drm/amd/powerplay: enable gpu_metrics export on legacy powerplay routines
Subsystems (2)
PATCH 1/9 drm/amd/powerplay: enable gpu_metrics export on legacy powerplay routines
Enable gpu_metrics support on legacy powerplay routines.
--- drivers/gpu/drm/amd/powerplay/amd_powerplay.c | 19 +++++++++++++++++++ drivers/gpu/drm/amd/powerplay/inc/hwmgr.h | 1 + 2 files changed, 20 insertions(+)
drivers/gpu/drm/amd/powerplay/amd_powerplay.c +19 -0
@@ -1598,6 +1598,24 @@ static int pp_set_xgmi_pstate(void *handle, uint32_t pstate)return 0 ; } +static ssize_t pp_get_gpu_metrics (void *handle,void **table) +{ +struct pp_hwmgr *hwmgr = handle; +ssize_t size; + +if (!hwmgr) +return -EINVAL ; + +if (!hwmgr->pm_en || !hwmgr->hwmgr_func->get_gpu_metrics) +return -EOPNOTSUPP ; + +mutex_lock (&hwmgr->smu_lock); + size = hwmgr->hwmgr_func->get_gpu_metrics (hwmgr, table); +mutex_unlock (&hwmgr->smu_lock); + +return size; +} +static const struct amd_pm_funcs pp_dpm_funcs = { .load_firmware = pp_dpm_load_fw, .wait_for_fw_loading_complete = pp_dpm_fw_loading_complete,
@@ -1658,4 +1676,5 @@ static const struct amd_pm_funcs pp_dpm_funcs = { .smu_i2c_bus_access = pp_smu_i2c_bus_access, .set_df_cstate = pp_set_df_cstate, .set_xgmi_pstate = pp_set_xgmi_pstate, + .get_gpu_metrics = pp_get_gpu_metrics, };
drivers/gpu/drm/amd/powerplay/inc/hwmgr.h +1 -0
@@ -359,6 +359,7 @@ struct pp_hwmgr_func {int (*set_xgmi_pstate)(struct pp_hwmgr *hwmgr,uint32_t pstate);int (*disable_power_features_for_compute_performance)(struct pp_hwmgr *hwmgr,bool disable); +ssize_t (*get_gpu_metrics)(struct pp_hwmgr *hwmgr,void **table); };struct pp_table_func {
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 2/9 drm/amd/powerplay: add Vega20 support for gpu metrics export
Add Vega20 gpu metrics export interface.
--- .../drm/amd/powerplay/hwmgr/vega20_hwmgr.c | 121 +++++++++++++++++- .../drm/amd/powerplay/hwmgr/vega20_hwmgr.h | 1 + 2 files changed, 115 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/powerplay/hwmgr/vega20_hwmgr.c +114 -7
@@ -55,6 +55,11 @@#define smnPCIE_LC_SPEED_CNTL 0x11140290#define smnPCIE_LC_LINK_WIDTH_CNTL 0x11140288 +#define LINK_WIDTH_MAX 6 +#define LINK_SPEED_MAX 3 +static int link_width[] = {0 ,1 ,2 ,4 ,8 ,12 ,16 }; +static int link_speed[] = {25 ,50 ,80 ,160 }; +static void vega20_set_default_registry_data (struct pp_hwmgr *hwmgr) { struct vega20_hwmgr *data =
@@ -3265,6 +3270,46 @@ static int vega20_set_ppfeature_status(struct pp_hwmgr *hwmgr, uint64_t new_ppfereturn 0 ; } +static int vega20_get_current_pcie_link_width_level (struct pp_hwmgr *hwmgr) +{ +struct amdgpu_device *adev = hwmgr->adev; + +return (RREG32_PCIE (smnPCIE_LC_LINK_WIDTH_CNTL) & +PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD_MASK ) + >>PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD__SHIFT ; +} + +static int vega20_get_current_pcie_link_width (struct pp_hwmgr *hwmgr) +{ +uint32_t width_level; + + width_level =vega20_get_current_pcie_link_width_level (hwmgr); +if (width_level >LINK_WIDTH_MAX ) + width_level =0 ; + +return link_width[width_level]; +} + +static int vega20_get_current_pcie_link_speed_level (struct pp_hwmgr *hwmgr) +{ +struct amdgpu_device *adev = hwmgr->adev; + +return (RREG32_PCIE (smnPCIE_LC_SPEED_CNTL) & +PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE_MASK ) + >>PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE__SHIFT ; +} + +static int vega20_get_current_pcie_link_speed (struct pp_hwmgr *hwmgr) +{ +uint32_t speed_level; + + speed_level =vega20_get_current_pcie_link_speed_level (hwmgr); +if (speed_level >LINK_SPEED_MAX ) + speed_level =0 ; + +return link_speed[speed_level]; +} +static int vega20_print_clock_levels (struct pp_hwmgr *hwmgr,enum pp_clock_type type,char *buf) {
@@ -3277,7 +3322,6 @@ static int vega20_print_clock_levels(struct pp_hwmgr *hwmgr,struct phm_ppt_v3_information *pptable_information = (struct phm_ppt_v3_information *)hwmgr->pptable;PPTable_t *pptable = (PPTable_t *)pptable_information->smc_pptable; -struct amdgpu_device *adev = hwmgr->adev;struct pp_clock_levels_with_latency clocks;struct vega20_single_dpm_table *fclk_dpm_table = &(data->dpm_table.fclk_table);
