4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
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.
11 ******************************************************************************
13 ** This file contains the VFS implementation for unix-like operating systems
14 ** include Linux, MacOSX, *BSD, QNX, VxWorks, AIX, HPUX, and others.
16 ** There are actually several different VFS implementations in this file.
17 ** The differences are in the way that file locking is done. The default
18 ** implementation uses Posix Advisory Locks. Alternative implementations
19 ** use flock(), dot-files, various proprietary locking schemas, or simply
20 ** skip locking all together.
22 ** This source file is organized into divisions where the logic for various
23 ** subfunctions is contained within the appropriate division. PLEASE
24 ** KEEP THE STRUCTURE OF THIS FILE INTACT. New code should be placed
25 ** in the correct division and should be clearly labeled.
27 ** The layout of divisions is as follows:
29 ** * General-purpose declarations and utility functions.
30 ** * Unique file ID logic used by VxWorks.
31 ** * Various locking primitive implementations (all except proxy locking):
32 ** + for Posix Advisory Locks
34 ** + for dot-file locks
35 ** + for flock() locking
36 ** + for named semaphore locks (VxWorks only)
37 ** + for AFP filesystem locks (MacOSX only)
38 ** * sqlite3_file methods not associated with locking.
39 ** * Definitions of sqlite3_io_methods objects for all locking
40 ** methods plus "finder" functions for each locking method.
41 ** * sqlite3_vfs method implementations.
42 ** * Locking primitives for the proxy uber-locking-method. (MacOSX only)
43 ** * Definitions of sqlite3_vfs objects for all locking methods
44 ** plus implementations of sqlite3_os_init() and sqlite3_os_end().
46 #include "sqliteInt.h"
47 #if SQLITE_OS_UNIX /* This file is used on unix only */
50 ** There are various methods for file locking used for concurrency
53 ** 1. POSIX locking (the default),
55 ** 3. Dot-file locking,
56 ** 4. flock() locking,
57 ** 5. AFP locking (OSX only),
58 ** 6. Named POSIX semaphores (VXWorks only),
59 ** 7. proxy locking. (OSX only)
61 ** Styles 4, 5, and 7 are only available of SQLITE_ENABLE_LOCKING_STYLE
62 ** is defined to 1. The SQLITE_ENABLE_LOCKING_STYLE also enables automatic
63 ** selection of the appropriate locking style based on the filesystem
64 ** where the database is located.
66 #if !defined(SQLITE_ENABLE_LOCKING_STYLE)
67 # if defined(__APPLE__)
68 # define SQLITE_ENABLE_LOCKING_STYLE 1
70 # define SQLITE_ENABLE_LOCKING_STYLE 0
74 /* Use pread() and pwrite() if they are available */
75 #if defined(__APPLE__)
77 # define HAVE_PWRITE 1
79 #if defined(HAVE_PREAD64) && defined(HAVE_PWRITE64)
81 # define USE_PREAD64 1
82 #elif defined(HAVE_PREAD) && defined(HAVE_PWRITE)
88 ** standard include files.
90 #include <sys/types.h>
93 #include <sys/ioctl.h>
98 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
99 # include <sys/mman.h>
102 #if SQLITE_ENABLE_LOCKING_STYLE
103 # include <sys/ioctl.h>
104 # include <sys/file.h>
105 # include <sys/param.h>
106 #endif /* SQLITE_ENABLE_LOCKING_STYLE */
108 #if defined(__APPLE__) && ((__MAC_OS_X_VERSION_MIN_REQUIRED > 1050) || \
109 (__IPHONE_OS_VERSION_MIN_REQUIRED > 2000))
110 # if (!defined(TARGET_OS_EMBEDDED) || (TARGET_OS_EMBEDDED==0)) \
111 && (!defined(TARGET_IPHONE_SIMULATOR) || (TARGET_IPHONE_SIMULATOR==0))
112 # define HAVE_GETHOSTUUID 1
114 # warning "gethostuuid() is disabled."
120 # include <sys/ioctl.h>
121 # include <semaphore.h>
123 #endif /* OS_VXWORKS */
125 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
126 # include <sys/mount.h>
134 ** Allowed values of unixFile.fsFlags
136 #define SQLITE_FSFLAGS_IS_MSDOS 0x1
139 ** If we are to be thread-safe, include the pthreads header and define
140 ** the SQLITE_UNIX_THREADS macro.
142 #if SQLITE_THREADSAFE
143 # include <pthread.h>
144 # define SQLITE_UNIX_THREADS 1
148 ** Default permissions when creating a new file
150 #ifndef SQLITE_DEFAULT_FILE_PERMISSIONS
151 # define SQLITE_DEFAULT_FILE_PERMISSIONS 0644
155 ** Default permissions when creating auto proxy dir
157 #ifndef SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
158 # define SQLITE_DEFAULT_PROXYDIR_PERMISSIONS 0755
162 ** Maximum supported path-length.
164 #define MAX_PATHNAME 512
167 ** Maximum supported symbolic links
169 #define SQLITE_MAX_SYMLINKS 100
171 /* Always cast the getpid() return type for compatibility with
172 ** kernel modules in VxWorks. */
173 #define osGetpid(X) (pid_t)getpid()
176 ** Only set the lastErrno if the error code is a real error and not
177 ** a normal expected return code of SQLITE_BUSY or SQLITE_OK
179 #define IS_LOCK_ERROR(x) ((x != SQLITE_OK) && (x != SQLITE_BUSY))
181 /* Forward references */
182 typedef struct unixShm unixShm
; /* Connection shared memory */
183 typedef struct unixShmNode unixShmNode
; /* Shared memory instance */
184 typedef struct unixInodeInfo unixInodeInfo
; /* An i-node */
185 typedef struct UnixUnusedFd UnixUnusedFd
; /* An unused file descriptor */
188 ** Sometimes, after a file handle is closed by SQLite, the file descriptor
189 ** cannot be closed immediately. In these cases, instances of the following
190 ** structure are used to store the file descriptor while waiting for an
191 ** opportunity to either close or reuse it.
193 struct UnixUnusedFd
{
194 int fd
; /* File descriptor to close */
195 int flags
; /* Flags this file descriptor was opened with */
196 UnixUnusedFd
*pNext
; /* Next unused file descriptor on same file */
200 ** The unixFile structure is subclass of sqlite3_file specific to the unix
201 ** VFS implementations.
203 typedef struct unixFile unixFile
;
205 sqlite3_io_methods
const *pMethod
; /* Always the first entry */
206 sqlite3_vfs
*pVfs
; /* The VFS that created this unixFile */
207 unixInodeInfo
*pInode
; /* Info about locks on this inode */
208 int h
; /* The file descriptor */
209 unsigned char eFileLock
; /* The type of lock held on this fd */
210 unsigned short int ctrlFlags
; /* Behavioral bits. UNIXFILE_* flags */
211 int lastErrno
; /* The unix errno from last I/O error */
212 void *lockingContext
; /* Locking style specific state */
213 UnixUnusedFd
*pPreallocatedUnused
; /* Pre-allocated UnixUnusedFd */
214 const char *zPath
; /* Name of the file */
215 unixShm
*pShm
; /* Shared memory segment information */
216 int szChunk
; /* Configured by FCNTL_CHUNK_SIZE */
217 #if SQLITE_MAX_MMAP_SIZE>0
218 int nFetchOut
; /* Number of outstanding xFetch refs */
219 sqlite3_int64 mmapSize
; /* Usable size of mapping at pMapRegion */
220 sqlite3_int64 mmapSizeActual
; /* Actual size of mapping at pMapRegion */
221 sqlite3_int64 mmapSizeMax
; /* Configured FCNTL_MMAP_SIZE value */
222 void *pMapRegion
; /* Memory mapped region */
224 int sectorSize
; /* Device sector size */
225 int deviceCharacteristics
; /* Precomputed device characteristics */
226 #if SQLITE_ENABLE_LOCKING_STYLE
227 int openFlags
; /* The flags specified at open() */
229 #if SQLITE_ENABLE_LOCKING_STYLE || defined(__APPLE__)
230 unsigned fsFlags
; /* cached details from statfs() */
233 struct vxworksFileId
*pId
; /* Unique file ID */
236 /* The next group of variables are used to track whether or not the
237 ** transaction counter in bytes 24-27 of database files are updated
238 ** whenever any part of the database changes. An assertion fault will
239 ** occur if a file is updated without also updating the transaction
240 ** counter. This test is made to avoid new problems similar to the
241 ** one described by ticket #3584.
243 unsigned char transCntrChng
; /* True if the transaction counter changed */
244 unsigned char dbUpdate
; /* True if any part of database file changed */
245 unsigned char inNormalWrite
; /* True if in a normal write operation */
250 /* In test mode, increase the size of this structure a bit so that
251 ** it is larger than the struct CrashFile defined in test6.c.
257 /* This variable holds the process id (pid) from when the xRandomness()
258 ** method was called. If xOpen() is called from a different process id,
259 ** indicating that a fork() has occurred, the PRNG will be reset.
261 static pid_t randomnessPid
= 0;
264 ** Allowed values for the unixFile.ctrlFlags bitmask:
266 #define UNIXFILE_EXCL 0x01 /* Connections from one process only */
267 #define UNIXFILE_RDONLY 0x02 /* Connection is read only */
268 #define UNIXFILE_PERSIST_WAL 0x04 /* Persistent WAL mode */
269 #ifndef SQLITE_DISABLE_DIRSYNC
270 # define UNIXFILE_DIRSYNC 0x08 /* Directory sync needed */
272 # define UNIXFILE_DIRSYNC 0x00
274 #define UNIXFILE_PSOW 0x10 /* SQLITE_IOCAP_POWERSAFE_OVERWRITE */
275 #define UNIXFILE_DELETE 0x20 /* Delete on close */
276 #define UNIXFILE_URI 0x40 /* Filename might have query parameters */
277 #define UNIXFILE_NOLOCK 0x80 /* Do no file locking */
280 ** Include code that is common to all os_*.c files
282 #include "os_common.h"
285 ** Define various macros that are missing from some systems.
288 # define O_LARGEFILE 0
290 #ifdef SQLITE_DISABLE_LFS
292 # define O_LARGEFILE 0
295 # define O_NOFOLLOW 0
302 ** The threadid macro resolves to the thread-id or to 0. Used for
303 ** testing and debugging only.
305 #if SQLITE_THREADSAFE
306 #define threadid pthread_self()
312 ** HAVE_MREMAP defaults to true on Linux and false everywhere else.
314 #if !defined(HAVE_MREMAP)
315 # if defined(__linux__) && defined(_GNU_SOURCE)
316 # define HAVE_MREMAP 1
318 # define HAVE_MREMAP 0
323 ** Explicitly call the 64-bit version of lseek() on Android. Otherwise, lseek()
324 ** is the 32-bit version, even if _FILE_OFFSET_BITS=64 is defined.
327 # define lseek lseek64
332 ** Linux-specific IOCTL magic numbers used for controlling F2FS
334 #define F2FS_IOCTL_MAGIC 0xf5
335 #define F2FS_IOC_START_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 1)
336 #define F2FS_IOC_COMMIT_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 2)
337 #define F2FS_IOC_START_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 3)
338 #define F2FS_IOC_ABORT_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 5)
339 #define F2FS_IOC_GET_FEATURES _IOR(F2FS_IOCTL_MAGIC, 12, u32)
340 #define F2FS_FEATURE_ATOMIC_WRITE 0x0004
341 #endif /* __linux__ */
345 ** Different Unix systems declare open() in different ways. Same use
346 ** open(const char*,int,mode_t). Others use open(const char*,int,...).
347 ** The difference is important when using a pointer to the function.
349 ** The safest way to deal with the problem is to always use this wrapper
350 ** which always has the same well-defined interface.
352 static int posixOpen(const char *zFile
, int flags
, int mode
){
353 return open(zFile
, flags
, mode
);
356 /* Forward reference */
357 static int openDirectory(const char*, int*);
358 static int unixGetpagesize(void);
361 ** Many system calls are accessed through pointer-to-functions so that
362 ** they may be overridden at runtime to facilitate fault injection during
363 ** testing and sandboxing. The following array holds the names and pointers
364 ** to all overrideable system calls.
366 static struct unix_syscall
{
367 const char *zName
; /* Name of the system call */
368 sqlite3_syscall_ptr pCurrent
; /* Current value of the system call */
369 sqlite3_syscall_ptr pDefault
; /* Default value */
371 { "open", (sqlite3_syscall_ptr
)posixOpen
, 0 },
372 #define osOpen ((int(*)(const char*,int,int))aSyscall[0].pCurrent)
374 { "close", (sqlite3_syscall_ptr
)close
, 0 },
375 #define osClose ((int(*)(int))aSyscall[1].pCurrent)
377 { "access", (sqlite3_syscall_ptr
)access
, 0 },
378 #define osAccess ((int(*)(const char*,int))aSyscall[2].pCurrent)
380 { "getcwd", (sqlite3_syscall_ptr
)getcwd
, 0 },
381 #define osGetcwd ((char*(*)(char*,size_t))aSyscall[3].pCurrent)
383 { "stat", (sqlite3_syscall_ptr
)stat
, 0 },
384 #define osStat ((int(*)(const char*,struct stat*))aSyscall[4].pCurrent)
387 ** The DJGPP compiler environment looks mostly like Unix, but it
388 ** lacks the fcntl() system call. So redefine fcntl() to be something
389 ** that always succeeds. This means that locking does not occur under
390 ** DJGPP. But it is DOS - what did you expect?
394 #define osFstat(a,b,c) 0
396 { "fstat", (sqlite3_syscall_ptr
)fstat
, 0 },
397 #define osFstat ((int(*)(int,struct stat*))aSyscall[5].pCurrent)
400 { "ftruncate", (sqlite3_syscall_ptr
)ftruncate
, 0 },
401 #define osFtruncate ((int(*)(int,off_t))aSyscall[6].pCurrent)
403 { "fcntl", (sqlite3_syscall_ptr
)fcntl
, 0 },
404 #define osFcntl ((int(*)(int,int,...))aSyscall[7].pCurrent)
406 { "read", (sqlite3_syscall_ptr
)read
, 0 },
407 #define osRead ((ssize_t(*)(int,void*,size_t))aSyscall[8].pCurrent)
409 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
410 { "pread", (sqlite3_syscall_ptr
)pread
, 0 },
412 { "pread", (sqlite3_syscall_ptr
)0, 0 },
414 #define osPread ((ssize_t(*)(int,void*,size_t,off_t))aSyscall[9].pCurrent)
416 #if defined(USE_PREAD64)
417 { "pread64", (sqlite3_syscall_ptr
)pread64
, 0 },
419 { "pread64", (sqlite3_syscall_ptr
)0, 0 },
421 #define osPread64 ((ssize_t(*)(int,void*,size_t,off64_t))aSyscall[10].pCurrent)
423 { "write", (sqlite3_syscall_ptr
)write
, 0 },
424 #define osWrite ((ssize_t(*)(int,const void*,size_t))aSyscall[11].pCurrent)
426 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
427 { "pwrite", (sqlite3_syscall_ptr
)pwrite
, 0 },
429 { "pwrite", (sqlite3_syscall_ptr
)0, 0 },
431 #define osPwrite ((ssize_t(*)(int,const void*,size_t,off_t))\
432 aSyscall[12].pCurrent)
434 #if defined(USE_PREAD64)
435 { "pwrite64", (sqlite3_syscall_ptr
)pwrite64
, 0 },
437 { "pwrite64", (sqlite3_syscall_ptr
)0, 0 },
439 #define osPwrite64 ((ssize_t(*)(int,const void*,size_t,off64_t))\
440 aSyscall[13].pCurrent)
442 { "fchmod", (sqlite3_syscall_ptr
)fchmod
, 0 },
443 #define osFchmod ((int(*)(int,mode_t))aSyscall[14].pCurrent)
445 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
446 { "fallocate", (sqlite3_syscall_ptr
)posix_fallocate
, 0 },
448 { "fallocate", (sqlite3_syscall_ptr
)0, 0 },
450 #define osFallocate ((int(*)(int,off_t,off_t))aSyscall[15].pCurrent)
452 { "unlink", (sqlite3_syscall_ptr
)unlink
, 0 },
453 #define osUnlink ((int(*)(const char*))aSyscall[16].pCurrent)
455 { "openDirectory", (sqlite3_syscall_ptr
)openDirectory
, 0 },
456 #define osOpenDirectory ((int(*)(const char*,int*))aSyscall[17].pCurrent)
458 { "mkdir", (sqlite3_syscall_ptr
)mkdir
, 0 },
459 #define osMkdir ((int(*)(const char*,mode_t))aSyscall[18].pCurrent)
461 { "rmdir", (sqlite3_syscall_ptr
)rmdir
, 0 },
462 #define osRmdir ((int(*)(const char*))aSyscall[19].pCurrent)
464 #if defined(HAVE_FCHOWN)
465 { "fchown", (sqlite3_syscall_ptr
)fchown
, 0 },
467 { "fchown", (sqlite3_syscall_ptr
)0, 0 },
469 #define osFchown ((int(*)(int,uid_t,gid_t))aSyscall[20].pCurrent)
471 #if defined(HAVE_FCHOWN)
472 { "geteuid", (sqlite3_syscall_ptr
)geteuid
, 0 },
474 { "geteuid", (sqlite3_syscall_ptr
)0, 0 },
476 #define osGeteuid ((uid_t(*)(void))aSyscall[21].pCurrent)
478 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
479 { "mmap", (sqlite3_syscall_ptr
)mmap
, 0 },
481 { "mmap", (sqlite3_syscall_ptr
)0, 0 },
483 #define osMmap ((void*(*)(void*,size_t,int,int,int,off_t))aSyscall[22].pCurrent)
485 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
486 { "munmap", (sqlite3_syscall_ptr
)munmap
, 0 },
488 { "munmap", (sqlite3_syscall_ptr
)0, 0 },
490 #define osMunmap ((int(*)(void*,size_t))aSyscall[23].pCurrent)
492 #if HAVE_MREMAP && (!defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0)
493 { "mremap", (sqlite3_syscall_ptr
)mremap
, 0 },
495 { "mremap", (sqlite3_syscall_ptr
)0, 0 },
497 #define osMremap ((void*(*)(void*,size_t,size_t,int,...))aSyscall[24].pCurrent)
499 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
500 { "getpagesize", (sqlite3_syscall_ptr
)unixGetpagesize
, 0 },
502 { "getpagesize", (sqlite3_syscall_ptr
)0, 0 },
504 #define osGetpagesize ((int(*)(void))aSyscall[25].pCurrent)
506 #if defined(HAVE_READLINK)
507 { "readlink", (sqlite3_syscall_ptr
)readlink
, 0 },
509 { "readlink", (sqlite3_syscall_ptr
)0, 0 },
511 #define osReadlink ((ssize_t(*)(const char*,char*,size_t))aSyscall[26].pCurrent)
513 #if defined(HAVE_LSTAT)
514 { "lstat", (sqlite3_syscall_ptr
)lstat
, 0 },
516 { "lstat", (sqlite3_syscall_ptr
)0, 0 },
518 #define osLstat ((int(*)(const char*,struct stat*))aSyscall[27].pCurrent)
520 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
521 { "ioctl", (sqlite3_syscall_ptr
)ioctl
, 0 },
523 { "ioctl", (sqlite3_syscall_ptr
)0, 0 },
525 #define osIoctl ((int(*)(int,int,...))aSyscall[28].pCurrent)
527 }; /* End of the overrideable system calls */
531 ** On some systems, calls to fchown() will trigger a message in a security
532 ** log if they come from non-root processes. So avoid calling fchown() if
533 ** we are not running as root.
535 static int robustFchown(int fd
, uid_t uid
, gid_t gid
){
536 #if defined(HAVE_FCHOWN)
537 return osGeteuid() ? 0 : osFchown(fd
,uid
,gid
);
544 ** This is the xSetSystemCall() method of sqlite3_vfs for all of the
545 ** "unix" VFSes. Return SQLITE_OK opon successfully updating the
546 ** system call pointer, or SQLITE_NOTFOUND if there is no configurable
547 ** system call named zName.
549 static int unixSetSystemCall(
550 sqlite3_vfs
*pNotUsed
, /* The VFS pointer. Not used */
551 const char *zName
, /* Name of system call to override */
552 sqlite3_syscall_ptr pNewFunc
/* Pointer to new system call value */
555 int rc
= SQLITE_NOTFOUND
;
557 UNUSED_PARAMETER(pNotUsed
);
559 /* If no zName is given, restore all system calls to their default
560 ** settings and return NULL
563 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
564 if( aSyscall
[i
].pDefault
){
565 aSyscall
[i
].pCurrent
= aSyscall
[i
].pDefault
;
569 /* If zName is specified, operate on only the one system call
572 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
573 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ){
574 if( aSyscall
[i
].pDefault
==0 ){
575 aSyscall
[i
].pDefault
= aSyscall
[i
].pCurrent
;
578 if( pNewFunc
==0 ) pNewFunc
= aSyscall
[i
].pDefault
;
579 aSyscall
[i
].pCurrent
= pNewFunc
;
588 ** Return the value of a system call. Return NULL if zName is not a
589 ** recognized system call name. NULL is also returned if the system call
590 ** is currently undefined.
592 static sqlite3_syscall_ptr
unixGetSystemCall(
593 sqlite3_vfs
*pNotUsed
,
598 UNUSED_PARAMETER(pNotUsed
);
599 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
600 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) return aSyscall
[i
].pCurrent
;
606 ** Return the name of the first system call after zName. If zName==NULL
607 ** then return the name of the first system call. Return NULL if zName
608 ** is the last system call or if zName is not the name of a valid
611 static const char *unixNextSystemCall(sqlite3_vfs
*p
, const char *zName
){
616 for(i
=0; i
<ArraySize(aSyscall
)-1; i
++){
617 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) break;
620 for(i
++; i
<ArraySize(aSyscall
); i
++){
621 if( aSyscall
[i
].pCurrent
!=0 ) return aSyscall
[i
].zName
;
627 ** Do not accept any file descriptor less than this value, in order to avoid
628 ** opening database file using file descriptors that are commonly used for
629 ** standard input, output, and error.
631 #ifndef SQLITE_MINIMUM_FILE_DESCRIPTOR
632 # define SQLITE_MINIMUM_FILE_DESCRIPTOR 3
636 ** Invoke open(). Do so multiple times, until it either succeeds or
637 ** fails for some reason other than EINTR.
639 ** If the file creation mode "m" is 0 then set it to the default for
640 ** SQLite. The default is SQLITE_DEFAULT_FILE_PERMISSIONS (normally
641 ** 0644) as modified by the system umask. If m is not 0, then
642 ** make the file creation mode be exactly m ignoring the umask.
644 ** The m parameter will be non-zero only when creating -wal, -journal,
645 ** and -shm files. We want those files to have *exactly* the same
646 ** permissions as their original database, unadulterated by the umask.
647 ** In that way, if a database file is -rw-rw-rw or -rw-rw-r-, and a
648 ** transaction crashes and leaves behind hot journals, then any
649 ** process that is able to write to the database will also be able to
650 ** recover the hot journals.
652 static int robust_open(const char *z
, int f
, mode_t m
){
654 mode_t m2
= m
? m
: SQLITE_DEFAULT_FILE_PERMISSIONS
;
656 #if defined(O_CLOEXEC)
657 fd
= osOpen(z
,f
|O_CLOEXEC
,m2
);
662 if( errno
==EINTR
) continue;
665 if( fd
>=SQLITE_MINIMUM_FILE_DESCRIPTOR
) break;
667 sqlite3_log(SQLITE_WARNING
,
668 "attempt to open \"%s\" as file descriptor %d", z
, fd
);
670 if( osOpen("/dev/null", f
, m
)<0 ) break;
675 if( osFstat(fd
, &statbuf
)==0
676 && statbuf
.st_size
==0
677 && (statbuf
.st_mode
&0777)!=m
682 #if defined(FD_CLOEXEC) && (!defined(O_CLOEXEC) || O_CLOEXEC==0)
683 osFcntl(fd
, F_SETFD
, osFcntl(fd
, F_GETFD
, 0) | FD_CLOEXEC
);
690 ** Helper functions to obtain and relinquish the global mutex. The
691 ** global mutex is used to protect the unixInodeInfo and
692 ** vxworksFileId objects used by this file, all of which may be
693 ** shared by multiple threads.
695 ** Function unixMutexHeld() is used to assert() that the global mutex
696 ** is held when required. This function is only used as part of assert()
700 ** assert( unixMutexHeld() );
703 static sqlite3_mutex
*unixBigLock
= 0;
704 static void unixEnterMutex(void){
705 sqlite3_mutex_enter(unixBigLock
);
707 static void unixLeaveMutex(void){
708 sqlite3_mutex_leave(unixBigLock
);
711 static int unixMutexHeld(void) {
712 return sqlite3_mutex_held(unixBigLock
);
717 #ifdef SQLITE_HAVE_OS_TRACE
719 ** Helper function for printing out trace information from debugging
720 ** binaries. This returns the string representation of the supplied
721 ** integer lock-type.
723 static const char *azFileLock(int eFileLock
){
725 case NO_LOCK
: return "NONE";
726 case SHARED_LOCK
: return "SHARED";
727 case RESERVED_LOCK
: return "RESERVED";
728 case PENDING_LOCK
: return "PENDING";
729 case EXCLUSIVE_LOCK
: return "EXCLUSIVE";
735 #ifdef SQLITE_LOCK_TRACE
737 ** Print out information about all locking operations.
739 ** This routine is used for troubleshooting locks on multithreaded
740 ** platforms. Enable by compiling with the -DSQLITE_LOCK_TRACE
741 ** command-line option on the compiler. This code is normally
744 static int lockTrace(int fd
, int op
, struct flock
*p
){
745 char *zOpName
, *zType
;
750 }else if( op
==F_SETLK
){
753 s
= osFcntl(fd
, op
, p
);
754 sqlite3DebugPrintf("fcntl unknown %d %d %d\n", fd
, op
, s
);
757 if( p
->l_type
==F_RDLCK
){
759 }else if( p
->l_type
==F_WRLCK
){
761 }else if( p
->l_type
==F_UNLCK
){
766 assert( p
->l_whence
==SEEK_SET
);
767 s
= osFcntl(fd
, op
, p
);
769 sqlite3DebugPrintf("fcntl %d %d %s %s %d %d %d %d\n",
770 threadid
, fd
, zOpName
, zType
, (int)p
->l_start
, (int)p
->l_len
,
772 if( s
==(-1) && op
==F_SETLK
&& (p
->l_type
==F_RDLCK
|| p
->l_type
==F_WRLCK
) ){
775 osFcntl(fd
, F_GETLK
, &l2
);
776 if( l2
.l_type
==F_RDLCK
){
778 }else if( l2
.l_type
==F_WRLCK
){
780 }else if( l2
.l_type
==F_UNLCK
){
785 sqlite3DebugPrintf("fcntl-failure-reason: %s %d %d %d\n",
786 zType
, (int)l2
.l_start
, (int)l2
.l_len
, (int)l2
.l_pid
);
792 #define osFcntl lockTrace
793 #endif /* SQLITE_LOCK_TRACE */
796 ** Retry ftruncate() calls that fail due to EINTR
798 ** All calls to ftruncate() within this file should be made through
799 ** this wrapper. On the Android platform, bypassing the logic below
800 ** could lead to a corrupt database.
802 static int robust_ftruncate(int h
, sqlite3_int64 sz
){
805 /* On Android, ftruncate() always uses 32-bit offsets, even if
806 ** _FILE_OFFSET_BITS=64 is defined. This means it is unsafe to attempt to
807 ** truncate a file to any size larger than 2GiB. Silently ignore any
809 if( sz
>(sqlite3_int64
)0x7FFFFFFF ){
813 do{ rc
= osFtruncate(h
,sz
); }while( rc
<0 && errno
==EINTR
);
818 ** This routine translates a standard POSIX errno code into something
819 ** useful to the clients of the sqlite3 functions. Specifically, it is
820 ** intended to translate a variety of "try again" errors into SQLITE_BUSY
821 ** and a variety of "please close the file descriptor NOW" errors into
824 ** Errors during initialization of locks, or file system support for locks,
825 ** should handle ENOLCK, ENOTSUP, EOPNOTSUPP separately.
827 static int sqliteErrorFromPosixError(int posixError
, int sqliteIOErr
) {
828 assert( (sqliteIOErr
== SQLITE_IOERR_LOCK
) ||
829 (sqliteIOErr
== SQLITE_IOERR_UNLOCK
) ||
830 (sqliteIOErr
== SQLITE_IOERR_RDLOCK
) ||
831 (sqliteIOErr
== SQLITE_IOERR_CHECKRESERVEDLOCK
) );
832 switch (posixError
) {
839 /* random NFS retry error, unless during file system support
840 * introspection, in which it actually means what it says */
852 /******************************************************************************
853 ****************** Begin Unique File ID Utility Used By VxWorks ***************
855 ** On most versions of unix, we can get a unique ID for a file by concatenating
856 ** the device number and the inode number. But this does not work on VxWorks.
857 ** On VxWorks, a unique file id must be based on the canonical filename.
859 ** A pointer to an instance of the following structure can be used as a
860 ** unique file ID in VxWorks. Each instance of this structure contains
861 ** a copy of the canonical filename. There is also a reference count.
862 ** The structure is reclaimed when the number of pointers to it drops to
865 ** There are never very many files open at one time and lookups are not
866 ** a performance-critical path, so it is sufficient to put these
867 ** structures on a linked list.
869 struct vxworksFileId
{
870 struct vxworksFileId
*pNext
; /* Next in a list of them all */
871 int nRef
; /* Number of references to this one */
872 int nName
; /* Length of the zCanonicalName[] string */
873 char *zCanonicalName
; /* Canonical filename */
878 ** All unique filenames are held on a linked list headed by this
881 static struct vxworksFileId
*vxworksFileList
= 0;
884 ** Simplify a filename into its canonical form
885 ** by making the following changes:
887 ** * removing any trailing and duplicate /
888 ** * convert /./ into just /
889 ** * convert /A/../ where A is any simple name into just /
891 ** Changes are made in-place. Return the new name length.
893 ** The original filename is in z[0..n-1]. Return the number of
894 ** characters in the simplified name.
896 static int vxworksSimplifyName(char *z
, int n
){
898 while( n
>1 && z
[n
-1]=='/' ){ n
--; }
899 for(i
=j
=0; i
<n
; i
++){
901 if( z
[i
+1]=='/' ) continue;
902 if( z
[i
+1]=='.' && i
+2<n
&& z
[i
+2]=='/' ){
906 if( z
[i
+1]=='.' && i
+3<n
&& z
[i
+2]=='.' && z
[i
+3]=='/' ){
907 while( j
>0 && z
[j
-1]!='/' ){ j
--; }
920 ** Find a unique file ID for the given absolute pathname. Return
921 ** a pointer to the vxworksFileId object. This pointer is the unique
924 ** The nRef field of the vxworksFileId object is incremented before
925 ** the object is returned. A new vxworksFileId object is created
926 ** and added to the global list if necessary.
928 ** If a memory allocation error occurs, return NULL.
930 static struct vxworksFileId
*vxworksFindFileId(const char *zAbsoluteName
){
931 struct vxworksFileId
*pNew
; /* search key and new file ID */
932 struct vxworksFileId
*pCandidate
; /* For looping over existing file IDs */
933 int n
; /* Length of zAbsoluteName string */
935 assert( zAbsoluteName
[0]=='/' );
936 n
= (int)strlen(zAbsoluteName
);
937 pNew
= sqlite3_malloc64( sizeof(*pNew
) + (n
+1) );
938 if( pNew
==0 ) return 0;
939 pNew
->zCanonicalName
= (char*)&pNew
[1];
940 memcpy(pNew
->zCanonicalName
, zAbsoluteName
, n
+1);
941 n
= vxworksSimplifyName(pNew
->zCanonicalName
, n
);
943 /* Search for an existing entry that matching the canonical name.
944 ** If found, increment the reference count and return a pointer to
945 ** the existing file ID.
948 for(pCandidate
=vxworksFileList
; pCandidate
; pCandidate
=pCandidate
->pNext
){
949 if( pCandidate
->nName
==n
950 && memcmp(pCandidate
->zCanonicalName
, pNew
->zCanonicalName
, n
)==0
959 /* No match was found. We will make a new file ID */
962 pNew
->pNext
= vxworksFileList
;
963 vxworksFileList
= pNew
;
969 ** Decrement the reference count on a vxworksFileId object. Free
970 ** the object when the reference count reaches zero.
972 static void vxworksReleaseFileId(struct vxworksFileId
*pId
){
974 assert( pId
->nRef
>0 );
977 struct vxworksFileId
**pp
;
978 for(pp
=&vxworksFileList
; *pp
&& *pp
!=pId
; pp
= &((*pp
)->pNext
)){}
985 #endif /* OS_VXWORKS */
986 /*************** End of Unique File ID Utility Used By VxWorks ****************
987 ******************************************************************************/
990 /******************************************************************************
991 *************************** Posix Advisory Locking ****************************
993 ** POSIX advisory locks are broken by design. ANSI STD 1003.1 (1996)
994 ** section 6.5.2.2 lines 483 through 490 specify that when a process
995 ** sets or clears a lock, that operation overrides any prior locks set
996 ** by the same process. It does not explicitly say so, but this implies
997 ** that it overrides locks set by the same process using a different
998 ** file descriptor. Consider this test case:
1000 ** int fd1 = open("./file1", O_RDWR|O_CREAT, 0644);
1001 ** int fd2 = open("./file2", O_RDWR|O_CREAT, 0644);
1003 ** Suppose ./file1 and ./file2 are really the same file (because
1004 ** one is a hard or symbolic link to the other) then if you set
1005 ** an exclusive lock on fd1, then try to get an exclusive lock
1006 ** on fd2, it works. I would have expected the second lock to
1007 ** fail since there was already a lock on the file due to fd1.
1008 ** But not so. Since both locks came from the same process, the
1009 ** second overrides the first, even though they were on different
1010 ** file descriptors opened on different file names.
1012 ** This means that we cannot use POSIX locks to synchronize file access
1013 ** among competing threads of the same process. POSIX locks will work fine
1014 ** to synchronize access for threads in separate processes, but not
1015 ** threads within the same process.
1017 ** To work around the problem, SQLite has to manage file locks internally
1018 ** on its own. Whenever a new database is opened, we have to find the
1019 ** specific inode of the database file (the inode is determined by the
1020 ** st_dev and st_ino fields of the stat structure that fstat() fills in)
1021 ** and check for locks already existing on that inode. When locks are
1022 ** created or removed, we have to look at our own internal record of the
1023 ** locks to see if another thread has previously set a lock on that same
1026 ** (Aside: The use of inode numbers as unique IDs does not work on VxWorks.
1027 ** For VxWorks, we have to use the alternative unique ID system based on
1028 ** canonical filename and implemented in the previous division.)
1030 ** The sqlite3_file structure for POSIX is no longer just an integer file
1031 ** descriptor. It is now a structure that holds the integer file
1032 ** descriptor and a pointer to a structure that describes the internal
1033 ** locks on the corresponding inode. There is one locking structure
1034 ** per inode, so if the same inode is opened twice, both unixFile structures
1035 ** point to the same locking structure. The locking structure keeps
1036 ** a reference count (so we will know when to delete it) and a "cnt"
1037 ** field that tells us its internal lock status. cnt==0 means the
1038 ** file is unlocked. cnt==-1 means the file has an exclusive lock.
1039 ** cnt>0 means there are cnt shared locks on the file.
1041 ** Any attempt to lock or unlock a file first checks the locking
1042 ** structure. The fcntl() system call is only invoked to set a
1043 ** POSIX lock if the internal lock structure transitions between
1044 ** a locked and an unlocked state.
1046 ** But wait: there are yet more problems with POSIX advisory locks.
1048 ** If you close a file descriptor that points to a file that has locks,
1049 ** all locks on that file that are owned by the current process are
1050 ** released. To work around this problem, each unixInodeInfo object
1051 ** maintains a count of the number of pending locks on tha inode.
1052 ** When an attempt is made to close an unixFile, if there are
1053 ** other unixFile open on the same inode that are holding locks, the call
1054 ** to close() the file descriptor is deferred until all of the locks clear.
1055 ** The unixInodeInfo structure keeps a list of file descriptors that need to
1056 ** be closed and that list is walked (and cleared) when the last lock
1059 ** Yet another problem: LinuxThreads do not play well with posix locks.
1061 ** Many older versions of linux use the LinuxThreads library which is
1062 ** not posix compliant. Under LinuxThreads, a lock created by thread
1063 ** A cannot be modified or overridden by a different thread B.
1064 ** Only thread A can modify the lock. Locking behavior is correct
1065 ** if the appliation uses the newer Native Posix Thread Library (NPTL)
1066 ** on linux - with NPTL a lock created by thread A can override locks
1067 ** in thread B. But there is no way to know at compile-time which
1068 ** threading library is being used. So there is no way to know at
1069 ** compile-time whether or not thread A can override locks on thread B.
1070 ** One has to do a run-time check to discover the behavior of the
1073 ** SQLite used to support LinuxThreads. But support for LinuxThreads
1074 ** was dropped beginning with version 3.7.0. SQLite will still work with
1075 ** LinuxThreads provided that (1) there is no more than one connection
1076 ** per database file in the same process and (2) database connections
1077 ** do not move across threads.
1081 ** An instance of the following structure serves as the key used
1082 ** to locate a particular unixInodeInfo object.
1085 dev_t dev
; /* Device number */
1087 struct vxworksFileId
*pId
; /* Unique file ID for vxworks. */
1089 /* We are told that some versions of Android contain a bug that
1090 ** sizes ino_t at only 32-bits instead of 64-bits. (See
1091 ** https://android-review.googlesource.com/#/c/115351/3/dist/sqlite3.c)
1092 ** To work around this, always allocate 64-bits for the inode number.
1093 ** On small machines that only have 32-bit inodes, this wastes 4 bytes,
1094 ** but that should not be a big deal. */
1095 /* WAS: ino_t ino; */
1096 u64 ino
; /* Inode number */
1101 ** An instance of the following structure is allocated for each open
1102 ** inode. Or, on LinuxThreads, there is one of these structures for
1103 ** each inode opened by each thread.
1105 ** A single inode can have multiple file descriptors, so each unixFile
1106 ** structure contains a pointer to an instance of this object and this
1107 ** object keeps a count of the number of unixFile pointing to it.
1109 struct unixInodeInfo
{
1110 struct unixFileId fileId
; /* The lookup key */
1111 int nShared
; /* Number of SHARED locks held */
1112 unsigned char eFileLock
; /* One of SHARED_LOCK, RESERVED_LOCK etc. */
1113 unsigned char bProcessLock
; /* An exclusive process lock is held */
1114 int nRef
; /* Number of pointers to this structure */
1115 unixShmNode
*pShmNode
; /* Shared memory associated with this inode */
1116 int nLock
; /* Number of outstanding file locks */
1117 UnixUnusedFd
*pUnused
; /* Unused file descriptors to close */
1118 unixInodeInfo
*pNext
; /* List of all unixInodeInfo objects */
1119 unixInodeInfo
*pPrev
; /* .... doubly linked */
1120 #if SQLITE_ENABLE_LOCKING_STYLE
1121 unsigned long long sharedByte
; /* for AFP simulated shared lock */
1124 sem_t
*pSem
; /* Named POSIX semaphore */
1125 char aSemName
[MAX_PATHNAME
+2]; /* Name of that semaphore */
1130 ** A lists of all unixInodeInfo objects.
1132 static unixInodeInfo
*inodeList
= 0; /* All unixInodeInfo objects */
1133 static unsigned int nUnusedFd
= 0; /* Total unused file descriptors */
1137 ** This function - unixLogErrorAtLine(), is only ever called via the macro
1140 ** It is invoked after an error occurs in an OS function and errno has been
1141 ** set. It logs a message using sqlite3_log() containing the current value of
1142 ** errno and, if possible, the human-readable equivalent from strerror() or
1145 ** The first argument passed to the macro should be the error code that
1146 ** will be returned to SQLite (e.g. SQLITE_IOERR_DELETE, SQLITE_CANTOPEN).
1147 ** The two subsequent arguments should be the name of the OS function that
1148 ** failed (e.g. "unlink", "open") and the associated file-system path,
1151 #define unixLogError(a,b,c) unixLogErrorAtLine(a,b,c,__LINE__)
1152 static int unixLogErrorAtLine(
1153 int errcode
, /* SQLite error code */
1154 const char *zFunc
, /* Name of OS function that failed */
1155 const char *zPath
, /* File path associated with error */
1156 int iLine
/* Source line number where error occurred */
1158 char *zErr
; /* Message from strerror() or equivalent */
1159 int iErrno
= errno
; /* Saved syscall error number */
1161 /* If this is not a threadsafe build (SQLITE_THREADSAFE==0), then use
1162 ** the strerror() function to obtain the human-readable error message
1163 ** equivalent to errno. Otherwise, use strerror_r().
1165 #if SQLITE_THREADSAFE && defined(HAVE_STRERROR_R)
1167 memset(aErr
, 0, sizeof(aErr
));
1170 /* If STRERROR_R_CHAR_P (set by autoconf scripts) or __USE_GNU is defined,
1171 ** assume that the system provides the GNU version of strerror_r() that
1172 ** returns a pointer to a buffer containing the error message. That pointer
1173 ** may point to aErr[], or it may point to some static storage somewhere.
1174 ** Otherwise, assume that the system provides the POSIX version of
1175 ** strerror_r(), which always writes an error message into aErr[].
1177 ** If the code incorrectly assumes that it is the POSIX version that is
1178 ** available, the error message will often be an empty string. Not a
1179 ** huge problem. Incorrectly concluding that the GNU version is available
1180 ** could lead to a segfault though.
1182 #if defined(STRERROR_R_CHAR_P) || defined(__USE_GNU)
1185 strerror_r(iErrno
, aErr
, sizeof(aErr
)-1);
1187 #elif SQLITE_THREADSAFE
1188 /* This is a threadsafe build, but strerror_r() is not available. */
1191 /* Non-threadsafe build, use strerror(). */
1192 zErr
= strerror(iErrno
);
1195 if( zPath
==0 ) zPath
= "";
1196 sqlite3_log(errcode
,
1197 "os_unix.c:%d: (%d) %s(%s) - %s",
1198 iLine
, iErrno
, zFunc
, zPath
, zErr
1205 ** Close a file descriptor.
1207 ** We assume that close() almost always works, since it is only in a
1208 ** very sick application or on a very sick platform that it might fail.
1209 ** If it does fail, simply leak the file descriptor, but do log the
1212 ** Note that it is not safe to retry close() after EINTR since the
1213 ** file descriptor might have already been reused by another thread.
1214 ** So we don't even try to recover from an EINTR. Just log the error
1217 static void robust_close(unixFile
*pFile
, int h
, int lineno
){
1219 unixLogErrorAtLine(SQLITE_IOERR_CLOSE
, "close",
1220 pFile
? pFile
->zPath
: 0, lineno
);
1225 ** Set the pFile->lastErrno. Do this in a subroutine as that provides
1226 ** a convenient place to set a breakpoint.
1228 static void storeLastErrno(unixFile
*pFile
, int error
){
1229 pFile
->lastErrno
= error
;
1233 ** Close all file descriptors accumuated in the unixInodeInfo->pUnused list.
1235 static void closePendingFds(unixFile
*pFile
){
1236 unixInodeInfo
*pInode
= pFile
->pInode
;
1238 UnixUnusedFd
*pNext
;
1239 for(p
=pInode
->pUnused
; p
; p
=pNext
){
1241 robust_close(pFile
, p
->fd
, __LINE__
);
1245 pInode
->pUnused
= 0;
1249 ** Release a unixInodeInfo structure previously allocated by findInodeInfo().
1251 ** The mutex entered using the unixEnterMutex() function must be held
1252 ** when this function is called.
1254 static void releaseInodeInfo(unixFile
*pFile
){
1255 unixInodeInfo
*pInode
= pFile
->pInode
;
1256 assert( unixMutexHeld() );
1257 if( ALWAYS(pInode
) ){
1259 if( pInode
->nRef
==0 ){
1260 assert( pInode
->pShmNode
==0 );
1261 closePendingFds(pFile
);
1262 if( pInode
->pPrev
){
1263 assert( pInode
->pPrev
->pNext
==pInode
);
1264 pInode
->pPrev
->pNext
= pInode
->pNext
;
1266 assert( inodeList
==pInode
);
1267 inodeList
= pInode
->pNext
;
1269 if( pInode
->pNext
){
1270 assert( pInode
->pNext
->pPrev
==pInode
);
1271 pInode
->pNext
->pPrev
= pInode
->pPrev
;
1273 sqlite3_free(pInode
);
1276 assert( inodeList
!=0 || nUnusedFd
==0 );
1280 ** Given a file descriptor, locate the unixInodeInfo object that
1281 ** describes that file descriptor. Create a new one if necessary. The
1282 ** return value might be uninitialized if an error occurs.
1284 ** The mutex entered using the unixEnterMutex() function must be held
1285 ** when this function is called.
1287 ** Return an appropriate error code.
1289 static int findInodeInfo(
1290 unixFile
*pFile
, /* Unix file with file desc used in the key */
1291 unixInodeInfo
**ppInode
/* Return the unixInodeInfo object here */
1293 int rc
; /* System call return code */
1294 int fd
; /* The file descriptor for pFile */
1295 struct unixFileId fileId
; /* Lookup key for the unixInodeInfo */
1296 struct stat statbuf
; /* Low-level file information */
1297 unixInodeInfo
*pInode
= 0; /* Candidate unixInodeInfo object */
1299 assert( unixMutexHeld() );
1301 /* Get low-level information about the file that we can used to
1302 ** create a unique name for the file.
1305 rc
= osFstat(fd
, &statbuf
);
1307 storeLastErrno(pFile
, errno
);
1308 #if defined(EOVERFLOW) && defined(SQLITE_DISABLE_LFS)
1309 if( pFile
->lastErrno
==EOVERFLOW
) return SQLITE_NOLFS
;
1311 return SQLITE_IOERR
;
1315 /* On OS X on an msdos filesystem, the inode number is reported
1316 ** incorrectly for zero-size files. See ticket #3260. To work
1317 ** around this problem (we consider it a bug in OS X, not SQLite)
1318 ** we always increase the file size to 1 by writing a single byte
1319 ** prior to accessing the inode number. The one byte written is
1320 ** an ASCII 'S' character which also happens to be the first byte
1321 ** in the header of every SQLite database. In this way, if there
1322 ** is a race condition such that another thread has already populated
1323 ** the first page of the database, no damage is done.
1325 if( statbuf
.st_size
==0 && (pFile
->fsFlags
& SQLITE_FSFLAGS_IS_MSDOS
)!=0 ){
1326 do{ rc
= osWrite(fd
, "S", 1); }while( rc
<0 && errno
==EINTR
);
1328 storeLastErrno(pFile
, errno
);
1329 return SQLITE_IOERR
;
1331 rc
= osFstat(fd
, &statbuf
);
1333 storeLastErrno(pFile
, errno
);
1334 return SQLITE_IOERR
;
1339 memset(&fileId
, 0, sizeof(fileId
));
1340 fileId
.dev
= statbuf
.st_dev
;
1342 fileId
.pId
= pFile
->pId
;
1344 fileId
.ino
= (u64
)statbuf
.st_ino
;
1346 assert( inodeList
!=0 || nUnusedFd
==0 );
1348 while( pInode
&& memcmp(&fileId
, &pInode
->fileId
, sizeof(fileId
)) ){
1349 pInode
= pInode
->pNext
;
1352 pInode
= sqlite3_malloc64( sizeof(*pInode
) );
1354 return SQLITE_NOMEM_BKPT
;
1356 memset(pInode
, 0, sizeof(*pInode
));
1357 memcpy(&pInode
->fileId
, &fileId
, sizeof(fileId
));
1359 pInode
->pNext
= inodeList
;
1361 if( inodeList
) inodeList
->pPrev
= pInode
;
1371 ** Return TRUE if pFile has been renamed or unlinked since it was first opened.
1373 static int fileHasMoved(unixFile
*pFile
){
1375 return pFile
->pInode
!=0 && pFile
->pId
!=pFile
->pInode
->fileId
.pId
;
1378 return pFile
->pInode
!=0 &&
1379 (osStat(pFile
->zPath
, &buf
)!=0
1380 || (u64
)buf
.st_ino
!=pFile
->pInode
->fileId
.ino
);
1386 ** Check a unixFile that is a database. Verify the following:
1388 ** (1) There is exactly one hard link on the file
1389 ** (2) The file is not a symbolic link
1390 ** (3) The file has not been renamed or unlinked
1392 ** Issue sqlite3_log(SQLITE_WARNING,...) messages if anything is not right.
1394 static void verifyDbFile(unixFile
*pFile
){
1398 /* These verifications occurs for the main database only */
1399 if( pFile
->ctrlFlags
& UNIXFILE_NOLOCK
) return;
1401 rc
= osFstat(pFile
->h
, &buf
);
1403 sqlite3_log(SQLITE_WARNING
, "cannot fstat db file %s", pFile
->zPath
);
1406 if( buf
.st_nlink
==0 ){
1407 sqlite3_log(SQLITE_WARNING
, "file unlinked while open: %s", pFile
->zPath
);
1410 if( buf
.st_nlink
>1 ){
1411 sqlite3_log(SQLITE_WARNING
, "multiple links to file: %s", pFile
->zPath
);
1414 if( fileHasMoved(pFile
) ){
1415 sqlite3_log(SQLITE_WARNING
, "file renamed while open: %s", pFile
->zPath
);
1422 ** This routine checks if there is a RESERVED lock held on the specified
1423 ** file by this or any other process. If such a lock is held, set *pResOut
1424 ** to a non-zero value otherwise *pResOut is set to zero. The return value
1425 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
1427 static int unixCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
1430 unixFile
*pFile
= (unixFile
*)id
;
1432 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
1435 assert( pFile
->eFileLock
<=SHARED_LOCK
);
1436 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
1438 /* Check if a thread in this process holds such a lock */
1439 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
1443 /* Otherwise see if some other process holds it.
1446 if( !reserved
&& !pFile
->pInode
->bProcessLock
){
1448 lock
.l_whence
= SEEK_SET
;
1449 lock
.l_start
= RESERVED_BYTE
;
1451 lock
.l_type
= F_WRLCK
;
1452 if( osFcntl(pFile
->h
, F_GETLK
, &lock
) ){
1453 rc
= SQLITE_IOERR_CHECKRESERVEDLOCK
;
1454 storeLastErrno(pFile
, errno
);
1455 } else if( lock
.l_type
!=F_UNLCK
){
1462 OSTRACE(("TEST WR-LOCK %d %d %d (unix)\n", pFile
->h
, rc
, reserved
));
1464 *pResOut
= reserved
;
1469 ** Attempt to set a system-lock on the file pFile. The lock is
1470 ** described by pLock.
1472 ** If the pFile was opened read/write from unix-excl, then the only lock
1473 ** ever obtained is an exclusive lock, and it is obtained exactly once
1474 ** the first time any lock is attempted. All subsequent system locking
1475 ** operations become no-ops. Locking operations still happen internally,
1476 ** in order to coordinate access between separate database connections
1477 ** within this process, but all of that is handled in memory and the
1478 ** operating system does not participate.
1480 ** This function is a pass-through to fcntl(F_SETLK) if pFile is using
1481 ** any VFS other than "unix-excl" or if pFile is opened on "unix-excl"
1482 ** and is read-only.
1484 ** Zero is returned if the call completes successfully, or -1 if a call
1485 ** to fcntl() fails. In this case, errno is set appropriately (by fcntl()).
1487 static int unixFileLock(unixFile
*pFile
, struct flock
*pLock
){
1489 unixInodeInfo
*pInode
= pFile
->pInode
;
1490 assert( unixMutexHeld() );
1491 assert( pInode
!=0 );
1492 if( (pFile
->ctrlFlags
& (UNIXFILE_EXCL
|UNIXFILE_RDONLY
))==UNIXFILE_EXCL
){
1493 if( pInode
->bProcessLock
==0 ){
1495 assert( pInode
->nLock
==0 );
1496 lock
.l_whence
= SEEK_SET
;
1497 lock
.l_start
= SHARED_FIRST
;
1498 lock
.l_len
= SHARED_SIZE
;
1499 lock
.l_type
= F_WRLCK
;
1500 rc
= osFcntl(pFile
->h
, F_SETLK
, &lock
);
1501 if( rc
<0 ) return rc
;
1502 pInode
->bProcessLock
= 1;
1508 rc
= osFcntl(pFile
->h
, F_SETLK
, pLock
);
1514 ** Lock the file with the lock specified by parameter eFileLock - one
1515 ** of the following:
1518 ** (2) RESERVED_LOCK
1520 ** (4) EXCLUSIVE_LOCK
1522 ** Sometimes when requesting one lock state, additional lock states
1523 ** are inserted in between. The locking might fail on one of the later
1524 ** transitions leaving the lock state different from what it started but
1525 ** still short of its goal. The following chart shows the allowed
1526 ** transitions and the inserted intermediate states:
1528 ** UNLOCKED -> SHARED
1529 ** SHARED -> RESERVED
1530 ** SHARED -> (PENDING) -> EXCLUSIVE
1531 ** RESERVED -> (PENDING) -> EXCLUSIVE
1532 ** PENDING -> EXCLUSIVE
1534 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
1535 ** routine to lower a locking level.
1537 static int unixLock(sqlite3_file
*id
, int eFileLock
){
1538 /* The following describes the implementation of the various locks and
1539 ** lock transitions in terms of the POSIX advisory shared and exclusive
1540 ** lock primitives (called read-locks and write-locks below, to avoid
1541 ** confusion with SQLite lock names). The algorithms are complicated
1542 ** slightly in order to be compatible with Windows95 systems simultaneously
1543 ** accessing the same database file, in case that is ever required.
1545 ** Symbols defined in os.h indentify the 'pending byte' and the 'reserved
1546 ** byte', each single bytes at well known offsets, and the 'shared byte
1547 ** range', a range of 510 bytes at a well known offset.
1549 ** To obtain a SHARED lock, a read-lock is obtained on the 'pending
1550 ** byte'. If this is successful, 'shared byte range' is read-locked
1551 ** and the lock on the 'pending byte' released. (Legacy note: When
1552 ** SQLite was first developed, Windows95 systems were still very common,
1553 ** and Widnows95 lacks a shared-lock capability. So on Windows95, a
1554 ** single randomly selected by from the 'shared byte range' is locked.
1555 ** Windows95 is now pretty much extinct, but this work-around for the
1556 ** lack of shared-locks on Windows95 lives on, for backwards
1559 ** A process may only obtain a RESERVED lock after it has a SHARED lock.
1560 ** A RESERVED lock is implemented by grabbing a write-lock on the
1563 ** A process may only obtain a PENDING lock after it has obtained a
1564 ** SHARED lock. A PENDING lock is implemented by obtaining a write-lock
1565 ** on the 'pending byte'. This ensures that no new SHARED locks can be
1566 ** obtained, but existing SHARED locks are allowed to persist. A process
1567 ** does not have to obtain a RESERVED lock on the way to a PENDING lock.
1568 ** This property is used by the algorithm for rolling back a journal file
1571 ** An EXCLUSIVE lock, obtained after a PENDING lock is held, is
1572 ** implemented by obtaining a write-lock on the entire 'shared byte
1573 ** range'. Since all other locks require a read-lock on one of the bytes
1574 ** within this range, this ensures that no other locks are held on the
1578 unixFile
*pFile
= (unixFile
*)id
;
1579 unixInodeInfo
*pInode
;
1584 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (unix)\n", pFile
->h
,
1585 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
1586 azFileLock(pFile
->pInode
->eFileLock
), pFile
->pInode
->nShared
,
1589 /* If there is already a lock of this type or more restrictive on the
1590 ** unixFile, do nothing. Don't use the end_lock: exit path, as
1591 ** unixEnterMutex() hasn't been called yet.
1593 if( pFile
->eFileLock
>=eFileLock
){
1594 OSTRACE(("LOCK %d %s ok (already held) (unix)\n", pFile
->h
,
1595 azFileLock(eFileLock
)));
1599 /* Make sure the locking sequence is correct.
1600 ** (1) We never move from unlocked to anything higher than shared lock.
1601 ** (2) SQLite never explicitly requests a pendig lock.
1602 ** (3) A shared lock is always held when a reserve lock is requested.
1604 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
1605 assert( eFileLock
!=PENDING_LOCK
);
1606 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
1608 /* This mutex is needed because pFile->pInode is shared across threads
1611 pInode
= pFile
->pInode
;
1613 /* If some thread using this PID has a lock via a different unixFile*
1614 ** handle that precludes the requested lock, return BUSY.
1616 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
1617 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
1623 /* If a SHARED lock is requested, and some thread using this PID already
1624 ** has a SHARED or RESERVED lock, then increment reference counts and
1625 ** return SQLITE_OK.
1627 if( eFileLock
==SHARED_LOCK
&&
1628 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
1629 assert( eFileLock
==SHARED_LOCK
);
1630 assert( pFile
->eFileLock
==0 );
1631 assert( pInode
->nShared
>0 );
1632 pFile
->eFileLock
= SHARED_LOCK
;
1639 /* A PENDING lock is needed before acquiring a SHARED lock and before
1640 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
1644 lock
.l_whence
= SEEK_SET
;
1645 if( eFileLock
==SHARED_LOCK
1646 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
1648 lock
.l_type
= (eFileLock
==SHARED_LOCK
?F_RDLCK
:F_WRLCK
);
1649 lock
.l_start
= PENDING_BYTE
;
1650 if( unixFileLock(pFile
, &lock
) ){
1652 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1653 if( rc
!=SQLITE_BUSY
){
1654 storeLastErrno(pFile
, tErrno
);
1661 /* If control gets to this point, then actually go ahead and make
1662 ** operating system calls for the specified lock.
1664 if( eFileLock
==SHARED_LOCK
){
1665 assert( pInode
->nShared
==0 );
1666 assert( pInode
->eFileLock
==0 );
1667 assert( rc
==SQLITE_OK
);
1669 /* Now get the read-lock */
1670 lock
.l_start
= SHARED_FIRST
;
1671 lock
.l_len
= SHARED_SIZE
;
1672 if( unixFileLock(pFile
, &lock
) ){
1674 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1677 /* Drop the temporary PENDING lock */
1678 lock
.l_start
= PENDING_BYTE
;
1680 lock
.l_type
= F_UNLCK
;
1681 if( unixFileLock(pFile
, &lock
) && rc
==SQLITE_OK
){
1682 /* This could happen with a network mount */
1684 rc
= SQLITE_IOERR_UNLOCK
;
1688 if( rc
!=SQLITE_BUSY
){
1689 storeLastErrno(pFile
, tErrno
);
1693 pFile
->eFileLock
= SHARED_LOCK
;
1695 pInode
->nShared
= 1;
1697 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
1698 /* We are trying for an exclusive lock but another thread in this
1699 ** same process is still holding a shared lock. */
1702 /* The request was for a RESERVED or EXCLUSIVE lock. It is
1703 ** assumed that there is a SHARED or greater lock on the file
1706 assert( 0!=pFile
->eFileLock
);
1707 lock
.l_type
= F_WRLCK
;
1709 assert( eFileLock
==RESERVED_LOCK
|| eFileLock
==EXCLUSIVE_LOCK
);
1710 if( eFileLock
==RESERVED_LOCK
){
1711 lock
.l_start
= RESERVED_BYTE
;
1714 lock
.l_start
= SHARED_FIRST
;
1715 lock
.l_len
= SHARED_SIZE
;
1718 if( unixFileLock(pFile
, &lock
) ){
1720 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1721 if( rc
!=SQLITE_BUSY
){
1722 storeLastErrno(pFile
, tErrno
);
1729 /* Set up the transaction-counter change checking flags when
1730 ** transitioning from a SHARED to a RESERVED lock. The change
1731 ** from SHARED to RESERVED marks the beginning of a normal
1732 ** write operation (not a hot journal rollback).
1735 && pFile
->eFileLock
<=SHARED_LOCK
1736 && eFileLock
==RESERVED_LOCK
1738 pFile
->transCntrChng
= 0;
1739 pFile
->dbUpdate
= 0;
1740 pFile
->inNormalWrite
= 1;
1745 if( rc
==SQLITE_OK
){
1746 pFile
->eFileLock
= eFileLock
;
1747 pInode
->eFileLock
= eFileLock
;
1748 }else if( eFileLock
==EXCLUSIVE_LOCK
){
1749 pFile
->eFileLock
= PENDING_LOCK
;
1750 pInode
->eFileLock
= PENDING_LOCK
;
1755 OSTRACE(("LOCK %d %s %s (unix)\n", pFile
->h
, azFileLock(eFileLock
),
1756 rc
==SQLITE_OK
? "ok" : "failed"));
1761 ** Add the file descriptor used by file handle pFile to the corresponding
1764 static void setPendingFd(unixFile
*pFile
){
1765 unixInodeInfo
*pInode
= pFile
->pInode
;
1766 UnixUnusedFd
*p
= pFile
->pPreallocatedUnused
;
1767 p
->pNext
= pInode
->pUnused
;
1768 pInode
->pUnused
= p
;
1770 pFile
->pPreallocatedUnused
= 0;
1775 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1776 ** must be either NO_LOCK or SHARED_LOCK.
1778 ** If the locking level of the file descriptor is already at or below
1779 ** the requested locking level, this routine is a no-op.
1781 ** If handleNFSUnlock is true, then on downgrading an EXCLUSIVE_LOCK to SHARED
1782 ** the byte range is divided into 2 parts and the first part is unlocked then
1783 ** set to a read lock, then the other part is simply unlocked. This works
1784 ** around a bug in BSD NFS lockd (also seen on MacOSX 10.3+) that fails to
1785 ** remove the write lock on a region when a read lock is set.
1787 static int posixUnlock(sqlite3_file
*id
, int eFileLock
, int handleNFSUnlock
){
1788 unixFile
*pFile
= (unixFile
*)id
;
1789 unixInodeInfo
*pInode
;
1794 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (unix)\n", pFile
->h
, eFileLock
,
1795 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
1798 assert( eFileLock
<=SHARED_LOCK
);
1799 if( pFile
->eFileLock
<=eFileLock
){
1803 pInode
= pFile
->pInode
;
1804 assert( pInode
->nShared
!=0 );
1805 if( pFile
->eFileLock
>SHARED_LOCK
){
1806 assert( pInode
->eFileLock
==pFile
->eFileLock
);
1809 /* When reducing a lock such that other processes can start
1810 ** reading the database file again, make sure that the
1811 ** transaction counter was updated if any part of the database
1812 ** file changed. If the transaction counter is not updated,
1813 ** other connections to the same file might not realize that
1814 ** the file has changed and hence might not know to flush their
1815 ** cache. The use of a stale cache can lead to database corruption.
1817 pFile
->inNormalWrite
= 0;
1820 /* downgrading to a shared lock on NFS involves clearing the write lock
1821 ** before establishing the readlock - to avoid a race condition we downgrade
1822 ** the lock in 2 blocks, so that part of the range will be covered by a
1823 ** write lock until the rest is covered by a read lock:
1829 if( eFileLock
==SHARED_LOCK
){
1830 #if !defined(__APPLE__) || !SQLITE_ENABLE_LOCKING_STYLE
1831 (void)handleNFSUnlock
;
1832 assert( handleNFSUnlock
==0 );
1834 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
1835 if( handleNFSUnlock
){
1836 int tErrno
; /* Error code from system call errors */
1837 off_t divSize
= SHARED_SIZE
- 1;
1839 lock
.l_type
= F_UNLCK
;
1840 lock
.l_whence
= SEEK_SET
;
1841 lock
.l_start
= SHARED_FIRST
;
1842 lock
.l_len
= divSize
;
1843 if( unixFileLock(pFile
, &lock
)==(-1) ){
1845 rc
= SQLITE_IOERR_UNLOCK
;
1846 storeLastErrno(pFile
, tErrno
);
1849 lock
.l_type
= F_RDLCK
;
1850 lock
.l_whence
= SEEK_SET
;
1851 lock
.l_start
= SHARED_FIRST
;
1852 lock
.l_len
= divSize
;
1853 if( unixFileLock(pFile
, &lock
)==(-1) ){
1855 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_RDLOCK
);
1856 if( IS_LOCK_ERROR(rc
) ){
1857 storeLastErrno(pFile
, tErrno
);
1861 lock
.l_type
= F_UNLCK
;
1862 lock
.l_whence
= SEEK_SET
;
1863 lock
.l_start
= SHARED_FIRST
+divSize
;
1864 lock
.l_len
= SHARED_SIZE
-divSize
;
1865 if( unixFileLock(pFile
, &lock
)==(-1) ){
1867 rc
= SQLITE_IOERR_UNLOCK
;
1868 storeLastErrno(pFile
, tErrno
);
1872 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
1874 lock
.l_type
= F_RDLCK
;
1875 lock
.l_whence
= SEEK_SET
;
1876 lock
.l_start
= SHARED_FIRST
;
1877 lock
.l_len
= SHARED_SIZE
;
1878 if( unixFileLock(pFile
, &lock
) ){
1879 /* In theory, the call to unixFileLock() cannot fail because another
1880 ** process is holding an incompatible lock. If it does, this
1881 ** indicates that the other process is not following the locking
1882 ** protocol. If this happens, return SQLITE_IOERR_RDLOCK. Returning
1883 ** SQLITE_BUSY would confuse the upper layer (in practice it causes
1884 ** an assert to fail). */
1885 rc
= SQLITE_IOERR_RDLOCK
;
1886 storeLastErrno(pFile
, errno
);
1891 lock
.l_type
= F_UNLCK
;
1892 lock
.l_whence
= SEEK_SET
;
1893 lock
.l_start
= PENDING_BYTE
;
1894 lock
.l_len
= 2L; assert( PENDING_BYTE
+1==RESERVED_BYTE
);
1895 if( unixFileLock(pFile
, &lock
)==0 ){
1896 pInode
->eFileLock
= SHARED_LOCK
;
1898 rc
= SQLITE_IOERR_UNLOCK
;
1899 storeLastErrno(pFile
, errno
);
1903 if( eFileLock
==NO_LOCK
){
1904 /* Decrement the shared lock counter. Release the lock using an
1905 ** OS call only when all threads in this same process have released
1909 if( pInode
->nShared
==0 ){
1910 lock
.l_type
= F_UNLCK
;
1911 lock
.l_whence
= SEEK_SET
;
1912 lock
.l_start
= lock
.l_len
= 0L;
1913 if( unixFileLock(pFile
, &lock
)==0 ){
1914 pInode
->eFileLock
= NO_LOCK
;
1916 rc
= SQLITE_IOERR_UNLOCK
;
1917 storeLastErrno(pFile
, errno
);
1918 pInode
->eFileLock
= NO_LOCK
;
1919 pFile
->eFileLock
= NO_LOCK
;
1923 /* Decrement the count of locks against this same file. When the
1924 ** count reaches zero, close any other file descriptors whose close
1925 ** was deferred because of outstanding locks.
1928 assert( pInode
->nLock
>=0 );
1929 if( pInode
->nLock
==0 ){
1930 closePendingFds(pFile
);
1936 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
1941 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1942 ** must be either NO_LOCK or SHARED_LOCK.
1944 ** If the locking level of the file descriptor is already at or below
1945 ** the requested locking level, this routine is a no-op.
1947 static int unixUnlock(sqlite3_file
*id
, int eFileLock
){
1948 #if SQLITE_MAX_MMAP_SIZE>0
1949 assert( eFileLock
==SHARED_LOCK
|| ((unixFile
*)id
)->nFetchOut
==0 );
1951 return posixUnlock(id
, eFileLock
, 0);
1954 #if SQLITE_MAX_MMAP_SIZE>0
1955 static int unixMapfile(unixFile
*pFd
, i64 nByte
);
1956 static void unixUnmapfile(unixFile
*pFd
);
1960 ** This function performs the parts of the "close file" operation
1961 ** common to all locking schemes. It closes the directory and file
1962 ** handles, if they are valid, and sets all fields of the unixFile
1965 ** It is *not* necessary to hold the mutex when this routine is called,
1966 ** even on VxWorks. A mutex will be acquired on VxWorks by the
1967 ** vxworksReleaseFileId() routine.
1969 static int closeUnixFile(sqlite3_file
*id
){
1970 unixFile
*pFile
= (unixFile
*)id
;
1971 #if SQLITE_MAX_MMAP_SIZE>0
1972 unixUnmapfile(pFile
);
1975 robust_close(pFile
, pFile
->h
, __LINE__
);
1980 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
1981 osUnlink(pFile
->pId
->zCanonicalName
);
1983 vxworksReleaseFileId(pFile
->pId
);
1987 #ifdef SQLITE_UNLINK_AFTER_CLOSE
1988 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
1989 osUnlink(pFile
->zPath
);
1990 sqlite3_free(*(char**)&pFile
->zPath
);
1994 OSTRACE(("CLOSE %-3d\n", pFile
->h
));
1996 sqlite3_free(pFile
->pPreallocatedUnused
);
1997 memset(pFile
, 0, sizeof(unixFile
));
2004 static int unixClose(sqlite3_file
*id
){
2006 unixFile
*pFile
= (unixFile
*)id
;
2007 verifyDbFile(pFile
);
2008 unixUnlock(id
, NO_LOCK
);
2011 /* unixFile.pInode is always valid here. Otherwise, a different close
2012 ** routine (e.g. nolockClose()) would be called instead.
2014 assert( pFile
->pInode
->nLock
>0 || pFile
->pInode
->bProcessLock
==0 );
2015 if( ALWAYS(pFile
->pInode
) && pFile
->pInode
->nLock
){
2016 /* If there are outstanding locks, do not actually close the file just
2017 ** yet because that would clear those locks. Instead, add the file
2018 ** descriptor to pInode->pUnused list. It will be automatically closed
2019 ** when the last lock is cleared.
2021 setPendingFd(pFile
);
2023 releaseInodeInfo(pFile
);
2024 rc
= closeUnixFile(id
);
2029 /************** End of the posix advisory lock implementation *****************
2030 ******************************************************************************/
2032 /******************************************************************************
2033 ****************************** No-op Locking **********************************
2035 ** Of the various locking implementations available, this is by far the
2036 ** simplest: locking is ignored. No attempt is made to lock the database
2037 ** file for reading or writing.
2039 ** This locking mode is appropriate for use on read-only databases
2040 ** (ex: databases that are burned into CD-ROM, for example.) It can
2041 ** also be used if the application employs some external mechanism to
2042 ** prevent simultaneous access of the same database by two or more
2043 ** database connections. But there is a serious risk of database
2044 ** corruption if this locking mode is used in situations where multiple
2045 ** database connections are accessing the same database file at the same
2046 ** time and one or more of those connections are writing.
2049 static int nolockCheckReservedLock(sqlite3_file
*NotUsed
, int *pResOut
){
2050 UNUSED_PARAMETER(NotUsed
);
2054 static int nolockLock(sqlite3_file
*NotUsed
, int NotUsed2
){
2055 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2058 static int nolockUnlock(sqlite3_file
*NotUsed
, int NotUsed2
){
2059 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2066 static int nolockClose(sqlite3_file
*id
) {
2067 return closeUnixFile(id
);
2070 /******************* End of the no-op lock implementation *********************
2071 ******************************************************************************/
2073 /******************************************************************************
2074 ************************* Begin dot-file Locking ******************************
2076 ** The dotfile locking implementation uses the existence of separate lock
2077 ** files (really a directory) to control access to the database. This works
2078 ** on just about every filesystem imaginable. But there are serious downsides:
2080 ** (1) There is zero concurrency. A single reader blocks all other
2081 ** connections from reading or writing the database.
2083 ** (2) An application crash or power loss can leave stale lock files
2084 ** sitting around that need to be cleared manually.
2086 ** Nevertheless, a dotlock is an appropriate locking mode for use if no
2087 ** other locking strategy is available.
2089 ** Dotfile locking works by creating a subdirectory in the same directory as
2090 ** the database and with the same name but with a ".lock" extension added.
2091 ** The existence of a lock directory implies an EXCLUSIVE lock. All other
2092 ** lock types (SHARED, RESERVED, PENDING) are mapped into EXCLUSIVE.
2096 ** The file suffix added to the data base filename in order to create the
2099 #define DOTLOCK_SUFFIX ".lock"
2102 ** This routine checks if there is a RESERVED lock held on the specified
2103 ** file by this or any other process. If such a lock is held, set *pResOut
2104 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2105 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2107 ** In dotfile locking, either a lock exists or it does not. So in this
2108 ** variation of CheckReservedLock(), *pResOut is set to true if any lock
2109 ** is held on the file and false if the file is unlocked.
2111 static int dotlockCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2114 unixFile
*pFile
= (unixFile
*)id
;
2116 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2119 reserved
= osAccess((const char*)pFile
->lockingContext
, 0)==0;
2120 OSTRACE(("TEST WR-LOCK %d %d %d (dotlock)\n", pFile
->h
, rc
, reserved
));
2121 *pResOut
= reserved
;
2126 ** Lock the file with the lock specified by parameter eFileLock - one
2127 ** of the following:
2130 ** (2) RESERVED_LOCK
2132 ** (4) EXCLUSIVE_LOCK
2134 ** Sometimes when requesting one lock state, additional lock states
2135 ** are inserted in between. The locking might fail on one of the later
2136 ** transitions leaving the lock state different from what it started but
2137 ** still short of its goal. The following chart shows the allowed
2138 ** transitions and the inserted intermediate states:
2140 ** UNLOCKED -> SHARED
2141 ** SHARED -> RESERVED
2142 ** SHARED -> (PENDING) -> EXCLUSIVE
2143 ** RESERVED -> (PENDING) -> EXCLUSIVE
2144 ** PENDING -> EXCLUSIVE
2146 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2147 ** routine to lower a locking level.
2149 ** With dotfile locking, we really only support state (4): EXCLUSIVE.
2150 ** But we track the other locking levels internally.
2152 static int dotlockLock(sqlite3_file
*id
, int eFileLock
) {
2153 unixFile
*pFile
= (unixFile
*)id
;
2154 char *zLockFile
= (char *)pFile
->lockingContext
;
2158 /* If we have any lock, then the lock file already exists. All we have
2159 ** to do is adjust our internal record of the lock level.
2161 if( pFile
->eFileLock
> NO_LOCK
){
2162 pFile
->eFileLock
= eFileLock
;
2163 /* Always update the timestamp on the old file */
2165 utime(zLockFile
, NULL
);
2167 utimes(zLockFile
, NULL
);
2172 /* grab an exclusive lock */
2173 rc
= osMkdir(zLockFile
, 0777);
2175 /* failed to open/create the lock directory */
2177 if( EEXIST
== tErrno
){
2180 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2181 if( rc
!=SQLITE_BUSY
){
2182 storeLastErrno(pFile
, tErrno
);
2188 /* got it, set the type and return ok */
2189 pFile
->eFileLock
= eFileLock
;
2194 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2195 ** must be either NO_LOCK or SHARED_LOCK.
2197 ** If the locking level of the file descriptor is already at or below
2198 ** the requested locking level, this routine is a no-op.
2200 ** When the locking level reaches NO_LOCK, delete the lock file.
2202 static int dotlockUnlock(sqlite3_file
*id
, int eFileLock
) {
2203 unixFile
*pFile
= (unixFile
*)id
;
2204 char *zLockFile
= (char *)pFile
->lockingContext
;
2208 OSTRACE(("UNLOCK %d %d was %d pid=%d (dotlock)\n", pFile
->h
, eFileLock
,
2209 pFile
->eFileLock
, osGetpid(0)));
2210 assert( eFileLock
<=SHARED_LOCK
);
2212 /* no-op if possible */
2213 if( pFile
->eFileLock
==eFileLock
){
2217 /* To downgrade to shared, simply update our internal notion of the
2218 ** lock state. No need to mess with the file on disk.
2220 if( eFileLock
==SHARED_LOCK
){
2221 pFile
->eFileLock
= SHARED_LOCK
;
2225 /* To fully unlock the database, delete the lock file */
2226 assert( eFileLock
==NO_LOCK
);
2227 rc
= osRmdir(zLockFile
);
2230 if( tErrno
==ENOENT
){
2233 rc
= SQLITE_IOERR_UNLOCK
;
2234 storeLastErrno(pFile
, tErrno
);
2238 pFile
->eFileLock
= NO_LOCK
;
2243 ** Close a file. Make sure the lock has been released before closing.
2245 static int dotlockClose(sqlite3_file
*id
) {
2246 unixFile
*pFile
= (unixFile
*)id
;
2248 dotlockUnlock(id
, NO_LOCK
);
2249 sqlite3_free(pFile
->lockingContext
);
2250 return closeUnixFile(id
);
2252 /****************** End of the dot-file lock implementation *******************
2253 ******************************************************************************/
2255 /******************************************************************************
2256 ************************** Begin flock Locking ********************************
2258 ** Use the flock() system call to do file locking.
2260 ** flock() locking is like dot-file locking in that the various
2261 ** fine-grain locking levels supported by SQLite are collapsed into
2262 ** a single exclusive lock. In other words, SHARED, RESERVED, and
2263 ** PENDING locks are the same thing as an EXCLUSIVE lock. SQLite
2264 ** still works when you do this, but concurrency is reduced since
2265 ** only a single process can be reading the database at a time.
2267 ** Omit this section if SQLITE_ENABLE_LOCKING_STYLE is turned off
2269 #if SQLITE_ENABLE_LOCKING_STYLE
2272 ** Retry flock() calls that fail with EINTR
2275 static int robust_flock(int fd
, int op
){
2277 do{ rc
= flock(fd
,op
); }while( rc
<0 && errno
==EINTR
);
2281 # define robust_flock(a,b) flock(a,b)
2286 ** This routine checks if there is a RESERVED lock held on the specified
2287 ** file by this or any other process. If such a lock is held, set *pResOut
2288 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2289 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2291 static int flockCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2294 unixFile
*pFile
= (unixFile
*)id
;
2296 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2300 /* Check if a thread in this process holds such a lock */
2301 if( pFile
->eFileLock
>SHARED_LOCK
){
2305 /* Otherwise see if some other process holds it. */
2307 /* attempt to get the lock */
2308 int lrc
= robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
);
2310 /* got the lock, unlock it */
2311 lrc
= robust_flock(pFile
->h
, LOCK_UN
);
2314 /* unlock failed with an error */
2315 lrc
= SQLITE_IOERR_UNLOCK
;
2316 storeLastErrno(pFile
, tErrno
);
2322 /* someone else might have it reserved */
2323 lrc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2324 if( IS_LOCK_ERROR(lrc
) ){
2325 storeLastErrno(pFile
, tErrno
);
2330 OSTRACE(("TEST WR-LOCK %d %d %d (flock)\n", pFile
->h
, rc
, reserved
));
2332 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2333 if( (rc
& 0xff) == SQLITE_IOERR
){
2337 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2338 *pResOut
= reserved
;
2343 ** Lock the file with the lock specified by parameter eFileLock - one
2344 ** of the following:
2347 ** (2) RESERVED_LOCK
2349 ** (4) EXCLUSIVE_LOCK
2351 ** Sometimes when requesting one lock state, additional lock states
2352 ** are inserted in between. The locking might fail on one of the later
2353 ** transitions leaving the lock state different from what it started but
2354 ** still short of its goal. The following chart shows the allowed
2355 ** transitions and the inserted intermediate states:
2357 ** UNLOCKED -> SHARED
2358 ** SHARED -> RESERVED
2359 ** SHARED -> (PENDING) -> EXCLUSIVE
2360 ** RESERVED -> (PENDING) -> EXCLUSIVE
2361 ** PENDING -> EXCLUSIVE
2363 ** flock() only really support EXCLUSIVE locks. We track intermediate
2364 ** lock states in the sqlite3_file structure, but all locks SHARED or
2365 ** above are really EXCLUSIVE locks and exclude all other processes from
2368 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2369 ** routine to lower a locking level.
2371 static int flockLock(sqlite3_file
*id
, int eFileLock
) {
2373 unixFile
*pFile
= (unixFile
*)id
;
2377 /* if we already have a lock, it is exclusive.
2378 ** Just adjust level and punt on outta here. */
2379 if (pFile
->eFileLock
> NO_LOCK
) {
2380 pFile
->eFileLock
= eFileLock
;
2384 /* grab an exclusive lock */
2386 if (robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
)) {
2388 /* didn't get, must be busy */
2389 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2390 if( IS_LOCK_ERROR(rc
) ){
2391 storeLastErrno(pFile
, tErrno
);
2394 /* got it, set the type and return ok */
2395 pFile
->eFileLock
= eFileLock
;
2397 OSTRACE(("LOCK %d %s %s (flock)\n", pFile
->h
, azFileLock(eFileLock
),
2398 rc
==SQLITE_OK
? "ok" : "failed"));
2399 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2400 if( (rc
& 0xff) == SQLITE_IOERR
){
2403 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2409 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2410 ** must be either NO_LOCK or SHARED_LOCK.
2412 ** If the locking level of the file descriptor is already at or below
2413 ** the requested locking level, this routine is a no-op.
2415 static int flockUnlock(sqlite3_file
*id
, int eFileLock
) {
2416 unixFile
*pFile
= (unixFile
*)id
;
2419 OSTRACE(("UNLOCK %d %d was %d pid=%d (flock)\n", pFile
->h
, eFileLock
,
2420 pFile
->eFileLock
, osGetpid(0)));
2421 assert( eFileLock
<=SHARED_LOCK
);
2423 /* no-op if possible */
2424 if( pFile
->eFileLock
==eFileLock
){
2428 /* shared can just be set because we always have an exclusive */
2429 if (eFileLock
==SHARED_LOCK
) {
2430 pFile
->eFileLock
= eFileLock
;
2434 /* no, really, unlock. */
2435 if( robust_flock(pFile
->h
, LOCK_UN
) ){
2436 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2438 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2439 return SQLITE_IOERR_UNLOCK
;
2441 pFile
->eFileLock
= NO_LOCK
;
2449 static int flockClose(sqlite3_file
*id
) {
2451 flockUnlock(id
, NO_LOCK
);
2452 return closeUnixFile(id
);
2455 #endif /* SQLITE_ENABLE_LOCKING_STYLE && !OS_VXWORK */
2457 /******************* End of the flock lock implementation *********************
2458 ******************************************************************************/
2460 /******************************************************************************
2461 ************************ Begin Named Semaphore Locking ************************
2463 ** Named semaphore locking is only supported on VxWorks.
2465 ** Semaphore locking is like dot-lock and flock in that it really only
2466 ** supports EXCLUSIVE locking. Only a single process can read or write
2467 ** the database file at a time. This reduces potential concurrency, but
2468 ** makes the lock implementation much easier.
2473 ** This routine checks if there is a RESERVED lock held on the specified
2474 ** file by this or any other process. If such a lock is held, set *pResOut
2475 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2476 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2478 static int semXCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2481 unixFile
*pFile
= (unixFile
*)id
;
2483 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2487 /* Check if a thread in this process holds such a lock */
2488 if( pFile
->eFileLock
>SHARED_LOCK
){
2492 /* Otherwise see if some other process holds it. */
2494 sem_t
*pSem
= pFile
->pInode
->pSem
;
2496 if( sem_trywait(pSem
)==-1 ){
2498 if( EAGAIN
!= tErrno
){
2499 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_CHECKRESERVEDLOCK
);
2500 storeLastErrno(pFile
, tErrno
);
2502 /* someone else has the lock when we are in NO_LOCK */
2503 reserved
= (pFile
->eFileLock
< SHARED_LOCK
);
2506 /* we could have it if we want it */
2510 OSTRACE(("TEST WR-LOCK %d %d %d (sem)\n", pFile
->h
, rc
, reserved
));
2512 *pResOut
= reserved
;
2517 ** Lock the file with the lock specified by parameter eFileLock - one
2518 ** of the following:
2521 ** (2) RESERVED_LOCK
2523 ** (4) EXCLUSIVE_LOCK
2525 ** Sometimes when requesting one lock state, additional lock states
2526 ** are inserted in between. The locking might fail on one of the later
2527 ** transitions leaving the lock state different from what it started but
2528 ** still short of its goal. The following chart shows the allowed
2529 ** transitions and the inserted intermediate states:
2531 ** UNLOCKED -> SHARED
2532 ** SHARED -> RESERVED
2533 ** SHARED -> (PENDING) -> EXCLUSIVE
2534 ** RESERVED -> (PENDING) -> EXCLUSIVE
2535 ** PENDING -> EXCLUSIVE
2537 ** Semaphore locks only really support EXCLUSIVE locks. We track intermediate
2538 ** lock states in the sqlite3_file structure, but all locks SHARED or
2539 ** above are really EXCLUSIVE locks and exclude all other processes from
2542 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2543 ** routine to lower a locking level.
2545 static int semXLock(sqlite3_file
*id
, int eFileLock
) {
2546 unixFile
*pFile
= (unixFile
*)id
;
2547 sem_t
*pSem
= pFile
->pInode
->pSem
;
2550 /* if we already have a lock, it is exclusive.
2551 ** Just adjust level and punt on outta here. */
2552 if (pFile
->eFileLock
> NO_LOCK
) {
2553 pFile
->eFileLock
= eFileLock
;
2558 /* lock semaphore now but bail out when already locked. */
2559 if( sem_trywait(pSem
)==-1 ){
2564 /* got it, set the type and return ok */
2565 pFile
->eFileLock
= eFileLock
;
2572 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2573 ** must be either NO_LOCK or SHARED_LOCK.
2575 ** If the locking level of the file descriptor is already at or below
2576 ** the requested locking level, this routine is a no-op.
2578 static int semXUnlock(sqlite3_file
*id
, int eFileLock
) {
2579 unixFile
*pFile
= (unixFile
*)id
;
2580 sem_t
*pSem
= pFile
->pInode
->pSem
;
2584 OSTRACE(("UNLOCK %d %d was %d pid=%d (sem)\n", pFile
->h
, eFileLock
,
2585 pFile
->eFileLock
, osGetpid(0)));
2586 assert( eFileLock
<=SHARED_LOCK
);
2588 /* no-op if possible */
2589 if( pFile
->eFileLock
==eFileLock
){
2593 /* shared can just be set because we always have an exclusive */
2594 if (eFileLock
==SHARED_LOCK
) {
2595 pFile
->eFileLock
= eFileLock
;
2599 /* no, really unlock. */
2600 if ( sem_post(pSem
)==-1 ) {
2601 int rc
, tErrno
= errno
;
2602 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_UNLOCK
);
2603 if( IS_LOCK_ERROR(rc
) ){
2604 storeLastErrno(pFile
, tErrno
);
2608 pFile
->eFileLock
= NO_LOCK
;
2615 static int semXClose(sqlite3_file
*id
) {
2617 unixFile
*pFile
= (unixFile
*)id
;
2618 semXUnlock(id
, NO_LOCK
);
2621 releaseInodeInfo(pFile
);
2628 #endif /* OS_VXWORKS */
2630 ** Named semaphore locking is only available on VxWorks.
2632 *************** End of the named semaphore lock implementation ****************
2633 ******************************************************************************/
2636 /******************************************************************************
2637 *************************** Begin AFP Locking *********************************
2639 ** AFP is the Apple Filing Protocol. AFP is a network filesystem found
2640 ** on Apple Macintosh computers - both OS9 and OSX.
2642 ** Third-party implementations of AFP are available. But this code here
2643 ** only works on OSX.
2646 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
2648 ** The afpLockingContext structure contains all afp lock specific state
2650 typedef struct afpLockingContext afpLockingContext
;
2651 struct afpLockingContext
{
2653 const char *dbPath
; /* Name of the open file */
2656 struct ByteRangeLockPB2
2658 unsigned long long offset
; /* offset to first byte to lock */
2659 unsigned long long length
; /* nbr of bytes to lock */
2660 unsigned long long retRangeStart
; /* nbr of 1st byte locked if successful */
2661 unsigned char unLockFlag
; /* 1 = unlock, 0 = lock */
2662 unsigned char startEndFlag
; /* 1=rel to end of fork, 0=rel to start */
2663 int fd
; /* file desc to assoc this lock with */
2666 #define afpfsByteRangeLock2FSCTL _IOWR('z', 23, struct ByteRangeLockPB2)
2669 ** This is a utility for setting or clearing a bit-range lock on an
2672 ** Return SQLITE_OK on success, SQLITE_BUSY on failure.
2674 static int afpSetLock(
2675 const char *path
, /* Name of the file to be locked or unlocked */
2676 unixFile
*pFile
, /* Open file descriptor on path */
2677 unsigned long long offset
, /* First byte to be locked */
2678 unsigned long long length
, /* Number of bytes to lock */
2679 int setLockFlag
/* True to set lock. False to clear lock */
2681 struct ByteRangeLockPB2 pb
;
2684 pb
.unLockFlag
= setLockFlag
? 0 : 1;
2685 pb
.startEndFlag
= 0;
2690 OSTRACE(("AFPSETLOCK [%s] for %d%s in range %llx:%llx\n",
2691 (setLockFlag
?"ON":"OFF"), pFile
->h
, (pb
.fd
==-1?"[testval-1]":""),
2693 err
= fsctl(path
, afpfsByteRangeLock2FSCTL
, &pb
, 0);
2697 OSTRACE(("AFPSETLOCK failed to fsctl() '%s' %d %s\n",
2698 path
, tErrno
, strerror(tErrno
)));
2699 #ifdef SQLITE_IGNORE_AFP_LOCK_ERRORS
2702 rc
= sqliteErrorFromPosixError(tErrno
,
2703 setLockFlag
? SQLITE_IOERR_LOCK
: SQLITE_IOERR_UNLOCK
);
2704 #endif /* SQLITE_IGNORE_AFP_LOCK_ERRORS */
2705 if( IS_LOCK_ERROR(rc
) ){
2706 storeLastErrno(pFile
, tErrno
);
2715 ** This routine checks if there is a RESERVED lock held on the specified
2716 ** file by this or any other process. If such a lock is held, set *pResOut
2717 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2718 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2720 static int afpCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2723 unixFile
*pFile
= (unixFile
*)id
;
2724 afpLockingContext
*context
;
2726 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2729 context
= (afpLockingContext
*) pFile
->lockingContext
;
2730 if( context
->reserved
){
2734 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
2736 /* Check if a thread in this process holds such a lock */
2737 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
2741 /* Otherwise see if some other process holds it.
2744 /* lock the RESERVED byte */
2745 int lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2746 if( SQLITE_OK
==lrc
){
2747 /* if we succeeded in taking the reserved lock, unlock it to restore
2748 ** the original state */
2749 lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
2751 /* if we failed to get the lock then someone else must have it */
2754 if( IS_LOCK_ERROR(lrc
) ){
2760 OSTRACE(("TEST WR-LOCK %d %d %d (afp)\n", pFile
->h
, rc
, reserved
));
2762 *pResOut
= reserved
;
2767 ** Lock the file with the lock specified by parameter eFileLock - one
2768 ** of the following:
2771 ** (2) RESERVED_LOCK
2773 ** (4) EXCLUSIVE_LOCK
2775 ** Sometimes when requesting one lock state, additional lock states
2776 ** are inserted in between. The locking might fail on one of the later
2777 ** transitions leaving the lock state different from what it started but
2778 ** still short of its goal. The following chart shows the allowed
2779 ** transitions and the inserted intermediate states:
2781 ** UNLOCKED -> SHARED
2782 ** SHARED -> RESERVED
2783 ** SHARED -> (PENDING) -> EXCLUSIVE
2784 ** RESERVED -> (PENDING) -> EXCLUSIVE
2785 ** PENDING -> EXCLUSIVE
2787 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2788 ** routine to lower a locking level.
2790 static int afpLock(sqlite3_file
*id
, int eFileLock
){
2792 unixFile
*pFile
= (unixFile
*)id
;
2793 unixInodeInfo
*pInode
= pFile
->pInode
;
2794 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
2797 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (afp)\n", pFile
->h
,
2798 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
2799 azFileLock(pInode
->eFileLock
), pInode
->nShared
, osGetpid(0)));
2801 /* If there is already a lock of this type or more restrictive on the
2802 ** unixFile, do nothing. Don't use the afp_end_lock: exit path, as
2803 ** unixEnterMutex() hasn't been called yet.
2805 if( pFile
->eFileLock
>=eFileLock
){
2806 OSTRACE(("LOCK %d %s ok (already held) (afp)\n", pFile
->h
,
2807 azFileLock(eFileLock
)));
2811 /* Make sure the locking sequence is correct
2812 ** (1) We never move from unlocked to anything higher than shared lock.
2813 ** (2) SQLite never explicitly requests a pendig lock.
2814 ** (3) A shared lock is always held when a reserve lock is requested.
2816 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
2817 assert( eFileLock
!=PENDING_LOCK
);
2818 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
2820 /* This mutex is needed because pFile->pInode is shared across threads
2823 pInode
= pFile
->pInode
;
2825 /* If some thread using this PID has a lock via a different unixFile*
2826 ** handle that precludes the requested lock, return BUSY.
2828 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
2829 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
2835 /* If a SHARED lock is requested, and some thread using this PID already
2836 ** has a SHARED or RESERVED lock, then increment reference counts and
2837 ** return SQLITE_OK.
2839 if( eFileLock
==SHARED_LOCK
&&
2840 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
2841 assert( eFileLock
==SHARED_LOCK
);
2842 assert( pFile
->eFileLock
==0 );
2843 assert( pInode
->nShared
>0 );
2844 pFile
->eFileLock
= SHARED_LOCK
;
2850 /* A PENDING lock is needed before acquiring a SHARED lock and before
2851 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
2854 if( eFileLock
==SHARED_LOCK
2855 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
2858 failed
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 1);
2865 /* If control gets to this point, then actually go ahead and make
2866 ** operating system calls for the specified lock.
2868 if( eFileLock
==SHARED_LOCK
){
2869 int lrc1
, lrc2
, lrc1Errno
= 0;
2872 assert( pInode
->nShared
==0 );
2873 assert( pInode
->eFileLock
==0 );
2875 mask
= (sizeof(long)==8) ? LARGEST_INT64
: 0x7fffffff;
2876 /* Now get the read-lock SHARED_LOCK */
2877 /* note that the quality of the randomness doesn't matter that much */
2879 pInode
->sharedByte
= (lk
& mask
)%(SHARED_SIZE
- 1);
2880 lrc1
= afpSetLock(context
->dbPath
, pFile
,
2881 SHARED_FIRST
+pInode
->sharedByte
, 1, 1);
2882 if( IS_LOCK_ERROR(lrc1
) ){
2883 lrc1Errno
= pFile
->lastErrno
;
2885 /* Drop the temporary PENDING lock */
2886 lrc2
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
2888 if( IS_LOCK_ERROR(lrc1
) ) {
2889 storeLastErrno(pFile
, lrc1Errno
);
2892 } else if( IS_LOCK_ERROR(lrc2
) ){
2895 } else if( lrc1
!= SQLITE_OK
) {
2898 pFile
->eFileLock
= SHARED_LOCK
;
2900 pInode
->nShared
= 1;
2902 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
2903 /* We are trying for an exclusive lock but another thread in this
2904 ** same process is still holding a shared lock. */
2907 /* The request was for a RESERVED or EXCLUSIVE lock. It is
2908 ** assumed that there is a SHARED or greater lock on the file
2912 assert( 0!=pFile
->eFileLock
);
2913 if (eFileLock
>= RESERVED_LOCK
&& pFile
->eFileLock
< RESERVED_LOCK
) {
2914 /* Acquire a RESERVED lock */
2915 failed
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2917 context
->reserved
= 1;
2920 if (!failed
&& eFileLock
== EXCLUSIVE_LOCK
) {
2921 /* Acquire an EXCLUSIVE lock */
2923 /* Remove the shared lock before trying the range. we'll need to
2924 ** reestablish the shared lock if we can't get the afpUnlock
2926 if( !(failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
+
2927 pInode
->sharedByte
, 1, 0)) ){
2928 int failed2
= SQLITE_OK
;
2929 /* now attemmpt to get the exclusive lock range */
2930 failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
,
2932 if( failed
&& (failed2
= afpSetLock(context
->dbPath
, pFile
,
2933 SHARED_FIRST
+ pInode
->sharedByte
, 1, 1)) ){
2934 /* Can't reestablish the shared lock. Sqlite can't deal, this is
2935 ** a critical I/O error
2937 rc
= ((failed
& 0xff) == SQLITE_IOERR
) ? failed2
:
2950 if( rc
==SQLITE_OK
){
2951 pFile
->eFileLock
= eFileLock
;
2952 pInode
->eFileLock
= eFileLock
;
2953 }else if( eFileLock
==EXCLUSIVE_LOCK
){
2954 pFile
->eFileLock
= PENDING_LOCK
;
2955 pInode
->eFileLock
= PENDING_LOCK
;
2960 OSTRACE(("LOCK %d %s %s (afp)\n", pFile
->h
, azFileLock(eFileLock
),
2961 rc
==SQLITE_OK
? "ok" : "failed"));
2966 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2967 ** must be either NO_LOCK or SHARED_LOCK.
2969 ** If the locking level of the file descriptor is already at or below
2970 ** the requested locking level, this routine is a no-op.
2972 static int afpUnlock(sqlite3_file
*id
, int eFileLock
) {
2974 unixFile
*pFile
= (unixFile
*)id
;
2975 unixInodeInfo
*pInode
;
2976 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
2983 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (afp)\n", pFile
->h
, eFileLock
,
2984 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
2987 assert( eFileLock
<=SHARED_LOCK
);
2988 if( pFile
->eFileLock
<=eFileLock
){
2992 pInode
= pFile
->pInode
;
2993 assert( pInode
->nShared
!=0 );
2994 if( pFile
->eFileLock
>SHARED_LOCK
){
2995 assert( pInode
->eFileLock
==pFile
->eFileLock
);
2996 SimulateIOErrorBenign(1);
2997 SimulateIOError( h
=(-1) )
2998 SimulateIOErrorBenign(0);
3001 /* When reducing a lock such that other processes can start
3002 ** reading the database file again, make sure that the
3003 ** transaction counter was updated if any part of the database
3004 ** file changed. If the transaction counter is not updated,
3005 ** other connections to the same file might not realize that
3006 ** the file has changed and hence might not know to flush their
3007 ** cache. The use of a stale cache can lead to database corruption.
3009 assert( pFile
->inNormalWrite
==0
3010 || pFile
->dbUpdate
==0
3011 || pFile
->transCntrChng
==1 );
3012 pFile
->inNormalWrite
= 0;
3015 if( pFile
->eFileLock
==EXCLUSIVE_LOCK
){
3016 rc
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
, SHARED_SIZE
, 0);
3017 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1) ){
3018 /* only re-establish the shared lock if necessary */
3019 int sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3020 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 1);
3025 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=PENDING_LOCK
){
3026 rc
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
3028 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=RESERVED_LOCK
&& context
->reserved
){
3029 rc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
3031 context
->reserved
= 0;
3034 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1)){
3035 pInode
->eFileLock
= SHARED_LOCK
;
3038 if( rc
==SQLITE_OK
&& eFileLock
==NO_LOCK
){
3040 /* Decrement the shared lock counter. Release the lock using an
3041 ** OS call only when all threads in this same process have released
3044 unsigned long long sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3046 if( pInode
->nShared
==0 ){
3047 SimulateIOErrorBenign(1);
3048 SimulateIOError( h
=(-1) )
3049 SimulateIOErrorBenign(0);
3051 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 0);
3054 pInode
->eFileLock
= NO_LOCK
;
3055 pFile
->eFileLock
= NO_LOCK
;
3058 if( rc
==SQLITE_OK
){
3060 assert( pInode
->nLock
>=0 );
3061 if( pInode
->nLock
==0 ){
3062 closePendingFds(pFile
);
3068 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
3073 ** Close a file & cleanup AFP specific locking context
3075 static int afpClose(sqlite3_file
*id
) {
3077 unixFile
*pFile
= (unixFile
*)id
;
3079 afpUnlock(id
, NO_LOCK
);
3081 if( pFile
->pInode
&& pFile
->pInode
->nLock
){
3082 /* If there are outstanding locks, do not actually close the file just
3083 ** yet because that would clear those locks. Instead, add the file
3084 ** descriptor to pInode->aPending. It will be automatically closed when
3085 ** the last lock is cleared.
3087 setPendingFd(pFile
);
3089 releaseInodeInfo(pFile
);
3090 sqlite3_free(pFile
->lockingContext
);
3091 rc
= closeUnixFile(id
);
3096 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3098 ** The code above is the AFP lock implementation. The code is specific
3099 ** to MacOSX and does not work on other unix platforms. No alternative
3100 ** is available. If you don't compile for a mac, then the "unix-afp"
3101 ** VFS is not available.
3103 ********************* End of the AFP lock implementation **********************
3104 ******************************************************************************/
3106 /******************************************************************************
3107 *************************** Begin NFS Locking ********************************/
3109 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
3111 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
3112 ** must be either NO_LOCK or SHARED_LOCK.
3114 ** If the locking level of the file descriptor is already at or below
3115 ** the requested locking level, this routine is a no-op.
3117 static int nfsUnlock(sqlite3_file
*id
, int eFileLock
){
3118 return posixUnlock(id
, eFileLock
, 1);
3121 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3123 ** The code above is the NFS lock implementation. The code is specific
3124 ** to MacOSX and does not work on other unix platforms. No alternative
3127 ********************* End of the NFS lock implementation **********************
3128 ******************************************************************************/
3130 /******************************************************************************
3131 **************** Non-locking sqlite3_file methods *****************************
3133 ** The next division contains implementations for all methods of the
3134 ** sqlite3_file object other than the locking methods. The locking
3135 ** methods were defined in divisions above (one locking method per
3136 ** division). Those methods that are common to all locking modes
3137 ** are gather together into this division.
3141 ** Seek to the offset passed as the second argument, then read cnt
3142 ** bytes into pBuf. Return the number of bytes actually read.
3144 ** NB: If you define USE_PREAD or USE_PREAD64, then it might also
3145 ** be necessary to define _XOPEN_SOURCE to be 500. This varies from
3146 ** one system to another. Since SQLite does not define USE_PREAD
3147 ** in any form by default, we will not attempt to define _XOPEN_SOURCE.
3148 ** See tickets #2741 and #2681.
3150 ** To avoid stomping the errno value on a failed read the lastErrno value
3151 ** is set before returning.
3153 static int seekAndRead(unixFile
*id
, sqlite3_int64 offset
, void *pBuf
, int cnt
){
3156 #if (!defined(USE_PREAD) && !defined(USE_PREAD64))
3160 assert( cnt
==(cnt
&0x1ffff) );
3163 #if defined(USE_PREAD)
3164 got
= osPread(id
->h
, pBuf
, cnt
, offset
);
3165 SimulateIOError( got
= -1 );
3166 #elif defined(USE_PREAD64)
3167 got
= osPread64(id
->h
, pBuf
, cnt
, offset
);
3168 SimulateIOError( got
= -1 );
3170 newOffset
= lseek(id
->h
, offset
, SEEK_SET
);
3171 SimulateIOError( newOffset
= -1 );
3173 storeLastErrno((unixFile
*)id
, errno
);
3176 got
= osRead(id
->h
, pBuf
, cnt
);
3178 if( got
==cnt
) break;
3180 if( errno
==EINTR
){ got
= 1; continue; }
3182 storeLastErrno((unixFile
*)id
, errno
);
3188 pBuf
= (void*)(got
+ (char*)pBuf
);
3192 OSTRACE(("READ %-3d %5d %7lld %llu\n",
3193 id
->h
, got
+prior
, offset
-prior
, TIMER_ELAPSED
));
3198 ** Read data from a file into a buffer. Return SQLITE_OK if all
3199 ** bytes were read successfully and SQLITE_IOERR if anything goes
3202 static int unixRead(
3206 sqlite3_int64 offset
3208 unixFile
*pFile
= (unixFile
*)id
;
3211 assert( offset
>=0 );
3214 /* If this is a database file (not a journal, master-journal or temp
3215 ** file), the bytes in the locking range should never be read or written. */
3217 assert( pFile
->pPreallocatedUnused
==0
3218 || offset
>=PENDING_BYTE
+512
3219 || offset
+amt
<=PENDING_BYTE
3223 #if SQLITE_MAX_MMAP_SIZE>0
3224 /* Deal with as much of this read request as possible by transfering
3225 ** data from the memory mapping using memcpy(). */
3226 if( offset
<pFile
->mmapSize
){
3227 if( offset
+amt
<= pFile
->mmapSize
){
3228 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], amt
);
3231 int nCopy
= pFile
->mmapSize
- offset
;
3232 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], nCopy
);
3233 pBuf
= &((u8
*)pBuf
)[nCopy
];
3240 got
= seekAndRead(pFile
, offset
, pBuf
, amt
);
3244 /* lastErrno set by seekAndRead */
3245 return SQLITE_IOERR_READ
;
3247 storeLastErrno(pFile
, 0); /* not a system error */
3248 /* Unread parts of the buffer must be zero-filled */
3249 memset(&((char*)pBuf
)[got
], 0, amt
-got
);
3250 return SQLITE_IOERR_SHORT_READ
;
3255 ** Attempt to seek the file-descriptor passed as the first argument to
3256 ** absolute offset iOff, then attempt to write nBuf bytes of data from
3257 ** pBuf to it. If an error occurs, return -1 and set *piErrno. Otherwise,
3258 ** return the actual number of bytes written (which may be less than
3261 static int seekAndWriteFd(
3262 int fd
, /* File descriptor to write to */
3263 i64 iOff
, /* File offset to begin writing at */
3264 const void *pBuf
, /* Copy data from this buffer to the file */
3265 int nBuf
, /* Size of buffer pBuf in bytes */
3266 int *piErrno
/* OUT: Error number if error occurs */
3268 int rc
= 0; /* Value returned by system call */
3270 assert( nBuf
==(nBuf
&0x1ffff) );
3272 assert( piErrno
!=0 );
3276 #if defined(USE_PREAD)
3277 do{ rc
= (int)osPwrite(fd
, pBuf
, nBuf
, iOff
); }while( rc
<0 && errno
==EINTR
);
3278 #elif defined(USE_PREAD64)
3279 do{ rc
= (int)osPwrite64(fd
, pBuf
, nBuf
, iOff
);}while( rc
<0 && errno
==EINTR
);
3282 i64 iSeek
= lseek(fd
, iOff
, SEEK_SET
);
3283 SimulateIOError( iSeek
= -1 );
3288 rc
= osWrite(fd
, pBuf
, nBuf
);
3289 }while( rc
<0 && errno
==EINTR
);
3293 OSTRACE(("WRITE %-3d %5d %7lld %llu\n", fd
, rc
, iOff
, TIMER_ELAPSED
));
3295 if( rc
<0 ) *piErrno
= errno
;
3301 ** Seek to the offset in id->offset then read cnt bytes into pBuf.
3302 ** Return the number of bytes actually read. Update the offset.
3304 ** To avoid stomping the errno value on a failed write the lastErrno value
3305 ** is set before returning.
3307 static int seekAndWrite(unixFile
*id
, i64 offset
, const void *pBuf
, int cnt
){
3308 return seekAndWriteFd(id
->h
, offset
, pBuf
, cnt
, &id
->lastErrno
);
3313 ** Write data from a buffer into a file. Return SQLITE_OK on success
3314 ** or some other error code on failure.
3316 static int unixWrite(
3320 sqlite3_int64 offset
3322 unixFile
*pFile
= (unixFile
*)id
;
3327 /* If this is a database file (not a journal, master-journal or temp
3328 ** file), the bytes in the locking range should never be read or written. */
3330 assert( pFile
->pPreallocatedUnused
==0
3331 || offset
>=PENDING_BYTE
+512
3332 || offset
+amt
<=PENDING_BYTE
3337 /* If we are doing a normal write to a database file (as opposed to
3338 ** doing a hot-journal rollback or a write to some file other than a
3339 ** normal database file) then record the fact that the database
3340 ** has changed. If the transaction counter is modified, record that
3343 if( pFile
->inNormalWrite
){
3344 pFile
->dbUpdate
= 1; /* The database has been modified */
3345 if( offset
<=24 && offset
+amt
>=27 ){
3348 SimulateIOErrorBenign(1);
3349 rc
= seekAndRead(pFile
, 24, oldCntr
, 4);
3350 SimulateIOErrorBenign(0);
3351 if( rc
!=4 || memcmp(oldCntr
, &((char*)pBuf
)[24-offset
], 4)!=0 ){
3352 pFile
->transCntrChng
= 1; /* The transaction counter has changed */
3358 #if defined(SQLITE_MMAP_READWRITE) && SQLITE_MAX_MMAP_SIZE>0
3359 /* Deal with as much of this write request as possible by transfering
3360 ** data from the memory mapping using memcpy(). */
3361 if( offset
<pFile
->mmapSize
){
3362 if( offset
+amt
<= pFile
->mmapSize
){
3363 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, amt
);
3366 int nCopy
= pFile
->mmapSize
- offset
;
3367 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, nCopy
);
3368 pBuf
= &((u8
*)pBuf
)[nCopy
];
3375 while( (wrote
= seekAndWrite(pFile
, offset
, pBuf
, amt
))<amt
&& wrote
>0 ){
3378 pBuf
= &((char*)pBuf
)[wrote
];
3380 SimulateIOError(( wrote
=(-1), amt
=1 ));
3381 SimulateDiskfullError(( wrote
=0, amt
=1 ));
3384 if( wrote
<0 && pFile
->lastErrno
!=ENOSPC
){
3385 /* lastErrno set by seekAndWrite */
3386 return SQLITE_IOERR_WRITE
;
3388 storeLastErrno(pFile
, 0); /* not a system error */
3398 ** Count the number of fullsyncs and normal syncs. This is used to test
3399 ** that syncs and fullsyncs are occurring at the right times.
3401 int sqlite3_sync_count
= 0;
3402 int sqlite3_fullsync_count
= 0;
3406 ** We do not trust systems to provide a working fdatasync(). Some do.
3407 ** Others do no. To be safe, we will stick with the (slightly slower)
3408 ** fsync(). If you know that your system does support fdatasync() correctly,
3409 ** then simply compile with -Dfdatasync=fdatasync or -DHAVE_FDATASYNC
3411 #if !defined(fdatasync) && !HAVE_FDATASYNC
3412 # define fdatasync fsync
3416 ** Define HAVE_FULLFSYNC to 0 or 1 depending on whether or not
3417 ** the F_FULLFSYNC macro is defined. F_FULLFSYNC is currently
3418 ** only available on Mac OS X. But that could change.
3421 # define HAVE_FULLFSYNC 1
3423 # define HAVE_FULLFSYNC 0
3428 ** The fsync() system call does not work as advertised on many
3429 ** unix systems. The following procedure is an attempt to make
3432 ** The SQLITE_NO_SYNC macro disables all fsync()s. This is useful
3433 ** for testing when we want to run through the test suite quickly.
3434 ** You are strongly advised *not* to deploy with SQLITE_NO_SYNC
3435 ** enabled, however, since with SQLITE_NO_SYNC enabled, an OS crash
3436 ** or power failure will likely corrupt the database file.
3438 ** SQLite sets the dataOnly flag if the size of the file is unchanged.
3439 ** The idea behind dataOnly is that it should only write the file content
3440 ** to disk, not the inode. We only set dataOnly if the file size is
3441 ** unchanged since the file size is part of the inode. However,
3442 ** Ted Ts'o tells us that fdatasync() will also write the inode if the
3443 ** file size has changed. The only real difference between fdatasync()
3444 ** and fsync(), Ted tells us, is that fdatasync() will not flush the
3445 ** inode if the mtime or owner or other inode attributes have changed.
3446 ** We only care about the file size, not the other file attributes, so
3447 ** as far as SQLite is concerned, an fdatasync() is always adequate.
3448 ** So, we always use fdatasync() if it is available, regardless of
3449 ** the value of the dataOnly flag.
3451 static int full_fsync(int fd
, int fullSync
, int dataOnly
){
3454 /* The following "ifdef/elif/else/" block has the same structure as
3455 ** the one below. It is replicated here solely to avoid cluttering
3456 ** up the real code with the UNUSED_PARAMETER() macros.
3458 #ifdef SQLITE_NO_SYNC
3459 UNUSED_PARAMETER(fd
);
3460 UNUSED_PARAMETER(fullSync
);
3461 UNUSED_PARAMETER(dataOnly
);
3462 #elif HAVE_FULLFSYNC
3463 UNUSED_PARAMETER(dataOnly
);
3465 UNUSED_PARAMETER(fullSync
);
3466 UNUSED_PARAMETER(dataOnly
);
3469 /* Record the number of times that we do a normal fsync() and
3470 ** FULLSYNC. This is used during testing to verify that this procedure
3471 ** gets called with the correct arguments.
3474 if( fullSync
) sqlite3_fullsync_count
++;
3475 sqlite3_sync_count
++;
3478 /* If we compiled with the SQLITE_NO_SYNC flag, then syncing is a
3479 ** no-op. But go ahead and call fstat() to validate the file
3480 ** descriptor as we need a method to provoke a failure during
3481 ** coverate testing.
3483 #ifdef SQLITE_NO_SYNC
3486 rc
= osFstat(fd
, &buf
);
3488 #elif HAVE_FULLFSYNC
3490 rc
= osFcntl(fd
, F_FULLFSYNC
, 0);
3494 /* If the FULLFSYNC failed, fall back to attempting an fsync().
3495 ** It shouldn't be possible for fullfsync to fail on the local
3496 ** file system (on OSX), so failure indicates that FULLFSYNC
3497 ** isn't supported for this file system. So, attempt an fsync
3498 ** and (for now) ignore the overhead of a superfluous fcntl call.
3499 ** It'd be better to detect fullfsync support once and avoid
3500 ** the fcntl call every time sync is called.
3502 if( rc
) rc
= fsync(fd
);
3504 #elif defined(__APPLE__)
3505 /* fdatasync() on HFS+ doesn't yet flush the file size if it changed correctly
3506 ** so currently we default to the macro that redefines fdatasync to fsync
3512 if( rc
==-1 && errno
==ENOTSUP
){
3515 #endif /* OS_VXWORKS */
3516 #endif /* ifdef SQLITE_NO_SYNC elif HAVE_FULLFSYNC */
3518 if( OS_VXWORKS
&& rc
!= -1 ){
3525 ** Open a file descriptor to the directory containing file zFilename.
3526 ** If successful, *pFd is set to the opened file descriptor and
3527 ** SQLITE_OK is returned. If an error occurs, either SQLITE_NOMEM
3528 ** or SQLITE_CANTOPEN is returned and *pFd is set to an undefined
3531 ** The directory file descriptor is used for only one thing - to
3532 ** fsync() a directory to make sure file creation and deletion events
3533 ** are flushed to disk. Such fsyncs are not needed on newer
3534 ** journaling filesystems, but are required on older filesystems.
3536 ** This routine can be overridden using the xSetSysCall interface.
3537 ** The ability to override this routine was added in support of the
3538 ** chromium sandbox. Opening a directory is a security risk (we are
3539 ** told) so making it overrideable allows the chromium sandbox to
3540 ** replace this routine with a harmless no-op. To make this routine
3541 ** a no-op, replace it with a stub that returns SQLITE_OK but leaves
3542 ** *pFd set to a negative number.
3544 ** If SQLITE_OK is returned, the caller is responsible for closing
3545 ** the file descriptor *pFd using close().
3547 static int openDirectory(const char *zFilename
, int *pFd
){
3550 char zDirname
[MAX_PATHNAME
+1];
3552 sqlite3_snprintf(MAX_PATHNAME
, zDirname
, "%s", zFilename
);
3553 for(ii
=(int)strlen(zDirname
); ii
>0 && zDirname
[ii
]!='/'; ii
--);
3555 zDirname
[ii
] = '\0';
3557 if( zDirname
[0]!='/' ) zDirname
[0] = '.';
3560 fd
= robust_open(zDirname
, O_RDONLY
|O_BINARY
, 0);
3562 OSTRACE(("OPENDIR %-3d %s\n", fd
, zDirname
));
3565 if( fd
>=0 ) return SQLITE_OK
;
3566 return unixLogError(SQLITE_CANTOPEN_BKPT
, "openDirectory", zDirname
);
3570 ** Make sure all writes to a particular file are committed to disk.
3572 ** If dataOnly==0 then both the file itself and its metadata (file
3573 ** size, access time, etc) are synced. If dataOnly!=0 then only the
3574 ** file data is synced.
3576 ** Under Unix, also make sure that the directory entry for the file
3577 ** has been created by fsync-ing the directory that contains the file.
3578 ** If we do not do this and we encounter a power failure, the directory
3579 ** entry for the journal might not exist after we reboot. The next
3580 ** SQLite to access the file will not know that the journal exists (because
3581 ** the directory entry for the journal was never created) and the transaction
3582 ** will not roll back - possibly leading to database corruption.
3584 static int unixSync(sqlite3_file
*id
, int flags
){
3586 unixFile
*pFile
= (unixFile
*)id
;
3588 int isDataOnly
= (flags
&SQLITE_SYNC_DATAONLY
);
3589 int isFullsync
= (flags
&0x0F)==SQLITE_SYNC_FULL
;
3591 /* Check that one of SQLITE_SYNC_NORMAL or FULL was passed */
3592 assert((flags
&0x0F)==SQLITE_SYNC_NORMAL
3593 || (flags
&0x0F)==SQLITE_SYNC_FULL
3596 /* Unix cannot, but some systems may return SQLITE_FULL from here. This
3597 ** line is to test that doing so does not cause any problems.
3599 SimulateDiskfullError( return SQLITE_FULL
);
3602 OSTRACE(("SYNC %-3d\n", pFile
->h
));
3603 rc
= full_fsync(pFile
->h
, isFullsync
, isDataOnly
);
3604 SimulateIOError( rc
=1 );
3606 storeLastErrno(pFile
, errno
);
3607 return unixLogError(SQLITE_IOERR_FSYNC
, "full_fsync", pFile
->zPath
);
3610 /* Also fsync the directory containing the file if the DIRSYNC flag
3611 ** is set. This is a one-time occurrence. Many systems (examples: AIX)
3612 ** are unable to fsync a directory, so ignore errors on the fsync.
3614 if( pFile
->ctrlFlags
& UNIXFILE_DIRSYNC
){
3616 OSTRACE(("DIRSYNC %s (have_fullfsync=%d fullsync=%d)\n", pFile
->zPath
,
3617 HAVE_FULLFSYNC
, isFullsync
));
3618 rc
= osOpenDirectory(pFile
->zPath
, &dirfd
);
3619 if( rc
==SQLITE_OK
){
3620 full_fsync(dirfd
, 0, 0);
3621 robust_close(pFile
, dirfd
, __LINE__
);
3623 assert( rc
==SQLITE_CANTOPEN
);
3626 pFile
->ctrlFlags
&= ~UNIXFILE_DIRSYNC
;
3632 ** Truncate an open file to a specified size
3634 static int unixTruncate(sqlite3_file
*id
, i64 nByte
){
3635 unixFile
*pFile
= (unixFile
*)id
;
3638 SimulateIOError( return SQLITE_IOERR_TRUNCATE
);
3640 /* If the user has configured a chunk-size for this file, truncate the
3641 ** file so that it consists of an integer number of chunks (i.e. the
3642 ** actual file size after the operation may be larger than the requested
3645 if( pFile
->szChunk
>0 ){
3646 nByte
= ((nByte
+ pFile
->szChunk
- 1)/pFile
->szChunk
) * pFile
->szChunk
;
3649 rc
= robust_ftruncate(pFile
->h
, nByte
);
3651 storeLastErrno(pFile
, errno
);
3652 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3655 /* If we are doing a normal write to a database file (as opposed to
3656 ** doing a hot-journal rollback or a write to some file other than a
3657 ** normal database file) and we truncate the file to zero length,
3658 ** that effectively updates the change counter. This might happen
3659 ** when restoring a database using the backup API from a zero-length
3662 if( pFile
->inNormalWrite
&& nByte
==0 ){
3663 pFile
->transCntrChng
= 1;
3667 #if SQLITE_MAX_MMAP_SIZE>0
3668 /* If the file was just truncated to a size smaller than the currently
3669 ** mapped region, reduce the effective mapping size as well. SQLite will
3670 ** use read() and write() to access data beyond this point from now on.
3672 if( nByte
<pFile
->mmapSize
){
3673 pFile
->mmapSize
= nByte
;
3682 ** Determine the current size of a file in bytes
3684 static int unixFileSize(sqlite3_file
*id
, i64
*pSize
){
3688 rc
= osFstat(((unixFile
*)id
)->h
, &buf
);
3689 SimulateIOError( rc
=1 );
3691 storeLastErrno((unixFile
*)id
, errno
);
3692 return SQLITE_IOERR_FSTAT
;
3694 *pSize
= buf
.st_size
;
3696 /* When opening a zero-size database, the findInodeInfo() procedure
3697 ** writes a single byte into that file in order to work around a bug
3698 ** in the OS-X msdos filesystem. In order to avoid problems with upper
3699 ** layers, we need to report this file size as zero even though it is
3700 ** really 1. Ticket #3260.
3702 if( *pSize
==1 ) *pSize
= 0;
3708 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3710 ** Handler for proxy-locking file-control verbs. Defined below in the
3711 ** proxying locking division.
3713 static int proxyFileControl(sqlite3_file
*,int,void*);
3717 ** This function is called to handle the SQLITE_FCNTL_SIZE_HINT
3718 ** file-control operation. Enlarge the database to nBytes in size
3719 ** (rounded up to the next chunk-size). If the database is already
3720 ** nBytes or larger, this routine is a no-op.
3722 static int fcntlSizeHint(unixFile
*pFile
, i64 nByte
){
3723 if( pFile
->szChunk
>0 ){
3724 i64 nSize
; /* Required file size */
3725 struct stat buf
; /* Used to hold return values of fstat() */
3727 if( osFstat(pFile
->h
, &buf
) ){
3728 return SQLITE_IOERR_FSTAT
;
3731 nSize
= ((nByte
+pFile
->szChunk
-1) / pFile
->szChunk
) * pFile
->szChunk
;
3732 if( nSize
>(i64
)buf
.st_size
){
3734 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
3735 /* The code below is handling the return value of osFallocate()
3736 ** correctly. posix_fallocate() is defined to "returns zero on success,
3737 ** or an error number on failure". See the manpage for details. */
3740 err
= osFallocate(pFile
->h
, buf
.st_size
, nSize
-buf
.st_size
);
3741 }while( err
==EINTR
);
3742 if( err
) return SQLITE_IOERR_WRITE
;
3744 /* If the OS does not have posix_fallocate(), fake it. Write a
3745 ** single byte to the last byte in each block that falls entirely
3746 ** within the extended region. Then, if required, a single byte
3747 ** at offset (nSize-1), to set the size of the file correctly.
3748 ** This is a similar technique to that used by glibc on systems
3749 ** that do not have a real fallocate() call.
3751 int nBlk
= buf
.st_blksize
; /* File-system block size */
3752 int nWrite
= 0; /* Number of bytes written by seekAndWrite */
3753 i64 iWrite
; /* Next offset to write to */
3755 iWrite
= (buf
.st_size
/nBlk
)*nBlk
+ nBlk
- 1;
3756 assert( iWrite
>=buf
.st_size
);
3757 assert( ((iWrite
+1)%nBlk
)==0 );
3758 for(/*no-op*/; iWrite
<nSize
+nBlk
-1; iWrite
+=nBlk
){
3759 if( iWrite
>=nSize
) iWrite
= nSize
- 1;
3760 nWrite
= seekAndWrite(pFile
, iWrite
, "", 1);
3761 if( nWrite
!=1 ) return SQLITE_IOERR_WRITE
;
3767 #if SQLITE_MAX_MMAP_SIZE>0
3768 if( pFile
->mmapSizeMax
>0 && nByte
>pFile
->mmapSize
){
3770 if( pFile
->szChunk
<=0 ){
3771 if( robust_ftruncate(pFile
->h
, nByte
) ){
3772 storeLastErrno(pFile
, errno
);
3773 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3777 rc
= unixMapfile(pFile
, nByte
);
3786 ** If *pArg is initially negative then this is a query. Set *pArg to
3787 ** 1 or 0 depending on whether or not bit mask of pFile->ctrlFlags is set.
3789 ** If *pArg is 0 or 1, then clear or set the mask bit of pFile->ctrlFlags.
3791 static void unixModeBit(unixFile
*pFile
, unsigned char mask
, int *pArg
){
3793 *pArg
= (pFile
->ctrlFlags
& mask
)!=0;
3794 }else if( (*pArg
)==0 ){
3795 pFile
->ctrlFlags
&= ~mask
;
3797 pFile
->ctrlFlags
|= mask
;
3801 /* Forward declaration */
3802 static int unixGetTempname(int nBuf
, char *zBuf
);
3805 ** Information and control of an open file handle.
3807 static int unixFileControl(sqlite3_file
*id
, int op
, void *pArg
){
3808 unixFile
*pFile
= (unixFile
*)id
;
3810 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3811 case SQLITE_FCNTL_BEGIN_ATOMIC_WRITE
: {
3812 int rc
= osIoctl(pFile
->h
, F2FS_IOC_START_ATOMIC_WRITE
);
3813 return rc
? SQLITE_IOERR_BEGIN_ATOMIC
: SQLITE_OK
;
3815 case SQLITE_FCNTL_COMMIT_ATOMIC_WRITE
: {
3816 int rc
= osIoctl(pFile
->h
, F2FS_IOC_COMMIT_ATOMIC_WRITE
);
3817 return rc
? SQLITE_IOERR_COMMIT_ATOMIC
: SQLITE_OK
;
3819 case SQLITE_FCNTL_ROLLBACK_ATOMIC_WRITE
: {
3820 int rc
= osIoctl(pFile
->h
, F2FS_IOC_ABORT_VOLATILE_WRITE
);
3821 return rc
? SQLITE_IOERR_ROLLBACK_ATOMIC
: SQLITE_OK
;
3823 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3825 case SQLITE_FCNTL_LOCKSTATE
: {
3826 *(int*)pArg
= pFile
->eFileLock
;
3829 case SQLITE_FCNTL_LAST_ERRNO
: {
3830 *(int*)pArg
= pFile
->lastErrno
;
3833 case SQLITE_FCNTL_CHUNK_SIZE
: {
3834 pFile
->szChunk
= *(int *)pArg
;
3837 case SQLITE_FCNTL_SIZE_HINT
: {
3839 SimulateIOErrorBenign(1);
3840 rc
= fcntlSizeHint(pFile
, *(i64
*)pArg
);
3841 SimulateIOErrorBenign(0);
3844 case SQLITE_FCNTL_PERSIST_WAL
: {
3845 unixModeBit(pFile
, UNIXFILE_PERSIST_WAL
, (int*)pArg
);
3848 case SQLITE_FCNTL_POWERSAFE_OVERWRITE
: {
3849 unixModeBit(pFile
, UNIXFILE_PSOW
, (int*)pArg
);
3852 case SQLITE_FCNTL_VFSNAME
: {
3853 *(char**)pArg
= sqlite3_mprintf("%s", pFile
->pVfs
->zName
);
3856 case SQLITE_FCNTL_TEMPFILENAME
: {
3857 char *zTFile
= sqlite3_malloc64( pFile
->pVfs
->mxPathname
);
3859 unixGetTempname(pFile
->pVfs
->mxPathname
, zTFile
);
3860 *(char**)pArg
= zTFile
;
3864 case SQLITE_FCNTL_HAS_MOVED
: {
3865 *(int*)pArg
= fileHasMoved(pFile
);
3868 #if SQLITE_MAX_MMAP_SIZE>0
3869 case SQLITE_FCNTL_MMAP_SIZE
: {
3870 i64 newLimit
= *(i64
*)pArg
;
3872 if( newLimit
>sqlite3GlobalConfig
.mxMmap
){
3873 newLimit
= sqlite3GlobalConfig
.mxMmap
;
3876 /* The value of newLimit may be eventually cast to (size_t) and passed
3877 ** to mmap(). Restrict its value to 2GB if (size_t) is not at least a
3879 if( newLimit
>0 && sizeof(size_t)<8 ){
3880 newLimit
= (newLimit
& 0x7FFFFFFF);
3883 *(i64
*)pArg
= pFile
->mmapSizeMax
;
3884 if( newLimit
>=0 && newLimit
!=pFile
->mmapSizeMax
&& pFile
->nFetchOut
==0 ){
3885 pFile
->mmapSizeMax
= newLimit
;
3886 if( pFile
->mmapSize
>0 ){
3887 unixUnmapfile(pFile
);
3888 rc
= unixMapfile(pFile
, -1);
3895 /* The pager calls this method to signal that it has done
3896 ** a rollback and that the database is therefore unchanged and
3897 ** it hence it is OK for the transaction change counter to be
3900 case SQLITE_FCNTL_DB_UNCHANGED
: {
3901 ((unixFile
*)id
)->dbUpdate
= 0;
3905 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3906 case SQLITE_FCNTL_SET_LOCKPROXYFILE
:
3907 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
3908 return proxyFileControl(id
,op
,pArg
);
3910 #endif /* SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__) */
3912 return SQLITE_NOTFOUND
;
3916 ** If pFd->sectorSize is non-zero when this function is called, it is a
3917 ** no-op. Otherwise, the values of pFd->sectorSize and
3918 ** pFd->deviceCharacteristics are set according to the file-system
3921 ** There are two versions of this function. One for QNX and one for all
3925 static void setDeviceCharacteristics(unixFile
*pFd
){
3926 assert( pFd
->deviceCharacteristics
==0 || pFd
->sectorSize
!=0 );
3927 if( pFd
->sectorSize
==0 ){
3928 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3932 /* Check for support for F2FS atomic batch writes. */
3933 res
= osIoctl(pFd
->h
, F2FS_IOC_GET_FEATURES
, &f
);
3934 if( res
==0 && (f
& F2FS_FEATURE_ATOMIC_WRITE
) ){
3935 pFd
->deviceCharacteristics
= SQLITE_IOCAP_BATCH_ATOMIC
;
3937 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3939 /* Set the POWERSAFE_OVERWRITE flag if requested. */
3940 if( pFd
->ctrlFlags
& UNIXFILE_PSOW
){
3941 pFd
->deviceCharacteristics
|= SQLITE_IOCAP_POWERSAFE_OVERWRITE
;
3944 pFd
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
3948 #include <sys/dcmd_blk.h>
3949 #include <sys/statvfs.h>
3950 static void setDeviceCharacteristics(unixFile
*pFile
){
3951 if( pFile
->sectorSize
== 0 ){
3952 struct statvfs fsInfo
;
3954 /* Set defaults for non-supported filesystems */
3955 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
3956 pFile
->deviceCharacteristics
= 0;
3957 if( fstatvfs(pFile
->h
, &fsInfo
) == -1 ) {
3961 if( !strcmp(fsInfo
.f_basetype
, "tmp") ) {
3962 pFile
->sectorSize
= fsInfo
.f_bsize
;
3963 pFile
->deviceCharacteristics
=
3964 SQLITE_IOCAP_ATOMIC4K
| /* All ram filesystem writes are atomic */
3965 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3966 ** the write succeeds */
3967 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3968 ** so it is ordered */
3970 }else if( strstr(fsInfo
.f_basetype
, "etfs") ){
3971 pFile
->sectorSize
= fsInfo
.f_bsize
;
3972 pFile
->deviceCharacteristics
=
3973 /* etfs cluster size writes are atomic */
3974 (pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) |
3975 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3976 ** the write succeeds */
3977 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3978 ** so it is ordered */
3980 }else if( !strcmp(fsInfo
.f_basetype
, "qnx6") ){
3981 pFile
->sectorSize
= fsInfo
.f_bsize
;
3982 pFile
->deviceCharacteristics
=
3983 SQLITE_IOCAP_ATOMIC
| /* All filesystem writes are atomic */
3984 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3985 ** the write succeeds */
3986 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3987 ** so it is ordered */
3989 }else if( !strcmp(fsInfo
.f_basetype
, "qnx4") ){
3990 pFile
->sectorSize
= fsInfo
.f_bsize
;
3991 pFile
->deviceCharacteristics
=
3992 /* full bitset of atomics from max sector size and smaller */
3993 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
3994 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3995 ** so it is ordered */
3997 }else if( strstr(fsInfo
.f_basetype
, "dos") ){
3998 pFile
->sectorSize
= fsInfo
.f_bsize
;
3999 pFile
->deviceCharacteristics
=
4000 /* full bitset of atomics from max sector size and smaller */
4001 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
4002 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4003 ** so it is ordered */
4006 pFile
->deviceCharacteristics
=
4007 SQLITE_IOCAP_ATOMIC512
| /* blocks are atomic */
4008 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4009 ** the write succeeds */
4013 /* Last chance verification. If the sector size isn't a multiple of 512
4014 ** then it isn't valid.*/
4015 if( pFile
->sectorSize
% 512 != 0 ){
4016 pFile
->deviceCharacteristics
= 0;
4017 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
4023 ** Return the sector size in bytes of the underlying block device for
4024 ** the specified file. This is almost always 512 bytes, but may be
4025 ** larger for some devices.
4027 ** SQLite code assumes this function cannot fail. It also assumes that
4028 ** if two files are created in the same file-system directory (i.e.
4029 ** a database and its journal file) that the sector size will be the
4032 static int unixSectorSize(sqlite3_file
*id
){
4033 unixFile
*pFd
= (unixFile
*)id
;
4034 setDeviceCharacteristics(pFd
);
4035 return pFd
->sectorSize
;
4039 ** Return the device characteristics for the file.
4041 ** This VFS is set up to return SQLITE_IOCAP_POWERSAFE_OVERWRITE by default.
4042 ** However, that choice is controversial since technically the underlying
4043 ** file system does not always provide powersafe overwrites. (In other
4044 ** words, after a power-loss event, parts of the file that were never
4045 ** written might end up being altered.) However, non-PSOW behavior is very,
4046 ** very rare. And asserting PSOW makes a large reduction in the amount
4047 ** of required I/O for journaling, since a lot of padding is eliminated.
4048 ** Hence, while POWERSAFE_OVERWRITE is on by default, there is a file-control
4049 ** available to turn it off and URI query parameter available to turn it off.
4051 static int unixDeviceCharacteristics(sqlite3_file
*id
){
4052 unixFile
*pFd
= (unixFile
*)id
;
4053 setDeviceCharacteristics(pFd
);
4054 return pFd
->deviceCharacteristics
;
4057 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
4060 ** Return the system page size.
4062 ** This function should not be called directly by other code in this file.
4063 ** Instead, it should be called via macro osGetpagesize().
4065 static int unixGetpagesize(void){
4068 #elif defined(_BSD_SOURCE)
4069 return getpagesize();
4071 return (int)sysconf(_SC_PAGESIZE
);
4075 #endif /* !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0 */
4077 #ifndef SQLITE_OMIT_WAL
4080 ** Object used to represent an shared memory buffer.
4082 ** When multiple threads all reference the same wal-index, each thread
4083 ** has its own unixShm object, but they all point to a single instance
4084 ** of this unixShmNode object. In other words, each wal-index is opened
4085 ** only once per process.
4087 ** Each unixShmNode object is connected to a single unixInodeInfo object.
4088 ** We could coalesce this object into unixInodeInfo, but that would mean
4089 ** every open file that does not use shared memory (in other words, most
4090 ** open files) would have to carry around this extra information. So
4091 ** the unixInodeInfo object contains a pointer to this unixShmNode object
4092 ** and the unixShmNode object is created only when needed.
4094 ** unixMutexHeld() must be true when creating or destroying
4095 ** this object or while reading or writing the following fields:
4099 ** The following fields are read-only after the object is created:
4104 ** Either unixShmNode.mutex must be held or unixShmNode.nRef==0 and
4105 ** unixMutexHeld() is true when reading or writing any other field
4106 ** in this structure.
4108 struct unixShmNode
{
4109 unixInodeInfo
*pInode
; /* unixInodeInfo that owns this SHM node */
4110 sqlite3_mutex
*mutex
; /* Mutex to access this object */
4111 char *zFilename
; /* Name of the mmapped file */
4112 int h
; /* Open file descriptor */
4113 int szRegion
; /* Size of shared-memory regions */
4114 u16 nRegion
; /* Size of array apRegion */
4115 u8 isReadonly
; /* True if read-only */
4116 u8 isUnlocked
; /* True if no DMS lock held */
4117 char **apRegion
; /* Array of mapped shared-memory regions */
4118 int nRef
; /* Number of unixShm objects pointing to this */
4119 unixShm
*pFirst
; /* All unixShm objects pointing to this */
4121 u8 exclMask
; /* Mask of exclusive locks held */
4122 u8 sharedMask
; /* Mask of shared locks held */
4123 u8 nextShmId
; /* Next available unixShm.id value */
4128 ** Structure used internally by this VFS to record the state of an
4129 ** open shared memory connection.
4131 ** The following fields are initialized when this object is created and
4132 ** are read-only thereafter:
4137 ** All other fields are read/write. The unixShm.pFile->mutex must be held
4138 ** while accessing any read/write fields.
4141 unixShmNode
*pShmNode
; /* The underlying unixShmNode object */
4142 unixShm
*pNext
; /* Next unixShm with the same unixShmNode */
4143 u8 hasMutex
; /* True if holding the unixShmNode mutex */
4144 u8 id
; /* Id of this connection within its unixShmNode */
4145 u16 sharedMask
; /* Mask of shared locks held */
4146 u16 exclMask
; /* Mask of exclusive locks held */
4150 ** Constants used for locking
4152 #define UNIX_SHM_BASE ((22+SQLITE_SHM_NLOCK)*4) /* first lock byte */
4153 #define UNIX_SHM_DMS (UNIX_SHM_BASE+SQLITE_SHM_NLOCK) /* deadman switch */
4156 ** Apply posix advisory locks for all bytes from ofst through ofst+n-1.
4158 ** Locks block if the mask is exactly UNIX_SHM_C and are non-blocking
4161 static int unixShmSystemLock(
4162 unixFile
*pFile
, /* Open connection to the WAL file */
4163 int lockType
, /* F_UNLCK, F_RDLCK, or F_WRLCK */
4164 int ofst
, /* First byte of the locking range */
4165 int n
/* Number of bytes to lock */
4167 unixShmNode
*pShmNode
; /* Apply locks to this open shared-memory segment */
4168 struct flock f
; /* The posix advisory locking structure */
4169 int rc
= SQLITE_OK
; /* Result code form fcntl() */
4171 /* Access to the unixShmNode object is serialized by the caller */
4172 pShmNode
= pFile
->pInode
->pShmNode
;
4173 assert( pShmNode
->nRef
==0 || sqlite3_mutex_held(pShmNode
->mutex
) );
4175 /* Shared locks never span more than one byte */
4176 assert( n
==1 || lockType
!=F_RDLCK
);
4178 /* Locks are within range */
4179 assert( n
>=1 && n
<=SQLITE_SHM_NLOCK
);
4181 if( pShmNode
->h
>=0 ){
4182 /* Initialize the locking parameters */
4183 memset(&f
, 0, sizeof(f
));
4184 f
.l_type
= lockType
;
4185 f
.l_whence
= SEEK_SET
;
4189 rc
= osFcntl(pShmNode
->h
, F_SETLK
, &f
);
4190 rc
= (rc
!=(-1)) ? SQLITE_OK
: SQLITE_BUSY
;
4193 /* Update the global lock state and do debug tracing */
4196 OSTRACE(("SHM-LOCK "));
4197 mask
= ofst
>31 ? 0xffff : (1<<(ofst
+n
)) - (1<<ofst
);
4198 if( rc
==SQLITE_OK
){
4199 if( lockType
==F_UNLCK
){
4200 OSTRACE(("unlock %d ok", ofst
));
4201 pShmNode
->exclMask
&= ~mask
;
4202 pShmNode
->sharedMask
&= ~mask
;
4203 }else if( lockType
==F_RDLCK
){
4204 OSTRACE(("read-lock %d ok", ofst
));
4205 pShmNode
->exclMask
&= ~mask
;
4206 pShmNode
->sharedMask
|= mask
;
4208 assert( lockType
==F_WRLCK
);
4209 OSTRACE(("write-lock %d ok", ofst
));
4210 pShmNode
->exclMask
|= mask
;
4211 pShmNode
->sharedMask
&= ~mask
;
4214 if( lockType
==F_UNLCK
){
4215 OSTRACE(("unlock %d failed", ofst
));
4216 }else if( lockType
==F_RDLCK
){
4217 OSTRACE(("read-lock failed"));
4219 assert( lockType
==F_WRLCK
);
4220 OSTRACE(("write-lock %d failed", ofst
));
4223 OSTRACE((" - afterwards %03x,%03x\n",
4224 pShmNode
->sharedMask
, pShmNode
->exclMask
));
4232 ** Return the minimum number of 32KB shm regions that should be mapped at
4233 ** a time, assuming that each mapping must be an integer multiple of the
4234 ** current system page-size.
4236 ** Usually, this is 1. The exception seems to be systems that are configured
4237 ** to use 64KB pages - in this case each mapping must cover at least two
4240 static int unixShmRegionPerMap(void){
4241 int shmsz
= 32*1024; /* SHM region size */
4242 int pgsz
= osGetpagesize(); /* System page size */
4243 assert( ((pgsz
-1)&pgsz
)==0 ); /* Page size must be a power of 2 */
4244 if( pgsz
<shmsz
) return 1;
4249 ** Purge the unixShmNodeList list of all entries with unixShmNode.nRef==0.
4251 ** This is not a VFS shared-memory method; it is a utility function called
4252 ** by VFS shared-memory methods.
4254 static void unixShmPurge(unixFile
*pFd
){
4255 unixShmNode
*p
= pFd
->pInode
->pShmNode
;
4256 assert( unixMutexHeld() );
4257 if( p
&& ALWAYS(p
->nRef
==0) ){
4258 int nShmPerMap
= unixShmRegionPerMap();
4260 assert( p
->pInode
==pFd
->pInode
);
4261 sqlite3_mutex_free(p
->mutex
);
4262 for(i
=0; i
<p
->nRegion
; i
+=nShmPerMap
){
4264 osMunmap(p
->apRegion
[i
], p
->szRegion
);
4266 sqlite3_free(p
->apRegion
[i
]);
4269 sqlite3_free(p
->apRegion
);
4271 robust_close(pFd
, p
->h
, __LINE__
);
4274 p
->pInode
->pShmNode
= 0;
4280 ** The DMS lock has not yet been taken on shm file pShmNode. Attempt to
4281 ** take it now. Return SQLITE_OK if successful, or an SQLite error
4284 ** If the DMS cannot be locked because this is a readonly_shm=1
4285 ** connection and no other process already holds a lock, return
4286 ** SQLITE_READONLY_CANTINIT and set pShmNode->isUnlocked=1.
4288 static int unixLockSharedMemory(unixFile
*pDbFd
, unixShmNode
*pShmNode
){
4292 /* Use F_GETLK to determine the locks other processes are holding
4293 ** on the DMS byte. If it indicates that another process is holding
4294 ** a SHARED lock, then this process may also take a SHARED lock
4295 ** and proceed with opening the *-shm file.
4297 ** Or, if no other process is holding any lock, then this process
4298 ** is the first to open it. In this case take an EXCLUSIVE lock on the
4299 ** DMS byte and truncate the *-shm file to zero bytes in size. Then
4300 ** downgrade to a SHARED lock on the DMS byte.
4302 ** If another process is holding an EXCLUSIVE lock on the DMS byte,
4303 ** return SQLITE_BUSY to the caller (it will try again). An earlier
4304 ** version of this code attempted the SHARED lock at this point. But
4305 ** this introduced a subtle race condition: if the process holding
4306 ** EXCLUSIVE failed just before truncating the *-shm file, then this
4307 ** process might open and use the *-shm file without truncating it.
4308 ** And if the *-shm file has been corrupted by a power failure or
4309 ** system crash, the database itself may also become corrupt. */
4310 lock
.l_whence
= SEEK_SET
;
4311 lock
.l_start
= UNIX_SHM_DMS
;
4313 lock
.l_type
= F_WRLCK
;
4314 if( osFcntl(pShmNode
->h
, F_GETLK
, &lock
)!=0 ) {
4315 rc
= SQLITE_IOERR_LOCK
;
4316 }else if( lock
.l_type
==F_UNLCK
){
4317 if( pShmNode
->isReadonly
){
4318 pShmNode
->isUnlocked
= 1;
4319 rc
= SQLITE_READONLY_CANTINIT
;
4321 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, UNIX_SHM_DMS
, 1);
4322 if( rc
==SQLITE_OK
&& robust_ftruncate(pShmNode
->h
, 0) ){
4323 rc
= unixLogError(SQLITE_IOERR_SHMOPEN
,"ftruncate",pShmNode
->zFilename
);
4326 }else if( lock
.l_type
==F_WRLCK
){
4330 if( rc
==SQLITE_OK
){
4331 assert( lock
.l_type
==F_UNLCK
|| lock
.l_type
==F_RDLCK
);
4332 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, UNIX_SHM_DMS
, 1);
4338 ** Open a shared-memory area associated with open database file pDbFd.
4339 ** This particular implementation uses mmapped files.
4341 ** The file used to implement shared-memory is in the same directory
4342 ** as the open database file and has the same name as the open database
4343 ** file with the "-shm" suffix added. For example, if the database file
4344 ** is "/home/user1/config.db" then the file that is created and mmapped
4345 ** for shared memory will be called "/home/user1/config.db-shm".
4347 ** Another approach to is to use files in /dev/shm or /dev/tmp or an
4348 ** some other tmpfs mount. But if a file in a different directory
4349 ** from the database file is used, then differing access permissions
4350 ** or a chroot() might cause two different processes on the same
4351 ** database to end up using different files for shared memory -
4352 ** meaning that their memory would not really be shared - resulting
4353 ** in database corruption. Nevertheless, this tmpfs file usage
4354 ** can be enabled at compile-time using -DSQLITE_SHM_DIRECTORY="/dev/shm"
4355 ** or the equivalent. The use of the SQLITE_SHM_DIRECTORY compile-time
4356 ** option results in an incompatible build of SQLite; builds of SQLite
4357 ** that with differing SQLITE_SHM_DIRECTORY settings attempt to use the
4358 ** same database file at the same time, database corruption will likely
4359 ** result. The SQLITE_SHM_DIRECTORY compile-time option is considered
4360 ** "unsupported" and may go away in a future SQLite release.
4362 ** When opening a new shared-memory file, if no other instances of that
4363 ** file are currently open, in this process or in other processes, then
4364 ** the file must be truncated to zero length or have its header cleared.
4366 ** If the original database file (pDbFd) is using the "unix-excl" VFS
4367 ** that means that an exclusive lock is held on the database file and
4368 ** that no other processes are able to read or write the database. In
4369 ** that case, we do not really need shared memory. No shared memory
4370 ** file is created. The shared memory will be simulated with heap memory.
4372 static int unixOpenSharedMemory(unixFile
*pDbFd
){
4373 struct unixShm
*p
= 0; /* The connection to be opened */
4374 struct unixShmNode
*pShmNode
; /* The underlying mmapped file */
4375 int rc
= SQLITE_OK
; /* Result code */
4376 unixInodeInfo
*pInode
; /* The inode of fd */
4377 char *zShm
; /* Name of the file used for SHM */
4378 int nShmFilename
; /* Size of the SHM filename in bytes */
4380 /* Allocate space for the new unixShm object. */
4381 p
= sqlite3_malloc64( sizeof(*p
) );
4382 if( p
==0 ) return SQLITE_NOMEM_BKPT
;
4383 memset(p
, 0, sizeof(*p
));
4384 assert( pDbFd
->pShm
==0 );
4386 /* Check to see if a unixShmNode object already exists. Reuse an existing
4387 ** one if present. Create a new one if necessary.
4390 pInode
= pDbFd
->pInode
;
4391 pShmNode
= pInode
->pShmNode
;
4393 struct stat sStat
; /* fstat() info for database file */
4394 #ifndef SQLITE_SHM_DIRECTORY
4395 const char *zBasePath
= pDbFd
->zPath
;
4398 /* Call fstat() to figure out the permissions on the database file. If
4399 ** a new *-shm file is created, an attempt will be made to create it
4400 ** with the same permissions.
4402 if( osFstat(pDbFd
->h
, &sStat
) ){
4403 rc
= SQLITE_IOERR_FSTAT
;
4407 #ifdef SQLITE_SHM_DIRECTORY
4408 nShmFilename
= sizeof(SQLITE_SHM_DIRECTORY
) + 31;
4410 nShmFilename
= 6 + (int)strlen(zBasePath
);
4412 pShmNode
= sqlite3_malloc64( sizeof(*pShmNode
) + nShmFilename
);
4414 rc
= SQLITE_NOMEM_BKPT
;
4417 memset(pShmNode
, 0, sizeof(*pShmNode
)+nShmFilename
);
4418 zShm
= pShmNode
->zFilename
= (char*)&pShmNode
[1];
4419 #ifdef SQLITE_SHM_DIRECTORY
4420 sqlite3_snprintf(nShmFilename
, zShm
,
4421 SQLITE_SHM_DIRECTORY
"/sqlite-shm-%x-%x",
4422 (u32
)sStat
.st_ino
, (u32
)sStat
.st_dev
);
4424 sqlite3_snprintf(nShmFilename
, zShm
, "%s-shm", zBasePath
);
4425 sqlite3FileSuffix3(pDbFd
->zPath
, zShm
);
4428 pDbFd
->pInode
->pShmNode
= pShmNode
;
4429 pShmNode
->pInode
= pDbFd
->pInode
;
4430 if( sqlite3GlobalConfig
.bCoreMutex
){
4431 pShmNode
->mutex
= sqlite3_mutex_alloc(SQLITE_MUTEX_FAST
);
4432 if( pShmNode
->mutex
==0 ){
4433 rc
= SQLITE_NOMEM_BKPT
;
4438 if( pInode
->bProcessLock
==0 ){
4439 if( 0==sqlite3_uri_boolean(pDbFd
->zPath
, "readonly_shm", 0) ){
4440 pShmNode
->h
= robust_open(zShm
, O_RDWR
|O_CREAT
, (sStat
.st_mode
&0777));
4442 if( pShmNode
->h
<0 ){
4443 pShmNode
->h
= robust_open(zShm
, O_RDONLY
, (sStat
.st_mode
&0777));
4444 if( pShmNode
->h
<0 ){
4445 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zShm
);
4448 pShmNode
->isReadonly
= 1;
4451 /* If this process is running as root, make sure that the SHM file
4452 ** is owned by the same user that owns the original database. Otherwise,
4453 ** the original owner will not be able to connect.
4455 robustFchown(pShmNode
->h
, sStat
.st_uid
, sStat
.st_gid
);
4457 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4458 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_READONLY_CANTINIT
) goto shm_open_err
;
4462 /* Make the new connection a child of the unixShmNode */
4463 p
->pShmNode
= pShmNode
;
4465 p
->id
= pShmNode
->nextShmId
++;
4471 /* The reference count on pShmNode has already been incremented under
4472 ** the cover of the unixEnterMutex() mutex and the pointer from the
4473 ** new (struct unixShm) object to the pShmNode has been set. All that is
4474 ** left to do is to link the new object into the linked list starting
4475 ** at pShmNode->pFirst. This must be done while holding the pShmNode->mutex
4478 sqlite3_mutex_enter(pShmNode
->mutex
);
4479 p
->pNext
= pShmNode
->pFirst
;
4480 pShmNode
->pFirst
= p
;
4481 sqlite3_mutex_leave(pShmNode
->mutex
);
4484 /* Jump here on any error */
4486 unixShmPurge(pDbFd
); /* This call frees pShmNode if required */
4493 ** This function is called to obtain a pointer to region iRegion of the
4494 ** shared-memory associated with the database file fd. Shared-memory regions
4495 ** are numbered starting from zero. Each shared-memory region is szRegion
4498 ** If an error occurs, an error code is returned and *pp is set to NULL.
4500 ** Otherwise, if the bExtend parameter is 0 and the requested shared-memory
4501 ** region has not been allocated (by any client, including one running in a
4502 ** separate process), then *pp is set to NULL and SQLITE_OK returned. If
4503 ** bExtend is non-zero and the requested shared-memory region has not yet
4504 ** been allocated, it is allocated by this function.
4506 ** If the shared-memory region has already been allocated or is allocated by
4507 ** this call as described above, then it is mapped into this processes
4508 ** address space (if it is not already), *pp is set to point to the mapped
4509 ** memory and SQLITE_OK returned.
4511 static int unixShmMap(
4512 sqlite3_file
*fd
, /* Handle open on database file */
4513 int iRegion
, /* Region to retrieve */
4514 int szRegion
, /* Size of regions */
4515 int bExtend
, /* True to extend file if necessary */
4516 void volatile **pp
/* OUT: Mapped memory */
4518 unixFile
*pDbFd
= (unixFile
*)fd
;
4520 unixShmNode
*pShmNode
;
4522 int nShmPerMap
= unixShmRegionPerMap();
4525 /* If the shared-memory file has not yet been opened, open it now. */
4526 if( pDbFd
->pShm
==0 ){
4527 rc
= unixOpenSharedMemory(pDbFd
);
4528 if( rc
!=SQLITE_OK
) return rc
;
4532 pShmNode
= p
->pShmNode
;
4533 sqlite3_mutex_enter(pShmNode
->mutex
);
4534 if( pShmNode
->isUnlocked
){
4535 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4536 if( rc
!=SQLITE_OK
) goto shmpage_out
;
4537 pShmNode
->isUnlocked
= 0;
4539 assert( szRegion
==pShmNode
->szRegion
|| pShmNode
->nRegion
==0 );
4540 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4541 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4542 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4544 /* Minimum number of regions required to be mapped. */
4545 nReqRegion
= ((iRegion
+nShmPerMap
) / nShmPerMap
) * nShmPerMap
;
4547 if( pShmNode
->nRegion
<nReqRegion
){
4548 char **apNew
; /* New apRegion[] array */
4549 int nByte
= nReqRegion
*szRegion
; /* Minimum required file size */
4550 struct stat sStat
; /* Used by fstat() */
4552 pShmNode
->szRegion
= szRegion
;
4554 if( pShmNode
->h
>=0 ){
4555 /* The requested region is not mapped into this processes address space.
4556 ** Check to see if it has been allocated (i.e. if the wal-index file is
4557 ** large enough to contain the requested region).
4559 if( osFstat(pShmNode
->h
, &sStat
) ){
4560 rc
= SQLITE_IOERR_SHMSIZE
;
4564 if( sStat
.st_size
<nByte
){
4565 /* The requested memory region does not exist. If bExtend is set to
4566 ** false, exit early. *pp will be set to NULL and SQLITE_OK returned.
4572 /* Alternatively, if bExtend is true, extend the file. Do this by
4573 ** writing a single byte to the end of each (OS) page being
4574 ** allocated or extended. Technically, we need only write to the
4575 ** last page in order to extend the file. But writing to all new
4576 ** pages forces the OS to allocate them immediately, which reduces
4577 ** the chances of SIGBUS while accessing the mapped region later on.
4580 static const int pgsz
= 4096;
4583 /* Write to the last byte of each newly allocated or extended page */
4584 assert( (nByte
% pgsz
)==0 );
4585 for(iPg
=(sStat
.st_size
/pgsz
); iPg
<(nByte
/pgsz
); iPg
++){
4587 if( seekAndWriteFd(pShmNode
->h
, iPg
*pgsz
+ pgsz
-1, "", 1, &x
)!=1 ){
4588 const char *zFile
= pShmNode
->zFilename
;
4589 rc
= unixLogError(SQLITE_IOERR_SHMSIZE
, "write", zFile
);
4597 /* Map the requested memory region into this processes address space. */
4598 apNew
= (char **)sqlite3_realloc(
4599 pShmNode
->apRegion
, nReqRegion
*sizeof(char *)
4602 rc
= SQLITE_IOERR_NOMEM_BKPT
;
4605 pShmNode
->apRegion
= apNew
;
4606 while( pShmNode
->nRegion
<nReqRegion
){
4607 int nMap
= szRegion
*nShmPerMap
;
4610 if( pShmNode
->h
>=0 ){
4611 pMem
= osMmap(0, nMap
,
4612 pShmNode
->isReadonly
? PROT_READ
: PROT_READ
|PROT_WRITE
,
4613 MAP_SHARED
, pShmNode
->h
, szRegion
*(i64
)pShmNode
->nRegion
4615 if( pMem
==MAP_FAILED
){
4616 rc
= unixLogError(SQLITE_IOERR_SHMMAP
, "mmap", pShmNode
->zFilename
);
4620 pMem
= sqlite3_malloc64(szRegion
);
4622 rc
= SQLITE_NOMEM_BKPT
;
4625 memset(pMem
, 0, szRegion
);
4628 for(i
=0; i
<nShmPerMap
; i
++){
4629 pShmNode
->apRegion
[pShmNode
->nRegion
+i
] = &((char*)pMem
)[szRegion
*i
];
4631 pShmNode
->nRegion
+= nShmPerMap
;
4636 if( pShmNode
->nRegion
>iRegion
){
4637 *pp
= pShmNode
->apRegion
[iRegion
];
4641 if( pShmNode
->isReadonly
&& rc
==SQLITE_OK
) rc
= SQLITE_READONLY
;
4642 sqlite3_mutex_leave(pShmNode
->mutex
);
4647 ** Change the lock state for a shared-memory segment.
4649 ** Note that the relationship between SHAREd and EXCLUSIVE locks is a little
4650 ** different here than in posix. In xShmLock(), one can go from unlocked
4651 ** to shared and back or from unlocked to exclusive and back. But one may
4652 ** not go from shared to exclusive or from exclusive to shared.
4654 static int unixShmLock(
4655 sqlite3_file
*fd
, /* Database file holding the shared memory */
4656 int ofst
, /* First lock to acquire or release */
4657 int n
, /* Number of locks to acquire or release */
4658 int flags
/* What to do with the lock */
4660 unixFile
*pDbFd
= (unixFile
*)fd
; /* Connection holding shared memory */
4661 unixShm
*p
= pDbFd
->pShm
; /* The shared memory being locked */
4662 unixShm
*pX
; /* For looping over all siblings */
4663 unixShmNode
*pShmNode
= p
->pShmNode
; /* The underlying file iNode */
4664 int rc
= SQLITE_OK
; /* Result code */
4665 u16 mask
; /* Mask of locks to take or release */
4667 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4668 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4669 assert( ofst
>=0 && ofst
+n
<=SQLITE_SHM_NLOCK
);
4671 assert( flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_SHARED
)
4672 || flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_EXCLUSIVE
)
4673 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_SHARED
)
4674 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_EXCLUSIVE
) );
4675 assert( n
==1 || (flags
& SQLITE_SHM_EXCLUSIVE
)!=0 );
4676 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4677 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4679 mask
= (1<<(ofst
+n
)) - (1<<ofst
);
4680 assert( n
>1 || mask
==(1<<ofst
) );
4681 sqlite3_mutex_enter(pShmNode
->mutex
);
4682 if( flags
& SQLITE_SHM_UNLOCK
){
4683 u16 allMask
= 0; /* Mask of locks held by siblings */
4685 /* See if any siblings hold this same lock */
4686 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4687 if( pX
==p
) continue;
4688 assert( (pX
->exclMask
& (p
->exclMask
|p
->sharedMask
))==0 );
4689 allMask
|= pX
->sharedMask
;
4692 /* Unlock the system-level locks */
4693 if( (mask
& allMask
)==0 ){
4694 rc
= unixShmSystemLock(pDbFd
, F_UNLCK
, ofst
+UNIX_SHM_BASE
, n
);
4699 /* Undo the local locks */
4700 if( rc
==SQLITE_OK
){
4701 p
->exclMask
&= ~mask
;
4702 p
->sharedMask
&= ~mask
;
4704 }else if( flags
& SQLITE_SHM_SHARED
){
4705 u16 allShared
= 0; /* Union of locks held by connections other than "p" */
4707 /* Find out which shared locks are already held by sibling connections.
4708 ** If any sibling already holds an exclusive lock, go ahead and return
4711 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4712 if( (pX
->exclMask
& mask
)!=0 ){
4716 allShared
|= pX
->sharedMask
;
4719 /* Get shared locks at the system level, if necessary */
4720 if( rc
==SQLITE_OK
){
4721 if( (allShared
& mask
)==0 ){
4722 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, ofst
+UNIX_SHM_BASE
, n
);
4728 /* Get the local shared locks */
4729 if( rc
==SQLITE_OK
){
4730 p
->sharedMask
|= mask
;
4733 /* Make sure no sibling connections hold locks that will block this
4734 ** lock. If any do, return SQLITE_BUSY right away.
4736 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4737 if( (pX
->exclMask
& mask
)!=0 || (pX
->sharedMask
& mask
)!=0 ){
4743 /* Get the exclusive locks at the system level. Then if successful
4744 ** also mark the local connection as being locked.
4746 if( rc
==SQLITE_OK
){
4747 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, ofst
+UNIX_SHM_BASE
, n
);
4748 if( rc
==SQLITE_OK
){
4749 assert( (p
->sharedMask
& mask
)==0 );
4750 p
->exclMask
|= mask
;
4754 sqlite3_mutex_leave(pShmNode
->mutex
);
4755 OSTRACE(("SHM-LOCK shmid-%d, pid-%d got %03x,%03x\n",
4756 p
->id
, osGetpid(0), p
->sharedMask
, p
->exclMask
));
4761 ** Implement a memory barrier or memory fence on shared memory.
4763 ** All loads and stores begun before the barrier must complete before
4764 ** any load or store begun after the barrier.
4766 static void unixShmBarrier(
4767 sqlite3_file
*fd
/* Database file holding the shared memory */
4769 UNUSED_PARAMETER(fd
);
4770 sqlite3MemoryBarrier(); /* compiler-defined memory barrier */
4771 unixEnterMutex(); /* Also mutex, for redundancy */
4776 ** Close a connection to shared-memory. Delete the underlying
4777 ** storage if deleteFlag is true.
4779 ** If there is no shared memory associated with the connection then this
4780 ** routine is a harmless no-op.
4782 static int unixShmUnmap(
4783 sqlite3_file
*fd
, /* The underlying database file */
4784 int deleteFlag
/* Delete shared-memory if true */
4786 unixShm
*p
; /* The connection to be closed */
4787 unixShmNode
*pShmNode
; /* The underlying shared-memory file */
4788 unixShm
**pp
; /* For looping over sibling connections */
4789 unixFile
*pDbFd
; /* The underlying database file */
4791 pDbFd
= (unixFile
*)fd
;
4793 if( p
==0 ) return SQLITE_OK
;
4794 pShmNode
= p
->pShmNode
;
4796 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4797 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4799 /* Remove connection p from the set of connections associated
4801 sqlite3_mutex_enter(pShmNode
->mutex
);
4802 for(pp
=&pShmNode
->pFirst
; (*pp
)!=p
; pp
= &(*pp
)->pNext
){}
4805 /* Free the connection p */
4808 sqlite3_mutex_leave(pShmNode
->mutex
);
4810 /* If pShmNode->nRef has reached 0, then close the underlying
4811 ** shared-memory file, too */
4813 assert( pShmNode
->nRef
>0 );
4815 if( pShmNode
->nRef
==0 ){
4816 if( deleteFlag
&& pShmNode
->h
>=0 ){
4817 osUnlink(pShmNode
->zFilename
);
4819 unixShmPurge(pDbFd
);
4828 # define unixShmMap 0
4829 # define unixShmLock 0
4830 # define unixShmBarrier 0
4831 # define unixShmUnmap 0
4832 #endif /* #ifndef SQLITE_OMIT_WAL */
4834 #if SQLITE_MAX_MMAP_SIZE>0
4836 ** If it is currently memory mapped, unmap file pFd.
4838 static void unixUnmapfile(unixFile
*pFd
){
4839 assert( pFd
->nFetchOut
==0 );
4840 if( pFd
->pMapRegion
){
4841 osMunmap(pFd
->pMapRegion
, pFd
->mmapSizeActual
);
4842 pFd
->pMapRegion
= 0;
4844 pFd
->mmapSizeActual
= 0;
4849 ** Attempt to set the size of the memory mapping maintained by file
4850 ** descriptor pFd to nNew bytes. Any existing mapping is discarded.
4852 ** If successful, this function sets the following variables:
4854 ** unixFile.pMapRegion
4855 ** unixFile.mmapSize
4856 ** unixFile.mmapSizeActual
4858 ** If unsuccessful, an error message is logged via sqlite3_log() and
4859 ** the three variables above are zeroed. In this case SQLite should
4860 ** continue accessing the database using the xRead() and xWrite()
4863 static void unixRemapfile(
4864 unixFile
*pFd
, /* File descriptor object */
4865 i64 nNew
/* Required mapping size */
4867 const char *zErr
= "mmap";
4868 int h
= pFd
->h
; /* File descriptor open on db file */
4869 u8
*pOrig
= (u8
*)pFd
->pMapRegion
; /* Pointer to current file mapping */
4870 i64 nOrig
= pFd
->mmapSizeActual
; /* Size of pOrig region in bytes */
4871 u8
*pNew
= 0; /* Location of new mapping */
4872 int flags
= PROT_READ
; /* Flags to pass to mmap() */
4874 assert( pFd
->nFetchOut
==0 );
4875 assert( nNew
>pFd
->mmapSize
);
4876 assert( nNew
<=pFd
->mmapSizeMax
);
4878 assert( pFd
->mmapSizeActual
>=pFd
->mmapSize
);
4879 assert( MAP_FAILED
!=0 );
4881 #ifdef SQLITE_MMAP_READWRITE
4882 if( (pFd
->ctrlFlags
& UNIXFILE_RDONLY
)==0 ) flags
|= PROT_WRITE
;
4887 i64 nReuse
= pFd
->mmapSize
;
4889 const int szSyspage
= osGetpagesize();
4890 i64 nReuse
= (pFd
->mmapSize
& ~(szSyspage
-1));
4892 u8
*pReq
= &pOrig
[nReuse
];
4894 /* Unmap any pages of the existing mapping that cannot be reused. */
4895 if( nReuse
!=nOrig
){
4896 osMunmap(pReq
, nOrig
-nReuse
);
4900 pNew
= osMremap(pOrig
, nReuse
, nNew
, MREMAP_MAYMOVE
);
4903 pNew
= osMmap(pReq
, nNew
-nReuse
, flags
, MAP_SHARED
, h
, nReuse
);
4904 if( pNew
!=MAP_FAILED
){
4906 osMunmap(pNew
, nNew
- nReuse
);
4914 /* The attempt to extend the existing mapping failed. Free it. */
4915 if( pNew
==MAP_FAILED
|| pNew
==0 ){
4916 osMunmap(pOrig
, nReuse
);
4920 /* If pNew is still NULL, try to create an entirely new mapping. */
4922 pNew
= osMmap(0, nNew
, flags
, MAP_SHARED
, h
, 0);
4925 if( pNew
==MAP_FAILED
){
4928 unixLogError(SQLITE_OK
, zErr
, pFd
->zPath
);
4930 /* If the mmap() above failed, assume that all subsequent mmap() calls
4931 ** will probably fail too. Fall back to using xRead/xWrite exclusively
4933 pFd
->mmapSizeMax
= 0;
4935 pFd
->pMapRegion
= (void *)pNew
;
4936 pFd
->mmapSize
= pFd
->mmapSizeActual
= nNew
;
4940 ** Memory map or remap the file opened by file-descriptor pFd (if the file
4941 ** is already mapped, the existing mapping is replaced by the new). Or, if
4942 ** there already exists a mapping for this file, and there are still
4943 ** outstanding xFetch() references to it, this function is a no-op.
4945 ** If parameter nByte is non-negative, then it is the requested size of
4946 ** the mapping to create. Otherwise, if nByte is less than zero, then the
4947 ** requested size is the size of the file on disk. The actual size of the
4948 ** created mapping is either the requested size or the value configured
4949 ** using SQLITE_FCNTL_MMAP_LIMIT, whichever is smaller.
4951 ** SQLITE_OK is returned if no error occurs (even if the mapping is not
4952 ** recreated as a result of outstanding references) or an SQLite error
4955 static int unixMapfile(unixFile
*pFd
, i64 nMap
){
4956 assert( nMap
>=0 || pFd
->nFetchOut
==0 );
4957 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
4958 if( pFd
->nFetchOut
>0 ) return SQLITE_OK
;
4961 struct stat statbuf
; /* Low-level file information */
4962 if( osFstat(pFd
->h
, &statbuf
) ){
4963 return SQLITE_IOERR_FSTAT
;
4965 nMap
= statbuf
.st_size
;
4967 if( nMap
>pFd
->mmapSizeMax
){
4968 nMap
= pFd
->mmapSizeMax
;
4971 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
4972 if( nMap
!=pFd
->mmapSize
){
4973 unixRemapfile(pFd
, nMap
);
4978 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
4981 ** If possible, return a pointer to a mapping of file fd starting at offset
4982 ** iOff. The mapping must be valid for at least nAmt bytes.
4984 ** If such a pointer can be obtained, store it in *pp and return SQLITE_OK.
4985 ** Or, if one cannot but no error occurs, set *pp to 0 and return SQLITE_OK.
4986 ** Finally, if an error does occur, return an SQLite error code. The final
4987 ** value of *pp is undefined in this case.
4989 ** If this function does return a pointer, the caller must eventually
4990 ** release the reference by calling unixUnfetch().
4992 static int unixFetch(sqlite3_file
*fd
, i64 iOff
, int nAmt
, void **pp
){
4993 #if SQLITE_MAX_MMAP_SIZE>0
4994 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
4998 #if SQLITE_MAX_MMAP_SIZE>0
4999 if( pFd
->mmapSizeMax
>0 ){
5000 if( pFd
->pMapRegion
==0 ){
5001 int rc
= unixMapfile(pFd
, -1);
5002 if( rc
!=SQLITE_OK
) return rc
;
5004 if( pFd
->mmapSize
>= iOff
+nAmt
){
5005 *pp
= &((u8
*)pFd
->pMapRegion
)[iOff
];
5014 ** If the third argument is non-NULL, then this function releases a
5015 ** reference obtained by an earlier call to unixFetch(). The second
5016 ** argument passed to this function must be the same as the corresponding
5017 ** argument that was passed to the unixFetch() invocation.
5019 ** Or, if the third argument is NULL, then this function is being called
5020 ** to inform the VFS layer that, according to POSIX, any existing mapping
5021 ** may now be invalid and should be unmapped.
5023 static int unixUnfetch(sqlite3_file
*fd
, i64 iOff
, void *p
){
5024 #if SQLITE_MAX_MMAP_SIZE>0
5025 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
5026 UNUSED_PARAMETER(iOff
);
5028 /* If p==0 (unmap the entire file) then there must be no outstanding
5029 ** xFetch references. Or, if p!=0 (meaning it is an xFetch reference),
5030 ** then there must be at least one outstanding. */
5031 assert( (p
==0)==(pFd
->nFetchOut
==0) );
5033 /* If p!=0, it must match the iOff value. */
5034 assert( p
==0 || p
==&((u8
*)pFd
->pMapRegion
)[iOff
] );
5042 assert( pFd
->nFetchOut
>=0 );
5044 UNUSED_PARAMETER(fd
);
5045 UNUSED_PARAMETER(p
);
5046 UNUSED_PARAMETER(iOff
);
5052 ** Here ends the implementation of all sqlite3_file methods.
5054 ********************** End sqlite3_file Methods *******************************
5055 ******************************************************************************/
5058 ** This division contains definitions of sqlite3_io_methods objects that
5059 ** implement various file locking strategies. It also contains definitions
5060 ** of "finder" functions. A finder-function is used to locate the appropriate
5061 ** sqlite3_io_methods object for a particular database file. The pAppData
5062 ** field of the sqlite3_vfs VFS objects are initialized to be pointers to
5063 ** the correct finder-function for that VFS.
5065 ** Most finder functions return a pointer to a fixed sqlite3_io_methods
5066 ** object. The only interesting finder-function is autolockIoFinder, which
5067 ** looks at the filesystem type and tries to guess the best locking
5068 ** strategy from that.
5070 ** For finder-function F, two objects are created:
5072 ** (1) The real finder-function named "FImpt()".
5074 ** (2) A constant pointer to this function named just "F".
5077 ** A pointer to the F pointer is used as the pAppData value for VFS
5078 ** objects. We have to do this instead of letting pAppData point
5079 ** directly at the finder-function since C90 rules prevent a void*
5080 ** from be cast into a function pointer.
5083 ** Each instance of this macro generates two objects:
5085 ** * A constant sqlite3_io_methods object call METHOD that has locking
5086 ** methods CLOSE, LOCK, UNLOCK, CKRESLOCK.
5088 ** * An I/O method finder function called FINDER that returns a pointer
5089 ** to the METHOD object in the previous bullet.
5091 #define IOMETHODS(FINDER,METHOD,VERSION,CLOSE,LOCK,UNLOCK,CKLOCK,SHMMAP) \
5092 static const sqlite3_io_methods METHOD = { \
5093 VERSION, /* iVersion */ \
5094 CLOSE, /* xClose */ \
5095 unixRead, /* xRead */ \
5096 unixWrite, /* xWrite */ \
5097 unixTruncate, /* xTruncate */ \
5098 unixSync, /* xSync */ \
5099 unixFileSize, /* xFileSize */ \
5101 UNLOCK, /* xUnlock */ \
5102 CKLOCK, /* xCheckReservedLock */ \
5103 unixFileControl, /* xFileControl */ \
5104 unixSectorSize, /* xSectorSize */ \
5105 unixDeviceCharacteristics, /* xDeviceCapabilities */ \
5106 SHMMAP, /* xShmMap */ \
5107 unixShmLock, /* xShmLock */ \
5108 unixShmBarrier, /* xShmBarrier */ \
5109 unixShmUnmap, /* xShmUnmap */ \
5110 unixFetch, /* xFetch */ \
5111 unixUnfetch, /* xUnfetch */ \
5113 static const sqlite3_io_methods *FINDER##Impl(const char *z, unixFile *p){ \
5114 UNUSED_PARAMETER(z); UNUSED_PARAMETER(p); \
5117 static const sqlite3_io_methods *(*const FINDER)(const char*,unixFile *p) \
5121 ** Here are all of the sqlite3_io_methods objects for each of the
5122 ** locking strategies. Functions that return pointers to these methods
5123 ** are also created.
5126 posixIoFinder
, /* Finder function name */
5127 posixIoMethods
, /* sqlite3_io_methods object name */
5128 3, /* shared memory and mmap are enabled */
5129 unixClose
, /* xClose method */
5130 unixLock
, /* xLock method */
5131 unixUnlock
, /* xUnlock method */
5132 unixCheckReservedLock
, /* xCheckReservedLock method */
5133 unixShmMap
/* xShmMap method */
5136 nolockIoFinder
, /* Finder function name */
5137 nolockIoMethods
, /* sqlite3_io_methods object name */
5138 3, /* shared memory is disabled */
5139 nolockClose
, /* xClose method */
5140 nolockLock
, /* xLock method */
5141 nolockUnlock
, /* xUnlock method */
5142 nolockCheckReservedLock
, /* xCheckReservedLock method */
5143 0 /* xShmMap method */
5146 dotlockIoFinder
, /* Finder function name */
5147 dotlockIoMethods
, /* sqlite3_io_methods object name */
5148 1, /* shared memory is disabled */
5149 dotlockClose
, /* xClose method */
5150 dotlockLock
, /* xLock method */
5151 dotlockUnlock
, /* xUnlock method */
5152 dotlockCheckReservedLock
, /* xCheckReservedLock method */
5153 0 /* xShmMap method */
5156 #if SQLITE_ENABLE_LOCKING_STYLE
5158 flockIoFinder
, /* Finder function name */
5159 flockIoMethods
, /* sqlite3_io_methods object name */
5160 1, /* shared memory is disabled */
5161 flockClose
, /* xClose method */
5162 flockLock
, /* xLock method */
5163 flockUnlock
, /* xUnlock method */
5164 flockCheckReservedLock
, /* xCheckReservedLock method */
5165 0 /* xShmMap method */
5171 semIoFinder
, /* Finder function name */
5172 semIoMethods
, /* sqlite3_io_methods object name */
5173 1, /* shared memory is disabled */
5174 semXClose
, /* xClose method */
5175 semXLock
, /* xLock method */
5176 semXUnlock
, /* xUnlock method */
5177 semXCheckReservedLock
, /* xCheckReservedLock method */
5178 0 /* xShmMap method */
5182 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5184 afpIoFinder
, /* Finder function name */
5185 afpIoMethods
, /* sqlite3_io_methods object name */
5186 1, /* shared memory is disabled */
5187 afpClose
, /* xClose method */
5188 afpLock
, /* xLock method */
5189 afpUnlock
, /* xUnlock method */
5190 afpCheckReservedLock
, /* xCheckReservedLock method */
5191 0 /* xShmMap method */
5196 ** The proxy locking method is a "super-method" in the sense that it
5197 ** opens secondary file descriptors for the conch and lock files and
5198 ** it uses proxy, dot-file, AFP, and flock() locking methods on those
5199 ** secondary files. For this reason, the division that implements
5200 ** proxy locking is located much further down in the file. But we need
5201 ** to go ahead and define the sqlite3_io_methods and finder function
5202 ** for proxy locking here. So we forward declare the I/O methods.
5204 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5205 static int proxyClose(sqlite3_file
*);
5206 static int proxyLock(sqlite3_file
*, int);
5207 static int proxyUnlock(sqlite3_file
*, int);
5208 static int proxyCheckReservedLock(sqlite3_file
*, int*);
5210 proxyIoFinder
, /* Finder function name */
5211 proxyIoMethods
, /* sqlite3_io_methods object name */
5212 1, /* shared memory is disabled */
5213 proxyClose
, /* xClose method */
5214 proxyLock
, /* xLock method */
5215 proxyUnlock
, /* xUnlock method */
5216 proxyCheckReservedLock
, /* xCheckReservedLock method */
5217 0 /* xShmMap method */
5221 /* nfs lockd on OSX 10.3+ doesn't clear write locks when a read lock is set */
5222 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5224 nfsIoFinder
, /* Finder function name */
5225 nfsIoMethods
, /* sqlite3_io_methods object name */
5226 1, /* shared memory is disabled */
5227 unixClose
, /* xClose method */
5228 unixLock
, /* xLock method */
5229 nfsUnlock
, /* xUnlock method */
5230 unixCheckReservedLock
, /* xCheckReservedLock method */
5231 0 /* xShmMap method */
5235 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5237 ** This "finder" function attempts to determine the best locking strategy
5238 ** for the database file "filePath". It then returns the sqlite3_io_methods
5239 ** object that implements that strategy.
5241 ** This is for MacOSX only.
5243 static const sqlite3_io_methods
*autolockIoFinderImpl(
5244 const char *filePath
, /* name of the database file */
5245 unixFile
*pNew
/* open file object for the database file */
5247 static const struct Mapping
{
5248 const char *zFilesystem
; /* Filesystem type name */
5249 const sqlite3_io_methods
*pMethods
; /* Appropriate locking method */
5251 { "hfs", &posixIoMethods
},
5252 { "ufs", &posixIoMethods
},
5253 { "afpfs", &afpIoMethods
},
5254 { "smbfs", &afpIoMethods
},
5255 { "webdav", &nolockIoMethods
},
5259 struct statfs fsInfo
;
5260 struct flock lockInfo
;
5263 /* If filePath==NULL that means we are dealing with a transient file
5264 ** that does not need to be locked. */
5265 return &nolockIoMethods
;
5267 if( statfs(filePath
, &fsInfo
) != -1 ){
5268 if( fsInfo
.f_flags
& MNT_RDONLY
){
5269 return &nolockIoMethods
;
5271 for(i
=0; aMap
[i
].zFilesystem
; i
++){
5272 if( strcmp(fsInfo
.f_fstypename
, aMap
[i
].zFilesystem
)==0 ){
5273 return aMap
[i
].pMethods
;
5278 /* Default case. Handles, amongst others, "nfs".
5279 ** Test byte-range lock using fcntl(). If the call succeeds,
5280 ** assume that the file-system supports POSIX style locks.
5283 lockInfo
.l_start
= 0;
5284 lockInfo
.l_whence
= SEEK_SET
;
5285 lockInfo
.l_type
= F_RDLCK
;
5286 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5287 if( strcmp(fsInfo
.f_fstypename
, "nfs")==0 ){
5288 return &nfsIoMethods
;
5290 return &posixIoMethods
;
5293 return &dotlockIoMethods
;
5296 static const sqlite3_io_methods
5297 *(*const autolockIoFinder
)(const char*,unixFile
*) = autolockIoFinderImpl
;
5299 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
5303 ** This "finder" function for VxWorks checks to see if posix advisory
5304 ** locking works. If it does, then that is what is used. If it does not
5305 ** work, then fallback to named semaphore locking.
5307 static const sqlite3_io_methods
*vxworksIoFinderImpl(
5308 const char *filePath
, /* name of the database file */
5309 unixFile
*pNew
/* the open file object */
5311 struct flock lockInfo
;
5314 /* If filePath==NULL that means we are dealing with a transient file
5315 ** that does not need to be locked. */
5316 return &nolockIoMethods
;
5319 /* Test if fcntl() is supported and use POSIX style locks.
5320 ** Otherwise fall back to the named semaphore method.
5323 lockInfo
.l_start
= 0;
5324 lockInfo
.l_whence
= SEEK_SET
;
5325 lockInfo
.l_type
= F_RDLCK
;
5326 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5327 return &posixIoMethods
;
5329 return &semIoMethods
;
5332 static const sqlite3_io_methods
5333 *(*const vxworksIoFinder
)(const char*,unixFile
*) = vxworksIoFinderImpl
;
5335 #endif /* OS_VXWORKS */
5338 ** An abstract type for a pointer to an IO method finder function:
5340 typedef const sqlite3_io_methods
*(*finder_type
)(const char*,unixFile
*);
5343 /****************************************************************************
5344 **************************** sqlite3_vfs methods ****************************
5346 ** This division contains the implementation of methods on the
5347 ** sqlite3_vfs object.
5351 ** Initialize the contents of the unixFile structure pointed to by pId.
5353 static int fillInUnixFile(
5354 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
5355 int h
, /* Open file descriptor of file being opened */
5356 sqlite3_file
*pId
, /* Write to the unixFile structure here */
5357 const char *zFilename
, /* Name of the file being opened */
5358 int ctrlFlags
/* Zero or more UNIXFILE_* values */
5360 const sqlite3_io_methods
*pLockingStyle
;
5361 unixFile
*pNew
= (unixFile
*)pId
;
5364 assert( pNew
->pInode
==NULL
);
5366 /* No locking occurs in temporary files */
5367 assert( zFilename
!=0 || (ctrlFlags
& UNIXFILE_NOLOCK
)!=0 );
5369 OSTRACE(("OPEN %-3d %s\n", h
, zFilename
));
5372 pNew
->zPath
= zFilename
;
5373 pNew
->ctrlFlags
= (u8
)ctrlFlags
;
5374 #if SQLITE_MAX_MMAP_SIZE>0
5375 pNew
->mmapSizeMax
= sqlite3GlobalConfig
.szMmap
;
5377 if( sqlite3_uri_boolean(((ctrlFlags
& UNIXFILE_URI
) ? zFilename
: 0),
5378 "psow", SQLITE_POWERSAFE_OVERWRITE
) ){
5379 pNew
->ctrlFlags
|= UNIXFILE_PSOW
;
5381 if( strcmp(pVfs
->zName
,"unix-excl")==0 ){
5382 pNew
->ctrlFlags
|= UNIXFILE_EXCL
;
5386 pNew
->pId
= vxworksFindFileId(zFilename
);
5388 ctrlFlags
|= UNIXFILE_NOLOCK
;
5389 rc
= SQLITE_NOMEM_BKPT
;
5393 if( ctrlFlags
& UNIXFILE_NOLOCK
){
5394 pLockingStyle
= &nolockIoMethods
;
5396 pLockingStyle
= (**(finder_type
*)pVfs
->pAppData
)(zFilename
, pNew
);
5397 #if SQLITE_ENABLE_LOCKING_STYLE
5398 /* Cache zFilename in the locking context (AFP and dotlock override) for
5399 ** proxyLock activation is possible (remote proxy is based on db name)
5400 ** zFilename remains valid until file is closed, to support */
5401 pNew
->lockingContext
= (void*)zFilename
;
5405 if( pLockingStyle
== &posixIoMethods
5406 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5407 || pLockingStyle
== &nfsIoMethods
5411 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5412 if( rc
!=SQLITE_OK
){
5413 /* If an error occurred in findInodeInfo(), close the file descriptor
5414 ** immediately, before releasing the mutex. findInodeInfo() may fail
5415 ** in two scenarios:
5417 ** (a) A call to fstat() failed.
5418 ** (b) A malloc failed.
5420 ** Scenario (b) may only occur if the process is holding no other
5421 ** file descriptors open on the same file. If there were other file
5422 ** descriptors on this file, then no malloc would be required by
5423 ** findInodeInfo(). If this is the case, it is quite safe to close
5424 ** handle h - as it is guaranteed that no posix locks will be released
5427 ** If scenario (a) caused the error then things are not so safe. The
5428 ** implicit assumption here is that if fstat() fails, things are in
5429 ** such bad shape that dropping a lock or two doesn't matter much.
5431 robust_close(pNew
, h
, __LINE__
);
5437 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5438 else if( pLockingStyle
== &afpIoMethods
){
5439 /* AFP locking uses the file path so it needs to be included in
5440 ** the afpLockingContext.
5442 afpLockingContext
*pCtx
;
5443 pNew
->lockingContext
= pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
5445 rc
= SQLITE_NOMEM_BKPT
;
5447 /* NB: zFilename exists and remains valid until the file is closed
5448 ** according to requirement F11141. So we do not need to make a
5449 ** copy of the filename. */
5450 pCtx
->dbPath
= zFilename
;
5454 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5455 if( rc
!=SQLITE_OK
){
5456 sqlite3_free(pNew
->lockingContext
);
5457 robust_close(pNew
, h
, __LINE__
);
5465 else if( pLockingStyle
== &dotlockIoMethods
){
5466 /* Dotfile locking uses the file path so it needs to be included in
5467 ** the dotlockLockingContext
5471 assert( zFilename
!=0 );
5472 nFilename
= (int)strlen(zFilename
) + 6;
5473 zLockFile
= (char *)sqlite3_malloc64(nFilename
);
5475 rc
= SQLITE_NOMEM_BKPT
;
5477 sqlite3_snprintf(nFilename
, zLockFile
, "%s" DOTLOCK_SUFFIX
, zFilename
);
5479 pNew
->lockingContext
= zLockFile
;
5483 else if( pLockingStyle
== &semIoMethods
){
5484 /* Named semaphore locking uses the file path so it needs to be
5485 ** included in the semLockingContext
5488 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5489 if( (rc
==SQLITE_OK
) && (pNew
->pInode
->pSem
==NULL
) ){
5490 char *zSemName
= pNew
->pInode
->aSemName
;
5492 sqlite3_snprintf(MAX_PATHNAME
, zSemName
, "/%s.sem",
5493 pNew
->pId
->zCanonicalName
);
5494 for( n
=1; zSemName
[n
]; n
++ )
5495 if( zSemName
[n
]=='/' ) zSemName
[n
] = '_';
5496 pNew
->pInode
->pSem
= sem_open(zSemName
, O_CREAT
, 0666, 1);
5497 if( pNew
->pInode
->pSem
== SEM_FAILED
){
5498 rc
= SQLITE_NOMEM_BKPT
;
5499 pNew
->pInode
->aSemName
[0] = '\0';
5506 storeLastErrno(pNew
, 0);
5508 if( rc
!=SQLITE_OK
){
5509 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5511 osUnlink(zFilename
);
5512 pNew
->ctrlFlags
|= UNIXFILE_DELETE
;
5515 if( rc
!=SQLITE_OK
){
5516 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5518 pNew
->pMethod
= pLockingStyle
;
5526 ** Return the name of a directory in which to put temporary files.
5527 ** If no suitable temporary file directory can be found, return NULL.
5529 static const char *unixTempFileDir(void){
5530 static const char *azDirs
[] = {
5540 const char *zDir
= sqlite3_temp_directory
;
5542 if( !azDirs
[0] ) azDirs
[0] = getenv("SQLITE_TMPDIR");
5543 if( !azDirs
[1] ) azDirs
[1] = getenv("TMPDIR");
5546 && osStat(zDir
, &buf
)==0
5547 && S_ISDIR(buf
.st_mode
)
5548 && osAccess(zDir
, 03)==0
5552 if( i
>=sizeof(azDirs
)/sizeof(azDirs
[0]) ) break;
5559 ** Create a temporary file name in zBuf. zBuf must be allocated
5560 ** by the calling process and must be big enough to hold at least
5561 ** pVfs->mxPathname bytes.
5563 static int unixGetTempname(int nBuf
, char *zBuf
){
5567 /* It's odd to simulate an io-error here, but really this is just
5568 ** using the io-error infrastructure to test that SQLite handles this
5569 ** function failing.
5572 SimulateIOError( return SQLITE_IOERR
);
5574 zDir
= unixTempFileDir();
5575 if( zDir
==0 ) return SQLITE_IOERR_GETTEMPPATH
;
5578 sqlite3_randomness(sizeof(r
), &r
);
5581 sqlite3_snprintf(nBuf
, zBuf
, "%s/"SQLITE_TEMP_FILE_PREFIX
"%llx%c",
5583 if( zBuf
[nBuf
-2]!=0 || (iLimit
++)>10 ) return SQLITE_ERROR
;
5584 }while( osAccess(zBuf
,0)==0 );
5588 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5590 ** Routine to transform a unixFile into a proxy-locking unixFile.
5591 ** Implementation in the proxy-lock division, but used by unixOpen()
5592 ** if SQLITE_PREFER_PROXY_LOCKING is defined.
5594 static int proxyTransformUnixFile(unixFile
*, const char*);
5598 ** Search for an unused file descriptor that was opened on the database
5599 ** file (not a journal or master-journal file) identified by pathname
5600 ** zPath with SQLITE_OPEN_XXX flags matching those passed as the second
5601 ** argument to this function.
5603 ** Such a file descriptor may exist if a database connection was closed
5604 ** but the associated file descriptor could not be closed because some
5605 ** other file descriptor open on the same file is holding a file-lock.
5606 ** Refer to comments in the unixClose() function and the lengthy comment
5607 ** describing "Posix Advisory Locking" at the start of this file for
5608 ** further details. Also, ticket #4018.
5610 ** If a suitable file descriptor is found, then it is returned. If no
5611 ** such file descriptor is located, -1 is returned.
5613 static UnixUnusedFd
*findReusableFd(const char *zPath
, int flags
){
5614 UnixUnusedFd
*pUnused
= 0;
5616 /* Do not search for an unused file descriptor on vxworks. Not because
5617 ** vxworks would not benefit from the change (it might, we're not sure),
5618 ** but because no way to test it is currently available. It is better
5619 ** not to risk breaking vxworks support for the sake of such an obscure
5622 struct stat sStat
; /* Results of stat() call */
5626 /* A stat() call may fail for various reasons. If this happens, it is
5627 ** almost certain that an open() call on the same path will also fail.
5628 ** For this reason, if an error occurs in the stat() call here, it is
5629 ** ignored and -1 is returned. The caller will try to open a new file
5630 ** descriptor on the same path, fail, and return an error to SQLite.
5632 ** Even if a subsequent open() call does succeed, the consequences of
5633 ** not searching for a reusable file descriptor are not dire. */
5634 if( nUnusedFd
>0 && 0==osStat(zPath
, &sStat
) ){
5635 unixInodeInfo
*pInode
;
5638 while( pInode
&& (pInode
->fileId
.dev
!=sStat
.st_dev
5639 || pInode
->fileId
.ino
!=(u64
)sStat
.st_ino
) ){
5640 pInode
= pInode
->pNext
;
5644 for(pp
=&pInode
->pUnused
; *pp
&& (*pp
)->flags
!=flags
; pp
=&((*pp
)->pNext
));
5648 *pp
= pUnused
->pNext
;
5653 #endif /* if !OS_VXWORKS */
5658 ** Find the mode, uid and gid of file zFile.
5660 static int getFileMode(
5661 const char *zFile
, /* File name */
5662 mode_t
*pMode
, /* OUT: Permissions of zFile */
5663 uid_t
*pUid
, /* OUT: uid of zFile. */
5664 gid_t
*pGid
/* OUT: gid of zFile. */
5666 struct stat sStat
; /* Output of stat() on database file */
5668 if( 0==osStat(zFile
, &sStat
) ){
5669 *pMode
= sStat
.st_mode
& 0777;
5670 *pUid
= sStat
.st_uid
;
5671 *pGid
= sStat
.st_gid
;
5673 rc
= SQLITE_IOERR_FSTAT
;
5679 ** This function is called by unixOpen() to determine the unix permissions
5680 ** to create new files with. If no error occurs, then SQLITE_OK is returned
5681 ** and a value suitable for passing as the third argument to open(2) is
5682 ** written to *pMode. If an IO error occurs, an SQLite error code is
5683 ** returned and the value of *pMode is not modified.
5685 ** In most cases, this routine sets *pMode to 0, which will become
5686 ** an indication to robust_open() to create the file using
5687 ** SQLITE_DEFAULT_FILE_PERMISSIONS adjusted by the umask.
5688 ** But if the file being opened is a WAL or regular journal file, then
5689 ** this function queries the file-system for the permissions on the
5690 ** corresponding database file and sets *pMode to this value. Whenever
5691 ** possible, WAL and journal files are created using the same permissions
5692 ** as the associated database file.
5694 ** If the SQLITE_ENABLE_8_3_NAMES option is enabled, then the
5695 ** original filename is unavailable. But 8_3_NAMES is only used for
5696 ** FAT filesystems and permissions do not matter there, so just use
5697 ** the default permissions.
5699 static int findCreateFileMode(
5700 const char *zPath
, /* Path of file (possibly) being created */
5701 int flags
, /* Flags passed as 4th argument to xOpen() */
5702 mode_t
*pMode
, /* OUT: Permissions to open file with */
5703 uid_t
*pUid
, /* OUT: uid to set on the file */
5704 gid_t
*pGid
/* OUT: gid to set on the file */
5706 int rc
= SQLITE_OK
; /* Return Code */
5710 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5711 char zDb
[MAX_PATHNAME
+1]; /* Database file path */
5712 int nDb
; /* Number of valid bytes in zDb */
5714 /* zPath is a path to a WAL or journal file. The following block derives
5715 ** the path to the associated database file from zPath. This block handles
5716 ** the following naming conventions:
5718 ** "<path to db>-journal"
5719 ** "<path to db>-wal"
5720 ** "<path to db>-journalNN"
5721 ** "<path to db>-walNN"
5723 ** where NN is a decimal number. The NN naming schemes are
5724 ** used by the test_multiplex.c module.
5726 nDb
= sqlite3Strlen30(zPath
) - 1;
5727 while( zPath
[nDb
]!='-' ){
5728 /* In normal operation, the journal file name will always contain
5729 ** a '-' character. However in 8+3 filename mode, or if a corrupt
5730 ** rollback journal specifies a master journal with a goofy name, then
5731 ** the '-' might be missing. */
5732 if( nDb
==0 || zPath
[nDb
]=='.' ) return SQLITE_OK
;
5735 memcpy(zDb
, zPath
, nDb
);
5738 rc
= getFileMode(zDb
, pMode
, pUid
, pGid
);
5739 }else if( flags
& SQLITE_OPEN_DELETEONCLOSE
){
5741 }else if( flags
& SQLITE_OPEN_URI
){
5742 /* If this is a main database file and the file was opened using a URI
5743 ** filename, check for the "modeof" parameter. If present, interpret
5744 ** its value as a filename and try to copy the mode, uid and gid from
5746 const char *z
= sqlite3_uri_parameter(zPath
, "modeof");
5748 rc
= getFileMode(z
, pMode
, pUid
, pGid
);
5755 ** Open the file zPath.
5757 ** Previously, the SQLite OS layer used three functions in place of this
5760 ** sqlite3OsOpenReadWrite();
5761 ** sqlite3OsOpenReadOnly();
5762 ** sqlite3OsOpenExclusive();
5764 ** These calls correspond to the following combinations of flags:
5766 ** ReadWrite() -> (READWRITE | CREATE)
5767 ** ReadOnly() -> (READONLY)
5768 ** OpenExclusive() -> (READWRITE | CREATE | EXCLUSIVE)
5770 ** The old OpenExclusive() accepted a boolean argument - "delFlag". If
5771 ** true, the file was configured to be automatically deleted when the
5772 ** file handle closed. To achieve the same effect using this new
5773 ** interface, add the DELETEONCLOSE flag to those specified above for
5776 static int unixOpen(
5777 sqlite3_vfs
*pVfs
, /* The VFS for which this is the xOpen method */
5778 const char *zPath
, /* Pathname of file to be opened */
5779 sqlite3_file
*pFile
, /* The file descriptor to be filled in */
5780 int flags
, /* Input flags to control the opening */
5781 int *pOutFlags
/* Output flags returned to SQLite core */
5783 unixFile
*p
= (unixFile
*)pFile
;
5784 int fd
= -1; /* File descriptor returned by open() */
5785 int openFlags
= 0; /* Flags to pass to open() */
5786 int eType
= flags
&0xFFFFFF00; /* Type of file to open */
5787 int noLock
; /* True to omit locking primitives */
5788 int rc
= SQLITE_OK
; /* Function Return Code */
5789 int ctrlFlags
= 0; /* UNIXFILE_* flags */
5791 int isExclusive
= (flags
& SQLITE_OPEN_EXCLUSIVE
);
5792 int isDelete
= (flags
& SQLITE_OPEN_DELETEONCLOSE
);
5793 int isCreate
= (flags
& SQLITE_OPEN_CREATE
);
5794 int isReadonly
= (flags
& SQLITE_OPEN_READONLY
);
5795 int isReadWrite
= (flags
& SQLITE_OPEN_READWRITE
);
5796 #if SQLITE_ENABLE_LOCKING_STYLE
5797 int isAutoProxy
= (flags
& SQLITE_OPEN_AUTOPROXY
);
5799 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
5800 struct statfs fsInfo
;
5803 /* If creating a master or main-file journal, this function will open
5804 ** a file-descriptor on the directory too. The first time unixSync()
5805 ** is called the directory file descriptor will be fsync()ed and close()d.
5807 int isNewJrnl
= (isCreate
&& (
5808 eType
==SQLITE_OPEN_MASTER_JOURNAL
5809 || eType
==SQLITE_OPEN_MAIN_JOURNAL
5810 || eType
==SQLITE_OPEN_WAL
5813 /* If argument zPath is a NULL pointer, this function is required to open
5814 ** a temporary file. Use this buffer to store the file name in.
5816 char zTmpname
[MAX_PATHNAME
+2];
5817 const char *zName
= zPath
;
5819 /* Check the following statements are true:
5821 ** (a) Exactly one of the READWRITE and READONLY flags must be set, and
5822 ** (b) if CREATE is set, then READWRITE must also be set, and
5823 ** (c) if EXCLUSIVE is set, then CREATE must also be set.
5824 ** (d) if DELETEONCLOSE is set, then CREATE must also be set.
