SQLite

  p->db = db;
  p->nColumn = nCol;
  p->nPendingData = 0;
  p->azColumn = (char **)&p[1];
  p->pTokenizer = pTokenizer;
  p->nNodeSize = 1000;
  p->nMaxPendingData = FTS3_MAX_PENDING_DATA;
  zCsr = (char *)&p->azColumn[nCol];

  fts3HashInit(&p->pendingTerms, FTS3_HASH_STRING, 1);

  /* Fill in the zName and zDb fields of the vtab structure. */
  p->zName = zCsr;
  memcpy(zCsr, argv[2], nName);

#ifdef SQLITE_ENABLE_ICU
     || (pIcu && sqlite3Fts3HashInsert(pHash, "icu", 4, (void *)pIcu))
#endif
    ){
      rc = SQLITE_NOMEM;
    }
  }

#ifdef SQLITE_TEST
  sqlite3Fts3ExprInitTestInterface(db);
#endif

  /* Create the virtual table wrapper around the hash-table and overload 
  ** the two scalar functions. If this is successful, register the
  ** module with sqlite.
  */
  if( SQLITE_OK==rc 
   && SQLITE_OK==(rc = sqlite3Fts3InitHashTable(db, pHash, "fts3_tokenizer"))

  sqlite3_stmt **aLeavesStmt;     /* Array of prepared zSelectLeaves stmts */

  int nNodeSize;                  /* Soft limit for node size */

  /* The following hash table is used to buffer pending index updates during
  ** transactions. Variable nPendingData estimates the memory size of the 
  ** pending data, including hash table overhead, but not malloc overhead. 
  ** When nPendingData exceeds nMaxPendingData, the buffer is flushed 
  ** automatically. Variable iPrevDocid is the docid of the most recently
  ** inserted record.
  */
  int nMaxPendingData;
  int nPendingData;
  sqlite_int64 iPrevDocid;
  Fts3Hash pendingTerms;
};

/*
** When the core wants to read from the virtual table, it creates a

    rc = sqlite3_step(pStmt); 
    if( rc!=SQLITE_ROW ){
      return (rc==SQLITE_DONE ? SQLITE_CORRUPT : rc);
    }
  
    *pnBlock = sqlite3_column_bytes(pStmt, 0);
    *pzBlock = (char *)sqlite3_column_blob(pStmt, 0);
    if( sqlite3_column_type(pStmt, 0)!=SQLITE_BLOB ){
      return SQLITE_CORRUPT;
    }
  }
  return SQLITE_OK;
}

/*
** Set *ppStmt to a statement handle that may be used to iterate through

static int fts3PendingTermsDocid(Fts3Table *p, sqlite_int64 iDocid){
  /* TODO(shess) Explore whether partially flushing the buffer on
  ** forced-flush would provide better performance.  I suspect that if
  ** we ordered the doclists by size and flushed the largest until the
  ** buffer was half empty, that would let the less frequent terms
  ** generate longer doclists.
  */
  if( iDocid<=p->iPrevDocid || p->nPendingData>p->nMaxPendingData ){
    int rc = sqlite3Fts3PendingTermsFlush(p);
    if( rc!=SQLITE_OK ) return rc;
  }
  p->iPrevDocid = iDocid;
  return SQLITE_OK;
}

    if( rc==SQLITE_DONE || rc==SQLITE_OK ){
      rc = SQLITE_OK;
      sqlite3Fts3PendingTermsClear(p);
    }
#ifdef SQLITE_TEST
  }else if( nVal>9 && 0==sqlite3_strnicmp(zVal, "nodesize=", 9) ){
    p->nNodeSize = atoi(&zVal[9]);
    rc = SQLITE_OK;
  }else if( nVal>11 && 0==sqlite3_strnicmp(zVal, "maxpending=", 9) ){
    p->nMaxPendingData = atoi(&zVal[11]);
    rc = SQLITE_OK;
#endif
  }else{
    rc = SQLITE_ERROR;
  }

  return rc;

  x += y * (*q++);
  *v = (sqlite_int64) x;
  return (int) (q - (unsigned char *)p);
}


/*
** USAGE:  read_fts3varint BLOB VARNAME
**
** Read a varint from the start of BLOB. Set variable VARNAME to contain
** the interpreted value. Return the number of bytes of BLOB consumed.
*/
static int read_fts3varint(
  void * clientData,
  Tcl_Interp *interp,
  int objc,
  Tcl_Obj *CONST objv[]
){
  int nBlob;
  unsigned char *zBlob;

