mariadb-columnstore-engine/writeengine/shared/we_dbrootextenttracker.cpp

/* Copyright (C) 2014 InfiniDB, Inc.

   This program is free software; you can redistribute it and/or
   modify it under the terms of the GNU General Public License
   as published by the Free Software Foundation; version 2 of
   the License.

   This program is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; if not, write to the Free Software
   Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
   MA 02110-1301, USA. */

/*
 * $Id: we_dbrootextenttracker.cpp 4631 2013-05-02 15:21:09Z dcathey $
 */
#include "we_dbrootextenttracker.h"

#include <algorithm>
#include <sstream>

#include "we_brm.h"
#include "we_config.h"
#include "we_log.h"

namespace
{
const char* stateStrings[] = {"initState", "PartialExtent", "EmptyDbRoot", "ExtentBoundary", "OutOfService"};
}

namespace WriteEngine
{
//------------------------------------------------------------------------------
// DBRootExtentInfo constructor
//------------------------------------------------------------------------------
DBRootExtentInfo::DBRootExtentInfo(uint16_t dbRoot, uint32_t partition, uint16_t segment,
                                   BRM::LBID_t startLbid, HWM localHwm, uint64_t dbrootTotalBlocks,
                                   DBRootExtentInfoState state)
 : fPartition(partition)
 , fDbRoot(dbRoot)
 , fSegment(segment)
 , fStartLbid(startLbid)
 , fLocalHwm(localHwm)
 , fDBRootTotalBlocks(dbrootTotalBlocks)
 , fState(state)
{
}

//------------------------------------------------------------------------------
// LessThan operator used to sort DBRootExtentInfo objects by DBRoot.
//------------------------------------------------------------------------------
bool DBRootExtentInfo::operator<(const DBRootExtentInfo& entry) const
{
  if (fDbRoot < entry.fDbRoot)
    return true;

  return false;
}

//------------------------------------------------------------------------------
// DBRootExtentTracker constructor
//
// Mutex lock not needed in this function as it is only called from main thread
// before processing threads are spawned.
//------------------------------------------------------------------------------
DBRootExtentTracker::DBRootExtentTracker(OID oid, const std::vector<int>& colWidths,
                                         const std::vector<BRM::EmDbRootHWMInfo_v>& dbRootHWMInfoColVec,
                                         unsigned int columnIdx, Log* logger)
 : fOID(oid)
 , fLog(logger)
 , fCurrentDBRootIdx(-1)
 , fEmptyOrDisabledPM(false)
 , fEmptyPM(true)
 , fDisabledHWM(false)
{
  const BRM::EmDbRootHWMInfo_v& emDbRootHWMInfo = dbRootHWMInfoColVec[columnIdx];
  int colWidth = colWidths[columnIdx];

  fBlksPerExtent =
      (long long)BRMWrapper::getInstance()->getExtentRows() * (long long)colWidth / (long long)BYTE_PER_BLOCK;

  std::vector<bool> resetState;

  for (unsigned int i = 0; i < emDbRootHWMInfo.size(); i++)
  {
    resetState.push_back(false);
    DBRootExtentInfoState state = determineState(colWidths[columnIdx], emDbRootHWMInfo[i].localHWM,
                                                 emDbRootHWMInfo[i].totalBlocks, emDbRootHWMInfo[i].status);

    // For a full extent...
    // check to see if any of the column HWMs are partially full, in which
    // case we consider all the columns for that DBRoot to be partially
    // full.  (This can happen if a table has columns with varying widths,
    // as the HWM may be at the last extent block for a shorter column, and
    // still have free blocks for wider columns.)
    if (state == DBROOT_EXTENT_EXTENT_BOUNDARY)
    {
      for (unsigned int kCol = 0; kCol < dbRootHWMInfoColVec.size(); kCol++)
      {
        const BRM::EmDbRootHWMInfo_v& emDbRootHWMInfo2 = dbRootHWMInfoColVec[kCol];
        DBRootExtentInfoState state2 =
            determineState(colWidths[kCol], emDbRootHWMInfo2[i].localHWM, emDbRootHWMInfo2[i].totalBlocks,
                           emDbRootHWMInfo2[i].status);

        if (state2 == DBROOT_EXTENT_PARTIAL_EXTENT)
        {
          state = DBROOT_EXTENT_PARTIAL_EXTENT;
          resetState[resetState.size() - 1] = true;
          break;
        }
      }
    }