5826 assert((isReadonly
==0 || isReadWrite
==0) && (isReadWrite
|| isReadonly
));
5827 assert(isCreate
==0 || isReadWrite
);
5828 assert(isExclusive
==0 || isCreate
);
5829 assert(isDelete
==0 || isCreate
);
5831 /* The main DB, main journal, WAL file and master journal are never
5832 ** automatically deleted. Nor are they ever temporary files. */
5833 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_DB
);
5834 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_JOURNAL
);
5835 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MASTER_JOURNAL
);
5836 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_WAL
);
5838 /* Assert that the upper layer has set one of the "file-type" flags. */
5839 assert( eType
==SQLITE_OPEN_MAIN_DB
|| eType
==SQLITE_OPEN_TEMP_DB
5840 || eType
==SQLITE_OPEN_MAIN_JOURNAL
|| eType
==SQLITE_OPEN_TEMP_JOURNAL
5841 || eType
==SQLITE_OPEN_SUBJOURNAL
|| eType
==SQLITE_OPEN_MASTER_JOURNAL
5842 || eType
==SQLITE_OPEN_TRANSIENT_DB
|| eType
==SQLITE_OPEN_WAL
5845 /* Detect a pid change and reset the PRNG. There is a race condition
5846 ** here such that two or more threads all trying to open databases at
5847 ** the same instant might all reset the PRNG. But multiple resets
5850 if( randomnessPid
!=osGetpid(0) ){
5851 randomnessPid
= osGetpid(0);
5852 sqlite3_randomness(0,0);
5854 memset(p
, 0, sizeof(unixFile
));
5856 if( eType
==SQLITE_OPEN_MAIN_DB
){
5857 UnixUnusedFd
*pUnused
;
5858 pUnused
= findReusableFd(zName
, flags
);
5862 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
5864 return SQLITE_NOMEM_BKPT
;
5867 p
->pPreallocatedUnused
= pUnused
;
5869 /* Database filenames are double-zero terminated if they are not
5870 ** URIs with parameters. Hence, they can always be passed into
5871 ** sqlite3_uri_parameter(). */
5872 assert( (flags
& SQLITE_OPEN_URI
) || zName
[strlen(zName
)+1]==0 );
5875 /* If zName is NULL, the upper layer is requesting a temp file. */
5876 assert(isDelete
&& !isNewJrnl
);
5877 rc
= unixGetTempname(pVfs
->mxPathname
, zTmpname
);
5878 if( rc
!=SQLITE_OK
){
5883 /* Generated temporary filenames are always double-zero terminated
5884 ** for use by sqlite3_uri_parameter(). */
5885 assert( zName
[strlen(zName
)+1]==0 );
5888 /* Determine the value of the flags parameter passed to POSIX function
5889 ** open(). These must be calculated even if open() is not called, as
5890 ** they may be stored as part of the file handle and used by the
5891 ** 'conch file' locking functions later on. */
5892 if( isReadonly
) openFlags
|= O_RDONLY
;
5893 if( isReadWrite
) openFlags
|= O_RDWR
;
5894 if( isCreate
) openFlags
|= O_CREAT
;
5895 if( isExclusive
) openFlags
|= (O_EXCL
|O_NOFOLLOW
);
5896 openFlags
|= (O_LARGEFILE
|O_BINARY
);
5899 mode_t openMode
; /* Permissions to create file with */
5900 uid_t uid
; /* Userid for the file */
5901 gid_t gid
; /* Groupid for the file */
5902 rc
= findCreateFileMode(zName
, flags
, &openMode
, &uid
, &gid
);
5903 if( rc
!=SQLITE_OK
){
5904 assert( !p
->pPreallocatedUnused
);
5905 assert( eType
==SQLITE_OPEN_WAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
);
5908 fd
= robust_open(zName
, openFlags
, openMode
);
5909 OSTRACE(("OPENX %-3d %s 0%o\n", fd
, zName
, openFlags
));
5910 assert( !isExclusive
|| (openFlags
& O_CREAT
)!=0 );
5912 if( isNewJrnl
&& errno
==EACCES
&& osAccess(zName
, F_OK
) ){
5913 /* If unable to create a journal because the directory is not
5914 ** writable, change the error code to indicate that. */
5915 rc
= SQLITE_READONLY_DIRECTORY
;
5916 }else if( errno
!=EISDIR
&& isReadWrite
){
5917 /* Failed to open the file for read/write access. Try read-only. */
5918 flags
&= ~(SQLITE_OPEN_READWRITE
|SQLITE_OPEN_CREATE
);
5919 openFlags
&= ~(O_RDWR
|O_CREAT
);
5920 flags
|= SQLITE_OPEN_READONLY
;
5921 openFlags
|= O_RDONLY
;
5923 fd
= robust_open(zName
, openFlags
, openMode
);
5927 int rc2
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zName
);
5928 if( rc
==SQLITE_OK
) rc
= rc2
;
5932 /* If this process is running as root and if creating a new rollback
5933 ** journal or WAL file, set the ownership of the journal or WAL to be
5934 ** the same as the original database.
5936 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5937 robustFchown(fd
, uid
, gid
);
5945 if( p
->pPreallocatedUnused
){
5946 p
->pPreallocatedUnused
->fd
= fd
;
5947 p
->pPreallocatedUnused
->flags
= flags
;
5953 #elif defined(SQLITE_UNLINK_AFTER_CLOSE)
5954 zPath
= sqlite3_mprintf("%s", zName
);
5956 robust_close(p
, fd
, __LINE__
);
5957 return SQLITE_NOMEM_BKPT
;
5963 #if SQLITE_ENABLE_LOCKING_STYLE
5965 p
->openFlags
= openFlags
;
5969 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
5970 if( fstatfs(fd
, &fsInfo
) == -1 ){
5971 storeLastErrno(p
, errno
);
5972 robust_close(p
, fd
, __LINE__
);
5973 return SQLITE_IOERR_ACCESS
;
5975 if (0 == strncmp("msdos", fsInfo
.f_fstypename
, 5)) {
5976 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
5978 if (0 == strncmp("exfat", fsInfo
.f_fstypename
, 5)) {
5979 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
5983 /* Set up appropriate ctrlFlags */
5984 if( isDelete
) ctrlFlags
|= UNIXFILE_DELETE
;
5985 if( isReadonly
) ctrlFlags
|= UNIXFILE_RDONLY
;
5986 noLock
= eType
!=SQLITE_OPEN_MAIN_DB
;
5987 if( noLock
) ctrlFlags
|= UNIXFILE_NOLOCK
;
5988 if( isNewJrnl
) ctrlFlags
|= UNIXFILE_DIRSYNC
;
5989 if( flags
& SQLITE_OPEN_URI
) ctrlFlags
|= UNIXFILE_URI
;
5991 #if SQLITE_ENABLE_LOCKING_STYLE
5992 #if SQLITE_PREFER_PROXY_LOCKING
5995 if( isAutoProxy
&& (zPath
!=NULL
) && (!noLock
) && pVfs
->xOpen
){
5996 char *envforce
= getenv("SQLITE_FORCE_PROXY_LOCKING");
5999 /* SQLITE_FORCE_PROXY_LOCKING==1 means force always use proxy, 0 means
6000 ** never use proxy, NULL means use proxy for non-local files only. */
6001 if( envforce
!=NULL
){
6002 useProxy
= atoi(envforce
)>0;
6004 useProxy
= !(fsInfo
.f_flags
&MNT_LOCAL
);
6007 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
6008 if( rc
==SQLITE_OK
){
6009 rc
= proxyTransformUnixFile((unixFile
*)pFile
, ":auto:");
6010 if( rc
!=SQLITE_OK
){
6011 /* Use unixClose to clean up the resources added in fillInUnixFile
6012 ** and clear all the structure's references. Specifically,
6013 ** pFile->pMethods will be NULL so sqlite3OsClose will be a no-op
6024 assert( zPath
==0 || zPath
[0]=='/'
6025 || eType
==SQLITE_OPEN_MASTER_JOURNAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
6027 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
6030 if( rc
!=SQLITE_OK
){
6031 sqlite3_free(p
->pPreallocatedUnused
);
6038 ** Delete the file at zPath. If the dirSync argument is true, fsync()
6039 ** the directory after deleting the file.
6041 static int unixDelete(
6042 sqlite3_vfs
*NotUsed
, /* VFS containing this as the xDelete method */
6043 const char *zPath
, /* Name of file to be deleted */
6044 int dirSync
/* If true, fsync() directory after deleting file */
6047 UNUSED_PARAMETER(NotUsed
);
6048 SimulateIOError(return SQLITE_IOERR_DELETE
);
6049 if( osUnlink(zPath
)==(-1) ){
6052 || osAccess(zPath
,0)!=0
6055 rc
= SQLITE_IOERR_DELETE_NOENT
;
6057 rc
= unixLogError(SQLITE_IOERR_DELETE
, "unlink", zPath
);
6061 #ifndef SQLITE_DISABLE_DIRSYNC
6062 if( (dirSync
& 1)!=0 ){
6064 rc
= osOpenDirectory(zPath
, &fd
);
6065 if( rc
==SQLITE_OK
){
6066 if( full_fsync(fd
,0,0) ){
6067 rc
= unixLogError(SQLITE_IOERR_DIR_FSYNC
, "fsync", zPath
);
6069 robust_close(0, fd
, __LINE__
);
6071 assert( rc
==SQLITE_CANTOPEN
);
6080 ** Test the existence of or access permissions of file zPath. The
6081 ** test performed depends on the value of flags:
6083 ** SQLITE_ACCESS_EXISTS: Return 1 if the file exists
6084 ** SQLITE_ACCESS_READWRITE: Return 1 if the file is read and writable.
6085 ** SQLITE_ACCESS_READONLY: Return 1 if the file is readable.
6087 ** Otherwise return 0.
6089 static int unixAccess(
6090 sqlite3_vfs
*NotUsed
, /* The VFS containing this xAccess method */
6091 const char *zPath
, /* Path of the file to examine */
6092 int flags
, /* What do we want to learn about the zPath file? */
6093 int *pResOut
/* Write result boolean here */
6095 UNUSED_PARAMETER(NotUsed
);
6096 SimulateIOError( return SQLITE_IOERR_ACCESS
; );
6097 assert( pResOut
!=0 );
6099 /* The spec says there are three possible values for flags. But only
6100 ** two of them are actually used */
6101 assert( flags
==SQLITE_ACCESS_EXISTS
|| flags
==SQLITE_ACCESS_READWRITE
);
6103 if( flags
==SQLITE_ACCESS_EXISTS
){
6105 *pResOut
= (0==osStat(zPath
, &buf
) && buf
.st_size
>0);
6107 *pResOut
= osAccess(zPath
, W_OK
|R_OK
)==0;
6115 static int mkFullPathname(
6116 const char *zPath
, /* Input path */
6117 char *zOut
, /* Output buffer */
6118 int nOut
/* Allocated size of buffer zOut */
6120 int nPath
= sqlite3Strlen30(zPath
);
6122 if( zPath
[0]!='/' ){
6123 if( osGetcwd(zOut
, nOut
-2)==0 ){
6124 return unixLogError(SQLITE_CANTOPEN_BKPT
, "getcwd", zPath
);
6126 iOff
= sqlite3Strlen30(zOut
);
6129 if( (iOff
+nPath
+1)>nOut
){
6130 /* SQLite assumes that xFullPathname() nul-terminates the output buffer
6131 ** even if it returns an error. */
6133 return SQLITE_CANTOPEN_BKPT
;
6135 sqlite3_snprintf(nOut
-iOff
, &zOut
[iOff
], "%s", zPath
);
6140 ** Turn a relative pathname into a full pathname. The relative path
6141 ** is stored as a nul-terminated string in the buffer pointed to by
6144 ** zOut points to a buffer of at least sqlite3_vfs.mxPathname bytes
6145 ** (in this case, MAX_PATHNAME bytes). The full-path is written to
6146 ** this buffer before returning.
6148 static int unixFullPathname(
6149 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
6150 const char *zPath
, /* Possibly relative input path */
6151 int nOut
, /* Size of output buffer in bytes */
6152 char *zOut
/* Output buffer */
6154 #if !defined(HAVE_READLINK) || !defined(HAVE_LSTAT)
6155 return mkFullPathname(zPath
, zOut
, nOut
);
6159 int nLink
= 1; /* Number of symbolic links followed so far */
6160 const char *zIn
= zPath
; /* Input path for each iteration of loop */
6163 assert( pVfs
->mxPathname
==MAX_PATHNAME
);
6164 UNUSED_PARAMETER(pVfs
);
6166 /* It's odd to simulate an io-error here, but really this is just
6167 ** using the io-error infrastructure to test that SQLite handles this
6168 ** function failing. This function could fail if, for example, the
6169 ** current working directory has been unlinked.
6171 SimulateIOError( return SQLITE_ERROR
);
6175 /* Call stat() on path zIn. Set bLink to true if the path is a symbolic
6176 ** link, or false otherwise. */
6179 if( osLstat(zIn
, &buf
)!=0 ){
6180 if( errno
!=ENOENT
){
6181 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "lstat", zIn
);
6184 bLink
= S_ISLNK(buf
.st_mode
);
6189 zDel
= sqlite3_malloc(nOut
);
6190 if( zDel
==0 ) rc
= SQLITE_NOMEM_BKPT
;
6191 }else if( ++nLink
>SQLITE_MAX_SYMLINKS
){
6192 rc
= SQLITE_CANTOPEN_BKPT
;
6195 if( rc
==SQLITE_OK
){
6196 nByte
= osReadlink(zIn
, zDel
, nOut
-1);
6198 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "readlink", zIn
);
6202 for(n
= sqlite3Strlen30(zIn
); n
>0 && zIn
[n
-1]!='/'; n
--);
6203 if( nByte
+n
+1>nOut
){
6204 rc
= SQLITE_CANTOPEN_BKPT
;
6206 memmove(&zDel
[n
], zDel
, nByte
+1);
6207 memcpy(zDel
, zIn
, n
);
6218 assert( rc
!=SQLITE_OK
|| zIn
!=zOut
|| zIn
[0]=='/' );
6219 if( rc
==SQLITE_OK
&& zIn
!=zOut
){
6220 rc
= mkFullPathname(zIn
, zOut
, nOut
);
6222 if( bLink
==0 ) break;
6224 }while( rc
==SQLITE_OK
);
6228 #endif /* HAVE_READLINK && HAVE_LSTAT */
6232 #ifndef SQLITE_OMIT_LOAD_EXTENSION
6234 ** Interfaces for opening a shared library, finding entry points
6235 ** within the shared library, and closing the shared library.
6238 static void *unixDlOpen(sqlite3_vfs
*NotUsed
, const char *zFilename
){
6239 UNUSED_PARAMETER(NotUsed
);
6240 return dlopen(zFilename
, RTLD_NOW
| RTLD_GLOBAL
);
6244 ** SQLite calls this function immediately after a call to unixDlSym() or
6245 ** unixDlOpen() fails (returns a null pointer). If a more detailed error
6246 ** message is available, it is written to zBufOut. If no error message
6247 ** is available, zBufOut is left unmodified and SQLite uses a default
6250 static void unixDlError(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBufOut
){
6252 UNUSED_PARAMETER(NotUsed
);
6256 sqlite3_snprintf(nBuf
, zBufOut
, "%s", zErr
);
6260 static void (*unixDlSym(sqlite3_vfs
*NotUsed
, void *p
, const char*zSym
))(void){
6262 ** GCC with -pedantic-errors says that C90 does not allow a void* to be
6263 ** cast into a pointer to a function. And yet the library dlsym() routine
6264 ** returns a void* which is really a pointer to a function. So how do we
6265 ** use dlsym() with -pedantic-errors?
6267 ** Variable x below is defined to be a pointer to a function taking
6268 ** parameters void* and const char* and returning a pointer to a function.
6269 ** We initialize x by assigning it a pointer to the dlsym() function.
6270 ** (That assignment requires a cast.) Then we call the function that
6273 ** This work-around is unlikely to work correctly on any system where
6274 ** you really cannot cast a function pointer into void*. But then, on the
6275 ** other hand, dlsym() will not work on such a system either, so we have
6276 ** not really lost anything.
6278 void (*(*x
)(void*,const char*))(void);
6279 UNUSED_PARAMETER(NotUsed
);
6280 x
= (void(*(*)(void*,const char*))(void))dlsym
;
6281 return (*x
)(p
, zSym
);
6283 static void unixDlClose(sqlite3_vfs
*NotUsed
, void *pHandle
){
6284 UNUSED_PARAMETER(NotUsed
);
6287 #else /* if SQLITE_OMIT_LOAD_EXTENSION is defined: */
6288 #define unixDlOpen 0
6289 #define unixDlError 0
6291 #define unixDlClose 0
6295 ** Write nBuf bytes of random data to the supplied buffer zBuf.
6297 static int unixRandomness(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBuf
){
6298 UNUSED_PARAMETER(NotUsed
);
6299 assert((size_t)nBuf
>=(sizeof(time_t)+sizeof(int)));
6301 /* We have to initialize zBuf to prevent valgrind from reporting
6302 ** errors. The reports issued by valgrind are incorrect - we would
6303 ** prefer that the randomness be increased by making use of the
6304 ** uninitialized space in zBuf - but valgrind errors tend to worry
6305 ** some users. Rather than argue, it seems easier just to initialize
6306 ** the whole array and silence valgrind, even if that means less randomness
6307 ** in the random seed.
6309 ** When testing, initializing zBuf[] to zero is all we do. That means
6310 ** that we always use the same random number sequence. This makes the
6311 ** tests repeatable.
6313 memset(zBuf
, 0, nBuf
);
6314 randomnessPid
= osGetpid(0);
6315 #if !defined(SQLITE_TEST) && !defined(SQLITE_OMIT_RANDOMNESS)
6318 fd
= robust_open("/dev/urandom", O_RDONLY
, 0);
6322 memcpy(zBuf
, &t
, sizeof(t
));
6323 memcpy(&zBuf
[sizeof(t
)], &randomnessPid
, sizeof(randomnessPid
));
6324 assert( sizeof(t
)+sizeof(randomnessPid
)<=(size_t)nBuf
);
6325 nBuf
= sizeof(t
) + sizeof(randomnessPid
);
6327 do{ got
= osRead(fd
, zBuf
, nBuf
); }while( got
<0 && errno
==EINTR
);
6328 robust_close(0, fd
, __LINE__
);
6337 ** Sleep for a little while. Return the amount of time slept.
6338 ** The argument is the number of microseconds we want to sleep.
6339 ** The return value is the number of microseconds of sleep actually
6340 ** requested from the underlying operating system, a number which
6341 ** might be greater than or equal to the argument, but not less
6342 ** than the argument.
6344 static int unixSleep(sqlite3_vfs
*NotUsed
, int microseconds
){
6348 sp
.tv_sec
= microseconds
/ 1000000;
6349 sp
.tv_nsec
= (microseconds
% 1000000) * 1000;
6350 nanosleep(&sp
, NULL
);
6351 UNUSED_PARAMETER(NotUsed
);
6352 return microseconds
;
6353 #elif defined(HAVE_USLEEP) && HAVE_USLEEP
6354 usleep(microseconds
);
6355 UNUSED_PARAMETER(NotUsed
);
6356 return microseconds
;
6358 int seconds
= (microseconds
+999999)/1000000;
6360 UNUSED_PARAMETER(NotUsed
);
6361 return seconds
*1000000;
6366 ** The following variable, if set to a non-zero value, is interpreted as
6367 ** the number of seconds since 1970 and is used to set the result of
6368 ** sqlite3OsCurrentTime() during testing.
6371 int sqlite3_current_time
= 0; /* Fake system time in seconds since 1970. */
6375 ** Find the current time (in Universal Coordinated Time). Write into *piNow
6376 ** the current time and date as a Julian Day number times 86_400_000. In
6377 ** other words, write into *piNow the number of milliseconds since the Julian
6378 ** epoch of noon in Greenwich on November 24, 4714 B.C according to the
6379 ** proleptic Gregorian calendar.
6381 ** On success, return SQLITE_OK. Return SQLITE_ERROR if the time and date
6384 static int unixCurrentTimeInt64(sqlite3_vfs
*NotUsed
, sqlite3_int64
*piNow
){
6385 static const sqlite3_int64 unixEpoch
= 24405875*(sqlite3_int64
)8640000;
6387 #if defined(NO_GETTOD)
6390 *piNow
= ((sqlite3_int64
)t
)*1000 + unixEpoch
;
6392 struct timespec sNow
;
6393 clock_gettime(CLOCK_REALTIME
, &sNow
);
6394 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_nsec
/1000000;
6396 struct timeval sNow
;
6397 (void)gettimeofday(&sNow
, 0); /* Cannot fail given valid arguments */
6398 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_usec
/1000;
6402 if( sqlite3_current_time
){
6403 *piNow
= 1000*(sqlite3_int64
)sqlite3_current_time
+ unixEpoch
;
6406 UNUSED_PARAMETER(NotUsed
);
6410 #ifndef SQLITE_OMIT_DEPRECATED
6412 ** Find the current time (in Universal Coordinated Time). Write the
6413 ** current time and date as a Julian Day number into *prNow and
6414 ** return 0. Return 1 if the time and date cannot be found.
6416 static int unixCurrentTime(sqlite3_vfs
*NotUsed
, double *prNow
){
6417 sqlite3_int64 i
= 0;
6419 UNUSED_PARAMETER(NotUsed
);
6420 rc
= unixCurrentTimeInt64(0, &i
);
6421 *prNow
= i
/86400000.0;
6425 # define unixCurrentTime 0
6429 ** The xGetLastError() method is designed to return a better
6430 ** low-level error message when operating-system problems come up
6431 ** during SQLite operation. Only the integer return code is currently
6434 static int unixGetLastError(sqlite3_vfs
*NotUsed
, int NotUsed2
, char *NotUsed3
){
6435 UNUSED_PARAMETER(NotUsed
);
6436 UNUSED_PARAMETER(NotUsed2
);
6437 UNUSED_PARAMETER(NotUsed3
);
6443 ************************ End of sqlite3_vfs methods ***************************
6444 ******************************************************************************/
6446 /******************************************************************************
6447 ************************** Begin Proxy Locking ********************************
6449 ** Proxy locking is a "uber-locking-method" in this sense: It uses the
6450 ** other locking methods on secondary lock files. Proxy locking is a
6451 ** meta-layer over top of the primitive locking implemented above. For
6452 ** this reason, the division that implements of proxy locking is deferred
6453 ** until late in the file (here) after all of the other I/O methods have
6454 ** been defined - so that the primitive locking methods are available
6455 ** as services to help with the implementation of proxy locking.
6459 ** The default locking schemes in SQLite use byte-range locks on the
6460 ** database file to coordinate safe, concurrent access by multiple readers
6461 ** and writers [http://sqlite.org/lockingv3.html]. The five file locking
6462 ** states (UNLOCKED, PENDING, SHARED, RESERVED, EXCLUSIVE) are implemented
6463 ** as POSIX read & write locks over fixed set of locations (via fsctl),
6464 ** on AFP and SMB only exclusive byte-range locks are available via fsctl
6465 ** with _IOWR('z', 23, struct ByteRangeLockPB2) to track the same 5 states.
6466 ** To simulate a F_RDLCK on the shared range, on AFP a randomly selected
6467 ** address in the shared range is taken for a SHARED lock, the entire
6468 ** shared range is taken for an EXCLUSIVE lock):
6470 ** PENDING_BYTE 0x40000000
6471 ** RESERVED_BYTE 0x40000001
6472 ** SHARED_RANGE 0x40000002 -> 0x40000200
6474 ** This works well on the local file system, but shows a nearly 100x
6475 ** slowdown in read performance on AFP because the AFP client disables
6476 ** the read cache when byte-range locks are present. Enabling the read
6477 ** cache exposes a cache coherency problem that is present on all OS X
6478 ** supported network file systems. NFS and AFP both observe the
6479 ** close-to-open semantics for ensuring cache coherency
6480 ** [http://nfs.sourceforge.net/#faq_a8], which does not effectively
6481 ** address the requirements for concurrent database access by multiple
6482 ** readers and writers
6483 ** [http://www.nabble.com/SQLite-on-NFS-cache-coherency-td15655701.html].
6485 ** To address the performance and cache coherency issues, proxy file locking
6486 ** changes the way database access is controlled by limiting access to a
6487 ** single host at a time and moving file locks off of the database file
6488 ** and onto a proxy file on the local file system.
6491 ** Using proxy locks
6492 ** -----------------
6496 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_SET_LOCKPROXYFILE,
6497 ** <proxy_path> | ":auto:");
6498 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_GET_LOCKPROXYFILE,
6504 ** PRAGMA [database.]lock_proxy_file=<proxy_path> | :auto:
6505 ** PRAGMA [database.]lock_proxy_file
6507 ** Specifying ":auto:" means that if there is a conch file with a matching
6508 ** host ID in it, the proxy path in the conch file will be used, otherwise
6509 ** a proxy path based on the user's temp dir
6510 ** (via confstr(_CS_DARWIN_USER_TEMP_DIR,...)) will be used and the
6511 ** actual proxy file name is generated from the name and path of the
6512 ** database file. For example:
6514 ** For database path "/Users/me/foo.db"
6515 ** The lock path will be "<tmpdir>/sqliteplocks/_Users_me_foo.db:auto:")
6517 ** Once a lock proxy is configured for a database connection, it can not
6518 ** be removed, however it may be switched to a different proxy path via
6519 ** the above APIs (assuming the conch file is not being held by another
6520 ** connection or process).
6523 ** How proxy locking works
6524 ** -----------------------
6526 ** Proxy file locking relies primarily on two new supporting files:
6528 ** * conch file to limit access to the database file to a single host
6531 ** * proxy file to act as a proxy for the advisory locks normally
6532 ** taken on the database
6534 ** The conch file - to use a proxy file, sqlite must first "hold the conch"
6535 ** by taking an sqlite-style shared lock on the conch file, reading the
6536 ** contents and comparing the host's unique host ID (see below) and lock
6537 ** proxy path against the values stored in the conch. The conch file is
6538 ** stored in the same directory as the database file and the file name
6539 ** is patterned after the database file name as ".<databasename>-conch".
6540 ** If the conch file does not exist, or its contents do not match the
6541 ** host ID and/or proxy path, then the lock is escalated to an exclusive
6542 ** lock and the conch file contents is updated with the host ID and proxy
6543 ** path and the lock is downgraded to a shared lock again. If the conch
6544 ** is held by another process (with a shared lock), the exclusive lock
6545 ** will fail and SQLITE_BUSY is returned.
6547 ** The proxy file - a single-byte file used for all advisory file locks
6548 ** normally taken on the database file. This allows for safe sharing
6549 ** of the database file for multiple readers and writers on the same
6550 ** host (the conch ensures that they all use the same local lock file).
6552 ** Requesting the lock proxy does not immediately take the conch, it is
6553 ** only taken when the first request to lock database file is made.
6554 ** This matches the semantics of the traditional locking behavior, where
6555 ** opening a connection to a database file does not take a lock on it.
6556 ** The shared lock and an open file descriptor are maintained until
6557 ** the connection to the database is closed.
6559 ** The proxy file and the lock file are never deleted so they only need
6560 ** to be created the first time they are used.
6562 ** Configuration options
6563 ** ---------------------
6565 ** SQLITE_PREFER_PROXY_LOCKING
6567 ** Database files accessed on non-local file systems are
6568 ** automatically configured for proxy locking, lock files are
6569 ** named automatically using the same logic as
6570 ** PRAGMA lock_proxy_file=":auto:"
6572 ** SQLITE_PROXY_DEBUG
6574 ** Enables the logging of error messages during host id file
6575 ** retrieval and creation
6579 ** Overrides the default directory used for lock proxy files that
6580 ** are named automatically via the ":auto:" setting
6582 ** SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
6584 ** Permissions to use when creating a directory for storing the
6585 ** lock proxy files, only used when LOCKPROXYDIR is not set.
6588 ** As mentioned above, when compiled with SQLITE_PREFER_PROXY_LOCKING,
6589 ** setting the environment variable SQLITE_FORCE_PROXY_LOCKING to 1 will
6590 ** force proxy locking to be used for every database file opened, and 0
6591 ** will force automatic proxy locking to be disabled for all database
6592 ** files (explicitly calling the SQLITE_FCNTL_SET_LOCKPROXYFILE pragma or
6593 ** sqlite_file_control API is not affected by SQLITE_FORCE_PROXY_LOCKING).
6597 ** Proxy locking is only available on MacOSX
6599 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
6602 ** The proxyLockingContext has the path and file structures for the remote
6603 ** and local proxy files in it
6605 typedef struct proxyLockingContext proxyLockingContext
;
6606 struct proxyLockingContext
{
6607 unixFile
*conchFile
; /* Open conch file */
6608 char *conchFilePath
; /* Name of the conch file */
6609 unixFile
*lockProxy
; /* Open proxy lock file */
6610 char *lockProxyPath
; /* Name of the proxy lock file */
6611 char *dbPath
; /* Name of the open file */
6612 int conchHeld
; /* 1 if the conch is held, -1 if lockless */
6613 int nFails
; /* Number of conch taking failures */
6614 void *oldLockingContext
; /* Original lockingcontext to restore on close */
6615 sqlite3_io_methods
const *pOldMethod
; /* Original I/O methods for close */
6619 ** The proxy lock file path for the database at dbPath is written into lPath,
6620 ** which must point to valid, writable memory large enough for a maxLen length
6623 static int proxyGetLockPath(const char *dbPath
, char *lPath
, size_t maxLen
){
6629 len
= strlcpy(lPath
, LOCKPROXYDIR
, maxLen
);
6631 # ifdef _CS_DARWIN_USER_TEMP_DIR
6633 if( !confstr(_CS_DARWIN_USER_TEMP_DIR
, lPath
, maxLen
) ){
6634 OSTRACE(("GETLOCKPATH failed %s errno=%d pid=%d\n",
6635 lPath
, errno
, osGetpid(0)));
6636 return SQLITE_IOERR_LOCK
;
6638 len
= strlcat(lPath
, "sqliteplocks", maxLen
);
6641 len
= strlcpy(lPath
, "/tmp/", maxLen
);
6645 if( lPath
[len
-1]!='/' ){
6646 len
= strlcat(lPath
, "/", maxLen
);
6649 /* transform the db path to a unique cache name */
6650 dbLen
= (int)strlen(dbPath
);
6651 for( i
=0; i
<dbLen
&& (i
+len
+7)<(int)maxLen
; i
++){
6653 lPath
[i
+len
] = (c
=='/')?'_':c
;
6656 strlcat(lPath
, ":auto:", maxLen
);
6657 OSTRACE(("GETLOCKPATH proxy lock path=%s pid=%d\n", lPath
, osGetpid(0)));
6662 ** Creates the lock file and any missing directories in lockPath
6664 static int proxyCreateLockPath(const char *lockPath
){
6666 char buf
[MAXPATHLEN
];
6669 assert(lockPath
!=NULL
);
6670 /* try to create all the intermediate directories */
6671 len
= (int)strlen(lockPath
);
6672 buf
[0] = lockPath
[0];
6673 for( i
=1; i
<len
; i
++ ){
6674 if( lockPath
[i
] == '/' && (i
- start
> 0) ){
6675 /* only mkdir if leaf dir != "." or "/" or ".." */
6676 if( i
-start
>2 || (i
-start
==1 && buf
[start
] != '.' && buf
[start
] != '/')
6677 || (i
-start
==2 && buf
[start
] != '.' && buf
[start
+1] != '.') ){
6679 if( osMkdir(buf
, SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
) ){
6682 OSTRACE(("CREATELOCKPATH FAILED creating %s, "
6683 "'%s' proxy lock path=%s pid=%d\n",
6684 buf
, strerror(err
), lockPath
, osGetpid(0)));
6691 buf
[i
] = lockPath
[i
];
6693 OSTRACE(("CREATELOCKPATH proxy lock path=%s pid=%d\n",lockPath
,osGetpid(0)));
6698 ** Create a new VFS file descriptor (stored in memory obtained from
6699 ** sqlite3_malloc) and open the file named "path" in the file descriptor.
6701 ** The caller is responsible not only for closing the file descriptor
6702 ** but also for freeing the memory associated with the file descriptor.
6704 static int proxyCreateUnixFile(
6705 const char *path
, /* path for the new unixFile */
6706 unixFile
**ppFile
, /* unixFile created and returned by ref */
6707 int islockfile
/* if non zero missing dirs will be created */
6712 int openFlags
= O_RDWR
| O_CREAT
;
6713 sqlite3_vfs dummyVfs
;
6715 UnixUnusedFd
*pUnused
= NULL
;
6717 /* 1. first try to open/create the file
6718 ** 2. if that fails, and this is a lock file (not-conch), try creating
6719 ** the parent directories and then try again.
6720 ** 3. if that fails, try to open the file read-only
6721 ** otherwise return BUSY (if lock file) or CANTOPEN for the conch file
6723 pUnused
= findReusableFd(path
, openFlags
);
6727 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
6729 return SQLITE_NOMEM_BKPT
;
6733 fd
= robust_open(path
, openFlags
, 0);
6735 if( fd
<0 && errno
==ENOENT
&& islockfile
){
6736 if( proxyCreateLockPath(path
) == SQLITE_OK
){
6737 fd
= robust_open(path
, openFlags
, 0);
6742 openFlags
= O_RDONLY
;
6743 fd
= robust_open(path
, openFlags
, 0);
6754 return SQLITE_IOERR_LOCK
; /* even though it is the conch */
6756 return SQLITE_CANTOPEN_BKPT
;
6760 pNew
= (unixFile
*)sqlite3_malloc64(sizeof(*pNew
));
6762 rc
= SQLITE_NOMEM_BKPT
;
6763 goto end_create_proxy
;
6765 memset(pNew
, 0, sizeof(unixFile
));
6766 pNew
->openFlags
= openFlags
;
6767 memset(&dummyVfs
, 0, sizeof(dummyVfs
));
6768 dummyVfs
.pAppData
= (void*)&autolockIoFinder
;
6769 dummyVfs
.zName
= "dummy";
6771 pUnused
->flags
= openFlags
;
6772 pNew
->pPreallocatedUnused
= pUnused
;
6774 rc
= fillInUnixFile(&dummyVfs
, fd
, (sqlite3_file
*)pNew
, path
, 0);
6775 if( rc
==SQLITE_OK
){
6780 robust_close(pNew
, fd
, __LINE__
);
6782 sqlite3_free(pUnused
);
6787 /* simulate multiple hosts by creating unique hostid file paths */
6788 int sqlite3_hostid_num
= 0;
6791 #define PROXY_HOSTIDLEN 16 /* conch file host id length */
6793 #ifdef HAVE_GETHOSTUUID
6794 /* Not always defined in the headers as it ought to be */
6795 extern int gethostuuid(uuid_t id
, const struct timespec
*wait
);
6798 /* get the host ID via gethostuuid(), pHostID must point to PROXY_HOSTIDLEN
6799 ** bytes of writable memory.
