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() */
232 #ifdef SQLITE_ENABLE_SETLK_TIMEOUT
233 unsigned iBusyTimeout
; /* Wait this many millisec on locks */
236 struct vxworksFileId
*pId
; /* Unique file ID */
239 /* The next group of variables are used to track whether or not the
240 ** transaction counter in bytes 24-27 of database files are updated
241 ** whenever any part of the database changes. An assertion fault will
242 ** occur if a file is updated without also updating the transaction
243 ** counter. This test is made to avoid new problems similar to the
244 ** one described by ticket #3584.
246 unsigned char transCntrChng
; /* True if the transaction counter changed */
247 unsigned char dbUpdate
; /* True if any part of database file changed */
248 unsigned char inNormalWrite
; /* True if in a normal write operation */
253 /* In test mode, increase the size of this structure a bit so that
254 ** it is larger than the struct CrashFile defined in test6.c.
260 /* This variable holds the process id (pid) from when the xRandomness()
261 ** method was called. If xOpen() is called from a different process id,
262 ** indicating that a fork() has occurred, the PRNG will be reset.
264 static pid_t randomnessPid
= 0;
267 ** Allowed values for the unixFile.ctrlFlags bitmask:
269 #define UNIXFILE_EXCL 0x01 /* Connections from one process only */
270 #define UNIXFILE_RDONLY 0x02 /* Connection is read only */
271 #define UNIXFILE_PERSIST_WAL 0x04 /* Persistent WAL mode */
272 #ifndef SQLITE_DISABLE_DIRSYNC
273 # define UNIXFILE_DIRSYNC 0x08 /* Directory sync needed */
275 # define UNIXFILE_DIRSYNC 0x00
277 #define UNIXFILE_PSOW 0x10 /* SQLITE_IOCAP_POWERSAFE_OVERWRITE */
278 #define UNIXFILE_DELETE 0x20 /* Delete on close */
279 #define UNIXFILE_URI 0x40 /* Filename might have query parameters */
280 #define UNIXFILE_NOLOCK 0x80 /* Do no file locking */
283 ** Include code that is common to all os_*.c files
285 #include "os_common.h"
288 ** Define various macros that are missing from some systems.
291 # define O_LARGEFILE 0
293 #ifdef SQLITE_DISABLE_LFS
295 # define O_LARGEFILE 0
298 # define O_NOFOLLOW 0
305 ** The threadid macro resolves to the thread-id or to 0. Used for
306 ** testing and debugging only.
308 #if SQLITE_THREADSAFE
309 #define threadid pthread_self()
315 ** HAVE_MREMAP defaults to true on Linux and false everywhere else.
317 #if !defined(HAVE_MREMAP)
318 # if defined(__linux__) && defined(_GNU_SOURCE)
319 # define HAVE_MREMAP 1
321 # define HAVE_MREMAP 0
326 ** Explicitly call the 64-bit version of lseek() on Android. Otherwise, lseek()
327 ** is the 32-bit version, even if _FILE_OFFSET_BITS=64 is defined.
330 # define lseek lseek64
335 ** Linux-specific IOCTL magic numbers used for controlling F2FS
337 #define F2FS_IOCTL_MAGIC 0xf5
338 #define F2FS_IOC_START_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 1)
339 #define F2FS_IOC_COMMIT_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 2)
340 #define F2FS_IOC_START_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 3)
341 #define F2FS_IOC_ABORT_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 5)
342 #define F2FS_IOC_GET_FEATURES _IOR(F2FS_IOCTL_MAGIC, 12, u32)
343 #define F2FS_FEATURE_ATOMIC_WRITE 0x0004
344 #endif /* __linux__ */
348 ** Different Unix systems declare open() in different ways. Same use
349 ** open(const char*,int,mode_t). Others use open(const char*,int,...).
350 ** The difference is important when using a pointer to the function.
352 ** The safest way to deal with the problem is to always use this wrapper
353 ** which always has the same well-defined interface.
355 static int posixOpen(const char *zFile
, int flags
, int mode
){
356 return open(zFile
, flags
, mode
);
359 /* Forward reference */
360 static int openDirectory(const char*, int*);
361 static int unixGetpagesize(void);
364 ** Many system calls are accessed through pointer-to-functions so that
365 ** they may be overridden at runtime to facilitate fault injection during
366 ** testing and sandboxing. The following array holds the names and pointers
367 ** to all overrideable system calls.
369 static struct unix_syscall
{
370 const char *zName
; /* Name of the system call */
371 sqlite3_syscall_ptr pCurrent
; /* Current value of the system call */
372 sqlite3_syscall_ptr pDefault
; /* Default value */
374 { "open", (sqlite3_syscall_ptr
)posixOpen
, 0 },
375 #define osOpen ((int(*)(const char*,int,int))aSyscall[0].pCurrent)
377 { "close", (sqlite3_syscall_ptr
)close
, 0 },
378 #define osClose ((int(*)(int))aSyscall[1].pCurrent)
380 { "access", (sqlite3_syscall_ptr
)access
, 0 },
381 #define osAccess ((int(*)(const char*,int))aSyscall[2].pCurrent)
383 { "getcwd", (sqlite3_syscall_ptr
)getcwd
, 0 },
384 #define osGetcwd ((char*(*)(char*,size_t))aSyscall[3].pCurrent)
386 { "stat", (sqlite3_syscall_ptr
)stat
, 0 },
387 #define osStat ((int(*)(const char*,struct stat*))aSyscall[4].pCurrent)
390 ** The DJGPP compiler environment looks mostly like Unix, but it
391 ** lacks the fcntl() system call. So redefine fcntl() to be something
392 ** that always succeeds. This means that locking does not occur under
393 ** DJGPP. But it is DOS - what did you expect?
397 #define osFstat(a,b,c) 0
399 { "fstat", (sqlite3_syscall_ptr
)fstat
, 0 },
400 #define osFstat ((int(*)(int,struct stat*))aSyscall[5].pCurrent)
403 { "ftruncate", (sqlite3_syscall_ptr
)ftruncate
, 0 },
404 #define osFtruncate ((int(*)(int,off_t))aSyscall[6].pCurrent)
406 { "fcntl", (sqlite3_syscall_ptr
)fcntl
, 0 },
407 #define osFcntl ((int(*)(int,int,...))aSyscall[7].pCurrent)
409 { "read", (sqlite3_syscall_ptr
)read
, 0 },
410 #define osRead ((ssize_t(*)(int,void*,size_t))aSyscall[8].pCurrent)
412 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
413 { "pread", (sqlite3_syscall_ptr
)pread
, 0 },
415 { "pread", (sqlite3_syscall_ptr
)0, 0 },
417 #define osPread ((ssize_t(*)(int,void*,size_t,off_t))aSyscall[9].pCurrent)
419 #if defined(USE_PREAD64)
420 { "pread64", (sqlite3_syscall_ptr
)pread64
, 0 },
422 { "pread64", (sqlite3_syscall_ptr
)0, 0 },
424 #define osPread64 ((ssize_t(*)(int,void*,size_t,off64_t))aSyscall[10].pCurrent)
426 { "write", (sqlite3_syscall_ptr
)write
, 0 },
427 #define osWrite ((ssize_t(*)(int,const void*,size_t))aSyscall[11].pCurrent)
429 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
430 { "pwrite", (sqlite3_syscall_ptr
)pwrite
, 0 },
432 { "pwrite", (sqlite3_syscall_ptr
)0, 0 },
434 #define osPwrite ((ssize_t(*)(int,const void*,size_t,off_t))\
435 aSyscall[12].pCurrent)
437 #if defined(USE_PREAD64)
438 { "pwrite64", (sqlite3_syscall_ptr
)pwrite64
, 0 },
440 { "pwrite64", (sqlite3_syscall_ptr
)0, 0 },
442 #define osPwrite64 ((ssize_t(*)(int,const void*,size_t,off64_t))\
443 aSyscall[13].pCurrent)
445 { "fchmod", (sqlite3_syscall_ptr
)fchmod
, 0 },
446 #define osFchmod ((int(*)(int,mode_t))aSyscall[14].pCurrent)
448 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
449 { "fallocate", (sqlite3_syscall_ptr
)posix_fallocate
, 0 },
451 { "fallocate", (sqlite3_syscall_ptr
)0, 0 },
453 #define osFallocate ((int(*)(int,off_t,off_t))aSyscall[15].pCurrent)
455 { "unlink", (sqlite3_syscall_ptr
)unlink
, 0 },
456 #define osUnlink ((int(*)(const char*))aSyscall[16].pCurrent)
458 { "openDirectory", (sqlite3_syscall_ptr
)openDirectory
, 0 },
459 #define osOpenDirectory ((int(*)(const char*,int*))aSyscall[17].pCurrent)
461 { "mkdir", (sqlite3_syscall_ptr
)mkdir
, 0 },
462 #define osMkdir ((int(*)(const char*,mode_t))aSyscall[18].pCurrent)
464 { "rmdir", (sqlite3_syscall_ptr
)rmdir
, 0 },
465 #define osRmdir ((int(*)(const char*))aSyscall[19].pCurrent)
467 #if defined(HAVE_FCHOWN)
468 { "fchown", (sqlite3_syscall_ptr
)fchown
, 0 },
470 { "fchown", (sqlite3_syscall_ptr
)0, 0 },
472 #define osFchown ((int(*)(int,uid_t,gid_t))aSyscall[20].pCurrent)
474 #if defined(HAVE_FCHOWN)
475 { "geteuid", (sqlite3_syscall_ptr
)geteuid
, 0 },
477 { "geteuid", (sqlite3_syscall_ptr
)0, 0 },
479 #define osGeteuid ((uid_t(*)(void))aSyscall[21].pCurrent)
481 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
482 { "mmap", (sqlite3_syscall_ptr
)mmap
, 0 },
484 { "mmap", (sqlite3_syscall_ptr
)0, 0 },
486 #define osMmap ((void*(*)(void*,size_t,int,int,int,off_t))aSyscall[22].pCurrent)
488 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
489 { "munmap", (sqlite3_syscall_ptr
)munmap
, 0 },
491 { "munmap", (sqlite3_syscall_ptr
)0, 0 },
493 #define osMunmap ((int(*)(void*,size_t))aSyscall[23].pCurrent)
495 #if HAVE_MREMAP && (!defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0)
496 { "mremap", (sqlite3_syscall_ptr
)mremap
, 0 },
498 { "mremap", (sqlite3_syscall_ptr
)0, 0 },
500 #define osMremap ((void*(*)(void*,size_t,size_t,int,...))aSyscall[24].pCurrent)
502 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
503 { "getpagesize", (sqlite3_syscall_ptr
)unixGetpagesize
, 0 },
505 { "getpagesize", (sqlite3_syscall_ptr
)0, 0 },
507 #define osGetpagesize ((int(*)(void))aSyscall[25].pCurrent)
509 #if defined(HAVE_READLINK)
510 { "readlink", (sqlite3_syscall_ptr
)readlink
, 0 },
512 { "readlink", (sqlite3_syscall_ptr
)0, 0 },
514 #define osReadlink ((ssize_t(*)(const char*,char*,size_t))aSyscall[26].pCurrent)
516 #if defined(HAVE_LSTAT)
517 { "lstat", (sqlite3_syscall_ptr
)lstat
, 0 },
519 { "lstat", (sqlite3_syscall_ptr
)0, 0 },
521 #define osLstat ((int(*)(const char*,struct stat*))aSyscall[27].pCurrent)
523 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
524 { "ioctl", (sqlite3_syscall_ptr
)ioctl
, 0 },
526 { "ioctl", (sqlite3_syscall_ptr
)0, 0 },
528 #define osIoctl ((int(*)(int,int,...))aSyscall[28].pCurrent)
530 }; /* End of the overrideable system calls */
534 ** On some systems, calls to fchown() will trigger a message in a security
535 ** log if they come from non-root processes. So avoid calling fchown() if
536 ** we are not running as root.
538 static int robustFchown(int fd
, uid_t uid
, gid_t gid
){
539 #if defined(HAVE_FCHOWN)
540 return osGeteuid() ? 0 : osFchown(fd
,uid
,gid
);
547 ** This is the xSetSystemCall() method of sqlite3_vfs for all of the
548 ** "unix" VFSes. Return SQLITE_OK opon successfully updating the
549 ** system call pointer, or SQLITE_NOTFOUND if there is no configurable
550 ** system call named zName.
552 static int unixSetSystemCall(
553 sqlite3_vfs
*pNotUsed
, /* The VFS pointer. Not used */
554 const char *zName
, /* Name of system call to override */
555 sqlite3_syscall_ptr pNewFunc
/* Pointer to new system call value */
558 int rc
= SQLITE_NOTFOUND
;
560 UNUSED_PARAMETER(pNotUsed
);
562 /* If no zName is given, restore all system calls to their default
563 ** settings and return NULL
566 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
567 if( aSyscall
[i
].pDefault
){
568 aSyscall
[i
].pCurrent
= aSyscall
[i
].pDefault
;
572 /* If zName is specified, operate on only the one system call
575 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
576 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ){
577 if( aSyscall
[i
].pDefault
==0 ){
578 aSyscall
[i
].pDefault
= aSyscall
[i
].pCurrent
;
581 if( pNewFunc
==0 ) pNewFunc
= aSyscall
[i
].pDefault
;
582 aSyscall
[i
].pCurrent
= pNewFunc
;
591 ** Return the value of a system call. Return NULL if zName is not a
592 ** recognized system call name. NULL is also returned if the system call
593 ** is currently undefined.
595 static sqlite3_syscall_ptr
unixGetSystemCall(
596 sqlite3_vfs
*pNotUsed
,
601 UNUSED_PARAMETER(pNotUsed
);
602 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
603 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) return aSyscall
[i
].pCurrent
;
609 ** Return the name of the first system call after zName. If zName==NULL
610 ** then return the name of the first system call. Return NULL if zName
611 ** is the last system call or if zName is not the name of a valid
614 static const char *unixNextSystemCall(sqlite3_vfs
*p
, const char *zName
){
619 for(i
=0; i
<ArraySize(aSyscall
)-1; i
++){
620 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) break;
623 for(i
++; i
<ArraySize(aSyscall
); i
++){
624 if( aSyscall
[i
].pCurrent
!=0 ) return aSyscall
[i
].zName
;
630 ** Do not accept any file descriptor less than this value, in order to avoid
631 ** opening database file using file descriptors that are commonly used for
632 ** standard input, output, and error.
634 #ifndef SQLITE_MINIMUM_FILE_DESCRIPTOR
635 # define SQLITE_MINIMUM_FILE_DESCRIPTOR 3
639 ** Invoke open(). Do so multiple times, until it either succeeds or
640 ** fails for some reason other than EINTR.
642 ** If the file creation mode "m" is 0 then set it to the default for
643 ** SQLite. The default is SQLITE_DEFAULT_FILE_PERMISSIONS (normally
644 ** 0644) as modified by the system umask. If m is not 0, then
645 ** make the file creation mode be exactly m ignoring the umask.
647 ** The m parameter will be non-zero only when creating -wal, -journal,
648 ** and -shm files. We want those files to have *exactly* the same
649 ** permissions as their original database, unadulterated by the umask.
650 ** In that way, if a database file is -rw-rw-rw or -rw-rw-r-, and a
651 ** transaction crashes and leaves behind hot journals, then any
652 ** process that is able to write to the database will also be able to
653 ** recover the hot journals.
655 static int robust_open(const char *z
, int f
, mode_t m
){
657 mode_t m2
= m
? m
: SQLITE_DEFAULT_FILE_PERMISSIONS
;
659 #if defined(O_CLOEXEC)
660 fd
= osOpen(z
,f
|O_CLOEXEC
,m2
);
665 if( errno
==EINTR
) continue;
668 if( fd
>=SQLITE_MINIMUM_FILE_DESCRIPTOR
) break;
670 sqlite3_log(SQLITE_WARNING
,
671 "attempt to open \"%s\" as file descriptor %d", z
, fd
);
673 if( osOpen("/dev/null", f
, m
)<0 ) break;
678 if( osFstat(fd
, &statbuf
)==0
679 && statbuf
.st_size
==0
680 && (statbuf
.st_mode
&0777)!=m
685 #if defined(FD_CLOEXEC) && (!defined(O_CLOEXEC) || O_CLOEXEC==0)
686 osFcntl(fd
, F_SETFD
, osFcntl(fd
, F_GETFD
, 0) | FD_CLOEXEC
);
693 ** Helper functions to obtain and relinquish the global mutex. The
694 ** global mutex is used to protect the unixInodeInfo and
695 ** vxworksFileId objects used by this file, all of which may be
696 ** shared by multiple threads.
698 ** Function unixMutexHeld() is used to assert() that the global mutex
699 ** is held when required. This function is only used as part of assert()
703 ** assert( unixMutexHeld() );
706 static sqlite3_mutex
*unixBigLock
= 0;
707 static void unixEnterMutex(void){
708 sqlite3_mutex_enter(unixBigLock
);
710 static void unixLeaveMutex(void){
711 sqlite3_mutex_leave(unixBigLock
);
714 static int unixMutexHeld(void) {
715 return sqlite3_mutex_held(unixBigLock
);
720 #ifdef SQLITE_HAVE_OS_TRACE
722 ** Helper function for printing out trace information from debugging
723 ** binaries. This returns the string representation of the supplied
724 ** integer lock-type.
726 static const char *azFileLock(int eFileLock
){
728 case NO_LOCK
: return "NONE";
729 case SHARED_LOCK
: return "SHARED";
730 case RESERVED_LOCK
: return "RESERVED";
731 case PENDING_LOCK
: return "PENDING";
732 case EXCLUSIVE_LOCK
: return "EXCLUSIVE";
738 #ifdef SQLITE_LOCK_TRACE
740 ** Print out information about all locking operations.
742 ** This routine is used for troubleshooting locks on multithreaded
743 ** platforms. Enable by compiling with the -DSQLITE_LOCK_TRACE
744 ** command-line option on the compiler. This code is normally
747 static int lockTrace(int fd
, int op
, struct flock
*p
){
748 char *zOpName
, *zType
;
753 }else if( op
==F_SETLK
){
756 s
= osFcntl(fd
, op
, p
);
757 sqlite3DebugPrintf("fcntl unknown %d %d %d\n", fd
, op
, s
);
760 if( p
->l_type
==F_RDLCK
){
762 }else if( p
->l_type
==F_WRLCK
){
764 }else if( p
->l_type
==F_UNLCK
){
769 assert( p
->l_whence
==SEEK_SET
);
770 s
= osFcntl(fd
, op
, p
);
772 sqlite3DebugPrintf("fcntl %d %d %s %s %d %d %d %d\n",
773 threadid
, fd
, zOpName
, zType
, (int)p
->l_start
, (int)p
->l_len
,
775 if( s
==(-1) && op
==F_SETLK
&& (p
->l_type
==F_RDLCK
|| p
->l_type
==F_WRLCK
) ){
778 osFcntl(fd
, F_GETLK
, &l2
);
779 if( l2
.l_type
==F_RDLCK
){
781 }else if( l2
.l_type
==F_WRLCK
){
783 }else if( l2
.l_type
==F_UNLCK
){
788 sqlite3DebugPrintf("fcntl-failure-reason: %s %d %d %d\n",
789 zType
, (int)l2
.l_start
, (int)l2
.l_len
, (int)l2
.l_pid
);
795 #define osFcntl lockTrace
796 #endif /* SQLITE_LOCK_TRACE */
799 ** Retry ftruncate() calls that fail due to EINTR
801 ** All calls to ftruncate() within this file should be made through
802 ** this wrapper. On the Android platform, bypassing the logic below
803 ** could lead to a corrupt database.
805 static int robust_ftruncate(int h
, sqlite3_int64 sz
){
808 /* On Android, ftruncate() always uses 32-bit offsets, even if
809 ** _FILE_OFFSET_BITS=64 is defined. This means it is unsafe to attempt to
810 ** truncate a file to any size larger than 2GiB. Silently ignore any
812 if( sz
>(sqlite3_int64
)0x7FFFFFFF ){
816 do{ rc
= osFtruncate(h
,sz
); }while( rc
<0 && errno
==EINTR
);
821 ** This routine translates a standard POSIX errno code into something
822 ** useful to the clients of the sqlite3 functions. Specifically, it is
823 ** intended to translate a variety of "try again" errors into SQLITE_BUSY
824 ** and a variety of "please close the file descriptor NOW" errors into
827 ** Errors during initialization of locks, or file system support for locks,
828 ** should handle ENOLCK, ENOTSUP, EOPNOTSUPP separately.
830 static int sqliteErrorFromPosixError(int posixError
, int sqliteIOErr
) {
831 assert( (sqliteIOErr
== SQLITE_IOERR_LOCK
) ||
832 (sqliteIOErr
== SQLITE_IOERR_UNLOCK
) ||
833 (sqliteIOErr
== SQLITE_IOERR_RDLOCK
) ||
834 (sqliteIOErr
== SQLITE_IOERR_CHECKRESERVEDLOCK
) );
835 switch (posixError
) {
842 /* random NFS retry error, unless during file system support
843 * introspection, in which it actually means what it says */
855 /******************************************************************************
856 ****************** Begin Unique File ID Utility Used By VxWorks ***************
858 ** On most versions of unix, we can get a unique ID for a file by concatenating
859 ** the device number and the inode number. But this does not work on VxWorks.
860 ** On VxWorks, a unique file id must be based on the canonical filename.
862 ** A pointer to an instance of the following structure can be used as a
863 ** unique file ID in VxWorks. Each instance of this structure contains
864 ** a copy of the canonical filename. There is also a reference count.
865 ** The structure is reclaimed when the number of pointers to it drops to
868 ** There are never very many files open at one time and lookups are not
869 ** a performance-critical path, so it is sufficient to put these
870 ** structures on a linked list.
872 struct vxworksFileId
{
873 struct vxworksFileId
*pNext
; /* Next in a list of them all */
874 int nRef
; /* Number of references to this one */
875 int nName
; /* Length of the zCanonicalName[] string */
876 char *zCanonicalName
; /* Canonical filename */
881 ** All unique filenames are held on a linked list headed by this
884 static struct vxworksFileId
*vxworksFileList
= 0;
887 ** Simplify a filename into its canonical form
888 ** by making the following changes:
890 ** * removing any trailing and duplicate /
891 ** * convert /./ into just /
892 ** * convert /A/../ where A is any simple name into just /
894 ** Changes are made in-place. Return the new name length.
896 ** The original filename is in z[0..n-1]. Return the number of
897 ** characters in the simplified name.
899 static int vxworksSimplifyName(char *z
, int n
){
901 while( n
>1 && z
[n
-1]=='/' ){ n
--; }
902 for(i
=j
=0; i
<n
; i
++){
904 if( z
[i
+1]=='/' ) continue;
905 if( z
[i
+1]=='.' && i
+2<n
&& z
[i
+2]=='/' ){
909 if( z
[i
+1]=='.' && i
+3<n
&& z
[i
+2]=='.' && z
[i
+3]=='/' ){
910 while( j
>0 && z
[j
-1]!='/' ){ j
--; }
923 ** Find a unique file ID for the given absolute pathname. Return
924 ** a pointer to the vxworksFileId object. This pointer is the unique
927 ** The nRef field of the vxworksFileId object is incremented before
928 ** the object is returned. A new vxworksFileId object is created
929 ** and added to the global list if necessary.
931 ** If a memory allocation error occurs, return NULL.
933 static struct vxworksFileId
*vxworksFindFileId(const char *zAbsoluteName
){
934 struct vxworksFileId
*pNew
; /* search key and new file ID */
935 struct vxworksFileId
*pCandidate
; /* For looping over existing file IDs */
936 int n
; /* Length of zAbsoluteName string */
938 assert( zAbsoluteName
[0]=='/' );
939 n
= (int)strlen(zAbsoluteName
);
940 pNew
= sqlite3_malloc64( sizeof(*pNew
) + (n
+1) );
941 if( pNew
==0 ) return 0;
942 pNew
->zCanonicalName
= (char*)&pNew
[1];
943 memcpy(pNew
->zCanonicalName
, zAbsoluteName
, n
+1);
944 n
= vxworksSimplifyName(pNew
->zCanonicalName
, n
);
946 /* Search for an existing entry that matching the canonical name.
947 ** If found, increment the reference count and return a pointer to
948 ** the existing file ID.
951 for(pCandidate
=vxworksFileList
; pCandidate
; pCandidate
=pCandidate
->pNext
){
952 if( pCandidate
->nName
==n
953 && memcmp(pCandidate
->zCanonicalName
, pNew
->zCanonicalName
, n
)==0
962 /* No match was found. We will make a new file ID */
965 pNew
->pNext
= vxworksFileList
;
966 vxworksFileList
= pNew
;
972 ** Decrement the reference count on a vxworksFileId object. Free
973 ** the object when the reference count reaches zero.
975 static void vxworksReleaseFileId(struct vxworksFileId
*pId
){
977 assert( pId
->nRef
>0 );
980 struct vxworksFileId
**pp
;
981 for(pp
=&vxworksFileList
; *pp
&& *pp
!=pId
; pp
= &((*pp
)->pNext
)){}
988 #endif /* OS_VXWORKS */
989 /*************** End of Unique File ID Utility Used By VxWorks ****************
990 ******************************************************************************/
993 /******************************************************************************
994 *************************** Posix Advisory Locking ****************************
996 ** POSIX advisory locks are broken by design. ANSI STD 1003.1 (1996)
997 ** section 6.5.2.2 lines 483 through 490 specify that when a process
998 ** sets or clears a lock, that operation overrides any prior locks set
999 ** by the same process. It does not explicitly say so, but this implies
1000 ** that it overrides locks set by the same process using a different
1001 ** file descriptor. Consider this test case:
1003 ** int fd1 = open("./file1", O_RDWR|O_CREAT, 0644);
1004 ** int fd2 = open("./file2", O_RDWR|O_CREAT, 0644);
1006 ** Suppose ./file1 and ./file2 are really the same file (because
1007 ** one is a hard or symbolic link to the other) then if you set
1008 ** an exclusive lock on fd1, then try to get an exclusive lock
1009 ** on fd2, it works. I would have expected the second lock to
1010 ** fail since there was already a lock on the file due to fd1.
1011 ** But not so. Since both locks came from the same process, the
1012 ** second overrides the first, even though they were on different
1013 ** file descriptors opened on different file names.
1015 ** This means that we cannot use POSIX locks to synchronize file access
1016 ** among competing threads of the same process. POSIX locks will work fine
1017 ** to synchronize access for threads in separate processes, but not
1018 ** threads within the same process.
1020 ** To work around the problem, SQLite has to manage file locks internally
1021 ** on its own. Whenever a new database is opened, we have to find the
1022 ** specific inode of the database file (the inode is determined by the
1023 ** st_dev and st_ino fields of the stat structure that fstat() fills in)
1024 ** and check for locks already existing on that inode. When locks are
1025 ** created or removed, we have to look at our own internal record of the
1026 ** locks to see if another thread has previously set a lock on that same
1029 ** (Aside: The use of inode numbers as unique IDs does not work on VxWorks.
1030 ** For VxWorks, we have to use the alternative unique ID system based on
1031 ** canonical filename and implemented in the previous division.)
1033 ** The sqlite3_file structure for POSIX is no longer just an integer file
1034 ** descriptor. It is now a structure that holds the integer file
1035 ** descriptor and a pointer to a structure that describes the internal
1036 ** locks on the corresponding inode. There is one locking structure
1037 ** per inode, so if the same inode is opened twice, both unixFile structures
1038 ** point to the same locking structure. The locking structure keeps
1039 ** a reference count (so we will know when to delete it) and a "cnt"
1040 ** field that tells us its internal lock status. cnt==0 means the
1041 ** file is unlocked. cnt==-1 means the file has an exclusive lock.
1042 ** cnt>0 means there are cnt shared locks on the file.
1044 ** Any attempt to lock or unlock a file first checks the locking
1045 ** structure. The fcntl() system call is only invoked to set a
1046 ** POSIX lock if the internal lock structure transitions between
1047 ** a locked and an unlocked state.
1049 ** But wait: there are yet more problems with POSIX advisory locks.
1051 ** If you close a file descriptor that points to a file that has locks,
1052 ** all locks on that file that are owned by the current process are
1053 ** released. To work around this problem, each unixInodeInfo object
1054 ** maintains a count of the number of pending locks on tha inode.
1055 ** When an attempt is made to close an unixFile, if there are
1056 ** other unixFile open on the same inode that are holding locks, the call
1057 ** to close() the file descriptor is deferred until all of the locks clear.
1058 ** The unixInodeInfo structure keeps a list of file descriptors that need to
1059 ** be closed and that list is walked (and cleared) when the last lock
1062 ** Yet another problem: LinuxThreads do not play well with posix locks.
1064 ** Many older versions of linux use the LinuxThreads library which is
1065 ** not posix compliant. Under LinuxThreads, a lock created by thread
1066 ** A cannot be modified or overridden by a different thread B.
1067 ** Only thread A can modify the lock. Locking behavior is correct
1068 ** if the appliation uses the newer Native Posix Thread Library (NPTL)
1069 ** on linux - with NPTL a lock created by thread A can override locks
1070 ** in thread B. But there is no way to know at compile-time which
1071 ** threading library is being used. So there is no way to know at
1072 ** compile-time whether or not thread A can override locks on thread B.
1073 ** One has to do a run-time check to discover the behavior of the
1076 ** SQLite used to support LinuxThreads. But support for LinuxThreads
1077 ** was dropped beginning with version 3.7.0. SQLite will still work with
1078 ** LinuxThreads provided that (1) there is no more than one connection
1079 ** per database file in the same process and (2) database connections
1080 ** do not move across threads.
1084 ** An instance of the following structure serves as the key used
1085 ** to locate a particular unixInodeInfo object.
1088 dev_t dev
; /* Device number */
1090 struct vxworksFileId
*pId
; /* Unique file ID for vxworks. */
1092 /* We are told that some versions of Android contain a bug that
1093 ** sizes ino_t at only 32-bits instead of 64-bits. (See
1094 ** https://android-review.googlesource.com/#/c/115351/3/dist/sqlite3.c)
1095 ** To work around this, always allocate 64-bits for the inode number.
