src/memjournal.c

   1 /*
   2 ** 2008 October 7
   3 **
   4 ** The author disclaims copyright to this source code.  In place of
   5 ** a legal notice, here is a blessing:
   6 **
   7 **    May you do good and not evil.
   8 **    May you find forgiveness for yourself and forgive others.
   9 **    May you share freely, never taking more than you give.
  10 **
  11 *************************************************************************
  12 **
  13 ** This file contains code use to implement an in-memory rollback journal.
  14 ** The in-memory rollback journal is used to journal transactions for
  15 ** ":memory:" databases and when the journal_mode=MEMORY pragma is used.
  16 **
  17 ** Update:  The in-memory journal is also used to temporarily cache
  18 ** smaller journals that are not critical for power-loss recovery.
  19 ** For example, statement journals that are not too big will be held
  20 ** entirely in memory, thus reducing the number of file I/O calls, and
  21 ** more importantly, reducing temporary file creation events.  If these
  22 ** journals become too large for memory, they are spilled to disk.  But
  23 ** in the common case, they are usually small and no file I/O needs to
  24 ** occur.
  25 */
  26 #include "sqliteInt.h"
  27
  28 /* Forward references to internal structures */
  29 typedef struct MemJournal MemJournal;
  30 typedef struct FilePoint FilePoint;
  31 typedef struct FileChunk FileChunk;
  32
  33 /*
  34 ** The rollback journal is composed of a linked list of these structures.
  35 **
  36 ** The zChunk array is always at least 8 bytes in size - usually much more.
  37 ** Its actual size is stored in the MemJournal.nChunkSize variable.
  38 */
  39 struct FileChunk {
  40   FileChunk *pNext;               /* Next chunk in the journal */
  41   u8 zChunk[8];                   /* Content of this chunk */
  42 };
  43
  44 /*
  45 ** By default, allocate this many bytes of memory for each FileChunk object.
  46 */
  47 #define MEMJOURNAL_DFLT_FILECHUNKSIZE 1024
  48
  49 /*
  50 ** For chunk size nChunkSize, return the number of bytes that should
  51 ** be allocated for each FileChunk structure.
  52 */
  53 #define fileChunkSize(nChunkSize) (sizeof(FileChunk) + ((nChunkSize)-8))
  54
  55 /*
  56 ** An instance of this object serves as a cursor into the rollback journal.
  57 ** The cursor can be either for reading or writing.
  58 */
  59 struct FilePoint {
  60   sqlite3_int64 iOffset;          /* Offset from the beginning of the file */
  61   FileChunk *pChunk;              /* Specific chunk into which cursor points */
  62 };
  63
  64 /*
  65 ** This structure is a subclass of sqlite3_file. Each open memory-journal
  66 ** is an instance of this class.
  67 */
  68 struct MemJournal {
  69   const sqlite3_io_methods *pMethod; /* Parent class. MUST BE FIRST */
  70   int nChunkSize;                 /* In-memory chunk-size */
  71
  72   int nSpill;                     /* Bytes of data before flushing */
  73   FileChunk *pFirst;              /* Head of in-memory chunk-list */
  74   FilePoint endpoint;             /* Pointer to the end of the file */
  75   FilePoint readpoint;            /* Pointer to the end of the last xRead() */
  76
  77   int flags;                      /* xOpen flags */
  78   sqlite3_vfs *pVfs;              /* The "real" underlying VFS */
  79   const char *zJournal;           /* Name of the journal file */
  80 };
  81
  82 /*
  83 ** Read data from the in-memory journal file.  This is the implementation
  84 ** of the sqlite3_vfs.xRead method.
