/*

** 2001 September 15

**

** 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 is the implementation of the page cache subsystem or "pager".

** 

** 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 1.469 2008/08/02 03:50:39 drh Exp $

*/

#ifndef SQLITE_OMIT_DISKIO

#include "sqliteInt.h"

#include <assert.h>

#include <string.h>

/*

** Macros for troubleshooting.  Normally turned off

*/

#if 0

#define sqlite3DebugPrintf printf

#define PAGERTRACE1(X)       sqlite3DebugPrintf(X)

#define PAGERTRACE2(X,Y)     sqlite3DebugPrintf(X,Y)

#define PAGERTRACE3(X,Y,Z)   sqlite3DebugPrintf(X,Y,Z)

#define PAGERTRACE4(X,Y,Z,W) sqlite3DebugPrintf(X,Y,Z,W)

#define PAGERTRACE5(X,Y,Z,W,V) sqlite3DebugPrintf(X,Y,Z,W,V)

#else

#define PAGERTRACE1(X)

#define PAGERTRACE2(X,Y)

#define PAGERTRACE3(X,Y,Z)

#define PAGERTRACE4(X,Y,Z,W)

#define PAGERTRACE5(X,Y,Z,W,V)

#endif

/*

** The following two macros are used within the PAGERTRACEX() macros above

** to print out file-descriptors. 

**

** PAGERID() takes a pointer to a Pager struct as its argument. The

** associated file-descriptor is returned. FILEHANDLEID() takes an sqlite3_file

** struct as its argument.

*/

#define PAGERID(p) ((int)(p->fd))

#define FILEHANDLEID(fd) ((int)fd)

/*

** The page cache as a whole is always in one of the following

** states:

**

**   PAGER_UNLOCK        The page cache is not currently reading or 

**                       writing the database file.  There is no

**                       data held in memory.  This is the initial

**                       state.

**

**   PAGER_SHARED        The page cache is reading the database.

**                       Writing is not permitted.  There can be

**                       multiple readers accessing the same database

**                       file at the same time.

**

**   PAGER_RESERVED      This process has reserved the database for writing

**                       but has not yet made any changes.  Only one process

**                       at a time can reserve the database.  The original

**                       database file has not been modified so other

**                       processes may still be reading the on-disk

**                       database file.

**

**   PAGER_EXCLUSIVE     The page cache is writing the database.

**                       Access is exclusive.  No other processes or

**                       threads can be reading or writing while one

**                       process is writing.

**

**   PAGER_SYNCED        The pager moves to this state from PAGER_EXCLUSIVE

**                       after all dirty pages have been written to the

**                       database file and the file has been synced to

**                       disk. All that remains to do is to remove or

**                       truncate the journal file and the transaction 

**                       will be committed.

**

** The page cache comes up in PAGER_UNLOCK.  The first time a

** sqlite3PagerGet() occurs, the state transitions to PAGER_SHARED.

** After all pages have been released using sqlite_page_unref(),

** the state transitions back to PAGER_UNLOCK.  The first time

** that sqlite3PagerWrite() is called, the state transitions to

** PAGER_RESERVED.  (Note that sqlite3PagerWrite() can only be

** called on an outstanding page which means that the pager must

** be in PAGER_SHARED before it transitions to PAGER_RESERVED.)

** PAGER_RESERVED means that there is an open rollback journal.

** The transition to PAGER_EXCLUSIVE occurs before any changes

** are made to the database file, though writes to the rollback

** journal occurs with just PAGER_RESERVED.  After an sqlite3PagerRollback()

** or sqlite3PagerCommitPhaseTwo(), the state can go back to PAGER_SHARED,

** or it can stay at PAGER_EXCLUSIVE if we are in exclusive access mode.

*/

#define PAGER_UNLOCK      0

#define PAGER_SHARED      1   /* same as SHARED_LOCK */

#define PAGER_RESERVED    2   /* same as RESERVED_LOCK */

#define PAGER_EXCLUSIVE   4   /* same as EXCLUSIVE_LOCK */

#define PAGER_SYNCED      5

/*

** If the SQLITE_BUSY_RESERVED_LOCK macro is set to true at compile-time,

** then failed attempts to get a reserved lock will invoke the busy callback.

** This is off by default.  To see why, consider the following scenario:

** 

** Suppose thread A already has a shared lock and wants a reserved lock.

** Thread B already has a reserved lock and wants an exclusive lock.  If

** both threads are using their busy callbacks, it might be a long time

** be for one of the threads give up and allows the other to proceed.

** But if the thread trying to get the reserved lock gives up quickly

** (if it never invokes its busy callback) then the contention will be

** resolved quickly.

*/

#ifndef SQLITE_BUSY_RESERVED_LOCK

# define SQLITE_BUSY_RESERVED_LOCK 0

#endif

/*

** This macro rounds values up so that if the value is an address it

** is guaranteed to be an address that is aligned to an 8-byte boundary.

*/

#define FORCE_ALIGNMENT(X)   (((X)+7)&~7)

typedef struct PgHdr PgHdr;

