VRM Vdd Thermal Control Loop

Change-Id: Ie5771b33c13cd57822a8b6f1406a92c4516088c0
RTC: 180397
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/48563
Tested-by: FSP CI Jenkins <fsp-CI-jenkins+hostboot@us.ibm.com>
Reviewed-by: William A. Bryan <wilbryan@us.ibm.com>
Reviewed-by: Andres A. Lugo-Reyes <aalugore@us.ibm.com>
Reviewed-by: Christopher J. Cain <cjcain@us.ibm.com>
Reviewed-by: Martha Broyles <mbroyles@us.ibm.com>

11 files changed, 224 insertions, 40 deletions

diff --git a/src/occ_405/amec/amec_controller.c b/src/occ_405/amec/amec_controller.c
index 2154d07..f310208 100644
--- a/src/occ_405/amec/amec_controller.c
+++ b/src/occ_405/amec/amec_controller.c
@@ -156,6 +156,88 @@ void amec_controller_proc_thermal()
     }
 }
 
+// Function Specification
+//
+// Name: amec_controller_vrm_vdd_thermal
+//
+// Description: This function implements the Proportional Controller for the
+//              VRM Vdd thermal control. Although it doesn't return any
+//              results, it populates the thermal vote in the field
+//              g_amec->thermalvdd.speed_request.
+//
+// Task Flags:
+//
+// End Function Specification
+void amec_controller_vrm_vdd_thermal()
+{
+    /*------------------------------------------------------------------------*/
+    /*  Local Variables                                                       */
+    /*------------------------------------------------------------------------*/
+    uint16_t                      l_vdd_temp_tenthsC = 0;
+    uint16_t                      l_residue = 0;
+    uint16_t                      l_old_residue = 0;
+    int16_t                       l_thermal_diff = 0;
+    int16_t                       l_cpu_speed = 0;
+    int16_t                       l_throttle_chg = 0;
+    int32_t                       l_throttle = 0;
+    sensor_t                    * l_sensor = NULL;
+
+    /*------------------------------------------------------------------------*/
+    /*  Code                                                                  */
+    /*------------------------------------------------------------------------*/
+    // Get VRM Vdd temperature sensor
+    l_sensor = getSensorByGsid(TEMPVDD);
+
+    // Convert current Vdd temperature to 0.1 degrees C
+    l_vdd_temp_tenthsC = l_sensor->sample * 10;
+
+    // Calculate the temperature difference from the DVFS setpoint
+    l_thermal_diff = g_amec->thermalvdd.setpoint - l_vdd_temp_tenthsC;
+
+    // Proportional Controller for the thermal control loop
+    l_throttle = (int32_t) l_thermal_diff * g_amec->thermalvdd.Pgain;
+    l_residue = (uint16_t) l_throttle;
+    l_throttle_chg = (int16_t) (l_throttle >> 16);
+
+    // don't allow a throttle change more than step limit
+    if ((int16_t) l_throttle_chg > (int16_t) g_amec->sys.speed_step_limit)
+    {
+        l_throttle_chg = g_amec->sys.speed_step_limit;
+    }
+    else
+    {
+        if ((int16_t) l_throttle_chg < ((int16_t) (-g_amec->sys.speed_step_limit)))
+        {
+            l_throttle_chg = (int16_t)(-g_amec->sys.speed_step_limit);
+        }
+    }
+
+    // Calculate the new thermal CPU speed request
+    l_cpu_speed = g_amec->thermalvdd.speed_request +
+        (int16_t)(l_throttle_chg * g_amec->sys.speed_step);
+
+    // Proceed with residue summation to correctly follow set-point
+    l_old_residue = g_amec->thermalvdd.total_res;
+    g_amec->thermalvdd.total_res += l_residue;
+    if (g_amec->thermalvdd.total_res < l_old_residue)
+    {
+        l_cpu_speed += g_amec->sys.speed_step;
+    }
+
+    // Enforce actuator saturation limits
+    if (l_cpu_speed > g_amec->sys.max_speed)
+        l_cpu_speed = g_amec->sys.max_speed;
+    if (l_cpu_speed < g_amec->sys.min_speed)
+        l_cpu_speed = g_amec->sys.min_speed;
+
+    // Generate the new thermal speed request
+    g_amec->thermalvdd.speed_request = l_cpu_speed;
+    // Calculate frequency request based on thermal speed request
+    g_amec->thermalvdd.freq_request = amec_controller_speed2freq(
+            g_amec->thermalvdd.speed_request,
+            g_amec->sys.fmax);
+}
+
 //*************************************************************************
 // Function Specification
 //
diff --git a/src/occ_405/amec/amec_controller.h b/src/occ_405/amec/amec_controller.h
index 76e70ef..29a8061 100644
--- a/src/occ_405/amec/amec_controller.h
+++ b/src/occ_405/amec/amec_controller.h
@@ -91,6 +91,16 @@ typedef struct amec_controller
 void amec_controller_proc_thermal();
 