1096 ** On small machines that only have 32-bit inodes, this wastes 4 bytes,
1097 ** but that should not be a big deal. */
1098 /* WAS: ino_t ino; */
1099 u64 ino
; /* Inode number */
1104 ** An instance of the following structure is allocated for each open
1105 ** inode. Or, on LinuxThreads, there is one of these structures for
1106 ** each inode opened by each thread.
1108 ** A single inode can have multiple file descriptors, so each unixFile
1109 ** structure contains a pointer to an instance of this object and this
1110 ** object keeps a count of the number of unixFile pointing to it.
1112 struct unixInodeInfo
{
1113 struct unixFileId fileId
; /* The lookup key */
1114 int nShared
; /* Number of SHARED locks held */
1115 unsigned char eFileLock
; /* One of SHARED_LOCK, RESERVED_LOCK etc. */
1116 unsigned char bProcessLock
; /* An exclusive process lock is held */
1117 int nRef
; /* Number of pointers to this structure */
1118 unixShmNode
*pShmNode
; /* Shared memory associated with this inode */
1119 int nLock
; /* Number of outstanding file locks */
1120 UnixUnusedFd
*pUnused
; /* Unused file descriptors to close */
1121 unixInodeInfo
*pNext
; /* List of all unixInodeInfo objects */
1122 unixInodeInfo
*pPrev
; /* .... doubly linked */
1123 #if SQLITE_ENABLE_LOCKING_STYLE
1124 unsigned long long sharedByte
; /* for AFP simulated shared lock */
1127 sem_t
*pSem
; /* Named POSIX semaphore */
1128 char aSemName
[MAX_PATHNAME
+2]; /* Name of that semaphore */
1133 ** A lists of all unixInodeInfo objects.
1135 static unixInodeInfo
*inodeList
= 0; /* All unixInodeInfo objects */
1136 static unsigned int nUnusedFd
= 0; /* Total unused file descriptors */
1140 ** This function - unixLogErrorAtLine(), is only ever called via the macro
1143 ** It is invoked after an error occurs in an OS function and errno has been
1144 ** set. It logs a message using sqlite3_log() containing the current value of
1145 ** errno and, if possible, the human-readable equivalent from strerror() or
1148 ** The first argument passed to the macro should be the error code that
1149 ** will be returned to SQLite (e.g. SQLITE_IOERR_DELETE, SQLITE_CANTOPEN).
1150 ** The two subsequent arguments should be the name of the OS function that
1151 ** failed (e.g. "unlink", "open") and the associated file-system path,
1154 #define unixLogError(a,b,c) unixLogErrorAtLine(a,b,c,__LINE__)
1155 static int unixLogErrorAtLine(
1156 int errcode
, /* SQLite error code */
1157 const char *zFunc
, /* Name of OS function that failed */
1158 const char *zPath
, /* File path associated with error */
1159 int iLine
/* Source line number where error occurred */
1161 char *zErr
; /* Message from strerror() or equivalent */
1162 int iErrno
= errno
; /* Saved syscall error number */
1164 /* If this is not a threadsafe build (SQLITE_THREADSAFE==0), then use
1165 ** the strerror() function to obtain the human-readable error message
1166 ** equivalent to errno. Otherwise, use strerror_r().
1168 #if SQLITE_THREADSAFE && defined(HAVE_STRERROR_R)
1170 memset(aErr
, 0, sizeof(aErr
));
1173 /* If STRERROR_R_CHAR_P (set by autoconf scripts) or __USE_GNU is defined,
1174 ** assume that the system provides the GNU version of strerror_r() that
1175 ** returns a pointer to a buffer containing the error message. That pointer
1176 ** may point to aErr[], or it may point to some static storage somewhere.
1177 ** Otherwise, assume that the system provides the POSIX version of
1178 ** strerror_r(), which always writes an error message into aErr[].
1180 ** If the code incorrectly assumes that it is the POSIX version that is
1181 ** available, the error message will often be an empty string. Not a
1182 ** huge problem. Incorrectly concluding that the GNU version is available
1183 ** could lead to a segfault though.
1185 #if defined(STRERROR_R_CHAR_P) || defined(__USE_GNU)
1188 strerror_r(iErrno
, aErr
, sizeof(aErr
)-1);
1190 #elif SQLITE_THREADSAFE
1191 /* This is a threadsafe build, but strerror_r() is not available. */
1194 /* Non-threadsafe build, use strerror(). */
1195 zErr
= strerror(iErrno
);
1198 if( zPath
==0 ) zPath
= "";
1199 sqlite3_log(errcode
,
1200 "os_unix.c:%d: (%d) %s(%s) - %s",
1201 iLine
, iErrno
, zFunc
, zPath
, zErr
1208 ** Close a file descriptor.
1210 ** We assume that close() almost always works, since it is only in a
1211 ** very sick application or on a very sick platform that it might fail.
1212 ** If it does fail, simply leak the file descriptor, but do log the
1215 ** Note that it is not safe to retry close() after EINTR since the
1216 ** file descriptor might have already been reused by another thread.
1217 ** So we don't even try to recover from an EINTR. Just log the error
1220 static void robust_close(unixFile
*pFile
, int h
, int lineno
){
1222 unixLogErrorAtLine(SQLITE_IOERR_CLOSE
, "close",
1223 pFile
? pFile
->zPath
: 0, lineno
);
1228 ** Set the pFile->lastErrno. Do this in a subroutine as that provides
1229 ** a convenient place to set a breakpoint.
1231 static void storeLastErrno(unixFile
*pFile
, int error
){
1232 pFile
->lastErrno
= error
;
1236 ** Close all file descriptors accumuated in the unixInodeInfo->pUnused list.
1238 static void closePendingFds(unixFile
*pFile
){
1239 unixInodeInfo
*pInode
= pFile
->pInode
;
1241 UnixUnusedFd
*pNext
;
1242 for(p
=pInode
->pUnused
; p
; p
=pNext
){
1244 robust_close(pFile
, p
->fd
, __LINE__
);
1248 pInode
->pUnused
= 0;
1252 ** Release a unixInodeInfo structure previously allocated by findInodeInfo().
1254 ** The mutex entered using the unixEnterMutex() function must be held
1255 ** when this function is called.
1257 static void releaseInodeInfo(unixFile
*pFile
){
1258 unixInodeInfo
*pInode
= pFile
->pInode
;
1259 assert( unixMutexHeld() );
1260 if( ALWAYS(pInode
) ){
1262 if( pInode
->nRef
==0 ){
1263 assert( pInode
->pShmNode
==0 );
1264 closePendingFds(pFile
);
1265 if( pInode
->pPrev
){
1266 assert( pInode
->pPrev
->pNext
==pInode
);
1267 pInode
->pPrev
->pNext
= pInode
->pNext
;
1269 assert( inodeList
==pInode
);
1270 inodeList
= pInode
->pNext
;
1272 if( pInode
->pNext
){
1273 assert( pInode
->pNext
->pPrev
==pInode
);
1274 pInode
->pNext
->pPrev
= pInode
->pPrev
;
1276 sqlite3_free(pInode
);
1279 assert( inodeList
!=0 || nUnusedFd
==0 );
1283 ** Given a file descriptor, locate the unixInodeInfo object that
1284 ** describes that file descriptor. Create a new one if necessary. The
1285 ** return value might be uninitialized if an error occurs.
1287 ** The mutex entered using the unixEnterMutex() function must be held
1288 ** when this function is called.
1290 ** Return an appropriate error code.
1292 static int findInodeInfo(
1293 unixFile
*pFile
, /* Unix file with file desc used in the key */
1294 unixInodeInfo
**ppInode
/* Return the unixInodeInfo object here */
1296 int rc
; /* System call return code */
1297 int fd
; /* The file descriptor for pFile */
1298 struct unixFileId fileId
; /* Lookup key for the unixInodeInfo */
1299 struct stat statbuf
; /* Low-level file information */
1300 unixInodeInfo
*pInode
= 0; /* Candidate unixInodeInfo object */
1302 assert( unixMutexHeld() );
1304 /* Get low-level information about the file that we can used to
1305 ** create a unique name for the file.
1308 rc
= osFstat(fd
, &statbuf
);
1310 storeLastErrno(pFile
, errno
);
1311 #if defined(EOVERFLOW) && defined(SQLITE_DISABLE_LFS)
1312 if( pFile
->lastErrno
==EOVERFLOW
) return SQLITE_NOLFS
;
1314 return SQLITE_IOERR
;
1318 /* On OS X on an msdos filesystem, the inode number is reported
1319 ** incorrectly for zero-size files. See ticket #3260. To work
1320 ** around this problem (we consider it a bug in OS X, not SQLite)
1321 ** we always increase the file size to 1 by writing a single byte
1322 ** prior to accessing the inode number. The one byte written is
1323 ** an ASCII 'S' character which also happens to be the first byte
1324 ** in the header of every SQLite database. In this way, if there
1325 ** is a race condition such that another thread has already populated
1326 ** the first page of the database, no damage is done.
1328 if( statbuf
.st_size
==0 && (pFile
->fsFlags
& SQLITE_FSFLAGS_IS_MSDOS
)!=0 ){
1329 do{ rc
= osWrite(fd
, "S", 1); }while( rc
<0 && errno
==EINTR
);
1331 storeLastErrno(pFile
, errno
);
1332 return SQLITE_IOERR
;
1334 rc
= osFstat(fd
, &statbuf
);
1336 storeLastErrno(pFile
, errno
);
1337 return SQLITE_IOERR
;
1342 memset(&fileId
, 0, sizeof(fileId
));
1343 fileId
.dev
= statbuf
.st_dev
;
1345 fileId
.pId
= pFile
->pId
;
1347 fileId
.ino
= (u64
)statbuf
.st_ino
;
1349 assert( inodeList
!=0 || nUnusedFd
==0 );
1351 while( pInode
&& memcmp(&fileId
, &pInode
->fileId
, sizeof(fileId
)) ){
1352 pInode
= pInode
->pNext
;
1355 pInode
= sqlite3_malloc64( sizeof(*pInode
) );
1357 return SQLITE_NOMEM_BKPT
;
1359 memset(pInode
, 0, sizeof(*pInode
));
1360 memcpy(&pInode
->fileId
, &fileId
, sizeof(fileId
));
1362 pInode
->pNext
= inodeList
;
1364 if( inodeList
) inodeList
->pPrev
= pInode
;
1374 ** Return TRUE if pFile has been renamed or unlinked since it was first opened.
1376 static int fileHasMoved(unixFile
*pFile
){
1378 return pFile
->pInode
!=0 && pFile
->pId
!=pFile
->pInode
->fileId
.pId
;
1381 return pFile
->pInode
!=0 &&
1382 (osStat(pFile
->zPath
, &buf
)!=0
1383 || (u64
)buf
.st_ino
!=pFile
->pInode
->fileId
.ino
);
1389 ** Check a unixFile that is a database. Verify the following:
1391 ** (1) There is exactly one hard link on the file
1392 ** (2) The file is not a symbolic link
1393 ** (3) The file has not been renamed or unlinked
1395 ** Issue sqlite3_log(SQLITE_WARNING,...) messages if anything is not right.
1397 static void verifyDbFile(unixFile
*pFile
){
1401 /* These verifications occurs for the main database only */
1402 if( pFile
->ctrlFlags
& UNIXFILE_NOLOCK
) return;
1404 rc
= osFstat(pFile
->h
, &buf
);
1406 sqlite3_log(SQLITE_WARNING
, "cannot fstat db file %s", pFile
->zPath
);
1409 if( buf
.st_nlink
==0 ){
1410 sqlite3_log(SQLITE_WARNING
, "file unlinked while open: %s", pFile
->zPath
);
1413 if( buf
.st_nlink
>1 ){
1414 sqlite3_log(SQLITE_WARNING
, "multiple links to file: %s", pFile
->zPath
);
1417 if( fileHasMoved(pFile
) ){
1418 sqlite3_log(SQLITE_WARNING
, "file renamed while open: %s", pFile
->zPath
);
1425 ** This routine checks if there is a RESERVED lock held on the specified
1426 ** file by this or any other process. If such a lock is held, set *pResOut
1427 ** to a non-zero value otherwise *pResOut is set to zero. The return value
1428 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
1430 static int unixCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
1433 unixFile
*pFile
= (unixFile
*)id
;
1435 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
1438 assert( pFile
->eFileLock
<=SHARED_LOCK
);
1439 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
1441 /* Check if a thread in this process holds such a lock */
1442 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
1446 /* Otherwise see if some other process holds it.
1449 if( !reserved
&& !pFile
->pInode
->bProcessLock
){
1451 lock
.l_whence
= SEEK_SET
;
1452 lock
.l_start
= RESERVED_BYTE
;
1454 lock
.l_type
= F_WRLCK
;
1455 if( osFcntl(pFile
->h
, F_GETLK
, &lock
) ){
1456 rc
= SQLITE_IOERR_CHECKRESERVEDLOCK
;
1457 storeLastErrno(pFile
, errno
);
1458 } else if( lock
.l_type
!=F_UNLCK
){
1465 OSTRACE(("TEST WR-LOCK %d %d %d (unix)\n", pFile
->h
, rc
, reserved
));
1467 *pResOut
= reserved
;
1472 ** Set a posix-advisory-lock.
1474 ** There are two versions of this routine. If compiled with
1475 ** SQLITE_ENABLE_SETLK_TIMEOUT then the routine has an extra parameter
1476 ** which is a pointer to a unixFile. If the unixFile->iBusyTimeout
1477 ** value is set, then it is the number of milliseconds to wait before
1478 ** failing the lock. The iBusyTimeout value is always reset back to
1479 ** zero on each call.
1481 ** If SQLITE_ENABLE_SETLK_TIMEOUT is not defined, then do a non-blocking
1482 ** attempt to set the lock.
1484 #ifndef SQLITE_ENABLE_SETLK_TIMEOUT
1485 # define osSetPosixAdvisoryLock(h,x,t) osFcntl(h,F_SETLK,x)
1487 static int osSetPosixAdvisoryLock(
1488 int h
, /* The file descriptor on which to take the lock */
1489 struct flock
*pLock
, /* The description of the lock */
1490 unixFile
*pFile
/* Structure holding timeout value */
1492 int rc
= osFcntl(h
,F_SETLK
,pLock
);
1493 while( rc
<0 && pFile
->iBusyTimeout
>0 ){
1494 /* On systems that support some kind of blocking file lock with a timeout,
1495 ** make appropriate changes here to invoke that blocking file lock. On
1496 ** generic posix, however, there is no such API. So we simply try the
1497 ** lock once every millisecond until either the timeout expires, or until
1498 ** the lock is obtained. */
1500 rc
= osFcntl(h
,F_SETLK
,pLock
);
1501 pFile
->iBusyTimeout
--;
1505 #endif /* SQLITE_ENABLE_SETLK_TIMEOUT */
1509 ** Attempt to set a system-lock on the file pFile. The lock is
1510 ** described by pLock.
1512 ** If the pFile was opened read/write from unix-excl, then the only lock
1513 ** ever obtained is an exclusive lock, and it is obtained exactly once
1514 ** the first time any lock is attempted. All subsequent system locking
1515 ** operations become no-ops. Locking operations still happen internally,
1516 ** in order to coordinate access between separate database connections
1517 ** within this process, but all of that is handled in memory and the
1518 ** operating system does not participate.
1520 ** This function is a pass-through to fcntl(F_SETLK) if pFile is using
1521 ** any VFS other than "unix-excl" or if pFile is opened on "unix-excl"
1522 ** and is read-only.
1524 ** Zero is returned if the call completes successfully, or -1 if a call
1525 ** to fcntl() fails. In this case, errno is set appropriately (by fcntl()).
1527 static int unixFileLock(unixFile
*pFile
, struct flock
*pLock
){
1529 unixInodeInfo
*pInode
= pFile
->pInode
;
1530 assert( unixMutexHeld() );
1531 assert( pInode
!=0 );
1532 if( (pFile
->ctrlFlags
& (UNIXFILE_EXCL
|UNIXFILE_RDONLY
))==UNIXFILE_EXCL
){
1533 if( pInode
->bProcessLock
==0 ){
1535 assert( pInode
->nLock
==0 );
1536 lock
.l_whence
= SEEK_SET
;
1537 lock
.l_start
= SHARED_FIRST
;
1538 lock
.l_len
= SHARED_SIZE
;
1539 lock
.l_type
= F_WRLCK
;
1540 rc
= osSetPosixAdvisoryLock(pFile
->h
, &lock
, pFile
);
1541 if( rc
<0 ) return rc
;
1542 pInode
->bProcessLock
= 1;
1548 rc
= osSetPosixAdvisoryLock(pFile
->h
, pLock
, pFile
);
1554 ** Lock the file with the lock specified by parameter eFileLock - one
1555 ** of the following:
1558 ** (2) RESERVED_LOCK
1560 ** (4) EXCLUSIVE_LOCK
1562 ** Sometimes when requesting one lock state, additional lock states
1563 ** are inserted in between. The locking might fail on one of the later
1564 ** transitions leaving the lock state different from what it started but
1565 ** still short of its goal. The following chart shows the allowed
1566 ** transitions and the inserted intermediate states:
1568 ** UNLOCKED -> SHARED
1569 ** SHARED -> RESERVED
1570 ** SHARED -> (PENDING) -> EXCLUSIVE
1571 ** RESERVED -> (PENDING) -> EXCLUSIVE
1572 ** PENDING -> EXCLUSIVE
1574 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
1575 ** routine to lower a locking level.
1577 static int unixLock(sqlite3_file
*id
, int eFileLock
){
1578 /* The following describes the implementation of the various locks and
1579 ** lock transitions in terms of the POSIX advisory shared and exclusive
1580 ** lock primitives (called read-locks and write-locks below, to avoid
1581 ** confusion with SQLite lock names). The algorithms are complicated
1582 ** slightly in order to be compatible with Windows95 systems simultaneously
1583 ** accessing the same database file, in case that is ever required.
1585 ** Symbols defined in os.h indentify the 'pending byte' and the 'reserved
1586 ** byte', each single bytes at well known offsets, and the 'shared byte
1587 ** range', a range of 510 bytes at a well known offset.
1589 ** To obtain a SHARED lock, a read-lock is obtained on the 'pending
1590 ** byte'. If this is successful, 'shared byte range' is read-locked
1591 ** and the lock on the 'pending byte' released. (Legacy note: When
1592 ** SQLite was first developed, Windows95 systems were still very common,
1593 ** and Widnows95 lacks a shared-lock capability. So on Windows95, a
1594 ** single randomly selected by from the 'shared byte range' is locked.
1595 ** Windows95 is now pretty much extinct, but this work-around for the
1596 ** lack of shared-locks on Windows95 lives on, for backwards
1599 ** A process may only obtain a RESERVED lock after it has a SHARED lock.
1600 ** A RESERVED lock is implemented by grabbing a write-lock on the
1603 ** A process may only obtain a PENDING lock after it has obtained a
1604 ** SHARED lock. A PENDING lock is implemented by obtaining a write-lock
1605 ** on the 'pending byte'. This ensures that no new SHARED locks can be
1606 ** obtained, but existing SHARED locks are allowed to persist. A process
1607 ** does not have to obtain a RESERVED lock on the way to a PENDING lock.
1608 ** This property is used by the algorithm for rolling back a journal file
1611 ** An EXCLUSIVE lock, obtained after a PENDING lock is held, is
1612 ** implemented by obtaining a write-lock on the entire 'shared byte
1613 ** range'. Since all other locks require a read-lock on one of the bytes
1614 ** within this range, this ensures that no other locks are held on the
1618 unixFile
*pFile
= (unixFile
*)id
;
1619 unixInodeInfo
*pInode
;
1624 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (unix)\n", pFile
->h
,
1625 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
1626 azFileLock(pFile
->pInode
->eFileLock
), pFile
->pInode
->nShared
,
1629 /* If there is already a lock of this type or more restrictive on the
1630 ** unixFile, do nothing. Don't use the end_lock: exit path, as
1631 ** unixEnterMutex() hasn't been called yet.
1633 if( pFile
->eFileLock
>=eFileLock
){
1634 OSTRACE(("LOCK %d %s ok (already held) (unix)\n", pFile
->h
,
1635 azFileLock(eFileLock
)));
1639 /* Make sure the locking sequence is correct.
1640 ** (1) We never move from unlocked to anything higher than shared lock.
1641 ** (2) SQLite never explicitly requests a pendig lock.
1642 ** (3) A shared lock is always held when a reserve lock is requested.
1644 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
1645 assert( eFileLock
!=PENDING_LOCK
);
1646 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
1648 /* This mutex is needed because pFile->pInode is shared across threads
1651 pInode
= pFile
->pInode
;
1653 /* If some thread using this PID has a lock via a different unixFile*
1654 ** handle that precludes the requested lock, return BUSY.
1656 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
1657 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
1663 /* If a SHARED lock is requested, and some thread using this PID already
1664 ** has a SHARED or RESERVED lock, then increment reference counts and
1665 ** return SQLITE_OK.
1667 if( eFileLock
==SHARED_LOCK
&&
1668 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
1669 assert( eFileLock
==SHARED_LOCK
);
1670 assert( pFile
->eFileLock
==0 );
1671 assert( pInode
->nShared
>0 );
1672 pFile
->eFileLock
= SHARED_LOCK
;
1679 /* A PENDING lock is needed before acquiring a SHARED lock and before
1680 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
1684 lock
.l_whence
= SEEK_SET
;
1685 if( eFileLock
==SHARED_LOCK
1686 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
1688 lock
.l_type
= (eFileLock
==SHARED_LOCK
?F_RDLCK
:F_WRLCK
);
1689 lock
.l_start
= PENDING_BYTE
;
1690 if( unixFileLock(pFile
, &lock
) ){
1692 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1693 if( rc
!=SQLITE_BUSY
){
1694 storeLastErrno(pFile
, tErrno
);
1701 /* If control gets to this point, then actually go ahead and make
1702 ** operating system calls for the specified lock.
1704 if( eFileLock
==SHARED_LOCK
){
1705 assert( pInode
->nShared
==0 );
1706 assert( pInode
->eFileLock
==0 );
1707 assert( rc
==SQLITE_OK
);
1709 /* Now get the read-lock */
1710 lock
.l_start
= SHARED_FIRST
;
1711 lock
.l_len
= SHARED_SIZE
;
1712 if( unixFileLock(pFile
, &lock
) ){
1714 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1717 /* Drop the temporary PENDING lock */
1718 lock
.l_start
= PENDING_BYTE
;
1720 lock
.l_type
= F_UNLCK
;
1721 if( unixFileLock(pFile
, &lock
) && rc
==SQLITE_OK
){
1722 /* This could happen with a network mount */
1724 rc
= SQLITE_IOERR_UNLOCK
;
1728 if( rc
!=SQLITE_BUSY
){
1729 storeLastErrno(pFile
, tErrno
);
1733 pFile
->eFileLock
= SHARED_LOCK
;
1735 pInode
->nShared
= 1;
1737 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
1738 /* We are trying for an exclusive lock but another thread in this
1739 ** same process is still holding a shared lock. */
1742 /* The request was for a RESERVED or EXCLUSIVE lock. It is
1743 ** assumed that there is a SHARED or greater lock on the file
1746 assert( 0!=pFile
->eFileLock
);
1747 lock
.l_type
= F_WRLCK
;
1749 assert( eFileLock
==RESERVED_LOCK
|| eFileLock
==EXCLUSIVE_LOCK
);
1750 if( eFileLock
==RESERVED_LOCK
){
1751 lock
.l_start
= RESERVED_BYTE
;
1754 lock
.l_start
= SHARED_FIRST
;
1755 lock
.l_len
= SHARED_SIZE
;
1758 if( unixFileLock(pFile
, &lock
) ){
1760 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1761 if( rc
!=SQLITE_BUSY
){
1762 storeLastErrno(pFile
, tErrno
);
1769 /* Set up the transaction-counter change checking flags when
1770 ** transitioning from a SHARED to a RESERVED lock. The change
1771 ** from SHARED to RESERVED marks the beginning of a normal
1772 ** write operation (not a hot journal rollback).
1775 && pFile
->eFileLock
<=SHARED_LOCK
1776 && eFileLock
==RESERVED_LOCK
1778 pFile
->transCntrChng
= 0;
1779 pFile
->dbUpdate
= 0;
1780 pFile
->inNormalWrite
= 1;
1785 if( rc
==SQLITE_OK
){
1786 pFile
->eFileLock
= eFileLock
;
1787 pInode
->eFileLock
= eFileLock
;
1788 }else if( eFileLock
==EXCLUSIVE_LOCK
){
1789 pFile
->eFileLock
= PENDING_LOCK
;
1790 pInode
->eFileLock
= PENDING_LOCK
;
1795 OSTRACE(("LOCK %d %s %s (unix)\n", pFile
->h
, azFileLock(eFileLock
),
1796 rc
==SQLITE_OK
? "ok" : "failed"));
1801 ** Add the file descriptor used by file handle pFile to the corresponding
1804 static void setPendingFd(unixFile
*pFile
){
1805 unixInodeInfo
*pInode
= pFile
->pInode
;
1806 UnixUnusedFd
*p
= pFile
->pPreallocatedUnused
;
1807 p
->pNext
= pInode
->pUnused
;
1808 pInode
->pUnused
= p
;
1810 pFile
->pPreallocatedUnused
= 0;
1815 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1816 ** must be either NO_LOCK or SHARED_LOCK.
1818 ** If the locking level of the file descriptor is already at or below
1819 ** the requested locking level, this routine is a no-op.
1821 ** If handleNFSUnlock is true, then on downgrading an EXCLUSIVE_LOCK to SHARED
1822 ** the byte range is divided into 2 parts and the first part is unlocked then
1823 ** set to a read lock, then the other part is simply unlocked. This works
1824 ** around a bug in BSD NFS lockd (also seen on MacOSX 10.3+) that fails to
1825 ** remove the write lock on a region when a read lock is set.
1827 static int posixUnlock(sqlite3_file
*id
, int eFileLock
, int handleNFSUnlock
){
1828 unixFile
*pFile
= (unixFile
*)id
;
1829 unixInodeInfo
*pInode
;
1834 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (unix)\n", pFile
->h
, eFileLock
,
1835 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
1838 assert( eFileLock
<=SHARED_LOCK
);
1839 if( pFile
->eFileLock
<=eFileLock
){
1843 pInode
= pFile
->pInode
;
1844 assert( pInode
->nShared
!=0 );
1845 if( pFile
->eFileLock
>SHARED_LOCK
){
1846 assert( pInode
->eFileLock
==pFile
->eFileLock
);
1849 /* When reducing a lock such that other processes can start
1850 ** reading the database file again, make sure that the
1851 ** transaction counter was updated if any part of the database
1852 ** file changed. If the transaction counter is not updated,
1853 ** other connections to the same file might not realize that
1854 ** the file has changed and hence might not know to flush their
1855 ** cache. The use of a stale cache can lead to database corruption.
1857 pFile
->inNormalWrite
= 0;
1860 /* downgrading to a shared lock on NFS involves clearing the write lock
1861 ** before establishing the readlock - to avoid a race condition we downgrade
1862 ** the lock in 2 blocks, so that part of the range will be covered by a
1863 ** write lock until the rest is covered by a read lock:
1869 if( eFileLock
==SHARED_LOCK
){
1870 #if !defined(__APPLE__) || !SQLITE_ENABLE_LOCKING_STYLE
1871 (void)handleNFSUnlock
;
1872 assert( handleNFSUnlock
==0 );
1874 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
1875 if( handleNFSUnlock
){
1876 int tErrno
; /* Error code from system call errors */
1877 off_t divSize
= SHARED_SIZE
- 1;
1879 lock
.l_type
= F_UNLCK
;
1880 lock
.l_whence
= SEEK_SET
;
1881 lock
.l_start
= SHARED_FIRST
;
1882 lock
.l_len
= divSize
;
1883 if( unixFileLock(pFile
, &lock
)==(-1) ){
1885 rc
= SQLITE_IOERR_UNLOCK
;
1886 storeLastErrno(pFile
, tErrno
);
1889 lock
.l_type
= F_RDLCK
;
1890 lock
.l_whence
= SEEK_SET
;
1891 lock
.l_start
= SHARED_FIRST
;
1892 lock
.l_len
= divSize
;
1893 if( unixFileLock(pFile
, &lock
)==(-1) ){
1895 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_RDLOCK
);
1896 if( IS_LOCK_ERROR(rc
) ){
1897 storeLastErrno(pFile
, tErrno
);
1901 lock
.l_type
= F_UNLCK
;
1902 lock
.l_whence
= SEEK_SET
;
1903 lock
.l_start
= SHARED_FIRST
+divSize
;
1904 lock
.l_len
= SHARED_SIZE
-divSize
;
1905 if( unixFileLock(pFile
, &lock
)==(-1) ){
1907 rc
= SQLITE_IOERR_UNLOCK
;
1908 storeLastErrno(pFile
, tErrno
);
1912 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
1914 lock
.l_type
= F_RDLCK
;
1915 lock
.l_whence
= SEEK_SET
;
1916 lock
.l_start
= SHARED_FIRST
;
1917 lock
.l_len
= SHARED_SIZE
;
1918 if( unixFileLock(pFile
, &lock
) ){
1919 /* In theory, the call to unixFileLock() cannot fail because another
1920 ** process is holding an incompatible lock. If it does, this
1921 ** indicates that the other process is not following the locking
1922 ** protocol. If this happens, return SQLITE_IOERR_RDLOCK. Returning
1923 ** SQLITE_BUSY would confuse the upper layer (in practice it causes
1924 ** an assert to fail). */
1925 rc
= SQLITE_IOERR_RDLOCK
;
1926 storeLastErrno(pFile
, errno
);
1931 lock
.l_type
= F_UNLCK
;
1932 lock
.l_whence
= SEEK_SET
;
1933 lock
.l_start
= PENDING_BYTE
;
1934 lock
.l_len
= 2L; assert( PENDING_BYTE
+1==RESERVED_BYTE
);
1935 if( unixFileLock(pFile
, &lock
)==0 ){
1936 pInode
->eFileLock
= SHARED_LOCK
;
1938 rc
= SQLITE_IOERR_UNLOCK
;
1939 storeLastErrno(pFile
, errno
);
1943 if( eFileLock
==NO_LOCK
){
1944 /* Decrement the shared lock counter. Release the lock using an
1945 ** OS call only when all threads in this same process have released
1949 if( pInode
->nShared
==0 ){
1950 lock
.l_type
= F_UNLCK
;
1951 lock
.l_whence
= SEEK_SET
;
1952 lock
.l_start
= lock
.l_len
= 0L;
1953 if( unixFileLock(pFile
, &lock
)==0 ){
1954 pInode
->eFileLock
= NO_LOCK
;
1956 rc
= SQLITE_IOERR_UNLOCK
;
1957 storeLastErrno(pFile
, errno
);
1958 pInode
->eFileLock
= NO_LOCK
;
1959 pFile
->eFileLock
= NO_LOCK
;
1963 /* Decrement the count of locks against this same file. When the
1964 ** count reaches zero, close any other file descriptors whose close
1965 ** was deferred because of outstanding locks.
