Merge mysql.com:/space/pekka/ndb/version/my51

into  mysql.com:/space/pekka/ndb/version/my51-rbr

storage/ndb/include/ndbapi/NdbDictionary.hpp

@@ -1027,7 +1027,8 @@ public:
       _TE_CREATE=6,
       _TE_GCP_COMPLETE=7,
       _TE_CLUSTER_FAILURE=8,
-      _TE_STOP=9
+      _TE_STOP=9,
+      _TE_NUL=10 // internal (INS o DEL within same GCI)
     };
 #endif
     /**

storage/ndb/include/ndbapi/NdbEventOperation.hpp

@@ -93,6 +93,12 @@ public:
    * Retrieve current state of the NdbEventOperation object
    */
   State getState();
+  /**
+   * By default events on same NdbEventOperation within same GCI
+   * are merged into a single event.  This can be changed with
+   * separateEvents(true).
+   */
+  void separateEvents(bool flag);
 
   /**
    * Activates the NdbEventOperation to start receiving events. The

storage/ndb/ndbapi-examples/ndbapi_event/Makefile

 TARGET = ndbapi_event
 SRCS = ndbapi_event.cpp
 OBJS = ndbapi_event.o
-CXX = g++
+CXX = g++ -g
 CFLAGS = -c -Wall -fno-rtti -fno-exceptions
 CXXFLAGS = 
 DEBUG = 
@@ -17,7 +17,7 @@ $(TARGET): $(OBJS)
 	$(CXX) $(CXXFLAGS) $(LFLAGS) $(LIB_DIR) $(OBJS) -lndbclient -lmysqlclient_r -lmysys -lmystrings -lz $(SYS_LIB) -o $(TARGET)
 
 $(TARGET).o: $(SRCS)
-	$(CXX) $(CFLAGS) -I$(INCLUDE_DIR) -I$(INCLUDE_DIR)/ndbapi $(SRCS)
+	$(CXX) $(CFLAGS) -I$(INCLUDE_DIR) -I$(INCLUDE_DIR)/ndbapi -I$(TOP_SRCDIR)/include $(SRCS)
 
 clean:
 	rm -f *.o $(TARGET)

storage/ndb/ndbapi-examples/ndbapi_event/ndbapi_event.cpp

@@ -58,24 +58,29 @@
 
 /**
  *
- * Assume that there is a table TAB0 which is being updated by 
+ * Assume that there is a table t0 which is being updated by 
  * another process (e.g. flexBench -l 0 -stdtables).
- * We want to monitor what happens with columns COL0, COL2, COL11
+ * We want to monitor what happens with columns c0,c1,c2,c3.
  *
  * or together with the mysql client;
  *
  * shell> mysql -u root
  * mysql> create database TEST_DB;
  * mysql> use TEST_DB;
- * mysql> create table TAB0 (COL0 int primary key, COL1 int, COL11 int) engine=ndb;
+ * mysql> create table t0 (c0 int, c1 int, c2 char(4), c3 char(4),
+ *        primary key(c0, c2)) engine ndb charset latin1;
  *
  * In another window start ndbapi_event, wait until properly started
  *
-   insert into TAB0 values (1,2,3);
-   insert into TAB0 values (2,2,3);
-   insert into TAB0 values (3,2,9);
-   update TAB0 set COL1=10 where COL0=1;
-   delete from TAB0 where COL0=1;
+   insert into t0 values (1, 2, 'a', 'b');
+   insert into t0 values (3, 4, 'c', 'd');
+   update t0 set c3 = 'e' where c0 = 1 and c2 = 'a'; -- use pk
+   update t0 set c3 = 'f'; -- use scan
+   update t0 set c3 = 'F'; -- use scan update to 'same'
+   update t0 set c2 = 'g' where c0 = 1; -- update pk part
+   update t0 set c2 = 'G' where c0 = 1; -- update pk part to 'same'
+   update t0 set c0 = 5, c2 = 'H' where c0 = 3; -- update full PK
+   delete from t0;
  *
  * you should see the data popping up in the example window
  *
@@ -92,9 +97,10 @@ int myCreateEvent(Ndb* myNdb,
 		  const char **eventColumnName,
 		  const int noEventColumnName);
 
-int main()
+int main(int argc, char** argv)
 {
   ndb_init();
+  bool sep = argc > 1 && strcmp(argv[1], "-s") == 0;
 
   Ndb_cluster_connection *cluster_connection=
     new Ndb_cluster_connection(); // Object representing the cluster
@@ -126,13 +132,15 @@ int main()
 
   if (myNdb->init() == -1) APIERROR(myNdb->getNdbError());
 