  85 */
  86 static int memjrnlRead(
  87   sqlite3_file *pJfd,    /* The journal file from which to read */
  88   void *zBuf,            /* Put the results here */
  89   int iAmt,              /* Number of bytes to read */
  90   sqlite_int64 iOfst     /* Begin reading at this offset */
  91 ){
  92   MemJournal *p = (MemJournal *)pJfd;
  93   u8 *zOut = zBuf;
  94   int nRead = iAmt;
  95   int iChunkOffset;
  96   FileChunk *pChunk;
  97
  98   if( (iAmt+iOfst)>p->endpoint.iOffset ){
  99     return SQLITE_IOERR_SHORT_READ;
 100   }
 101   assert( p->readpoint.iOffset==0 || p->readpoint.pChunk!=0 );
 102   if( p->readpoint.iOffset!=iOfst || iOfst==0 ){
 103     sqlite3_int64 iOff = 0;
 104     for(pChunk=p->pFirst;
 105         ALWAYS(pChunk) && (iOff+p->nChunkSize)<=iOfst;
 106         pChunk=pChunk->pNext
 107     ){
 108       iOff += p->nChunkSize;
 109     }
 110   }else{
 111     pChunk = p->readpoint.pChunk;
 112     assert( pChunk!=0 );
 113   }
 114
 115   iChunkOffset = (int)(iOfst%p->nChunkSize);
 116   do {
 117     int iSpace = p->nChunkSize - iChunkOffset;
 118     int nCopy = MIN(nRead, (p->nChunkSize - iChunkOffset));
 119     memcpy(zOut, (u8*)pChunk->zChunk + iChunkOffset, nCopy);
 120     zOut += nCopy;
 121     nRead -= iSpace;
 122     iChunkOffset = 0;
 123   } while( nRead>=0 && (pChunk=pChunk->pNext)!=0 && nRead>0 );
 124   p->readpoint.iOffset = pChunk ? iOfst+iAmt : 0;
 125   p->readpoint.pChunk = pChunk;
 126
 127   return SQLITE_OK;
 128 }
 129
 130 /*
 131 ** Free the list of FileChunk structures headed at MemJournal.pFirst.
 132 */
 133 static void memjrnlFreeChunks(FileChunk *pFirst){
 134   FileChunk *pIter;
 135   FileChunk *pNext;
 136   for(pIter=pFirst; pIter; pIter=pNext){
 137     pNext = pIter->pNext;
 138     sqlite3_free(pIter);
 139   }
 140 }
 141
 142 /*
 143 ** Flush the contents of memory to a real file on disk.
 144 */
 145 static int memjrnlCreateFile(MemJournal *p){
 146   int rc;
 147   sqlite3_file *pReal = (sqlite3_file*)p;
 148   MemJournal copy = *p;
 149
 150   memset(p, 0, sizeof(MemJournal));
 151   rc = sqlite3OsOpen(copy.pVfs, copy.zJournal, pReal, copy.flags, 0);
 152   if( rc==SQLITE_OK ){
 153     int nChunk = copy.nChunkSize;
 154     i64 iOff = 0;
 155     FileChunk *pIter;
 156     for(pIter=copy.pFirst; pIter; pIter=pIter->pNext){
 157       if( iOff + nChunk > copy.endpoint.iOffset ){
 158         nChunk = copy.endpoint.iOffset - iOff;
 159       }
 160       rc = sqlite3OsWrite(pReal, (u8*)pIter->zChunk, nChunk, iOff);
 161       if( rc ) break;
 162       iOff += nChunk;
 163     }
 164     if( rc==SQLITE_OK ){
 165       /* No error has occurred. Free the in-memory buffers. */
 166       memjrnlFreeChunks(copy.pFirst);
 167     }
 168   }
 169   if( rc!=SQLITE_OK ){
 170     /* If an error occurred while creating or writing to the file, restore
 171     ** the original before returning. This way, SQLite uses the in-memory
 172     ** journal data to roll back changes made to the internal page-cache
 173     ** before this function was called.  */
 174     sqlite3OsClose(pReal);
 175     *p = copy;
 176   }
 177   return rc;
 178 }
 179
 180
 181 /*
 182 ** Write data to the file.
 183 */
 184 static int memjrnlWrite(
 185   sqlite3_file *pJfd,    /* The journal file into which to write */
 186   const void *zBuf,      /* Take data to be written from here */
 187   int iAmt,              /* Number of bytes to write */
 188   sqlite_int64 iOfst     /* Begin writing at this offset into the file */
 189 ){
 190   MemJournal *p = (MemJournal *)pJfd;
 191   int nWrite = iAmt;
 192   u8 *zWrite = (u8 *)zBuf;
 193
 194   /* If the file should be created now, create it and write the new data
 195   ** into the file on disk. */
 196   if( p->nSpill>0 && (iAmt+iOfst)>p->nSpill ){
 197     int rc = memjrnlCreateFile(p);
 198     if( rc==SQLITE_OK ){
 199       rc = sqlite3OsWrite(pJfd, zBuf, iAmt, iOfst);
 200     }
 201     return rc;
 202   }
 203
 204   /* If the contents of this write should be stored in memory */
 205   else{
 206     /* An in-memory journal file should only ever be appended to. Random
 207     ** access writes are not required. The only exception to this is when
 208     ** the in-memory journal is being used by a connection using the