    DBRootExtentInfo dbRootExtent(emDbRootHWMInfo[i].dbRoot, emDbRootHWMInfo[i].partitionNum,
                                  emDbRootHWMInfo[i].segmentNum, emDbRootHWMInfo[i].startLbid,
                                  emDbRootHWMInfo[i].localHWM, emDbRootHWMInfo[i].totalBlocks, state);

    fDBRootExtentList.push_back(dbRootExtent);
  }

  std::sort(fDBRootExtentList.begin(), fDBRootExtentList.end());

  if (fLog)
  {
    // Always log this info for now; may control with debug later
    // if (fLog->isDebug(DEBUG_1))
    {
      std::ostringstream oss;
      oss << "Starting DBRoot info for OID " << fOID;

      for (unsigned int k = 0; k < fDBRootExtentList.size(); k++)
      {
        oss << std::endl;
        oss << "  DBRoot-" << fDBRootExtentList[k].fDbRoot
            << ", part/seg/hwm/LBID/totBlks/state: " << fDBRootExtentList[k].fPartition << "/"
            << fDBRootExtentList[k].fSegment << "/" << fDBRootExtentList[k].fLocalHwm << "/"
            << fDBRootExtentList[k].fStartLbid << "/" << fDBRootExtentList[k].fDBRootTotalBlocks << "/"
            << stateStrings[fDBRootExtentList[k].fState];

        if (resetState[k])
          oss << ".";
      }

      fLog->logMsg(oss.str(), MSGLVL_INFO2);
    }
  }
}

//------------------------------------------------------------------------------
// Determines the state of the HWM extent (for a DBRoot); considering the
// current BRM status, HWM, and total block count for the DBRoot.
//------------------------------------------------------------------------------
DBRootExtentInfoState DBRootExtentTracker::determineState(int colWidth, HWM localHwm,
                                                          uint64_t dbRootTotalBlocks, int16_t status)
{
  DBRootExtentInfoState extentState;

  if (status == BRM::EXTENTOUTOFSERVICE)
  {
    extentState = DBROOT_EXTENT_OUT_OF_SERVICE;
  }
  else
  {
    if (dbRootTotalBlocks == 0)
    {
      extentState = DBROOT_EXTENT_EMPTY_DBROOT;
    }
    else
    {
      extentState = DBROOT_EXTENT_PARTIAL_EXTENT;

      // See if local hwm is on an extent bndry,in which case the extent
      // is full and we won't be adding rows to the current HWM extent;
      // we will instead need to allocate a new extent in order to begin
      // adding any rows.
      long long nRows = ((long long)(localHwm + 1) * (long long)BYTE_PER_BLOCK) / (long long)colWidth;
      long long nRem = nRows % BRMWrapper::getInstance()->getExtentRows();

      if (nRem == 0)
      {
        extentState = DBROOT_EXTENT_EXTENT_BOUNDARY;
      }
    }
  }

  return extentState;
}

//------------------------------------------------------------------------------
// Select the first segment file to add rows to for the local PM.
// Function will first try to find the HWM extent with the fewest blocks.
// If all the HWM extents are full, then we select the DBRoot/segment file
// with the fewest total overall blocks for that DBRoot.
// Return value is 0 upon success, else non zero if no eligible entries found.
//
// Mutex lock not needed in this function as it is only called from main thread
// before processing threads are spawned.
//------------------------------------------------------------------------------
int DBRootExtentTracker::selectFirstSegFile(DBRootExtentInfo& dbRootExtent, bool& bNoStartExtentOnThisPM,
                                            bool& bEmptyPM, std::string& errMsg)
{
  int startExtentIdx = -1;
  int fewestLocalBlocksIdx = -1;  // track HWM extent with fewest blocks
  int fewestTotalBlocksIdx = -1;  // track DBRoot with fewest total blocks
  bNoStartExtentOnThisPM = false;

  unsigned int fewestTotalBlks = UINT_MAX;
  unsigned int fewestLocalBlks = UINT_MAX;
  uint16_t fewestTotalBlkSegNum = USHRT_MAX;
  uint16_t fewestLocalBlkSegNum = USHRT_MAX;