6801 static int proxyGetHostID(unsigned char *pHostID
, int *pError
){
6802 assert(PROXY_HOSTIDLEN
== sizeof(uuid_t
));
6803 memset(pHostID
, 0, PROXY_HOSTIDLEN
);
6804 #ifdef HAVE_GETHOSTUUID
6806 struct timespec timeout
= {1, 0}; /* 1 sec timeout */
6807 if( gethostuuid(pHostID
, &timeout
) ){
6812 return SQLITE_IOERR
;
6816 UNUSED_PARAMETER(pError
);
6819 /* simulate multiple hosts by creating unique hostid file paths */
6820 if( sqlite3_hostid_num
!= 0){
6821 pHostID
[0] = (char)(pHostID
[0] + (char)(sqlite3_hostid_num
& 0xFF));
6828 /* The conch file contains the header, host id and lock file path
6830 #define PROXY_CONCHVERSION 2 /* 1-byte header, 16-byte host id, path */
6831 #define PROXY_HEADERLEN 1 /* conch file header length */
6832 #define PROXY_PATHINDEX (PROXY_HEADERLEN+PROXY_HOSTIDLEN)
6833 #define PROXY_MAXCONCHLEN (PROXY_HEADERLEN+PROXY_HOSTIDLEN+MAXPATHLEN)
6836 ** Takes an open conch file, copies the contents to a new path and then moves
6837 ** it back. The newly created file's file descriptor is assigned to the
6838 ** conch file structure and finally the original conch file descriptor is
6839 ** closed. Returns zero if successful.
6841 static int proxyBreakConchLock(unixFile
*pFile
, uuid_t myHostID
){
6842 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6843 unixFile
*conchFile
= pCtx
->conchFile
;
6844 char tPath
[MAXPATHLEN
];
6845 char buf
[PROXY_MAXCONCHLEN
];
6846 char *cPath
= pCtx
->conchFilePath
;
6849 char errmsg
[64] = "";
6852 UNUSED_PARAMETER(myHostID
);
6854 /* create a new path by replace the trailing '-conch' with '-break' */
6855 pathLen
= strlcpy(tPath
, cPath
, MAXPATHLEN
);
6856 if( pathLen
>MAXPATHLEN
|| pathLen
<6 ||
6857 (strlcpy(&tPath
[pathLen
-5], "break", 6) != 5) ){
6858 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"path error (len %d)",(int)pathLen
);
6861 /* read the conch content */
6862 readLen
= osPread(conchFile
->h
, buf
, PROXY_MAXCONCHLEN
, 0);
6863 if( readLen
<PROXY_PATHINDEX
){
6864 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"read error (len %d)",(int)readLen
);
6867 /* write it out to the temporary break file */
6868 fd
= robust_open(tPath
, (O_RDWR
|O_CREAT
|O_EXCL
), 0);
6870 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "create failed (%d)", errno
);
6873 if( osPwrite(fd
, buf
, readLen
, 0) != (ssize_t
)readLen
){
6874 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "write failed (%d)", errno
);
6877 if( rename(tPath
, cPath
) ){
6878 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "rename failed (%d)", errno
);
6882 fprintf(stderr
, "broke stale lock on %s\n", cPath
);
6883 robust_close(pFile
, conchFile
->h
, __LINE__
);
6885 conchFile
->openFlags
= O_RDWR
| O_CREAT
;
6891 robust_close(pFile
, fd
, __LINE__
);
6893 fprintf(stderr
, "failed to break stale lock on %s, %s\n", cPath
, errmsg
);
6898 /* Take the requested lock on the conch file and break a stale lock if the
6901 static int proxyConchLock(unixFile
*pFile
, uuid_t myHostID
, int lockType
){
6902 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6903 unixFile
*conchFile
= pCtx
->conchFile
;
6906 struct timespec conchModTime
;
6908 memset(&conchModTime
, 0, sizeof(conchModTime
));
6910 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
6912 if( rc
==SQLITE_BUSY
){
6913 /* If the lock failed (busy):
6914 * 1st try: get the mod time of the conch, wait 0.5s and try again.
6915 * 2nd try: fail if the mod time changed or host id is different, wait
6916 * 10 sec and try again
6917 * 3rd try: break the lock unless the mod time has changed.
6920 if( osFstat(conchFile
->h
, &buf
) ){
6921 storeLastErrno(pFile
, errno
);
6922 return SQLITE_IOERR_LOCK
;
6926 conchModTime
= buf
.st_mtimespec
;
6927 usleep(500000); /* wait 0.5 sec and try the lock again*/
6932 if( conchModTime
.tv_sec
!= buf
.st_mtimespec
.tv_sec
||
6933 conchModTime
.tv_nsec
!= buf
.st_mtimespec
.tv_nsec
){
6938 char tBuf
[PROXY_MAXCONCHLEN
];
6939 int len
= osPread(conchFile
->h
, tBuf
, PROXY_MAXCONCHLEN
, 0);
6941 storeLastErrno(pFile
, errno
);
6942 return SQLITE_IOERR_LOCK
;
6944 if( len
>PROXY_PATHINDEX
&& tBuf
[0]==(char)PROXY_CONCHVERSION
){
6945 /* don't break the lock if the host id doesn't match */
6946 if( 0!=memcmp(&tBuf
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
) ){
6950 /* don't break the lock on short read or a version mismatch */
6953 usleep(10000000); /* wait 10 sec and try the lock again */
6957 assert( nTries
==3 );
6958 if( 0==proxyBreakConchLock(pFile
, myHostID
) ){
6960 if( lockType
==EXCLUSIVE_LOCK
){
6961 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
6964 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
6968 } while( rc
==SQLITE_BUSY
&& nTries
<3 );
6973 /* Takes the conch by taking a shared lock and read the contents conch, if
6974 ** lockPath is non-NULL, the host ID and lock file path must match. A NULL
6975 ** lockPath means that the lockPath in the conch file will be used if the
6976 ** host IDs match, or a new lock path will be generated automatically
6977 ** and written to the conch file.
6979 static int proxyTakeConch(unixFile
*pFile
){
6980 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6982 if( pCtx
->conchHeld
!=0 ){
6985 unixFile
*conchFile
= pCtx
->conchFile
;
6988 char readBuf
[PROXY_MAXCONCHLEN
];
6989 char lockPath
[MAXPATHLEN
];
6990 char *tempLockPath
= NULL
;
6992 int createConch
= 0;
6993 int hostIdMatch
= 0;
6995 int tryOldLockPath
= 0;
6996 int forceNewLockPath
= 0;
6998 OSTRACE(("TAKECONCH %d for %s pid=%d\n", conchFile
->h
,
6999 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
7002 rc
= proxyGetHostID(myHostID
, &pError
);
7003 if( (rc
&0xff)==SQLITE_IOERR
){
7004 storeLastErrno(pFile
, pError
);
7007 rc
= proxyConchLock(pFile
, myHostID
, SHARED_LOCK
);
7008 if( rc
!=SQLITE_OK
){
7011 /* read the existing conch file */
7012 readLen
= seekAndRead((unixFile
*)conchFile
, 0, readBuf
, PROXY_MAXCONCHLEN
);
7014 /* I/O error: lastErrno set by seekAndRead */
7015 storeLastErrno(pFile
, conchFile
->lastErrno
);
7016 rc
= SQLITE_IOERR_READ
;
7018 }else if( readLen
<=(PROXY_HEADERLEN
+PROXY_HOSTIDLEN
) ||
7019 readBuf
[0]!=(char)PROXY_CONCHVERSION
){
7020 /* a short read or version format mismatch means we need to create a new
7025 /* if the host id matches and the lock path already exists in the conch
7026 ** we'll try to use the path there, if we can't open that path, we'll
7027 ** retry with a new auto-generated path
7029 do { /* in case we need to try again for an :auto: named lock file */
7031 if( !createConch
&& !forceNewLockPath
){
7032 hostIdMatch
= !memcmp(&readBuf
[PROXY_HEADERLEN
], myHostID
,
7034 /* if the conch has data compare the contents */
7035 if( !pCtx
->lockProxyPath
){
7036 /* for auto-named local lock file, just check the host ID and we'll
7037 ** use the local lock file path that's already in there
7040 size_t pathLen
= (readLen
- PROXY_PATHINDEX
);
7042 if( pathLen
>=MAXPATHLEN
){
7043 pathLen
=MAXPATHLEN
-1;
7045 memcpy(lockPath
, &readBuf
[PROXY_PATHINDEX
], pathLen
);
7046 lockPath
[pathLen
] = 0;
7047 tempLockPath
= lockPath
;
7049 /* create a copy of the lock path if the conch is taken */
7052 }else if( hostIdMatch
7053 && !strncmp(pCtx
->lockProxyPath
, &readBuf
[PROXY_PATHINDEX
],
7054 readLen
-PROXY_PATHINDEX
)
7056 /* conch host and lock path match */
7061 /* if the conch isn't writable and doesn't match, we can't take it */
7062 if( (conchFile
->openFlags
&O_RDWR
) == 0 ){
7067 /* either the conch didn't match or we need to create a new one */
7068 if( !pCtx
->lockProxyPath
){
7069 proxyGetLockPath(pCtx
->dbPath
, lockPath
, MAXPATHLEN
);
7070 tempLockPath
= lockPath
;
7071 /* create a copy of the lock path _only_ if the conch is taken */
7074 /* update conch with host and path (this will fail if other process
7075 ** has a shared lock already), if the host id matches, use the big
7078 futimes(conchFile
->h
, NULL
);
7079 if( hostIdMatch
&& !createConch
){
7080 if( conchFile
->pInode
&& conchFile
->pInode
->nShared
>1 ){
7081 /* We are trying for an exclusive lock but another thread in this
7082 ** same process is still holding a shared lock. */
7085 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7088 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7090 if( rc
==SQLITE_OK
){
7091 char writeBuffer
[PROXY_MAXCONCHLEN
];
7094 writeBuffer
[0] = (char)PROXY_CONCHVERSION
;
7095 memcpy(&writeBuffer
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
);
7096 if( pCtx
->lockProxyPath
!=NULL
){
7097 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], pCtx
->lockProxyPath
,
7100 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], tempLockPath
, MAXPATHLEN
);
7102 writeSize
= PROXY_PATHINDEX
+ strlen(&writeBuffer
[PROXY_PATHINDEX
]);
7103 robust_ftruncate(conchFile
->h
, writeSize
);
7104 rc
= unixWrite((sqlite3_file
*)conchFile
, writeBuffer
, writeSize
, 0);
7105 full_fsync(conchFile
->h
,0,0);
7106 /* If we created a new conch file (not just updated the contents of a
7107 ** valid conch file), try to match the permissions of the database
7109 if( rc
==SQLITE_OK
&& createConch
){
7111 int err
= osFstat(pFile
->h
, &buf
);
7113 mode_t cmode
= buf
.st_mode
&(S_IRUSR
|S_IWUSR
| S_IRGRP
|S_IWGRP
|
7115 /* try to match the database file R/W permissions, ignore failure */
7116 #ifndef SQLITE_PROXY_DEBUG
7117 osFchmod(conchFile
->h
, cmode
);
7120 rc
= osFchmod(conchFile
->h
, cmode
);
7121 }while( rc
==(-1) && errno
==EINTR
);
7124 fprintf(stderr
, "fchmod %o FAILED with %d %s\n",
7125 cmode
, code
, strerror(code
));
7127 fprintf(stderr
, "fchmod %o SUCCEDED\n",cmode
);
7131 fprintf(stderr
, "STAT FAILED[%d] with %d %s\n",
7132 err
, code
, strerror(code
));
7137 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
7140 OSTRACE(("TRANSPROXY: CLOSE %d\n", pFile
->h
));
7141 if( rc
==SQLITE_OK
&& pFile
->openFlags
){
7144 robust_close(pFile
, pFile
->h
, __LINE__
);
7147 fd
= robust_open(pCtx
->dbPath
, pFile
->openFlags
, 0);
7148 OSTRACE(("TRANSPROXY: OPEN %d\n", fd
));
7152 rc
=SQLITE_CANTOPEN_BKPT
; /* SQLITE_BUSY? proxyTakeConch called
7156 if( rc
==SQLITE_OK
&& !pCtx
->lockProxy
){
7157 char *path
= tempLockPath
? tempLockPath
: pCtx
->lockProxyPath
;
7158 rc
= proxyCreateUnixFile(path
, &pCtx
->lockProxy
, 1);
7159 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_NOMEM
&& tryOldLockPath
){
7160 /* we couldn't create the proxy lock file with the old lock file path
7161 ** so try again via auto-naming
7163 forceNewLockPath
= 1;
7165 continue; /* go back to the do {} while start point, try again */
7168 if( rc
==SQLITE_OK
){
7169 /* Need to make a copy of path if we extracted the value
7170 ** from the conch file or the path was allocated on the stack
7173 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, tempLockPath
);
7174 if( !pCtx
->lockProxyPath
){
7175 rc
= SQLITE_NOMEM_BKPT
;
7179 if( rc
==SQLITE_OK
){
7180 pCtx
->conchHeld
= 1;
7182 if( pCtx
->lockProxy
->pMethod
== &afpIoMethods
){
7183 afpLockingContext
*afpCtx
;
7184 afpCtx
= (afpLockingContext
*)pCtx
->lockProxy
->lockingContext
;
7185 afpCtx
->dbPath
= pCtx
->lockProxyPath
;
7188 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7190 OSTRACE(("TAKECONCH %d %s\n", conchFile
->h
,
7191 rc
==SQLITE_OK
?"ok":"failed"));
7193 } while (1); /* in case we need to retry the :auto: lock file -
7194 ** we should never get here except via the 'continue' call. */
7199 ** If pFile holds a lock on a conch file, then release that lock.
7201 static int proxyReleaseConch(unixFile
*pFile
){
7202 int rc
= SQLITE_OK
; /* Subroutine return code */
7203 proxyLockingContext
*pCtx
; /* The locking context for the proxy lock */
7204 unixFile
*conchFile
; /* Name of the conch file */
7206 pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7207 conchFile
= pCtx
->conchFile
;
7208 OSTRACE(("RELEASECONCH %d for %s pid=%d\n", conchFile
->h
,
7209 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
7211 if( pCtx
->conchHeld
>0 ){
7212 rc
= conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7214 pCtx
->conchHeld
= 0;
7215 OSTRACE(("RELEASECONCH %d %s\n", conchFile
->h
,
7216 (rc
==SQLITE_OK
? "ok" : "failed")));
7221 ** Given the name of a database file, compute the name of its conch file.
7222 ** Store the conch filename in memory obtained from sqlite3_malloc64().
7223 ** Make *pConchPath point to the new name. Return SQLITE_OK on success
7224 ** or SQLITE_NOMEM if unable to obtain memory.
7226 ** The caller is responsible for ensuring that the allocated memory
7227 ** space is eventually freed.
7229 ** *pConchPath is set to NULL if a memory allocation error occurs.
7231 static int proxyCreateConchPathname(char *dbPath
, char **pConchPath
){
7232 int i
; /* Loop counter */
7233 int len
= (int)strlen(dbPath
); /* Length of database filename - dbPath */
7234 char *conchPath
; /* buffer in which to construct conch name */
7236 /* Allocate space for the conch filename and initialize the name to
7237 ** the name of the original database file. */
7238 *pConchPath
= conchPath
= (char *)sqlite3_malloc64(len
+ 8);
7240 return SQLITE_NOMEM_BKPT
;
7242 memcpy(conchPath
, dbPath
, len
+1);
7244 /* now insert a "." before the last / character */
7245 for( i
=(len
-1); i
>=0; i
-- ){
7246 if( conchPath
[i
]=='/' ){
7253 conchPath
[i
+1]=dbPath
[i
];
7257 /* append the "-conch" suffix to the file */
7258 memcpy(&conchPath
[i
+1], "-conch", 7);
7259 assert( (int)strlen(conchPath
) == len
+7 );
7265 /* Takes a fully configured proxy locking-style unix file and switches
7266 ** the local lock file path
7268 static int switchLockProxyPath(unixFile
*pFile
, const char *path
) {
7269 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7270 char *oldPath
= pCtx
->lockProxyPath
;
7273 if( pFile
->eFileLock
!=NO_LOCK
){
7277 /* nothing to do if the path is NULL, :auto: or matches the existing path */
7278 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ||
7279 (oldPath
&& !strncmp(oldPath
, path
, MAXPATHLEN
)) ){
7282 unixFile
*lockProxy
= pCtx
->lockProxy
;
7283 pCtx
->lockProxy
=NULL
;
7284 pCtx
->conchHeld
= 0;
7285 if( lockProxy
!=NULL
){
7286 rc
=lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7288 sqlite3_free(lockProxy
);
7290 sqlite3_free(oldPath
);
7291 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, path
);
7298 ** pFile is a file that has been opened by a prior xOpen call. dbPath
7299 ** is a string buffer at least MAXPATHLEN+1 characters in size.
7301 ** This routine find the filename associated with pFile and writes it
7304 static int proxyGetDbPathForUnixFile(unixFile
*pFile
, char *dbPath
){
7305 #if defined(__APPLE__)
7306 if( pFile
->pMethod
== &afpIoMethods
){
7307 /* afp style keeps a reference to the db path in the filePath field
7309 assert( (int)strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7310 strlcpy(dbPath
, ((afpLockingContext
*)pFile
->lockingContext
)->dbPath
,
7314 if( pFile
->pMethod
== &dotlockIoMethods
){
7315 /* dot lock style uses the locking context to store the dot lock
7317 int len
= strlen((char *)pFile
->lockingContext
) - strlen(DOTLOCK_SUFFIX
);
7318 memcpy(dbPath
, (char *)pFile
->lockingContext
, len
+ 1);
7320 /* all other styles use the locking context to store the db file path */
7321 assert( strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7322 strlcpy(dbPath
, (char *)pFile
->lockingContext
, MAXPATHLEN
);
7328 ** Takes an already filled in unix file and alters it so all file locking
7329 ** will be performed on the local proxy lock file. The following fields
7330 ** are preserved in the locking context so that they can be restored and
7331 ** the unix structure properly cleaned up at close time:
7335 static int proxyTransformUnixFile(unixFile
*pFile
, const char *path
) {
7336 proxyLockingContext
*pCtx
;
7337 char dbPath
[MAXPATHLEN
+1]; /* Name of the database file */
7338 char *lockPath
=NULL
;
7341 if( pFile
->eFileLock
!=NO_LOCK
){
7344 proxyGetDbPathForUnixFile(pFile
, dbPath
);
7345 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ){
7348 lockPath
=(char *)path
;
7351 OSTRACE(("TRANSPROXY %d for %s pid=%d\n", pFile
->h
,
7352 (lockPath
? lockPath
: ":auto:"), osGetpid(0)));
7354 pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
7356 return SQLITE_NOMEM_BKPT
;
7358 memset(pCtx
, 0, sizeof(*pCtx
));
7360 rc
= proxyCreateConchPathname(dbPath
, &pCtx
->conchFilePath
);
7361 if( rc
==SQLITE_OK
){
7362 rc
= proxyCreateUnixFile(pCtx
->conchFilePath
, &pCtx
->conchFile
, 0);
7363 if( rc
==SQLITE_CANTOPEN
&& ((pFile
->openFlags
&O_RDWR
) == 0) ){
7364 /* if (a) the open flags are not O_RDWR, (b) the conch isn't there, and
7365 ** (c) the file system is read-only, then enable no-locking access.
7366 ** Ugh, since O_RDONLY==0x0000 we test for !O_RDWR since unixOpen asserts
7367 ** that openFlags will have only one of O_RDONLY or O_RDWR.
7369 struct statfs fsInfo
;
7370 struct stat conchInfo
;
7373 if( osStat(pCtx
->conchFilePath
, &conchInfo
) == -1 ) {
7375 if( (err
==ENOENT
) && (statfs(dbPath
, &fsInfo
) != -1) ){
7376 goLockless
= (fsInfo
.f_flags
&MNT_RDONLY
) == MNT_RDONLY
;
7380 pCtx
->conchHeld
= -1; /* read only FS/ lockless */
7385 if( rc
==SQLITE_OK
&& lockPath
){
7386 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, lockPath
);
7389 if( rc
==SQLITE_OK
){
7390 pCtx
->dbPath
= sqlite3DbStrDup(0, dbPath
);
7391 if( pCtx
->dbPath
==NULL
){
7392 rc
= SQLITE_NOMEM_BKPT
;
7395 if( rc
==SQLITE_OK
){
7396 /* all memory is allocated, proxys are created and assigned,
7397 ** switch the locking context and pMethod then return.
7399 pCtx
->oldLockingContext
= pFile
->lockingContext
;
7400 pFile
->lockingContext
= pCtx
;
7401 pCtx
->pOldMethod
= pFile
->pMethod
;
7402 pFile
->pMethod
= &proxyIoMethods
;
7404 if( pCtx
->conchFile
){
7405 pCtx
->conchFile
->pMethod
->xClose((sqlite3_file
*)pCtx
->conchFile
);
7406 sqlite3_free(pCtx
->conchFile
);
7408 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7409 sqlite3_free(pCtx
->conchFilePath
);
7412 OSTRACE(("TRANSPROXY %d %s\n", pFile
->h
,
7413 (rc
==SQLITE_OK
? "ok" : "failed")));
7419 ** This routine handles sqlite3_file_control() calls that are specific
7420 ** to proxy locking.
7422 static int proxyFileControl(sqlite3_file
*id
, int op
, void *pArg
){
7424 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
7425 unixFile
*pFile
= (unixFile
*)id
;
7426 if( pFile
->pMethod
== &proxyIoMethods
){
7427 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7428 proxyTakeConch(pFile
);
7429 if( pCtx
->lockProxyPath
){
7430 *(const char **)pArg
= pCtx
->lockProxyPath
;
7432 *(const char **)pArg
= ":auto: (not held)";
7435 *(const char **)pArg
= NULL
;
7439 case SQLITE_FCNTL_SET_LOCKPROXYFILE
: {
7440 unixFile
*pFile
= (unixFile
*)id
;
7442 int isProxyStyle
= (pFile
->pMethod
== &proxyIoMethods
);
7443 if( pArg
==NULL
|| (const char *)pArg
==0 ){
7445 /* turn off proxy locking - not supported. If support is added for
7446 ** switching proxy locking mode off then it will need to fail if
7447 ** the journal mode is WAL mode.
7449 rc
= SQLITE_ERROR
/*SQLITE_PROTOCOL? SQLITE_MISUSE?*/;
7451 /* turn off proxy locking - already off - NOOP */
7455 const char *proxyPath
= (const char *)pArg
;
7457 proxyLockingContext
*pCtx
=
7458 (proxyLockingContext
*)pFile
->lockingContext
;
7459 if( !strcmp(pArg
, ":auto:")
7460 || (pCtx
->lockProxyPath
&&
7461 !strncmp(pCtx
->lockProxyPath
, proxyPath
, MAXPATHLEN
))
7465 rc
= switchLockProxyPath(pFile
, proxyPath
);
7468 /* turn on proxy file locking */
7469 rc
= proxyTransformUnixFile(pFile
, proxyPath
);
7475 assert( 0 ); /* The call assures that only valid opcodes are sent */
7479 return SQLITE_ERROR
;
7483 ** Within this division (the proxying locking implementation) the procedures
7484 ** above this point are all utilities. The lock-related methods of the
7485 ** proxy-locking sqlite3_io_method object follow.
7490 ** This routine checks if there is a RESERVED lock held on the specified
7491 ** file by this or any other process. If such a lock is held, set *pResOut
7492 ** to a non-zero value otherwise *pResOut is set to zero. The return value
7493 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
7495 static int proxyCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
7496 unixFile
*pFile
= (unixFile
*)id
;
7497 int rc
= proxyTakeConch(pFile
);
7498 if( rc
==SQLITE_OK
){
7499 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7500 if( pCtx
->conchHeld
>0 ){
7501 unixFile
*proxy
= pCtx
->lockProxy
;
7502 return proxy
->pMethod
->xCheckReservedLock((sqlite3_file
*)proxy
, pResOut
);
7503 }else{ /* conchHeld < 0 is lockless */
7511 ** Lock the file with the lock specified by parameter eFileLock - one
7512 ** of the following:
7515 ** (2) RESERVED_LOCK
7517 ** (4) EXCLUSIVE_LOCK
7519 ** Sometimes when requesting one lock state, additional lock states
7520 ** are inserted in between. The locking might fail on one of the later
7521 ** transitions leaving the lock state different from what it started but
7522 ** still short of its goal. The following chart shows the allowed
7523 ** transitions and the inserted intermediate states:
7525 ** UNLOCKED -> SHARED
7526 ** SHARED -> RESERVED
7527 ** SHARED -> (PENDING) -> EXCLUSIVE
7528 ** RESERVED -> (PENDING) -> EXCLUSIVE
7529 ** PENDING -> EXCLUSIVE
7531 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
7532 ** routine to lower a locking level.
7534 static int proxyLock(sqlite3_file
*id
, int eFileLock
) {
7535 unixFile
*pFile
= (unixFile
*)id
;
7536 int rc
= proxyTakeConch(pFile
);
7537 if( rc
==SQLITE_OK
){
7538 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7539 if( pCtx
->conchHeld
>0 ){
7540 unixFile
*proxy
= pCtx
->lockProxy
;
7541 rc
= proxy
->pMethod
->xLock((sqlite3_file
*)proxy
, eFileLock
);
7542 pFile
->eFileLock
= proxy
->eFileLock
;
7544 /* conchHeld < 0 is lockless */
7552 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
7553 ** must be either NO_LOCK or SHARED_LOCK.
7555 ** If the locking level of the file descriptor is already at or below
7556 ** the requested locking level, this routine is a no-op.
7558 static int proxyUnlock(sqlite3_file
*id
, int eFileLock
) {
7559 unixFile
*pFile
= (unixFile
*)id
;
7560 int rc
= proxyTakeConch(pFile
);
7561 if( rc
==SQLITE_OK
){
7562 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7563 if( pCtx
->conchHeld
>0 ){
7564 unixFile
*proxy
= pCtx
->lockProxy
;
7565 rc
= proxy
->pMethod
->xUnlock((sqlite3_file
*)proxy
, eFileLock
);
7566 pFile
->eFileLock
= proxy
->eFileLock
;
7568 /* conchHeld < 0 is lockless */
7575 ** Close a file that uses proxy locks.
7577 static int proxyClose(sqlite3_file
*id
) {
7579 unixFile
*pFile
= (unixFile
*)id
;
7580 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7581 unixFile
*lockProxy
= pCtx
->lockProxy
;
7582 unixFile
*conchFile
= pCtx
->conchFile
;
7586 rc
= lockProxy
->pMethod
->xUnlock((sqlite3_file
*)lockProxy
, NO_LOCK
);
7588 rc
= lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7590 sqlite3_free(lockProxy
);
7591 pCtx
->lockProxy
= 0;
7594 if( pCtx
->conchHeld
){
7595 rc
= proxyReleaseConch(pFile
);
7598 rc
= conchFile
->pMethod
->xClose((sqlite3_file
*)conchFile
);
7600 sqlite3_free(conchFile
);
7602 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7603 sqlite3_free(pCtx
->conchFilePath
);
7604 sqlite3DbFree(0, pCtx
->dbPath
);
7605 /* restore the original locking context and pMethod then close it */
7606 pFile
->lockingContext
= pCtx
->oldLockingContext
;
7607 pFile
->pMethod
= pCtx
->pOldMethod
;
7609 return pFile
->pMethod
->xClose(id
);
7616 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
7618 ** The proxy locking style is intended for use with AFP filesystems.
7619 ** And since AFP is only supported on MacOSX, the proxy locking is also
7620 ** restricted to MacOSX.
7623 ******************* End of the proxy lock implementation **********************
7624 ******************************************************************************/
7627 ** Initialize the operating system interface.
7629 ** This routine registers all VFS implementations for unix-like operating
7630 ** systems. This routine, and the sqlite3_os_end() routine that follows,
7631 ** should be the only routines in this file that are visible from other
7634 ** This routine is called once during SQLite initialization and by a
7635 ** single thread. The memory allocation and mutex subsystems have not
7636 ** necessarily been initialized when this routine is called, and so they
7637 ** should not be used.
7639 int sqlite3_os_init(void){
7641 ** The following macro defines an initializer for an sqlite3_vfs object.
7642 ** The name of the VFS is NAME. The pAppData is a pointer to a pointer
7643 ** to the "finder" function. (pAppData is a pointer to a pointer because
7644 ** silly C90 rules prohibit a void* from being cast to a function pointer
7645 ** and so we have to go through the intermediate pointer to avoid problems
7646 ** when compiling with -pedantic-errors on GCC.)
7648 ** The FINDER parameter to this macro is the name of the pointer to the
7649 ** finder-function. The finder-function returns a pointer to the
7650 ** sqlite_io_methods object that implements the desired locking
7651 ** behaviors. See the division above that contains the IOMETHODS
7652 ** macro for addition information on finder-functions.
7654 ** Most finders simply return a pointer to a fixed sqlite3_io_methods
7655 ** object. But the "autolockIoFinder" available on MacOSX does a little
7656 ** more than that; it looks at the filesystem type that hosts the
7657 ** database file and tries to choose an locking method appropriate for
7658 ** that filesystem time.
7660 #define UNIXVFS(VFSNAME, FINDER) { \
7662 sizeof(unixFile), /* szOsFile */ \
7663 MAX_PATHNAME, /* mxPathname */ \
7665 VFSNAME, /* zName */ \
7666 (void*)&FINDER, /* pAppData */ \
7667 unixOpen, /* xOpen */ \
7668 unixDelete, /* xDelete */ \
7669 unixAccess, /* xAccess */ \
7670 unixFullPathname, /* xFullPathname */ \
7671 unixDlOpen, /* xDlOpen */ \
7672 unixDlError, /* xDlError */ \
7673 unixDlSym, /* xDlSym */ \
7674 unixDlClose, /* xDlClose */ \
7675 unixRandomness, /* xRandomness */ \
7676 unixSleep, /* xSleep */ \
7677 unixCurrentTime, /* xCurrentTime */ \
7678 unixGetLastError, /* xGetLastError */ \
7679 unixCurrentTimeInt64, /* xCurrentTimeInt64 */ \
7680 unixSetSystemCall, /* xSetSystemCall */ \
7681 unixGetSystemCall, /* xGetSystemCall */ \
7682 unixNextSystemCall, /* xNextSystemCall */ \
7686 ** All default VFSes for unix are contained in the following array.
7688 ** Note that the sqlite3_vfs.pNext field of the VFS object is modified
7689 ** by the SQLite core when the VFS is registered. So the following
7690 ** array cannot be const.
7692 static sqlite3_vfs aVfs
[] = {
7693 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7694 UNIXVFS("unix", autolockIoFinder
),
7696 UNIXVFS("unix", vxworksIoFinder
),
7698 UNIXVFS("unix", posixIoFinder
),
7700 UNIXVFS("unix-none", nolockIoFinder
),
7701 UNIXVFS("unix-dotfile", dotlockIoFinder
),
7702 UNIXVFS("unix-excl", posixIoFinder
),
7704 UNIXVFS("unix-namedsem", semIoFinder
),
7706 #if SQLITE_ENABLE_LOCKING_STYLE || OS_VXWORKS
7707 UNIXVFS("unix-posix", posixIoFinder
),
7709 #if SQLITE_ENABLE_LOCKING_STYLE
7710 UNIXVFS("unix-flock", flockIoFinder
),
7712 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7713 UNIXVFS("unix-afp", afpIoFinder
),
7714 UNIXVFS("unix-nfs", nfsIoFinder
),
7715 UNIXVFS("unix-proxy", proxyIoFinder
),
7718 unsigned int i
; /* Loop counter */
7720 /* Double-check that the aSyscall[] array has been constructed
7721 ** correctly. See ticket [bb3a86e890c8e96ab] */
7722 assert( ArraySize(aSyscall
)==29 );
7724 /* Register all VFSes defined in the aVfs[] array */
7725 for(i
=0; i
<(sizeof(aVfs
)/sizeof(sqlite3_vfs
)); i
++){
7726 sqlite3_vfs_register(&aVfs
[i
], i
==0);
7728 unixBigLock
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_VFS1
);
7733 ** Shutdown the operating system interface.
7735 ** Some operating systems might need to do some cleanup in this routine,
7736 ** to release dynamically allocated objects. But not on unix.
7737 ** This routine is a no-op for unix.
7739 int sqlite3_os_end(void){
7744 #endif /* SQLITE_OS_UNIX */