-  const char *eventName= "CHNG_IN_TAB0";
-  const char *eventTableName= "TAB0";
-  const int noEventColumnName= 3;
+  const char *eventName= "CHNG_IN_t0";
+  const char *eventTableName= "t0";
+  const int noEventColumnName= 4;
   const char *eventColumnName[noEventColumnName]=
-    {"COL0",
-     "COL1",
-     "COL11"};
+    {"c0",
+     "c1",
+     "c2",
+     "c3"
+    };
   
   // Create events
   myCreateEvent(myNdb,
@@ -142,13 +150,14 @@ int main()
 		noEventColumnName);
 
   int j= 0;
-  while (j < 5) {
+  while (j < 99) {
 
     // Start "transaction" for handling events
     NdbEventOperation* op;
     printf("create EventOperation\n");
     if ((op = myNdb->createEventOperation(eventName)) == NULL)
       APIERROR(myNdb->getNdbError());
+    op->separateEvents(sep);
 
     printf("get values\n");
     NdbRecAttr* recAttr[noEventColumnName];
@@ -175,32 +184,43 @@ int main()
 	  i++;
 	  switch (op->getEventType()) {
 	  case NdbDictionary::Event::TE_INSERT:
-	    printf("%u INSERT: ", i);
+	    printf("%u INSERT", i);
 	    break;
 	  case NdbDictionary::Event::TE_DELETE:
-	    printf("%u DELETE: ", i);
+	    printf("%u DELETE", i);
 	    break;
 	  case NdbDictionary::Event::TE_UPDATE:
-	    printf("%u UPDATE: ", i);
+	    printf("%u UPDATE", i);
 	    break;
 	  default:
 	    abort(); // should not happen
 	  }
-	  for (int i = 1; i < noEventColumnName; i++) {
+          printf(" gci=%d\n", op->getGCI());
+          printf("post:  ");
+	  for (int i = 0; i < noEventColumnName; i++) {
 	    if (recAttr[i]->isNULL() >= 0) { // we have a value
-	      printf(" post[%u]=", i);
-	      if (recAttr[i]->isNULL() == 0) // we have a non-null value
-		printf("%u", recAttr[i]->u_32_value());
-	      else                           // we have a null value
-		printf("NULL");
+	      if (recAttr[i]->isNULL() == 0) { // we have a non-null value
+                if (i < 2)
+                  printf("%-5u", recAttr[i]->u_32_value());
+                else
+                  printf("%-5.4s", recAttr[i]->aRef());
+              } else                           // we have a null value
+		printf("%-5s", "NULL");
+	    } else
+              printf("%-5s", "-");
           }
+          printf("\npre :  ");
+	  for (int i = 0; i < noEventColumnName; i++) {
 	    if (recAttrPre[i]->isNULL() >= 0) { // we have a value
-	      printf(" pre[%u]=", i);
-	      if (recAttrPre[i]->isNULL() == 0) // we have a non-null value
-		printf("%u", recAttrPre[i]->u_32_value());
-	      else                              // we have a null value
-		printf("NULL");
-	    }
+	      if (recAttrPre[i]->isNULL() == 0) { // we have a non-null value
+                if (i < 2)
+                  printf("%-5u", recAttrPre[i]->u_32_value());
+                else
+                  printf("%-5.4s", recAttrPre[i]->aRef());
+              } else                              // we have a null value
+		printf("%-5s", "NULL");
+	    } else
+              printf("%-5s", "-");
 	  }
           printf("\n");
 	}

storage/ndb/src/ndbapi/NdbEventOperation.cpp

@@ -38,6 +38,11 @@ NdbEventOperation::State NdbEventOperation::getState()
   return m_impl.getState();
 }
 
+void NdbEventOperation::separateEvents(bool flag)
+{
+  m_impl.m_separateEvents = flag;
+}
+
 NdbRecAttr *
 NdbEventOperation::getValue(const char *colName, char *aValue)
 {

storage/ndb/src/ndbapi/NdbEventOperationImpl.cpp

@@ -104,6 +104,8 @@ NdbEventOperationImpl::NdbEventOperationImpl(NdbEventOperation &N,
 
   m_state= EO_CREATED;
 
+  m_separateEvents = false;
+
   m_has_error= 0;
 
   DBUG_PRINT("exit",("this: 0x%x oid: %u", this, m_oid));
@@ -693,6 +695,21 @@ NdbEventBuffer::pollEvents(int aMillisecondNumber, Uint64 *latestGCI)
   return ret;
 }
 
+#ifdef VM_TRACE
+static void
+print_std(const char* tag, const SubTableData * sdata, LinearSectionPtr ptr[3])
+{
+  printf("%s\n", tag);
+  printf("addr=%p gci=%d op=%d\n", (void*)sdata, sdata->gci, sdata->operation);
+  for (int i = 0; i <= 2; i++) {
+    printf("sec=%d addr=%p sz=%d\n", i, (void*)ptr[i].p, ptr[i].sz);
+    for (int j = 0; j < ptr[i].sz; j++)
+      printf("%08x ", ptr[i].p[j]);
+    printf("\n");
+  }
+}
+#endif
+
 NdbEventOperation *
 NdbEventBuffer::nextEvent()
 {
@@ -734,6 +751,10 @@ NdbEventBuffer::nextEvent()
     op->m_data_done_count++;
 #endif
 