  } aObjCmd[] = {
     { "hexio_read",                   hexio_read            },
     { "hexio_write",                  hexio_write           },
     { "hexio_get_int",                hexio_get_int         },
     { "hexio_render_int16",           hexio_render_int16    },
     { "hexio_render_int32",           hexio_render_int32    },
     { "utf8_to_utf8",                 utf8_to_utf8          },
     { "read_fts3varint",              read_fts3varint       },
  };
  int i;
  for(i=0; i<sizeof(aObjCmd)/sizeof(aObjCmd[0]); i++){
    Tcl_CreateObjCommand(interp, aObjCmd[i].zName, aObjCmd[i].xProc, 0, 0);
  }
  return TCL_OK;
}

  join $lDoc " "
}

###########################################################################

proc gobble_varint {varname} {
  upvar $varname blob
  set n [read_fts3varint $blob ret]
  set blob [string range $blob $n end]
  return $ret
}
proc gobble_string {varname nLength} {
  upvar $varname blob
  set ret [string range $blob 0 [expr $nLength-1]]
  set blob [string range $blob $nLength end]

# by parameter $result, or (b) TCL throws an "out of memory" error.
#
# If DO_MALLOC_TEST is defined and set to zero, then the SELECT statement
# is executed just once. In this case the test case passes if the results
# match the expected results passed via parameter $result.
#
proc do_select_test {name sql result} {
  uplevel [list doPassiveTest 0 $name $sql [list 0 $result]]
}

proc do_restart_select_test {name sql result} {
  uplevel [list doPassiveTest 1 $name $sql [list 0 $result]]
}

proc do_error_test {name sql error} {
  uplevel [list doPassiveTest 0 $name $sql [list 1 $error]]
}

proc doPassiveTest {isRestart name sql catchres} {
  if {![info exists ::DO_MALLOC_TEST]} { set ::DO_MALLOC_TEST 1 }

  switch $::DO_MALLOC_TEST {
    0 { # No malloc failures.
      do_test $name [list catchsql $sql] $catchres
      return
    }
    1 { # Simulate transient failures.
      set nRepeat 1
      set zName "transient"
      set nStartLimit 100000
      set nBackup 1
    }
    2 { # Simulate persistent failures.
      set nRepeat 1
      set zName "persistent"
      set nStartLimit 100000
      set nBackup 1
    }
    3 { # Simulate transient failures with extra brute force.
      set nRepeat 100000
      set zName "ridiculous"
      set nStartLimit 1
      set nBackup 10
    }
  }

  # The set of acceptable results from running [catchsql $sql].
  #
  set answers [list {1 {out of memory}} $catchres]
  set str [join $answers " OR "]

  set nFail 1
  for {set iLimit $nStartLimit} {$nFail} {incr iLimit} {
    for {set iFail 1} {$nFail && $iFail<=$iLimit} {incr iFail} {
      for {set iTest 0} {$iTest<$nBackup && ($iFail-$iTest)>0} {incr iTest} {

        if {$isRestart} { sqlite3 db test.db }

        sqlite3_memdebug_fail [expr $iFail-$iTest] -repeat $nRepeat
        set res [uplevel [list catchsql $sql]]
        if {[lsearch -exact $answers $res]>=0} { set res $str }


        set testname "$name.$zName.$iFail"

        do_test "$name.$zName.$iLimit.$iFail" [list set {} $res] $str

        set nFail [sqlite3_memdebug_fail -1 -benigncnt nBenign]

      }
    }
  }
}


#-------------------------------------------------------------------------
# Test a single write to the database. In this case a  "write" is a 

# 2009 December 03
#
#    May you do good and not evil.
#    May you find forgiveness for yourself and forgive others.
#    May you share freely, never taking more than you give.
#
#***********************************************************************
#
# The tests in this file are structural coverage tests. They are designed
# to complement the tests in fts3rnd.test and fts3doc.test. Between them,
# the three files should provide full coverage of the fts3 extension code.
#

set testdir [file dirname $argv0]
source $testdir/tester.tcl

# If this build does not include FTS3, skip the tests in this file.
#
ifcapable !fts3 { finish_test ; return }
source $testdir/fts3_common.tcl

set DO_MALLOC_TEST 0

#--------------------------------------------------------------------------
# When it first needs to read a block from the %_segments table, the FTS3 
# module compiles an SQL statement for that purpose. The statement is 
# stored and reused each subsequent time a block is read. This test case 
# tests the effects of an OOM error occuring while compiling the statement.
#
# Similarly, when FTS3 first needs to scan through a set of segment leaves
# to find a set of documents that matches a term, it allocates a string
# containing the text of the required SQL, and compiles one or more 
# statements to traverse the leaves. This test case tests that OOM errors
# that occur while allocating this string and statement are handled correctly
# also.
#
do_test fts3cov-1.1 {
  execsql { 
    CREATE VIRTUAL TABLE t1 USING fts3(x);
    INSERT INTO t1(t1) VALUES('nodesize=24');
    BEGIN;
      INSERT INTO t1 VALUES('Is the night chilly and dark?');
      INSERT INTO t1 VALUES('The night is chilly, but not dark.');
      INSERT INTO t1 VALUES('The thin gray cloud is spread on high,');
      INSERT INTO t1 VALUES('It covers but not hides the sky.');
    COMMIT;
    SELECT count(*)>0 FROM t1_segments;
  }
} {1}

set DO_MALLOC_TEST 1
do_restart_select_test fts3cov-1.2 {
  SELECT docid FROM t1 WHERE t1 MATCH 'chilly';
} {1 2}
set DO_MALLOC_TEST 0

#--------------------------------------------------------------------------
# When querying the full-text index, if an expected internal node block is 
# missing from the %_segments table, or if a NULL value is stored in the 
# %_segments table instead of a binary blob, database corruption should be 
# reported.
#
# Even with tiny 24 byte nodes, it takes a fair bit of data to produce a
# segment b-tree that uses the %_segments table to store internal nodes. 
#
do_test fts3cov-2.1 {
  execsql {
    INSERT INTO t1(t1) VALUES('nodesize=24');
    BEGIN;
      INSERT INTO t1 VALUES('The moon is behind, and at the full;');
      INSERT INTO t1 VALUES('And yet she looks both small and dull.');
      INSERT INTO t1 VALUES('The night is chill, the cloud is gray:');
      INSERT INTO t1 VALUES('''T is a month before the month of May,');
      INSERT INTO t1 VALUES('And the Spring comes slowly up this way.');
      INSERT INTO t1 VALUES('The lovely lady, Christabel,');
      INSERT INTO t1 VALUES('Whom her father loves so well,');
      INSERT INTO t1 VALUES('What makes her in the wood so late,');
      INSERT INTO t1 VALUES('A furlong from the castle gate?');
      INSERT INTO t1 VALUES('She had dreams all yesternight');
      INSERT INTO t1 VALUES('Of her own betrothed knight;');
      INSERT INTO t1 VALUES('And she in the midnight wood will pray');
      INSERT INTO t1 VALUES('For the weal of her lover that''s far away.');
    COMMIT;

    INSERT INTO t1(t1) VALUES('optimize');
    SELECT substr(hex(root), 1, 2) FROM t1_segdir;
  }
} {03}

# Test the "missing entry" case:
do_test fts3cov-2.1 {
  set root [db one {SELECT root FROM t1_segdir}]
  read_fts3varint [string range $root 1 end] left_child
  execsql { DELETE FROM t1_segments WHERE blockid = $left_child }
} {}
do_error_test fts3cov-2.2 {
  SELECT * FROM t1 WHERE t1 MATCH 'c*'
} {database disk image is malformed}

# Test the "replaced with NULL" case:
do_test fts3cov-2.3 {
  execsql { INSERT INTO t1_segments VALUES($left_child, NULL) }
} {}
do_error_test fts3cov-2.4 {
  SELECT * FROM t1 WHERE t1 MATCH 'cloud'
} {database disk image is malformed}

#--------------------------------------------------------------------------
# The following tests are to test the effects of OOM errors while storing
# terms in the pending-hash table. Specifically, while creating doclist
# blobs to store in the table. More specifically, to test OOM errors while
# appending column numbers to doclists. For example, if a doclist consists
# of:
#
#   <docid> <column 0 offset-list> 0x01 <column N> <column N offset-list>
#
# The following tests check that malloc errors encountered while appending
# the "0x01 <column N>" data to the dynamically growable blob used to 
# accumulate the doclist in memory are handled correctly.
#
do_test fts3cov-3.1 {
  set cols [list]
  set vals [list]
  for {set i 0} {$i < 120} {incr i} {
    lappend cols "col$i"
    lappend vals "'word'"
  }
  execsql "CREATE VIRTUAL TABLE t2 USING fts3([join $cols ,])"
} {}
set DO_MALLOC_TEST 1 
do_write_test fts3cov-3.2 t2_content "
  INSERT INTO t2(docid, [join $cols ,]) VALUES(1, [join $vals ,])
"
do_write_test fts3cov-3.3 t2_content "
  INSERT INTO t2(docid, [join $cols ,]) VALUES(200, [join $vals ,])
"
do_write_test fts3cov-3.4 t2_content "
  INSERT INTO t2(docid, [join $cols ,]) VALUES(60000, [join $vals ,])
"

#-------------------------------------------------------------------------
# If too much data accumulates in the pending-terms hash table, it is
# flushed to the database automatically, even if the transaction has not
# finished. The following tests check the effects of encountering an OOM 
# while doing this.
#
do_test fts3cov-4.1 {
  execsql {
    CREATE VIRTUAL TABLE t3 USING fts3(x);
    INSERT INTO t3(t3) VALUES('nodesize=24');
    INSERT INTO t3(t3) VALUES('maxpending=100');
  }
} {}
set DO_MALLOC_TEST 1 
do_write_test fts3cov-4.2 t3_content {
  INSERT INTO t3(docid, x)
    SELECT 1, 'Then Christabel stretched forth her hand,' UNION ALL
    SELECT 3, 'And comforted fair Geraldine:'             UNION ALL
    SELECT 4, '''O well, bright dame, may you command'    UNION ALL
    SELECT 5, 'The service of Sir Leoline;'               UNION ALL
    SELECT 2, 'And gladly our stout chivalry'             UNION ALL
    SELECT 7, 'Will he send forth, and friends withal,'   UNION ALL
    SELECT 8, 'To guide and guard you safe and free'      UNION ALL
    SELECT 6, 'Home to your noble father''s hall.'''
}

finish_test