SQLite

** The pager is used to access a database disk file.  It implements
** atomic commit and rollback through the use of a journal file that
** is separate from the database file.  The pager also implements file
** locking to prevent two processes from writing the same database
** file simultaneously, or one process from reading the database while
** another is writing.
**
** @(#) $Id: pager.c,v 1.542 2009/01/08 17:50:46 danielk1977 Exp $
*/
#ifndef SQLITE_OMIT_DISKIO
#include "sqliteInt.h"

/*
** Macros for troubleshooting.  Normally turned off
*/

** ---------------------------------------
** 0                         0
** 512                       512
** 100                       512
** 2000                      2048
** 
*/
static i64 journalHdrOffset(Pager *pPager){
  i64 offset = 0;
  i64 c = pPager->journalOff;
  if( c ){
    offset = ((c-1)/JOURNAL_HDR_SZ(pPager) + 1) * JOURNAL_HDR_SZ(pPager);
  }
  assert( offset%JOURNAL_HDR_SZ(pPager)==0 );
  assert( offset>=c );
  assert( (offset-c)<JOURNAL_HDR_SZ(pPager) );
  return offset;
}
static void seekJournalHdr(Pager *pPager){
  pPager->journalOff = journalHdrOffset(pPager);
}

/*
** Write zeros over the header of the journal file.  This has the
** effect of invalidating the journal file and committing the
** transaction.
*/

  if( pPager->needSync ){
    assert( !pPager->tempFile );
    if( pPager->journalMode!=PAGER_JOURNALMODE_MEMORY ){
      int iDc = sqlite3OsDeviceCharacteristics(pPager->fd);
      assert( pPager->journalOpen );

      if( 0==(iDc&SQLITE_IOCAP_SAFE_APPEND) ){
        i64 jrnlOff = journalHdrOffset(pPager);
        u8 zMagic[8];

        /* This block deals with an obscure problem. If the last connection
        ** that wrote to this database was operating in persistent-journal
        ** mode, then the journal file may at this point actually be larger
        ** than Pager.journalOff bytes. If the next thing in the journal
        ** file happens to be a journal-header (written as part of the
        ** previous connections transaction), and a crash or power-failure 
        ** occurs after nRec is updated but before this connection writes 
        ** anything else to the journal file (or commits/rolls back its 
        ** transaction), then SQLite may become confused when doing the 
        ** hot-journal rollback following recovery. It may roll back all
        ** of this connections data, then proceed to rolling back the old,
        ** out-of-date data that follows it. Database corruption.
        **
        ** To work around this, if the journal file does appear to contain
        ** a valid header following Pager.journalOff, then write a 0x00
        ** byte to the start of it to prevent it from being recognized.
        */
        rc = sqlite3OsRead(pPager->jfd, zMagic, 8, jrnlOff);
        if( rc==SQLITE_OK && 0==memcmp(zMagic, aJournalMagic, 8) ){
          static const u8 zerobyte = 0;
          rc = sqlite3OsWrite(pPager->jfd, &zerobyte, 1, jrnlOff);
        }
        if( rc!=SQLITE_OK && rc!=SQLITE_IOERR_SHORT_READ ){
          return rc;
        }

        /* Write the nRec value into the journal file header. If in
        ** full-synchronous mode, sync the journal first. This ensures that
        ** all data has really hit the disk before nRec is updated to mark
        ** it as a candidate for rollback.
        **
        ** This is not required if the persistent media supports the
        ** SAFE_APPEND property. Because in this case it is not possible 
        ** for garbage data to be appended to the file, the nRec field
        ** is populated with 0xFFFFFFFF when the journal header is written
        ** and never needs to be updated.
        */

        if( pPager->fullSync && 0==(iDc&SQLITE_IOCAP_SEQUENTIAL) ){
          PAGERTRACE(("SYNC journal of %d\n", PAGERID(pPager)));
          IOTRACE(("JSYNC %p\n", pPager))
          rc = sqlite3OsSync(pPager->jfd, pPager->sync_flags);
          if( rc!=0 ) return rc;
        }

      pPager->journalOpen = 1;
      pPager->journalStarted = 0;
      pPager->journalOff = 0;
      pPager->setMaster = 0;
      pPager->journalHdr = 0;
 