 /**
+ *  Thermal Control Loop based on VRM Vdd temperature
+ *
+ *  This function implements the Proportional Controller for the VRM Vdd thermal
+ *  control loop. Although it doesn't return any results, it populates the
+ *  thermal vote in the field g_amec->thermalvdd.speed_request.
+ *
+ */
+void amec_controller_vrm_vdd_thermal();
+
+/**
  *  Thermal Control Loop based on Centaur temperatures.
  *
  *  This function implements a Proportional Controller for the
diff --git a/src/occ_405/amec/amec_freq.c b/src/occ_405/amec/amec_freq.c
index 03bc22c..9150328 100755
--- a/src/occ_405/amec/amec_freq.c
+++ b/src/occ_405/amec/amec_freq.c
@@ -282,6 +282,9 @@ void amec_slv_proc_voting_box(void)
     amec_proc_voting_reason_t       l_kvm_throt_reason = NO_THROTTLE;
     amec_part_t                     *l_part = NULL;
 
+    // frequency threshold for reporting throttling
+    uint16_t l_report_throttle_freq = G_sysConfigData.system_type.report_dvfs_nom ?  G_sysConfigData.sys_mode_freq.table[OCC_MODE_NOMINAL] : G_sysConfigData.sys_mode_freq.table[OCC_MODE_TURBO];
+
     /*------------------------------------------------------------------------*/
     /*  Code                                                                  */
     /*------------------------------------------------------------------------*/
@@ -307,9 +310,9 @@ void amec_slv_proc_voting_box(void)
         l_chip_fmax = g_amec->proc[0].pwr_votes.ppb_fmax;
         l_chip_reason = AMEC_VOTING_REASON_PPB;
 
-        if( G_sysConfigData.sys_mode_freq.table[OCC_MODE_TURBO] < l_chip_fmax)
+        if(l_report_throttle_freq <= l_chip_fmax)
         {
-            l_kvm_throt_reason = PCAP_EXCEED_PTURBO;
+            l_kvm_throt_reason = PCAP_EXCEED_REPORT;
         }
         else
         {
@@ -340,9 +343,9 @@ void amec_slv_proc_voting_box(void)
         l_chip_fmax = g_amec->thermalproc.freq_request;
         l_chip_reason = AMEC_VOTING_REASON_PROC_THRM;
 
-        if( G_sysConfigData.sys_mode_freq.table[OCC_MODE_TURBO] < l_chip_fmax)
+        if( l_report_throttle_freq <= l_chip_fmax)
         {
-            l_kvm_throt_reason = PROC_OVERTEMP_EXCEED_PTURBO;
+            l_kvm_throt_reason = PROC_OVERTEMP_EXCEED_REPORT;
         }
         else
         {
@@ -350,15 +353,31 @@ void amec_slv_proc_voting_box(void)
         }
     }
 
+    //Thermal controller input based on VRM Vdd temperature
+    if(g_amec->thermalvdd.freq_request < l_chip_fmax)
+    {
+        l_chip_fmax = g_amec->thermalvdd.freq_request;
+        l_chip_reason = AMEC_VOTING_REASON_VDD_THRM;
+
+        if( l_report_throttle_freq <= l_chip_fmax)
+        {
+            l_kvm_throt_reason = VDD_OVERTEMP_EXCEED_REPORT;
+        }
+        else
+        {
+            l_kvm_throt_reason = VDD_OVERTEMP;
+        }
+    }
+
     // Controller request based on VRHOT signal from processor regulator
     if(g_amec->vrhotproc.freq_request < l_chip_fmax)
     {
         l_chip_fmax = g_amec->vrhotproc.freq_request;
         l_chip_reason = AMEC_VOTING_REASON_VRHOT_THRM;
 