+    // NUL event is not returned
+    if (data->sdata->operation == NdbDictionary::Event::_TE_NUL)
+      continue;
+
     int r= op->receive_event();
     if (r > 0)
     {
@@ -1099,13 +1120,15 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
   DBUG_ENTER("NdbEventBuffer::insertDataL");
 
   Uint64 gci= sdata->gci;
-  EventBufData *data= m_free_data;
 
   if ( likely((Uint32)op->mi_type & 1 << (Uint32)sdata->operation) )
   {
     Gci_container* bucket= find_bucket(&m_active_gci, gci);
       
     DBUG_PRINT("info", ("data insertion in eventId %d", op->m_eventId));
+    DBUG_PRINT("info", ("gci=%d tab=%d op=%d node=%d",
+                        sdata->gci, sdata->tableId, sdata->operation,
+                        sdata->req_nodeid));
 
     if (unlikely(bucket == 0))
     {
@@ -1116,6 +1139,84 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
       DBUG_RETURN(0);
     }
 
+    bool use_hash =
+      ! op->m_separateEvents &&
+      sdata->operation < NdbDictionary::Event::_TE_FIRST_NON_DATA_EVENT;
+
+    // find position in bucket hash table
+    EventBufData* data = 0;
+    EventBufData_hash::Pos hpos;
+    if (use_hash)
+    {
+      bucket->m_data_hash.search(hpos, op, ptr);
+      data = hpos.data;
+    }
+
+    if (data == 0)
+    {
+      // allocate new result buffer
+      data = alloc_data();
+      if (unlikely(data == 0))
+      {
+        op->m_has_error = 2;
+        DBUG_RETURN(-1);
+      }
+
+      if (unlikely(copy_data(sdata, ptr, data)))
+      {
+        op->m_has_error = 3;
+        DBUG_RETURN(-1);
+      }
+      // add it to list and hash table
+      bucket->m_data.append(data);
+      if (use_hash)
+      {
+        bucket->m_data_hash.append(hpos, data);
+      }
+#ifdef VM_TRACE
+      op->m_data_count++;
+#endif
+    }
+    else
+    {
+      // event with same op, PK found, merge into old buffer
+      if (unlikely(merge_data(sdata, ptr, data)))
+      {
+        op->m_has_error = 3;
+        DBUG_RETURN(-1);
+      }
+    }
+    data->m_event_op = op;
+    if (use_hash)
+    {
+      data->m_pkhash = hpos.pkhash;
+    }
+    DBUG_RETURN(0);
+  }
+
+#ifdef VM_TRACE
+  if ((Uint32)op->m_eventImpl->mi_type & 1 << (Uint32)sdata->operation)
+  {
+    // XXX never reached
+    DBUG_PRINT("info",("Data arrived before ready eventId", op->m_eventId));
+    DBUG_RETURN(0);
+  }
+  else {
+    DBUG_PRINT("info",("skipped"));
+    DBUG_RETURN(0);
+  }
+#else
+  return 0;
+#endif
+}
+
+// allocate EventBufData
+EventBufData*
+NdbEventBuffer::alloc_data()
+{
+  DBUG_ENTER("alloc_data");
+  EventBufData* data = m_free_data;
+
   if (unlikely(data == 0))
   {
 #ifdef VM_TRACE
@@ -1125,7 +1226,7 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
     expand(4000);
     reportStatus();
 
-      data= m_free_data;
+    data = m_free_data;
     if (unlikely(data == 0))
     {
 #ifdef VM_TRACE
@@ -1139,125 +1240,295 @@ NdbEventBuffer::insertDataL(NdbEventOperationImpl *op,
              m_available_data.m_tail ? m_available_data.m_tail->sdata->gci : 0);
       printf("m_used_data_count %d\n", m_used_data.m_count);
 #endif
-	op->m_has_error= 2;
-	DBUG_RETURN(-1); // TODO handle this, overrun, or, skip?
+      DBUG_RETURN(0); // TODO handle this, overrun, or, skip?
     }
   }
 
   // remove data from free list
-    m_free_data= data->m_next;
+  m_free_data = data->m_next;
+  data->m_next = 0;
 #ifdef VM_TRACE
   m_free_data_count--;
   assert(m_free_data_sz >= data->sz);
 #endif
-    m_free_data_sz-= data->sz;
+  m_free_data_sz -= data->sz;
+  DBUG_RETURN(data);
+}
 