@@ -3371,12 +3415,10 @@ static int vega20_print_clock_levels(struct pp_hwmgr *hwmgr,break ;case PP_PCIE : - current_gen_speed = (RREG32_PCIE (smnPCIE_LC_SPEED_CNTL) & -PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE_MASK ) - >>PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE__SHIFT ; - current_lane_width = (RREG32_PCIE (smnPCIE_LC_LINK_WIDTH_CNTL) & -PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD_MASK ) - >>PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD__SHIFT ; + current_gen_speed = +vega20_get_current_pcie_link_speed_level (hwmgr); + current_lane_width = +vega20_get_current_pcie_link_width_level (hwmgr);for (i =0 ; i <NUM_LINK_LEVELS ; i++) {if (i ==1 && data->pcie_parameters_override) { gen_speed = data->pcie_gen_level1;
@@ -4218,6 +4260,70 @@ static int vega20_set_xgmi_pstate(struct pp_hwmgr *hwmgr,return ret; } +static void vega20_init_gpu_metrics_v1_0 (struct gpu_metrics_v1_0 *gpu_metrics) +{ +memset (gpu_metrics,0xFF ,sizeof (struct gpu_metrics_v1_0 )); + + gpu_metrics->common_header.structure_size = +sizeof (struct gpu_metrics_v1_0 ); + gpu_metrics->common_header.format_revision =1 ; + gpu_metrics->common_header.content_revision =0 ; +} + +static ssize_t vega20_get_gpu_metrics (struct pp_hwmgr *hwmgr, +void **table) +{ +struct vega20_hwmgr *data = + (struct vega20_hwmgr *)(hwmgr->backend); +struct gpu_metrics_v1_0 *gpu_metrics = + &data->gpu_metrics_table; +SmuMetrics_t metrics; +uint32_t fan_speed_rpm; +int ret; + + ret =vega20_get_metrics_table (hwmgr, &metrics); +if (ret) +return ret; + +vega20_init_gpu_metrics_v1_0 (gpu_metrics); + + gpu_metrics->temperature_edge = metrics.TemperatureEdge; + gpu_metrics->temperature_hotspot = metrics.TemperatureHotspot; + gpu_metrics->temperature_mem = metrics.TemperatureHBM; + gpu_metrics->temperature_vrgfx = metrics.TemperatureVrGfx; + gpu_metrics->temperature_vrsoc = metrics.TemperatureVrSoc; + gpu_metrics->temperature_vrmem = metrics.TemperatureVrMem0; + + gpu_metrics->average_gfx_activity = metrics.AverageGfxActivity; + gpu_metrics->average_umc_activity = metrics.AverageUclkActivity; + + gpu_metrics->average_socket_power = metrics.AverageSocketPower; + + gpu_metrics->average_gfxclk_frequency = metrics.AverageGfxclkFrequency; + gpu_metrics->average_socclk_frequency = metrics.AverageSocclkFrequency; + gpu_metrics->average_uclk_frequency = metrics.AverageUclkFrequency; + + gpu_metrics->current_gfxclk = metrics.CurrClock[PPCLK_GFXCLK ]; + gpu_metrics->current_socclk = metrics.CurrClock[PPCLK_SOCCLK ]; + gpu_metrics->current_uclk = metrics.CurrClock[PPCLK_UCLK ]; + gpu_metrics->current_vclk0 = metrics.CurrClock[PPCLK_VCLK ]; + gpu_metrics->current_dclk0 = metrics.CurrClock[PPCLK_DCLK ]; + + gpu_metrics->throttle_status = metrics.ThrottlerStatus; + +vega20_fan_ctrl_get_fan_speed_rpm (hwmgr, &fan_speed_rpm); + gpu_metrics->current_fan_speed = (uint16_t )fan_speed_rpm; + + gpu_metrics->pcie_link_width = +vega20_get_current_pcie_link_width (hwmgr); + gpu_metrics->pcie_link_speed = +vega20_get_current_pcie_link_speed (hwmgr); + + *table = (void *)gpu_metrics; + +return sizeof (struct gpu_metrics_v1_0 ); +} +static const struct pp_hwmgr_func vega20_hwmgr_funcs = {/* init/fini related */ .backend_init = vega20_hwmgr_backend_init,
@@ -4288,6 +4394,7 @@ static const struct pp_hwmgr_func vega20_hwmgr_funcs = { .smu_i2c_bus_access = vega20_smu_i2c_bus_access, .set_df_cstate = vega20_set_df_cstate, .set_xgmi_pstate = vega20_set_xgmi_pstate, + .get_gpu_metrics = vega20_get_gpu_metrics, };int vega20_hwmgr_init (struct pp_hwmgr *hwmgr)
drivers/gpu/drm/amd/powerplay/hwmgr/vega20_hwmgr.h +1 -0
@@ -527,6 +527,7 @@ struct vega20_hwmgr {unsigned long metrics_time;SmuMetrics_t metrics_table; +struct gpu_metrics_v1_0 gpu_metrics_table;bool pcie_parameters_override;uint32_t pcie_gen_level1;
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 3/9 drm/amd/powerplay: add Vega12 support for gpu metrics export
In reply to Evan QuanAdd Vega12 gpu metrics export interface.
