Fix DIMM temperature error handling for poll response

RTC:155187

Change-Id: I38039dc18de9bfc5b9194f63b3b869bf7c16991f
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/26067
Tested-by: FSP CI Jenkins <fsp-CI-jenkins+hostboot@us.ibm.com>
Reviewed-by: Christopher J. Cain <cjcain@us.ibm.com>
Reviewed-by: Martha Broyles <mbroyles@us.ibm.com>
Reviewed-by: William A. Bryan <wilbryan@us.ibm.com>

4 files changed, 51 insertions, 43 deletions

diff --git a/src/occ_405/amec/amec_controller.c b/src/occ_405/amec/amec_controller.c
index d0ecfc3..fe27587 100644
--- a/src/occ_405/amec/amec_controller.c
+++ b/src/occ_405/amec/amec_controller.c
@@ -5,7 +5,7 @@
 /*                                                                        */
 /* OpenPOWER OnChipController Project                                     */
 /*                                                                        */
-/* Contributors Listed Below - COPYRIGHT 2011,2015                        */
+/* Contributors Listed Below - COPYRIGHT 2011,2016                        */
 /* [+] International Business Machines Corp.                              */
 /*                                                                        */
 /*                                                                        */
@@ -29,11 +29,12 @@
 #include <occ_common.h>
 #include <sensor.h>
 #include <amec_sys.h>
+#include <centaur_data.h>
 
 //*************************************************************************
 // Externs
 //*************************************************************************
-extern uint8_t G_dimm_temp_expired_bitmap;
+extern cent_sensor_flags_t G_dimm_temp_expired_bitmap;
 extern uint8_t G_cent_temp_expired_bitmap;
 //*************************************************************************
 // Macros
@@ -178,7 +179,7 @@ void amec_controller_dimm_thermal()
     // Get TEMP2MSDIMM sensor value
     l_sensor = getSensorByGsid(TEMP2MSDIMM);
 
-    if(G_dimm_temp_expired_bitmap)
+    if(G_dimm_temp_expired_bitmap.bigword)
     {
        //we were not able to read one or more dimm temperatures.
        //Assume temperature is at the setpoint plus 1 degree C.
diff --git a/src/occ_405/amec/amec_freq.c b/src/occ_405/amec/amec_freq.c
index 0afe676..93a43fb 100755
--- a/src/occ_405/amec/amec_freq.c
+++ b/src/occ_405/amec/amec_freq.c
@@ -1,11 +1,11 @@
 /* IBM_PROLOG_BEGIN_TAG                                                   */
 /* This is an automatically generated prolog.                             */
 /*                                                                        */
-/* $Source: src/occ/amec/amec_freq.c $                                    */
+/* $Source: src/occ_405/amec/amec_freq.c $                                */
 /*                                                                        */
 /* OpenPOWER OnChipController Project                                     */
 /*                                                                        */
-/* Contributors Listed Below - COPYRIGHT 2011,2015                        */
+/* Contributors Listed Below - COPYRIGHT 2011,2016                        */
 /* [+] International Business Machines Corp.                              */
 /*                                                                        */
 /*                                                                        */
@@ -48,12 +48,13 @@
 #include <amec_data.h>
 #include <amec_freq.h>
 #include "pss_constants.h"
+#include <centaur_data.h>
 
 //*************************************************************************
 // Externs
 //*************************************************************************
 extern uint8_t G_cent_temp_expired_bitmap;
-extern uint8_t G_dimm_temp_expired_bitmap;
+extern cent_sensor_flags_t G_dimm_temp_expired_bitmap;
 
 //*************************************************************************
 // Defines/Enums
@@ -78,7 +79,7 @@ const uint32_t G_pmc_ffdc_scom_addrs[PMC_FFDC_SCOM_ADDRS_SIZE] =
     PMC_LFIR_ERR_MASK_REG,
     OCB_OCCLFIR,
     PBA_FIR,
-    TOD_VALUE_REG 
+    TOD_VALUE_REG
 };
 
 //FFDC OCI addresses as requested by Greg Still in defect SW247927
@@ -618,14 +619,12 @@ void amec_slv_mem_voting_box(void)
         if(!L_throttle_traced)
         {
             L_throttle_traced = TRUE;
-// @TODO - TEMP - No dimm temp Colection yet
-/*
-            TRAC_INFO("Memory is being throttled. reason[%d] vote[%d] cent_expired[0x%02x] dimm_expired[0x%02x]",
+            TRAC_INFO("Memory is being throttled. reason[%d] vote[%d] cent_expired[0x%02x] dimm_expired[0x%08x%08x]",
                        l_reason,
                        l_vote,
                        G_cent_temp_expired_bitmap,
-                       G_dimm_temp_expired_bitmap);
-*/
+                       G_dimm_temp_expired_bitmap.words[0],
+                       G_dimm_temp_expired_bitmap.words[1]);
         }
     }
     else
diff --git a/src/occ_405/amec/amec_health.c b/src/occ_405/amec/amec_health.c
index 1e061bf..a559f04 100755
--- a/src/occ_405/amec/amec_health.c
+++ b/src/occ_405/amec/amec_health.c
@@ -23,9 +23,9 @@
 /*                                                                        */
 /* IBM_PROLOG_END_TAG                                                     */
 