-        if(G_sysConfigData.sys_mode_freq.table[OCC_MODE_TURBO] < l_chip_fmax)
+        if(l_report_throttle_freq <= l_chip_fmax)
         {
-            l_kvm_throt_reason = PROC_OVERTEMP_EXCEED_PTURBO;
+            l_kvm_throt_reason = PROC_OVERTEMP_EXCEED_REPORT;
         }
         else
         {
@@ -393,6 +412,7 @@ void amec_slv_proc_voting_box(void)
                         // Before enforcing a soft Fmin, make sure we don't
                         // have a thermal or power emergency
                         if(!(l_chip_reason & (AMEC_VOTING_REASON_PROC_THRM |
+                                              AMEC_VOTING_REASON_VDD_THRM |
                                               AMEC_VOTING_REASON_VRHOT_THRM |
                                               AMEC_VOTING_REASON_PPB |
                                               AMEC_VOTING_REASON_PMAX |
@@ -428,9 +448,9 @@ void amec_slv_proc_voting_box(void)
                     l_core_freq = g_amec->proc[0].pwr_votes.proc_pcap_vote;
                     l_core_reason = AMEC_VOTING_REASON_PWR;
 
-                    if(G_sysConfigData.sys_mode_freq.table[OCC_MODE_TURBO] < l_core_freq)
+                    if(l_report_throttle_freq <= l_core_freq)
                     {
-                        l_kvm_throt_reason = PCAP_EXCEED_PTURBO;
+                        l_kvm_throt_reason = PCAP_EXCEED_REPORT;
                     }
                     else
                     {
@@ -544,19 +564,17 @@ void amec_slv_proc_voting_box(void)
         }
     }//End of for loop
 
-    //convert POWERCAP reason to POWER_SUPPLY_FAILURE if ovs is asserted
-    if((l_kvm_throt_reason == POWERCAP) && AMEC_INTF_GET_OVERSUBSCRIPTION())
+    //check if there was a throttle reason change
+    if(l_kvm_throt_reason != G_amec_opal_proc_throt_reason)
     {
-        l_kvm_throt_reason = POWER_SUPPLY_FAILURE;
-    }
-
-    //check if we need to update the throttle reason in homer
-    if(G_sysConfigData.system_type.kvm &&
-       (l_kvm_throt_reason != G_amec_opal_proc_throt_reason))
-    {
-        //Notify dcom thread to update the table
+        //Always update G_amec_opal_proc_throt_reason, this is used to set poll rsp bits for all system types
         G_amec_opal_proc_throt_reason = l_kvm_throt_reason;
-        ssx_semaphore_post(&G_dcomThreadWakeupSem);
+
+        // Only if running OPAL need to notify dcom thread to update the table in HOMER for OPAL
+        if(G_sysConfigData.system_type.kvm)
+        {
+            ssx_semaphore_post(&G_dcomThreadWakeupSem);
+        }
     }
 }
 
diff --git a/src/occ_405/amec/amec_freq.h b/src/occ_405/amec/amec_freq.h
index d2b9c90..2cf3c85 100644
--- a/src/occ_405/amec/amec_freq.h
+++ b/src/occ_405/amec/amec_freq.h
@@ -81,6 +81,7 @@ typedef enum
     AMEC_VOTING_REASON_OVERRIDE_CORE    = 0x00020000,
     AMEC_VOTING_REASON_IPS              = 0x00040000,
     AMEC_VOTING_REASON_APSS_PMAX        = 0x00080000,
+    AMEC_VOTING_REASON_VDD_THRM         = 0x00100000,
 }amec_freq_voting_reason_t;
 
