/*

** 2004 May 22

**

** 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 file contains code that modified the OS layer in order to simulate

** the effect on the database file of an OS crash or power failure.  This

** is used to test the ability of SQLite to recover from those situations.

**

** $Id: test6.c,v 1.39 2008/06/06 11:11:26 danielk1977 Exp $

*/

#if SQLITE_TEST          /* This file is used for testing only */

#include "sqliteInt.h"

#include "tcl.h"

#ifndef SQLITE_OMIT_DISKIO  /* This file is a no-op if disk I/O is disabled */

/* #define TRACE_CRASHTEST */

typedef struct CrashFile CrashFile;

typedef struct CrashGlobal CrashGlobal;

typedef struct WriteBuffer WriteBuffer;

/*

** Method:

**

**   This layer is implemented as a wrapper around the "real" 

**   sqlite3_file object for the host system. Each time data is 

**   written to the file object, instead of being written to the

**   underlying file, the write operation is stored in an in-memory 

**   structure (type WriteBuffer). This structure is placed at the

**   end of a global ordered list (the write-list).

**

**   When data is read from a file object, the requested region is

**   first retrieved from the real file. The write-list is then 

**   traversed and data copied from any overlapping WriteBuffer 

**   structures to the output buffer. i.e. a read() operation following

**   one or more write() operations works as expected, even if no

**   data has actually been written out to the real file.

**

**   When a fsync() operation is performed, an operating system crash 

**   may be simulated, in which case exit(-1) is called (the call to 

**   xSync() never returns). Whether or not a crash is simulated,

**   the data associated with a subset of the WriteBuffer structures 

**   stored in the write-list is written to the real underlying files 

**   and the entries removed from the write-list. If a crash is simulated,

**   a subset of the buffers may be corrupted before the data is written.

**

**   The exact subset of the write-list written and/or corrupted is

**   determined by the simulated device characteristics and sector-size.

**

** "Normal" mode:

**

**   Normal mode is used when the simulated device has none of the

**   SQLITE_IOCAP_XXX flags set.

**

**   In normal mode, if the fsync() is not a simulated crash, the 

**   write-list is traversed from beginning to end. Each WriteBuffer

**   structure associated with the file handle used to call xSync()

**   is written to the real file and removed from the write-list.

**

**   If a crash is simulated, one of the following takes place for 

**   each WriteBuffer in the write-list, regardless of which 

**   file-handle it is associated with:

**

**     1. The buffer is correctly written to the file, just as if

**        a crash were not being simulated.

**

**     2. Nothing is done.

**

**     3. Garbage data is written to all sectors of the file that 

**        overlap the region specified by the WriteBuffer. Or garbage

**        data is written to some contiguous section within the 

**        overlapped sectors.

**

** Device Characteristic flag handling:

**

**   If the IOCAP_ATOMIC flag is set, then option (3) above is 

**   never selected.

**

**   If the IOCAP_ATOMIC512 flag is set, and the WriteBuffer represents

**   an aligned write() of an integer number of 512 byte regions, then

**   option (3) above is never selected. Instead, each 512 byte region

**   is either correctly written or left completely untouched. Similar

**   logic governs the behaviour if any of the other ATOMICXXX flags

**   is set.

**

**   If either the IOCAP_SAFEAPPEND or IOCAP_SEQUENTIAL flags are set

**   and a crash is being simulated, then an entry of the write-list is

**   selected at random. Everything in the list after the selected entry 

**   is discarded before processing begins.

**

**   If IOCAP_SEQUENTIAL is set and a crash is being simulated, option 

**   (1) is selected for all write-list entries except the last. If a 

**   crash is not being simulated, then all entries in the write-list

**   that occur before at least one write() on the file-handle specified

**   as part of the xSync() are written to their associated real files.

**

**   If IOCAP_SAFEAPPEND is set and the first byte written by the write()

**   operation is one byte past the current end of the file, then option

**   (1) is always selected.

*/

/*

** Each write operation in the write-list is represented by an instance

** of the following structure.

**

** If zBuf is 0, then this structure represents a call to xTruncate(), 

** not xWrite(). In that case, iOffset is the size that the file is

** truncated to.

*/

struct WriteBuffer {

  i64 iOffset;                 /* Byte offset of the start of this write() */

  int nBuf;                    /* Number of bytes written */

  u8 *zBuf;                    /* Pointer to copy of written data */

  CrashFile *pFile;            /* File this write() applies to */

  WriteBuffer *pNext;          /* Next in CrashGlobal.pWriteList */

};

struct CrashFile {

  const sqlite3_io_methods *pMethod;   /* Must be first */

  sqlite3_file *pRealFile;             /* Underlying "real" file handle */

  char *zName;