--- .../drm/amd/powerplay/hwmgr/vega12_hwmgr.c | 109 ++++++++++++++++++ .../drm/amd/powerplay/hwmgr/vega12_hwmgr.h | 1 + 2 files changed, 110 insertions(+)
drivers/gpu/drm/amd/powerplay/hwmgr/vega12_hwmgr.c +109 -0
@@ -47,6 +47,13 @@#include "pp_thermal.h" #include "vega12_baco.h" +#define smnPCIE_LC_SPEED_CNTL 0x11140290 +#define smnPCIE_LC_LINK_WIDTH_CNTL 0x11140288 + +#define LINK_WIDTH_MAX 6 +#define LINK_SPEED_MAX 3 +static int link_width[] = {0 ,1 ,2 ,4 ,8 ,12 ,16 }; +static int link_speed[] = {25 ,50 ,80 ,160 };static int vega12_force_clock_level (struct pp_hwmgr *hwmgr,enum pp_clock_type type,uint32_t mask);
@@ -2095,6 +2102,46 @@ static int vega12_set_ppfeature_status(struct pp_hwmgr *hwmgr, uint64_t new_ppfereturn 0 ; } +static int vega12_get_current_pcie_link_width_level (struct pp_hwmgr *hwmgr) +{ +struct amdgpu_device *adev = hwmgr->adev; + +return (RREG32_PCIE (smnPCIE_LC_LINK_WIDTH_CNTL) & +PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD_MASK ) + >>PCIE_LC_LINK_WIDTH_CNTL__LC_LINK_WIDTH_RD__SHIFT ; +} + +static int vega12_get_current_pcie_link_width (struct pp_hwmgr *hwmgr) +{ +uint32_t width_level; + + width_level =vega12_get_current_pcie_link_width_level (hwmgr); +if (width_level >LINK_WIDTH_MAX ) + width_level =0 ; + +return link_width[width_level]; +} + +static int vega12_get_current_pcie_link_speed_level (struct pp_hwmgr *hwmgr) +{ +struct amdgpu_device *adev = hwmgr->adev; + +return (RREG32_PCIE (smnPCIE_LC_SPEED_CNTL) & +PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE_MASK ) + >>PSWUSP0_PCIE_LC_SPEED_CNTL__LC_CURRENT_DATA_RATE__SHIFT ; +} + +static int vega12_get_current_pcie_link_speed (struct pp_hwmgr *hwmgr) +{ +uint32_t speed_level; + + speed_level =vega12_get_current_pcie_link_speed_level (hwmgr); +if (speed_level >LINK_SPEED_MAX ) + speed_level =0 ; + +return link_speed[speed_level]; +} +static int vega12_print_clock_levels (struct pp_hwmgr *hwmgr,enum pp_clock_type type,char *buf) {
@@ -2682,6 +2729,67 @@ static int vega12_set_mp1_state(struct pp_hwmgr *hwmgr,return 0 ; } +static void vega12_init_gpu_metrics_v1_0 (struct gpu_metrics_v1_0 *gpu_metrics) +{ +memset (gpu_metrics,0xFF ,sizeof (struct gpu_metrics_v1_0 )); + + gpu_metrics->common_header.structure_size = +sizeof (struct gpu_metrics_v1_0 ); + gpu_metrics->common_header.format_revision =1 ; + gpu_metrics->common_header.content_revision =0 ; +} + +static ssize_t vega12_get_gpu_metrics (struct pp_hwmgr *hwmgr, +void **table) +{ +struct vega12_hwmgr *data = + (struct vega12_hwmgr *)(hwmgr->backend); +struct gpu_metrics_v1_0 *gpu_metrics = + &data->gpu_metrics_table; +SmuMetrics_t metrics; +uint32_t fan_speed_rpm; +int ret; + + ret =vega12_get_metrics_table (hwmgr, &metrics); +if (ret) +return ret; + +vega12_init_gpu_metrics_v1_0 (gpu_metrics); + + gpu_metrics->temperature_edge = metrics.TemperatureEdge; + gpu_metrics->temperature_hotspot = metrics.TemperatureHotspot; + gpu_metrics->temperature_mem = metrics.TemperatureHBM; + gpu_metrics->temperature_vrgfx = metrics.TemperatureVrGfx; + gpu_metrics->temperature_vrmem = metrics.TemperatureVrMem; + + gpu_metrics->average_gfx_activity = metrics.AverageGfxActivity; + gpu_metrics->average_umc_activity = metrics.AverageUclkActivity; + + gpu_metrics->average_gfxclk_frequency = metrics.AverageGfxclkFrequency; + gpu_metrics->average_socclk_frequency = metrics.AverageSocclkFrequency; + gpu_metrics->average_uclk_frequency = metrics.AverageUclkFrequency; + + gpu_metrics->current_gfxclk = metrics.CurrClock[PPCLK_GFXCLK ]; + gpu_metrics->current_socclk = metrics.CurrClock[PPCLK_SOCCLK ]; + gpu_metrics->current_uclk = metrics.CurrClock[PPCLK_UCLK ]; + gpu_metrics->current_vclk0 = metrics.CurrClock[PPCLK_VCLK ]; + gpu_metrics->current_dclk0 = metrics.CurrClock[PPCLK_DCLK ]; + + gpu_metrics->throttle_status = metrics.ThrottlerStatus; + +vega12_fan_ctrl_get_fan_speed_rpm (hwmgr, &fan_speed_rpm); + gpu_metrics->current_fan_speed = (uint16_t )fan_speed_rpm; + + gpu_metrics->pcie_link_width = +vega12_get_current_pcie_link_width (hwmgr); + gpu_metrics->pcie_link_speed = +vega12_get_current_pcie_link_speed (hwmgr); + + *table = (void *)gpu_metrics; + +return sizeof (struct gpu_metrics_v1_0 ); +} +static const struct pp_hwmgr_func vega12_hwmgr_funcs = { .backend_init = vega12_hwmgr_backend_init, .backend_fini = vega12_hwmgr_backend_fini,
@@ -2739,6 +2847,7 @@ static const struct pp_hwmgr_func vega12_hwmgr_funcs = { .get_ppfeature_status = vega12_get_ppfeature_status, .set_ppfeature_status = vega12_set_ppfeature_status, .set_mp1_state = vega12_set_mp1_state, + .get_gpu_metrics = vega12_get_gpu_metrics, };int vega12_hwmgr_init (struct pp_hwmgr *hwmgr)
drivers/gpu/drm/amd/powerplay/hwmgr/vega12_hwmgr.h +1 -0
@@ -399,6 +399,7 @@ struct vega12_hwmgr {unsigned long metrics_time;SmuMetrics_t metrics_table; +struct gpu_metrics_v1_0 gpu_metrics_table; };#define VEGA12_DPM2_NEAR_TDP_DEC 10
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 4/9 drm/amd/powerplay: add control method to bypass metrics cache on Arcturus
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- drivers/gpu/drm/amd/powerplay/arcturus_ppt.c | 74 +++++++++++++------- 1 file changed, 49 insertions(+), 25 deletions(-)
drivers/gpu/drm/amd/powerplay/arcturus_ppt.c +49 -25
@@ -541,18 +541,16 @@ static int arcturus_freqs_in_same_level(int32_t frequency1,return (abs (frequency1 - frequency2) <=EPSILON ); } -static int arcturus_get_smu_metrics_data (struct smu_context *smu, -MetricsMember_t member, -uint32_t *value) +static int arcturus_get_metrics_table_locked (struct smu_context *smu, +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct smu_table_context *smu_table= &smu->smu_table; -SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table;int ret =0 ; -mutex_lock (&smu->metrics_lock); - -if (!smu_table->metrics_time || -time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { +if (bypass_cache || + !smu_table->metrics_time || +time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { ret = smu_cmn_update_table(smu, SMU_TABLE_SMU_METRICS, 0,
@@ -560,12 +558,50 @@ static int arcturus_get_smu_metrics_data(struct smu_context *smu, false);if (ret) {dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" ); -mutex_unlock (&smu->metrics_lock);return ret; } smu_table->metrics_time = jiffies; } +if (metrics_table) +memcpy (metrics_table, smu_table->metrics_table,sizeof (SmuMetrics_t)); + +return 0 ; +} + +static int arcturus_get_metrics_table (struct smu_context *smu, +SmuMetrics_t *metrics_table, +bool bypass_cache) +{ +int ret =0 ; + +mutex_lock (&smu->metrics_lock); + ret =arcturus_get_metrics_table_locked (smu, + metrics_table, + bypass_cache); +mutex_unlock (&smu->metrics_lock); + +return ret; +} + +static int arcturus_get_smu_metrics_data (struct smu_context *smu, +MetricsMember_t member, +uint32_t *value) +{ +struct smu_table_context *smu_table= &smu->smu_table; +SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table; +int ret =0 ; + +mutex_lock (&smu->metrics_lock); + + ret =arcturus_get_metrics_table_locked (smu, +NULL , + false); +if (ret) { +mutex_unlock (&smu->metrics_lock); +return ret; + } +switch (member) {case METRICS_CURR_GFXCLK : *value = metrics->CurrClock[PPCLK_GFXCLK ];