/*

** Each pager stores all currently unreferenced pages in a list sorted

** in least-recently-used (LRU) order (i.e. the first item on the list has 

** not been referenced in a long time, the last item has been recently

** used). An instance of this structure is included as part of each

** pager structure for this purpose (variable Pager.lru).

**

** Additionally, if memory-management is enabled, all unreferenced pages 

** are stored in a global LRU list (global variable sqlite3LruPageList).

**

** In both cases, the PagerLruList.pFirstSynced variable points to

** the first page in the corresponding list that does not require an

** fsync() operation before its memory can be reclaimed. If no such

** page exists, PagerLruList.pFirstSynced is set to NULL.

*/

typedef struct PagerLruList PagerLruList;

struct PagerLruList {

  PgHdr *pFirst;         /* First page in LRU list */

  PgHdr *pLast;          /* Last page in LRU list (the most recently used) */

  PgHdr *pFirstSynced;   /* First page in list with PgHdr.needSync==0 */

};

/*

** The following structure contains the next and previous pointers used

** to link a PgHdr structure into a PagerLruList linked list. 

*/

typedef struct PagerLruLink PagerLruLink;

struct PagerLruLink {

  PgHdr *pNext;

  PgHdr *pPrev;

};

/*

** Each in-memory image of a page begins with the following header.

** This header is only visible to this pager module.  The client

** code that calls pager sees only the data that follows the header.

**

** Client code should call sqlite3PagerWrite() on a page prior to making

** any modifications to that page.  The first time sqlite3PagerWrite()

** is called, the original page contents are written into the rollback

** journal and PgHdr.inJournal and PgHdr.needSync are set.  Later, once

** the journal page has made it onto the disk surface, PgHdr.needSync

** is cleared.  The modified page cannot be written back into the original

** database file until the journal pages has been synced to disk and the

** PgHdr.needSync has been cleared.

**

** The PgHdr.dirty flag is set when sqlite3PagerWrite() is called and

** is cleared again when the page content is written back to the original

** database file.

**

** Details of important structure elements:

**

** needSync

**

**     If this is true, this means that it is not safe to write the page

**     content to the database because the original content needed

**     for rollback has not by synced to the main rollback journal.

**     The original content may have been written to the rollback journal

**     but it has not yet been synced.  So we cannot write to the database

**     file because power failure might cause the page in the journal file

**     to never reach the disk.  It is as if the write to the journal file

**     does not occur until the journal file is synced.

**     

**     This flag is false if the page content exactly matches what

**     currently exists in the database file.  The needSync flag is also

**     false if the original content has been written to the main rollback

**     journal and synced.  If the page represents a new page that has

**     been added onto the end of the database during the current

**     transaction, the needSync flag is true until the original database

**     size in the journal header has been synced to disk.

**

** inJournal

**

**     This is true if the original page has been written into the main

**     rollback journal.  This is always false for new pages added to

**     the end of the database file during the current transaction.

**     And this flag says nothing about whether or not the journal

**     has been synced to disk.  For pages that are in the original

**     database file, the following expression should always be true:

**

**       inJournal = sqlite3BitvecTest(pPager->pInJournal, pgno)

**

**     The pPager->pInJournal object is only valid for the original

**     pages of the database, not new pages that are added to the end

**     of the database, so obviously the above expression cannot be

**     valid for new pages.  For new pages inJournal is always 0.

**

** dirty

**

**     When true, this means that the content of the page has been

**     modified and needs to be written back to the database file.