1968 assert( pInode
->nLock
>=0 );
1969 if( pInode
->nLock
==0 ){
1970 closePendingFds(pFile
);
1976 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
1981 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1982 ** must be either NO_LOCK or SHARED_LOCK.
1984 ** If the locking level of the file descriptor is already at or below
1985 ** the requested locking level, this routine is a no-op.
1987 static int unixUnlock(sqlite3_file
*id
, int eFileLock
){
1988 #if SQLITE_MAX_MMAP_SIZE>0
1989 assert( eFileLock
==SHARED_LOCK
|| ((unixFile
*)id
)->nFetchOut
==0 );
1991 return posixUnlock(id
, eFileLock
, 0);
1994 #if SQLITE_MAX_MMAP_SIZE>0
1995 static int unixMapfile(unixFile
*pFd
, i64 nByte
);
1996 static void unixUnmapfile(unixFile
*pFd
);
2000 ** This function performs the parts of the "close file" operation
2001 ** common to all locking schemes. It closes the directory and file
2002 ** handles, if they are valid, and sets all fields of the unixFile
2005 ** It is *not* necessary to hold the mutex when this routine is called,
2006 ** even on VxWorks. A mutex will be acquired on VxWorks by the
2007 ** vxworksReleaseFileId() routine.
2009 static int closeUnixFile(sqlite3_file
*id
){
2010 unixFile
*pFile
= (unixFile
*)id
;
2011 #if SQLITE_MAX_MMAP_SIZE>0
2012 unixUnmapfile(pFile
);
2015 robust_close(pFile
, pFile
->h
, __LINE__
);
2020 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
2021 osUnlink(pFile
->pId
->zCanonicalName
);
2023 vxworksReleaseFileId(pFile
->pId
);
2027 #ifdef SQLITE_UNLINK_AFTER_CLOSE
2028 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
2029 osUnlink(pFile
->zPath
);
2030 sqlite3_free(*(char**)&pFile
->zPath
);
2034 OSTRACE(("CLOSE %-3d\n", pFile
->h
));
2036 sqlite3_free(pFile
->pPreallocatedUnused
);
2037 memset(pFile
, 0, sizeof(unixFile
));
2044 static int unixClose(sqlite3_file
*id
){
2046 unixFile
*pFile
= (unixFile
*)id
;
2047 verifyDbFile(pFile
);
2048 unixUnlock(id
, NO_LOCK
);
2051 /* unixFile.pInode is always valid here. Otherwise, a different close
2052 ** routine (e.g. nolockClose()) would be called instead.
2054 assert( pFile
->pInode
->nLock
>0 || pFile
->pInode
->bProcessLock
==0 );
2055 if( ALWAYS(pFile
->pInode
) && pFile
->pInode
->nLock
){
2056 /* If there are outstanding locks, do not actually close the file just
2057 ** yet because that would clear those locks. Instead, add the file
2058 ** descriptor to pInode->pUnused list. It will be automatically closed
2059 ** when the last lock is cleared.
2061 setPendingFd(pFile
);
2063 releaseInodeInfo(pFile
);
2064 rc
= closeUnixFile(id
);
2069 /************** End of the posix advisory lock implementation *****************
2070 ******************************************************************************/
2072 /******************************************************************************
2073 ****************************** No-op Locking **********************************
2075 ** Of the various locking implementations available, this is by far the
2076 ** simplest: locking is ignored. No attempt is made to lock the database
2077 ** file for reading or writing.
2079 ** This locking mode is appropriate for use on read-only databases
2080 ** (ex: databases that are burned into CD-ROM, for example.) It can
2081 ** also be used if the application employs some external mechanism to
2082 ** prevent simultaneous access of the same database by two or more
2083 ** database connections. But there is a serious risk of database
2084 ** corruption if this locking mode is used in situations where multiple
2085 ** database connections are accessing the same database file at the same
2086 ** time and one or more of those connections are writing.
2089 static int nolockCheckReservedLock(sqlite3_file
*NotUsed
, int *pResOut
){
2090 UNUSED_PARAMETER(NotUsed
);
2094 static int nolockLock(sqlite3_file
*NotUsed
, int NotUsed2
){
2095 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2098 static int nolockUnlock(sqlite3_file
*NotUsed
, int NotUsed2
){
2099 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2106 static int nolockClose(sqlite3_file
*id
) {
2107 return closeUnixFile(id
);
2110 /******************* End of the no-op lock implementation *********************
2111 ******************************************************************************/
2113 /******************************************************************************
2114 ************************* Begin dot-file Locking ******************************
2116 ** The dotfile locking implementation uses the existence of separate lock
2117 ** files (really a directory) to control access to the database. This works
2118 ** on just about every filesystem imaginable. But there are serious downsides:
2120 ** (1) There is zero concurrency. A single reader blocks all other
2121 ** connections from reading or writing the database.
2123 ** (2) An application crash or power loss can leave stale lock files
2124 ** sitting around that need to be cleared manually.
2126 ** Nevertheless, a dotlock is an appropriate locking mode for use if no
2127 ** other locking strategy is available.
2129 ** Dotfile locking works by creating a subdirectory in the same directory as
2130 ** the database and with the same name but with a ".lock" extension added.
2131 ** The existence of a lock directory implies an EXCLUSIVE lock. All other
2132 ** lock types (SHARED, RESERVED, PENDING) are mapped into EXCLUSIVE.
2136 ** The file suffix added to the data base filename in order to create the
2139 #define DOTLOCK_SUFFIX ".lock"
2142 ** This routine checks if there is a RESERVED lock held on the specified
2143 ** file by this or any other process. If such a lock is held, set *pResOut
2144 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2145 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2147 ** In dotfile locking, either a lock exists or it does not. So in this
2148 ** variation of CheckReservedLock(), *pResOut is set to true if any lock
2149 ** is held on the file and false if the file is unlocked.
2151 static int dotlockCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2154 unixFile
*pFile
= (unixFile
*)id
;
2156 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2159 reserved
= osAccess((const char*)pFile
->lockingContext
, 0)==0;
2160 OSTRACE(("TEST WR-LOCK %d %d %d (dotlock)\n", pFile
->h
, rc
, reserved
));
2161 *pResOut
= reserved
;
2166 ** Lock the file with the lock specified by parameter eFileLock - one
2167 ** of the following:
2170 ** (2) RESERVED_LOCK
2172 ** (4) EXCLUSIVE_LOCK
2174 ** Sometimes when requesting one lock state, additional lock states
2175 ** are inserted in between. The locking might fail on one of the later
2176 ** transitions leaving the lock state different from what it started but
2177 ** still short of its goal. The following chart shows the allowed
2178 ** transitions and the inserted intermediate states:
2180 ** UNLOCKED -> SHARED
2181 ** SHARED -> RESERVED
2182 ** SHARED -> (PENDING) -> EXCLUSIVE
2183 ** RESERVED -> (PENDING) -> EXCLUSIVE
2184 ** PENDING -> EXCLUSIVE
2186 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2187 ** routine to lower a locking level.
2189 ** With dotfile locking, we really only support state (4): EXCLUSIVE.
2190 ** But we track the other locking levels internally.
2192 static int dotlockLock(sqlite3_file
*id
, int eFileLock
) {
2193 unixFile
*pFile
= (unixFile
*)id
;
2194 char *zLockFile
= (char *)pFile
->lockingContext
;
2198 /* If we have any lock, then the lock file already exists. All we have
2199 ** to do is adjust our internal record of the lock level.
2201 if( pFile
->eFileLock
> NO_LOCK
){
2202 pFile
->eFileLock
= eFileLock
;
2203 /* Always update the timestamp on the old file */
2205 utime(zLockFile
, NULL
);
2207 utimes(zLockFile
, NULL
);
2212 /* grab an exclusive lock */
2213 rc
= osMkdir(zLockFile
, 0777);
2215 /* failed to open/create the lock directory */
2217 if( EEXIST
== tErrno
){
2220 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2221 if( rc
!=SQLITE_BUSY
){
2222 storeLastErrno(pFile
, tErrno
);
2228 /* got it, set the type and return ok */
2229 pFile
->eFileLock
= eFileLock
;
2234 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2235 ** must be either NO_LOCK or SHARED_LOCK.
2237 ** If the locking level of the file descriptor is already at or below
2238 ** the requested locking level, this routine is a no-op.
2240 ** When the locking level reaches NO_LOCK, delete the lock file.
2242 static int dotlockUnlock(sqlite3_file
*id
, int eFileLock
) {
2243 unixFile
*pFile
= (unixFile
*)id
;
2244 char *zLockFile
= (char *)pFile
->lockingContext
;
2248 OSTRACE(("UNLOCK %d %d was %d pid=%d (dotlock)\n", pFile
->h
, eFileLock
,
2249 pFile
->eFileLock
, osGetpid(0)));
2250 assert( eFileLock
<=SHARED_LOCK
);
2252 /* no-op if possible */
2253 if( pFile
->eFileLock
==eFileLock
){
2257 /* To downgrade to shared, simply update our internal notion of the
2258 ** lock state. No need to mess with the file on disk.
2260 if( eFileLock
==SHARED_LOCK
){
2261 pFile
->eFileLock
= SHARED_LOCK
;
2265 /* To fully unlock the database, delete the lock file */
2266 assert( eFileLock
==NO_LOCK
);
2267 rc
= osRmdir(zLockFile
);
2270 if( tErrno
==ENOENT
){
2273 rc
= SQLITE_IOERR_UNLOCK
;
2274 storeLastErrno(pFile
, tErrno
);
2278 pFile
->eFileLock
= NO_LOCK
;
2283 ** Close a file. Make sure the lock has been released before closing.
2285 static int dotlockClose(sqlite3_file
*id
) {
2286 unixFile
*pFile
= (unixFile
*)id
;
2288 dotlockUnlock(id
, NO_LOCK
);
2289 sqlite3_free(pFile
->lockingContext
);
2290 return closeUnixFile(id
);
2292 /****************** End of the dot-file lock implementation *******************
2293 ******************************************************************************/
2295 /******************************************************************************
2296 ************************** Begin flock Locking ********************************
2298 ** Use the flock() system call to do file locking.
2300 ** flock() locking is like dot-file locking in that the various
2301 ** fine-grain locking levels supported by SQLite are collapsed into
2302 ** a single exclusive lock. In other words, SHARED, RESERVED, and
2303 ** PENDING locks are the same thing as an EXCLUSIVE lock. SQLite
2304 ** still works when you do this, but concurrency is reduced since
2305 ** only a single process can be reading the database at a time.
2307 ** Omit this section if SQLITE_ENABLE_LOCKING_STYLE is turned off
2309 #if SQLITE_ENABLE_LOCKING_STYLE
2312 ** Retry flock() calls that fail with EINTR
2315 static int robust_flock(int fd
, int op
){
2317 do{ rc
= flock(fd
,op
); }while( rc
<0 && errno
==EINTR
);
2321 # define robust_flock(a,b) flock(a,b)
2326 ** This routine checks if there is a RESERVED lock held on the specified
2327 ** file by this or any other process. If such a lock is held, set *pResOut
2328 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2329 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2331 static int flockCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2334 unixFile
*pFile
= (unixFile
*)id
;
2336 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2340 /* Check if a thread in this process holds such a lock */
2341 if( pFile
->eFileLock
>SHARED_LOCK
){
2345 /* Otherwise see if some other process holds it. */
2347 /* attempt to get the lock */
2348 int lrc
= robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
);
2350 /* got the lock, unlock it */
2351 lrc
= robust_flock(pFile
->h
, LOCK_UN
);
2354 /* unlock failed with an error */
2355 lrc
= SQLITE_IOERR_UNLOCK
;
2356 storeLastErrno(pFile
, tErrno
);
2362 /* someone else might have it reserved */
2363 lrc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2364 if( IS_LOCK_ERROR(lrc
) ){
2365 storeLastErrno(pFile
, tErrno
);
2370 OSTRACE(("TEST WR-LOCK %d %d %d (flock)\n", pFile
->h
, rc
, reserved
));
2372 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2373 if( (rc
& 0xff) == SQLITE_IOERR
){
2377 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2378 *pResOut
= reserved
;
2383 ** Lock the file with the lock specified by parameter eFileLock - one
2384 ** of the following:
2387 ** (2) RESERVED_LOCK
2389 ** (4) EXCLUSIVE_LOCK
2391 ** Sometimes when requesting one lock state, additional lock states
2392 ** are inserted in between. The locking might fail on one of the later
2393 ** transitions leaving the lock state different from what it started but
2394 ** still short of its goal. The following chart shows the allowed
2395 ** transitions and the inserted intermediate states:
2397 ** UNLOCKED -> SHARED
2398 ** SHARED -> RESERVED
2399 ** SHARED -> (PENDING) -> EXCLUSIVE
2400 ** RESERVED -> (PENDING) -> EXCLUSIVE
2401 ** PENDING -> EXCLUSIVE
2403 ** flock() only really support EXCLUSIVE locks. We track intermediate
2404 ** lock states in the sqlite3_file structure, but all locks SHARED or
2405 ** above are really EXCLUSIVE locks and exclude all other processes from
2408 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2409 ** routine to lower a locking level.
2411 static int flockLock(sqlite3_file
*id
, int eFileLock
) {
2413 unixFile
*pFile
= (unixFile
*)id
;
2417 /* if we already have a lock, it is exclusive.
2418 ** Just adjust level and punt on outta here. */
2419 if (pFile
->eFileLock
> NO_LOCK
) {
2420 pFile
->eFileLock
= eFileLock
;
2424 /* grab an exclusive lock */
2426 if (robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
)) {
2428 /* didn't get, must be busy */
2429 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2430 if( IS_LOCK_ERROR(rc
) ){
2431 storeLastErrno(pFile
, tErrno
);
2434 /* got it, set the type and return ok */
2435 pFile
->eFileLock
= eFileLock
;
2437 OSTRACE(("LOCK %d %s %s (flock)\n", pFile
->h
, azFileLock(eFileLock
),
2438 rc
==SQLITE_OK
? "ok" : "failed"));
2439 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2440 if( (rc
& 0xff) == SQLITE_IOERR
){
2443 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2449 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2450 ** must be either NO_LOCK or SHARED_LOCK.
2452 ** If the locking level of the file descriptor is already at or below
2453 ** the requested locking level, this routine is a no-op.
2455 static int flockUnlock(sqlite3_file
*id
, int eFileLock
) {
2456 unixFile
*pFile
= (unixFile
*)id
;
2459 OSTRACE(("UNLOCK %d %d was %d pid=%d (flock)\n", pFile
->h
, eFileLock
,
2460 pFile
->eFileLock
, osGetpid(0)));
2461 assert( eFileLock
<=SHARED_LOCK
);
2463 /* no-op if possible */
2464 if( pFile
->eFileLock
==eFileLock
){
2468 /* shared can just be set because we always have an exclusive */
2469 if (eFileLock
==SHARED_LOCK
) {
2470 pFile
->eFileLock
= eFileLock
;
2474 /* no, really, unlock. */
2475 if( robust_flock(pFile
->h
, LOCK_UN
) ){
2476 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2478 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2479 return SQLITE_IOERR_UNLOCK
;
2481 pFile
->eFileLock
= NO_LOCK
;
2489 static int flockClose(sqlite3_file
*id
) {
2491 flockUnlock(id
, NO_LOCK
);
2492 return closeUnixFile(id
);
2495 #endif /* SQLITE_ENABLE_LOCKING_STYLE && !OS_VXWORK */
2497 /******************* End of the flock lock implementation *********************
2498 ******************************************************************************/
2500 /******************************************************************************
2501 ************************ Begin Named Semaphore Locking ************************
2503 ** Named semaphore locking is only supported on VxWorks.
2505 ** Semaphore locking is like dot-lock and flock in that it really only
2506 ** supports EXCLUSIVE locking. Only a single process can read or write
2507 ** the database file at a time. This reduces potential concurrency, but
2508 ** makes the lock implementation much easier.
2513 ** This routine checks if there is a RESERVED lock held on the specified
2514 ** file by this or any other process. If such a lock is held, set *pResOut
2515 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2516 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2518 static int semXCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2521 unixFile
*pFile
= (unixFile
*)id
;
2523 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2527 /* Check if a thread in this process holds such a lock */
2528 if( pFile
->eFileLock
>SHARED_LOCK
){
2532 /* Otherwise see if some other process holds it. */
2534 sem_t
*pSem
= pFile
->pInode
->pSem
;
2536 if( sem_trywait(pSem
)==-1 ){
2538 if( EAGAIN
!= tErrno
){
2539 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_CHECKRESERVEDLOCK
);
2540 storeLastErrno(pFile
, tErrno
);
2542 /* someone else has the lock when we are in NO_LOCK */
2543 reserved
= (pFile
->eFileLock
< SHARED_LOCK
);
2546 /* we could have it if we want it */
2550 OSTRACE(("TEST WR-LOCK %d %d %d (sem)\n", pFile
->h
, rc
, reserved
));
2552 *pResOut
= reserved
;
2557 ** Lock the file with the lock specified by parameter eFileLock - one
2558 ** of the following:
2561 ** (2) RESERVED_LOCK
2563 ** (4) EXCLUSIVE_LOCK
2565 ** Sometimes when requesting one lock state, additional lock states
2566 ** are inserted in between. The locking might fail on one of the later
2567 ** transitions leaving the lock state different from what it started but
2568 ** still short of its goal. The following chart shows the allowed
2569 ** transitions and the inserted intermediate states:
2571 ** UNLOCKED -> SHARED
2572 ** SHARED -> RESERVED
2573 ** SHARED -> (PENDING) -> EXCLUSIVE
2574 ** RESERVED -> (PENDING) -> EXCLUSIVE
2575 ** PENDING -> EXCLUSIVE
2577 ** Semaphore locks only really support EXCLUSIVE locks. We track intermediate
2578 ** lock states in the sqlite3_file structure, but all locks SHARED or
2579 ** above are really EXCLUSIVE locks and exclude all other processes from
2582 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2583 ** routine to lower a locking level.
2585 static int semXLock(sqlite3_file
*id
, int eFileLock
) {
2586 unixFile
*pFile
= (unixFile
*)id
;
2587 sem_t
*pSem
= pFile
->pInode
->pSem
;
2590 /* if we already have a lock, it is exclusive.
2591 ** Just adjust level and punt on outta here. */
2592 if (pFile
->eFileLock
> NO_LOCK
) {
2593 pFile
->eFileLock
= eFileLock
;
2598 /* lock semaphore now but bail out when already locked. */
2599 if( sem_trywait(pSem
)==-1 ){
2604 /* got it, set the type and return ok */
2605 pFile
->eFileLock
= eFileLock
;
2612 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2613 ** must be either NO_LOCK or SHARED_LOCK.
2615 ** If the locking level of the file descriptor is already at or below
2616 ** the requested locking level, this routine is a no-op.
2618 static int semXUnlock(sqlite3_file
*id
, int eFileLock
) {
2619 unixFile
*pFile
= (unixFile
*)id
;
2620 sem_t
*pSem
= pFile
->pInode
->pSem
;
2624 OSTRACE(("UNLOCK %d %d was %d pid=%d (sem)\n", pFile
->h
, eFileLock
,
2625 pFile
->eFileLock
, osGetpid(0)));
2626 assert( eFileLock
<=SHARED_LOCK
);
2628 /* no-op if possible */
2629 if( pFile
->eFileLock
==eFileLock
){
2633 /* shared can just be set because we always have an exclusive */
2634 if (eFileLock
==SHARED_LOCK
) {
2635 pFile
->eFileLock
= eFileLock
;
2639 /* no, really unlock. */
2640 if ( sem_post(pSem
)==-1 ) {
2641 int rc
, tErrno
= errno
;
2642 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_UNLOCK
);
2643 if( IS_LOCK_ERROR(rc
) ){
2644 storeLastErrno(pFile
, tErrno
);
2648 pFile
->eFileLock
= NO_LOCK
;
2655 static int semXClose(sqlite3_file
*id
) {
2657 unixFile
*pFile
= (unixFile
*)id
;
2658 semXUnlock(id
, NO_LOCK
);
2661 releaseInodeInfo(pFile
);
2668 #endif /* OS_VXWORKS */
2670 ** Named semaphore locking is only available on VxWorks.
2672 *************** End of the named semaphore lock implementation ****************
2673 ******************************************************************************/
2676 /******************************************************************************
2677 *************************** Begin AFP Locking *********************************
2679 ** AFP is the Apple Filing Protocol. AFP is a network filesystem found
2680 ** on Apple Macintosh computers - both OS9 and OSX.
2682 ** Third-party implementations of AFP are available. But this code here
2683 ** only works on OSX.
2686 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
2688 ** The afpLockingContext structure contains all afp lock specific state
2690 typedef struct afpLockingContext afpLockingContext
;
2691 struct afpLockingContext
{
2693 const char *dbPath
; /* Name of the open file */
2696 struct ByteRangeLockPB2
2698 unsigned long long offset
; /* offset to first byte to lock */
2699 unsigned long long length
; /* nbr of bytes to lock */
2700 unsigned long long retRangeStart
; /* nbr of 1st byte locked if successful */
2701 unsigned char unLockFlag
; /* 1 = unlock, 0 = lock */
2702 unsigned char startEndFlag
; /* 1=rel to end of fork, 0=rel to start */
2703 int fd
; /* file desc to assoc this lock with */
2706 #define afpfsByteRangeLock2FSCTL _IOWR('z', 23, struct ByteRangeLockPB2)
2709 ** This is a utility for setting or clearing a bit-range lock on an
2712 ** Return SQLITE_OK on success, SQLITE_BUSY on failure.
2714 static int afpSetLock(
2715 const char *path
, /* Name of the file to be locked or unlocked */
2716 unixFile
*pFile
, /* Open file descriptor on path */
2717 unsigned long long offset
, /* First byte to be locked */
2718 unsigned long long length
, /* Number of bytes to lock */
2719 int setLockFlag
/* True to set lock. False to clear lock */
2721 struct ByteRangeLockPB2 pb
;
2724 pb
.unLockFlag
= setLockFlag
? 0 : 1;
2725 pb
.startEndFlag
= 0;
2730 OSTRACE(("AFPSETLOCK [%s] for %d%s in range %llx:%llx\n",
2731 (setLockFlag
?"ON":"OFF"), pFile
->h
, (pb
.fd
==-1?"[testval-1]":""),
2733 err
= fsctl(path
, afpfsByteRangeLock2FSCTL
, &pb
, 0);
2737 OSTRACE(("AFPSETLOCK failed to fsctl() '%s' %d %s\n",
2738 path
, tErrno
, strerror(tErrno
)));
2739 #ifdef SQLITE_IGNORE_AFP_LOCK_ERRORS
2742 rc
= sqliteErrorFromPosixError(tErrno
,
2743 setLockFlag
? SQLITE_IOERR_LOCK
: SQLITE_IOERR_UNLOCK
);
2744 #endif /* SQLITE_IGNORE_AFP_LOCK_ERRORS */
2745 if( IS_LOCK_ERROR(rc
) ){
2746 storeLastErrno(pFile
, tErrno
);
2755 ** This routine checks if there is a RESERVED lock held on the specified
2756 ** file by this or any other process. If such a lock is held, set *pResOut
2757 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2758 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2760 static int afpCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2763 unixFile
*pFile
= (unixFile
*)id
;
2764 afpLockingContext
*context
;
2766 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2769 context
= (afpLockingContext
*) pFile
->lockingContext
;
2770 if( context
->reserved
){
2774 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
2776 /* Check if a thread in this process holds such a lock */
2777 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
2781 /* Otherwise see if some other process holds it.
2784 /* lock the RESERVED byte */
2785 int lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2786 if( SQLITE_OK
==lrc
){
2787 /* if we succeeded in taking the reserved lock, unlock it to restore
2788 ** the original state */
2789 lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
2791 /* if we failed to get the lock then someone else must have it */
2794 if( IS_LOCK_ERROR(lrc
) ){
2800 OSTRACE(("TEST WR-LOCK %d %d %d (afp)\n", pFile
->h
, rc
, reserved
));
2802 *pResOut
= reserved
;
2807 ** Lock the file with the lock specified by parameter eFileLock - one
2808 ** of the following:
2811 ** (2) RESERVED_LOCK
2813 ** (4) EXCLUSIVE_LOCK
2815 ** Sometimes when requesting one lock state, additional lock states
2816 ** are inserted in between. The locking might fail on one of the later
2817 ** transitions leaving the lock state different from what it started but
2818 ** still short of its goal. The following chart shows the allowed
2819 ** transitions and the inserted intermediate states:
2821 ** UNLOCKED -> SHARED
2822 ** SHARED -> RESERVED
2823 ** SHARED -> (PENDING) -> EXCLUSIVE
2824 ** RESERVED -> (PENDING) -> EXCLUSIVE
2825 ** PENDING -> EXCLUSIVE
2827 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2828 ** routine to lower a locking level.
2830 static int afpLock(sqlite3_file
*id
, int eFileLock
){
2832 unixFile
*pFile
= (unixFile
*)id
;
2833 unixInodeInfo
*pInode
= pFile
->pInode
;
2834 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
2837 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (afp)\n", pFile
->h
,
2838 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
2839 azFileLock(pInode
->eFileLock
), pInode
->nShared
, osGetpid(0)));
2841 /* If there is already a lock of this type or more restrictive on the
2842 ** unixFile, do nothing. Don't use the afp_end_lock: exit path, as
2843 ** unixEnterMutex() hasn't been called yet.
2845 if( pFile
->eFileLock
>=eFileLock
){
2846 OSTRACE(("LOCK %d %s ok (already held) (afp)\n", pFile
->h
,
2847 azFileLock(eFileLock
)));
2851 /* Make sure the locking sequence is correct
2852 ** (1) We never move from unlocked to anything higher than shared lock.
2853 ** (2) SQLite never explicitly requests a pendig lock.
2854 ** (3) A shared lock is always held when a reserve lock is requested.
2856 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
2857 assert( eFileLock
!=PENDING_LOCK
);
2858 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
2860 /* This mutex is needed because pFile->pInode is shared across threads
2863 pInode
= pFile
->pInode
;
2865 /* If some thread using this PID has a lock via a different unixFile*
2866 ** handle that precludes the requested lock, return BUSY.
2868 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
2869 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
2875 /* If a SHARED lock is requested, and some thread using this PID already
2876 ** has a SHARED or RESERVED lock, then increment reference counts and
2877 ** return SQLITE_OK.
2879 if( eFileLock
==SHARED_LOCK
&&
2880 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
2881 assert( eFileLock
==SHARED_LOCK
);
2882 assert( pFile
->eFileLock
==0 );
2883 assert( pInode
->nShared
>0 );
2884 pFile
->eFileLock
= SHARED_LOCK
;
2890 /* A PENDING lock is needed before acquiring a SHARED lock and before
2891 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
2894 if( eFileLock
==SHARED_LOCK
2895 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
2898 failed
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 1);
2905 /* If control gets to this point, then actually go ahead and make
2906 ** operating system calls for the specified lock.
2908 if( eFileLock
==SHARED_LOCK
){
2909 int lrc1
, lrc2
, lrc1Errno
= 0;
2912 assert( pInode
->nShared
==0 );
2913 assert( pInode
->eFileLock
==0 );
2915 mask
= (sizeof(long)==8) ? LARGEST_INT64
: 0x7fffffff;
2916 /* Now get the read-lock SHARED_LOCK */
2917 /* note that the quality of the randomness doesn't matter that much */
2919 pInode
->sharedByte
= (lk
& mask
)%(SHARED_SIZE
- 1);
2920 lrc1
= afpSetLock(context
->dbPath
, pFile
,
2921 SHARED_FIRST
+pInode
->sharedByte
, 1, 1);
2922 if( IS_LOCK_ERROR(lrc1
) ){
2923 lrc1Errno
= pFile
->lastErrno
;
2925 /* Drop the temporary PENDING lock */
2926 lrc2
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
2928 if( IS_LOCK_ERROR(lrc1
) ) {
2929 storeLastErrno(pFile
, lrc1Errno
);
2932 } else if( IS_LOCK_ERROR(lrc2
) ){
2935 } else if( lrc1
!= SQLITE_OK
) {
2938 pFile
->eFileLock
= SHARED_LOCK
;
2940 pInode
->nShared
= 1;
2942 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
2943 /* We are trying for an exclusive lock but another thread in this
2944 ** same process is still holding a shared lock. */
2947 /* The request was for a RESERVED or EXCLUSIVE lock. It is
2948 ** assumed that there is a SHARED or greater lock on the file
2952 assert( 0!=pFile
->eFileLock
);
2953 if (eFileLock
>= RESERVED_LOCK
&& pFile
->eFileLock
< RESERVED_LOCK
) {
2954 /* Acquire a RESERVED lock */
2955 failed
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2957 context
->reserved
= 1;
2960 if (!failed
&& eFileLock
== EXCLUSIVE_LOCK
) {
2961 /* Acquire an EXCLUSIVE lock */
2963 /* Remove the shared lock before trying the range. we'll need to
2964 ** reestablish the shared lock if we can't get the afpUnlock
2966 if( !(failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
+
2967 pInode
->sharedByte
, 1, 0)) ){
2968 int failed2
= SQLITE_OK
;
2969 /* now attemmpt to get the exclusive lock range */
2970 failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
,
2972 if( failed
&& (failed2
= afpSetLock(context
->dbPath
, pFile
,
2973 SHARED_FIRST
+ pInode
->sharedByte
, 1, 1)) ){
2974 /* Can't reestablish the shared lock. Sqlite can't deal, this is
2975 ** a critical I/O error
2977 rc
= ((failed
& 0xff) == SQLITE_IOERR
) ? failed2
:
2990 if( rc
==SQLITE_OK
){
2991 pFile
->eFileLock
= eFileLock
;
2992 pInode
->eFileLock
= eFileLock
;
2993 }else if( eFileLock
==EXCLUSIVE_LOCK
){
2994 pFile
->eFileLock
= PENDING_LOCK
;
2995 pInode
->eFileLock
= PENDING_LOCK
;
3000 OSTRACE(("LOCK %d %s %s (afp)\n", pFile
->h
, azFileLock(eFileLock
),
3001 rc
==SQLITE_OK
? "ok" : "failed"));
3006 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
3007 ** must be either NO_LOCK or SHARED_LOCK.