 209     ** atomic-write optimization. In this case the first 28 bytes of the
 210     ** journal file may be written as part of committing the transaction. */
 211     assert( iOfst==p->endpoint.iOffset || iOfst==0 );
 212 #if defined(SQLITE_ENABLE_ATOMIC_WRITE) \
 213  || defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
 214     if( iOfst==0 && p->pFirst ){
 215       assert( p->nChunkSize>iAmt );
 216       memcpy((u8*)p->pFirst->zChunk, zBuf, iAmt);
 217     }else
 218 #else
 219     assert( iOfst>0 || p->pFirst==0 );
 220 #endif
 221     {
 222       while( nWrite>0 ){
 223         FileChunk *pChunk = p->endpoint.pChunk;
 224         int iChunkOffset = (int)(p->endpoint.iOffset%p->nChunkSize);
 225         int iSpace = MIN(nWrite, p->nChunkSize - iChunkOffset);
 226
 227         if( iChunkOffset==0 ){
 228           /* New chunk is required to extend the file. */
 229           FileChunk *pNew = sqlite3_malloc(fileChunkSize(p->nChunkSize));
 230           if( !pNew ){
 231             return SQLITE_IOERR_NOMEM_BKPT;
 232           }
 233           pNew->pNext = 0;
 234           if( pChunk ){
 235             assert( p->pFirst );
 236             pChunk->pNext = pNew;
 237           }else{
 238             assert( !p->pFirst );
 239             p->pFirst = pNew;
 240           }
 241           p->endpoint.pChunk = pNew;
 242         }
 243
 244         memcpy((u8*)p->endpoint.pChunk->zChunk + iChunkOffset, zWrite, iSpace);
 245         zWrite += iSpace;
 246         nWrite -= iSpace;
 247         p->endpoint.iOffset += iSpace;
 248       }
 249     }
 250   }
 251
 252   return SQLITE_OK;
 253 }
 254
 255 /*
 256 ** Truncate the in-memory file.
 257 */
 258 static int memjrnlTruncate(sqlite3_file *pJfd, sqlite_int64 size){
 259   MemJournal *p = (MemJournal *)pJfd;
 260   assert( p->endpoint.pChunk==0 || p->endpoint.pChunk->pNext==0 );
 261   if( size<p->endpoint.iOffset ){
 262     FileChunk *pIter = 0;
 263     if( size==0 ){
 264       memjrnlFreeChunks(p->pFirst);
 265       p->pFirst = 0;
 266     }else{
 267       i64 iOff = p->nChunkSize;
 268       for(pIter=p->pFirst; ALWAYS(pIter) && iOff<size; pIter=pIter->pNext){
 269         iOff += p->nChunkSize;
 270       }
 271       if( ALWAYS(pIter) ){
 272         memjrnlFreeChunks(pIter->pNext);
 273         pIter->pNext = 0;
 274       }
 275     }
 276
 277     p->endpoint.pChunk = pIter;
 278     p->endpoint.iOffset = size;
 279     p->readpoint.pChunk = 0;
 280     p->readpoint.iOffset = 0;
 281   }
 282   return SQLITE_OK;
 283 }
 284
 285 /*
 286 ** Close the file.
 287 */
 288 static int memjrnlClose(sqlite3_file *pJfd){
 289   MemJournal *p = (MemJournal *)pJfd;
 290   memjrnlFreeChunks(p->pFirst);
 291   return SQLITE_OK;
 292 }
 293
 294 /*
 295 ** Sync the file.
 296 **
 297 ** If the real file has been created, call its xSync method. Otherwise,
 298 ** syncing an in-memory journal is a no-op.
 299 */
 300 static int memjrnlSync(sqlite3_file *pJfd, int flags){
 301   UNUSED_PARAMETER2(pJfd, flags);
 302   return SQLITE_OK;
 303 }
 304
 305 /*
 306 ** Query the size of the file in bytes.
 307 */
 308 static int memjrnlFileSize(sqlite3_file *pJfd, sqlite_int64 *pSize){
 309   MemJournal *p = (MemJournal *)pJfd;
 310   *pSize = (sqlite_int64) p->endpoint.iOffset;
 311   return SQLITE_OK;
 312 }
 313
 314 /*
 315 ** Table of methods for MemJournal sqlite3_file object.
 316 */
 317 static const struct sqlite3_io_methods MemJournalMethods = {
 318   1,                /* iVersion */
 319   memjrnlClose,     /* xClose */
 320   memjrnlRead,      /* xRead */
 321   memjrnlWrite,     /* xWrite */
 322   memjrnlTruncate,  /* xTruncate */
 323   memjrnlSync,      /* xSync */
 324   memjrnlFileSize,  /* xFileSize */
 325   0,                /* xLock */
 326   0,                /* xUnlock */
 327   0,                /* xCheckReservedLock */
 328   0,                /* xFileControl */
 329   0,                /* xSectorSize */
 330   0,                /* xDeviceCharacteristics */
 331   0,                /* xShmMap */
 332   0,                /* xShmLock */
 333   0,                /* xShmBarrier */
 334   0,                /* xShmUnmap */
 335   0,                /* xFetch */
 336   0                 /* xUnfetch */
 337 };
 338
 339 /*
 340 ** Open a journal file.
