
@@ -857,11 +893,26 @@ #interrupt-cells = <3>; interrupt-controller; interrupts = <GIC_PPI 9 IRQ_TYPE_LEVEL_HIGH>; }; - ddr-pmu@3d800000 { + ddrc: dram-controller@3d400000 { + compatible = "fsl,imx8mm-ddrc", "fsl,imx8m-ddrc"; + reg = <0x3d400000 0x400000>; + clock-names = "dram_core", + "dram_pll", + "dram_alt", + "dram_apb"; + clocks = <&clk IMX8MM_CLK_DRAM_CORE>, + <&clk IMX8MM_DRAM_PLL>, + <&clk IMX8MM_CLK_DRAM_ALT>, + <&clk IMX8MM_CLK_DRAM_APB>; + devfreq-events = <&ddr_pmu>; + operating-points-v2 = <&ddrc_opp_table>; + }; + + ddr_pmu: ddr-pmu@3d800000 { compatible = "fsl,imx8mm-ddr-pmu", "fsl,imx8m-ddr-pmu"; reg = <0x3d800000 0x400000>; interrupt-parent = <&gic>; interrupts = <GIC_SPI 98 IRQ_TYPE_LEVEL_HIGH>; };
Signature · 1 line
2.17.1
Stephen Boyd
maintainerRe: [PATCH v2 1/7] clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
In reply to Leonard CrestezQuoting Leonard Crestez (2019-08-20 08:45:06)
Quoted 3 lines
Dram frequency changes required modifying these clocks outside the control of clk framework. Mark them as
CLK_GET_RATE_NOCACHEso that rates are always read back from registers.
Why can't we control the clks from the clk framework? Please add that information in the commit text here.
Quoted 14 lines
Signed-off-by Leonard Crestez <[email protected]>--- drivers/clk/imx/clk-imx8mm.c | 6 ++++-- drivers/clk/imx/clk-imx8mn.c | 6 ++++-- drivers/clk/imx/clk-imx8mq.c | 7 ++++--- 3 files changed, 12 insertions(+), 7 deletions(-)drivers/clk/imx/clk-imx8mm.c +0
-0@@ -526,12 +526,14 @@ static int imx8mm_clocks_probe(struct platform_device *pdev)/* IPG */
drivers/clk/imx/clk-imx8mm.c static int imx8mm_clocks_probe(struct platform_device *pdev) · line 534
Quoted 10 lines
drivers/clk/imx/clk-imx8mm.c +4
-2clks[IMX8MM_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MM_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 );/* IP */ - clks[IMX8MM_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 ); - clks[IMX8MM_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 ); + clks[IMX8MM_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MM_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE );
Also, add a comment to this effect about why it can't be done from the clk framework wherever the CLK_GET_RATE_NOCACHE flag is set. Basically this flag is a hack and is an example of something that we need to fix.
drivers/clk/imx/clk-imx8mm.c static int imx8mm_clocks_probe(struct platform_device *pdev) · line 539
Quoted 5 lines
drivers/clk/imx/clk-imx8mm.c +0
-0clks[IMX8MM_CLK_VPU_G1 ] =imx8m_clk_composite ("vpu_g1" , imx8mm_vpu_g1_sels, base +0xa100 ); clks[IMX8MM_CLK_VPU_G2 ] =imx8m_clk_composite ("vpu_g2" , imx8mm_vpu_g2_sels, base +0xa180 ); clks[IMX8MM_CLK_DISP_DTRC ] =imx8m_clk_composite ("disp_dtrc" , imx8mm_disp_dtrc_sels, base +0xa200 ); clks[IMX8MM_CLK_DISP_DC8000 ] =imx8m_clk_composite ("disp_dc8000" , imx8mm_disp_dc8000_sels, base +0xa280 ); clks[IMX8MM_CLK_PCIE1_CTRL ] =imx8m_clk_composite ("pcie1_ctrl" , imx8mm_pcie1_ctrl_sels, base +0xa300 );
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v2 1/7] clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
Quoted 7 lines · Stephen Boyd
Quoting Leonard Crestez (2019-08-20 08:45:06)
Quoted 3 lines
Dram frequency changes required modifying these clocks outside the control of clk framework. Mark them as
CLK_GET_RATE_NOCACHEso that rates are always read back from registers.Why can't we control the clks from the clk framework? Please add that information in the commit text here.
