Prevent calling out Centaurs on clock failover

CQ: SW437405

Change-Id: I1057d70bc6673b9d08a95573f00c9268f00dd126
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/62157
Tested-by: FSP CI Jenkins <fsp-CI-jenkins+hostboot@us.ibm.com>
Reviewed-by: Christopher J. Cain <cjcain@us.ibm.com>
Reviewed-by: Douglas R. Gilbert <dgilbert@us.ibm.com>
Reviewed-by: Martha Broyles <mbroyles@us.ibm.com>

6 files changed, 26 insertions, 89 deletions

diff --git a/src/occ_405/cent/centaur_control.c b/src/occ_405/cent/centaur_control.c
index e002bed..407bbdd 100755
--- a/src/occ_405/cent/centaur_control.c
+++ b/src/occ_405/cent/centaur_control.c
@@ -440,69 +440,4 @@ void centaur_control_init( void )
     return;
 }
 
-bool check_centaur_checkstop(memory_control_task_t * i_memControlTask )
-{
-    errlHndl_t l_err       = NULL;
-    int cent = i_memControlTask->curMemIndex;
-    // Check if the centaur has a channel checkstop. If it does,
-    // then do not log any errors. We also don't want to throttle
-    // a centaur that is in this condition.
-    if(G_centaur_control_reg_parms.error.rc != CENTAUR_CHANNEL_CHECKSTOP)
-    {
-        TRAC_ERR("task_memory_control: IPC_ST_CENTAUR_SCOM failed. "
-                 "cent=%d rc=%x, index=0x%08x",
-                 cent, G_centaur_control_reg_parms.error.rc,
-                 G_centaur_control_reg_parms.error.addr);
-
-        /* @
-         * @errortype
-         * @moduleid    CENT_CONTROL_MOD
-         * @reasoncode  CENT_SCOM_ERROR
-         * @userdata1   rc - Return code of scom operation
-         * @userdata2   index of scom operation that failed
-         * @userdata4   OCC_NO_EXTENDED_RC
-         * @devdesc     OCC access to centaur failed
-         */
-        l_err = createErrl(
-            CENT_CONTROL_MOD,                       // modId
-            CENT_SCOM_ERROR,                        // reasoncode
-            OCC_NO_EXTENDED_RC,                     // Extended reason code
-            ERRL_SEV_PREDICTIVE,                    // Severity
-            NULL,                                   // Trace Buf
-            DEFAULT_TRACE_SIZE,                     // Trace Size
-            G_centaur_control_reg_parms.error.rc,   // userdata1
-            G_centaur_control_reg_parms.error.addr  // userdata2
-            );
-
-        addUsrDtlsToErrl(l_err,                                  //io_err
-                         (uint8_t *) &(i_memControlTask->gpe_req.ffdc),    //i_dataPtr,
-                         sizeof(GpeFfdc),                                  //i_size
-                         ERRL_USR_DTL_STRUCT_VERSION_1,                    //version
-                         ERRL_USR_DTL_BINARY_DATA);                        //type
-
-        //callout the centaur
-        addCalloutToErrl(l_err,
-                         ERRL_CALLOUT_TYPE_HUID,
-                         G_sysConfigData.centaur_huids[cent],
-                         ERRL_CALLOUT_PRIORITY_MED);
-
-        //callout the processor
-        addCalloutToErrl(l_err,
-                         ERRL_CALLOUT_TYPE_HUID,
-                         G_sysConfigData.proc_huid,
-                         ERRL_CALLOUT_PRIORITY_MED);
-
-        commitErrl(&l_err);
-
-        return FALSE; // error was not a channel checkstop
-    }
-    else
-    {
-        // Remove the centaur sensor and all dimm sensors behind it.
-        cent_chan_checkstop(cent);
-    }
-    return TRUE;    // Centaur channel checkstop
-
-}
-
 
diff --git a/src/occ_405/cent/centaur_control.h b/src/occ_405/cent/centaur_control.h
index b6bb817..08832de 100755
--- a/src/occ_405/cent/centaur_control.h
+++ b/src/occ_405/cent/centaur_control.h
@@ -67,6 +67,4 @@ bool centaur_control( memory_control_task_t * i_memControlTask );
 //void centaur_control_init( void ) INIT_SECTION;
 void centaur_control_init( void );
 
-bool check_centaur_checkstop( memory_control_task_t * i_memControlTask );
-
 #endif //_CENTAUR_CONTROL_H
diff --git a/src/occ_405/cent/centaur_data.c b/src/occ_405/cent/centaur_data.c
index 8b08030..ec27b9d 100755
--- a/src/occ_405/cent/centaur_data.c
+++ b/src/occ_405/cent/centaur_data.c
@@ -328,7 +328,9 @@ void cent_recovery(uint32_t i_cent)
                                  ERRL_CALLOUT_TYPE_HUID,
                                  G_sysConfigData.proc_huid,
                                  ERRL_CALLOUT_PRIORITY_MED);
-                commitErrl(&l_err);
+
+                // recovery is failing, ask for OCC reset to try to recover
+                REQUEST_RESET(l_err);
             }
         }
 
