PRD: Add ECC checking for maint cmd complete attentions

Change-Id: I195e96ef91f495cbbc0cef262b5a040b24179d3f
RTC: 157892
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/32509
Tested-by: Jenkins Server <pfd-jenkins+hostboot@us.ibm.com>
Reviewed-by: Caleb N. Palmer <cnpalmer@us.ibm.com>
Reviewed-by: Benjamin J. Weisenbeck <bweisenb@us.ibm.com>
Reviewed-by: Zane C. Shelley <zshelle@us.ibm.com>
Reviewed-on: http://ralgit01.raleigh.ibm.com/gerrit1/33079
Tested-by: FSP CI Jenkins <fsp-CI-jenkins+hostboot@us.ibm.com>

5 files changed, 279 insertions, 5 deletions

diff --git a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.C b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.C
index 5c543ad79..6823e8254 100644
--- a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.C
+++ b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.C
@@ -83,16 +83,21 @@ uint32_t MemTdCtlr<T>::handleCmdComplete( STEP_CODE_DATA_STRUCT & io_sc )
                 break;
             }
 
-            // TODO: RTC 157892 Check why the command stopped and take actions
-            //       appropriately. Note that since nothing is happening here at
-            //       the moment, the code will simply assume the command stopped
-            //       at the end of memory with no errors.
+            // Then, check for ECC errors, if they exist.
+            bool errorsFound = false;
+            o_rc = checkEcc( errorsFound, io_sc );
+            if ( SUCCESS != o_rc )
+            {
+                PRDF_ERR( PRDF_FUNC "checkEcc(0x%08x) failed",
+                          iv_chip->getHuid() );
+                break;
+            }
 
             // If the command completed successfully with no error, the error
             // log will not have any useful information. Therefore, do not
             // commit the error log. This is done to avoid useless
             // informational error logs.
-            io_sc.service_data->setDontCommitErrl();
+            if ( !errorsFound ) io_sc.service_data->setDontCommitErrl();
         }
 
         // Move onto the next step in the state machine.
diff --git a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.H b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.H
index bd2f66718..c853b2031 100644
--- a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.H
+++ b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr.H
@@ -172,6 +172,17 @@ class MemTdCtlr
      */
     uint32_t initStoppedRank();
 
+    /**
+     * @brief  This is called when handling a command complete attention for a
+     *         non-TD command to check for ECC errors. This must be called after
+     *         initStoppedRank() to ensure iv_stoppedRank is initialized.
+     * @param  o_errorsFound True if errors where found and handled. False
+     *                       otherwise.
+     * @param  io_sc         The step code data struct.
+     * @return Non-SUCCESS if an internal function fails, SUCCESS otherwise.
+     */
+    uint32_t checkEcc( bool & o_errorsFound, STEP_CODE_DATA_STRUCT & io_sc );
+
     #ifdef __HOSTBOOT_RUNTIME
 
     /**
diff --git a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_ipl.C b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_ipl.C
index a34e4f199..4e49e423e 100644
--- a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_ipl.C
+++ b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_ipl.C
@@ -30,7 +30,11 @@
 #include <prdfMemTdCtlr.H>
 
 // Platform includes
+#include <prdfMemMark.H>
+#include <prdfMemoryMru.H>
 #include <prdfMemScrubUtils.H>
+#include <prdfMemVcm_ipl.H>
+#include <prdfP9McaExtraSig.H>
 
 using namespace TARGETING;
 
@@ -99,6 +103,174 @@ uint32_t MemTdCtlr<T>::defaultStep( STEP_CODE_DATA_STRUCT & io_sc )
 