3009 ** If the locking level of the file descriptor is already at or below
3010 ** the requested locking level, this routine is a no-op.
3012 static int afpUnlock(sqlite3_file
*id
, int eFileLock
) {
3014 unixFile
*pFile
= (unixFile
*)id
;
3015 unixInodeInfo
*pInode
;
3016 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
3023 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (afp)\n", pFile
->h
, eFileLock
,
3024 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
3027 assert( eFileLock
<=SHARED_LOCK
);
3028 if( pFile
->eFileLock
<=eFileLock
){
3032 pInode
= pFile
->pInode
;
3033 assert( pInode
->nShared
!=0 );
3034 if( pFile
->eFileLock
>SHARED_LOCK
){
3035 assert( pInode
->eFileLock
==pFile
->eFileLock
);
3036 SimulateIOErrorBenign(1);
3037 SimulateIOError( h
=(-1) )
3038 SimulateIOErrorBenign(0);
3041 /* When reducing a lock such that other processes can start
3042 ** reading the database file again, make sure that the
3043 ** transaction counter was updated if any part of the database
3044 ** file changed. If the transaction counter is not updated,
3045 ** other connections to the same file might not realize that
3046 ** the file has changed and hence might not know to flush their
3047 ** cache. The use of a stale cache can lead to database corruption.
3049 assert( pFile
->inNormalWrite
==0
3050 || pFile
->dbUpdate
==0
3051 || pFile
->transCntrChng
==1 );
3052 pFile
->inNormalWrite
= 0;
3055 if( pFile
->eFileLock
==EXCLUSIVE_LOCK
){
3056 rc
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
, SHARED_SIZE
, 0);
3057 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1) ){
3058 /* only re-establish the shared lock if necessary */
3059 int sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3060 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 1);
3065 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=PENDING_LOCK
){
3066 rc
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
3068 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=RESERVED_LOCK
&& context
->reserved
){
3069 rc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
3071 context
->reserved
= 0;
3074 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1)){
3075 pInode
->eFileLock
= SHARED_LOCK
;
3078 if( rc
==SQLITE_OK
&& eFileLock
==NO_LOCK
){
3080 /* Decrement the shared lock counter. Release the lock using an
3081 ** OS call only when all threads in this same process have released
3084 unsigned long long sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3086 if( pInode
->nShared
==0 ){
3087 SimulateIOErrorBenign(1);
3088 SimulateIOError( h
=(-1) )
3089 SimulateIOErrorBenign(0);
3091 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 0);
3094 pInode
->eFileLock
= NO_LOCK
;
3095 pFile
->eFileLock
= NO_LOCK
;
3098 if( rc
==SQLITE_OK
){
3100 assert( pInode
->nLock
>=0 );
3101 if( pInode
->nLock
==0 ){
3102 closePendingFds(pFile
);
3108 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
3113 ** Close a file & cleanup AFP specific locking context
3115 static int afpClose(sqlite3_file
*id
) {
3117 unixFile
*pFile
= (unixFile
*)id
;
3119 afpUnlock(id
, NO_LOCK
);
3121 if( pFile
->pInode
&& pFile
->pInode
->nLock
){
3122 /* If there are outstanding locks, do not actually close the file just
3123 ** yet because that would clear those locks. Instead, add the file
3124 ** descriptor to pInode->aPending. It will be automatically closed when
3125 ** the last lock is cleared.
3127 setPendingFd(pFile
);
3129 releaseInodeInfo(pFile
);
3130 sqlite3_free(pFile
->lockingContext
);
3131 rc
= closeUnixFile(id
);
3136 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3138 ** The code above is the AFP lock implementation. The code is specific
3139 ** to MacOSX and does not work on other unix platforms. No alternative
3140 ** is available. If you don't compile for a mac, then the "unix-afp"
3141 ** VFS is not available.
3143 ********************* End of the AFP lock implementation **********************
3144 ******************************************************************************/
3146 /******************************************************************************
3147 *************************** Begin NFS Locking ********************************/
3149 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
3151 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
3152 ** must be either NO_LOCK or SHARED_LOCK.
3154 ** If the locking level of the file descriptor is already at or below
3155 ** the requested locking level, this routine is a no-op.
3157 static int nfsUnlock(sqlite3_file
*id
, int eFileLock
){
3158 return posixUnlock(id
, eFileLock
, 1);
3161 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3163 ** The code above is the NFS lock implementation. The code is specific
3164 ** to MacOSX and does not work on other unix platforms. No alternative
3167 ********************* End of the NFS lock implementation **********************
3168 ******************************************************************************/
3170 /******************************************************************************
3171 **************** Non-locking sqlite3_file methods *****************************
3173 ** The next division contains implementations for all methods of the
3174 ** sqlite3_file object other than the locking methods. The locking
3175 ** methods were defined in divisions above (one locking method per
3176 ** division). Those methods that are common to all locking modes
3177 ** are gather together into this division.
3181 ** Seek to the offset passed as the second argument, then read cnt
3182 ** bytes into pBuf. Return the number of bytes actually read.
3184 ** NB: If you define USE_PREAD or USE_PREAD64, then it might also
3185 ** be necessary to define _XOPEN_SOURCE to be 500. This varies from
3186 ** one system to another. Since SQLite does not define USE_PREAD
3187 ** in any form by default, we will not attempt to define _XOPEN_SOURCE.
3188 ** See tickets #2741 and #2681.
3190 ** To avoid stomping the errno value on a failed read the lastErrno value
3191 ** is set before returning.
3193 static int seekAndRead(unixFile
*id
, sqlite3_int64 offset
, void *pBuf
, int cnt
){
3196 #if (!defined(USE_PREAD) && !defined(USE_PREAD64))
3200 assert( cnt
==(cnt
&0x1ffff) );
3203 #if defined(USE_PREAD)
3204 got
= osPread(id
->h
, pBuf
, cnt
, offset
);
3205 SimulateIOError( got
= -1 );
3206 #elif defined(USE_PREAD64)
3207 got
= osPread64(id
->h
, pBuf
, cnt
, offset
);
3208 SimulateIOError( got
= -1 );
3210 newOffset
= lseek(id
->h
, offset
, SEEK_SET
);
3211 SimulateIOError( newOffset
= -1 );
3213 storeLastErrno((unixFile
*)id
, errno
);
3216 got
= osRead(id
->h
, pBuf
, cnt
);
3218 if( got
==cnt
) break;
3220 if( errno
==EINTR
){ got
= 1; continue; }
3222 storeLastErrno((unixFile
*)id
, errno
);
3228 pBuf
= (void*)(got
+ (char*)pBuf
);
3232 OSTRACE(("READ %-3d %5d %7lld %llu\n",
3233 id
->h
, got
+prior
, offset
-prior
, TIMER_ELAPSED
));
3238 ** Read data from a file into a buffer. Return SQLITE_OK if all
3239 ** bytes were read successfully and SQLITE_IOERR if anything goes
3242 static int unixRead(
3246 sqlite3_int64 offset
3248 unixFile
*pFile
= (unixFile
*)id
;
3251 assert( offset
>=0 );
3254 /* If this is a database file (not a journal, master-journal or temp
3255 ** file), the bytes in the locking range should never be read or written. */
3257 assert( pFile
->pPreallocatedUnused
==0
3258 || offset
>=PENDING_BYTE
+512
3259 || offset
+amt
<=PENDING_BYTE
3263 #if SQLITE_MAX_MMAP_SIZE>0
3264 /* Deal with as much of this read request as possible by transfering
3265 ** data from the memory mapping using memcpy(). */
3266 if( offset
<pFile
->mmapSize
){
3267 if( offset
+amt
<= pFile
->mmapSize
){
3268 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], amt
);
3271 int nCopy
= pFile
->mmapSize
- offset
;
3272 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], nCopy
);
3273 pBuf
= &((u8
*)pBuf
)[nCopy
];
3280 got
= seekAndRead(pFile
, offset
, pBuf
, amt
);
3284 /* lastErrno set by seekAndRead */
3285 return SQLITE_IOERR_READ
;
3287 storeLastErrno(pFile
, 0); /* not a system error */
3288 /* Unread parts of the buffer must be zero-filled */
3289 memset(&((char*)pBuf
)[got
], 0, amt
-got
);
3290 return SQLITE_IOERR_SHORT_READ
;
3295 ** Attempt to seek the file-descriptor passed as the first argument to
3296 ** absolute offset iOff, then attempt to write nBuf bytes of data from
3297 ** pBuf to it. If an error occurs, return -1 and set *piErrno. Otherwise,
3298 ** return the actual number of bytes written (which may be less than
3301 static int seekAndWriteFd(
3302 int fd
, /* File descriptor to write to */
3303 i64 iOff
, /* File offset to begin writing at */
3304 const void *pBuf
, /* Copy data from this buffer to the file */
3305 int nBuf
, /* Size of buffer pBuf in bytes */
3306 int *piErrno
/* OUT: Error number if error occurs */
3308 int rc
= 0; /* Value returned by system call */
3310 assert( nBuf
==(nBuf
&0x1ffff) );
3312 assert( piErrno
!=0 );
3316 #if defined(USE_PREAD)
3317 do{ rc
= (int)osPwrite(fd
, pBuf
, nBuf
, iOff
); }while( rc
<0 && errno
==EINTR
);
3318 #elif defined(USE_PREAD64)
3319 do{ rc
= (int)osPwrite64(fd
, pBuf
, nBuf
, iOff
);}while( rc
<0 && errno
==EINTR
);
3322 i64 iSeek
= lseek(fd
, iOff
, SEEK_SET
);
3323 SimulateIOError( iSeek
= -1 );
3328 rc
= osWrite(fd
, pBuf
, nBuf
);
3329 }while( rc
<0 && errno
==EINTR
);
3333 OSTRACE(("WRITE %-3d %5d %7lld %llu\n", fd
, rc
, iOff
, TIMER_ELAPSED
));
3335 if( rc
<0 ) *piErrno
= errno
;
3341 ** Seek to the offset in id->offset then read cnt bytes into pBuf.
3342 ** Return the number of bytes actually read. Update the offset.
3344 ** To avoid stomping the errno value on a failed write the lastErrno value
3345 ** is set before returning.
3347 static int seekAndWrite(unixFile
*id
, i64 offset
, const void *pBuf
, int cnt
){
3348 return seekAndWriteFd(id
->h
, offset
, pBuf
, cnt
, &id
->lastErrno
);
3353 ** Write data from a buffer into a file. Return SQLITE_OK on success
3354 ** or some other error code on failure.
3356 static int unixWrite(
3360 sqlite3_int64 offset
3362 unixFile
*pFile
= (unixFile
*)id
;
3367 /* If this is a database file (not a journal, master-journal or temp
3368 ** file), the bytes in the locking range should never be read or written. */
3370 assert( pFile
->pPreallocatedUnused
==0
3371 || offset
>=PENDING_BYTE
+512
3372 || offset
+amt
<=PENDING_BYTE
3377 /* If we are doing a normal write to a database file (as opposed to
3378 ** doing a hot-journal rollback or a write to some file other than a
3379 ** normal database file) then record the fact that the database
3380 ** has changed. If the transaction counter is modified, record that
3383 if( pFile
->inNormalWrite
){
3384 pFile
->dbUpdate
= 1; /* The database has been modified */
3385 if( offset
<=24 && offset
+amt
>=27 ){
3388 SimulateIOErrorBenign(1);
3389 rc
= seekAndRead(pFile
, 24, oldCntr
, 4);
3390 SimulateIOErrorBenign(0);
3391 if( rc
!=4 || memcmp(oldCntr
, &((char*)pBuf
)[24-offset
], 4)!=0 ){
3392 pFile
->transCntrChng
= 1; /* The transaction counter has changed */
3398 #if defined(SQLITE_MMAP_READWRITE) && SQLITE_MAX_MMAP_SIZE>0
3399 /* Deal with as much of this write request as possible by transfering
3400 ** data from the memory mapping using memcpy(). */
3401 if( offset
<pFile
->mmapSize
){
3402 if( offset
+amt
<= pFile
->mmapSize
){
3403 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, amt
);
3406 int nCopy
= pFile
->mmapSize
- offset
;
3407 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, nCopy
);
3408 pBuf
= &((u8
*)pBuf
)[nCopy
];
3415 while( (wrote
= seekAndWrite(pFile
, offset
, pBuf
, amt
))<amt
&& wrote
>0 ){
3418 pBuf
= &((char*)pBuf
)[wrote
];
3420 SimulateIOError(( wrote
=(-1), amt
=1 ));
3421 SimulateDiskfullError(( wrote
=0, amt
=1 ));
3424 if( wrote
<0 && pFile
->lastErrno
!=ENOSPC
){
3425 /* lastErrno set by seekAndWrite */
3426 return SQLITE_IOERR_WRITE
;
3428 storeLastErrno(pFile
, 0); /* not a system error */
3438 ** Count the number of fullsyncs and normal syncs. This is used to test
3439 ** that syncs and fullsyncs are occurring at the right times.
3441 int sqlite3_sync_count
= 0;
3442 int sqlite3_fullsync_count
= 0;
3446 ** We do not trust systems to provide a working fdatasync(). Some do.
3447 ** Others do no. To be safe, we will stick with the (slightly slower)
3448 ** fsync(). If you know that your system does support fdatasync() correctly,
3449 ** then simply compile with -Dfdatasync=fdatasync or -DHAVE_FDATASYNC
3451 #if !defined(fdatasync) && !HAVE_FDATASYNC
3452 # define fdatasync fsync
3456 ** Define HAVE_FULLFSYNC to 0 or 1 depending on whether or not
3457 ** the F_FULLFSYNC macro is defined. F_FULLFSYNC is currently
3458 ** only available on Mac OS X. But that could change.
3461 # define HAVE_FULLFSYNC 1
3463 # define HAVE_FULLFSYNC 0
3468 ** The fsync() system call does not work as advertised on many
3469 ** unix systems. The following procedure is an attempt to make
3472 ** The SQLITE_NO_SYNC macro disables all fsync()s. This is useful
3473 ** for testing when we want to run through the test suite quickly.
3474 ** You are strongly advised *not* to deploy with SQLITE_NO_SYNC
3475 ** enabled, however, since with SQLITE_NO_SYNC enabled, an OS crash
3476 ** or power failure will likely corrupt the database file.
3478 ** SQLite sets the dataOnly flag if the size of the file is unchanged.
3479 ** The idea behind dataOnly is that it should only write the file content
3480 ** to disk, not the inode. We only set dataOnly if the file size is
3481 ** unchanged since the file size is part of the inode. However,
3482 ** Ted Ts'o tells us that fdatasync() will also write the inode if the
3483 ** file size has changed. The only real difference between fdatasync()
3484 ** and fsync(), Ted tells us, is that fdatasync() will not flush the
3485 ** inode if the mtime or owner or other inode attributes have changed.
3486 ** We only care about the file size, not the other file attributes, so
3487 ** as far as SQLite is concerned, an fdatasync() is always adequate.
3488 ** So, we always use fdatasync() if it is available, regardless of
3489 ** the value of the dataOnly flag.
3491 static int full_fsync(int fd
, int fullSync
, int dataOnly
){
3494 /* The following "ifdef/elif/else/" block has the same structure as
3495 ** the one below. It is replicated here solely to avoid cluttering
3496 ** up the real code with the UNUSED_PARAMETER() macros.
3498 #ifdef SQLITE_NO_SYNC
3499 UNUSED_PARAMETER(fd
);
3500 UNUSED_PARAMETER(fullSync
);
3501 UNUSED_PARAMETER(dataOnly
);
3502 #elif HAVE_FULLFSYNC
3503 UNUSED_PARAMETER(dataOnly
);
3505 UNUSED_PARAMETER(fullSync
);
3506 UNUSED_PARAMETER(dataOnly
);
3509 /* Record the number of times that we do a normal fsync() and
3510 ** FULLSYNC. This is used during testing to verify that this procedure
3511 ** gets called with the correct arguments.
3514 if( fullSync
) sqlite3_fullsync_count
++;
3515 sqlite3_sync_count
++;
3518 /* If we compiled with the SQLITE_NO_SYNC flag, then syncing is a
3519 ** no-op. But go ahead and call fstat() to validate the file
3520 ** descriptor as we need a method to provoke a failure during
3521 ** coverate testing.
3523 #ifdef SQLITE_NO_SYNC
3526 rc
= osFstat(fd
, &buf
);
3528 #elif HAVE_FULLFSYNC
3530 rc
= osFcntl(fd
, F_FULLFSYNC
, 0);
3534 /* If the FULLFSYNC failed, fall back to attempting an fsync().
3535 ** It shouldn't be possible for fullfsync to fail on the local
3536 ** file system (on OSX), so failure indicates that FULLFSYNC
3537 ** isn't supported for this file system. So, attempt an fsync
3538 ** and (for now) ignore the overhead of a superfluous fcntl call.
3539 ** It'd be better to detect fullfsync support once and avoid
3540 ** the fcntl call every time sync is called.
3542 if( rc
) rc
= fsync(fd
);
3544 #elif defined(__APPLE__)
3545 /* fdatasync() on HFS+ doesn't yet flush the file size if it changed correctly
3546 ** so currently we default to the macro that redefines fdatasync to fsync
3552 if( rc
==-1 && errno
==ENOTSUP
){
3555 #endif /* OS_VXWORKS */
3556 #endif /* ifdef SQLITE_NO_SYNC elif HAVE_FULLFSYNC */
3558 if( OS_VXWORKS
&& rc
!= -1 ){
3565 ** Open a file descriptor to the directory containing file zFilename.
3566 ** If successful, *pFd is set to the opened file descriptor and
3567 ** SQLITE_OK is returned. If an error occurs, either SQLITE_NOMEM
3568 ** or SQLITE_CANTOPEN is returned and *pFd is set to an undefined
3571 ** The directory file descriptor is used for only one thing - to
3572 ** fsync() a directory to make sure file creation and deletion events
3573 ** are flushed to disk. Such fsyncs are not needed on newer
3574 ** journaling filesystems, but are required on older filesystems.
3576 ** This routine can be overridden using the xSetSysCall interface.
3577 ** The ability to override this routine was added in support of the
3578 ** chromium sandbox. Opening a directory is a security risk (we are
3579 ** told) so making it overrideable allows the chromium sandbox to
3580 ** replace this routine with a harmless no-op. To make this routine
3581 ** a no-op, replace it with a stub that returns SQLITE_OK but leaves
3582 ** *pFd set to a negative number.
3584 ** If SQLITE_OK is returned, the caller is responsible for closing
3585 ** the file descriptor *pFd using close().
3587 static int openDirectory(const char *zFilename
, int *pFd
){
3590 char zDirname
[MAX_PATHNAME
+1];
3592 sqlite3_snprintf(MAX_PATHNAME
, zDirname
, "%s", zFilename
);
3593 for(ii
=(int)strlen(zDirname
); ii
>0 && zDirname
[ii
]!='/'; ii
--);
3595 zDirname
[ii
] = '\0';
3597 if( zDirname
[0]!='/' ) zDirname
[0] = '.';
3600 fd
= robust_open(zDirname
, O_RDONLY
|O_BINARY
, 0);
3602 OSTRACE(("OPENDIR %-3d %s\n", fd
, zDirname
));
3605 if( fd
>=0 ) return SQLITE_OK
;
3606 return unixLogError(SQLITE_CANTOPEN_BKPT
, "openDirectory", zDirname
);
3610 ** Make sure all writes to a particular file are committed to disk.
3612 ** If dataOnly==0 then both the file itself and its metadata (file
3613 ** size, access time, etc) are synced. If dataOnly!=0 then only the
3614 ** file data is synced.
3616 ** Under Unix, also make sure that the directory entry for the file
3617 ** has been created by fsync-ing the directory that contains the file.
3618 ** If we do not do this and we encounter a power failure, the directory
3619 ** entry for the journal might not exist after we reboot. The next
3620 ** SQLite to access the file will not know that the journal exists (because
3621 ** the directory entry for the journal was never created) and the transaction
3622 ** will not roll back - possibly leading to database corruption.
3624 static int unixSync(sqlite3_file
*id
, int flags
){
3626 unixFile
*pFile
= (unixFile
*)id
;
3628 int isDataOnly
= (flags
&SQLITE_SYNC_DATAONLY
);
3629 int isFullsync
= (flags
&0x0F)==SQLITE_SYNC_FULL
;
3631 /* Check that one of SQLITE_SYNC_NORMAL or FULL was passed */
3632 assert((flags
&0x0F)==SQLITE_SYNC_NORMAL
3633 || (flags
&0x0F)==SQLITE_SYNC_FULL
3636 /* Unix cannot, but some systems may return SQLITE_FULL from here. This
3637 ** line is to test that doing so does not cause any problems.
3639 SimulateDiskfullError( return SQLITE_FULL
);
3642 OSTRACE(("SYNC %-3d\n", pFile
->h
));
3643 rc
= full_fsync(pFile
->h
, isFullsync
, isDataOnly
);
3644 SimulateIOError( rc
=1 );
3646 storeLastErrno(pFile
, errno
);
3647 return unixLogError(SQLITE_IOERR_FSYNC
, "full_fsync", pFile
->zPath
);
3650 /* Also fsync the directory containing the file if the DIRSYNC flag
3651 ** is set. This is a one-time occurrence. Many systems (examples: AIX)
3652 ** are unable to fsync a directory, so ignore errors on the fsync.
3654 if( pFile
->ctrlFlags
& UNIXFILE_DIRSYNC
){
3656 OSTRACE(("DIRSYNC %s (have_fullfsync=%d fullsync=%d)\n", pFile
->zPath
,
3657 HAVE_FULLFSYNC
, isFullsync
));
3658 rc
= osOpenDirectory(pFile
->zPath
, &dirfd
);
3659 if( rc
==SQLITE_OK
){
3660 full_fsync(dirfd
, 0, 0);
3661 robust_close(pFile
, dirfd
, __LINE__
);
3663 assert( rc
==SQLITE_CANTOPEN
);
3666 pFile
->ctrlFlags
&= ~UNIXFILE_DIRSYNC
;
3672 ** Truncate an open file to a specified size
3674 static int unixTruncate(sqlite3_file
*id
, i64 nByte
){
3675 unixFile
*pFile
= (unixFile
*)id
;
3678 SimulateIOError( return SQLITE_IOERR_TRUNCATE
);
3680 /* If the user has configured a chunk-size for this file, truncate the
3681 ** file so that it consists of an integer number of chunks (i.e. the
3682 ** actual file size after the operation may be larger than the requested
3685 if( pFile
->szChunk
>0 ){
3686 nByte
= ((nByte
+ pFile
->szChunk
- 1)/pFile
->szChunk
) * pFile
->szChunk
;
3689 rc
= robust_ftruncate(pFile
->h
, nByte
);
3691 storeLastErrno(pFile
, errno
);
3692 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3695 /* If we are doing a normal write to a database file (as opposed to
3696 ** doing a hot-journal rollback or a write to some file other than a
3697 ** normal database file) and we truncate the file to zero length,
3698 ** that effectively updates the change counter. This might happen
3699 ** when restoring a database using the backup API from a zero-length
3702 if( pFile
->inNormalWrite
&& nByte
==0 ){
3703 pFile
->transCntrChng
= 1;
3707 #if SQLITE_MAX_MMAP_SIZE>0
3708 /* If the file was just truncated to a size smaller than the currently
3709 ** mapped region, reduce the effective mapping size as well. SQLite will
3710 ** use read() and write() to access data beyond this point from now on.
3712 if( nByte
<pFile
->mmapSize
){
3713 pFile
->mmapSize
= nByte
;
3722 ** Determine the current size of a file in bytes
3724 static int unixFileSize(sqlite3_file
*id
, i64
*pSize
){
3728 rc
= osFstat(((unixFile
*)id
)->h
, &buf
);
3729 SimulateIOError( rc
=1 );
3731 storeLastErrno((unixFile
*)id
, errno
);
3732 return SQLITE_IOERR_FSTAT
;
3734 *pSize
= buf
.st_size
;
3736 /* When opening a zero-size database, the findInodeInfo() procedure
3737 ** writes a single byte into that file in order to work around a bug
3738 ** in the OS-X msdos filesystem. In order to avoid problems with upper
3739 ** layers, we need to report this file size as zero even though it is
3740 ** really 1. Ticket #3260.
3742 if( *pSize
==1 ) *pSize
= 0;
3748 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3750 ** Handler for proxy-locking file-control verbs. Defined below in the
3751 ** proxying locking division.
3753 static int proxyFileControl(sqlite3_file
*,int,void*);
3757 ** This function is called to handle the SQLITE_FCNTL_SIZE_HINT
3758 ** file-control operation. Enlarge the database to nBytes in size
3759 ** (rounded up to the next chunk-size). If the database is already
3760 ** nBytes or larger, this routine is a no-op.
3762 static int fcntlSizeHint(unixFile
*pFile
, i64 nByte
){
3763 if( pFile
->szChunk
>0 ){
3764 i64 nSize
; /* Required file size */
3765 struct stat buf
; /* Used to hold return values of fstat() */
3767 if( osFstat(pFile
->h
, &buf
) ){
3768 return SQLITE_IOERR_FSTAT
;
3771 nSize
= ((nByte
+pFile
->szChunk
-1) / pFile
->szChunk
) * pFile
->szChunk
;
3772 if( nSize
>(i64
)buf
.st_size
){
3774 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
3775 /* The code below is handling the return value of osFallocate()
3776 ** correctly. posix_fallocate() is defined to "returns zero on success,
3777 ** or an error number on failure". See the manpage for details. */
3780 err
= osFallocate(pFile
->h
, buf
.st_size
, nSize
-buf
.st_size
);
3781 }while( err
==EINTR
);
3782 if( err
&& err
!=EINVAL
) return SQLITE_IOERR_WRITE
;
3784 /* If the OS does not have posix_fallocate(), fake it. Write a
3785 ** single byte to the last byte in each block that falls entirely
3786 ** within the extended region. Then, if required, a single byte
3787 ** at offset (nSize-1), to set the size of the file correctly.
3788 ** This is a similar technique to that used by glibc on systems
3789 ** that do not have a real fallocate() call.
3791 int nBlk
= buf
.st_blksize
; /* File-system block size */
3792 int nWrite
= 0; /* Number of bytes written by seekAndWrite */
3793 i64 iWrite
; /* Next offset to write to */
3795 iWrite
= (buf
.st_size
/nBlk
)*nBlk
+ nBlk
- 1;
3796 assert( iWrite
>=buf
.st_size
);
3797 assert( ((iWrite
+1)%nBlk
)==0 );
3798 for(/*no-op*/; iWrite
<nSize
+nBlk
-1; iWrite
+=nBlk
){
3799 if( iWrite
>=nSize
) iWrite
= nSize
- 1;
3800 nWrite
= seekAndWrite(pFile
, iWrite
, "", 1);
3801 if( nWrite
!=1 ) return SQLITE_IOERR_WRITE
;
3807 #if SQLITE_MAX_MMAP_SIZE>0
3808 if( pFile
->mmapSizeMax
>0 && nByte
>pFile
->mmapSize
){
3810 if( pFile
->szChunk
<=0 ){
3811 if( robust_ftruncate(pFile
->h
, nByte
) ){
3812 storeLastErrno(pFile
, errno
);
3813 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3817 rc
= unixMapfile(pFile
, nByte
);
3826 ** If *pArg is initially negative then this is a query. Set *pArg to
3827 ** 1 or 0 depending on whether or not bit mask of pFile->ctrlFlags is set.
3829 ** If *pArg is 0 or 1, then clear or set the mask bit of pFile->ctrlFlags.
3831 static void unixModeBit(unixFile
*pFile
, unsigned char mask
, int *pArg
){
3833 *pArg
= (pFile
->ctrlFlags
& mask
)!=0;
3834 }else if( (*pArg
)==0 ){
3835 pFile
->ctrlFlags
&= ~mask
;
3837 pFile
->ctrlFlags
|= mask
;
3841 /* Forward declaration */
3842 static int unixGetTempname(int nBuf
, char *zBuf
);
3845 ** Information and control of an open file handle.