      /* Playback and delete the journal.  Drop the database write
      ** lock and reacquire the read lock. Purge the cache before
      ** playing back the hot-journal so that we don't end up with
      */
      sqlite3PcacheClear(pPager->pPCache);
      rc = pager_playback(pPager, 1);
      if( rc!=SQLITE_OK ){
        rc = pager_error(pPager, rc);
        goto failed;
      }
      assert(pPager->state==PAGER_SHARED || 
          (pPager->exclusiveMode && pPager->state>PAGER_SHARED)

# 2009 January 8
#
# The author disclaims copyright to this source code.  In place of
# a legal notice, here is a blessing:
#
#    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.
#
#***********************************************************************
#
# This test verifies a couple of specific potential data corruption 
# scenarios involving crashes or power failures.
#
# $Id: crash8.test,v 1.1 2009/01/08 17:50:46 danielk1977 Exp $


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

ifcapable !crashtest {
  finish_test
  return
}

do_test crash8-1.1 {
  execsql {
    CREATE TABLE t1(a, b);
    CREATE INDEX i1 ON t1(a, b);
    INSERT INTO t1 VALUES(1, randstr(1000,1000));
    INSERT INTO t1 VALUES(2, randstr(1000,1000));
    INSERT INTO t1 VALUES(3, randstr(1000,1000));
    INSERT INTO t1 VALUES(4, randstr(1000,1000));
    INSERT INTO t1 VALUES(5, randstr(1000,1000));
    INSERT INTO t1 VALUES(6, randstr(1000,1000));
    CREATE TABLE t2(a, b);
    CREATE TABLE t3(a, b);
    CREATE TABLE t4(a, b);
    CREATE TABLE t5(a, b);
    CREATE TABLE t6(a, b);
    CREATE TABLE t7(a, b);
    CREATE TABLE t8(a, b);
    CREATE TABLE t9(a, b);
    CREATE TABLE t10(a, b);
    PRAGMA integrity_check
  }
} {ok}


# Potential corruption scenario 1. A second process opens the database 
# and modifies a large portion of it. It then opens a second transaction
# and modifies a small part of the database, but crashes before it commits
# the transaction. 
#
# When the first process accessed the database again, it was rolling back
# the aborted transaction, but was not purging its in-memory cache (which
# was loaded before the second process made its first, successful, 
# modification). Producing an inconsistent cache.
#
do_test crash8-1.2 {
  crashsql -delay 2 -file test.db {
    PRAGMA cache_size = 10;
    UPDATE t1 SET b = randstr(1000,1000);
    INSERT INTO t9 VALUES(1, 2);
  }
} {1 {child process exited abnormally}}
do_test crash8-1.3 {
  execsql {PRAGMA integrity_check}
} {ok}

# Potential corruption scenario 2. The second process, operating in
# persistent-journal mode, makes a large change to the database file
# with a small in-memory cache. Such that more than one journal-header
# was written to the file. It then opens a second transaction and makes
# a smaller change that requires only a single journal-header to be
# written to the journal file. The second change is such that the 
# journal content written to the persistent journal file exactly overwrites
# the first journal-header and set of subsequent records written by the
# first, successful, change. The second process crashes before it can
# commit its second change.
#
# When the first process accessed the database again, it was rolling back
# the second aborted transaction, then continuing to rollback the second
# and subsequent journal-headers written by the first, successful, change.
# Database corruption.
#
do_test crash8.2.1 {
  crashsql -delay 2 -file test.db {
    PRAGMA journal_mode = persist;
    PRAGMA cache_size = 10;
    UPDATE t1 SET b = randstr(1000,1000);
    PRAGMA cache_size = 100;
    BEGIN;
      INSERT INTO t2 VALUES('a', 'b');
      INSERT INTO t3 VALUES('a', 'b');
      INSERT INTO t4 VALUES('a', 'b');
      INSERT INTO t5 VALUES('a', 'b');
      INSERT INTO t6 VALUES('a', 'b');
      INSERT INTO t7 VALUES('a', 'b');
      INSERT INTO t8 VALUES('a', 'b');
      INSERT INTO t9 VALUES('a', 'b');
      INSERT INTO t10 VALUES('a', 'b');
    COMMIT;
  }
} {1 {child process exited abnormally}}

do_test crash8-2.3 {
  execsql {PRAGMA integrity_check}
} {ok}

finish_test