-//*************************************************************************
+//*************************************************************************/
 // Includes
-//*************************************************************************
+//*************************************************************************/
 #include "amec_health.h"
 #include "amec_sys.h"
 #include "amec_service_codes.h"
@@ -33,17 +33,17 @@
 #include <centaur_data.h>
 #include <proc_data.h>
 
-//*************************************************************************
+//*************************************************************************/
 // Externs
-//*************************************************************************
+//*************************************************************************/
 
-//*************************************************************************
+//*************************************************************************/
 // Defines/Enums
-//*************************************************************************
+//*************************************************************************/
 
-//*************************************************************************
+//*************************************************************************/
 // Globals
-//*************************************************************************
+//*************************************************************************/
 
 // Have we already called out the dimm for overtemp (bitmap of dimms)?
 cent_sensor_flags_t G_dimm_overtemp_logged_bitmap = {0};
@@ -51,9 +51,8 @@ cent_sensor_flags_t G_dimm_overtemp_logged_bitmap = {0};
 // Have we already called out the dimm for timeout (bitmap of dimms)?
 cent_sensor_flags_t G_dimm_timeout_logged_bitmap = {0};
 
-// Are any dimms currently in the timedout state (bitmap of centaurs)?
-// Note: this only tells you which centaur, not which dimm.
-uint8_t             G_dimm_temp_expired_bitmap = 0;
+// Are any dimms currently in the timedout state (bitmap of dimm)?
+cent_sensor_flags_t G_dimm_temp_expired_bitmap = {0};
 
 // Have we already called out the centaur for timeout (bitmap of centaurs)?
 uint8_t G_cent_timeout_logged_bitmap = 0;
@@ -67,13 +66,13 @@ uint8_t G_cent_temp_expired_bitmap = 0;
 // Array to store the update tag of each core's temperature sensor
 uint32_t G_core_temp_update_tag[MAX_NUM_CORES] = {0};
 
-//*************************************************************************
+//*************************************************************************/
 // Function Declarations
-//*************************************************************************
+//*************************************************************************/
 
-//*************************************************************************
+//*************************************************************************/
 // Functions
-//*************************************************************************
+//*************************************************************************/
 uint64_t amec_mem_get_huid(uint8_t i_cent, uint8_t i_dimm)
 {
     uint64_t l_huid;
@@ -380,16 +379,17 @@ void amec_health_check_dimm_timeout()
             break;
         }
 
-        //iterate across all centaurs incrementing dimm sensor timers as needed
+        //iterate across all centaurs/ports incrementing dimm sensor timers as needed
         for(l_cent = 0; l_cent < MAX_NUM_CENTAURS; l_cent++)
         {
             //any dimm timers behind this centaur need incrementing?
             if(!l_need_inc.bytes[l_cent])
             {
-                //all dimm sensors were updated for this centaur. Clear the dimm timeout bit for this centaur.
-                if(G_dimm_temp_expired_bitmap & (CENTAUR0_PRESENT_MASK >> l_cent))
+                // All dimm sensors were updated for this centaur/port
+                // Trace this fact and clear the expired byte for all DIMMs on this centaur/port
+                if(G_dimm_temp_expired_bitmap.bytes[l_cent])
                 {
-                    G_dimm_temp_expired_bitmap &= ~(CENTAUR0_PRESENT_MASK >> l_cent);
+                    G_dimm_temp_expired_bitmap.bytes[l_cent] = 0;
                     TRAC_INFO("All dimm sensors for centaur %d have been updated", l_cent);
                 }
                 continue;
@@ -398,9 +398,14 @@ void amec_health_check_dimm_timeout()
             //There's at least one dimm requiring an increment, find the dimm
             for(l_dimm = 0; l_dimm < NUM_DIMMS_PER_CENTAUR; l_dimm++)
             {
-                //not this one, go to next one
+                //not this one, check if we need to clear the dimm timeout and go to the next one
                 if(!(l_need_inc.bytes[l_cent] & (DIMM_SENSOR0 >> l_dimm)))
                 {
+                    // Clear this one if needed
+                    if(G_dimm_temp_expired_bitmap.bytes[l_cent] & (DIMM_SENSOR0 >> l_dimm))
+                    {
+                        G_dimm_temp_expired_bitmap.bytes[l_cent] &= ~(DIMM_SENSOR0 >> l_dimm);
+                    }
                     continue;
                 }
 