  /* Cache of the entire file. This is used to speed up OsRead() and 

  ** OsFileSize() calls. Although both could be done by traversing the

  ** write-list, in practice this is impractically slow.

  */

  int iSize;                           /* Size of file in bytes */

  int nData;                           /* Size of buffer allocated at zData */

  u8 *zData;                           /* Buffer containing file contents */

};

struct CrashGlobal {

  WriteBuffer *pWriteList;     /* Head of write-list */

  WriteBuffer *pWriteListEnd;  /* End of write-list */

  int iSectorSize;             /* Value of simulated sector size */

  int iDeviceCharacteristics;  /* Value of simulated device characteristics */

  int iCrash;                  /* Crash on the iCrash'th call to xSync() */

  char zCrashFile[500];        /* Crash during an xSync() on this file */ 

};

static CrashGlobal g = {0, 0, SQLITE_DEFAULT_SECTOR_SIZE, 0, 0};

/*

** Set this global variable to 1 to enable crash testing.

*/

static int sqlite3CrashTestEnable = 0;

static void *crash_malloc(int nByte){

  return (void *)Tcl_Alloc((size_t)nByte);

}

static void crash_free(void *p){

  Tcl_Free(p);

}

static void *crash_realloc(void *p, int n){

  return (void *)Tcl_Realloc(p, (size_t)n);

}

/*

** Flush the write-list as if xSync() had been called on file handle

** pFile. If isCrash is true, simulate a crash.

*/

static int writeListSync(CrashFile *pFile, int isCrash){

  int rc = SQLITE_OK;

  int iDc = g.iDeviceCharacteristics;

  WriteBuffer *pWrite;

  WriteBuffer **ppPtr;

  /* If this is not a crash simulation, set pFinal to point to the 

  ** last element of the write-list that is associated with file handle

  ** pFile.

  **

  ** If this is a crash simulation, set pFinal to an arbitrarily selected

  ** element of the write-list.

  */

  WriteBuffer *pFinal = 0;

  if( !isCrash ){

    for(pWrite=g.pWriteList; pWrite; pWrite=pWrite->pNext){

      if( pWrite->pFile==pFile ){

        pFinal = pWrite;

      }

    }

  }else if( iDc&(SQLITE_IOCAP_SEQUENTIAL|SQLITE_IOCAP_SAFE_APPEND) ){

    int nWrite = 0;

    int iFinal;

    for(pWrite=g.pWriteList; pWrite; pWrite=pWrite->pNext) nWrite++;

    sqlite3_randomness(sizeof(int), &iFinal);

    iFinal = ((iFinal<0)?-1*iFinal:iFinal)%nWrite;

    for(pWrite=g.pWriteList; iFinal>0; pWrite=pWrite->pNext) iFinal--;

    pFinal = pWrite;

  }

#ifdef TRACE_CRASHTEST

  printf("Sync %s (is %s crash)\n", pFile->zName, (isCrash?"a":"not a"));

#endif

  ppPtr = &g.pWriteList;

  for(pWrite=*ppPtr; rc==SQLITE_OK && pWrite; pWrite=*ppPtr){

    sqlite3_file *pRealFile = pWrite->pFile->pRealFile;

    /* (eAction==1)      -> write block out normally,

    ** (eAction==2)      -> do nothing,

    ** (eAction==3)      -> trash sectors.

    */

    int eAction = 0;

    if( !isCrash ){

      eAction = 2;

      if( (pWrite->pFile==pFile || iDc&SQLITE_IOCAP_SEQUENTIAL) ){

        eAction = 1;

      }

    }else{

      char random;

      sqlite3_randomness(1, &random);

      /* Do not select option 3 (sector trashing) if the IOCAP_ATOMIC flag 

      ** is set or this is an OsTruncate(), not an Oswrite().

      */

      if( (iDc&SQLITE_IOCAP_ATOMIC) || (pWrite->zBuf==0) ){

        random &= 0x01;

      }

      /* If IOCAP_SEQUENTIAL is set and this is not the final entry

      ** in the truncated write-list, always select option 1 (write

      ** out correctly).

      */

      if( (iDc&SQLITE_IOCAP_SEQUENTIAL && pWrite!=pFinal) ){

        random = 0;

      }

      /* If IOCAP_SAFE_APPEND is set and this OsWrite() operation is

      ** an append (first byte of the written region is 1 byte past the

      ** current EOF), always select option 1 (write out correctly).

      */

      if( iDc&SQLITE_IOCAP_SAFE_APPEND && pWrite->zBuf ){

        i64 iSize;

        sqlite3OsFileSize(pRealFile, &iSize);

        if( iSize==pWrite->iOffset ){

          random = 0;