**     If false, it means that either the content of the page is

**     unchanged or else the content is unimportant and we do not

**     care whether or not it is preserved.

**

** alwaysRollback

**

**     This means that the sqlite3PagerDontRollback() API should be

**     ignored for this page.  The DontRollback() API attempts to say

**     that the content of the page on disk is unimportant (it is an

**     unused page on the freelist) so that it is unnecessary to 

**     rollback changes to this page because the content of the page

**     can change without changing the meaning of the database.  This

**     flag overrides any DontRollback() attempt.  This flag is set

**     when a page that originally contained valid data is added to

**     the freelist.  Later in the same transaction, this page might

**     be pulled from the freelist and reused for something different

**     and at that point the DontRollback() API will be called because

**     pages taken from the freelist do not need to be protected by

**     the rollback journal.  But this flag says that the page was

**     not originally part of the freelist so that it still needs to

**     be rolled back in spite of any subsequent DontRollback() calls.

**

** needRead 

**

**     This flag means (when true) that the content of the page has

**     not yet been loaded from disk.  The in-memory content is just

**     garbage.  (Actually, we zero the content, but you should not

**     make any assumptions about the content nevertheless.)  If the

**     content is needed in the future, it should be read from the

**     original database file.

*/

struct PgHdr {

  Pager *pPager;                 /* The pager to which this page belongs */

  Pgno pgno;                     /* The page number for this page */

  PgHdr *pNextHash, *pPrevHash;  /* Hash collision chain for PgHdr.pgno */

  PagerLruLink free;             /* Next and previous free pages */

  PgHdr *pNextAll;               /* A list of all pages */

  u8 inJournal;                  /* TRUE if has been written to journal */

  u8 dirty;                      /* TRUE if we need to write back changes */

  u8 needSync;                   /* Sync journal before writing this page */

  u8 alwaysRollback;             /* Disable DontRollback() for this page */

  u8 needRead;                   /* Read content if PagerWrite() is called */

  short int nRef;                /* Number of users of this page */

  PgHdr *pDirty, *pPrevDirty;    /* Dirty pages */

#ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT

  PgHdr *pPrevAll;               /* A list of all pages */

  PagerLruLink gfree;            /* Global list of nRef==0 pages */

#endif

#ifdef SQLITE_CHECK_PAGES

  u32 pageHash;

#endif

  void *pData;                   /* Page data */

  /* Pager.nExtra bytes of local data appended to this header */

};

/*

** For an in-memory only database, some extra information is recorded about

** each page so that changes can be rolled back.  (Journal files are not

** used for in-memory databases.)  The following information is added to

** the end of every EXTRA block for in-memory databases.

**

** This information could have been added directly to the PgHdr structure.

** But then it would take up an extra 8 bytes of storage on every PgHdr

** even for disk-based databases.  Splitting it out saves 8 bytes.  This

** is only a savings of 0.8% but those percentages add up.

*/

typedef struct PgHistory PgHistory;

struct PgHistory {

  u8 *pOrig;     /* Original page text.  Restore to this on a full rollback */

  u8 *pStmt;     /* Text as it was at the beginning of the current statement */

  PgHdr *pNextStmt, *pPrevStmt;  /* List of pages in the statement journal */

  u8 inStmt;                     /* TRUE if in the statement subjournal */

};

/*

** A macro used for invoking the codec if there is one

*/

#ifdef SQLITE_HAS_CODEC

# define CODEC1(P,D,N,X) if( P->xCodec!=0 ){ P->xCodec(P->pCodecArg,D,N,X); }

# define CODEC2(P,D,N,X) ((char*)(P->xCodec!=0?P->xCodec(P->pCodecArg,D,N,X):D))

#else

# define CODEC1(P,D,N,X) /* NO-OP */

# define CODEC2(P,D,N,X) ((char*)D)

#endif

/*

** Convert a pointer to a PgHdr into a pointer to its data

** and back again.

*/

#define PGHDR_TO_DATA(P)    ((P)->pData)

#define PGHDR_TO_EXTRA(G,P) ((void*)&((G)[1]))

#define PGHDR_TO_HIST(P,PGR)  \

            ((PgHistory*)&((char*)(&(P)[1]))[(PGR)->nExtra])