@@ -430,12 +435,12 @@ void amec_health_check_dimm_timeout()
                     continue;
                 }
 
-                //temperature has expired.  Notify control algorithms which centaur.
-                if(!(G_dimm_temp_expired_bitmap & (CENTAUR0_PRESENT_MASK >> l_cent)))
+                //temperature has expired.  Notify control algorithms which DIMM
+                if(!(G_dimm_temp_expired_bitmap.bytes[l_cent] & (DIMM_SENSOR0 >> l_dimm)))
                 {
-                    G_dimm_temp_expired_bitmap |= CENTAUR0_PRESENT_MASK >> l_cent;
-                    TRAC_ERR("Timed out reading dimm temperature sensor on cent %d.",
-                             l_cent);
+                    G_dimm_temp_expired_bitmap.bytes[l_cent] |= (DIMM_SENSOR0 >> l_dimm);
+                    TRAC_ERR("Timed out reading dimm temperature sensor on cent %d dimm %d.",
+                             l_cent, l_dimm);
                 }
 
                 //If we've already logged an error for this FRU go to the next one.
@@ -460,7 +465,7 @@ void amec_health_check_dimm_timeout()
                      *
                      */
                     l_err = createErrl(AMEC_HEALTH_CHECK_DIMM_TIMEOUT,    //modId
-                                       FRU_TEMP_TIMEOUT,                 //reasoncode
+                                       FRU_TEMP_TIMEOUT,                  //reasoncode
                                        OCC_NO_EXTENDED_RC,                //Extended reason code
                                        ERRL_SEV_PREDICTIVE,               //Severity
                                        NULL,                              //Trace Buf
@@ -496,7 +501,7 @@ void amec_health_check_dimm_timeout()
                                      &G_dimm_timeout_logged_bitmap.bytes[l_cent]);
             } //iterate over all dimms
 
-        } //iterate over all centaurs
+        } //iterate over all centaurs/ports
 
         if(l_err)
         {
@@ -509,7 +514,7 @@ void amec_health_check_dimm_timeout()
             break;
         }
 
-        //iterate across all centaurs clearing dimm sensor timers as needed
+        //iterate across all centaurs/ports clearing dimm sensor timers as needed
         for(l_cent = 0; l_cent < MAX_NUM_CENTAURS; l_cent++)
         {
 
@@ -541,7 +546,7 @@ void amec_health_check_dimm_timeout()
                 }
 
             }//iterate over all dimms
-        }//iterate over all centaurs
+        }//iterate over all centaurs/ports
     }while(0);
     L_ran_once = TRUE;
 }
diff --git a/src/occ_405/cmdh/cmdh_fsp_cmds.c b/src/occ_405/cmdh/cmdh_fsp_cmds.c
index 3b7da38..48d1d5d 100755
--- a/src/occ_405/cmdh/cmdh_fsp_cmds.c
+++ b/src/occ_405/cmdh/cmdh_fsp_cmds.c
@@ -46,6 +46,9 @@
 #include "amec_master_smh.h"
 #include <proc_data.h>
 #include "homer.h"
+#include <centaur_data.h>
+
+extern cent_sensor_flags_t G_dimm_temp_expired_bitmap;
 
 // This table contains tunable parameter information that can be exposed to
 // customers (only Master OCC should access/control this table)
@@ -261,7 +264,7 @@ ERRL_RC cmdh_poll_v20(cmdh_fsp_rsp_t * o_rsp_ptr)
                     l_tempSensorList[l_sensorHeader.count].id = g_amec->proc[0].memctl[l_port].centaur.dimm_temps[l_dimm].temp_sid;
 
                     //If a dimm timed out long enough, we should return 0xFFFF for that sensor.
-                    if (G_dimm_timeout_logged_bitmap.bytes[l_port] & (DIMM_SENSOR0 >> l_dimm))
+                    if (G_dimm_temp_expired_bitmap.bytes[l_port] & (DIMM_SENSOR0 >> l_dimm))
                     {
                         l_tempSensorList[l_sensorHeader.count].value = 0xFFFF;
                     }
@@ -300,7 +303,7 @@ ERRL_RC cmdh_poll_v20(cmdh_fsp_rsp_t * o_rsp_ptr)
                     {
                         l_tempSensorList[l_sensorHeader.count].id = g_amec->proc[0].memctl[l_cent].centaur.dimm_temps[l_dimm].temp_sid;
                         //If a dimm timed out long enough, we should return 0xFFFF for that sensor.
-                        if (G_dimm_timeout_logged_bitmap.bytes[l_cent] & (DIMM_SENSOR0 >> l_dimm))
+                        if (G_dimm_temp_expired_bitmap.bytes[l_cent] & (DIMM_SENSOR0 >> l_dimm))
                         {
                             l_tempSensorList[l_sensorHeader.count].value = 0xFFFF;
                         }