@@ -2285,23 +2321,11 @@ static ssize_t arcturus_get_gpu_metrics(struct smu_context *smu,SmuMetrics_t metrics;int ret =0 ; -mutex_lock (&smu->metrics_lock); - - ret =smu_cmn_update_table (smu, -SMU_TABLE_SMU_METRICS , -0 , - smu_table->metrics_table, - false); -if (ret) { -dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" ); -mutex_unlock (&smu->metrics_lock); + ret =arcturus_get_metrics_table (smu, + &metrics, + true); +if (ret)return ret; - } - smu_table->metrics_time = jiffies; - -memcpy (&metrics, smu_table->metrics_table,sizeof (SmuMetrics_t)); - -mutex_unlock (&smu->metrics_lock);smu_v11_0_init_gpu_metrics_v1_0 (gpu_metrics);
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 5/9 drm/amd/powerplay: add control method to bypass metrics cache on Navi10
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- drivers/gpu/drm/amd/powerplay/navi10_ppt.c | 60 ++++++++++++++-------- 1 file changed, 38 insertions(+), 22 deletions(-)
drivers/gpu/drm/amd/powerplay/navi10_ppt.c +38 -22
@@ -504,22 +504,16 @@ static int navi10_tables_init(struct smu_context *smu)return -ENOMEM ; } -static int navi10_get_smu_metrics_data (struct smu_context *smu, -MetricsMember_t member, -uint32_t *value) +static int navi10_get_metrics_table_locked (struct smu_context *smu, +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct smu_table_context *smu_table= &smu->smu_table; -/* -* This works for NV12 also. As although NV12 uses a different -* SmuMetrics structure from other NV1X ASICs, they share the -* same offsets for the heading parts(those members used here). -*/ -SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table;int ret =0 ; -mutex_lock (&smu->metrics_lock); -if (!smu_table->metrics_time || -time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { +if (bypass_cache || + !smu_table->metrics_time || +time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { ret = smu_cmn_update_table(smu, SMU_TABLE_SMU_METRICS, 0,
@@ -527,12 +521,40 @@ static int navi10_get_smu_metrics_data(struct smu_context *smu, false);if (ret) {dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" ); -mutex_unlock (&smu->metrics_lock);return ret; } smu_table->metrics_time = jiffies; } +if (metrics_table) +memcpy (metrics_table, smu_table->metrics_table,sizeof (SmuMetrics_t)); + +return 0 ; +} + +static int navi10_get_smu_metrics_data (struct smu_context *smu, +MetricsMember_t member, +uint32_t *value) +{ +struct smu_table_context *smu_table= &smu->smu_table; +/* +* This works for NV12 also. As although NV12 uses a different +* SmuMetrics structure from other NV1X ASICs, they share the +* same offsets for the heading parts(those members used here). +*/ +SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table; +int ret =0 ; + +mutex_lock (&smu->metrics_lock); + + ret =navi10_get_metrics_table_locked (smu, +NULL , + false); +if (ret) { +mutex_unlock (&smu->metrics_lock); +return ret; + } +switch (member) {case METRICS_CURR_GFXCLK : *value = metrics->CurrClock[PPCLK_GFXCLK ];
@@ -2526,19 +2548,13 @@ static ssize_t navi10_get_gpu_metrics(struct smu_context *smu,mutex_lock (&smu->metrics_lock); - ret =smu_cmn_update_table (smu, -SMU_TABLE_SMU_METRICS , -0 , - smu_table->metrics_table, - false); + ret =navi10_get_metrics_table_locked (smu, + &metrics, + true);if (ret) { -dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" );mutex_unlock (&smu->metrics_lock);return ret; } - smu_table->metrics_time = jiffies; - -memcpy (&metrics, smu_table->metrics_table,sizeof (SmuMetrics_t));if (adev->asic_type ==CHIP_NAVI12 )memcpy (&nv12_metrics, smu_table->metrics_table,sizeof (SmuMetrics_NV12_t));
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 6/9 drm/amd/powerplay: add control method to bypass metrics cache on Sienna Cichlid
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- .../drm/amd/powerplay/sienna_cichlid_ppt.c | 73 +++++++++++++------ 1 file changed, 49 insertions(+), 24 deletions(-)