3847 static int unixFileControl(sqlite3_file
*id
, int op
, void *pArg
){
3848 unixFile
*pFile
= (unixFile
*)id
;
3850 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3851 case SQLITE_FCNTL_BEGIN_ATOMIC_WRITE
: {
3852 int rc
= osIoctl(pFile
->h
, F2FS_IOC_START_ATOMIC_WRITE
);
3853 return rc
? SQLITE_IOERR_BEGIN_ATOMIC
: SQLITE_OK
;
3855 case SQLITE_FCNTL_COMMIT_ATOMIC_WRITE
: {
3856 int rc
= osIoctl(pFile
->h
, F2FS_IOC_COMMIT_ATOMIC_WRITE
);
3857 return rc
? SQLITE_IOERR_COMMIT_ATOMIC
: SQLITE_OK
;
3859 case SQLITE_FCNTL_ROLLBACK_ATOMIC_WRITE
: {
3860 int rc
= osIoctl(pFile
->h
, F2FS_IOC_ABORT_VOLATILE_WRITE
);
3861 return rc
? SQLITE_IOERR_ROLLBACK_ATOMIC
: SQLITE_OK
;
3863 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3865 case SQLITE_FCNTL_LOCKSTATE
: {
3866 *(int*)pArg
= pFile
->eFileLock
;
3869 case SQLITE_FCNTL_LAST_ERRNO
: {
3870 *(int*)pArg
= pFile
->lastErrno
;
3873 case SQLITE_FCNTL_CHUNK_SIZE
: {
3874 pFile
->szChunk
= *(int *)pArg
;
3877 case SQLITE_FCNTL_SIZE_HINT
: {
3879 SimulateIOErrorBenign(1);
3880 rc
= fcntlSizeHint(pFile
, *(i64
*)pArg
);
3881 SimulateIOErrorBenign(0);
3884 case SQLITE_FCNTL_PERSIST_WAL
: {
3885 unixModeBit(pFile
, UNIXFILE_PERSIST_WAL
, (int*)pArg
);
3888 case SQLITE_FCNTL_POWERSAFE_OVERWRITE
: {
3889 unixModeBit(pFile
, UNIXFILE_PSOW
, (int*)pArg
);
3892 case SQLITE_FCNTL_VFSNAME
: {
3893 *(char**)pArg
= sqlite3_mprintf("%s", pFile
->pVfs
->zName
);
3896 case SQLITE_FCNTL_TEMPFILENAME
: {
3897 char *zTFile
= sqlite3_malloc64( pFile
->pVfs
->mxPathname
);
3899 unixGetTempname(pFile
->pVfs
->mxPathname
, zTFile
);
3900 *(char**)pArg
= zTFile
;
3904 case SQLITE_FCNTL_HAS_MOVED
: {
3905 *(int*)pArg
= fileHasMoved(pFile
);
3908 #ifdef SQLITE_ENABLE_SETLK_TIMEOUT
3909 case SQLITE_FCNTL_LOCK_TIMEOUT
: {
3910 pFile
->iBusyTimeout
= *(int*)pArg
;
3914 #if SQLITE_MAX_MMAP_SIZE>0
3915 case SQLITE_FCNTL_MMAP_SIZE
: {
3916 i64 newLimit
= *(i64
*)pArg
;
3918 if( newLimit
>sqlite3GlobalConfig
.mxMmap
){
3919 newLimit
= sqlite3GlobalConfig
.mxMmap
;
3922 /* The value of newLimit may be eventually cast to (size_t) and passed
3923 ** to mmap(). Restrict its value to 2GB if (size_t) is not at least a
3925 if( newLimit
>0 && sizeof(size_t)<8 ){
3926 newLimit
= (newLimit
& 0x7FFFFFFF);
3929 *(i64
*)pArg
= pFile
->mmapSizeMax
;
3930 if( newLimit
>=0 && newLimit
!=pFile
->mmapSizeMax
&& pFile
->nFetchOut
==0 ){
3931 pFile
->mmapSizeMax
= newLimit
;
3932 if( pFile
->mmapSize
>0 ){
3933 unixUnmapfile(pFile
);
3934 rc
= unixMapfile(pFile
, -1);
3941 /* The pager calls this method to signal that it has done
3942 ** a rollback and that the database is therefore unchanged and
3943 ** it hence it is OK for the transaction change counter to be
3946 case SQLITE_FCNTL_DB_UNCHANGED
: {
3947 ((unixFile
*)id
)->dbUpdate
= 0;
3951 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3952 case SQLITE_FCNTL_SET_LOCKPROXYFILE
:
3953 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
3954 return proxyFileControl(id
,op
,pArg
);
3956 #endif /* SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__) */
3958 return SQLITE_NOTFOUND
;
3962 ** If pFd->sectorSize is non-zero when this function is called, it is a
3963 ** no-op. Otherwise, the values of pFd->sectorSize and
3964 ** pFd->deviceCharacteristics are set according to the file-system
3967 ** There are two versions of this function. One for QNX and one for all
3971 static void setDeviceCharacteristics(unixFile
*pFd
){
3972 assert( pFd
->deviceCharacteristics
==0 || pFd
->sectorSize
!=0 );
3973 if( pFd
->sectorSize
==0 ){
3974 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3978 /* Check for support for F2FS atomic batch writes. */
3979 res
= osIoctl(pFd
->h
, F2FS_IOC_GET_FEATURES
, &f
);
3980 if( res
==0 && (f
& F2FS_FEATURE_ATOMIC_WRITE
) ){
3981 pFd
->deviceCharacteristics
= SQLITE_IOCAP_BATCH_ATOMIC
;
3983 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3985 /* Set the POWERSAFE_OVERWRITE flag if requested. */
3986 if( pFd
->ctrlFlags
& UNIXFILE_PSOW
){
3987 pFd
->deviceCharacteristics
|= SQLITE_IOCAP_POWERSAFE_OVERWRITE
;
3990 pFd
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
3994 #include <sys/dcmd_blk.h>
3995 #include <sys/statvfs.h>
3996 static void setDeviceCharacteristics(unixFile
*pFile
){
3997 if( pFile
->sectorSize
== 0 ){
3998 struct statvfs fsInfo
;
4000 /* Set defaults for non-supported filesystems */
4001 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
4002 pFile
->deviceCharacteristics
= 0;
4003 if( fstatvfs(pFile
->h
, &fsInfo
) == -1 ) {
4007 if( !strcmp(fsInfo
.f_basetype
, "tmp") ) {
4008 pFile
->sectorSize
= fsInfo
.f_bsize
;
4009 pFile
->deviceCharacteristics
=
4010 SQLITE_IOCAP_ATOMIC4K
| /* All ram filesystem writes are atomic */
4011 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4012 ** the write succeeds */
4013 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4014 ** so it is ordered */
4016 }else if( strstr(fsInfo
.f_basetype
, "etfs") ){
4017 pFile
->sectorSize
= fsInfo
.f_bsize
;
4018 pFile
->deviceCharacteristics
=
4019 /* etfs cluster size writes are atomic */
4020 (pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) |
4021 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4022 ** the write succeeds */
4023 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4024 ** so it is ordered */
4026 }else if( !strcmp(fsInfo
.f_basetype
, "qnx6") ){
4027 pFile
->sectorSize
= fsInfo
.f_bsize
;
4028 pFile
->deviceCharacteristics
=
4029 SQLITE_IOCAP_ATOMIC
| /* All filesystem writes are atomic */
4030 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4031 ** the write succeeds */
4032 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4033 ** so it is ordered */
4035 }else if( !strcmp(fsInfo
.f_basetype
, "qnx4") ){
4036 pFile
->sectorSize
= fsInfo
.f_bsize
;
4037 pFile
->deviceCharacteristics
=
4038 /* full bitset of atomics from max sector size and smaller */
4039 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
4040 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4041 ** so it is ordered */
4043 }else if( strstr(fsInfo
.f_basetype
, "dos") ){
4044 pFile
->sectorSize
= fsInfo
.f_bsize
;
4045 pFile
->deviceCharacteristics
=
4046 /* full bitset of atomics from max sector size and smaller */
4047 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
4048 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
4049 ** so it is ordered */
4052 pFile
->deviceCharacteristics
=
4053 SQLITE_IOCAP_ATOMIC512
| /* blocks are atomic */
4054 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4055 ** the write succeeds */
4059 /* Last chance verification. If the sector size isn't a multiple of 512
4060 ** then it isn't valid.*/
4061 if( pFile
->sectorSize
% 512 != 0 ){
4062 pFile
->deviceCharacteristics
= 0;
4063 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
4069 ** Return the sector size in bytes of the underlying block device for
4070 ** the specified file. This is almost always 512 bytes, but may be
4071 ** larger for some devices.
4073 ** SQLite code assumes this function cannot fail. It also assumes that
4074 ** if two files are created in the same file-system directory (i.e.
4075 ** a database and its journal file) that the sector size will be the
4078 static int unixSectorSize(sqlite3_file
*id
){
4079 unixFile
*pFd
= (unixFile
*)id
;
4080 setDeviceCharacteristics(pFd
);
4081 return pFd
->sectorSize
;
4085 ** Return the device characteristics for the file.
4087 ** This VFS is set up to return SQLITE_IOCAP_POWERSAFE_OVERWRITE by default.
4088 ** However, that choice is controversial since technically the underlying
4089 ** file system does not always provide powersafe overwrites. (In other
4090 ** words, after a power-loss event, parts of the file that were never
4091 ** written might end up being altered.) However, non-PSOW behavior is very,
4092 ** very rare. And asserting PSOW makes a large reduction in the amount
4093 ** of required I/O for journaling, since a lot of padding is eliminated.
4094 ** Hence, while POWERSAFE_OVERWRITE is on by default, there is a file-control
4095 ** available to turn it off and URI query parameter available to turn it off.
4097 static int unixDeviceCharacteristics(sqlite3_file
*id
){
4098 unixFile
*pFd
= (unixFile
*)id
;
4099 setDeviceCharacteristics(pFd
);
4100 return pFd
->deviceCharacteristics
;
4103 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
4106 ** Return the system page size.
4108 ** This function should not be called directly by other code in this file.
4109 ** Instead, it should be called via macro osGetpagesize().
4111 static int unixGetpagesize(void){
4114 #elif defined(_BSD_SOURCE)
4115 return getpagesize();
4117 return (int)sysconf(_SC_PAGESIZE
);
4121 #endif /* !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0 */
4123 #ifndef SQLITE_OMIT_WAL
4126 ** Object used to represent an shared memory buffer.
4128 ** When multiple threads all reference the same wal-index, each thread
4129 ** has its own unixShm object, but they all point to a single instance
4130 ** of this unixShmNode object. In other words, each wal-index is opened
4131 ** only once per process.
4133 ** Each unixShmNode object is connected to a single unixInodeInfo object.
4134 ** We could coalesce this object into unixInodeInfo, but that would mean
4135 ** every open file that does not use shared memory (in other words, most
4136 ** open files) would have to carry around this extra information. So
4137 ** the unixInodeInfo object contains a pointer to this unixShmNode object
4138 ** and the unixShmNode object is created only when needed.
4140 ** unixMutexHeld() must be true when creating or destroying
4141 ** this object or while reading or writing the following fields:
4145 ** The following fields are read-only after the object is created:
4150 ** Either unixShmNode.mutex must be held or unixShmNode.nRef==0 and
4151 ** unixMutexHeld() is true when reading or writing any other field
4152 ** in this structure.
4154 struct unixShmNode
{
4155 unixInodeInfo
*pInode
; /* unixInodeInfo that owns this SHM node */
4156 sqlite3_mutex
*mutex
; /* Mutex to access this object */
4157 char *zFilename
; /* Name of the mmapped file */
4158 int h
; /* Open file descriptor */
4159 int szRegion
; /* Size of shared-memory regions */
4160 u16 nRegion
; /* Size of array apRegion */
4161 u8 isReadonly
; /* True if read-only */
4162 u8 isUnlocked
; /* True if no DMS lock held */
4163 char **apRegion
; /* Array of mapped shared-memory regions */
4164 int nRef
; /* Number of unixShm objects pointing to this */
4165 unixShm
*pFirst
; /* All unixShm objects pointing to this */
4167 u8 exclMask
; /* Mask of exclusive locks held */
4168 u8 sharedMask
; /* Mask of shared locks held */
4169 u8 nextShmId
; /* Next available unixShm.id value */
4174 ** Structure used internally by this VFS to record the state of an
4175 ** open shared memory connection.
4177 ** The following fields are initialized when this object is created and
4178 ** are read-only thereafter:
4183 ** All other fields are read/write. The unixShm.pFile->mutex must be held
4184 ** while accessing any read/write fields.
4187 unixShmNode
*pShmNode
; /* The underlying unixShmNode object */
4188 unixShm
*pNext
; /* Next unixShm with the same unixShmNode */
4189 u8 hasMutex
; /* True if holding the unixShmNode mutex */
4190 u8 id
; /* Id of this connection within its unixShmNode */
4191 u16 sharedMask
; /* Mask of shared locks held */
4192 u16 exclMask
; /* Mask of exclusive locks held */
4196 ** Constants used for locking
4198 #define UNIX_SHM_BASE ((22+SQLITE_SHM_NLOCK)*4) /* first lock byte */
4199 #define UNIX_SHM_DMS (UNIX_SHM_BASE+SQLITE_SHM_NLOCK) /* deadman switch */
4202 ** Apply posix advisory locks for all bytes from ofst through ofst+n-1.
4204 ** Locks block if the mask is exactly UNIX_SHM_C and are non-blocking
4207 static int unixShmSystemLock(
4208 unixFile
*pFile
, /* Open connection to the WAL file */
4209 int lockType
, /* F_UNLCK, F_RDLCK, or F_WRLCK */
4210 int ofst
, /* First byte of the locking range */
4211 int n
/* Number of bytes to lock */
4213 unixShmNode
*pShmNode
; /* Apply locks to this open shared-memory segment */
4214 struct flock f
; /* The posix advisory locking structure */
4215 int rc
= SQLITE_OK
; /* Result code form fcntl() */
4217 /* Access to the unixShmNode object is serialized by the caller */
4218 pShmNode
= pFile
->pInode
->pShmNode
;
4219 assert( pShmNode
->nRef
==0 || sqlite3_mutex_held(pShmNode
->mutex
) );
4221 /* Shared locks never span more than one byte */
4222 assert( n
==1 || lockType
!=F_RDLCK
);
4224 /* Locks are within range */
4225 assert( n
>=1 && n
<=SQLITE_SHM_NLOCK
);
4227 if( pShmNode
->h
>=0 ){
4228 /* Initialize the locking parameters */
4229 f
.l_type
= lockType
;
4230 f
.l_whence
= SEEK_SET
;
4233 rc
= osSetPosixAdvisoryLock(pShmNode
->h
, &f
, pFile
);
4234 rc
= (rc
!=(-1)) ? SQLITE_OK
: SQLITE_BUSY
;
4237 /* Update the global lock state and do debug tracing */
4240 OSTRACE(("SHM-LOCK "));
4241 mask
= ofst
>31 ? 0xffff : (1<<(ofst
+n
)) - (1<<ofst
);
4242 if( rc
==SQLITE_OK
){
4243 if( lockType
==F_UNLCK
){
4244 OSTRACE(("unlock %d ok", ofst
));
4245 pShmNode
->exclMask
&= ~mask
;
4246 pShmNode
->sharedMask
&= ~mask
;
4247 }else if( lockType
==F_RDLCK
){
4248 OSTRACE(("read-lock %d ok", ofst
));
4249 pShmNode
->exclMask
&= ~mask
;
4250 pShmNode
->sharedMask
|= mask
;
4252 assert( lockType
==F_WRLCK
);
4253 OSTRACE(("write-lock %d ok", ofst
));
4254 pShmNode
->exclMask
|= mask
;
4255 pShmNode
->sharedMask
&= ~mask
;
4258 if( lockType
==F_UNLCK
){
4259 OSTRACE(("unlock %d failed", ofst
));
4260 }else if( lockType
==F_RDLCK
){
4261 OSTRACE(("read-lock failed"));
4263 assert( lockType
==F_WRLCK
);
4264 OSTRACE(("write-lock %d failed", ofst
));
4267 OSTRACE((" - afterwards %03x,%03x\n",
4268 pShmNode
->sharedMask
, pShmNode
->exclMask
));
4276 ** Return the minimum number of 32KB shm regions that should be mapped at
4277 ** a time, assuming that each mapping must be an integer multiple of the
4278 ** current system page-size.
4280 ** Usually, this is 1. The exception seems to be systems that are configured
4281 ** to use 64KB pages - in this case each mapping must cover at least two
4284 static int unixShmRegionPerMap(void){
4285 int shmsz
= 32*1024; /* SHM region size */
4286 int pgsz
= osGetpagesize(); /* System page size */
4287 assert( ((pgsz
-1)&pgsz
)==0 ); /* Page size must be a power of 2 */
4288 if( pgsz
<shmsz
) return 1;
4293 ** Purge the unixShmNodeList list of all entries with unixShmNode.nRef==0.
4295 ** This is not a VFS shared-memory method; it is a utility function called
4296 ** by VFS shared-memory methods.
4298 static void unixShmPurge(unixFile
*pFd
){
4299 unixShmNode
*p
= pFd
->pInode
->pShmNode
;
4300 assert( unixMutexHeld() );
4301 if( p
&& ALWAYS(p
->nRef
==0) ){
4302 int nShmPerMap
= unixShmRegionPerMap();
4304 assert( p
->pInode
==pFd
->pInode
);
4305 sqlite3_mutex_free(p
->mutex
);
4306 for(i
=0; i
<p
->nRegion
; i
+=nShmPerMap
){
4308 osMunmap(p
->apRegion
[i
], p
->szRegion
);
4310 sqlite3_free(p
->apRegion
[i
]);
4313 sqlite3_free(p
->apRegion
);
4315 robust_close(pFd
, p
->h
, __LINE__
);
4318 p
->pInode
->pShmNode
= 0;
4324 ** The DMS lock has not yet been taken on shm file pShmNode. Attempt to
4325 ** take it now. Return SQLITE_OK if successful, or an SQLite error
4328 ** If the DMS cannot be locked because this is a readonly_shm=1
4329 ** connection and no other process already holds a lock, return
4330 ** SQLITE_READONLY_CANTINIT and set pShmNode->isUnlocked=1.
4332 static int unixLockSharedMemory(unixFile
*pDbFd
, unixShmNode
*pShmNode
){
4336 /* Use F_GETLK to determine the locks other processes are holding
4337 ** on the DMS byte. If it indicates that another process is holding
4338 ** a SHARED lock, then this process may also take a SHARED lock
4339 ** and proceed with opening the *-shm file.
4341 ** Or, if no other process is holding any lock, then this process
4342 ** is the first to open it. In this case take an EXCLUSIVE lock on the
4343 ** DMS byte and truncate the *-shm file to zero bytes in size. Then
4344 ** downgrade to a SHARED lock on the DMS byte.
4346 ** If another process is holding an EXCLUSIVE lock on the DMS byte,
4347 ** return SQLITE_BUSY to the caller (it will try again). An earlier
4348 ** version of this code attempted the SHARED lock at this point. But
4349 ** this introduced a subtle race condition: if the process holding
4350 ** EXCLUSIVE failed just before truncating the *-shm file, then this
4351 ** process might open and use the *-shm file without truncating it.
4352 ** And if the *-shm file has been corrupted by a power failure or
4353 ** system crash, the database itself may also become corrupt. */
4354 lock
.l_whence
= SEEK_SET
;
4355 lock
.l_start
= UNIX_SHM_DMS
;
4357 lock
.l_type
= F_WRLCK
;
4358 if( osFcntl(pShmNode
->h
, F_GETLK
, &lock
)!=0 ) {
4359 rc
= SQLITE_IOERR_LOCK
;
4360 }else if( lock
.l_type
==F_UNLCK
){
4361 if( pShmNode
->isReadonly
){
4362 pShmNode
->isUnlocked
= 1;
4363 rc
= SQLITE_READONLY_CANTINIT
;
4365 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, UNIX_SHM_DMS
, 1);
4366 if( rc
==SQLITE_OK
&& robust_ftruncate(pShmNode
->h
, 0) ){
4367 rc
= unixLogError(SQLITE_IOERR_SHMOPEN
,"ftruncate",pShmNode
->zFilename
);
4370 }else if( lock
.l_type
==F_WRLCK
){
4374 if( rc
==SQLITE_OK
){
4375 assert( lock
.l_type
==F_UNLCK
|| lock
.l_type
==F_RDLCK
);
4376 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, UNIX_SHM_DMS
, 1);
4382 ** Open a shared-memory area associated with open database file pDbFd.
4383 ** This particular implementation uses mmapped files.
4385 ** The file used to implement shared-memory is in the same directory
4386 ** as the open database file and has the same name as the open database
4387 ** file with the "-shm" suffix added. For example, if the database file
4388 ** is "/home/user1/config.db" then the file that is created and mmapped
4389 ** for shared memory will be called "/home/user1/config.db-shm".
4391 ** Another approach to is to use files in /dev/shm or /dev/tmp or an
4392 ** some other tmpfs mount. But if a file in a different directory
4393 ** from the database file is used, then differing access permissions
4394 ** or a chroot() might cause two different processes on the same
4395 ** database to end up using different files for shared memory -
4396 ** meaning that their memory would not really be shared - resulting
4397 ** in database corruption. Nevertheless, this tmpfs file usage
4398 ** can be enabled at compile-time using -DSQLITE_SHM_DIRECTORY="/dev/shm"
4399 ** or the equivalent. The use of the SQLITE_SHM_DIRECTORY compile-time
4400 ** option results in an incompatible build of SQLite; builds of SQLite
4401 ** that with differing SQLITE_SHM_DIRECTORY settings attempt to use the
4402 ** same database file at the same time, database corruption will likely
4403 ** result. The SQLITE_SHM_DIRECTORY compile-time option is considered
4404 ** "unsupported" and may go away in a future SQLite release.
4406 ** When opening a new shared-memory file, if no other instances of that
4407 ** file are currently open, in this process or in other processes, then
4408 ** the file must be truncated to zero length or have its header cleared.
4410 ** If the original database file (pDbFd) is using the "unix-excl" VFS
4411 ** that means that an exclusive lock is held on the database file and
4412 ** that no other processes are able to read or write the database. In
4413 ** that case, we do not really need shared memory. No shared memory
4414 ** file is created. The shared memory will be simulated with heap memory.
4416 static int unixOpenSharedMemory(unixFile
*pDbFd
){
4417 struct unixShm
*p
= 0; /* The connection to be opened */
4418 struct unixShmNode
*pShmNode
; /* The underlying mmapped file */
4419 int rc
= SQLITE_OK
; /* Result code */
4420 unixInodeInfo
*pInode
; /* The inode of fd */
4421 char *zShm
; /* Name of the file used for SHM */
4422 int nShmFilename
; /* Size of the SHM filename in bytes */
4424 /* Allocate space for the new unixShm object. */
4425 p
= sqlite3_malloc64( sizeof(*p
) );
4426 if( p
==0 ) return SQLITE_NOMEM_BKPT
;
4427 memset(p
, 0, sizeof(*p
));
4428 assert( pDbFd
->pShm
==0 );
4430 /* Check to see if a unixShmNode object already exists. Reuse an existing
4431 ** one if present. Create a new one if necessary.
4434 pInode
= pDbFd
->pInode
;
4435 pShmNode
= pInode
->pShmNode
;
4437 struct stat sStat
; /* fstat() info for database file */
4438 #ifndef SQLITE_SHM_DIRECTORY
4439 const char *zBasePath
= pDbFd
->zPath
;
4442 /* Call fstat() to figure out the permissions on the database file. If
4443 ** a new *-shm file is created, an attempt will be made to create it
4444 ** with the same permissions.
4446 if( osFstat(pDbFd
->h
, &sStat
) ){
4447 rc
= SQLITE_IOERR_FSTAT
;
4451 #ifdef SQLITE_SHM_DIRECTORY
4452 nShmFilename
= sizeof(SQLITE_SHM_DIRECTORY
) + 31;
4454 nShmFilename
= 6 + (int)strlen(zBasePath
);
4456 pShmNode
= sqlite3_malloc64( sizeof(*pShmNode
) + nShmFilename
);
4458 rc
= SQLITE_NOMEM_BKPT
;
4461 memset(pShmNode
, 0, sizeof(*pShmNode
)+nShmFilename
);
4462 zShm
= pShmNode
->zFilename
= (char*)&pShmNode
[1];
4463 #ifdef SQLITE_SHM_DIRECTORY
4464 sqlite3_snprintf(nShmFilename
, zShm
,
4465 SQLITE_SHM_DIRECTORY
"/sqlite-shm-%x-%x",
4466 (u32
)sStat
.st_ino
, (u32
)sStat
.st_dev
);
4468 sqlite3_snprintf(nShmFilename
, zShm
, "%s-shm", zBasePath
);
4469 sqlite3FileSuffix3(pDbFd
->zPath
, zShm
);
4472 pDbFd
->pInode
->pShmNode
= pShmNode
;
4473 pShmNode
->pInode
= pDbFd
->pInode
;
4474 if( sqlite3GlobalConfig
.bCoreMutex
){
4475 pShmNode
->mutex
= sqlite3_mutex_alloc(SQLITE_MUTEX_FAST
);
4476 if( pShmNode
->mutex
==0 ){
4477 rc
= SQLITE_NOMEM_BKPT
;
4482 if( pInode
->bProcessLock
==0 ){
4483 if( 0==sqlite3_uri_boolean(pDbFd
->zPath
, "readonly_shm", 0) ){
4484 pShmNode
->h
= robust_open(zShm
, O_RDWR
|O_CREAT
, (sStat
.st_mode
&0777));
4486 if( pShmNode
->h
<0 ){
4487 pShmNode
->h
= robust_open(zShm
, O_RDONLY
, (sStat
.st_mode
&0777));
4488 if( pShmNode
->h
<0 ){
4489 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zShm
);
4492 pShmNode
->isReadonly
= 1;
4495 /* If this process is running as root, make sure that the SHM file
4496 ** is owned by the same user that owns the original database. Otherwise,
4497 ** the original owner will not be able to connect.
4499 robustFchown(pShmNode
->h
, sStat
.st_uid
, sStat
.st_gid
);
4501 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4502 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_READONLY_CANTINIT
) goto shm_open_err
;
4506 /* Make the new connection a child of the unixShmNode */
4507 p
->pShmNode
= pShmNode
;
4509 p
->id
= pShmNode
->nextShmId
++;
4515 /* The reference count on pShmNode has already been incremented under
4516 ** the cover of the unixEnterMutex() mutex and the pointer from the
4517 ** new (struct unixShm) object to the pShmNode has been set. All that is
4518 ** left to do is to link the new object into the linked list starting
4519 ** at pShmNode->pFirst. This must be done while holding the pShmNode->mutex
4522 sqlite3_mutex_enter(pShmNode
->mutex
);
4523 p
->pNext
= pShmNode
->pFirst
;
4524 pShmNode
->pFirst
= p
;
4525 sqlite3_mutex_leave(pShmNode
->mutex
);
4528 /* Jump here on any error */
4530 unixShmPurge(pDbFd
); /* This call frees pShmNode if required */
4537 ** This function is called to obtain a pointer to region iRegion of the
4538 ** shared-memory associated with the database file fd. Shared-memory regions
4539 ** are numbered starting from zero. Each shared-memory region is szRegion
4542 ** If an error occurs, an error code is returned and *pp is set to NULL.
4544 ** Otherwise, if the bExtend parameter is 0 and the requested shared-memory
4545 ** region has not been allocated (by any client, including one running in a
4546 ** separate process), then *pp is set to NULL and SQLITE_OK returned. If
4547 ** bExtend is non-zero and the requested shared-memory region has not yet
4548 ** been allocated, it is allocated by this function.
4550 ** If the shared-memory region has already been allocated or is allocated by
4551 ** this call as described above, then it is mapped into this processes
4552 ** address space (if it is not already), *pp is set to point to the mapped
4553 ** memory and SQLITE_OK returned.
4555 static int unixShmMap(
4556 sqlite3_file
*fd
, /* Handle open on database file */
4557 int iRegion
, /* Region to retrieve */
4558 int szRegion
, /* Size of regions */
4559 int bExtend
, /* True to extend file if necessary */
4560 void volatile **pp
/* OUT: Mapped memory */
4562 unixFile
*pDbFd
= (unixFile
*)fd
;
4564 unixShmNode
*pShmNode
;
4566 int nShmPerMap
= unixShmRegionPerMap();
4569 /* If the shared-memory file has not yet been opened, open it now. */
4570 if( pDbFd
->pShm
==0 ){
4571 rc
= unixOpenSharedMemory(pDbFd
);
4572 if( rc
!=SQLITE_OK
) return rc
;
4576 pShmNode
= p
->pShmNode
;
4577 sqlite3_mutex_enter(pShmNode
->mutex
);
4578 if( pShmNode
->isUnlocked
){
4579 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4580 if( rc
!=SQLITE_OK
) goto shmpage_out
;
4581 pShmNode
->isUnlocked
= 0;
4583 assert( szRegion
==pShmNode
->szRegion
|| pShmNode
->nRegion
==0 );
4584 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4585 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4586 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4588 /* Minimum number of regions required to be mapped. */
4589 nReqRegion
= ((iRegion
+nShmPerMap
) / nShmPerMap
) * nShmPerMap
;
4591 if( pShmNode
->nRegion
<nReqRegion
){
4592 char **apNew
; /* New apRegion[] array */
4593 int nByte
= nReqRegion
*szRegion
; /* Minimum required file size */
4594 struct stat sStat
; /* Used by fstat() */
4596 pShmNode
->szRegion
= szRegion
;
4598 if( pShmNode
->h
>=0 ){
4599 /* The requested region is not mapped into this processes address space.
4600 ** Check to see if it has been allocated (i.e. if the wal-index file is
4601 ** large enough to contain the requested region).
4603 if( osFstat(pShmNode
->h
, &sStat
) ){
4604 rc
= SQLITE_IOERR_SHMSIZE
;
4608 if( sStat
.st_size
<nByte
){
4609 /* The requested memory region does not exist. If bExtend is set to
4610 ** false, exit early. *pp will be set to NULL and SQLITE_OK returned.
4616 /* Alternatively, if bExtend is true, extend the file. Do this by
4617 ** writing a single byte to the end of each (OS) page being
4618 ** allocated or extended. Technically, we need only write to the
4619 ** last page in order to extend the file. But writing to all new
4620 ** pages forces the OS to allocate them immediately, which reduces
4621 ** the chances of SIGBUS while accessing the mapped region later on.