OK, I will update commit message and comments
These clocks are only modified for DRAM frequency switches during which DRAM is briefly inaccessible. The switch is performed with a SMC call to by TF-A which runs from a SRAM area. Upon returning to linux several clocks bits are modified and we need to update them.
For rate bits an easy solution is to just mark with CLK_GET_RATE_NOCACHE, muxes are handled explicitly.
Linux code performing the SMC call is also part of this series:
https://patchwork.kernel.org/patch/11104145/
Quoted 16 lines
Quoted 16 lines
Signed-off-by Leonard Crestez <[email protected]>--- drivers/clk/imx/clk-imx8mm.c | 6 ++++-- drivers/clk/imx/clk-imx8mn.c | 6 ++++-- drivers/clk/imx/clk-imx8mq.c | 7 ++++--- 3 files changed, 12 insertions(+), 7 deletions(-)drivers/clk/imx/clk-imx8mm.c +0
-0@@ -526,12 +526,14 @@ static int imx8mm_clocks_probe(struct platform_device *pdev)/* IPG */ clks[IMX8MM_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MM_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 );
Quoted 11 lines
drivers/clk/imx/clk-imx8mm.c static int imx8mm_clocks_probe(struct platform_device *pdev) · line 534
Quoted 7 lines
drivers/clk/imx/clk-imx8mm.c +4
-2/* IP */ - clks[IMX8MM_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 ); - clks[IMX8MM_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 ); + clks[IMX8MM_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MM_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE );Also, add a comment to this effect about why it can't be done from the clk framework wherever the
CLK_GET_RATE_NOCACHEflag is set. Basically this flag is a hack and is an example of something that we need to fix.
DRAM freq switch requires multiple clk changes to be performed atomically while DRAM itself is not accessible so it's not something to "fix".
--
Regards,
Leonard
Stephen Boyd
maintainerRe: [PATCH v2 1/7] clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
Quoting Leonard Crestez (2019-09-16 16:03:53)
Quoted 11 lines · Stephen Boyd
Quoted 7 lines
Quoting Leonard Crestez (2019-08-20 08:45:06)
Quoted 3 lines
Dram frequency changes required modifying these clocks outside the control of clk framework. Mark them as
CLK_GET_RATE_NOCACHEso that rates are always read back from registers.Why can't we control the clks from the clk framework? Please add that information in the commit text here.
OK, I will update commit message and comments
Quoted 7 lines
These clocks are only modified for DRAM frequency switches during which DRAM is briefly inaccessible. The switch is performed with a SMC call to by TF-A which runs from a SRAM area. Upon returning to linux several clocks bits are modified and we need to update them.
For rate bits an easy solution is to just mark with
CLK_GET_RATE_NOCACHE, muxes are handled explicitly.
Is there any reason to expose or control these clks from Linux then? It might be easier to just make any children clks of the DRAM frequency clk "root" clks and then ignore any frequency that they might have. Similarly, because the SMC call is used to change the frequency, it may be simpler to handle that completely outside of the clk framework (it may already be this way in this patch series but I haven't read everything here).