drivers/gpu/drm/amd/powerplay/sienna_cichlid_ppt.c +49 -24
@@ -407,17 +407,16 @@ static int sienna_cichlid_tables_init(struct smu_context *smu)return -ENOMEM ; } -static int sienna_cichlid_get_smu_metrics_data (struct smu_context *smu, -MetricsMember_t member, -uint32_t *value) +static int sienna_cichlid_get_metrics_table_locked (struct smu_context *smu, +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct smu_table_context *smu_table= &smu->smu_table; -SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table;int ret =0 ; -mutex_lock (&smu->metrics_lock); -if (!smu_table->metrics_time || -time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { +if (bypass_cache || + !smu_table->metrics_time || +time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (1 ))) { ret = smu_cmn_update_table(smu, SMU_TABLE_SMU_METRICS, 0,
@@ -425,12 +424,50 @@ static int sienna_cichlid_get_smu_metrics_data(struct smu_context *smu, false);if (ret) {dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" ); -mutex_unlock (&smu->metrics_lock);return ret; } smu_table->metrics_time = jiffies; } +if (metrics_table) +memcpy (metrics_table, smu_table->metrics_table,sizeof (SmuMetrics_t)); + +return 0 ; +} + +static int sienna_cichlid_get_metrics_table (struct smu_context *smu, +SmuMetrics_t *metrics_table, +bool bypass_cache) +{ +int ret =0 ; + +mutex_lock (&smu->metrics_lock); + ret =sienna_cichlid_get_metrics_table_locked (smu, + metrics_table, + bypass_cache); +mutex_unlock (&smu->metrics_lock); + +return ret; +} + +static int sienna_cichlid_get_smu_metrics_data (struct smu_context *smu, +MetricsMember_t member, +uint32_t *value) +{ +struct smu_table_context *smu_table= &smu->smu_table; +SmuMetrics_t *metrics = (SmuMetrics_t *)smu_table->metrics_table; +int ret =0 ; + +mutex_lock (&smu->metrics_lock); + + ret =sienna_cichlid_get_metrics_table_locked (smu, +NULL , + false); +if (ret) { +mutex_unlock (&smu->metrics_lock); +return ret; + } +switch (member) {case METRICS_CURR_GFXCLK : *value = metrics->CurrClock[PPCLK_GFXCLK ];
@@ -2684,23 +2721,11 @@ static ssize_t sienna_cichlid_get_gpu_metrics(struct smu_context *smu,SmuMetrics_t metrics;int ret =0 ; -mutex_lock (&smu->metrics_lock); - - ret =smu_cmn_update_table (smu, -SMU_TABLE_SMU_METRICS , -0 , - smu_table->metrics_table, - false); -if (ret) { -dev_info (smu->adev->dev,"Failed to export SMU metrics table!\n" ); -mutex_unlock (&smu->metrics_lock); + ret =sienna_cichlid_get_metrics_table (smu, + &metrics, + true); +if (ret)return ret; - } - smu_table->metrics_time = jiffies; - -memcpy (&metrics, smu_table->metrics_table,sizeof (SmuMetrics_t)); - -mutex_unlock (&smu->metrics_lock);smu_v11_0_init_gpu_metrics_v1_0 (gpu_metrics);
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 7/9 drm/amd/powerplay: add control method to bypass metrics cache on Renoir
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- drivers/gpu/drm/amd/powerplay/renoir_ppt.c | 22 ++++++++++++++-------- 1 file changed, 14 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/powerplay/renoir_ppt.c +14 -8
@@ -129,13 +129,17 @@ static struct cmn2asic_mapping renoir_workload_map[PP_SMC_POWER_PROFILE_COUNT] = };static int renoir_get_metrics_table (struct smu_context *smu, -SmuMetrics_t *metrics_table) +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct smu_table_context *smu_table= &smu->smu_table;int ret =0 ;mutex_lock (&smu->metrics_lock); -if (!smu_table->metrics_time ||time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (100 ))) { + +if (bypass_cache || + !smu_table->metrics_time || +time_after (jiffies, smu_table->metrics_time +msecs_to_jiffies (100 ))) { ret =smu_cmn_update_table (smu,SMU_TABLE_SMU_METRICS ,0 , (void *)smu_table->metrics_table, false);if (ret) {
