/*

** 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 is specific to Unix systems.

*/

#include "sqliteInt.h"

#include "os.h"

#if OS_UNIX              /* This file is used on unix only */

/*

** These #defines should enable >2GB file support on Posix if the

** underlying operating system supports it.  If the OS lacks

** large file support, these should be no-ops.

**

** Large file support can be disabled using the -DSQLITE_DISABLE_LFS switch

** on the compiler command line.  This is necessary if you are compiling

** on a recent machine (ex: RedHat 7.2) but you want your code to work

** on an older machine (ex: RedHat 6.0).  If you compile on RedHat 7.2

** without this option, LFS is enable.  But LFS does not exist in the kernel

** in RedHat 6.0, so the code won't work.  Hence, for maximum binary

** portability you should omit LFS.

*/

#ifndef SQLITE_DISABLE_LFS

# define _LARGE_FILE       1

# ifndef _FILE_OFFSET_BITS

#   define _FILE_OFFSET_BITS 64

# endif

# define _LARGEFILE_SOURCE 1

#endif

/*

** standard include files.

*/

#include <sys/types.h>

#include <sys/stat.h>

#include <fcntl.h>

#include <unistd.h>

#include <time.h>

#include <sys/time.h>

#include <errno.h>

/*

** If we are to be thread-safe, include the pthreads header and define

** the SQLITE_UNIX_THREADS macro.

*/

#if defined(THREADSAFE) && THREADSAFE

# include <pthread.h>

# define SQLITE_UNIX_THREADS 1

#endif

/*

** Default permissions when creating a new file

*/

#ifndef SQLITE_DEFAULT_FILE_PERMISSIONS

# define SQLITE_DEFAULT_FILE_PERMISSIONS 0644

#endif

/*

** The unixFile structure is subclass of OsFile specific for the unix

** protability layer.

*/

typedef struct unixFile unixFile;

struct unixFile {

  IoMethod const *pMethod;  /* Always the first entry */

  struct openCnt *pOpen;    /* Info about all open fd's on this inode */

  struct lockInfo *pLock;   /* Info about locks on this inode */

  int h;                    /* The file descriptor */

  unsigned char locktype;   /* The type of lock held on this fd */

  unsigned char isOpen;     /* True if needs to be closed */

  unsigned char fullSync;   /* Use F_FULLSYNC if available */

  int dirfd;                /* File descriptor for the directory */

  i64 offset;               /* Seek offset */

#ifdef SQLITE_UNIX_THREADS

  pthread_t tid;            /* The thread that "owns" this OsFile */

#endif

};

/*

** Provide the ability to override some OS-layer functions during

** testing.  This is used to simulate OS crashes to verify that 

** commits are atomic even in the event of an OS crash.

*/

#ifdef SQLITE_CRASH_TEST

  extern int sqlite3CrashTestEnable;

  extern int sqlite3CrashOpenReadWrite(const char*, OsFile**, int*);

  extern int sqlite3CrashOpenExclusive(const char*, OsFile**, int);

  extern int sqlite3CrashOpenReadOnly(const char*, OsFile**, int);

# define CRASH_TEST_OVERRIDE(X,A,B,C) \

    if(sqlite3CrashTestEnable){ return X(A,B,C); }

#else

# define CRASH_TEST_OVERRIDE(X,A,B,C)  /* no-op */

#endif

/*

** Include code that is common to all os_*.c files

*/

#include "os_common.h"

/*

** Do not include any of the File I/O interface procedures if the

** SQLITE_OMIT_DISKIO macro is defined (indicating that the database

** will be in-memory only)

*/

#ifndef SQLITE_OMIT_DISKIO

/*

** Define various macros that are missing from some systems.

*/

#ifndef O_LARGEFILE

# define O_LARGEFILE 0

#endif

#ifdef SQLITE_DISABLE_LFS

# undef O_LARGEFILE

# define O_LARGEFILE 0

#endif

#ifndef O_NOFOLLOW

# define O_NOFOLLOW 0

#endif

#ifndef O_BINARY

# define O_BINARY 0

#endif

/*

** The DJGPP compiler environment looks mostly like Unix, but it

** lacks the fcntl() system call.  So redefine fcntl() to be something

** that always succeeds.  This means that locking does not occur under

** DJGPP.  But it's DOS - what did you expect?

*/

#ifdef __DJGPP__

# define fcntl(A,B,C) 0

#endif

/*

** The threadid macro resolves to the thread-id or to 0.  Used for

** testing and debugging only.

*/

#ifdef SQLITE_UNIX_THREADS

#define threadid pthread_self()

#else

#define threadid 0

#endif

/*

** Set or check the OsFile.tid field.  This field is set when an OsFile

** is first opened.  All subsequent uses of the OsFile verify that the

** same thread is operating on the OsFile.  Some operating systems do

** not allow locks to be overridden by other threads and that restriction

** means that sqlite3* database handles cannot be moved from one thread

** to another.  This logic makes sure a user does not try to do that

** by mistake.

**

** Version 3.3.1 (2006-01-15):  OsFiles can be moved from one thread to

** another as long as we are running on a system that supports threads

** overriding each others locks (which now the most common behavior)

** or if no locks are held.  But the OsFile.pLock field needs to be

** recomputed because its key includes the thread-id.  See the 

** transferOwnership() function below for additional information

*/

#if defined(SQLITE_UNIX_THREADS)

# define SET_THREADID(X)   (X)->tid = pthread_self()

# define CHECK_THREADID(X) (threadsOverrideEachOthersLocks==0 && \

                            !pthread_equal((X)->tid, pthread_self()))

#else

# define SET_THREADID(X)

# define CHECK_THREADID(X) 0

#endif

/*

** Here is the dirt on POSIX advisory locks:  ANSI STD 1003.1 (1996)

** section 6.5.2.2 lines 483 through 490 specify that when a process

** sets or clears a lock, that operation overrides any prior locks set

** by the same process.  It does not explicitly say so, but this implies

** that it overrides locks set by the same process using a different

** file descriptor.  Consider this test case:

**

**       int fd1 = open("./file1", O_RDWR|O_CREAT, 0644);

**       int fd2 = open("./file2", O_RDWR|O_CREAT, 0644);

**

** Suppose ./file1 and ./file2 are really the same file (because

** one is a hard or symbolic link to the other) then if you set

** an exclusive lock on fd1, then try to get an exclusive lock

** on fd2, it works.  I would have expected the second lock to

** fail since there was already a lock on the file due to fd1.

** But not so.  Since both locks came from the same process, the

** second overrides the first, even though they were on different

** file descriptors opened on different file names.

**

** Bummer.  If you ask me, this is broken.  Badly broken.  It means

** that we cannot use POSIX locks to synchronize file access among

** competing threads of the same process.  POSIX locks will work fine

** to synchronize access for threads in separate processes, but not

** threads within the same process.

**

** To work around the problem, SQLite has to manage file locks internally

** on its own.  Whenever a new database is opened, we have to find the

** specific inode of the database file (the inode is determined by the

** st_dev and st_ino fields of the stat structure that fstat() fills in)

** and check for locks already existing on that inode.  When locks are

** created or removed, we have to look at our own internal record of the

** locks to see if another thread has previously set a lock on that same

** inode.

**

** The OsFile structure for POSIX is no longer just an integer file

** descriptor.  It is now a structure that holds the integer file

** descriptor and a pointer to a structure that describes the internal

** locks on the corresponding inode.  There is one locking structure

** per inode, so if the same inode is opened twice, both OsFile structures

** point to the same locking structure.  The locking structure keeps

** a reference count (so we will know when to delete it) and a "cnt"

** field that tells us its internal lock status.  cnt==0 means the

** file is unlocked.  cnt==-1 means the file has an exclusive lock.

** cnt>0 means there are cnt shared locks on the file.

**

** Any attempt to lock or unlock a file first checks the locking

** structure.  The fcntl() system call is only invoked to set a 

** POSIX lock if the internal lock structure transitions between

** a locked and an unlocked state.

**

** 2004-Jan-11:

** More recent discoveries about POSIX advisory locks.  (The more

** I discover, the more I realize the a POSIX advisory locks are

** an abomination.)

**

** If you close a file descriptor that points to a file that has locks,

** all locks on that file that are owned by the current process are

** released.  To work around this problem, each OsFile structure contains

** a pointer to an openCnt structure.  There is one openCnt structure

** per open inode, which means that multiple OsFiles can point to a single

** openCnt.  When an attempt is made to close an OsFile, if there are

** other OsFiles open on the same inode that are holding locks, the call

** to close() the file descriptor is deferred until all of the locks clear.

** The openCnt structure keeps a list of file descriptors that need to

** be closed and that list is walked (and cleared) when the last lock

** clears.

**

** First, under Linux threads, because each thread has a separate

** process ID, lock operations in one thread do not override locks

** to the same file in other threads.  Linux threads behave like

** separate processes in this respect.  But, if you close a file

** descriptor in linux threads, all locks are cleared, even locks

** on other threads and even though the other threads have different

** process IDs.  Linux threads is inconsistent in this respect.

** (I'm beginning to think that linux threads is an abomination too.)

** The consequence of this all is that the hash table for the lockInfo

** structure has to include the process id as part of its key because

** locks in different threads are treated as distinct.  But the 

** openCnt structure should not include the process id in its

** key because close() clears lock on all threads, not just the current

** thread.  Were it not for this goofiness in linux threads, we could

** combine the lockInfo and openCnt structures into a single structure.

**

** 2004-Jun-28:

** On some versions of linux, threads can override each others locks.

** On others not.  Sometimes you can change the behavior on the same

** system by setting the LD_ASSUME_KERNEL environment variable.  The

** POSIX standard is silent as to which behavior is correct, as far

** as I can tell, so other versions of unix might show the same

** inconsistency.  There is no little doubt in my mind that posix

** advisory locks and linux threads are profoundly broken.

**

** To work around the inconsistencies, we have to test at runtime 

** whether or not threads can override each others locks.  This test

** is run once, the first time any lock is attempted.  A static 

** variable is set to record the results of this test for future

** use.

*/

/*

** An instance of the following structure serves as the key used

** to locate a particular lockInfo structure given its inode.

**

** If threads cannot override each others locks, then we set the

** lockKey.tid field to the thread ID.  If threads can override

** each others locks then tid is always set to zero.  tid is omitted

** if we compile without threading support.

*/

struct lockKey {

  dev_t dev;       /* Device number */

  ino_t ino;       /* Inode number */

#ifdef SQLITE_UNIX_THREADS

  pthread_t tid;   /* Thread ID or zero if threads can override each other */

#endif

};

/*

** An instance of the following structure is allocated for each open

** inode on each thread with a different process ID.  (Threads have

** different process IDs on linux, but not on most other unixes.)

**

** A single inode can have multiple file descriptors, so each OsFile

** structure contains a pointer to an instance of this object and this

** object keeps a count of the number of OsFiles pointing to it.

*/

struct lockInfo {

  struct lockKey key;  /* The lookup key */

  int cnt;             /* Number of SHARED locks held */

  int locktype;        /* One of SHARED_LOCK, RESERVED_LOCK etc. */

  int nRef;            /* Number of pointers to this structure */

};

/*

** An instance of the following structure serves as the key used

** to locate a particular openCnt structure given its inode.  This

** is the same as the lockKey except that the thread ID is omitted.

*/

struct openKey {

  dev_t dev;   /* Device number */

  ino_t ino;   /* Inode number */

};

/*

** An instance of the following structure is allocated for each open

** inode.  This structure keeps track of the number of locks on that

** inode.  If a close is attempted against an inode that is holding

** locks, the close is deferred until all locks clear by adding the

** file descriptor to be closed to the pending list.

*/

struct openCnt {

  struct openKey key;   /* The lookup key */

  int nRef;             /* Number of pointers to this structure */

  int nLock;            /* Number of outstanding locks */

  int nPending;         /* Number of pending close() operations */

  int *aPending;        /* Malloced space holding fd's awaiting a close() */

};

/* 

** These hash tables map inodes and file descriptors (really, lockKey and

** openKey structures) into lockInfo and openCnt structures.  Access to 

** these hash tables must be protected by a mutex.

*/

static Hash lockHash = {SQLITE_HASH_BINARY, 0, 0, 0, 

    sqlite3ThreadSafeMalloc, sqlite3ThreadSafeFree, 0, 0};

static Hash openHash = {SQLITE_HASH_BINARY, 0, 0, 0, 

    sqlite3ThreadSafeMalloc, sqlite3ThreadSafeFree, 0, 0};

#ifdef SQLITE_UNIX_THREADS

/*

** This variable records whether or not threads can override each others

** locks.

**

**    0:  No.  Threads cannot override each others locks.

**    1:  Yes.  Threads can override each others locks.

**   -1:  We don't know yet.

**

** On some systems, we know at compile-time if threads can override each

** others locks.  On those systems, the SQLITE_THREAD_OVERRIDE_LOCK macro

** will be set appropriately.  On other systems, we have to check at

** runtime.  On these latter systems, SQLTIE_THREAD_OVERRIDE_LOCK is

** undefined.

**

** This variable normally has file scope only.  But during testing, we make

** it a global so that the test code can change its value in order to verify

** that the right stuff happens in either case.

*/

#ifndef SQLITE_THREAD_OVERRIDE_LOCK

# define SQLITE_THREAD_OVERRIDE_LOCK -1

#endif

#ifdef SQLITE_TEST

int threadsOverrideEachOthersLocks = SQLITE_THREAD_OVERRIDE_LOCK;

#else

static int threadsOverrideEachOthersLocks = SQLITE_THREAD_OVERRIDE_LOCK;

#endif

/*

** This structure holds information passed into individual test

** threads by the testThreadLockingBehavior() routine.

*/

struct threadTestData {

  int fd;                /* File to be locked */

  struct flock lock;     /* The locking operation */

  int result;            /* Result of the locking operation */

};

#ifdef SQLITE_LOCK_TRACE

/*

** Print out information about all locking operations.

**

** This routine is used for troubleshooting locks on multithreaded

** platforms.  Enable by compiling with the -DSQLITE_LOCK_TRACE

** command-line option on the compiler.  This code is normally

** turned off.

*/

static int lockTrace(int fd, int op, struct flock *p){

  char *zOpName, *zType;

  int s;

  int savedErrno;

  if( op==F_GETLK ){

    zOpName = "GETLK";

  }else if( op==F_SETLK ){

    zOpName = "SETLK";

  }else{

    s = fcntl(fd, op, p);

    sqlite3DebugPrintf("fcntl unknown %d %d %d\n", fd, op, s);

    return s;

  }

  if( p->l_type==F_RDLCK ){

    zType = "RDLCK";

  }else if( p->l_type==F_WRLCK ){

    zType = "WRLCK";

  }else if( p->l_type==F_UNLCK ){

    zType = "UNLCK";

  }else{

    assert( 0 );

  }

  assert( p->l_whence==SEEK_SET );

  s = fcntl(fd, op, p);

  savedErrno = errno;

  sqlite3DebugPrintf("fcntl %d %d %s %s %d %d %d %d\n",

     threadid, fd, zOpName, zType, (int)p->l_start, (int)p->l_len,

     (int)p->l_pid, s);

  if( s && op==F_SETLK && (p->l_type==F_RDLCK || p->l_type==F_WRLCK) ){

    struct flock l2;

    l2 = *p;

    fcntl(fd, F_GETLK, &l2);

    if( l2.l_type==F_RDLCK ){

      zType = "RDLCK";

    }else if( l2.l_type==F_WRLCK ){

      zType = "WRLCK";

    }else if( l2.l_type==F_UNLCK ){

      zType = "UNLCK";

    }else{

      assert( 0 );

    }

    sqlite3DebugPrintf("fcntl-failure-reason: %s %d %d %d\n",

       zType, (int)l2.l_start, (int)l2.l_len, (int)l2.l_pid);

  }

  errno = savedErrno;

  return s;

}

#define fcntl lockTrace

#endif /* SQLITE_LOCK_TRACE */

/*

** The testThreadLockingBehavior() routine launches two separate

** threads on this routine.  This routine attempts to lock a file

** descriptor then returns.  The success or failure of that attempt

** allows the testThreadLockingBehavior() procedure to determine

** whether or not threads can override each others locks.

*/

static void *threadLockingTest(void *pArg){

  struct threadTestData *pData = (struct threadTestData*)pArg;

  pData->result = fcntl(pData->fd, F_SETLK, &pData->lock);

  return pArg;

}

/*

** This procedure attempts to determine whether or not threads

** can override each others locks then sets the 

** threadsOverrideEachOthersLocks variable appropriately.

*/

static void testThreadLockingBehavior(int fd_orig){

  int fd;

  struct threadTestData d[2];

  pthread_t t[2];

  fd = dup(fd_orig);

  if( fd<0 ) return;

  memset(d, 0, sizeof(d));

  d[0].fd = fd;

  d[0].lock.l_type = F_RDLCK;

  d[0].lock.l_len = 1;

  d[0].lock.l_start = 0;

  d[0].lock.l_whence = SEEK_SET;

  d[1] = d[0];

  d[1].lock.l_type = F_WRLCK;

  pthread_create(&t[0], 0, threadLockingTest, &d[0]);

  pthread_create(&t[1], 0, threadLockingTest, &d[1]);

  pthread_join(t[0], 0);

  pthread_join(t[1], 0);

  close(fd);

  threadsOverrideEachOthersLocks =  d[0].result==0 && d[1].result==0;

}

#endif /* SQLITE_UNIX_THREADS */

/*

** Release a lockInfo structure previously allocated by findLockInfo().

*/

static void releaseLockInfo(struct lockInfo *pLock){

  assert( sqlite3OsInMutex(1) );

  pLock->nRef--;

  if( pLock->nRef==0 ){

    sqlite3HashInsert(&lockHash, &pLock->key, sizeof(pLock->key), 0);

    sqlite3ThreadSafeFree(pLock);

  }

}

/*

** Release a openCnt structure previously allocated by findLockInfo().

*/

static void releaseOpenCnt(struct openCnt *pOpen){

  assert( sqlite3OsInMutex(1) );

  pOpen->nRef--;

  if( pOpen->nRef==0 ){

    sqlite3HashInsert(&openHash, &pOpen->key, sizeof(pOpen->key), 0);

    free(pOpen->aPending);

    sqlite3ThreadSafeFree(pOpen);

  }

}

/*

** Given a file descriptor, locate lockInfo and openCnt structures that

** describes that file descriptor.  Create new ones if necessary.  The

** return values might be uninitialized if an error occurs.

**

** Return the number of errors.

*/

static int findLockInfo(

  int fd,                      /* The file descriptor used in the key */

  struct lockInfo **ppLock,    /* Return the lockInfo structure here */

  struct openCnt **ppOpen      /* Return the openCnt structure here */

){

  int rc;

  struct lockKey key1;

  struct openKey key2;

  struct stat statbuf;