os_unix.c

sqlite最新源码

#endif
#ifdef ENOTSUP
  case ENOTSUP: 
    /* invalid fd, unless during file system support introspection, in which 
     * it actually means what it says */
#endif
  case EIO:
  case EBADF:
  case EINVAL:
  case ENOTCONN:
  case ENODEV:
  case ENXIO:
  case ENOENT:
  case ESTALE:
  case ENOSYS:
    /* these should force the client to close the file and reconnect */
    
  default: 
    return sqliteIOErr;
  }
}



/******************************************************************************
****************** Begin Unique File ID Utility Used By VxWorks ***************
**
** On most versions of unix, we can get a unique ID for a file by concatenating
** the device number and the inode number.  But this does not work on VxWorks.
** On VxWorks, a unique file id must be based on the canonical filename.
**
** A pointer to an instance of the following structure can be used as a
** unique file ID in VxWorks.  Each instance of this structure contains
** a copy of the canonical filename.  There is also a reference count.  
** The structure is reclaimed when the number of pointers to it drops to
** zero.
**
** There are never very many files open at one time and lookups are not
** a performance-critical path, so it is sufficient to put these
** structures on a linked list.
*/
struct vxworksFileId {
  struct vxworksFileId *pNext;  /* Next in a list of them all */
  int nRef;                     /* Number of references to this one */
  int nName;                    /* Length of the zCanonicalName[] string */
  char *zCanonicalName;         /* Canonical filename */
};

#if OS_VXWORKS
/* 
** All unique filenames are held on a linked list headed by this
** variable:
*/
static struct vxworksFileId *vxworksFileList = 0;

/*
** Simplify a filename into its canonical form
** by making the following changes:
**
**  * removing any trailing and duplicate /
**  * convert /./ into just /
**  * convert /A/../ where A is any simple name into just /
**
** Changes are made in-place.  Return the new name length.
**
** The original filename is in z[0..n-1].  Return the number of
** characters in the simplified name.
*/
static int vxworksSimplifyName(char *z, int n){
  int i, j;
  while( n>1 && z[n-1]=='/' ){ n--; }
  for(i=j=0; i<n; i++){
    if( z[i]=='/' ){
      if( z[i+1]=='/' ) continue;
      if( z[i+1]=='.' && i+2<n && z[i+2]=='/' ){
        i += 1;
        continue;
      }
      if( z[i+1]=='.' && i+3<n && z[i+2]=='.' && z[i+3]=='/' ){
        while( j>0 && z[j-1]!='/' ){ j--; }
        if( j>0 ){ j--; }
        i += 2;
        continue;
      }
    }
    z[j++] = z[i];
  }
  z[j] = 0;
  return j;
}

/*
** Find a unique file ID for the given absolute pathname.  Return
** a pointer to the vxworksFileId object.  This pointer is the unique
** file ID.
**
** The nRef field of the vxworksFileId object is incremented before
** the object is returned.  A new vxworksFileId object is created
** and added to the global list if necessary.
**
** If a memory allocation error occurs, return NULL.
*/
static struct vxworksFileId *vxworksFindFileId(const char *zAbsoluteName){
  struct vxworksFileId *pNew;         /* search key and new file ID */
  struct vxworksFileId *pCandidate;   /* For looping over existing file IDs */
  int n;                              /* Length of zAbsoluteName string */

  assert( zAbsoluteName[0]=='/' );
  n = (int)strlen(zAbsoluteName);
  pNew = sqlite3_malloc( sizeof(*pNew) + (n+1) );
  if( pNew==0 ) return 0;
  pNew->zCanonicalName = (char*)&pNew[1];
  memcpy(pNew->zCanonicalName, zAbsoluteName, n+1);
  n = vxworksSimplifyName(pNew->zCanonicalName, n);

  /* Search for an existing entry that matching the canonical name.
  ** If found, increment the reference count and return a pointer to
  ** the existing file ID.
  */
  unixEnterMutex();
  for(pCandidate=vxworksFileList; pCandidate; pCandidate=pCandidate->pNext){
    if( pCandidate->nName==n 
     && memcmp(pCandidate->zCanonicalName, pNew->zCanonicalName, n)==0
    ){
       sqlite3_free(pNew);
       pCandidate->nRef++;
       unixLeaveMutex();
       return pCandidate;
    }
  }

  /* No match was found.  We will make a new file ID */
  pNew->nRef = 1;
  pNew->nName = n;
  pNew->pNext = vxworksFileList;
  vxworksFileList = pNew;
  unixLeaveMutex();
  return pNew;
}

/*
** Decrement the reference count on a vxworksFileId object.  Free
** the object when the reference count reaches zero.
*/
static void vxworksReleaseFileId(struct vxworksFileId *pId){
  unixEnterMutex();
  assert( pId->nRef>0 );
  pId->nRef--;
  if( pId->nRef==0 ){
    struct vxworksFileId **pp;
    for(pp=&vxworksFileList; *pp && *pp!=pId; pp = &((*pp)->pNext)){}
    assert( *pp==pId );
    *pp = pId->pNext;
    sqlite3_free(pId);
  }
  unixLeaveMutex();
}
#endif /* OS_VXWORKS */
/*************** End of Unique File ID Utility Used By VxWorks ****************
******************************************************************************/