4624 static const int pgsz
= 4096;
4627 /* Write to the last byte of each newly allocated or extended page */
4628 assert( (nByte
% pgsz
)==0 );
4629 for(iPg
=(sStat
.st_size
/pgsz
); iPg
<(nByte
/pgsz
); iPg
++){
4631 if( seekAndWriteFd(pShmNode
->h
, iPg
*pgsz
+ pgsz
-1, "", 1, &x
)!=1 ){
4632 const char *zFile
= pShmNode
->zFilename
;
4633 rc
= unixLogError(SQLITE_IOERR_SHMSIZE
, "write", zFile
);
4641 /* Map the requested memory region into this processes address space. */
4642 apNew
= (char **)sqlite3_realloc(
4643 pShmNode
->apRegion
, nReqRegion
*sizeof(char *)
4646 rc
= SQLITE_IOERR_NOMEM_BKPT
;
4649 pShmNode
->apRegion
= apNew
;
4650 while( pShmNode
->nRegion
<nReqRegion
){
4651 int nMap
= szRegion
*nShmPerMap
;
4654 if( pShmNode
->h
>=0 ){
4655 pMem
= osMmap(0, nMap
,
4656 pShmNode
->isReadonly
? PROT_READ
: PROT_READ
|PROT_WRITE
,
4657 MAP_SHARED
, pShmNode
->h
, szRegion
*(i64
)pShmNode
->nRegion
4659 if( pMem
==MAP_FAILED
){
4660 rc
= unixLogError(SQLITE_IOERR_SHMMAP
, "mmap", pShmNode
->zFilename
);
4664 pMem
= sqlite3_malloc64(szRegion
);
4666 rc
= SQLITE_NOMEM_BKPT
;
4669 memset(pMem
, 0, szRegion
);
4672 for(i
=0; i
<nShmPerMap
; i
++){
4673 pShmNode
->apRegion
[pShmNode
->nRegion
+i
] = &((char*)pMem
)[szRegion
*i
];
4675 pShmNode
->nRegion
+= nShmPerMap
;
4680 if( pShmNode
->nRegion
>iRegion
){
4681 *pp
= pShmNode
->apRegion
[iRegion
];
4685 if( pShmNode
->isReadonly
&& rc
==SQLITE_OK
) rc
= SQLITE_READONLY
;
4686 sqlite3_mutex_leave(pShmNode
->mutex
);
4691 ** Change the lock state for a shared-memory segment.
4693 ** Note that the relationship between SHAREd and EXCLUSIVE locks is a little
4694 ** different here than in posix. In xShmLock(), one can go from unlocked
4695 ** to shared and back or from unlocked to exclusive and back. But one may
4696 ** not go from shared to exclusive or from exclusive to shared.
4698 static int unixShmLock(
4699 sqlite3_file
*fd
, /* Database file holding the shared memory */
4700 int ofst
, /* First lock to acquire or release */
4701 int n
, /* Number of locks to acquire or release */
4702 int flags
/* What to do with the lock */
4704 unixFile
*pDbFd
= (unixFile
*)fd
; /* Connection holding shared memory */
4705 unixShm
*p
= pDbFd
->pShm
; /* The shared memory being locked */
4706 unixShm
*pX
; /* For looping over all siblings */
4707 unixShmNode
*pShmNode
= p
->pShmNode
; /* The underlying file iNode */
4708 int rc
= SQLITE_OK
; /* Result code */
4709 u16 mask
; /* Mask of locks to take or release */
4711 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4712 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4713 assert( ofst
>=0 && ofst
+n
<=SQLITE_SHM_NLOCK
);
4715 assert( flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_SHARED
)
4716 || flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_EXCLUSIVE
)
4717 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_SHARED
)
4718 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_EXCLUSIVE
) );
4719 assert( n
==1 || (flags
& SQLITE_SHM_EXCLUSIVE
)!=0 );
4720 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4721 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4723 mask
= (1<<(ofst
+n
)) - (1<<ofst
);
4724 assert( n
>1 || mask
==(1<<ofst
) );
4725 sqlite3_mutex_enter(pShmNode
->mutex
);
4726 if( flags
& SQLITE_SHM_UNLOCK
){
4727 u16 allMask
= 0; /* Mask of locks held by siblings */
4729 /* See if any siblings hold this same lock */
4730 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4731 if( pX
==p
) continue;
4732 assert( (pX
->exclMask
& (p
->exclMask
|p
->sharedMask
))==0 );
4733 allMask
|= pX
->sharedMask
;
4736 /* Unlock the system-level locks */
4737 if( (mask
& allMask
)==0 ){
4738 rc
= unixShmSystemLock(pDbFd
, F_UNLCK
, ofst
+UNIX_SHM_BASE
, n
);
4743 /* Undo the local locks */
4744 if( rc
==SQLITE_OK
){
4745 p
->exclMask
&= ~mask
;
4746 p
->sharedMask
&= ~mask
;
4748 }else if( flags
& SQLITE_SHM_SHARED
){
4749 u16 allShared
= 0; /* Union of locks held by connections other than "p" */
4751 /* Find out which shared locks are already held by sibling connections.
4752 ** If any sibling already holds an exclusive lock, go ahead and return
4755 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4756 if( (pX
->exclMask
& mask
)!=0 ){
4760 allShared
|= pX
->sharedMask
;
4763 /* Get shared locks at the system level, if necessary */
4764 if( rc
==SQLITE_OK
){
4765 if( (allShared
& mask
)==0 ){
4766 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, ofst
+UNIX_SHM_BASE
, n
);
4772 /* Get the local shared locks */
4773 if( rc
==SQLITE_OK
){
4774 p
->sharedMask
|= mask
;
4777 /* Make sure no sibling connections hold locks that will block this
4778 ** lock. If any do, return SQLITE_BUSY right away.
4780 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4781 if( (pX
->exclMask
& mask
)!=0 || (pX
->sharedMask
& mask
)!=0 ){
4787 /* Get the exclusive locks at the system level. Then if successful
4788 ** also mark the local connection as being locked.
4790 if( rc
==SQLITE_OK
){
4791 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, ofst
+UNIX_SHM_BASE
, n
);
4792 if( rc
==SQLITE_OK
){
4793 assert( (p
->sharedMask
& mask
)==0 );
4794 p
->exclMask
|= mask
;
4798 sqlite3_mutex_leave(pShmNode
->mutex
);
4799 OSTRACE(("SHM-LOCK shmid-%d, pid-%d got %03x,%03x\n",
4800 p
->id
, osGetpid(0), p
->sharedMask
, p
->exclMask
));
4805 ** Implement a memory barrier or memory fence on shared memory.
4807 ** All loads and stores begun before the barrier must complete before
4808 ** any load or store begun after the barrier.
4810 static void unixShmBarrier(
4811 sqlite3_file
*fd
/* Database file holding the shared memory */
4813 UNUSED_PARAMETER(fd
);
4814 sqlite3MemoryBarrier(); /* compiler-defined memory barrier */
4815 unixEnterMutex(); /* Also mutex, for redundancy */
4820 ** Close a connection to shared-memory. Delete the underlying
4821 ** storage if deleteFlag is true.
4823 ** If there is no shared memory associated with the connection then this
4824 ** routine is a harmless no-op.
4826 static int unixShmUnmap(
4827 sqlite3_file
*fd
, /* The underlying database file */
4828 int deleteFlag
/* Delete shared-memory if true */
4830 unixShm
*p
; /* The connection to be closed */
4831 unixShmNode
*pShmNode
; /* The underlying shared-memory file */
4832 unixShm
**pp
; /* For looping over sibling connections */
4833 unixFile
*pDbFd
; /* The underlying database file */
4835 pDbFd
= (unixFile
*)fd
;
4837 if( p
==0 ) return SQLITE_OK
;
4838 pShmNode
= p
->pShmNode
;
4840 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4841 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4843 /* Remove connection p from the set of connections associated
4845 sqlite3_mutex_enter(pShmNode
->mutex
);
4846 for(pp
=&pShmNode
->pFirst
; (*pp
)!=p
; pp
= &(*pp
)->pNext
){}
4849 /* Free the connection p */
4852 sqlite3_mutex_leave(pShmNode
->mutex
);
4854 /* If pShmNode->nRef has reached 0, then close the underlying
4855 ** shared-memory file, too */
4857 assert( pShmNode
->nRef
>0 );
4859 if( pShmNode
->nRef
==0 ){
4860 if( deleteFlag
&& pShmNode
->h
>=0 ){
4861 osUnlink(pShmNode
->zFilename
);
4863 unixShmPurge(pDbFd
);
4872 # define unixShmMap 0
4873 # define unixShmLock 0
4874 # define unixShmBarrier 0
4875 # define unixShmUnmap 0
4876 #endif /* #ifndef SQLITE_OMIT_WAL */
4878 #if SQLITE_MAX_MMAP_SIZE>0
4880 ** If it is currently memory mapped, unmap file pFd.
4882 static void unixUnmapfile(unixFile
*pFd
){
4883 assert( pFd
->nFetchOut
==0 );
4884 if( pFd
->pMapRegion
){
4885 osMunmap(pFd
->pMapRegion
, pFd
->mmapSizeActual
);
4886 pFd
->pMapRegion
= 0;
4888 pFd
->mmapSizeActual
= 0;
4893 ** Attempt to set the size of the memory mapping maintained by file
4894 ** descriptor pFd to nNew bytes. Any existing mapping is discarded.
4896 ** If successful, this function sets the following variables:
4898 ** unixFile.pMapRegion
4899 ** unixFile.mmapSize
4900 ** unixFile.mmapSizeActual
4902 ** If unsuccessful, an error message is logged via sqlite3_log() and
4903 ** the three variables above are zeroed. In this case SQLite should
4904 ** continue accessing the database using the xRead() and xWrite()
4907 static void unixRemapfile(
4908 unixFile
*pFd
, /* File descriptor object */
4909 i64 nNew
/* Required mapping size */
4911 const char *zErr
= "mmap";
4912 int h
= pFd
->h
; /* File descriptor open on db file */
4913 u8
*pOrig
= (u8
*)pFd
->pMapRegion
; /* Pointer to current file mapping */
4914 i64 nOrig
= pFd
->mmapSizeActual
; /* Size of pOrig region in bytes */
4915 u8
*pNew
= 0; /* Location of new mapping */
4916 int flags
= PROT_READ
; /* Flags to pass to mmap() */
4918 assert( pFd
->nFetchOut
==0 );
4919 assert( nNew
>pFd
->mmapSize
);
4920 assert( nNew
<=pFd
->mmapSizeMax
);
4922 assert( pFd
->mmapSizeActual
>=pFd
->mmapSize
);
4923 assert( MAP_FAILED
!=0 );
4925 #ifdef SQLITE_MMAP_READWRITE
4926 if( (pFd
->ctrlFlags
& UNIXFILE_RDONLY
)==0 ) flags
|= PROT_WRITE
;
4931 i64 nReuse
= pFd
->mmapSize
;
4933 const int szSyspage
= osGetpagesize();
4934 i64 nReuse
= (pFd
->mmapSize
& ~(szSyspage
-1));
4936 u8
*pReq
= &pOrig
[nReuse
];
4938 /* Unmap any pages of the existing mapping that cannot be reused. */
4939 if( nReuse
!=nOrig
){
4940 osMunmap(pReq
, nOrig
-nReuse
);
4944 pNew
= osMremap(pOrig
, nReuse
, nNew
, MREMAP_MAYMOVE
);
4947 pNew
= osMmap(pReq
, nNew
-nReuse
, flags
, MAP_SHARED
, h
, nReuse
);
4948 if( pNew
!=MAP_FAILED
){
4950 osMunmap(pNew
, nNew
- nReuse
);
4958 /* The attempt to extend the existing mapping failed. Free it. */
4959 if( pNew
==MAP_FAILED
|| pNew
==0 ){
4960 osMunmap(pOrig
, nReuse
);
4964 /* If pNew is still NULL, try to create an entirely new mapping. */
4966 pNew
= osMmap(0, nNew
, flags
, MAP_SHARED
, h
, 0);
4969 if( pNew
==MAP_FAILED
){
4972 unixLogError(SQLITE_OK
, zErr
, pFd
->zPath
);
4974 /* If the mmap() above failed, assume that all subsequent mmap() calls
4975 ** will probably fail too. Fall back to using xRead/xWrite exclusively
4977 pFd
->mmapSizeMax
= 0;
4979 pFd
->pMapRegion
= (void *)pNew
;
4980 pFd
->mmapSize
= pFd
->mmapSizeActual
= nNew
;
4984 ** Memory map or remap the file opened by file-descriptor pFd (if the file
4985 ** is already mapped, the existing mapping is replaced by the new). Or, if
4986 ** there already exists a mapping for this file, and there are still
4987 ** outstanding xFetch() references to it, this function is a no-op.
4989 ** If parameter nByte is non-negative, then it is the requested size of
4990 ** the mapping to create. Otherwise, if nByte is less than zero, then the
4991 ** requested size is the size of the file on disk. The actual size of the
4992 ** created mapping is either the requested size or the value configured
4993 ** using SQLITE_FCNTL_MMAP_LIMIT, whichever is smaller.
4995 ** SQLITE_OK is returned if no error occurs (even if the mapping is not
4996 ** recreated as a result of outstanding references) or an SQLite error
4999 static int unixMapfile(unixFile
*pFd
, i64 nMap
){
5000 assert( nMap
>=0 || pFd
->nFetchOut
==0 );
5001 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
5002 if( pFd
->nFetchOut
>0 ) return SQLITE_OK
;
5005 struct stat statbuf
; /* Low-level file information */
5006 if( osFstat(pFd
->h
, &statbuf
) ){
5007 return SQLITE_IOERR_FSTAT
;
5009 nMap
= statbuf
.st_size
;
5011 if( nMap
>pFd
->mmapSizeMax
){
5012 nMap
= pFd
->mmapSizeMax
;
5015 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
5016 if( nMap
!=pFd
->mmapSize
){
5017 unixRemapfile(pFd
, nMap
);
5022 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
5025 ** If possible, return a pointer to a mapping of file fd starting at offset
5026 ** iOff. The mapping must be valid for at least nAmt bytes.
5028 ** If such a pointer can be obtained, store it in *pp and return SQLITE_OK.
5029 ** Or, if one cannot but no error occurs, set *pp to 0 and return SQLITE_OK.
5030 ** Finally, if an error does occur, return an SQLite error code. The final
5031 ** value of *pp is undefined in this case.
5033 ** If this function does return a pointer, the caller must eventually
5034 ** release the reference by calling unixUnfetch().
5036 static int unixFetch(sqlite3_file
*fd
, i64 iOff
, int nAmt
, void **pp
){
5037 #if SQLITE_MAX_MMAP_SIZE>0
5038 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
5042 #if SQLITE_MAX_MMAP_SIZE>0
5043 if( pFd
->mmapSizeMax
>0 ){
5044 if( pFd
->pMapRegion
==0 ){
5045 int rc
= unixMapfile(pFd
, -1);
5046 if( rc
!=SQLITE_OK
) return rc
;
5048 if( pFd
->mmapSize
>= iOff
+nAmt
){
5049 *pp
= &((u8
*)pFd
->pMapRegion
)[iOff
];
5058 ** If the third argument is non-NULL, then this function releases a
5059 ** reference obtained by an earlier call to unixFetch(). The second
5060 ** argument passed to this function must be the same as the corresponding
5061 ** argument that was passed to the unixFetch() invocation.
5063 ** Or, if the third argument is NULL, then this function is being called
5064 ** to inform the VFS layer that, according to POSIX, any existing mapping
5065 ** may now be invalid and should be unmapped.
5067 static int unixUnfetch(sqlite3_file
*fd
, i64 iOff
, void *p
){
5068 #if SQLITE_MAX_MMAP_SIZE>0
5069 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
5070 UNUSED_PARAMETER(iOff
);
5072 /* If p==0 (unmap the entire file) then there must be no outstanding
5073 ** xFetch references. Or, if p!=0 (meaning it is an xFetch reference),
5074 ** then there must be at least one outstanding. */
5075 assert( (p
==0)==(pFd
->nFetchOut
==0) );
5077 /* If p!=0, it must match the iOff value. */
5078 assert( p
==0 || p
==&((u8
*)pFd
->pMapRegion
)[iOff
] );
5086 assert( pFd
->nFetchOut
>=0 );
5088 UNUSED_PARAMETER(fd
);
5089 UNUSED_PARAMETER(p
);
5090 UNUSED_PARAMETER(iOff
);
5096 ** Here ends the implementation of all sqlite3_file methods.
5098 ********************** End sqlite3_file Methods *******************************
5099 ******************************************************************************/
5102 ** This division contains definitions of sqlite3_io_methods objects that
5103 ** implement various file locking strategies. It also contains definitions
5104 ** of "finder" functions. A finder-function is used to locate the appropriate
5105 ** sqlite3_io_methods object for a particular database file. The pAppData
5106 ** field of the sqlite3_vfs VFS objects are initialized to be pointers to
5107 ** the correct finder-function for that VFS.
5109 ** Most finder functions return a pointer to a fixed sqlite3_io_methods
5110 ** object. The only interesting finder-function is autolockIoFinder, which
5111 ** looks at the filesystem type and tries to guess the best locking
5112 ** strategy from that.
5114 ** For finder-function F, two objects are created:
5116 ** (1) The real finder-function named "FImpt()".
5118 ** (2) A constant pointer to this function named just "F".
5121 ** A pointer to the F pointer is used as the pAppData value for VFS
5122 ** objects. We have to do this instead of letting pAppData point
5123 ** directly at the finder-function since C90 rules prevent a void*
5124 ** from be cast into a function pointer.
5127 ** Each instance of this macro generates two objects:
5129 ** * A constant sqlite3_io_methods object call METHOD that has locking
5130 ** methods CLOSE, LOCK, UNLOCK, CKRESLOCK.
5132 ** * An I/O method finder function called FINDER that returns a pointer
5133 ** to the METHOD object in the previous bullet.
5135 #define IOMETHODS(FINDER,METHOD,VERSION,CLOSE,LOCK,UNLOCK,CKLOCK,SHMMAP) \
5136 static const sqlite3_io_methods METHOD = { \
5137 VERSION, /* iVersion */ \
5138 CLOSE, /* xClose */ \
5139 unixRead, /* xRead */ \
5140 unixWrite, /* xWrite */ \
5141 unixTruncate, /* xTruncate */ \
5142 unixSync, /* xSync */ \
5143 unixFileSize, /* xFileSize */ \
5145 UNLOCK, /* xUnlock */ \
5146 CKLOCK, /* xCheckReservedLock */ \
5147 unixFileControl, /* xFileControl */ \
5148 unixSectorSize, /* xSectorSize */ \
5149 unixDeviceCharacteristics, /* xDeviceCapabilities */ \
5150 SHMMAP, /* xShmMap */ \
5151 unixShmLock, /* xShmLock */ \
5152 unixShmBarrier, /* xShmBarrier */ \
5153 unixShmUnmap, /* xShmUnmap */ \
5154 unixFetch, /* xFetch */ \
5155 unixUnfetch, /* xUnfetch */ \
5157 static const sqlite3_io_methods *FINDER##Impl(const char *z, unixFile *p){ \
5158 UNUSED_PARAMETER(z); UNUSED_PARAMETER(p); \
5161 static const sqlite3_io_methods *(*const FINDER)(const char*,unixFile *p) \
5165 ** Here are all of the sqlite3_io_methods objects for each of the
5166 ** locking strategies. Functions that return pointers to these methods
5167 ** are also created.
5170 posixIoFinder
, /* Finder function name */
5171 posixIoMethods
, /* sqlite3_io_methods object name */
5172 3, /* shared memory and mmap are enabled */
5173 unixClose
, /* xClose method */
5174 unixLock
, /* xLock method */
5175 unixUnlock
, /* xUnlock method */
5176 unixCheckReservedLock
, /* xCheckReservedLock method */
5177 unixShmMap
/* xShmMap method */
5180 nolockIoFinder
, /* Finder function name */
5181 nolockIoMethods
, /* sqlite3_io_methods object name */
5182 3, /* shared memory is disabled */
5183 nolockClose
, /* xClose method */
5184 nolockLock
, /* xLock method */
5185 nolockUnlock
, /* xUnlock method */
5186 nolockCheckReservedLock
, /* xCheckReservedLock method */
5187 0 /* xShmMap method */
5190 dotlockIoFinder
, /* Finder function name */
5191 dotlockIoMethods
, /* sqlite3_io_methods object name */
5192 1, /* shared memory is disabled */
5193 dotlockClose
, /* xClose method */
5194 dotlockLock
, /* xLock method */
5195 dotlockUnlock
, /* xUnlock method */
5196 dotlockCheckReservedLock
, /* xCheckReservedLock method */
5197 0 /* xShmMap method */
5200 #if SQLITE_ENABLE_LOCKING_STYLE
5202 flockIoFinder
, /* Finder function name */
5203 flockIoMethods
, /* sqlite3_io_methods object name */
5204 1, /* shared memory is disabled */
5205 flockClose
, /* xClose method */
5206 flockLock
, /* xLock method */
5207 flockUnlock
, /* xUnlock method */
5208 flockCheckReservedLock
, /* xCheckReservedLock method */
5209 0 /* xShmMap method */
5215 semIoFinder
, /* Finder function name */
5216 semIoMethods
, /* sqlite3_io_methods object name */
5217 1, /* shared memory is disabled */
5218 semXClose
, /* xClose method */
5219 semXLock
, /* xLock method */
5220 semXUnlock
, /* xUnlock method */
5221 semXCheckReservedLock
, /* xCheckReservedLock method */
5222 0 /* xShmMap method */
5226 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5228 afpIoFinder
, /* Finder function name */
5229 afpIoMethods
, /* sqlite3_io_methods object name */
5230 1, /* shared memory is disabled */
5231 afpClose
, /* xClose method */
5232 afpLock
, /* xLock method */
5233 afpUnlock
, /* xUnlock method */
5234 afpCheckReservedLock
, /* xCheckReservedLock method */
5235 0 /* xShmMap method */
5240 ** The proxy locking method is a "super-method" in the sense that it
5241 ** opens secondary file descriptors for the conch and lock files and
5242 ** it uses proxy, dot-file, AFP, and flock() locking methods on those
5243 ** secondary files. For this reason, the division that implements
5244 ** proxy locking is located much further down in the file. But we need
5245 ** to go ahead and define the sqlite3_io_methods and finder function
5246 ** for proxy locking here. So we forward declare the I/O methods.
5248 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5249 static int proxyClose(sqlite3_file
*);
5250 static int proxyLock(sqlite3_file
*, int);
5251 static int proxyUnlock(sqlite3_file
*, int);
5252 static int proxyCheckReservedLock(sqlite3_file
*, int*);
5254 proxyIoFinder
, /* Finder function name */
5255 proxyIoMethods
, /* sqlite3_io_methods object name */
5256 1, /* shared memory is disabled */
5257 proxyClose
, /* xClose method */
5258 proxyLock
, /* xLock method */
5259 proxyUnlock
, /* xUnlock method */
5260 proxyCheckReservedLock
, /* xCheckReservedLock method */
5261 0 /* xShmMap method */
5265 /* nfs lockd on OSX 10.3+ doesn't clear write locks when a read lock is set */
5266 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5268 nfsIoFinder
, /* Finder function name */
5269 nfsIoMethods
, /* sqlite3_io_methods object name */
5270 1, /* shared memory is disabled */
5271 unixClose
, /* xClose method */
5272 unixLock
, /* xLock method */
5273 nfsUnlock
, /* xUnlock method */
5274 unixCheckReservedLock
, /* xCheckReservedLock method */
5275 0 /* xShmMap method */
5279 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5281 ** This "finder" function attempts to determine the best locking strategy
5282 ** for the database file "filePath". It then returns the sqlite3_io_methods
5283 ** object that implements that strategy.
5285 ** This is for MacOSX only.
5287 static const sqlite3_io_methods
*autolockIoFinderImpl(
5288 const char *filePath
, /* name of the database file */
5289 unixFile
*pNew
/* open file object for the database file */
5291 static const struct Mapping
{
5292 const char *zFilesystem
; /* Filesystem type name */
5293 const sqlite3_io_methods
*pMethods
; /* Appropriate locking method */
5295 { "hfs", &posixIoMethods
},
5296 { "ufs", &posixIoMethods
},
5297 { "afpfs", &afpIoMethods
},
5298 { "smbfs", &afpIoMethods
},
5299 { "webdav", &nolockIoMethods
},
5303 struct statfs fsInfo
;
5304 struct flock lockInfo
;
5307 /* If filePath==NULL that means we are dealing with a transient file
5308 ** that does not need to be locked. */
5309 return &nolockIoMethods
;
5311 if( statfs(filePath
, &fsInfo
) != -1 ){
5312 if( fsInfo
.f_flags
& MNT_RDONLY
){
5313 return &nolockIoMethods
;
5315 for(i
=0; aMap
[i
].zFilesystem
; i
++){
5316 if( strcmp(fsInfo
.f_fstypename
, aMap
[i
].zFilesystem
)==0 ){
5317 return aMap
[i
].pMethods
;
5322 /* Default case. Handles, amongst others, "nfs".
5323 ** Test byte-range lock using fcntl(). If the call succeeds,
5324 ** assume that the file-system supports POSIX style locks.
5327 lockInfo
.l_start
= 0;
5328 lockInfo
.l_whence
= SEEK_SET
;
5329 lockInfo
.l_type
= F_RDLCK
;
5330 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5331 if( strcmp(fsInfo
.f_fstypename
, "nfs")==0 ){
5332 return &nfsIoMethods
;
5334 return &posixIoMethods
;
5337 return &dotlockIoMethods
;
5340 static const sqlite3_io_methods
5341 *(*const autolockIoFinder
)(const char*,unixFile
*) = autolockIoFinderImpl
;
5343 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
5347 ** This "finder" function for VxWorks checks to see if posix advisory
5348 ** locking works. If it does, then that is what is used. If it does not
5349 ** work, then fallback to named semaphore locking.
5351 static const sqlite3_io_methods
*vxworksIoFinderImpl(
5352 const char *filePath
, /* name of the database file */
5353 unixFile
*pNew
/* the open file object */
5355 struct flock lockInfo
;
5358 /* If filePath==NULL that means we are dealing with a transient file
5359 ** that does not need to be locked. */
5360 return &nolockIoMethods
;
5363 /* Test if fcntl() is supported and use POSIX style locks.
5364 ** Otherwise fall back to the named semaphore method.
5367 lockInfo
.l_start
= 0;
5368 lockInfo
.l_whence
= SEEK_SET
;
5369 lockInfo
.l_type
= F_RDLCK
;
5370 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5371 return &posixIoMethods
;
5373 return &semIoMethods
;
5376 static const sqlite3_io_methods
5377 *(*const vxworksIoFinder
)(const char*,unixFile
*) = vxworksIoFinderImpl
;
5379 #endif /* OS_VXWORKS */
5382 ** An abstract type for a pointer to an IO method finder function:
5384 typedef const sqlite3_io_methods
*(*finder_type
)(const char*,unixFile
*);
5387 /****************************************************************************
5388 **************************** sqlite3_vfs methods ****************************
5390 ** This division contains the implementation of methods on the
5391 ** sqlite3_vfs object.
5395 ** Initialize the contents of the unixFile structure pointed to by pId.
5397 static int fillInUnixFile(
5398 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
5399 int h
, /* Open file descriptor of file being opened */
5400 sqlite3_file
*pId
, /* Write to the unixFile structure here */
5401 const char *zFilename
, /* Name of the file being opened */
5402 int ctrlFlags
/* Zero or more UNIXFILE_* values */
5404 const sqlite3_io_methods
*pLockingStyle
;
5405 unixFile
*pNew
= (unixFile
*)pId
;
5408 assert( pNew
->pInode
==NULL
);
5410 /* No locking occurs in temporary files */
5411 assert( zFilename
!=0 || (ctrlFlags
& UNIXFILE_NOLOCK
)!=0 );
5413 OSTRACE(("OPEN %-3d %s\n", h
, zFilename
));
5416 pNew
->zPath
= zFilename
;
5417 pNew
->ctrlFlags
= (u8
)ctrlFlags
;
5418 #if SQLITE_MAX_MMAP_SIZE>0
5419 pNew
->mmapSizeMax
= sqlite3GlobalConfig
.szMmap
;
5421 if( sqlite3_uri_boolean(((ctrlFlags
& UNIXFILE_URI
) ? zFilename
: 0),
5422 "psow", SQLITE_POWERSAFE_OVERWRITE
) ){
5423 pNew
->ctrlFlags
|= UNIXFILE_PSOW
;
5425 if( strcmp(pVfs
->zName
,"unix-excl")==0 ){
5426 pNew
->ctrlFlags
|= UNIXFILE_EXCL
;
5430 pNew
->pId
= vxworksFindFileId(zFilename
);
5432 ctrlFlags
|= UNIXFILE_NOLOCK
;
5433 rc
= SQLITE_NOMEM_BKPT
;
5437 if( ctrlFlags
& UNIXFILE_NOLOCK
){
5438 pLockingStyle
= &nolockIoMethods
;
5440 pLockingStyle
= (**(finder_type
*)pVfs
->pAppData
)(zFilename
, pNew
);
5441 #if SQLITE_ENABLE_LOCKING_STYLE
5442 /* Cache zFilename in the locking context (AFP and dotlock override) for
5443 ** proxyLock activation is possible (remote proxy is based on db name)
5444 ** zFilename remains valid until file is closed, to support */
5445 pNew
->lockingContext
= (void*)zFilename
;
5449 if( pLockingStyle
== &posixIoMethods
5450 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5451 || pLockingStyle
== &nfsIoMethods
5455 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5456 if( rc
!=SQLITE_OK
){
5457 /* If an error occurred in findInodeInfo(), close the file descriptor
5458 ** immediately, before releasing the mutex. findInodeInfo() may fail
5459 ** in two scenarios:
5461 ** (a) A call to fstat() failed.