-    if (unlikely(copy_data_alloc(sdata, ptr, data)))
-    {
-      op->m_has_error= 3;
-      DBUG_RETURN(-1);
-    }
+// allocate initial or bigger memory area in EventBufData
+// takes sizes from given ptr and sets up data->ptr
+int
+NdbEventBuffer::alloc_mem(EventBufData* data, LinearSectionPtr ptr[3])
+{
+  const Uint32 min_alloc_size = 128;
 
-    // add it to received data
-    bucket->m_data.append(data);
+  Uint32 sz4 = (sizeof(SubTableData) + 3) >> 2;
+  Uint32 alloc_size = (sz4 + ptr[0].sz + ptr[1].sz + ptr[2].sz) << 2;
+  if (alloc_size < min_alloc_size)
+    alloc_size = min_alloc_size;
 
-    data->m_event_op= op;
-#ifdef VM_TRACE
-    op->m_data_count++;
-#endif
-    DBUG_RETURN(0);
+  if (data->sz < alloc_size)
+  {
+    NdbMem_Free((char*)data->memory);
+    assert(m_total_alloc >= data->sz);
+    m_total_alloc -= data->sz;
+    data->memory = 0;
+    data->sz = 0;
+
+    data->memory = (Uint32*)NdbMem_Allocate(alloc_size);
+    if (data->memory == 0)
+      return -1;
+    data->sz = alloc_size;
+    m_total_alloc += data->sz;
   }
 
-#ifdef VM_TRACE
-  if ((Uint32)op->m_eventImpl->mi_type & 1 << (Uint32)sdata->operation)
+  Uint32* memptr = data->memory;
+  memptr += sz4;
+  int i;
+  for (i = 0; i <= 2; i++)
   {
-    DBUG_PRINT("info",("Data arrived before ready eventId", op->m_eventId));
-    DBUG_RETURN(0);
+    data->ptr[i].p = memptr;
+    data->ptr[i].sz = ptr[i].sz;
+    memptr += ptr[i].sz;
   }
-  else {
-    DBUG_PRINT("info",("skipped"));
-    DBUG_RETURN(0);
-  }
-#else
+
   return 0;
-#endif
 }
 