/*

** A open page cache is an instance of the following structure.

**

** Pager.errCode may be set to SQLITE_IOERR, SQLITE_CORRUPT, or

** or SQLITE_FULL. Once one of the first three errors occurs, it persists

** and is returned as the result of every major pager API call.  The

** SQLITE_FULL return code is slightly different. It persists only until the

** next successful rollback is performed on the pager cache. Also,

** SQLITE_FULL does not affect the sqlite3PagerGet() and sqlite3PagerLookup()

** APIs, they may still be used successfully.

*/

struct Pager {

  sqlite3_vfs *pVfs;          /* OS functions to use for IO */

  u8 journalOpen;             /* True if journal file descriptors is valid */

  u8 journalStarted;          /* True if header of journal is synced */

  u8 useJournal;              /* Use a rollback journal on this file */

  u8 noReadlock;              /* Do not bother to obtain readlocks */

  u8 stmtOpen;                /* True if the statement subjournal is open */

  u8 stmtInUse;               /* True we are in a statement subtransaction */

  u8 stmtAutoopen;            /* Open stmt journal when main journal is opened*/

  u8 noSync;                  /* Do not sync the journal if true */

  u8 fullSync;                /* Do extra syncs of the journal for robustness */

  u8 sync_flags;              /* One of SYNC_NORMAL or SYNC_FULL */

  u8 state;                   /* PAGER_UNLOCK, _SHARED, _RESERVED, etc. */

  u8 tempFile;                /* zFilename is a temporary file */

  u8 readOnly;                /* True for a read-only database */

  u8 needSync;                /* True if an fsync() is needed on the journal */

  u8 dirtyCache;              /* True if cached pages have changed */

  u8 alwaysRollback;          /* Disable DontRollback() for all pages */

  u8 memDb;                   /* True to inhibit all file I/O */

  u8 setMaster;               /* True if a m-j name has been written to jrnl */

  u8 doNotSync;               /* Boolean. While true, do not spill the cache */

  u8 exclusiveMode;           /* Boolean. True if locking_mode==EXCLUSIVE */

  u8 journalMode;             /* On of the PAGER_JOURNALMODE_* values */

  u8 dbModified;              /* True if there are any changes to the Db */

  u8 changeCountDone;         /* Set after incrementing the change-counter */

  u32 vfsFlags;               /* Flags for sqlite3_vfs.xOpen() */

  int errCode;                /* One of several kinds of errors */

  int dbSize;                 /* Number of pages in the file */

  int origDbSize;             /* dbSize before the current change */

  int stmtSize;               /* Size of database (in pages) at stmt_begin() */

  int nRec;                   /* Number of pages written to the journal */

  u32 cksumInit;              /* Quasi-random value added to every checksum */

  int stmtNRec;               /* Number of records in stmt subjournal */

  int nExtra;                 /* Add this many bytes to each in-memory page */

  int pageSize;               /* Number of bytes in a page */

  int nPage;                  /* Total number of in-memory pages */

  int nRef;                   /* Number of in-memory pages with PgHdr.nRef>0 */

  int mxPage;                 /* Maximum number of pages to hold in cache */

  Pgno mxPgno;                /* Maximum allowed size of the database */

  Bitvec *pInJournal;         /* One bit for each page in the database file */

  Bitvec *pInStmt;            /* One bit for each page in the database */

  char *zFilename;            /* Name of the database file */

  char *zJournal;             /* Name of the journal file */

  char *zDirectory;           /* Directory hold database and journal files */

  sqlite3_file *fd, *jfd;     /* File descriptors for database and journal */

  sqlite3_file *stfd;         /* File descriptor for the statement subjournal*/

  BusyHandler *pBusyHandler;  /* Pointer to sqlite.busyHandler */

  PagerLruList lru;           /* LRU list of free pages */

  PgHdr *pAll;                /* List of all pages */

  PgHdr *pStmt;               /* List of pages in the statement subjournal */

  PgHdr *pDirty;              /* List of all dirty pages */

  i64 journalOff;             /* Current byte offset in the journal file */

  i64 journalHdr;             /* Byte offset to previous journal header */

  i64 stmtHdrOff;             /* First journal header written this statement */

  i64 stmtCksum;              /* cksumInit when statement was started */

  i64 stmtJSize;              /* Size of journal at stmt_begin() */

  int sectorSize;             /* Assumed sector size during rollback */

#ifdef SQLITE_TEST

  int nHit, nMiss;            /* Cache hits and missing */

  int nRead, nWrite;          /* Database pages read/written */

#endif

  void (*xDestructor)(DbPage*,int); /* Call this routine when freeing pages */

  void (*xReiniter)(DbPage*,int);   /* Call this routine when reloading pages */

#ifdef SQLITE_HAS_CODEC

  void *(*xCodec)(void*,void*,Pgno,int); /* Routine for en/decoding data */

  void *pCodecArg;            /* First argument to xCodec() */

#endif

  int nHash;                  /* Size of the pager hash table */

  PgHdr **aHash;              /* Hash table to map page number to PgHdr */

#ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT

  Pager *pNext;               /* Doubly linked list of pagers on which */

  Pager *pPrev;               /* sqlite3_release_memory() will work */

  volatile int iInUseMM;      /* Non-zero if unavailable to MM */

  volatile int iInUseDB;      /* Non-zero if in sqlite3_release_memory() */

#endif

  char *pTmpSpace;            /* Pager.pageSize bytes of space for tmp use */

  char dbFileVers[16];        /* Changes whenever database file changes */

  i64 journalSizeLimit;       /* Size limit for persistent journal files */

};