5462 ** (b) A malloc failed.
5464 ** Scenario (b) may only occur if the process is holding no other
5465 ** file descriptors open on the same file. If there were other file
5466 ** descriptors on this file, then no malloc would be required by
5467 ** findInodeInfo(). If this is the case, it is quite safe to close
5468 ** handle h - as it is guaranteed that no posix locks will be released
5471 ** If scenario (a) caused the error then things are not so safe. The
5472 ** implicit assumption here is that if fstat() fails, things are in
5473 ** such bad shape that dropping a lock or two doesn't matter much.
5475 robust_close(pNew
, h
, __LINE__
);
5481 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5482 else if( pLockingStyle
== &afpIoMethods
){
5483 /* AFP locking uses the file path so it needs to be included in
5484 ** the afpLockingContext.
5486 afpLockingContext
*pCtx
;
5487 pNew
->lockingContext
= pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
5489 rc
= SQLITE_NOMEM_BKPT
;
5491 /* NB: zFilename exists and remains valid until the file is closed
5492 ** according to requirement F11141. So we do not need to make a
5493 ** copy of the filename. */
5494 pCtx
->dbPath
= zFilename
;
5498 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5499 if( rc
!=SQLITE_OK
){
5500 sqlite3_free(pNew
->lockingContext
);
5501 robust_close(pNew
, h
, __LINE__
);
5509 else if( pLockingStyle
== &dotlockIoMethods
){
5510 /* Dotfile locking uses the file path so it needs to be included in
5511 ** the dotlockLockingContext
5515 assert( zFilename
!=0 );
5516 nFilename
= (int)strlen(zFilename
) + 6;
5517 zLockFile
= (char *)sqlite3_malloc64(nFilename
);
5519 rc
= SQLITE_NOMEM_BKPT
;
5521 sqlite3_snprintf(nFilename
, zLockFile
, "%s" DOTLOCK_SUFFIX
, zFilename
);
5523 pNew
->lockingContext
= zLockFile
;
5527 else if( pLockingStyle
== &semIoMethods
){
5528 /* Named semaphore locking uses the file path so it needs to be
5529 ** included in the semLockingContext
5532 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5533 if( (rc
==SQLITE_OK
) && (pNew
->pInode
->pSem
==NULL
) ){
5534 char *zSemName
= pNew
->pInode
->aSemName
;
5536 sqlite3_snprintf(MAX_PATHNAME
, zSemName
, "/%s.sem",
5537 pNew
->pId
->zCanonicalName
);
5538 for( n
=1; zSemName
[n
]; n
++ )
5539 if( zSemName
[n
]=='/' ) zSemName
[n
] = '_';
5540 pNew
->pInode
->pSem
= sem_open(zSemName
, O_CREAT
, 0666, 1);
5541 if( pNew
->pInode
->pSem
== SEM_FAILED
){
5542 rc
= SQLITE_NOMEM_BKPT
;
5543 pNew
->pInode
->aSemName
[0] = '\0';
5550 storeLastErrno(pNew
, 0);
5552 if( rc
!=SQLITE_OK
){
5553 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5555 osUnlink(zFilename
);
5556 pNew
->ctrlFlags
|= UNIXFILE_DELETE
;
5559 if( rc
!=SQLITE_OK
){
5560 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5562 pNew
->pMethod
= pLockingStyle
;
5570 ** Return the name of a directory in which to put temporary files.
5571 ** If no suitable temporary file directory can be found, return NULL.
5573 static const char *unixTempFileDir(void){
5574 static const char *azDirs
[] = {
5584 const char *zDir
= sqlite3_temp_directory
;
5586 if( !azDirs
[0] ) azDirs
[0] = getenv("SQLITE_TMPDIR");
5587 if( !azDirs
[1] ) azDirs
[1] = getenv("TMPDIR");
5590 && osStat(zDir
, &buf
)==0
5591 && S_ISDIR(buf
.st_mode
)
5592 && osAccess(zDir
, 03)==0
5596 if( i
>=sizeof(azDirs
)/sizeof(azDirs
[0]) ) break;
5603 ** Create a temporary file name in zBuf. zBuf must be allocated
5604 ** by the calling process and must be big enough to hold at least
5605 ** pVfs->mxPathname bytes.
5607 static int unixGetTempname(int nBuf
, char *zBuf
){
5611 /* It's odd to simulate an io-error here, but really this is just
5612 ** using the io-error infrastructure to test that SQLite handles this
5613 ** function failing.
5616 SimulateIOError( return SQLITE_IOERR
);
5618 zDir
= unixTempFileDir();
5619 if( zDir
==0 ) return SQLITE_IOERR_GETTEMPPATH
;
5622 sqlite3_randomness(sizeof(r
), &r
);
5625 sqlite3_snprintf(nBuf
, zBuf
, "%s/"SQLITE_TEMP_FILE_PREFIX
"%llx%c",
5627 if( zBuf
[nBuf
-2]!=0 || (iLimit
++)>10 ) return SQLITE_ERROR
;
5628 }while( osAccess(zBuf
,0)==0 );
5632 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5634 ** Routine to transform a unixFile into a proxy-locking unixFile.
5635 ** Implementation in the proxy-lock division, but used by unixOpen()
5636 ** if SQLITE_PREFER_PROXY_LOCKING is defined.
5638 static int proxyTransformUnixFile(unixFile
*, const char*);
5642 ** Search for an unused file descriptor that was opened on the database
5643 ** file (not a journal or master-journal file) identified by pathname
5644 ** zPath with SQLITE_OPEN_XXX flags matching those passed as the second
5645 ** argument to this function.
5647 ** Such a file descriptor may exist if a database connection was closed
5648 ** but the associated file descriptor could not be closed because some
5649 ** other file descriptor open on the same file is holding a file-lock.
5650 ** Refer to comments in the unixClose() function and the lengthy comment
5651 ** describing "Posix Advisory Locking" at the start of this file for
5652 ** further details. Also, ticket #4018.
5654 ** If a suitable file descriptor is found, then it is returned. If no
5655 ** such file descriptor is located, -1 is returned.
5657 static UnixUnusedFd
*findReusableFd(const char *zPath
, int flags
){
5658 UnixUnusedFd
*pUnused
= 0;
5660 /* Do not search for an unused file descriptor on vxworks. Not because
5661 ** vxworks would not benefit from the change (it might, we're not sure),
5662 ** but because no way to test it is currently available. It is better
5663 ** not to risk breaking vxworks support for the sake of such an obscure
5666 struct stat sStat
; /* Results of stat() call */
5670 /* A stat() call may fail for various reasons. If this happens, it is
5671 ** almost certain that an open() call on the same path will also fail.
5672 ** For this reason, if an error occurs in the stat() call here, it is
5673 ** ignored and -1 is returned. The caller will try to open a new file
5674 ** descriptor on the same path, fail, and return an error to SQLite.
5676 ** Even if a subsequent open() call does succeed, the consequences of
5677 ** not searching for a reusable file descriptor are not dire. */
5678 if( nUnusedFd
>0 && 0==osStat(zPath
, &sStat
) ){
5679 unixInodeInfo
*pInode
;
5682 while( pInode
&& (pInode
->fileId
.dev
!=sStat
.st_dev
5683 || pInode
->fileId
.ino
!=(u64
)sStat
.st_ino
) ){
5684 pInode
= pInode
->pNext
;
5688 for(pp
=&pInode
->pUnused
; *pp
&& (*pp
)->flags
!=flags
; pp
=&((*pp
)->pNext
));
5692 *pp
= pUnused
->pNext
;
5697 #endif /* if !OS_VXWORKS */
5702 ** Find the mode, uid and gid of file zFile.
5704 static int getFileMode(
5705 const char *zFile
, /* File name */
5706 mode_t
*pMode
, /* OUT: Permissions of zFile */
5707 uid_t
*pUid
, /* OUT: uid of zFile. */
5708 gid_t
*pGid
/* OUT: gid of zFile. */
5710 struct stat sStat
; /* Output of stat() on database file */
5712 if( 0==osStat(zFile
, &sStat
) ){
5713 *pMode
= sStat
.st_mode
& 0777;
5714 *pUid
= sStat
.st_uid
;
5715 *pGid
= sStat
.st_gid
;
5717 rc
= SQLITE_IOERR_FSTAT
;
5723 ** This function is called by unixOpen() to determine the unix permissions
5724 ** to create new files with. If no error occurs, then SQLITE_OK is returned
5725 ** and a value suitable for passing as the third argument to open(2) is
5726 ** written to *pMode. If an IO error occurs, an SQLite error code is
5727 ** returned and the value of *pMode is not modified.
5729 ** In most cases, this routine sets *pMode to 0, which will become
5730 ** an indication to robust_open() to create the file using
5731 ** SQLITE_DEFAULT_FILE_PERMISSIONS adjusted by the umask.
5732 ** But if the file being opened is a WAL or regular journal file, then
5733 ** this function queries the file-system for the permissions on the
5734 ** corresponding database file and sets *pMode to this value. Whenever
5735 ** possible, WAL and journal files are created using the same permissions
5736 ** as the associated database file.
5738 ** If the SQLITE_ENABLE_8_3_NAMES option is enabled, then the
5739 ** original filename is unavailable. But 8_3_NAMES is only used for
5740 ** FAT filesystems and permissions do not matter there, so just use
5741 ** the default permissions.
5743 static int findCreateFileMode(
5744 const char *zPath
, /* Path of file (possibly) being created */
5745 int flags
, /* Flags passed as 4th argument to xOpen() */
5746 mode_t
*pMode
, /* OUT: Permissions to open file with */
5747 uid_t
*pUid
, /* OUT: uid to set on the file */
5748 gid_t
*pGid
/* OUT: gid to set on the file */
5750 int rc
= SQLITE_OK
; /* Return Code */
5754 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5755 char zDb
[MAX_PATHNAME
+1]; /* Database file path */
5756 int nDb
; /* Number of valid bytes in zDb */
5758 /* zPath is a path to a WAL or journal file. The following block derives
5759 ** the path to the associated database file from zPath. This block handles
5760 ** the following naming conventions:
5762 ** "<path to db>-journal"
5763 ** "<path to db>-wal"
5764 ** "<path to db>-journalNN"
5765 ** "<path to db>-walNN"
5767 ** where NN is a decimal number. The NN naming schemes are
5768 ** used by the test_multiplex.c module.
5770 nDb
= sqlite3Strlen30(zPath
) - 1;
5771 while( zPath
[nDb
]!='-' ){
5772 /* In normal operation, the journal file name will always contain
5773 ** a '-' character. However in 8+3 filename mode, or if a corrupt
5774 ** rollback journal specifies a master journal with a goofy name, then
5775 ** the '-' might be missing. */
5776 if( nDb
==0 || zPath
[nDb
]=='.' ) return SQLITE_OK
;
5779 memcpy(zDb
, zPath
, nDb
);
5782 rc
= getFileMode(zDb
, pMode
, pUid
, pGid
);
5783 }else if( flags
& SQLITE_OPEN_DELETEONCLOSE
){
5785 }else if( flags
& SQLITE_OPEN_URI
){
5786 /* If this is a main database file and the file was opened using a URI
5787 ** filename, check for the "modeof" parameter. If present, interpret
5788 ** its value as a filename and try to copy the mode, uid and gid from
5790 const char *z
= sqlite3_uri_parameter(zPath
, "modeof");
5792 rc
= getFileMode(z
, pMode
, pUid
, pGid
);
5799 ** Open the file zPath.
5801 ** Previously, the SQLite OS layer used three functions in place of this
5804 ** sqlite3OsOpenReadWrite();
5805 ** sqlite3OsOpenReadOnly();
5806 ** sqlite3OsOpenExclusive();
5808 ** These calls correspond to the following combinations of flags:
5810 ** ReadWrite() -> (READWRITE | CREATE)
5811 ** ReadOnly() -> (READONLY)
5812 ** OpenExclusive() -> (READWRITE | CREATE | EXCLUSIVE)
5814 ** The old OpenExclusive() accepted a boolean argument - "delFlag". If
5815 ** true, the file was configured to be automatically deleted when the
5816 ** file handle closed. To achieve the same effect using this new
5817 ** interface, add the DELETEONCLOSE flag to those specified above for
5820 static int unixOpen(
5821 sqlite3_vfs
*pVfs
, /* The VFS for which this is the xOpen method */
5822 const char *zPath
, /* Pathname of file to be opened */
5823 sqlite3_file
*pFile
, /* The file descriptor to be filled in */
5824 int flags
, /* Input flags to control the opening */
5825 int *pOutFlags
/* Output flags returned to SQLite core */
5827 unixFile
*p
= (unixFile
*)pFile
;
5828 int fd
= -1; /* File descriptor returned by open() */
5829 int openFlags
= 0; /* Flags to pass to open() */
5830 int eType
= flags
&0xFFFFFF00; /* Type of file to open */
5831 int noLock
; /* True to omit locking primitives */
5832 int rc
= SQLITE_OK
; /* Function Return Code */
5833 int ctrlFlags
= 0; /* UNIXFILE_* flags */
5835 int isExclusive
= (flags
& SQLITE_OPEN_EXCLUSIVE
);
5836 int isDelete
= (flags
& SQLITE_OPEN_DELETEONCLOSE
);
5837 int isCreate
= (flags
& SQLITE_OPEN_CREATE
);
5838 int isReadonly
= (flags
& SQLITE_OPEN_READONLY
);
5839 int isReadWrite
= (flags
& SQLITE_OPEN_READWRITE
);
5840 #if SQLITE_ENABLE_LOCKING_STYLE
5841 int isAutoProxy
= (flags
& SQLITE_OPEN_AUTOPROXY
);
5843 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
5844 struct statfs fsInfo
;
5847 /* If creating a master or main-file journal, this function will open
5848 ** a file-descriptor on the directory too. The first time unixSync()
5849 ** is called the directory file descriptor will be fsync()ed and close()d.
5851 int isNewJrnl
= (isCreate
&& (
5852 eType
==SQLITE_OPEN_MASTER_JOURNAL
5853 || eType
==SQLITE_OPEN_MAIN_JOURNAL
5854 || eType
==SQLITE_OPEN_WAL
5857 /* If argument zPath is a NULL pointer, this function is required to open
5858 ** a temporary file. Use this buffer to store the file name in.
5860 char zTmpname
[MAX_PATHNAME
+2];
5861 const char *zName
= zPath
;
5863 /* Check the following statements are true:
5865 ** (a) Exactly one of the READWRITE and READONLY flags must be set, and
5866 ** (b) if CREATE is set, then READWRITE must also be set, and
5867 ** (c) if EXCLUSIVE is set, then CREATE must also be set.
5868 ** (d) if DELETEONCLOSE is set, then CREATE must also be set.
5870 assert((isReadonly
==0 || isReadWrite
==0) && (isReadWrite
|| isReadonly
));
5871 assert(isCreate
==0 || isReadWrite
);
5872 assert(isExclusive
==0 || isCreate
);
5873 assert(isDelete
==0 || isCreate
);
5875 /* The main DB, main journal, WAL file and master journal are never
5876 ** automatically deleted. Nor are they ever temporary files. */
5877 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_DB
);
5878 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_JOURNAL
);
5879 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MASTER_JOURNAL
);
5880 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_WAL
);
5882 /* Assert that the upper layer has set one of the "file-type" flags. */
5883 assert( eType
==SQLITE_OPEN_MAIN_DB
|| eType
==SQLITE_OPEN_TEMP_DB
5884 || eType
==SQLITE_OPEN_MAIN_JOURNAL
|| eType
==SQLITE_OPEN_TEMP_JOURNAL
5885 || eType
==SQLITE_OPEN_SUBJOURNAL
|| eType
==SQLITE_OPEN_MASTER_JOURNAL
5886 || eType
==SQLITE_OPEN_TRANSIENT_DB
|| eType
==SQLITE_OPEN_WAL
5889 /* Detect a pid change and reset the PRNG. There is a race condition
5890 ** here such that two or more threads all trying to open databases at
5891 ** the same instant might all reset the PRNG. But multiple resets
5894 if( randomnessPid
!=osGetpid(0) ){
5895 randomnessPid
= osGetpid(0);
5896 sqlite3_randomness(0,0);
5898 memset(p
, 0, sizeof(unixFile
));
5900 if( eType
==SQLITE_OPEN_MAIN_DB
){
5901 UnixUnusedFd
*pUnused
;
5902 pUnused
= findReusableFd(zName
, flags
);
5906 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
5908 return SQLITE_NOMEM_BKPT
;
5911 p
->pPreallocatedUnused
= pUnused
;
5913 /* Database filenames are double-zero terminated if they are not
5914 ** URIs with parameters. Hence, they can always be passed into
5915 ** sqlite3_uri_parameter(). */
5916 assert( (flags
& SQLITE_OPEN_URI
) || zName
[strlen(zName
)+1]==0 );
5919 /* If zName is NULL, the upper layer is requesting a temp file. */
5920 assert(isDelete
&& !isNewJrnl
);
5921 rc
= unixGetTempname(pVfs
->mxPathname
, zTmpname
);
5922 if( rc
!=SQLITE_OK
){
5927 /* Generated temporary filenames are always double-zero terminated
5928 ** for use by sqlite3_uri_parameter(). */
5929 assert( zName
[strlen(zName
)+1]==0 );
5932 /* Determine the value of the flags parameter passed to POSIX function
5933 ** open(). These must be calculated even if open() is not called, as
5934 ** they may be stored as part of the file handle and used by the
5935 ** 'conch file' locking functions later on. */
5936 if( isReadonly
) openFlags
|= O_RDONLY
;
5937 if( isReadWrite
) openFlags
|= O_RDWR
;
5938 if( isCreate
) openFlags
|= O_CREAT
;
5939 if( isExclusive
) openFlags
|= (O_EXCL
|O_NOFOLLOW
);
5940 openFlags
|= (O_LARGEFILE
|O_BINARY
);
5943 mode_t openMode
; /* Permissions to create file with */
5944 uid_t uid
; /* Userid for the file */
5945 gid_t gid
; /* Groupid for the file */
5946 rc
= findCreateFileMode(zName
, flags
, &openMode
, &uid
, &gid
);
5947 if( rc
!=SQLITE_OK
){
5948 assert( !p
->pPreallocatedUnused
);
5949 assert( eType
==SQLITE_OPEN_WAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
);
5952 fd
= robust_open(zName
, openFlags
, openMode
);
5953 OSTRACE(("OPENX %-3d %s 0%o\n", fd
, zName
, openFlags
));
5954 assert( !isExclusive
|| (openFlags
& O_CREAT
)!=0 );
5956 if( isNewJrnl
&& errno
==EACCES
&& osAccess(zName
, F_OK
) ){
5957 /* If unable to create a journal because the directory is not
5958 ** writable, change the error code to indicate that. */
5959 rc
= SQLITE_READONLY_DIRECTORY
;
5960 }else if( errno
!=EISDIR
&& isReadWrite
){
5961 /* Failed to open the file for read/write access. Try read-only. */
5962 flags
&= ~(SQLITE_OPEN_READWRITE
|SQLITE_OPEN_CREATE
);
5963 openFlags
&= ~(O_RDWR
|O_CREAT
);
5964 flags
|= SQLITE_OPEN_READONLY
;
5965 openFlags
|= O_RDONLY
;
5967 fd
= robust_open(zName
, openFlags
, openMode
);
5971 int rc2
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zName
);
5972 if( rc
==SQLITE_OK
) rc
= rc2
;
5976 /* If this process is running as root and if creating a new rollback
5977 ** journal or WAL file, set the ownership of the journal or WAL to be
5978 ** the same as the original database.
5980 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5981 robustFchown(fd
, uid
, gid
);
5989 if( p
->pPreallocatedUnused
){
5990 p
->pPreallocatedUnused
->fd
= fd
;
5991 p
->pPreallocatedUnused
->flags
= flags
;
5997 #elif defined(SQLITE_UNLINK_AFTER_CLOSE)
5998 zPath
= sqlite3_mprintf("%s", zName
);
6000 robust_close(p
, fd
, __LINE__
);
6001 return SQLITE_NOMEM_BKPT
;
6007 #if SQLITE_ENABLE_LOCKING_STYLE
6009 p
->openFlags
= openFlags
;
6013 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
6014 if( fstatfs(fd
, &fsInfo
) == -1 ){
6015 storeLastErrno(p
, errno
);
6016 robust_close(p
, fd
, __LINE__
);
6017 return SQLITE_IOERR_ACCESS
;
6019 if (0 == strncmp("msdos", fsInfo
.f_fstypename
, 5)) {
6020 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
6022 if (0 == strncmp("exfat", fsInfo
.f_fstypename
, 5)) {
6023 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
6027 /* Set up appropriate ctrlFlags */
6028 if( isDelete
) ctrlFlags
|= UNIXFILE_DELETE
;
6029 if( isReadonly
) ctrlFlags
|= UNIXFILE_RDONLY
;
6030 noLock
= eType
!=SQLITE_OPEN_MAIN_DB
;
6031 if( noLock
) ctrlFlags
|= UNIXFILE_NOLOCK
;
6032 if( isNewJrnl
) ctrlFlags
|= UNIXFILE_DIRSYNC
;
6033 if( flags
& SQLITE_OPEN_URI
) ctrlFlags
|= UNIXFILE_URI
;
6035 #if SQLITE_ENABLE_LOCKING_STYLE
6036 #if SQLITE_PREFER_PROXY_LOCKING
6039 if( isAutoProxy
&& (zPath
!=NULL
) && (!noLock
) && pVfs
->xOpen
){
6040 char *envforce
= getenv("SQLITE_FORCE_PROXY_LOCKING");
6043 /* SQLITE_FORCE_PROXY_LOCKING==1 means force always use proxy, 0 means
6044 ** never use proxy, NULL means use proxy for non-local files only. */
6045 if( envforce
!=NULL
){
6046 useProxy
= atoi(envforce
)>0;
6048 useProxy
= !(fsInfo
.f_flags
&MNT_LOCAL
);
6051 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
6052 if( rc
==SQLITE_OK
){
6053 rc
= proxyTransformUnixFile((unixFile
*)pFile
, ":auto:");
6054 if( rc
!=SQLITE_OK
){
6055 /* Use unixClose to clean up the resources added in fillInUnixFile
6056 ** and clear all the structure's references. Specifically,
6057 ** pFile->pMethods will be NULL so sqlite3OsClose will be a no-op
6068 assert( zPath
==0 || zPath
[0]=='/'
6069 || eType
==SQLITE_OPEN_MASTER_JOURNAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
6071 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
6074 if( rc
!=SQLITE_OK
){
6075 sqlite3_free(p
->pPreallocatedUnused
);
6082 ** Delete the file at zPath. If the dirSync argument is true, fsync()
6083 ** the directory after deleting the file.
6085 static int unixDelete(
6086 sqlite3_vfs
*NotUsed
, /* VFS containing this as the xDelete method */
6087 const char *zPath
, /* Name of file to be deleted */
6088 int dirSync
/* If true, fsync() directory after deleting file */
6091 UNUSED_PARAMETER(NotUsed
);
6092 SimulateIOError(return SQLITE_IOERR_DELETE
);
6093 if( osUnlink(zPath
)==(-1) ){
6096 || osAccess(zPath
,0)!=0
6099 rc
= SQLITE_IOERR_DELETE_NOENT
;
6101 rc
= unixLogError(SQLITE_IOERR_DELETE
, "unlink", zPath
);
6105 #ifndef SQLITE_DISABLE_DIRSYNC
6106 if( (dirSync
& 1)!=0 ){
6108 rc
= osOpenDirectory(zPath
, &fd
);
6109 if( rc
==SQLITE_OK
){
6110 if( full_fsync(fd
,0,0) ){
6111 rc
= unixLogError(SQLITE_IOERR_DIR_FSYNC
, "fsync", zPath
);
6113 robust_close(0, fd
, __LINE__
);
6115 assert( rc
==SQLITE_CANTOPEN
);
6124 ** Test the existence of or access permissions of file zPath. The
6125 ** test performed depends on the value of flags:
6127 ** SQLITE_ACCESS_EXISTS: Return 1 if the file exists
6128 ** SQLITE_ACCESS_READWRITE: Return 1 if the file is read and writable.
6129 ** SQLITE_ACCESS_READONLY: Return 1 if the file is readable.
6131 ** Otherwise return 0.
6133 static int unixAccess(
6134 sqlite3_vfs
*NotUsed
, /* The VFS containing this xAccess method */
6135 const char *zPath
, /* Path of the file to examine */
6136 int flags
, /* What do we want to learn about the zPath file? */
6137 int *pResOut
/* Write result boolean here */
6139 UNUSED_PARAMETER(NotUsed
);
6140 SimulateIOError( return SQLITE_IOERR_ACCESS
; );
6141 assert( pResOut
!=0 );
6143 /* The spec says there are three possible values for flags. But only
6144 ** two of them are actually used */
6145 assert( flags
==SQLITE_ACCESS_EXISTS
|| flags
==SQLITE_ACCESS_READWRITE
);
6147 if( flags
==SQLITE_ACCESS_EXISTS
){
6149 *pResOut
= (0==osStat(zPath
, &buf
) && buf
.st_size
>0);
6151 *pResOut
= osAccess(zPath
, W_OK
|R_OK
)==0;
6159 static int mkFullPathname(
6160 const char *zPath
, /* Input path */
6161 char *zOut
, /* Output buffer */
6162 int nOut
/* Allocated size of buffer zOut */
6164 int nPath
= sqlite3Strlen30(zPath
);
6166 if( zPath
[0]!='/' ){
6167 if( osGetcwd(zOut
, nOut
-2)==0 ){
6168 return unixLogError(SQLITE_CANTOPEN_BKPT
, "getcwd", zPath
);
6170 iOff
= sqlite3Strlen30(zOut
);
6173 if( (iOff
+nPath
+1)>nOut
){
6174 /* SQLite assumes that xFullPathname() nul-terminates the output buffer
6175 ** even if it returns an error. */
6177 return SQLITE_CANTOPEN_BKPT
;
6179 sqlite3_snprintf(nOut
-iOff
, &zOut
[iOff
], "%s", zPath
);
6184 ** Turn a relative pathname into a full pathname. The relative path
6185 ** is stored as a nul-terminated string in the buffer pointed to by
6188 ** zOut points to a buffer of at least sqlite3_vfs.mxPathname bytes
6189 ** (in this case, MAX_PATHNAME bytes). The full-path is written to
6190 ** this buffer before returning.
6192 static int unixFullPathname(
6193 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
6194 const char *zPath
, /* Possibly relative input path */
6195 int nOut
, /* Size of output buffer in bytes */
6196 char *zOut
/* Output buffer */
6198 #if !defined(HAVE_READLINK) || !defined(HAVE_LSTAT)
6199 return mkFullPathname(zPath
, zOut
, nOut
);
6203 int nLink
= 1; /* Number of symbolic links followed so far */
6204 const char *zIn
= zPath
; /* Input path for each iteration of loop */
6207 assert( pVfs
->mxPathname
==MAX_PATHNAME
);
6208 UNUSED_PARAMETER(pVfs
);
6210 /* It's odd to simulate an io-error here, but really this is just
6211 ** using the io-error infrastructure to test that SQLite handles this
6212 ** function failing. This function could fail if, for example, the
6213 ** current working directory has been unlinked.
6215 SimulateIOError( return SQLITE_ERROR
);
6219 /* Call stat() on path zIn. Set bLink to true if the path is a symbolic
6220 ** link, or false otherwise. */
6223 if( osLstat(zIn
, &buf
)!=0 ){
6224 if( errno
!=ENOENT
){
6225 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "lstat", zIn
);
6228 bLink
= S_ISLNK(buf
.st_mode
);
6233 zDel
= sqlite3_malloc(nOut
);
6234 if( zDel
==0 ) rc
= SQLITE_NOMEM_BKPT
;
6235 }else if( ++nLink
>SQLITE_MAX_SYMLINKS
){
6236 rc
= SQLITE_CANTOPEN_BKPT
;
6239 if( rc
==SQLITE_OK
){
6240 nByte
= osReadlink(zIn
, zDel
, nOut
-1);
6242 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "readlink", zIn
);
6246 for(n
= sqlite3Strlen30(zIn
); n
>0 && zIn
[n
-1]!='/'; n
--);
6247 if( nByte
+n
+1>nOut
){
6248 rc
= SQLITE_CANTOPEN_BKPT
;
6250 memmove(&zDel
[n
], zDel
, nByte
+1);
6251 memcpy(zDel
, zIn
, n
);
6262 assert( rc
!=SQLITE_OK
|| zIn
!=zOut
|| zIn
[0]=='/' );
6263 if( rc
==SQLITE_OK
&& zIn
!=zOut
){
6264 rc
= mkFullPathname(zIn
, zOut
, nOut
);
6266 if( bLink
==0 ) break;
6268 }while( rc
==SQLITE_OK
);
6272 #endif /* HAVE_READLINK && HAVE_LSTAT */
6276 #ifndef SQLITE_OMIT_LOAD_EXTENSION
6278 ** Interfaces for opening a shared library, finding entry points
6279 ** within the shared library, and closing the shared library.
6282 static void *unixDlOpen(sqlite3_vfs
*NotUsed
, const char *zFilename
){
6283 UNUSED_PARAMETER(NotUsed
);
6284 return dlopen(zFilename
, RTLD_NOW
| RTLD_GLOBAL
);
6288 ** SQLite calls this function immediately after a call to unixDlSym() or
6289 ** unixDlOpen() fails (returns a null pointer). If a more detailed error
6290 ** message is available, it is written to zBufOut. If no error message
6291 ** is available, zBufOut is left unmodified and SQLite uses a default
6294 static void unixDlError(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBufOut
){
6296 UNUSED_PARAMETER(NotUsed
);
6300 sqlite3_snprintf(nBuf
, zBufOut
, "%s", zErr
);
6304 static void (*unixDlSym(sqlite3_vfs
*NotUsed
, void *p
, const char*zSym
))(void){
6306 ** GCC with -pedantic-errors says that C90 does not allow a void* to be
6307 ** cast into a pointer to a function. And yet the library dlsym() routine
6308 ** returns a void* which is really a pointer to a function. So how do we
6309 ** use dlsym() with -pedantic-errors?