 int 
-NdbEventBuffer::copy_data_alloc(const SubTableData * const f_sdata,
-				LinearSectionPtr f_ptr[3],
-				EventBufData *ev_buf)
+NdbEventBuffer::copy_data(const SubTableData * const sdata,
+                          LinearSectionPtr ptr[3],
+                          EventBufData* data)
 {
-  DBUG_ENTER("NdbEventBuffer::copy_data_alloc");
-  const unsigned min_alloc_size= 128;
-  const unsigned sz4= (sizeof(SubTableData)+3)>>2;
-  Uint32 f_ptr_sz_0= f_ptr[0].sz;
-  Uint32 f_ptr_sz_1= f_ptr[1].sz;
-  Uint32 f_ptr_sz_2= f_ptr[2].sz;
-  LinearSectionPtr *t_ptr= ev_buf->ptr;
-  SubTableData *sdata= ev_buf->sdata;
-  const unsigned alloc_size= (sz4 +
-			      f_ptr_sz_0 +
-			      f_ptr_sz_1 +
-			      f_ptr_sz_2) * sizeof(Uint32);
-  Uint32 *ptr;
-  if (alloc_size > min_alloc_size)
-  {
-    if (sdata)
-    {
-      NdbMem_Free((char*)sdata);
-#ifdef VM_TRACE
-      assert(m_total_alloc >= ev_buf->sz);
-#endif
-      m_total_alloc-= ev_buf->sz;
-    }
-    ptr= (Uint32*)NdbMem_Allocate(alloc_size);
-    ev_buf->sdata= (SubTableData *)ptr;
-    ev_buf->sz= alloc_size;
-    m_total_alloc+= alloc_size;
-  }
-  else /* alloc_size <= min_alloc_size */
+  DBUG_ENTER("NdbEventBuffer::copy_data");
+
+  if (alloc_mem(data, ptr) != 0)
+    DBUG_RETURN(-1);
+  memcpy(data->sdata, sdata, sizeof(SubTableData));
+  int i;
+  for (i = 0; i <= 2; i++)
+    memcpy(data->ptr[i].p, ptr[i].p, ptr[i].sz << 2);
+  DBUG_RETURN(0);
+}
+
+static struct Ev_t {
+  enum {
+    INS = NdbDictionary::Event::_TE_INSERT,
+    DEL = NdbDictionary::Event::_TE_DELETE,
+    UPD = NdbDictionary::Event::_TE_UPDATE,
+    NUL = NdbDictionary::Event::_TE_NUL,
+    ERR = 255
+  };
+  int t1, t2, t3;
+} ev_t[] = {
+  { Ev_t::INS, Ev_t::INS, Ev_t::ERR },
+  { Ev_t::INS, Ev_t::DEL, Ev_t::NUL }, //ok
+  { Ev_t::INS, Ev_t::UPD, Ev_t::INS }, //ok
+  { Ev_t::DEL, Ev_t::INS, Ev_t::UPD }, //ok
+  { Ev_t::DEL, Ev_t::DEL, Ev_t::ERR },
+  { Ev_t::DEL, Ev_t::UPD, Ev_t::ERR },
+  { Ev_t::UPD, Ev_t::INS, Ev_t::ERR },
+  { Ev_t::UPD, Ev_t::DEL, Ev_t::DEL }, //ok
+  { Ev_t::UPD, Ev_t::UPD, Ev_t::UPD }  //ok
+};
+
+/*
+ *   | INS            | DEL              | UPD
+ * 0 | pk ah + all ah | pk ah            | pk ah + new ah 
+ * 1 | pk ad + all ad | old pk ad        | new pk ad + new ad 
+ * 2 | empty          | old non-pk ah+ad | old ah+ad
+ */
+
+static AttributeHeader
+copy_head(Uint32& i1, Uint32* p1, Uint32& i2, const Uint32* p2,
+          Uint32 flags)
+{
+  AttributeHeader ah(p2[i2]);
+  bool do_copy = (flags & 1);
+  if (do_copy)
+    p1[i1] = p2[i2];
+  i1++;
+  i2++;
+  return ah;
+}
+
+static void
+copy_attr(AttributeHeader ah,
+          Uint32& j1, Uint32* p1, Uint32& j2, const Uint32* p2,
+          Uint32 flags)
+{
+  bool do_copy = (flags & 1);
+  bool with_head = (flags & 2);
+  Uint32 n = with_head + ah.getDataSize();
+  if (do_copy)
   {
-    if (sdata)
-      ptr= (Uint32*)sdata;
+    Uint32 k;
+    for (k = 0; k < n; k++)
+      p1[j1++] = p2[j2++];
+  }
   else
   {
-      ptr= (Uint32*)NdbMem_Allocate(min_alloc_size);
-      ev_buf->sdata= (SubTableData *)ptr;
-      ev_buf->sz= min_alloc_size;
-      m_total_alloc+= min_alloc_size;
+    j1 += n;
+    j2 += n;
+  }
+}
+
+int 
+NdbEventBuffer::merge_data(const SubTableData * const sdata,
+                           LinearSectionPtr ptr2[3],
+                           EventBufData* data)
+{
+  DBUG_ENTER("NdbEventBuffer::merge_data");
+
+  Uint32 nkey = data->m_event_op->m_eventImpl->m_tableImpl->m_noOfKeys;
+
+  int t1 = data->sdata->operation;
+  int t2 = sdata->operation;
+  if (t1 == Ev_t::NUL)
+    DBUG_RETURN(copy_data(sdata, ptr2, data));
+
+  Ev_t* tp = 0;
+  int i;
+  for (i = 0; i < sizeof(ev_t)/sizeof(ev_t[0]); i++) {
+    if (ev_t[i].t1 == t1 && ev_t[i].t2 == t2) {
+      tp = &ev_t[i];
+      break;
     }
   }
+  assert(tp != 0 && tp->t3 != Ev_t::ERR);
 
-  memcpy(ptr,f_sdata,sizeof(SubTableData));
-  ptr+= sz4;
+  // save old data
+  EventBufData olddata = *data;
+  data->memory = 0;
+  data->sz = 0;
 
-  t_ptr->p= ptr;
-  t_ptr->sz= f_ptr_sz_0;
+  // compose ptr1 o ptr2 = ptr
+  LinearSectionPtr (&ptr1) [3] = olddata.ptr;
+  LinearSectionPtr (&ptr) [3] = data->ptr;
 
-  memcpy(ptr, f_ptr[0].p, sizeof(Uint32)*f_ptr_sz_0);
-  ptr+= f_ptr_sz_0;
-  t_ptr++;
+  // loop twice where first loop only sets sizes
+  int loop;
+  for (loop = 0; loop <= 1; loop++)
+  {
+    if (loop == 1)
+    {
+      if (alloc_mem(data, ptr) != 0)
+        DBUG_RETURN(-1);
+      *data->sdata = *sdata;
+      data->sdata->operation = tp->t3;
+    }
 
-  t_ptr->p= ptr;
-  t_ptr->sz= f_ptr_sz_1;
+    ptr[0].sz = ptr[1].sz = ptr[3].sz = 0;
 