/*

** The following global variables hold counters used for

** testing purposes only.  These variables do not exist in

** a non-testing build.  These variables are not thread-safe.

*/

#ifdef SQLITE_TEST

int sqlite3_pager_readdb_count = 0;    /* Number of full pages read from DB */

int sqlite3_pager_writedb_count = 0;   /* Number of full pages written to DB */

int sqlite3_pager_writej_count = 0;    /* Number of pages written to journal */

int sqlite3_pager_pgfree_count = 0;    /* Number of cache pages freed */

# define PAGER_INCR(v)  v++

#else

# define PAGER_INCR(v)

#endif

/*

** The following variable points to the head of a double-linked list

** of all pagers that are eligible for page stealing by the

** sqlite3_release_memory() interface.  Access to this list is

** protected by the SQLITE_MUTEX_STATIC_MEM2 mutex.

*/

#ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT

static Pager *sqlite3PagerList = 0;

static PagerLruList sqlite3LruPageList = {0, 0, 0};

#endif

/*

** Journal files begin with the following magic string.  The data

** was obtained from /dev/random.  It is used only as a sanity check.

**

** Since version 2.8.0, the journal format contains additional sanity

** checking information.  If the power fails while the journal is begin

** written, semi-random garbage data might appear in the journal

** file after power is restored.  If an attempt is then made

** to roll the journal back, the database could be corrupted.  The additional

** sanity checking data is an attempt to discover the garbage in the

** journal and ignore it.

**

** The sanity checking information for the new journal format consists

** of a 32-bit checksum on each page of data.  The checksum covers both

** the page number and the pPager->pageSize bytes of data for the page.

** This cksum is initialized to a 32-bit random value that appears in the

** journal file right after the header.  The random initializer is important,

** because garbage data that appears at the end of a journal is likely

** data that was once in other files that have now been deleted.  If the

** garbage data came from an obsolete journal file, the checksums might

** be correct.  But by initializing the checksum to random value which

** is different for every journal, we minimize that risk.

*/

static const unsigned char aJournalMagic[] = {

  0xd9, 0xd5, 0x05, 0xf9, 0x20, 0xa1, 0x63, 0xd7,

};

/*

** The size of the header and of each page in the journal is determined

** by the following macros.

*/

#define JOURNAL_PG_SZ(pPager)  ((pPager->pageSize) + 8)

/*

** The journal header size for this pager. In the future, this could be

** set to some value read from the disk controller. The important

** characteristic is that it is the same size as a disk sector.

*/

#define JOURNAL_HDR_SZ(pPager) (pPager->sectorSize)

/*

** The macro MEMDB is true if we are dealing with an in-memory database.

** We do this as a macro so that if the SQLITE_OMIT_MEMORYDB macro is set,

** the value of MEMDB will be a constant and the compiler will optimize

** out code that would never execute.

*/

#ifdef SQLITE_OMIT_MEMORYDB

# define MEMDB 0

#else

# define MEMDB pPager->memDb

#endif

/*

** Page number PAGER_MJ_PGNO is never used in an SQLite database (it is

** reserved for working around a windows/posix incompatibility). It is

** used in the journal to signify that the remainder of the journal file 

** is devoted to storing a master journal name - there are no more pages to

** roll back. See comments for function writeMasterJournal() for details.

*/

/* #define PAGER_MJ_PGNO(x) (PENDING_BYTE/((x)->pageSize)) */

#define PAGER_MJ_PGNO(x) ((PENDING_BYTE/((x)->pageSize))+1)

/*

** The maximum legal page number is (2^31 - 1).

*/

#define PAGER_MAX_PGNO 2147483647

/*

** The pagerEnter() and pagerLeave() routines acquire and release

** a mutex on each pager.  The mutex is recursive.

**

** This is a special-purpose mutex.  It only provides mutual exclusion

** between the Btree and the Memory Management sqlite3_release_memory()

** function.  It does not prevent, for example, two Btrees from accessing

** the same pager at the same time.  Other general-purpose mutexes in

** the btree layer handle that chore.

*/

#ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT

  static void pagerEnter(Pager *p){

    p->iInUseDB++;

    if( p->iInUseMM && p->iInUseDB==1 ){

#ifndef SQLITE_MUTEX_NOOP

      sqlite3_mutex *mutex;

      mutex = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MEM2);

#endif

      p->iInUseDB = 0;

      sqlite3_mutex_enter(mutex);

      p->iInUseDB = 1;

      sqlite3_mutex_leave(mutex);

    }