  // Find DBRoot having HWM extent with fewest blocks.  If all HWM extents
  // are full (remblks=0), then fall-back on selecting the DBRoot with fewest
  // total blks.
  //
  // Selecting HWM extent with fewest blocks should be straight forward, be-
  // cause all the DBRoots on a PM should end on an extent boundary except
  // for the current last extent.  But if the user has moved a DBRoot, then
  // we can end up with 2 partially filled HWM extents on 2 DBRoots, on the
  // same PM.  That's why we loop through the DBRoots to see if we have more
  // than 1 partially filled HWM extent.
  for (unsigned int iroot = 0; iroot < fDBRootExtentList.size(); iroot++)
  {
    // Skip over DBRoots which have no extents
    if (fDBRootExtentList[iroot].fState == DBROOT_EXTENT_EMPTY_DBROOT)
      continue;

    fEmptyPM = false;

    // Find DBRoot and segment file with most incomplete extent.
    // Break a tie by selecting the lowest segment number.
    long long remBlks = (long long)(fDBRootExtentList[iroot].fLocalHwm + 1) % fBlksPerExtent;

    if (remBlks > 0)
    {
      if ((remBlks < fewestLocalBlks) ||
          ((remBlks == fewestLocalBlks) && (fDBRootExtentList[iroot].fSegment < fewestLocalBlkSegNum)))
      {
        fewestLocalBlocksIdx = iroot;
        fewestLocalBlks = remBlks;
        fewestLocalBlkSegNum = fDBRootExtentList[iroot].fSegment;
      }
    }

    // Find DBRoot with fewest total of blocks.
    // Break a tie by selecting the highest segment number.
    if ((fDBRootExtentList[iroot].fDBRootTotalBlocks < fewestTotalBlks) ||
        ((fDBRootExtentList[iroot].fDBRootTotalBlocks == fewestTotalBlks) &&
         (fDBRootExtentList[iroot].fSegment > fewestTotalBlkSegNum)))
    {
      fewestTotalBlocksIdx = iroot;
      fewestTotalBlks = fDBRootExtentList[iroot].fDBRootTotalBlocks;
      fewestTotalBlkSegNum = fDBRootExtentList[iroot].fSegment;
    }
  }

  // Select HWM extent with fewest number of blocks;
  // If chosen extent is disabled, then treat like an empty PM,
  // meaning we have to allocate a new extent before adding any rows
  if (fewestLocalBlocksIdx != -1)
  {
    startExtentIdx = fewestLocalBlocksIdx;

    if (fDBRootExtentList[startExtentIdx].fState == DBROOT_EXTENT_OUT_OF_SERVICE)
    {
      fDisabledHWM = true;
    }
  }

  // If the HWM on each DBRoot ends on an extent boundary, then
  // select the DBRoot with the fewest total number of blocks;
  // If chosen extent is disabled, then treat like an empty PM,
  // meaning we have to allocate a new extent before adding any rows
  else if (fewestTotalBlocksIdx != -1)
  {
    startExtentIdx = fewestTotalBlocksIdx;

    if (fDBRootExtentList[startExtentIdx].fState == DBROOT_EXTENT_OUT_OF_SERVICE)
    {
      fDisabledHWM = true;
    }
  }

  // PM with no extents (or all extents disabled), so select DBRoot/segment
  // file from DBRoot list, where we will start inserting rows.
  // Select lowest segment file number.
  else
  {
    RETURN_ON_ERROR(selectFirstSegFileForEmptyPM(errMsg));

    startExtentIdx = fCurrentDBRootIdx;
  }

  if ((fEmptyOrDisabledPM) || (fDisabledHWM))
    bNoStartExtentOnThisPM = true;

  bEmptyPM = fEmptyPM;
  fCurrentDBRootIdx = startExtentIdx;

  // Finish Initializing DBRootExtentList for empty DBRoots w/o any extents
  initEmptyDBRoots();

  logFirstDBRootSelection();

  dbRootExtent = fDBRootExtentList[startExtentIdx];
  fDBRootExtentList[startExtentIdx].fState = DBROOT_EXTENT_EXTENT_BOUNDARY;

  return NO_ERROR;
}

//------------------------------------------------------------------------------
// If we have encountered a PM with no extents (or all extent disabled), then
// this function can be called to determine the DBRoot to be used for the 1st
// extent for the applicable PM.  First DBRoot for relevant PM is selected.
// At extent creation time, BRM will assign the segment number.
//
// Mutex lock not needed in this function as it is only called from main thread
// before processing threads are spawned.
//------------------------------------------------------------------------------
int DBRootExtentTracker::selectFirstSegFileForEmptyPM(std::string& errMsg)
{
  fEmptyOrDisabledPM = true;

  fCurrentDBRootIdx = 0;  // Start with first DBRoot for this PM

  // Always start empty PM with partition number 0.  If the DBRoot has a HWM
  // extent that is disabled, then BRM will override this partition number.
  fDBRootExtentList[fCurrentDBRootIdx].fPartition = 0;

  return NO_ERROR;
}

//------------------------------------------------------------------------------
// Finish Initializing fDBRootExtentList for any empty DBRoots w/o any extents.
//------------------------------------------------------------------------------
void DBRootExtentTracker::initEmptyDBRoots()
{
  int startExtentIdx = fCurrentDBRootIdx;
  bool bAnyChanges = false;  // If fDBRootExtentList changes, log the contents