 
@@ -116,8 +117,10 @@ typedef enum {
     POWER_SUPPLY_FAILURE            = 0x03,
     OVERCURRENT                     = 0x04,
     OCC_RESET                       = 0x05,
-    PCAP_EXCEED_PTURBO              = 0x06,
-    PROC_OVERTEMP_EXCEED_PTURBO     = 0x07,
+    PCAP_EXCEED_REPORT              = 0x06,
+    PROC_OVERTEMP_EXCEED_REPORT     = 0x07,
+    VDD_OVERTEMP                    = 0x08,
+    VDD_OVERTEMP_EXCEED_REPORT      = 0x09,
     MANUFACTURING_OVERRIDE          = 0xAA,
 }amec_proc_voting_reason_t;
 
diff --git a/src/occ_405/amec/amec_init.c b/src/occ_405/amec/amec_init.c
index 5ca4e66..4ba89fc 100644
--- a/src/occ_405/amec/amec_init.c
+++ b/src/occ_405/amec/amec_init.c
@@ -250,6 +250,13 @@ void amec_init_gamec_struct(void)
   g_amec->thermalproc.freq_request = -1; //unconstrained frequency vote
   g_amec->thermalproc.total_res = 0;
 
+  // Initialize thermal controller for VRM Vdd
+  g_amec->thermalvdd.setpoint = 850; // 850 = 85.0 C
+  g_amec->thermalvdd.Pgain = 1000;
+  g_amec->thermalvdd.speed_request = 1000;
+  g_amec->thermalvdd.freq_request = -1; //unconstrained frequency vote
+  g_amec->thermalvdd.total_res = 0;
+
   // Initialize thermal controller based on DIMM temperatures
   g_amec->thermaldimm.setpoint = 850; //In 0.1 degrees C -> 850 = 85.0 C
   g_amec->thermaldimm.Pgain = 30000;
diff --git a/src/occ_405/amec/amec_master_smh.c b/src/occ_405/amec/amec_master_smh.c
index ce84749..ab2a3af 100755
--- a/src/occ_405/amec/amec_master_smh.c
+++ b/src/occ_405/amec/amec_master_smh.c
@@ -59,8 +59,8 @@
 //Power cap failure threshold with no GPUs set to 32 ticks
 #define PCAP_FAILURE_THRESHOLD 32
 
-//Power cap failure threshold with GPUs set to number of ticks for 100ms
-#define PCAP_GPU_FAILURE_THRESHOLD (100000 / MICS_PER_TICK)
+//Power cap failure threshold with GPUs set to number of ticks for 2s
+#define PCAP_GPU_FAILURE_THRESHOLD (2000000 / MICS_PER_TICK)
 
 //*************************************************************************/
 // Structures
@@ -398,6 +398,9 @@ void amec_mst_check_under_pcap(void)
     errlHndl_t l_err = NULL;
     uint8_t i = 0;
     uint8_t l_apss_func_id = 0;
+    uint32_t l_trace[MAX_APSS_ADC_CHANNELS] = {0};  // used to trace per channel data
+    uint8_t l_trace_idx = 0;
+
 
     /*------------------------------------------------------------------------*/
     /*  Code                                                                  */
@@ -424,6 +427,7 @@ void amec_mst_check_under_pcap(void)
                      AMECSENSOR_PTR(PWRSYS)->sample);
 
             // Trace power per APSS channel to have the best breakdown for debug
+            // compress traces to 4 max to save space on OP systems
             for (i = 0; i < MAX_APSS_ADC_CHANNELS; i++)
             {
                 l_apss_func_id = G_apss_ch_to_function[i];
@@ -433,10 +437,37 @@ void amec_mst_check_under_pcap(void)
                    (l_apss_func_id != ADC_GND_REMOTE_SENSE) &&
                    (l_apss_func_id != ADC_12V_STANDBY_CURRENT) )
                 {
-                    TRAC_ERR("APSS channel %d Function ID = %d Power = %dW", i, l_apss_func_id,
-                              AMECSENSOR_PTR(PWRAPSSCH0 + i)->sample);
+                    l_trace[l_trace_idx] = (i << 24) | (l_apss_func_id << 16) | (AMECSENSOR_PTR(PWRAPSSCH0 + i)->sample);
+                    l_trace_idx++;
                 }
             }
+            while(l_trace_idx != 0)
+            {
+               if(l_trace_idx >=4)
+               {
+                  TRAC_ERR("APSS channel/FuncID/Power: [%08X], [%08X], [%08X], [%08X]",
+                             l_trace[l_trace_idx-1], l_trace[l_trace_idx-2], l_trace[l_trace_idx-3], l_trace[l_trace_idx-4]);
+                  l_trace_idx -= 4;
+               }
+               else if(l_trace_idx == 3)
+               {
+                  TRAC_ERR("APSS channel/FuncID/Power: [%08X], [%08X], [%08X]",
+                             l_trace[l_trace_idx-1], l_trace[l_trace_idx-2], l_trace[l_trace_idx-3]);
+                  l_trace_idx = 0;
+               }
+               else if(l_trace_idx == 2)
+               {
+                  TRAC_ERR("APSS channel/FuncID/Power: [%08X], [%08X]",
+                             l_trace[l_trace_idx-1], l_trace[l_trace_idx-2]);
+                  l_trace_idx = 0;
+               }
+               else // l_trace_idx == 1
+               {
+                  TRAC_ERR("APSS channel/FuncID/Power: [%08X]",
+                             l_trace[l_trace_idx-1]);
+                  l_trace_idx = 0;
+               }
+            }
 