/******************************************************************************
*************************** Posix Advisory Locking ****************************
**
** POSIX advisory locks are broken by design.  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.
**
** This 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.
**
** (Aside: The use of inode numbers as unique IDs does not work on VxWorks.
** For VxWorks, we have to use the alternative unique ID system based on
** canonical filename and implemented in the previous division.)
**
** The sqlite3_file 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 unixFile 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.
**
** But wait:  there are yet more problems with POSIX advisory locks.
**
** 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 unixFile structure contains
** a pointer to an unixOpenCnt structure.  There is one unixOpenCnt structure
** per open inode, which means that multiple unixFile can point to a single
** unixOpenCnt.  When an attempt is made to close an unixFile, if there are
** other unixFile 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 unixOpenCnt structure keeps a list of file descriptors that need to
** be closed and that list is walked (and cleared) when the last lock
** clears.
**
** Yet another problem:  LinuxThreads do not play well with posix locks.
**
** Many older versions of linux use the LinuxThreads library which is
** not posix compliant.  Under LinuxThreads, a lock created by thread
** A cannot be modified or overridden by a different thread B.
** Only thread A can modify the lock.  Locking behavior is correct
** if the appliation uses the newer Native Posix Thread Library (NPTL)
** on linux - with NPTL a lock created by thread A can override locks
** in thread B.  But there is no way to know at compile-time which
** threading library is being used.  So there is no way to know at
** compile-time whether or not thread A can override locks on thread B.
** We have to do a run-time check to discover the behavior of the
** current process.
**
** On systems where thread A is unable to modify locks created by
** thread B, we have to keep track of which thread created each
** lock.  Hence there is an extra field in the key to the unixLockInfo
** structure to record this information.  And on those systems it
** is illegal to begin a transaction in one thread and finish it
** in another.  For this latter restriction, there is no work-around.
** It is a limitation of LinuxThreads.
*/

/*
** Set or check the unixFile.tid field.  This field is set when an unixFile
** is first opened.  All subsequent uses of the unixFile verify that the
** same thread is operating on the unixFile.  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 while locks are held.
**
** Version 3.3.1 (2006-01-15):  unixFile 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 is now the most common behavior)
** or if no locks are held.  But the unixFile.pLock field needs to be
** recomputed because its key includes the thread-id.  See the 
** transferOwnership() function below for additional information
*/
#if SQLITE_THREADSAFE && defined(__linux__)
# 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

/*
** An instance of the following structure serves as the key used
** to locate a particular unixOpenCnt structure given its inode.  This
** is the same as the unixLockKey except that the thread ID is omitted.
*/
struct unixFileId {
  dev_t dev;                  /* Device number */
#if OS_VXWORKS
  struct vxworksFileId *pId;  /* Unique file ID for vxworks. */
#else
  ino_t ino;                  /* Inode number */
#endif
};

/*
** An instance of the following structure serves as the key used
** to locate a particular unixLockInfo structure given its inode.
**
** If threads cannot override each others locks (LinuxThreads), then we
** set the unixLockKey.tid field to the thread ID.  If threads can override
** each others locks (Posix and NPTL) then tid is always set to zero.
** tid is omitted if we compile without threading support or on an OS
** other than linux.
*/
struct unixLockKey {
  struct unixFileId fid;  /* Unique identifier for the file */
#if SQLITE_THREADSAFE && defined(__linux__)
  pthread_t tid;  /* Thread ID of lock owner. Zero if not using LinuxThreads */
#endif
};

/*
** An instance of the following structure is allocated for each open
** inode.  Or, on LinuxThreads, there is one of these structures for
** each inode opened by each thread.
**
** A single inode can have multiple file descriptors, so each unixFile
** structure contains a pointer to an instance of this object and this
** object keeps a count of the number of unixFile pointing to it.
*/
struct unixLockInfo {
  struct unixLockKey lockKey;     /* 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 */
  struct unixLockInfo *pNext;     /* List of all unixLockInfo objects */
  struct unixLockInfo *pPrev;     /*    .... doubly linked */
};

/*
** 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.
**
** TODO:  Consider changing this so that there is only a single file
** descriptor for each open file, even when it is opened multiple times.
** The close() system call would only occur when the last database
** using the file closes.
*/
struct unixOpenCnt {
  struct unixFileId fileId;   /* 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() */
#if OS_VXWORKS
  sem_t *pSem;                     /* Named POSIX semaphore */
  char aSemName[MAX_PATHNAME+1];   /* Name of that semaphore */
#endif
  struct unixOpenCnt *pNext, *pPrev;   /* List of all unixOpenCnt objects */
};

/*
** Lists of all unixLockInfo and unixOpenCnt objects.  These used to be hash
** tables.  But the number of objects is rarely more than a dozen and
** never exceeds a few thousand.  And lookup is not on a critical
** path so a simple linked list will suffice.
*/
static struct unixLockInfo *lockList = 0;
static struct unixOpenCnt *openList = 0;

/*
** This variable remembers whether or not threads can override each others
** locks.
**
**    0:  No.  Threads cannot override each others locks.  (LinuxThreads)
**    1:  Yes.  Threads can override each others locks.  (Posix & NLPT)
**   -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.
*/
#if SQLITE_THREADSAFE && defined(__linux__)
#  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
#endif

/*
** This structure holds information passed into individual test
** threads by the testThreadLockingBehavior() routine.
*/
struct threadTestData {