 //------------------------------------------------------------------------------
 
+template <TARGETING::TYPE T>
+uint32_t __checkEcc( ExtensibleChip * i_chip, const MemRank & i_rank,
+                     TdQueue & io_queue, bool & o_errorsFound,
+                     STEP_CODE_DATA_STRUCT & io_sc )
+{
+    #define PRDF_FUNC "[__checkEcc] "
+
+    uint32_t o_rc = SUCCESS;
+
+    o_errorsFound = true; // Assume true for unless nothing found.
+
+    TargetHandle_t trgt = i_chip->getTrgt();
+    HUID           huid = i_chip->getHuid();
+
+    do
+    {
+        // Check for ECC errors.
+        uint32_t eccAttns = 0;
+        o_rc = checkEccFirs<T>( i_chip, eccAttns );
+        if ( SUCCESS != o_rc )
+        {
+            PRDF_ERR( PRDF_FUNC "checkEccFirs<T>(0x%08x) failed", huid );
+            break;
+        }
+
+        if ( 0 != (eccAttns & MAINT_UE) )
+        {
+            // Add the signature to the multi-signature list. Also, since
+            // this will be a predictive callout, change the primary
+            // signature as well.
+            io_sc.service_data->AddSignatureList( trgt, PRDFSIG_MaintUE );
+            io_sc.service_data->setSignature(     huid, PRDFSIG_MaintUE );
+
+            // Add the rank to the callout list.
+            MemoryMru mm { trgt, i_rank, MemoryMruData::CALLOUT_RANK };
+            io_sc.service_data->SetCallout( mm );
+
+            // Make the error log predictive.
+            io_sc.service_data->setServiceCall();
+        }
+        else if ( 0 != (eccAttns & MAINT_MPE) )
+        {
+            io_sc.service_data->AddSignatureList( trgt, PRDFSIG_MaintMPE );
+
+            // Read the chip mark from markstore.
+            MemMark chipMark;
+            o_rc = MarkStore::readChipMark<T>( i_chip, i_rank, chipMark );
+            if ( SUCCESS != o_rc )
+            {
+                PRDF_ERR( PRDF_FUNC "readChipMark<T>(0x%08x,%d) failed",
+                          huid, i_rank.getMaster() );
+                break;
+            }
+
+            // If the chip mark is not valid, then somehow the chip mark was
+            // placed on a rank other than the rank in which the command
+            // stopped. This would most likely be a code bug.
+            PRDF_ASSERT( chipMark.isValid() );
+
+            // Add the mark to the callout list.
+            MemoryMru mm { trgt, i_rank, chipMark.getSymbol() };
+            io_sc.service_data->SetCallout( mm );
+
+            // Add a new VCM procedure to the queue.
+            TdEntry * e = new VcmEvent<T>{ i_chip, i_rank };
+            io_queue.push( e );
+        }
+        else if ( isMfgCeCheckingEnabled() &&
+                (0 != (eccAttns & MAINT_HARD_NCE_ETE)) )
+        {
+            io_sc.service_data->AddSignatureList( trgt, PRDFSIG_MaintHARD_CTE );
+
+            // TODO RTC 136128
+            // - Query the per-symbol counters for the hard CE symbol (there
+            //   should be only one).
+            // - Add the symbol to the callout list (via MemoryMru).
+            // - Add a TPS procedure to the queue.
+        }
+        else // Nothing found.
+        {
+            o_errorsFound = false;
+        }
+
+    } while (0);
+
+    return o_rc;
+
+    #undef PRDF_FUNC
+}
+
+//------------------------------------------------------------------------------
+
+template <>
+uint32_t MemTdCtlr<TYPE_MCBIST>::checkEcc( bool & o_errorsFound,
+                                           STEP_CODE_DATA_STRUCT & io_sc )
+{
+    #define PRDF_FUNC "[MemTdCtlr<TYPE_MCBIST>::checkEcc] "
+
+    uint32_t o_rc = SUCCESS;
+
+    o_errorsFound = false;
+
+    MemRank rank = iv_stoppedRank.getRank();
+
+    do
+    {
+        // Get all ports in which the command was run.
+        std::vector<ExtensibleChip *> portList;
+        o_rc = getMcbistMaintPort( iv_chip, portList );
+        if ( SUCCESS != o_rc )
+        {
+            PRDF_ERR( PRDF_FUNC "getMcbistMaintPort(0x%08x) failed",
+                      iv_chip->getHuid() );
+            break;
+        }
+
+        // Check each MCA for ECC errors.
+        for ( auto & mcaChip : portList )
+        {
+            bool errorsFound;
+            uint32_t l_rc = __checkEcc<TYPE_MCA>( mcaChip, rank, iv_queue,
+                                                  errorsFound, io_sc );
+            if ( SUCCESS != l_rc )
+            {
+                PRDF_ERR( PRDF_FUNC "__checkEcc<TYPE_MCA>(0x%08x,%d) failed",
+                          mcaChip->getHuid(), rank.getMaster() );
+                o_rc |= l_rc; continue; // Try the other MCAs.
+            }
+
+            if ( errorsFound ) o_errorsFound = true;
+        }
+        if ( SUCCESS != o_rc ) break;
+
+    } while (0);
+
+    return o_rc;
+
+    #undef PRDF_FUNC
+}
+
+//------------------------------------------------------------------------------
+
+template <>
+uint32_t MemTdCtlr<TYPE_MBA>::checkEcc( bool & o_errorsFound,
+                                        STEP_CODE_DATA_STRUCT & io_sc )
+{
+    #define PRDF_FUNC "[MemTdCtlr<TYPE_MBA>::checkEcc] "
+
+    uint32_t o_rc = SUCCESS;
+
+    o_errorsFound = false;
+
+    MemRank rank = iv_stoppedRank.getRank();
+
+    o_rc = __checkEcc<TYPE_MBA>( iv_chip, rank, iv_queue, o_errorsFound, io_sc);
+    if ( SUCCESS != o_rc )
+    {
+        PRDF_ERR( PRDF_FUNC "__checkEcc<TYPE_MBA>(0x%08x,%d) failed",
+                  iv_chip->getHuid(), rank.getMaster() );
+    }
+
+    return o_rc;
+
+    #undef PRDF_FUNC
+}
+
+//------------------------------------------------------------------------------
+
 // Avoid linker errors with the template.
 template class MemTdCtlr<TYPE_MCBIST>;
 template class MemTdCtlr<TYPE_MBA>;
diff --git a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_rt.C b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_rt.C
index 9254d3ee2..f525acebd 100644
--- a/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_rt.C
+++ b/src/usr/diag/prdf/plat/mem/prdfMemTdCtlr_rt.C
@@ -151,6 +151,87 @@ uint32_t MemTdCtlr<T>::defaultStep( STEP_CODE_DATA_STRUCT & io_sc )
 