  // Fill in starting partition for any DBRoots having no extents
  for (unsigned int iroot = 0; iroot < fDBRootExtentList.size(); iroot++)
  {
    if ((fDBRootExtentList[iroot].fState == DBROOT_EXTENT_EMPTY_DBROOT) &&
        ((int)iroot != startExtentIdx))  // skip over selected dbroot
    {
      if (fDBRootExtentList[iroot].fPartition != fDBRootExtentList[startExtentIdx].fPartition)
      {
        bAnyChanges = true;

        fDBRootExtentList[iroot].fPartition = fDBRootExtentList[startExtentIdx].fPartition;
      }
    }
  }

  // Log fDBRootExtentList if modifications were made
  if ((bAnyChanges) && (fLog))
  {
    // Always log this info for now; may control with debug later
    // if (fLog->isDebug(DEBUG_1))
    {
      std::ostringstream oss;
      oss << "Updated starting (empty) DBRoot info for OID " << fOID;

      for (unsigned int k = 0; k < fDBRootExtentList.size(); k++)
      {
        oss << std::endl;
        oss << "  DBRoot-" << fDBRootExtentList[k].fDbRoot
            << ", part/seg/hwm/LBID/totBlks/state: " << fDBRootExtentList[k].fPartition << "/"
            << fDBRootExtentList[k].fSegment << "/" << fDBRootExtentList[k].fLocalHwm << "/"
            << fDBRootExtentList[k].fStartLbid << "/" << fDBRootExtentList[k].fDBRootTotalBlocks << "/"
            << stateStrings[fDBRootExtentList[k].fState];
      }

      fLog->logMsg(oss.str(), MSGLVL_INFO2);
    }
  }
}

//------------------------------------------------------------------------------
// Assign the DBRoot/segment file to be used for extent loading based on the
// setting in the specified reference tracker (set for a reference column).
//
// Mutex lock not needed in this function as it is only called from main thread
// before processing threads are spawned.
//------------------------------------------------------------------------------
void DBRootExtentTracker::assignFirstSegFile(const DBRootExtentTracker& refTracker,
                                             DBRootExtentInfo& dbRootExtent)
{
  // Start with the same DBRoot index as the reference tracker; assumes that
  // DBRoots for each column are listed in same order in fDBRootExtentList.
  // That should be a safe assumption since DBRootExtentTracker constructor
  // sorts the entries in fDBRootExtentList by fDbRoot.
  int startExtentIdx = refTracker.fCurrentDBRootIdx;
  fEmptyOrDisabledPM = refTracker.fEmptyOrDisabledPM;
  fEmptyPM = refTracker.fEmptyPM;
  fDisabledHWM = refTracker.fDisabledHWM;

  // Always start empty PM with partition number 0.  If the DBRoot has a HWM
  // extent that is disabled, then BRM will override this partition number.
  if (fEmptyOrDisabledPM)
  {
    fDBRootExtentList[startExtentIdx].fPartition = 0;
  }

  fCurrentDBRootIdx = startExtentIdx;

  // Finish Initializing DBRootExtentList for empty DBRoots w/o any extents
  initEmptyDBRoots();

  logFirstDBRootSelection();

  dbRootExtent = fDBRootExtentList[startExtentIdx];
  fDBRootExtentList[startExtentIdx].fState = DBROOT_EXTENT_EXTENT_BOUNDARY;
}