 341 **
 342 ** The behaviour of the journal file depends on the value of parameter
 343 ** nSpill. If nSpill is 0, then the journal file is always create and
 344 ** accessed using the underlying VFS. If nSpill is less than zero, then
 345 ** all content is always stored in main-memory. Finally, if nSpill is a
 346 ** positive value, then the journal file is initially created in-memory
 347 ** but may be flushed to disk later on. In this case the journal file is
 348 ** flushed to disk either when it grows larger than nSpill bytes in size,
 349 ** or when sqlite3JournalCreate() is called.
 350 */
 351 int sqlite3JournalOpen(
 352   sqlite3_vfs *pVfs,         /* The VFS to use for actual file I/O */
 353   const char *zName,         /* Name of the journal file */
 354   sqlite3_file *pJfd,        /* Preallocated, blank file handle */
 355   int flags,                 /* Opening flags */
 356   int nSpill                 /* Bytes buffered before opening the file */
 357 ){
 358   MemJournal *p = (MemJournal*)pJfd;
 359
 360   /* Zero the file-handle object. If nSpill was passed zero, initialize
 361   ** it using the sqlite3OsOpen() function of the underlying VFS. In this
 362   ** case none of the code in this module is executed as a result of calls
 363   ** made on the journal file-handle.  */
 364   memset(p, 0, sizeof(MemJournal));
 365   if( nSpill==0 ){
 366     return sqlite3OsOpen(pVfs, zName, pJfd, flags, 0);
 367   }
 368
 369   if( nSpill>0 ){
 370     p->nChunkSize = nSpill;
 371   }else{
 372     p->nChunkSize = 8 + MEMJOURNAL_DFLT_FILECHUNKSIZE - sizeof(FileChunk);
 373     assert( MEMJOURNAL_DFLT_FILECHUNKSIZE==fileChunkSize(p->nChunkSize) );
 374   }
 375
 376   pJfd->pMethods = (const sqlite3_io_methods*)&MemJournalMethods;
 377   p->nSpill = nSpill;
 378   p->flags = flags;
 379   p->zJournal = zName;
 380   p->pVfs = pVfs;
 381   return SQLITE_OK;
 382 }
 383
 384 /*
 385 ** Open an in-memory journal file.
 386 */
 387 void sqlite3MemJournalOpen(sqlite3_file *pJfd){
 388   sqlite3JournalOpen(0, 0, pJfd, 0, -1);
 389 }
 390
 391 #if defined(SQLITE_ENABLE_ATOMIC_WRITE) \
 392  || defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
 393 /*
 394 ** If the argument p points to a MemJournal structure that is not an
 395 ** in-memory-only journal file (i.e. is one that was opened with a +ve
 396 ** nSpill parameter or as SQLITE_OPEN_MAIN_JOURNAL), and the underlying
 397 ** file has not yet been created, create it now.
 398 */
 399 int sqlite3JournalCreate(sqlite3_file *pJfd){
 400   int rc = SQLITE_OK;
 401   MemJournal *p = (MemJournal*)pJfd;
 402   if( pJfd->pMethods==&MemJournalMethods && (
 403 #ifdef SQLITE_ENABLE_ATOMIC_WRITE
 404      p->nSpill>0
 405 #else
 406      /* While this appears to not be possible without ATOMIC_WRITE, the
 407      ** paths are complex, so it seems prudent to leave the test in as
 408      ** a NEVER(), in case our analysis is subtly flawed. */
 409      NEVER(p->nSpill>0)
 410 #endif
 411 #ifdef SQLITE_ENABLE_BATCH_ATOMIC_WRITE
 412      || (p->flags & SQLITE_OPEN_MAIN_JOURNAL)
 413 #endif
 414   )){
 415     rc = memjrnlCreateFile(p);
 416   }
 417   return rc;
 418 }
 419 #endif
 420
 421 /*
 422 ** The file-handle passed as the only argument is open on a journal file.
 423 ** Return true if this "journal file" is currently stored in heap memory,
 424 ** or false otherwise.
 425 */
 426 int sqlite3JournalIsInMemory(sqlite3_file *p){
 427   return p->pMethods==&MemJournalMethods;
 428 }
 429
 430 /*
 431 ** Return the number of bytes required to store a JournalFile that uses vfs
 432 ** pVfs to create the underlying on-disk files.
 433 */
 434 int sqlite3JournalSize(sqlite3_vfs *pVfs){
 435   return MAX(pVfs->szOsFile, (int)sizeof(MemJournal));
 436 }