@@ -146,7 +150,9 @@ static int renoir_get_metrics_table(struct smu_context *smu, smu_table->metrics_time = jiffies; } -memcpy (metrics_table, smu_table->metrics_table,sizeof (SmuMetrics_t)); +if (metrics_table) +memcpy (metrics_table, smu_table->metrics_table,sizeof (SmuMetrics_t)); +mutex_unlock (&smu->metrics_lock);return ret;
@@ -375,7 +381,7 @@ static int renoir_print_clk_levels(struct smu_context *smu,memset (&metrics,0 ,sizeof (metrics)); - ret =renoir_get_metrics_table (smu, &metrics); + ret =renoir_get_metrics_table (smu, &metrics, false);if (ret)return ret;
@@ -529,7 +535,7 @@ static int renoir_get_current_clk_freq_by_table(struct smu_context *smu,int ret =0 , clk_id =0 ;SmuMetrics_t metrics; - ret =renoir_get_metrics_table (smu, &metrics); + ret =renoir_get_metrics_table (smu, &metrics, false);if (ret)return ret;
@@ -612,7 +618,7 @@ static int renoir_get_gpu_temperature(struct smu_context *smu, uint32_t *value)if (!value)return -EINVAL ; - ret =renoir_get_metrics_table (smu, &metrics); + ret =renoir_get_metrics_table (smu, &metrics, false);if (ret)return ret;
@@ -632,7 +638,7 @@ static int renoir_get_current_activity_percent(struct smu_context *smu,if (!value)return -EINVAL ; - ret =renoir_get_metrics_table (smu, &metrics); + ret =renoir_get_metrics_table (smu, &metrics, false);if (ret)return ret;
@@ -1018,7 +1024,7 @@ static ssize_t renoir_get_gpu_metrics(struct smu_context *smu,SmuMetrics_t metrics;int ret =0 ; - ret =renoir_get_metrics_table (smu, &metrics); + ret =renoir_get_metrics_table (smu, &metrics, true);if (ret)return ret;
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 8/9 drm/amd/powerplay: add control method to bypass metrics cache on Vega20
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- .../drm/amd/powerplay/hwmgr/vega20_hwmgr.c | 31 ++++++++++++------- 1 file changed, 19 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/powerplay/hwmgr/vega20_hwmgr.c +19 -12
@@ -2090,22 +2090,29 @@ static uint32_t vega20_dpm_get_mclk(struct pp_hwmgr *hwmgr, bool low)return (mem_clk *100 ); } -static int vega20_get_metrics_table (struct pp_hwmgr *hwmgr,SmuMetrics_t *metrics_table) +static int vega20_get_metrics_table (struct pp_hwmgr *hwmgr, +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct vega20_hwmgr *data = (struct vega20_hwmgr *)(hwmgr->backend);int ret =0 ; -if (!data->metrics_time ||time_after (jiffies, data->metrics_time +HZ /2 )) { - ret =smum_smc_table_manager (hwmgr, (uint8_t *)metrics_table, -TABLE_SMU_METRICS , true); +if (bypass_cache || + !data->metrics_time || +time_after (jiffies, data->metrics_time +HZ /2 )) { + ret =smum_smc_table_manager (hwmgr, + (uint8_t *)(&data->metrics_table), +TABLE_SMU_METRICS , + true);if (ret) {pr_info ("Failed to export SMU metrics table!\n" );return ret; } -memcpy (&data->metrics_table, metrics_table,sizeof (SmuMetrics_t)); data->metrics_time = jiffies; - }else + } + +if (metrics_table)memcpy (metrics_table, &data->metrics_table,sizeof (SmuMetrics_t));return ret;
@@ -2117,7 +2124,7 @@ static int vega20_get_gpu_power(struct pp_hwmgr *hwmgr,int ret =0 ;SmuMetrics_t metrics_table; - ret =vega20_get_metrics_table (hwmgr, &metrics_table); + ret =vega20_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -2155,7 +2162,7 @@ static int vega20_get_current_activity_percent(struct pp_hwmgr *hwmgr,int ret =0 ;SmuMetrics_t metrics_table; - ret =vega20_get_metrics_table (hwmgr, &metrics_table); + ret =vega20_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -2185,7 +2192,7 @@ static int vega20_read_sensor(struct pp_hwmgr *hwmgr, int idx,switch (idx) {case AMDGPU_PP_SENSOR_GFX_SCLK : - ret =vega20_get_metrics_table (hwmgr, &metrics_table); + ret =vega20_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -2210,7 +2217,7 @@ static int vega20_read_sensor(struct pp_hwmgr *hwmgr, int idx, *size =4 ;break ;case AMDGPU_PP_SENSOR_EDGE_TEMP : - ret =vega20_get_metrics_table (hwmgr, &metrics_table); + ret =vega20_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -2219,7 +2226,7 @@ static int vega20_read_sensor(struct pp_hwmgr *hwmgr, int idx, *size =4 ;break ;case AMDGPU_PP_SENSOR_MEM_TEMP : - ret =vega20_get_metrics_table (hwmgr, &metrics_table); + ret =vega20_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -4281,7 +4288,7 @@ static ssize_t vega20_get_gpu_metrics(struct pp_hwmgr *hwmgr,uint32_t fan_speed_rpm;int ret; - ret =vega20_get_metrics_table (hwmgr, &metrics); + ret =vega20_get_metrics_table (hwmgr, &metrics, true);if (ret)return ret;
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 9/9 drm/amd/powerplay: add control method to bypass metrics cache on Vega12
In reply to Evan QuanAs for the gpu metric export, metrics cache makes no sense. It's up to user to decide how often the metrics should be retrieved.
--- .../drm/amd/powerplay/hwmgr/vega12_hwmgr.c | 29 ++++++++++++------- 1 file changed, 18 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/powerplay/hwmgr/vega12_hwmgr.c +18 -11
@@ -1262,22 +1262,29 @@ static uint32_t vega12_dpm_get_mclk(struct pp_hwmgr *hwmgr, bool low)return (mem_clk *100 ); } -static int vega12_get_metrics_table (struct pp_hwmgr *hwmgr,SmuMetrics_t *metrics_table) +static int vega12_get_metrics_table (struct pp_hwmgr *hwmgr, +SmuMetrics_t *metrics_table, +bool bypass_cache) {struct vega12_hwmgr *data = (struct vega12_hwmgr *)(hwmgr->backend);int ret =0 ; -if (!data->metrics_time ||time_after (jiffies, data->metrics_time +HZ /2 )) { - ret =smum_smc_table_manager (hwmgr, (uint8_t *)metrics_table, -TABLE_SMU_METRICS , true); +if (bypass_cache || + !data->metrics_time || +time_after (jiffies, data->metrics_time +HZ /2 )) { + ret =smum_smc_table_manager (hwmgr, + (uint8_t *)(&data->metrics_table), +TABLE_SMU_METRICS , + true);if (ret) {pr_info ("Failed to export SMU metrics table!\n" );return ret; } -memcpy (&data->metrics_table, metrics_table,sizeof (SmuMetrics_t)); data->metrics_time = jiffies; - }else + } + +if (metrics_table)memcpy (metrics_table, &data->metrics_table,sizeof (SmuMetrics_t));return ret;
@@ -1288,7 +1295,7 @@ static int vega12_get_gpu_power(struct pp_hwmgr *hwmgr, uint32_t *query)SmuMetrics_t metrics_table;int ret =0 ; - ret =vega12_get_metrics_table (hwmgr, &metrics_table); + ret =vega12_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -1339,7 +1346,7 @@ static int vega12_get_current_activity_percent(SmuMetrics_t metrics_table;int ret =0 ; - ret =vega12_get_metrics_table (hwmgr, &metrics_table); + ret =vega12_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -1387,7 +1394,7 @@ static int vega12_read_sensor(struct pp_hwmgr *hwmgr, int idx, *size =4 ;break ;case AMDGPU_PP_SENSOR_HOTSPOT_TEMP : - ret =vega12_get_metrics_table (hwmgr, &metrics_table); + ret =vega12_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -1396,7 +1403,7 @@ static int vega12_read_sensor(struct pp_hwmgr *hwmgr, int idx, *size =4 ;break ;case AMDGPU_PP_SENSOR_MEM_TEMP : - ret =vega12_get_metrics_table (hwmgr, &metrics_table); + ret =vega12_get_metrics_table (hwmgr, &metrics_table, false);if (ret)return ret;
@@ -2750,7 +2757,7 @@ static ssize_t vega12_get_gpu_metrics(struct pp_hwmgr *hwmgr,uint32_t fan_speed_rpm;int ret; - ret =vega12_get_metrics_table (hwmgr, &metrics); + ret =vega12_get_metrics_table (hwmgr, &metrics, true);if (ret)return ret;
Signature · 6 lines
2.28.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx