os_unix.c

sqlite 嵌入式数据库的源码

/*
** 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 */


#include <time.h>
#include <errno.h>
#include <unistd.h>

/*
** Do not include any of the File I/O interface procedures if the
** SQLITE_OMIT_DISKIO macro is defined (indicating that there 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 its DOS - what did you expect?
*/
#ifdef __DJGPP__
# define fcntl(A,B,C) 0
#endif

/*
** Macros used to determine whether or not to use threads.  The
** SQLITE_UNIX_THREADS macro is defined if we are synchronizing for
** Posix threads and SQLITE_W32_THREADS is defined if we are
** synchronizing using Win32 threads.
*/
#if defined(THREADSAFE) && THREADSAFE
# include <pthread.h>
# define SQLITE_UNIX_THREADS 1
#endif


/*
** Include code that is common to all os_*.c files
*/
#include "os_common.h"

#if defined(THREADSAFE) && THREADSAFE && defined(__linux__)
#define getpid pthread_self
#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 also
** set to zero 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 cannot 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 table maps inodes and process IDs 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, 0, 0 };
static Hash openHash = { SQLITE_HASH_BINARY, 0, 0, 0, 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.
*/
static int threadsOverrideEachOthersLocks = -1;

/*
** 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 */
};

/*
** 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(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){
  pLock->nRef--;
  if( pLock->nRef==0 ){
    sqlite3HashInsert(&lockHash, &pLock->key, sizeof(pLock->key), 0);
    sqliteFree(pLock);
  }
}

/*
** Release a openCnt structure previously allocated by findLockInfo().
*/
static void releaseOpenCnt(struct openCnt *pOpen){
  pOpen->nRef--;
  if( pOpen->nRef==0 ){
    sqlite3HashInsert(&openHash, &pOpen->key, sizeof(pOpen->key), 0);
    sqliteFree(pOpen->aPending);
    sqliteFree(pOpen);
  }
}

/*
** Given a file descriptor, locate lockInfo and openCnt structures that
** describes that file descriptor.  Create a new ones if necessary.  The
** return values might be unset 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;
  struct lockInfo *pLock;
  struct openCnt *pOpen;
  rc = fstat(fd, &statbuf);
  if( rc!=0 ) return 1;
  memset(&key1, 0, sizeof(key1));
  key1.dev = statbuf.st_dev;
  key1.ino = statbuf.st_ino;
#ifdef SQLITE_UNIX_THREADS
  if( threadsOverrideEachOthersLocks<0 ){
    testThreadLockingBehavior(fd);
  }
  key1.tid = threadsOverrideEachOthersLocks ? 0 : pthread_self();
#endif
  memset(&key2, 0, sizeof(key2));
  key2.dev = statbuf.st_dev;
  key2.ino = statbuf.st_ino;
  pLock = (struct lockInfo*)sqlite3HashFind(&lockHash, &key1, sizeof(key1));
  if( pLock==0 ){
    struct lockInfo *pOld;
    pLock = sqliteMallocRaw( sizeof(*pLock) );
    if( pLock==0 ) return 1;
    pLock->key = key1;
    pLock->nRef = 1;
    pLock->cnt = 0;
    pLock->locktype = 0;
    pOld = sqlite3HashInsert(&lockHash, &pLock->key, sizeof(key1), pLock);
    if( pOld!=0 ){
      assert( pOld==pLock );
      sqliteFree(pLock);
      return 1;
    }
  }else{
    pLock->nRef++;
  }
  *ppLock = pLock;
  pOpen = (struct openCnt*)sqlite3HashFind(&openHash, &key2, sizeof(key2));
  if( pOpen==0 ){
    struct openCnt *pOld;
    pOpen = sqliteMallocRaw( sizeof(*pOpen) );
    if( pOpen==0 ){
      releaseLockInfo(pLock);
      return 1;
    }
    pOpen->key = key2;
    pOpen->nRef = 1;
    pOpen->nLock = 0;
    pOpen->nPending = 0;
    pOpen->aPending = 0;
    pOld = sqlite3HashInsert(&openHash, &pOpen->key, sizeof(key2), pOpen);
    if( pOld!=0 ){
      assert( pOld==pOpen );
      sqliteFree(pOpen);
      releaseLockInfo(pLock);
      return 1;
    }
  }else{
    pOpen->nRef++;
  }
  *ppOpen = pOpen;
  return 0;
}

/*
** Delete the named file
*/
int sqlite3OsDelete(const char *zFilename){
  unlink(zFilename);
  return SQLITE_OK;
}

/*
** Return TRUE if the named file exists.
*/
int sqlite3OsFileExists(const char *zFilename){
  return access(zFilename, 0)==0;
}

/*
** Attempt to open a file for both reading and writing.  If that
** fails, try opening it read-only.  If the file does not exist,
** try to create it.
**
** On success, a handle for the open file is written to *id
** and *pReadonly is set to 0 if the file was opened for reading and
** writing or 1 if the file was opened read-only.  The function returns
** SQLITE_OK.
**
** On failure, the function returns SQLITE_CANTOPEN and leaves
** *id and *pReadonly unchanged.
*/
int sqlite3OsOpenReadWrite(
  const char *zFilename,
  OsFile *id,
  int *pReadonly
){
  int rc;
  assert( !id->isOpen );
  id->dirfd = -1;
  id->h = open(zFilename, O_RDWR|O_CREAT|O_LARGEFILE|O_BINARY,
                          SQLITE_DEFAULT_FILE_PERMISSIONS);
  if( id->h<0 ){
#ifdef EISDIR