MCOL-498. Segment files extension uses fallocate() now to optimize load put on SSD disks.

utils/idbdatafile/BufferedFile.cpp

@@ -279,4 +279,21 @@ int BufferedFile::close()
     return ret;
 }
 
+int BufferedFile::fallocate(int mode, off64_t offset, off64_t length)
+{
+    int ret = 0;
+    int savedErrno = 0;
+
+    ret = ::fallocate( fileno(m_fp), mode, offset, length );
+    savedErrno = errno;
+
+    if ( ret == -1 && IDBLogger::isEnabled() )
+    {
+        IDBLogger::logNoArg(m_fname, this, "fallocate", errno);
+    }
+
+    errno = savedErrno;
+    return ret;
+}
+
 }

utils/idbdatafile/BufferedFile.h

@@ -50,6 +50,7 @@ public:
     /* virtual */ off64_t tell();
     /* virtual */ int flush();
     /* virtual */ time_t mtime();
+    /* virtual */ int fallocate(int mode, off64_t offset, off64_t length);
 
 protected:
     /* virtual */

utils/idbdatafile/IDBDataFile.h

@@ -186,6 +186,12 @@ public:
      */
     virtual time_t mtime() = 0;
 
+    /**
+     * The fallocate() method returns the modification time of the file in
+     * seconds.  Returns -1 on error.
+     */
+    virtual int fallocate(int mode, off64_t offset, off64_t length) = 0;
+
     int colWidth()
     {
         return m_fColWidth;

utils/idbdatafile/UnbufferedFile.cpp

@@ -329,4 +329,21 @@ int UnbufferedFile::close()
     return ret;
 }
 
+int UnbufferedFile::fallocate(int mode, off64_t offset, off64_t length)
+{
+    int ret = 0;
+    int savedErrno = 0;
+
+    ret = ::fallocate( m_fd, mode, offset, length );
+    savedErrno = errno;
+
+    if ( ret == -1 && IDBLogger::isEnabled() )
+    {
+        IDBLogger::logNoArg(m_fname, this, "fallocate", errno);
+    }
+
+    errno = savedErrno;
+    return ret;
+}
+
 }

utils/idbdatafile/UnbufferedFile.h

@@ -47,6 +47,7 @@ public:
     /* virtual */ off64_t tell();
     /* virtual */ int flush();
     /* virtual */ time_t mtime();
+    /* virtual */ int fallocate(int mode, off64_t offset, off64_t length);
 
 protected:
     /* virtual */

utils/idbhdfs/hdfs-shared/HdfsFile.cpp

@@ -348,4 +348,10 @@ int HdfsFile::close()
     return ret;
 }
 
+int HdfsFile::fallocate(int mode, off64_t offset, off64_t length)
+{
+    return 0;
+}
+
+
 }

utils/idbhdfs/hdfs-shared/HdfsFile.h

@@ -62,6 +62,7 @@ public:
     /* virtual */ off64_t tell();
     /* virtual */ int flush();
     /* virtual */ time_t mtime();
+    /* virtual */ int fallocate(int mode, off64_t offset, off64_t length);
 
 protected:
     /* virtual */

utils/idbhdfs/hdfs-shared/HdfsRdwrFileBuffer.cpp

@@ -317,4 +317,9 @@ int HdfsRdwrFileBuffer::close()
     return 0;
 }
 
+int HdfsRdwrFileBuffer::fallocate(int mode, off64_t offset, off64_t length)
+{
+    return 0;
+}
+
 }

utils/idbhdfs/hdfs-shared/HdfsRdwrFileBuffer.h

@@ -62,6 +62,7 @@ public:
     /* virtual */ off64_t tell();
     /* virtual */ int flush();
     /* virtual */ time_t mtime();
+    /* virtual*/ int fallocate(int mode, off64_t offset, off64_t length);
 
 protected:
     /* virtual */

utils/idbhdfs/hdfs-shared/HdfsRdwrMemBuffer.cpp

@@ -507,5 +507,10 @@ int HdfsRdwrMemBuffer::close()
     return 0;
 }
 
+int HdfsRdwrMemBuffer::fallocate(int mode, off64_t offset, off64_t length)
+{
+    return 0;
+}
+
 
 }

utils/idbhdfs/hdfs-shared/HdfsRdwrMemBuffer.h

@@ -62,6 +62,7 @@ public:
     /* virtual */ off64_t tell();
     /* virtual */ int flush();
     /* virtual */ time_t mtime();
+    /* virtual */ int fallocate(int mode, off64_t offset, off64_t length);
 
     // Returns the total size of all currently allocated HdfsRdwrMemBuffers
     static size_t getTotalBuff()

writeengine/shared/we_fileop.cpp

@@ -834,7 +834,8 @@ int FileOp::extendFile(
                            width,
                            newFile, // new or existing file
                            false,   // don't expand; new extent
-                           false ); // add full (not abbreviated) extent
+                           false,   // add full (not abbreviated) extent
+                           true);   // try to use fallocate first
 
     return rc;
 }
@@ -1016,6 +1017,7 @@ int FileOp::addExtentExactFile(
  *                   headers will be included "if" it is a compressed file.
  *    bExpandExtent (in) - Expand existing extent, or initialize a new one
  *    bAbbrevExtent(in) - if creating new extent, is it an abbreviated extent
+ *    bOptExtension(in) - full sized segment file allocation optimization flag
  * RETURN:
  *    returns ERR_FILE_WRITE if an error occurs,
  *    else returns NO_ERROR.
@@ -1028,7 +1030,8 @@ int FileOp::initColumnExtent(
     int      width,
     bool     bNewFile,
     bool     bExpandExtent,
-    bool     bAbbrevExtent )
+    bool     bAbbrevExtent,
+    bool     bOptExtension)
 {
     if ((bNewFile) && (m_compressionType))
     {
@@ -1071,6 +1074,7 @@ int FileOp::initColumnExtent(
         int writeSize = nBlocks * BYTE_PER_BLOCK; // 1M and 8M row extent size
         int loopCount = 1;
         int remWriteSize = 0;
+        off64_t currFileSize = pFile->size();
 
         if (nBlocks > MAX_NBLOCKS)                // 64M row extent size
         {
@@ -1099,10 +1103,14 @@ int FileOp::initColumnExtent(
 
         Stats::startParseEvent(WE_STATS_INIT_COL_EXTENT);
 #endif
-
+        int savedErrno = 0;
-        // Allocate buffer, and store in scoped_array to insure it's deletion.
+        // Try to fallocate the space - fallback to write if fallocate has failed
-        // Create scope {...} to manage deletion of writeBuf.
+        if (!bOptExtension || pFile->fallocate(0, currFileSize, writeSize))
         {
+            // Allocate buffer, store it in scoped_array to insure it's deletion.
+            // Create scope {...} to manage deletion of writeBuf.
+            // Save errno of the failed fallocate() to log it later.
+            savedErrno = errno;
             unsigned char* writeBuf = new unsigned char[writeSize];
             boost::scoped_array<unsigned char> writeBufPtr( writeBuf );
 
@@ -1159,6 +1167,18 @@ int FileOp::initColumnExtent(
             Stats::stopParseEvent(WE_STATS_CREATE_COL_EXTENT);
 
 #endif
+        // Log the fallocate() call result
+        std::ostringstream oss;
+        std::string errnoMsg;
+        Convertor::mapErrnoToString(savedErrno, errnoMsg);
+        oss << "FileOp::initColumnExtent(): fallocate(" << currFileSize <<
+            ", " << writeSize << "): errno = " << savedErrno <<
+            ": " << errnoMsg;
+        logging::Message::Args args;
+        args.add(oss.str());
+        SimpleSysLog::instance()->logMsg(args, logging::LOG_TYPE_INFO,
+             logging::M0006);
+
     }
 
     return NO_ERROR;
@@ -2790,7 +2810,8 @@ int FileOp::expandAbbrevColumnExtent(
     int rc = FileOp::initColumnExtent(pFile, dbRoot, blksToAdd, emptyVal, width,
                                       false,   // existing file
                                       true,    // expand existing extent
-                                      false);  // n/a since not adding new extent
+                                      false,  // n/a since not adding new extent
+                                      true);  // optimize segment file extension
 
     return rc;
 }

writeengine/shared/we_fileop.h

@@ -507,7 +507,8 @@ private:
                                           int      width,
                                           bool     bNewFile,
                                           bool     bExpandExtent,
-                                          bool     bAbbrevExtent );
+                                          bool     bAbbrevExtent,
+                                          bool     bOptExtension=false );
 
     static void         initDbRootExtentMutexes();
     static void         removeDbRootExtentMutexes();