//------------------------------------------------------------------------------
// Log information about the first DBRoot/segment file that is selected.
//------------------------------------------------------------------------------
void DBRootExtentTracker::logFirstDBRootSelection() const
{
  if (fLog)
  {
    int extentIdx = fCurrentDBRootIdx;

    if (fEmptyOrDisabledPM)
    {
      std::ostringstream oss;
      oss << "No active extents; will add partition to start adding "
             "rows for oid-"
          << fOID << "; DBRoot-" << fDBRootExtentList[extentIdx].fDbRoot;
      fLog->logMsg(oss.str(), MSGLVL_INFO2);
    }
    else if (fDisabledHWM)
    {
      std::ostringstream oss;
      oss << "HWM extent disabled; will add partition to start adding "
             "rows for oid-"
          << fOID << "; DBRoot-" << fDBRootExtentList[extentIdx].fDbRoot;
      fLog->logMsg(oss.str(), MSGLVL_INFO2);
    }
    else
    {
      std::ostringstream oss;
      oss << "Selecting existing segFile to begin adding rows: oid-" << fOID << "; DBRoot-"
          << fDBRootExtentList[extentIdx].fDbRoot
          << ", part/seg/hwm/LBID/totBlks/state: " << fDBRootExtentList[extentIdx].fPartition << "/"
          << fDBRootExtentList[extentIdx].fSegment << "/" << fDBRootExtentList[extentIdx].fLocalHwm << "/"
          << fDBRootExtentList[extentIdx].fStartLbid << "/" << fDBRootExtentList[extentIdx].fDBRootTotalBlocks
          << "/" << stateStrings[fDBRootExtentList[extentIdx].fState];
      fLog->logMsg(oss.str(), MSGLVL_INFO2);
    }
  }
}

//------------------------------------------------------------------------------
// Iterate/return next DBRoot to be used for the next extent
//
// If it is the "very" 1st extent for the specified DBRoot, then the applicable
// partition to be used for the first extent, is also returned.
// If a partial extent is to be filled in, then the current HWM
// and starting LBID for the relevant extent are returned.
// If a disabled extent is encountered, the return flag is true, so that
// we can allocate a new extent at the next physical partition.  BRM should
// take care of figuring out where to allocate this next extent.
//
// Returns true if new extent needs to be allocated, else false indicates that
// the extent is partially full.
//------------------------------------------------------------------------------
bool DBRootExtentTracker::nextSegFile(uint16_t& dbRoot, uint32_t& partition, uint16_t& segment, HWM& localHwm,
                                      BRM::LBID_t& startLbid)
{
  boost::mutex::scoped_lock lock(fDBRootExtTrkMutex);

  fCurrentDBRootIdx++;

  if ((unsigned int)fCurrentDBRootIdx >= fDBRootExtentList.size())
    fCurrentDBRootIdx = 0;

  dbRoot = fDBRootExtentList[fCurrentDBRootIdx].fDbRoot;
  segment = fDBRootExtentList[fCurrentDBRootIdx].fSegment;
  partition = fDBRootExtentList[fCurrentDBRootIdx].fPartition;
  localHwm = fDBRootExtentList[fCurrentDBRootIdx].fLocalHwm;
  startLbid = fDBRootExtentList[fCurrentDBRootIdx].fStartLbid;
  // std::cout << "NextSegFile: Current idx: " << fCurrentDBRootIdx <<
  //"; new dbroot: " << dbRoot    <<
  //"; segment: "    << segment   <<
  //"; partition: "  << partition <<
  //"; localHwm: "   << localHwm  <<
  //"; startLbid: "  << startLbid <<
  //"; state: "      << stateStrings[fDBRootExtentList[fCurrentDBRootIdx].fState]
  //  << std::endl;

  bool bAllocExtentFlag = true;

  if (fDBRootExtentList[fCurrentDBRootIdx].fState == DBROOT_EXTENT_PARTIAL_EXTENT)
    bAllocExtentFlag = false;

  // After we have taken care of the "first" extent for each DBRoot, we can
  // zero out everything.  The only thing we need to continue rotating thru
  // the DBRoots, is the DBRoot number itself.
  fDBRootExtentList[fCurrentDBRootIdx].fSegment = 0;
  fDBRootExtentList[fCurrentDBRootIdx].fPartition = 0;
  fDBRootExtentList[fCurrentDBRootIdx].fLocalHwm = 0;
  fDBRootExtentList[fCurrentDBRootIdx].fStartLbid = 0;
  fDBRootExtentList[fCurrentDBRootIdx].fDBRootTotalBlocks = 0;
  fDBRootExtentList[fCurrentDBRootIdx].fState = DBROOT_EXTENT_EXTENT_BOUNDARY;

  return bAllocExtentFlag;
}

const std::vector<DBRootExtentInfo>& DBRootExtentTracker::getDBRootExtentList()
{
  boost::mutex::scoped_lock lock(fDBRootExtTrkMutex);
  return fDBRootExtentList;
}

}  // namespace WriteEngine