             /* @
              * @errortype
diff --git a/src/occ_405/amec/amec_slave_smh.c b/src/occ_405/amec/amec_slave_smh.c
index 9142caa..4750336 100755
--- a/src/occ_405/amec/amec_slave_smh.c
+++ b/src/occ_405/amec/amec_slave_smh.c
@@ -547,6 +547,9 @@ void amec_slv_state_2(void)
   //-------------------------------------------------------
   amec_update_centaur_sensors(CENTAUR_2);
 */
+
+  // Call VRM Vdd thermal controller
+  amec_controller_vrm_vdd_thermal();
 }
 
 
diff --git a/src/occ_405/cmdh/cmdh_fsp_cmds.c b/src/occ_405/cmdh/cmdh_fsp_cmds.c
index 86ee360..5f27ba8 100755
--- a/src/occ_405/cmdh/cmdh_fsp_cmds.c
+++ b/src/occ_405/cmdh/cmdh_fsp_cmds.c
@@ -60,6 +60,7 @@ extern bool G_vrm_vdd_temp_expired;
 
 #include <gpe_export.h>
 extern gpe_shared_data_t G_shared_gpe_data;
+extern opal_proc_voting_reason_t G_amec_opal_proc_throt_reason;
 
 // This table contains tunable parameter information that can be exposed to
 // customers (only Master OCC should access/control this table)
@@ -170,12 +171,29 @@ ERRL_RC cmdh_poll_v20(cmdh_fsp_rsp_t * o_rsp_ptr)
         uint32_t l_freq_reason = g_amec->proc[0].core[k].f_reason;
         if ( l_freq_reason & (AMEC_VOTING_REASON_PROC_THRM | AMEC_VOTING_REASON_VRHOT_THRM) )
         {
-            l_poll_rsp->ext_status.dvfs_due_to_ot = 1;
+            // only set DVFS bit if throttling below frequency to report throttling
+            if(G_amec_opal_proc_throt_reason == CPU_OVERTEMP)
+            {
+               l_poll_rsp->ext_status.dvfs_due_to_ot = 1;
+            }
+        }
+
+        if ( l_freq_reason & AMEC_VOTING_REASON_VDD_THRM )
+        {
+            // only set DVFS bit if throttling below frequency to report throttling
+            if(G_amec_opal_proc_throt_reason == VDD_OVERTEMP)
+            {
+               l_poll_rsp->ext_status.dvfs_due_to_vdd_ot = 1;
+            }
         }
 
         if ( l_freq_reason & (AMEC_VOTING_REASON_PPB | AMEC_VOTING_REASON_PMAX | AMEC_VOTING_REASON_PWR) )
         {
-            l_poll_rsp->ext_status.dvfs_due_to_pwr = 1;
+            // only set DVFS bit if throttling below frequency to report throttling
+            if(G_amec_opal_proc_throt_reason == POWERCAP)
+            {
+               l_poll_rsp->ext_status.dvfs_due_to_pwr = 1;
+            }
         }
     }
 