        }

      }

      if( (random&0x06)==0x06 ){

        eAction = 3;

      }else{

        eAction = ((random&0x01)?2:1);

      }

    }

    switch( eAction ){

      case 1: {               /* Write out correctly */

        if( pWrite->zBuf ){

          rc = sqlite3OsWrite(

              pRealFile, pWrite->zBuf, pWrite->nBuf, pWrite->iOffset

          );

        }else{

          rc = sqlite3OsTruncate(pRealFile, pWrite->iOffset);

        }

        *ppPtr = pWrite->pNext;

#ifdef TRACE_CRASHTEST

        if( isCrash ){

          printf("Writing %d bytes @ %d (%s)\n", 

            pWrite->nBuf, (int)pWrite->iOffset, pWrite->pFile->zName

          );

        }

#endif

        crash_free(pWrite);

        break;

      }

      case 2: {               /* Do nothing */

        ppPtr = &pWrite->pNext;

#ifdef TRACE_CRASHTEST

        if( isCrash ){

          printf("Omiting %d bytes @ %d (%s)\n", 

            pWrite->nBuf, (int)pWrite->iOffset, pWrite->pFile->zName

          );

        }

#endif

        break;

      }

      case 3: {               /* Trash sectors */

        u8 *zGarbage;

        int iFirst = (pWrite->iOffset/g.iSectorSize);

        int iLast = (pWrite->iOffset+pWrite->nBuf-1)/g.iSectorSize;

        assert(pWrite->zBuf);

#ifdef TRACE_CRASHTEST

        printf("Trashing %d sectors @ sector %d (%s)\n", 

            1+iLast-iFirst, iFirst, pWrite->pFile->zName

        );

#endif

        zGarbage = crash_malloc(g.iSectorSize);

        if( zGarbage ){

          sqlite3_int64 i;

          for(i=iFirst; rc==SQLITE_OK && i<=iLast; i++){

            sqlite3_randomness(g.iSectorSize, zGarbage); 

            rc = sqlite3OsWrite(

              pRealFile, zGarbage, g.iSectorSize, i*g.iSectorSize

            );

          }

          crash_free(zGarbage);

        }else{

          rc = SQLITE_NOMEM;

        }

        ppPtr = &pWrite->pNext;

        break;

      }

      default:

    }

    if( pWrite==pFinal ) break;

  }

  if( rc==SQLITE_OK && isCrash ){

    exit(-1);

  }

  for(pWrite=g.pWriteList; pWrite && pWrite->pNext; pWrite=pWrite->pNext);

  g.pWriteListEnd = pWrite;

  return rc;

}

/*

** Add an entry to the end of the write-list.

*/

static int writeListAppend(

  sqlite3_file *pFile,

  sqlite3_int64 iOffset,

  const u8 *zBuf,

  int nBuf

){

  WriteBuffer *pNew;

  assert((zBuf && nBuf) || (!nBuf && !zBuf));

  pNew = (WriteBuffer *)crash_malloc(sizeof(WriteBuffer) + nBuf);

  if( pNew==0 ){

    fprintf(stderr, "out of memory in the crash simulator\n");

  }

  memset(pNew, 0, sizeof(WriteBuffer)+nBuf);

  pNew->iOffset = iOffset;

  pNew->nBuf = nBuf;

  pNew->pFile = (CrashFile *)pFile;

  if( zBuf ){

    pNew->zBuf = (u8 *)&pNew[1];

    memcpy(pNew->zBuf, zBuf, nBuf);

  }

  if( g.pWriteList ){

    assert(g.pWriteListEnd);

    g.pWriteListEnd->pNext = pNew;

  }else{

    g.pWriteList = pNew;

  }

  g.pWriteListEnd = pNew;

  return SQLITE_OK;

}

/*

** Close a crash-file.

*/

static int cfClose(sqlite3_file *pFile){

  CrashFile *pCrash = (CrashFile *)pFile;

  writeListSync(pCrash, 0);

  sqlite3OsClose(pCrash->pRealFile);

  return SQLITE_OK;

}

/*

** Read data from a crash-file.

*/

static int cfRead(

  sqlite3_file *pFile, 

  void *zBuf, 

  int iAmt, 

  sqlite_int64 iOfst

){

  CrashFile *pCrash = (CrashFile *)pFile;

  /* Check the file-size to see if this is a short-read */

  if( pCrash->iSize<(iOfst+iAmt) ){

    return SQLITE_IOERR_SHORT_READ;

  }

  memcpy(zBuf, &pCrash->zData[iOfst], iAmt);

  return SQLITE_OK;

}

/*

** Write data to a crash-file.

*/

static int cfWrite(

  sqlite3_file *pFile, 

  const void *zBuf, 

  int iAmt, 

  sqlite_int64 iOfst

){

  CrashFile *pCrash = (CrashFile *)pFile;

  if( iAmt+iOfst>pCrash->iSize ){

    pCrash->iSize = iAmt+iOfst;

  }

  while( pCrash->iSize>pCrash->nData ){

    u8 *zNew;

    int nNew = (pCrash->nData*2) + 4096;

    zNew = crash_realloc(pCrash->zData, nNew);

    if( !zNew ){

      return SQLITE_NOMEM;

    }

    memset(&zNew[pCrash->nData], 0, nNew-pCrash->nData);

    pCrash->nData = nNew;

    pCrash->zData = zNew;

  }

  memcpy(&pCrash->zData[iOfst], zBuf, iAmt);

  return writeListAppend(pFile, iOfst, zBuf, iAmt);

}

/*

** Truncate a crash-file.

*/

static int cfTruncate(sqlite3_file *pFile, sqlite_int64 size){

  CrashFile *pCrash = (CrashFile *)pFile;

  assert(size>=0);

  if( pCrash->iSize>size ){

    pCrash->iSize = size;

  }

  return writeListAppend(pFile, size, 0, 0);

}

/*

** Sync a crash-file.

*/

static int cfSync(sqlite3_file *pFile, int flags){

  CrashFile *pCrash = (CrashFile *)pFile;

  int isCrash = 0;

  const char *zName = pCrash->zName;

  const char *zCrashFile = g.zCrashFile;

  int nName = strlen(zName);

  int nCrashFile = strlen(zCrashFile);

  if( nCrashFile>0 && zCrashFile[nCrashFile-1]=='*' ){

    nCrashFile--;

    if( nName>nCrashFile ) nName = nCrashFile;

  }

  if( nName==nCrashFile && 0==memcmp(zName, zCrashFile, nName) ){

    if( (--g.iCrash)==0 ) isCrash = 1;

  }

  return writeListSync(pCrash, isCrash);

}

/*

** Return the current file-size of the crash-file.

*/

static int cfFileSize(sqlite3_file *pFile, sqlite_int64 *pSize){

  CrashFile *pCrash = (CrashFile *)pFile;

  *pSize = (i64)pCrash->iSize;

  return SQLITE_OK;

}

/*

** Calls related to file-locks are passed on to the real file handle.

*/

static int cfLock(sqlite3_file *pFile, int eLock){

  return sqlite3OsLock(((CrashFile *)pFile)->pRealFile, eLock);

}

static int cfUnlock(sqlite3_file *pFile, int eLock){

  return sqlite3OsUnlock(((CrashFile *)pFile)->pRealFile, eLock);

}

static int cfCheckReservedLock(sqlite3_file *pFile, int *pResOut){

  return sqlite3OsCheckReservedLock(((CrashFile *)pFile)->pRealFile, pResOut);

}

static int cfFileControl(sqlite3_file *pFile, int op, void *pArg){

  return sqlite3OsFileControl(((CrashFile *)pFile)->pRealFile, op, pArg);

}

/*

** The xSectorSize() and xDeviceCharacteristics() functions return

** the global values configured by the [sqlite_crashparams] tcl

*  interface.

*/

static int cfSectorSize(sqlite3_file *pFile){

  return g.iSectorSize;

}

static int cfDeviceCharacteristics(sqlite3_file *pFile){

  return g.iDeviceCharacteristics;

}

static const sqlite3_io_methods CrashFileVtab = {

  1,                            /* iVersion */

  cfClose,                      /* xClose */

  cfRead,                       /* xRead */

  cfWrite,                      /* xWrite */

  cfTruncate,                   /* xTruncate */

  cfSync,                       /* xSync */

  cfFileSize,                   /* xFileSize */

  cfLock,                       /* xLock */

  cfUnlock,                     /* xUnlock */

  cfCheckReservedLock,          /* xCheckReservedLock */

  cfFileControl,                /* xFileControl */

  cfSectorSize,                 /* xSectorSize */

  cfDeviceCharacteristics       /* xDeviceCharacteristics */

};

/*

** Application data for the crash VFS

*/

struct crashAppData {

  sqlite3_vfs *pOrig;                   /* Wrapped vfs structure */