@@ -671,9 +673,9 @@ void centaur_data( void )
                 else  // log the error if it was not a CENTAUR_CHANNEL_CHECKSTOP
                 {
                     //log an error the first time this happens but keep on running.
+                    //This should be informational (except mfg) since we are going to retry
                     //eventually, we will timeout on the dimm & centaur temps not being updated
-                    //and fans will go to max speed (probably won't be able to throttle for
-                    //same reason we can't access the centaur here).
+                    //if this is a hard failure which will call out the Centaur at that point.  
                     if(!L_gpe_error_logged)
                     {
                         L_gpe_error_logged = TRUE;
@@ -696,13 +698,16 @@ void centaur_data( void )
                                 CENT_TASK_DATA_MOD,                     //modId
                                 CENT_SCOM_ERROR,                        //reasoncode
                                 OCC_NO_EXTENDED_RC,                     //Extended reason code
-                                ERRL_SEV_PREDICTIVE,                    //Severity
+                                ERRL_SEV_INFORMATIONAL,                 //Severity
                                 NULL,                                   //Trace Buf
                                 DEFAULT_TRACE_SIZE,                     //Trace Size
                                 l_parms->error.rc,                            //userdata1
                                 0                                       //userdata2
                                 );
 
+                        //force severity to predictive if mfg ipl (allows callout to be added to info error)
+                        setErrlActions(l_err, ERRL_ACTIONS_MANUFACTURING_ERROR);
+
                         addUsrDtlsToErrl(l_err,                                   //io_err
                                 (uint8_t *) &(l_centaur_data_ptr->gpe_req.ffdc),  //i_dataPtr,
                                 sizeof(GpeFfdc),                                  //i_size
diff --git a/src/occ_405/main.c b/src/occ_405/main.c
index 2e36807..623fc87 100755
--- a/src/occ_405/main.c
+++ b/src/occ_405/main.c
@@ -1597,11 +1597,12 @@ void Main_thread_routine(void *private)
         // Look for FIR collection flag and status
         if (G_fir_collection_required && !L_fir_collection_completed)
         {
-            TRAC_IMP("fir data collection starting");
             // If this OCC is the FIR master and PNOR access is allowed perform
             // FIR collection
             if (OCC_IS_FIR_MASTER())
             {
+                TRAC_IMP("fir data collection starting");
+
                 //Need to schedule a task on GPE to start fir collection
                 if(!G_fir_collection_request_created) //Only need to create request once
                 {
@@ -1629,6 +1630,7 @@ void Main_thread_routine(void *private)
                         G_fir_collection_required = FALSE;
                     }
                 }
+                TRAC_IMP("fir data collection done");
             }
 
             // Error reporting is skipped while FIR collection is required so we
@@ -1640,7 +1642,6 @@ void Main_thread_routine(void *private)
             {
                 notify_host(INTR_REASON_HTMGT_SERVICE_REQUIRED);
             }
-            TRAC_IMP("fir data collection done");
         }
 
         if( l_ssxrc == SSX_OK)
diff --git a/src/occ_405/mem/memory.c b/src/occ_405/mem/memory.c
index d485fc8..f7f692d 100644
--- a/src/occ_405/mem/memory.c
+++ b/src/occ_405/mem/memory.c
@@ -87,7 +87,6 @@ void task_memory_control( task_t * i_task )
     int            rc        = 0;       // Return code
     uint8_t        memIndex;
     static bool    L_gpe_scheduled   = FALSE;
-    static uint8_t L_gpe_fail_logged = 0;
     static bool    L_gpe_idle_traced = FALSE;
     static bool    L_gpe_had_1_tick  = FALSE;
 
@@ -147,24 +146,23 @@ void task_memory_control( task_t * i_task )
         {
             if(!async_request_completed(&memControlTask->gpe_req.request) || gpe_rc)
             {
-                if (MEM_TYPE_CUMULUS ==  G_sysConfigData.mem_type)
+                // ignore error and stop monitoring this centaur if there is a channel checkstop
+                if( (MEM_TYPE_CUMULUS ==  G_sysConfigData.mem_type) &&
+                    (gpe_rc == CENTAUR_CHANNEL_CHECKSTOP) )
                 {
-                    if(!(L_gpe_fail_logged & (CENTAUR0_PRESENT_MASK >> memIndex)))
-                    {
-                        if (!check_centaur_checkstop(memControlTask))
-                        {
-                            L_gpe_fail_logged |= CENTAUR0_PRESENT_MASK >> memIndex;
-                        }
-                    }
+                    // Remove the centaur sensor and all dimm sensors behind it.
+                    cent_chan_checkstop(memControlTask->curMemIndex);
                 }
-                //Request failed. Keep count of failures and request a reset if we reach a
-                //max retry count
-                L_scom_timeout[memIndex]++;
-                if(L_scom_timeout[memIndex] == MEMORY_CONTROL_SCOM_TIMEOUT)
+                else
                 {
-                    break;
+                    //Request failed. Keep count of failures and request a reset if we reach a
+                    //max retry count
+                    L_scom_timeout[memIndex]++;
+                    if(L_scom_timeout[memIndex] == MEMORY_CONTROL_SCOM_TIMEOUT)
+                    {
+                        break;
+                    }
                 }
-
             }//if(!async_request_completed(&memControlTask->gpe_req.request) || l_parms->rc)
             else
             {
diff --git a/src/occ_405/occbuildname.c b/src/occ_405/occbuildname.c
index 21c2ecf..eb61723 100755
--- a/src/occ_405/occbuildname.c
+++ b/src/occ_405/occbuildname.c
@@ -34,6 +34,6 @@ volatile const char G_occ_buildname[16] __attribute__((section(".buildname"))) =
 
 #else
 
-volatile const char G_occ_buildname[16] __attribute__((section(".buildname"))) = /*<BuildName>*/  "op_occ_180629a\0"  /*</BuildName>*/ ;
+volatile const char G_occ_buildname[16] __attribute__((section(".buildname"))) = /*<BuildName>*/  "op_occ_180711a\0"  /*</BuildName>*/ ;
 
 #endif