-  memcpy(ptr, f_ptr[1].p, sizeof(Uint32)*f_ptr_sz_1);
-  ptr+= f_ptr_sz_1;
-  t_ptr++;
+    // copy pk from new version
+    {
+      AttributeHeader ah;
+      Uint32 i = 0;
+      Uint32 j = 0;
+      Uint32 i2 = 0;
+      Uint32 j2 = 0;
+      while (i < nkey)
+      {
+        ah = copy_head(i, ptr[0].p, i2, ptr2[0].p, loop);
+        copy_attr(ah, j, ptr[1].p, j2, ptr2[1].p, loop);
+      }
+      ptr[0].sz = i;
+      ptr[1].sz = j;
+    }
 
-  if (f_ptr_sz_2)
+    // merge after values, new version overrides
+    if (tp->t3 != Ev_t::DEL)
+    {
+      AttributeHeader ah;
+      Uint32 i = ptr[0].sz;
+      Uint32 j = ptr[1].sz;
+      Uint32 i1 = 0;
+      Uint32 j1 = 0;
+      Uint32 i2 = nkey;
+      Uint32 j2 = ptr[1].sz;
+      while (i1 < nkey)
+      {
+        j1 += AttributeHeader(ptr1[0].p[i1++]).getDataSize();
+      }
+      while (1)
+      {
+        bool b1 = (i1 < ptr1[0].sz);
+        bool b2 = (i2 < ptr2[0].sz);
+        if (b1 && b2)
+        {
+          Uint32 id1 = AttributeHeader(ptr1[0].p[i1]).getAttributeId();
+          Uint32 id2 = AttributeHeader(ptr2[0].p[i2]).getAttributeId();
+          if (id1 < id2)
+            b2 = false;
+          else if (id1 > id2)
+            b1 = false;
+          else
+          {
+            j1 += AttributeHeader(ptr1[0].p[i1++]).getDataSize();
+            b1 = false;
+          }
+        }
+        if (b1)
+        {
+          ah = copy_head(i, ptr[0].p, i1, ptr1[0].p, loop);
+          copy_attr(ah, j, ptr[1].p, j1, ptr1[1].p, loop);
+        }
+        else if (b2)
         {
-    t_ptr->p= ptr;
-    t_ptr->sz= f_ptr_sz_2;
-    memcpy(ptr, f_ptr[2].p, sizeof(Uint32)*f_ptr_sz_2);
+          ah = copy_head(i, ptr[0].p, i2, ptr2[0].p, loop);
+          copy_attr(ah, j, ptr[1].p, j2, ptr2[1].p, loop);
+        }
+        else
+          break;
       }
+      ptr[0].sz = i;
+      ptr[1].sz = j;
+    }
+
+    // merge before values, old version overrides
+    if (tp->t3 != Ev_t::INS)
+    {
+      AttributeHeader ah;
+      Uint32 k = 0;
+      Uint32 k1 = 0;
+      Uint32 k2 = 0;
+      while (1)
+      {
+        bool b1 = (k1 < ptr1[2].sz);
+        bool b2 = (k2 < ptr2[2].sz);
+        if (b1 && b2)
+        {
+          Uint32 id1 = AttributeHeader(ptr1[2].p[k1]).getAttributeId();
+          Uint32 id2 = AttributeHeader(ptr2[2].p[k2]).getAttributeId();
+          if (id1 < id2)
+            b2 = false;
+          else if (id1 > id2)
+            b1 = false;
           else
           {
-    t_ptr->p= 0;
-    t_ptr->sz= 0;
+            k2 += 1 + AttributeHeader(ptr2[2].p[k2]).getDataSize();
+            b2 = false;
+          }
+        }
+        if (b1)
+        {
+          ah = AttributeHeader(ptr1[2].p[k1]);
+          copy_attr(ah, k, ptr[2].p, k1, ptr1[2].p, loop | 2);
+        }
+        else if (b2)
+        {
+          ah = AttributeHeader(ptr2[2].p[k2]);
+          copy_attr(ah, k, ptr[2].p, k2, ptr2[2].p, loop | 2);
+        }
+        else
+          break;
+      }
+    }
   }
 
+  // free old data
+  NdbMem_Free((char*)olddata.memory);
+
   DBUG_RETURN(0);
 }
 