 //------------------------------------------------------------------------------
 
+template <>
+uint32_t MemTdCtlr<TYPE_MCBIST>::checkEcc( bool & o_errorsFound,
+                                           STEP_CODE_DATA_STRUCT & io_sc )
+{
+    #define PRDF_FUNC "[MemTdCtlr<TYPE_MCBIST>::checkEcc] "
+
+    uint32_t o_rc = SUCCESS;
+
+    o_errorsFound = false;
+
+    /* TODO: RTC 136126
+    MemRank rank = iv_stoppedRank.getRank();
+
+    do
+    {
+        // Get all ports in which the command was run.
+        std::vector<ExtensibleChip *> portList;
+        o_rc = getMcbistMaintPort( iv_chip, portList );
+        if ( SUCCESS != o_rc )
+        {
+            PRDF_ERR( PRDF_FUNC "getMcbistMaintPort(0x%08x) failed",
+                      iv_chip->getHuid() );
+            break;
+        }
+
+        // Check each MCA for ECC errors.
+        for ( auto & mcaChip : portList )
+        {
+            bool errorsFound;
+            uint32_t l_rc = __checkEcc<TYPE_MCA>( mcaChip, rank, iv_queue,
+                                                  io_sc, errorsFound );
+            if ( SUCCESS != l_rc )
+            {
+                PRDF_ERR( PRDF_FUNC "__checkEcc<TYPE_MCA>(0x%08x,%d) failed",
+                          mcaChip->getHuid(), rank.getMaster() );
+                o_rc |= l_rc; continue; // Try the other MCAs.
+            }
+
+            if ( errorsFound ) o_errorsFound = true;
+        }
+        if ( SUCCESS != o_rc ) break;
+
+    } while (0);
+    */
+
+    return o_rc;
+
+    #undef PRDF_FUNC
+}
+
+//------------------------------------------------------------------------------
+
+template <>
+uint32_t MemTdCtlr<TYPE_MBA>::checkEcc( bool & o_errorsFound,
+                                        STEP_CODE_DATA_STRUCT & io_sc )
+{
+    #define PRDF_FUNC "[MemTdCtlr<TYPE_MBA>::checkEcc] "
+
+    uint32_t o_rc = SUCCESS;
+
+    o_errorsFound = false;
+
+    /* TODO: RTC 136126
+    MemRank rank = iv_stoppedRank.getRank();
+
+    o_rc = __checkEcc<TYPE_MBA>( iv_chip, rank, iv_queue, io_sc,
+                                 o_errorsFound );
+    if ( SUCCESS != o_rc )
+    {
+        PRDF_ERR( PRDF_FUNC "__checkEcc<TYPE_MBA>(0x%08x,%d) failed",
+                  iv_chip->getHuid(), rank.getMaster() );
+    }
+    */
+
+    return o_rc;
+
+    #undef PRDF_FUNC
+}
+
+//------------------------------------------------------------------------------
+
 // Avoid linker errors with the template.
 template class MemTdCtlr<TYPE_MCBIST>;
 template class MemTdCtlr<TYPE_MBA>;
diff --git a/src/usr/diag/prdf/plat/mem/prdfMemTdQueue.H b/src/usr/diag/prdf/plat/mem/prdfMemTdQueue.H
index c46ef715c..5f33e5f3c 100755
--- a/src/usr/diag/prdf/plat/mem/prdfMemTdQueue.H
+++ b/src/usr/diag/prdf/plat/mem/prdfMemTdQueue.H
@@ -186,6 +186,11 @@ class TdQueue
         {
             iv_queue.push_back( i_e );
         }
+        else
+        {
+            // The event is already in the queue. So free up the memory.
+            delete i_e;
+        }
     }
 
     /**