diff --git a/src/occ_405/cmdh/cmdh_fsp_cmds.h b/src/occ_405/cmdh/cmdh_fsp_cmds.h
index 9dda8dc..a6c7083 100755
--- a/src/occ_405/cmdh/cmdh_fsp_cmds.h
+++ b/src/occ_405/cmdh/cmdh_fsp_cmds.h
@@ -116,14 +116,14 @@ typedef struct __attribute__ ((packed)) cmdh_poll_resp_v20
     {
         struct
         {
-            uint8_t dvfs_due_to_ot  : 1;   // 1 => OCC clipped max Pstate due to an over temp.
-            uint8_t dvfs_due_to_pwr : 1;   // 1 => OCC clipped max Psate due to reaching pcap limit.
-            uint8_t mthrot_due_to_ot: 1;   // 1 => OCC throttled memory due to an over temp.
-            uint8_t n_power         : 1;   // 1 => Server running without redundant power.
-            uint8_t _reserved_3     : 1;
-            uint8_t sync_request    : 1;   // 1 => OCC needs to restart snapshot buffers
-            uint8_t _reserved_1     : 1;
-            uint8_t _reserved_0     : 1;
+            uint8_t dvfs_due_to_ot     : 1;   // 1 => OCC clipped max Pstate due to a processor over temp.
+            uint8_t dvfs_due_to_pwr    : 1;   // 1 => OCC clipped max Psate due to reaching pcap limit.
+            uint8_t mthrot_due_to_ot   : 1;   // 1 => OCC throttled memory due to an over temp.
+            uint8_t n_power            : 1;   // 1 => Server running without redundant power.
+            uint8_t dvfs_due_to_vdd_ot : 1;   // 1 => OCC clipped max Pstate due to VRM Vdd over temp.
+            uint8_t sync_request       : 1;   // 1 => OCC needs to restart snapshot buffers
+            uint8_t _reserved_1        : 1;
+            uint8_t _reserved_0        : 1;
         };
         uint8_t word;
     } ext_status;
diff --git a/src/occ_405/cmdh/cmdh_fsp_cmds_datacnfg.c b/src/occ_405/cmdh/cmdh_fsp_cmds_datacnfg.c
index 52cb637..e60f3c7 100755
--- a/src/occ_405/cmdh/cmdh_fsp_cmds_datacnfg.c
+++ b/src/occ_405/cmdh/cmdh_fsp_cmds_datacnfg.c
@@ -1849,6 +1849,17 @@ errlHndl_t data_store_thrm_thresholds(const cmdh_fsp_cmd_t * i_cmd_ptr,
             break;
         }
 
+        // clear all FRU types to 0xFF (not defined) to prevent errors when (H)TMGT doesn't send thresholds for a FRU type
+        for(i=0; i<DATA_FRU_MAX; i++)
+        {
+            G_data_cnfg->thrm_thresh.data[i].fru_type = i;
+            G_data_cnfg->thrm_thresh.data[i].dvfs     = 0xFF;
+            G_data_cnfg->thrm_thresh.data[i].error    = 0xFF;
+            G_data_cnfg->thrm_thresh.data[i].pm_dvfs  = 0xFF;
+            G_data_cnfg->thrm_thresh.data[i].pm_error = 0xFF;
+            G_data_cnfg->thrm_thresh.data[i].max_read_timeout = 0xFF;
+        }
+
         // Store the base data
         G_data_cnfg->thrm_thresh.version          = l_cmd_ptr->version;
         G_data_cnfg->thrm_thresh.proc_core_weight = l_cmd_ptr->proc_core_weight;
diff --git a/src/occ_405/occ_sys_config.h b/src/occ_405/occ_sys_config.h
index f393256..6cfe2b1 100755
--- a/src/occ_405/occ_sys_config.h
+++ b/src/occ_405/occ_sys_config.h
@@ -74,10 +74,11 @@ typedef union
 {
     struct
     {
-        uint8_t     kvm:        1;
-        uint8_t     reserved:   5;
-        uint8_t     ite:        1;
-        uint8_t     single:     1;
+        uint8_t     kvm:              1;
+        uint8_t     reserved:         3;
+        uint8_t     report_dvfs_nom:  1;
+        uint8_t     reserved_2:       2;
+        uint8_t     single:           1;
     };
     uint8_t byte;
 } eSystemType;
@@ -332,7 +333,7 @@ typedef struct
 
   // Instead of system-type, lets try to send all system attributes
   // that matter instead of having tables in OCC code.
-  eSystemType system_type;  // OCC usage of this byte is TBD
+  eSystemType system_type;
 
   // Processor HUID - HUID for this OCC processor, used by OCC for processor error call out
   uint32_t proc_huid;