@@ -1399,5 +1670,107 @@ NdbEventBuffer::reportStatus()
 #endif
 }
 
+// hash table routines
+
+// could optimize the all-fixed case
+Uint32
+EventBufData_hash::getpkhash(NdbEventOperationImpl* op, LinearSectionPtr ptr[3])
+{
+  const NdbTableImpl* tab = op->m_eventImpl->m_tableImpl;
+
+  // in all cases ptr[0] = pk ah.. ptr[1] = pk ad..
+  // for pk update (to equivalent pk) post/pre values give same hash
+  Uint32 nkey = tab->m_noOfKeys;
+  assert(nkey != 0 && nkey <= ptr[0].sz);
+  const Uint32* hptr = ptr[0].p;
+  const uchar* dptr = (uchar*)ptr[1].p;
+
+  // hash registers
+  ulong nr1 = 0;
+  ulong nr2 = 0;
+  while (nkey-- != 0)
+  {
+    AttributeHeader ah(*hptr++);
+    Uint32 bytesize = ah.getByteSize();
+    assert(dptr + bytesize <= (uchar*)(ptr[1].p + ptr[1].sz));
+
+    Uint32 i = ah.getAttributeId();
+    const NdbColumnImpl* col = tab->getColumn(i);
+    assert(col != 0);
+
+    Uint32 lb, len;
+    bool ok = NdbSqlUtil::get_var_length(col->m_type, dptr, bytesize, lb, len);
+    assert(ok);
+
+    CHARSET_INFO* cs = col->m_cs ? col->m_cs : &my_charset_bin;
+    (*cs->coll->hash_sort)(cs, dptr + lb, len, &nr1, &nr2);
+    dptr += bytesize;
+  }
+  return nr1;
+}
+
+// this is seldom invoked
+bool
+EventBufData_hash::getpkequal(NdbEventOperationImpl* op, LinearSectionPtr ptr1[3], LinearSectionPtr ptr2[3])
+{
+  const NdbTableImpl* tab = op->m_eventImpl->m_tableImpl;
+
+  Uint32 nkey = tab->m_noOfKeys;
+  assert(nkey != 0 && nkey <= ptr1[0].sz && nkey <= ptr2[0].sz);
+  const Uint32* hptr1 = ptr1[0].p;
+  const Uint32* hptr2 = ptr2[0].p;
+  const uchar* dptr1 = (uchar*)ptr1[1].p;
+  const uchar* dptr2 = (uchar*)ptr2[1].p;
+
+  while (nkey-- != 0)
+  {
+    AttributeHeader ah1(*hptr1++);
+    AttributeHeader ah2(*hptr2++);
+    // sizes can differ on update of varchar endspace
+    Uint32 bytesize1 = ah1.getByteSize();
+    Uint32 bytesize2 = ah1.getByteSize();
+    assert(dptr1 + bytesize1 <= (uchar*)(ptr1[1].p + ptr1[1].sz));
+    assert(dptr2 + bytesize2 <= (uchar*)(ptr2[1].p + ptr2[1].sz));
+
+    assert(ah1.getAttributeId() == ah2.getAttributeId());
+    Uint32 i = ah1.getAttributeId();
+    const NdbColumnImpl* col = tab->getColumn(i);
+    assert(col != 0);
+
+    Uint32 lb1, len1;
+    bool ok1 = NdbSqlUtil::get_var_length(col->m_type, dptr1, bytesize1, lb1, len1);
+    Uint32 lb2, len2;
+    bool ok2 = NdbSqlUtil::get_var_length(col->m_type, dptr2, bytesize2, lb2, len2);
+    assert(ok1 && ok2 && lb1 == lb2);
+
+    CHARSET_INFO* cs = col->m_cs ? col->m_cs : &my_charset_bin;
+    int res = (cs->coll->strnncollsp)(cs, dptr1 + lb1, len1, dptr2 + lb2, len2, false);
+    if (res != 0)
+      return false;
+    dptr1 += bytesize1;
+    dptr2 += bytesize2;
+  }
+  return true;
+}
+
+void
+EventBufData_hash::search(Pos& hpos, NdbEventOperationImpl* op, LinearSectionPtr ptr[3])
+{
+  Uint32 pkhash = getpkhash(op, ptr);
+  Uint32 index = (op->m_oid ^ pkhash) % GCI_EVENT_HASH_SIZE;
+  EventBufData* data = m_hash[index];
+  while (data != 0)
+  {
+    if (data->m_event_op == op &&
+        data->m_pkhash == pkhash &&
+        getpkequal(op, data->ptr, ptr))
+      break;
+    data = data->m_next_hash;
+  }
+  hpos.index = index;
+  hpos.data = data;
+  hpos.pkhash = pkhash;
+}
+
 template class Vector<Gci_container>;
 template class Vector<NdbEventBuffer::EventBufData_chunk*>;

storage/ndb/src/ndbapi/NdbEventOperationImpl.hpp

@@ -25,16 +25,19 @@
 #define NDB_EVENT_OP_MAGIC_NUMBER 0xA9F301B4
 
 class NdbEventOperationImpl;
+
 struct EventBufData
 {
   union {
     SubTableData *sdata;
-    char *memory;
+    Uint32 *memory;
   };
   LinearSectionPtr ptr[3];
   unsigned sz;
   NdbEventOperationImpl *m_event_op;
   EventBufData *m_next; // Next wrt to global order
+  EventBufData *m_next_hash; // Next in per-GCI hash
+  Uint32 m_pkhash; // PK hash (without op) for fast compare
 };
 
 class EventBufData_list
@@ -116,6 +119,34 @@ void EventBufData_list::append(const EventBufData_list &list)
   m_sz+= list.m_sz;
 }
 