Quoted 33 lines
Linux code performing the SMC call is also part of this series:
https://patchwork.kernel.org/patch/11104145/Quoted 16 lines
Quoted 16 lines
Signed-off-by Leonard Crestez <[email protected]>--- drivers/clk/imx/clk-imx8mm.c | 6 ++++-- drivers/clk/imx/clk-imx8mn.c | 6 ++++-- drivers/clk/imx/clk-imx8mq.c | 7 ++++--- 3 files changed, 12 insertions(+), 7 deletions(-)drivers/clk/imx/clk-imx8mm.c +0
-0@@ -526,12 +526,14 @@ static int imx8mm_clocks_probe(struct platform_device *pdev)/* IPG */ clks[IMX8MM_CLK_IPG_ROOT ] =imx_clk_divider2 ("ipg_root" ,"ahb" , base +0x9080 ,0 ,1 ); clks[IMX8MM_CLK_IPG_AUDIO_ROOT ] =imx_clk_divider2 ("ipg_audio_root" ,"audio_ahb" , base +0x9180 ,0 ,1 );Quoted 11 lines
drivers/clk/imx/clk-imx8mm.c static int imx8mm_clocks_probe(struct platform_device *pdev) · line 534
Quoted 7 lines
drivers/clk/imx/clk-imx8mm.c +4
-2/* IP */ - clks[IMX8MM_CLK_DRAM_ALT ] =imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 ); - clks[IMX8MM_CLK_DRAM_APB ] =imx8m_clk_composite_critical ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 ); + clks[IMX8MM_CLK_DRAM_ALT ] =__imx8m_clk_composite ("dram_alt" , imx8mm_dram_alt_sels, base +0xa000 , +CLK_GET_RATE_NOCACHE ); + clks[IMX8MM_CLK_DRAM_APB ] =__imx8m_clk_composite ("dram_apb" , imx8mm_dram_apb_sels, base +0xa080 , +CLK_IS_CRITICAL |CLK_GET_RATE_NOCACHE );Also, add a comment to this effect about why it can't be done from the clk framework wherever the
CLK_GET_RATE_NOCACHEflag is set. Basically this flag is a hack and is an example of something that we need to fix.
Quoted 3 lines
DRAM freq switch requires multiple clk changes to be performed atomically while DRAM itself is not accessible so it's not something to "fix".
Ok. Fix may be the wrong word.
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v2 1/7] clk: imx8m: Set CLK_GET_RATE_NOCACHE on dram_alt/apb
Quoted 20 lines · Stephen Boyd
Quoting Leonard Crestez (2019-09-16 16:03:53)
Quoted 11 lines · Stephen Boyd
Quoted 7 lines
Quoting Leonard Crestez (2019-08-20 08:45:06)
Quoted 3 lines
Dram frequency changes required modifying these clocks outside the control of clk framework. Mark them as
CLK_GET_RATE_NOCACHEso that rates are always read back from registers.Why can't we control the clks from the clk framework? Please add that information in the commit text here.
OK, I will update commit message and comments
Quoted 7 lines
These clocks are only modified for DRAM frequency switches during which DRAM is briefly inaccessible. The switch is performed with a SMC call to by TF-A which runs from a SRAM area. Upon returning to linux several clocks bits are modified and we need to update them.
For rate bits an easy solution is to just mark with
CLK_GET_RATE_NOCACHE, muxes are handled explicitly.
Quoted 7 lines
Is there any reason to expose or control these clks from Linux then? It might be easier to just make any children clks of the DRAM frequency clk "root" clks and then ignore any frequency that they might have. Similarly, because the SMC call is used to change the frequency, it may be simpler to handle that completely outside of the clk framework (it may already be this way in this patch series but I haven't read everything here).
The dram alt/apb clocks are real imx8m composite clocks with the same HW implementation as used for peripherals. They also have mux parents which are under the control of the clock framework so the freq switching code takes care to properly enable the new parents before calling SMC.
See imx_ddrc_set_freq: https://patchwork.kernel.org/patch/11104145/
Removing dram alt/apb clocks from the tree would still require keeping possible parents enabled somehow. It wouldn't be simpler but a lot uglier.
--
Regards,
Leonard