+// GCI bucket has also a hash over data, with key event op, table PK.
+// It can only be appended to and is invalid after remove_first().
+class EventBufData_hash
+{
+public:
+  struct Pos { // search result
+    Uint32 index;       // index into hash array
+    EventBufData* data; // non-zero if found
+    Uint32 pkhash;      // PK hash
+  };
+
+  static Uint32 getpkhash(NdbEventOperationImpl* op, LinearSectionPtr ptr[3]);
+  static bool getpkequal(NdbEventOperationImpl* op, LinearSectionPtr ptr1[3], LinearSectionPtr ptr2[3]);
+
+  void search(Pos& hpos, NdbEventOperationImpl* op, LinearSectionPtr ptr[3]);
+  void append(Pos& hpos, EventBufData* data);
+
+  enum { GCI_EVENT_HASH_SIZE = 101 };
+  EventBufData* m_hash[GCI_EVENT_HASH_SIZE];
+};
+
+inline
+void EventBufData_hash::append(Pos& hpos, EventBufData* data)
+{
+  data->m_next_hash = m_hash[hpos.index];
+  m_hash[hpos.index] = data;
+}
+
 struct Gci_container
 {
   enum State 
@@ -127,6 +158,7 @@ struct Gci_container
   Uint32 m_gcp_complete_rep_count; // Remaining SUB_GCP_COMPLETE_REP until done
   Uint64 m_gci;                    // GCI
   EventBufData_list m_data;
+  EventBufData_hash m_data_hash;
 };
 
 class NdbEventOperationImpl : public NdbEventOperation {
@@ -174,6 +206,8 @@ public:
   Uint32 m_eventId;
   Uint32 m_oid;
 
+  bool m_separateEvents;
+  
   EventBufData *m_data_item;
 
   void *m_custom_data;
@@ -212,7 +246,6 @@ public:
   void add_op();
   void remove_op();
   void init_gci_containers();
-  Uint32 m_active_op_count;
 
   // accessed from the "receive thread"
   int insertDataL(NdbEventOperationImpl *op,
@@ -233,10 +266,15 @@ public:
 
   NdbEventOperationImpl *move_data();
 
-  // used by both user thread and receive thread
-  int copy_data_alloc(const SubTableData * const f_sdata,
-		      LinearSectionPtr f_ptr[3],
-		      EventBufData *ev_buf);
+  // routines to copy/merge events
+  EventBufData* alloc_data();
+  int alloc_mem(EventBufData* data, LinearSectionPtr ptr[3]);
+  int copy_data(const SubTableData * const sdata,
+                LinearSectionPtr ptr[3],
+                EventBufData* data);
+  int merge_data(const SubTableData * const sdata,
+                 LinearSectionPtr ptr[3],
+                 EventBufData* data);
 
   void free_list(EventBufData_list &list);
 
@@ -290,6 +328,8 @@ private:
   // dropped event operations that have not yet
   // been deleted
   NdbEventOperationImpl *m_dropped_ev_op;
+
+  Uint32 m_active_op_count;
 };
 
 inline

storage/ndb/test/ndbapi/test_event.cpp

@@ -169,6 +169,7 @@ eventOperation(Ndb* pNdb, const NdbDictionary::Table &tab, void* pstats, int rec
     g_err << function << "Event operation creation failed\n";
     return NDBT_FAILED;
   }
+  pOp->separateEvents(true);
 
   g_info << function << "get values\n";
   NdbRecAttr* recAttr[1024];
@@ -370,6 +371,7 @@ int runCreateDropEventOperation(NDBT_Context* ctx, NDBT_Step* step)
       g_err << "Event operation creation failed\n";
       return NDBT_FAILED;
     }
+    pOp->separateEvents(true);
 
     g_info << "dropping event operation" << endl;
     int res = pNdb->dropEventOperation(pOp);
@@ -540,6 +542,7 @@ int runEventApplier(NDBT_Context* ctx, NDBT_Step* step)
     g_err << "Event operation creation failed on %s" << buf << endl;
     DBUG_RETURN(NDBT_FAILED);
   }
+  pOp->separateEvents(true);
 
   int i;
   int n_columns= table->getNoOfColumns();
@@ -1185,6 +1188,7 @@ static int createEventOperations(Ndb * ndb)
     {
       DBUG_RETURN(NDBT_FAILED);
     }
+    pOp->separateEvents(true);
 
     int n_columns= pTabs[i]->getNoOfColumns();
     for (int j = 0; j < n_columns; j++)