6311 ** Variable x below is defined to be a pointer to a function taking
6312 ** parameters void* and const char* and returning a pointer to a function.
6313 ** We initialize x by assigning it a pointer to the dlsym() function.
6314 ** (That assignment requires a cast.) Then we call the function that
6317 ** This work-around is unlikely to work correctly on any system where
6318 ** you really cannot cast a function pointer into void*. But then, on the
6319 ** other hand, dlsym() will not work on such a system either, so we have
6320 ** not really lost anything.
6322 void (*(*x
)(void*,const char*))(void);
6323 UNUSED_PARAMETER(NotUsed
);
6324 x
= (void(*(*)(void*,const char*))(void))dlsym
;
6325 return (*x
)(p
, zSym
);
6327 static void unixDlClose(sqlite3_vfs
*NotUsed
, void *pHandle
){
6328 UNUSED_PARAMETER(NotUsed
);
6331 #else /* if SQLITE_OMIT_LOAD_EXTENSION is defined: */
6332 #define unixDlOpen 0
6333 #define unixDlError 0
6335 #define unixDlClose 0
6339 ** Write nBuf bytes of random data to the supplied buffer zBuf.
6341 static int unixRandomness(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBuf
){
6342 UNUSED_PARAMETER(NotUsed
);
6343 assert((size_t)nBuf
>=(sizeof(time_t)+sizeof(int)));
6345 /* We have to initialize zBuf to prevent valgrind from reporting
6346 ** errors. The reports issued by valgrind are incorrect - we would
6347 ** prefer that the randomness be increased by making use of the
6348 ** uninitialized space in zBuf - but valgrind errors tend to worry
6349 ** some users. Rather than argue, it seems easier just to initialize
6350 ** the whole array and silence valgrind, even if that means less randomness
6351 ** in the random seed.
6353 ** When testing, initializing zBuf[] to zero is all we do. That means
6354 ** that we always use the same random number sequence. This makes the
6355 ** tests repeatable.
6357 memset(zBuf
, 0, nBuf
);
6358 randomnessPid
= osGetpid(0);
6359 #if !defined(SQLITE_TEST) && !defined(SQLITE_OMIT_RANDOMNESS)
6362 fd
= robust_open("/dev/urandom", O_RDONLY
, 0);
6366 memcpy(zBuf
, &t
, sizeof(t
));
6367 memcpy(&zBuf
[sizeof(t
)], &randomnessPid
, sizeof(randomnessPid
));
6368 assert( sizeof(t
)+sizeof(randomnessPid
)<=(size_t)nBuf
);
6369 nBuf
= sizeof(t
) + sizeof(randomnessPid
);
6371 do{ got
= osRead(fd
, zBuf
, nBuf
); }while( got
<0 && errno
==EINTR
);
6372 robust_close(0, fd
, __LINE__
);
6381 ** Sleep for a little while. Return the amount of time slept.
6382 ** The argument is the number of microseconds we want to sleep.
6383 ** The return value is the number of microseconds of sleep actually
6384 ** requested from the underlying operating system, a number which
6385 ** might be greater than or equal to the argument, but not less
6386 ** than the argument.
6388 static int unixSleep(sqlite3_vfs
*NotUsed
, int microseconds
){
6392 sp
.tv_sec
= microseconds
/ 1000000;
6393 sp
.tv_nsec
= (microseconds
% 1000000) * 1000;
6394 nanosleep(&sp
, NULL
);
6395 UNUSED_PARAMETER(NotUsed
);
6396 return microseconds
;
6397 #elif defined(HAVE_USLEEP) && HAVE_USLEEP
6398 usleep(microseconds
);
6399 UNUSED_PARAMETER(NotUsed
);
6400 return microseconds
;
6402 int seconds
= (microseconds
+999999)/1000000;
6404 UNUSED_PARAMETER(NotUsed
);
6405 return seconds
*1000000;
6410 ** The following variable, if set to a non-zero value, is interpreted as
6411 ** the number of seconds since 1970 and is used to set the result of
6412 ** sqlite3OsCurrentTime() during testing.
6415 int sqlite3_current_time
= 0; /* Fake system time in seconds since 1970. */
6419 ** Find the current time (in Universal Coordinated Time). Write into *piNow
6420 ** the current time and date as a Julian Day number times 86_400_000. In
6421 ** other words, write into *piNow the number of milliseconds since the Julian
6422 ** epoch of noon in Greenwich on November 24, 4714 B.C according to the
6423 ** proleptic Gregorian calendar.
6425 ** On success, return SQLITE_OK. Return SQLITE_ERROR if the time and date
6428 static int unixCurrentTimeInt64(sqlite3_vfs
*NotUsed
, sqlite3_int64
*piNow
){
6429 static const sqlite3_int64 unixEpoch
= 24405875*(sqlite3_int64
)8640000;
6431 #if defined(NO_GETTOD)
6434 *piNow
= ((sqlite3_int64
)t
)*1000 + unixEpoch
;
6436 struct timespec sNow
;
6437 clock_gettime(CLOCK_REALTIME
, &sNow
);
6438 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_nsec
/1000000;
6440 struct timeval sNow
;
6441 (void)gettimeofday(&sNow
, 0); /* Cannot fail given valid arguments */
6442 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_usec
/1000;
6446 if( sqlite3_current_time
){
6447 *piNow
= 1000*(sqlite3_int64
)sqlite3_current_time
+ unixEpoch
;
6450 UNUSED_PARAMETER(NotUsed
);
6454 #ifndef SQLITE_OMIT_DEPRECATED
6456 ** Find the current time (in Universal Coordinated Time). Write the
6457 ** current time and date as a Julian Day number into *prNow and
6458 ** return 0. Return 1 if the time and date cannot be found.
6460 static int unixCurrentTime(sqlite3_vfs
*NotUsed
, double *prNow
){
6461 sqlite3_int64 i
= 0;
6463 UNUSED_PARAMETER(NotUsed
);
6464 rc
= unixCurrentTimeInt64(0, &i
);
6465 *prNow
= i
/86400000.0;
6469 # define unixCurrentTime 0
6473 ** The xGetLastError() method is designed to return a better
6474 ** low-level error message when operating-system problems come up
6475 ** during SQLite operation. Only the integer return code is currently
6478 static int unixGetLastError(sqlite3_vfs
*NotUsed
, int NotUsed2
, char *NotUsed3
){
6479 UNUSED_PARAMETER(NotUsed
);
6480 UNUSED_PARAMETER(NotUsed2
);
6481 UNUSED_PARAMETER(NotUsed3
);
6487 ************************ End of sqlite3_vfs methods ***************************
6488 ******************************************************************************/
6490 /******************************************************************************
6491 ************************** Begin Proxy Locking ********************************
6493 ** Proxy locking is a "uber-locking-method" in this sense: It uses the
6494 ** other locking methods on secondary lock files. Proxy locking is a
6495 ** meta-layer over top of the primitive locking implemented above. For
6496 ** this reason, the division that implements of proxy locking is deferred
6497 ** until late in the file (here) after all of the other I/O methods have
6498 ** been defined - so that the primitive locking methods are available
6499 ** as services to help with the implementation of proxy locking.
6503 ** The default locking schemes in SQLite use byte-range locks on the
6504 ** database file to coordinate safe, concurrent access by multiple readers
6505 ** and writers [http://sqlite.org/lockingv3.html]. The five file locking
6506 ** states (UNLOCKED, PENDING, SHARED, RESERVED, EXCLUSIVE) are implemented
6507 ** as POSIX read & write locks over fixed set of locations (via fsctl),
6508 ** on AFP and SMB only exclusive byte-range locks are available via fsctl
6509 ** with _IOWR('z', 23, struct ByteRangeLockPB2) to track the same 5 states.
6510 ** To simulate a F_RDLCK on the shared range, on AFP a randomly selected
6511 ** address in the shared range is taken for a SHARED lock, the entire
6512 ** shared range is taken for an EXCLUSIVE lock):
6514 ** PENDING_BYTE 0x40000000
6515 ** RESERVED_BYTE 0x40000001
6516 ** SHARED_RANGE 0x40000002 -> 0x40000200
6518 ** This works well on the local file system, but shows a nearly 100x
6519 ** slowdown in read performance on AFP because the AFP client disables
6520 ** the read cache when byte-range locks are present. Enabling the read
6521 ** cache exposes a cache coherency problem that is present on all OS X
6522 ** supported network file systems. NFS and AFP both observe the
6523 ** close-to-open semantics for ensuring cache coherency
6524 ** [http://nfs.sourceforge.net/#faq_a8], which does not effectively
6525 ** address the requirements for concurrent database access by multiple
6526 ** readers and writers
6527 ** [http://www.nabble.com/SQLite-on-NFS-cache-coherency-td15655701.html].
6529 ** To address the performance and cache coherency issues, proxy file locking
6530 ** changes the way database access is controlled by limiting access to a
6531 ** single host at a time and moving file locks off of the database file
6532 ** and onto a proxy file on the local file system.
6535 ** Using proxy locks
6536 ** -----------------
6540 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_SET_LOCKPROXYFILE,
6541 ** <proxy_path> | ":auto:");
6542 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_GET_LOCKPROXYFILE,
6548 ** PRAGMA [database.]lock_proxy_file=<proxy_path> | :auto:
6549 ** PRAGMA [database.]lock_proxy_file
6551 ** Specifying ":auto:" means that if there is a conch file with a matching
6552 ** host ID in it, the proxy path in the conch file will be used, otherwise
6553 ** a proxy path based on the user's temp dir
6554 ** (via confstr(_CS_DARWIN_USER_TEMP_DIR,...)) will be used and the
6555 ** actual proxy file name is generated from the name and path of the
6556 ** database file. For example:
6558 ** For database path "/Users/me/foo.db"
6559 ** The lock path will be "<tmpdir>/sqliteplocks/_Users_me_foo.db:auto:")
6561 ** Once a lock proxy is configured for a database connection, it can not
6562 ** be removed, however it may be switched to a different proxy path via
6563 ** the above APIs (assuming the conch file is not being held by another
6564 ** connection or process).
6567 ** How proxy locking works
6568 ** -----------------------
6570 ** Proxy file locking relies primarily on two new supporting files:
6572 ** * conch file to limit access to the database file to a single host
6575 ** * proxy file to act as a proxy for the advisory locks normally
6576 ** taken on the database
6578 ** The conch file - to use a proxy file, sqlite must first "hold the conch"
6579 ** by taking an sqlite-style shared lock on the conch file, reading the
6580 ** contents and comparing the host's unique host ID (see below) and lock
6581 ** proxy path against the values stored in the conch. The conch file is
6582 ** stored in the same directory as the database file and the file name
6583 ** is patterned after the database file name as ".<databasename>-conch".
6584 ** If the conch file does not exist, or its contents do not match the
6585 ** host ID and/or proxy path, then the lock is escalated to an exclusive
6586 ** lock and the conch file contents is updated with the host ID and proxy
6587 ** path and the lock is downgraded to a shared lock again. If the conch
6588 ** is held by another process (with a shared lock), the exclusive lock
6589 ** will fail and SQLITE_BUSY is returned.
6591 ** The proxy file - a single-byte file used for all advisory file locks
6592 ** normally taken on the database file. This allows for safe sharing
6593 ** of the database file for multiple readers and writers on the same
6594 ** host (the conch ensures that they all use the same local lock file).
6596 ** Requesting the lock proxy does not immediately take the conch, it is
6597 ** only taken when the first request to lock database file is made.
6598 ** This matches the semantics of the traditional locking behavior, where
6599 ** opening a connection to a database file does not take a lock on it.
6600 ** The shared lock and an open file descriptor are maintained until
6601 ** the connection to the database is closed.
6603 ** The proxy file and the lock file are never deleted so they only need
6604 ** to be created the first time they are used.
6606 ** Configuration options
6607 ** ---------------------
6609 ** SQLITE_PREFER_PROXY_LOCKING
6611 ** Database files accessed on non-local file systems are
6612 ** automatically configured for proxy locking, lock files are
6613 ** named automatically using the same logic as
6614 ** PRAGMA lock_proxy_file=":auto:"
6616 ** SQLITE_PROXY_DEBUG
6618 ** Enables the logging of error messages during host id file
6619 ** retrieval and creation
6623 ** Overrides the default directory used for lock proxy files that
6624 ** are named automatically via the ":auto:" setting
6626 ** SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
6628 ** Permissions to use when creating a directory for storing the
6629 ** lock proxy files, only used when LOCKPROXYDIR is not set.
6632 ** As mentioned above, when compiled with SQLITE_PREFER_PROXY_LOCKING,
6633 ** setting the environment variable SQLITE_FORCE_PROXY_LOCKING to 1 will
6634 ** force proxy locking to be used for every database file opened, and 0
6635 ** will force automatic proxy locking to be disabled for all database
6636 ** files (explicitly calling the SQLITE_FCNTL_SET_LOCKPROXYFILE pragma or
6637 ** sqlite_file_control API is not affected by SQLITE_FORCE_PROXY_LOCKING).
6641 ** Proxy locking is only available on MacOSX
6643 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
6646 ** The proxyLockingContext has the path and file structures for the remote
6647 ** and local proxy files in it
6649 typedef struct proxyLockingContext proxyLockingContext
;
6650 struct proxyLockingContext
{
6651 unixFile
*conchFile
; /* Open conch file */
6652 char *conchFilePath
; /* Name of the conch file */
6653 unixFile
*lockProxy
; /* Open proxy lock file */
6654 char *lockProxyPath
; /* Name of the proxy lock file */
6655 char *dbPath
; /* Name of the open file */
6656 int conchHeld
; /* 1 if the conch is held, -1 if lockless */
6657 int nFails
; /* Number of conch taking failures */
6658 void *oldLockingContext
; /* Original lockingcontext to restore on close */
6659 sqlite3_io_methods
const *pOldMethod
; /* Original I/O methods for close */
6663 ** The proxy lock file path for the database at dbPath is written into lPath,
6664 ** which must point to valid, writable memory large enough for a maxLen length
6667 static int proxyGetLockPath(const char *dbPath
, char *lPath
, size_t maxLen
){
6673 len
= strlcpy(lPath
, LOCKPROXYDIR
, maxLen
);
6675 # ifdef _CS_DARWIN_USER_TEMP_DIR
6677 if( !confstr(_CS_DARWIN_USER_TEMP_DIR
, lPath
, maxLen
) ){
6678 OSTRACE(("GETLOCKPATH failed %s errno=%d pid=%d\n",
6679 lPath
, errno
, osGetpid(0)));
6680 return SQLITE_IOERR_LOCK
;
6682 len
= strlcat(lPath
, "sqliteplocks", maxLen
);
6685 len
= strlcpy(lPath
, "/tmp/", maxLen
);
6689 if( lPath
[len
-1]!='/' ){
6690 len
= strlcat(lPath
, "/", maxLen
);
6693 /* transform the db path to a unique cache name */
6694 dbLen
= (int)strlen(dbPath
);
6695 for( i
=0; i
<dbLen
&& (i
+len
+7)<(int)maxLen
; i
++){
6697 lPath
[i
+len
] = (c
=='/')?'_':c
;
6700 strlcat(lPath
, ":auto:", maxLen
);
6701 OSTRACE(("GETLOCKPATH proxy lock path=%s pid=%d\n", lPath
, osGetpid(0)));
6706 ** Creates the lock file and any missing directories in lockPath
6708 static int proxyCreateLockPath(const char *lockPath
){
6710 char buf
[MAXPATHLEN
];
6713 assert(lockPath
!=NULL
);
6714 /* try to create all the intermediate directories */
6715 len
= (int)strlen(lockPath
);
6716 buf
[0] = lockPath
[0];
6717 for( i
=1; i
<len
; i
++ ){
6718 if( lockPath
[i
] == '/' && (i
- start
> 0) ){
6719 /* only mkdir if leaf dir != "." or "/" or ".." */
6720 if( i
-start
>2 || (i
-start
==1 && buf
[start
] != '.' && buf
[start
] != '/')
6721 || (i
-start
==2 && buf
[start
] != '.' && buf
[start
+1] != '.') ){
6723 if( osMkdir(buf
, SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
) ){
6726 OSTRACE(("CREATELOCKPATH FAILED creating %s, "
6727 "'%s' proxy lock path=%s pid=%d\n",
6728 buf
, strerror(err
), lockPath
, osGetpid(0)));
6735 buf
[i
] = lockPath
[i
];
6737 OSTRACE(("CREATELOCKPATH proxy lock path=%s pid=%d\n",lockPath
,osGetpid(0)));
6742 ** Create a new VFS file descriptor (stored in memory obtained from
6743 ** sqlite3_malloc) and open the file named "path" in the file descriptor.
6745 ** The caller is responsible not only for closing the file descriptor
6746 ** but also for freeing the memory associated with the file descriptor.
6748 static int proxyCreateUnixFile(
6749 const char *path
, /* path for the new unixFile */
6750 unixFile
**ppFile
, /* unixFile created and returned by ref */
6751 int islockfile
/* if non zero missing dirs will be created */
6756 int openFlags
= O_RDWR
| O_CREAT
;
6757 sqlite3_vfs dummyVfs
;
6759 UnixUnusedFd
*pUnused
= NULL
;
6761 /* 1. first try to open/create the file
6762 ** 2. if that fails, and this is a lock file (not-conch), try creating
6763 ** the parent directories and then try again.
6764 ** 3. if that fails, try to open the file read-only
6765 ** otherwise return BUSY (if lock file) or CANTOPEN for the conch file
6767 pUnused
= findReusableFd(path
, openFlags
);
6771 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
6773 return SQLITE_NOMEM_BKPT
;
6777 fd
= robust_open(path
, openFlags
, 0);
6779 if( fd
<0 && errno
==ENOENT
&& islockfile
){
6780 if( proxyCreateLockPath(path
) == SQLITE_OK
){
6781 fd
= robust_open(path
, openFlags
, 0);
6786 openFlags
= O_RDONLY
;
6787 fd
= robust_open(path
, openFlags
, 0);
6798 return SQLITE_IOERR_LOCK
; /* even though it is the conch */
6800 return SQLITE_CANTOPEN_BKPT
;
6804 pNew
= (unixFile
*)sqlite3_malloc64(sizeof(*pNew
));
6806 rc
= SQLITE_NOMEM_BKPT
;
6807 goto end_create_proxy
;
6809 memset(pNew
, 0, sizeof(unixFile
));
6810 pNew
->openFlags
= openFlags
;
6811 memset(&dummyVfs
, 0, sizeof(dummyVfs
));
6812 dummyVfs
.pAppData
= (void*)&autolockIoFinder
;
6813 dummyVfs
.zName
= "dummy";
6815 pUnused
->flags
= openFlags
;
6816 pNew
->pPreallocatedUnused
= pUnused
;
6818 rc
= fillInUnixFile(&dummyVfs
, fd
, (sqlite3_file
*)pNew
, path
, 0);
6819 if( rc
==SQLITE_OK
){
6824 robust_close(pNew
, fd
, __LINE__
);
6826 sqlite3_free(pUnused
);
6831 /* simulate multiple hosts by creating unique hostid file paths */
6832 int sqlite3_hostid_num
= 0;
6835 #define PROXY_HOSTIDLEN 16 /* conch file host id length */
6837 #ifdef HAVE_GETHOSTUUID
6838 /* Not always defined in the headers as it ought to be */
6839 extern int gethostuuid(uuid_t id
, const struct timespec
*wait
);
6842 /* get the host ID via gethostuuid(), pHostID must point to PROXY_HOSTIDLEN
6843 ** bytes of writable memory.
6845 static int proxyGetHostID(unsigned char *pHostID
, int *pError
){
6846 assert(PROXY_HOSTIDLEN
== sizeof(uuid_t
));
6847 memset(pHostID
, 0, PROXY_HOSTIDLEN
);
6848 #ifdef HAVE_GETHOSTUUID
6850 struct timespec timeout
= {1, 0}; /* 1 sec timeout */
6851 if( gethostuuid(pHostID
, &timeout
) ){
6856 return SQLITE_IOERR
;
6860 UNUSED_PARAMETER(pError
);
6863 /* simulate multiple hosts by creating unique hostid file paths */
6864 if( sqlite3_hostid_num
!= 0){
6865 pHostID
[0] = (char)(pHostID
[0] + (char)(sqlite3_hostid_num
& 0xFF));
6872 /* The conch file contains the header, host id and lock file path
6874 #define PROXY_CONCHVERSION 2 /* 1-byte header, 16-byte host id, path */
6875 #define PROXY_HEADERLEN 1 /* conch file header length */
6876 #define PROXY_PATHINDEX (PROXY_HEADERLEN+PROXY_HOSTIDLEN)
6877 #define PROXY_MAXCONCHLEN (PROXY_HEADERLEN+PROXY_HOSTIDLEN+MAXPATHLEN)
6880 ** Takes an open conch file, copies the contents to a new path and then moves
6881 ** it back. The newly created file's file descriptor is assigned to the
6882 ** conch file structure and finally the original conch file descriptor is
6883 ** closed. Returns zero if successful.
6885 static int proxyBreakConchLock(unixFile
*pFile
, uuid_t myHostID
){
6886 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6887 unixFile
*conchFile
= pCtx
->conchFile
;
6888 char tPath
[MAXPATHLEN
];
6889 char buf
[PROXY_MAXCONCHLEN
];
6890 char *cPath
= pCtx
->conchFilePath
;
6893 char errmsg
[64] = "";
6896 UNUSED_PARAMETER(myHostID
);
6898 /* create a new path by replace the trailing '-conch' with '-break' */
6899 pathLen
= strlcpy(tPath
, cPath
, MAXPATHLEN
);
6900 if( pathLen
>MAXPATHLEN
|| pathLen
<6 ||
6901 (strlcpy(&tPath
[pathLen
-5], "break", 6) != 5) ){
6902 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"path error (len %d)",(int)pathLen
);
6905 /* read the conch content */
6906 readLen
= osPread(conchFile
->h
, buf
, PROXY_MAXCONCHLEN
, 0);
6907 if( readLen
<PROXY_PATHINDEX
){
6908 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"read error (len %d)",(int)readLen
);
6911 /* write it out to the temporary break file */
6912 fd
= robust_open(tPath
, (O_RDWR
|O_CREAT
|O_EXCL
), 0);
6914 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "create failed (%d)", errno
);
6917 if( osPwrite(fd
, buf
, readLen
, 0) != (ssize_t
)readLen
){
6918 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "write failed (%d)", errno
);
6921 if( rename(tPath
, cPath
) ){
6922 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "rename failed (%d)", errno
);
6926 fprintf(stderr
, "broke stale lock on %s\n", cPath
);
6927 robust_close(pFile
, conchFile
->h
, __LINE__
);
6929 conchFile
->openFlags
= O_RDWR
| O_CREAT
;
6935 robust_close(pFile
, fd
, __LINE__
);
6937 fprintf(stderr
, "failed to break stale lock on %s, %s\n", cPath
, errmsg
);
6942 /* Take the requested lock on the conch file and break a stale lock if the
6945 static int proxyConchLock(unixFile
*pFile
, uuid_t myHostID
, int lockType
){
6946 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6947 unixFile
*conchFile
= pCtx
->conchFile
;
6950 struct timespec conchModTime
;
6952 memset(&conchModTime
, 0, sizeof(conchModTime
));
6954 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
6956 if( rc
==SQLITE_BUSY
){
6957 /* If the lock failed (busy):
6958 * 1st try: get the mod time of the conch, wait 0.5s and try again.
6959 * 2nd try: fail if the mod time changed or host id is different, wait
6960 * 10 sec and try again
6961 * 3rd try: break the lock unless the mod time has changed.
6964 if( osFstat(conchFile
->h
, &buf
) ){
6965 storeLastErrno(pFile
, errno
);
6966 return SQLITE_IOERR_LOCK
;
6970 conchModTime
= buf
.st_mtimespec
;
6971 usleep(500000); /* wait 0.5 sec and try the lock again*/
6976 if( conchModTime
.tv_sec
!= buf
.st_mtimespec
.tv_sec
||
6977 conchModTime
.tv_nsec
!= buf
.st_mtimespec
.tv_nsec
){
6982 char tBuf
[PROXY_MAXCONCHLEN
];
6983 int len
= osPread(conchFile
->h
, tBuf
, PROXY_MAXCONCHLEN
, 0);
6985 storeLastErrno(pFile
, errno
);
6986 return SQLITE_IOERR_LOCK
;
6988 if( len
>PROXY_PATHINDEX
&& tBuf
[0]==(char)PROXY_CONCHVERSION
){
6989 /* don't break the lock if the host id doesn't match */
6990 if( 0!=memcmp(&tBuf
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
) ){
6994 /* don't break the lock on short read or a version mismatch */
6997 usleep(10000000); /* wait 10 sec and try the lock again */
7001 assert( nTries
==3 );
7002 if( 0==proxyBreakConchLock(pFile
, myHostID
) ){
7004 if( lockType
==EXCLUSIVE_LOCK
){
7005 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
7008 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
7012 } while( rc
==SQLITE_BUSY
&& nTries
<3 );
7017 /* Takes the conch by taking a shared lock and read the contents conch, if
7018 ** lockPath is non-NULL, the host ID and lock file path must match. A NULL
7019 ** lockPath means that the lockPath in the conch file will be used if the
7020 ** host IDs match, or a new lock path will be generated automatically
7021 ** and written to the conch file.
7023 static int proxyTakeConch(unixFile
*pFile
){
7024 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7026 if( pCtx
->conchHeld
!=0 ){
7029 unixFile
*conchFile
= pCtx
->conchFile
;
7032 char readBuf
[PROXY_MAXCONCHLEN
];
7033 char lockPath
[MAXPATHLEN
];
7034 char *tempLockPath
= NULL
;
7036 int createConch
= 0;
7037 int hostIdMatch
= 0;
7039 int tryOldLockPath
= 0;
7040 int forceNewLockPath
= 0;
7042 OSTRACE(("TAKECONCH %d for %s pid=%d\n", conchFile
->h
,
7043 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
7046 rc
= proxyGetHostID(myHostID
, &pError
);
7047 if( (rc
&0xff)==SQLITE_IOERR
){
7048 storeLastErrno(pFile
, pError
);
7051 rc
= proxyConchLock(pFile
, myHostID
, SHARED_LOCK
);
7052 if( rc
!=SQLITE_OK
){
7055 /* read the existing conch file */
7056 readLen
= seekAndRead((unixFile
*)conchFile
, 0, readBuf
, PROXY_MAXCONCHLEN
);
7058 /* I/O error: lastErrno set by seekAndRead */
7059 storeLastErrno(pFile
, conchFile
->lastErrno
);
7060 rc
= SQLITE_IOERR_READ
;
7062 }else if( readLen
<=(PROXY_HEADERLEN
+PROXY_HOSTIDLEN
) ||
7063 readBuf
[0]!=(char)PROXY_CONCHVERSION
){
7064 /* a short read or version format mismatch means we need to create a new
7069 /* if the host id matches and the lock path already exists in the conch
7070 ** we'll try to use the path there, if we can't open that path, we'll
7071 ** retry with a new auto-generated path
7073 do { /* in case we need to try again for an :auto: named lock file */
7075 if( !createConch
&& !forceNewLockPath
){
7076 hostIdMatch
= !memcmp(&readBuf
[PROXY_HEADERLEN
], myHostID
,
7078 /* if the conch has data compare the contents */
7079 if( !pCtx
->lockProxyPath
){
7080 /* for auto-named local lock file, just check the host ID and we'll
7081 ** use the local lock file path that's already in there
7084 size_t pathLen
= (readLen
- PROXY_PATHINDEX
);
7086 if( pathLen
>=MAXPATHLEN
){
7087 pathLen
=MAXPATHLEN
-1;
7089 memcpy(lockPath
, &readBuf
[PROXY_PATHINDEX
], pathLen
);
7090 lockPath
[pathLen
] = 0;
7091 tempLockPath
= lockPath
;
7093 /* create a copy of the lock path if the conch is taken */
7096 }else if( hostIdMatch
7097 && !strncmp(pCtx
->lockProxyPath
, &readBuf
[PROXY_PATHINDEX
],
7098 readLen
-PROXY_PATHINDEX
)
7100 /* conch host and lock path match */
7105 /* if the conch isn't writable and doesn't match, we can't take it */
7106 if( (conchFile
->openFlags
&O_RDWR
) == 0 ){
7111 /* either the conch didn't match or we need to create a new one */
7112 if( !pCtx
->lockProxyPath
){
7113 proxyGetLockPath(pCtx
->dbPath
, lockPath
, MAXPATHLEN
);
7114 tempLockPath
= lockPath
;
7115 /* create a copy of the lock path _only_ if the conch is taken */
7118 /* update conch with host and path (this will fail if other process
7119 ** has a shared lock already), if the host id matches, use the big
7122 futimes(conchFile
->h
, NULL
);
7123 if( hostIdMatch
&& !createConch
){
7124 if( conchFile
->pInode
&& conchFile
->pInode
->nShared
>1 ){
7125 /* We are trying for an exclusive lock but another thread in this
7126 ** same process is still holding a shared lock. */
7129 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7132 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7134 if( rc
==SQLITE_OK
){
7135 char writeBuffer
[PROXY_MAXCONCHLEN
];
7138 writeBuffer
[0] = (char)PROXY_CONCHVERSION
;
7139 memcpy(&writeBuffer
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
);
7140 if( pCtx
->lockProxyPath
!=NULL
){
7141 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], pCtx
->lockProxyPath
,
7144 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], tempLockPath
, MAXPATHLEN
);
7146 writeSize
= PROXY_PATHINDEX
+ strlen(&writeBuffer
[PROXY_PATHINDEX
]);
7147 robust_ftruncate(conchFile
->h
, writeSize
);
7148 rc
= unixWrite((sqlite3_file
*)conchFile
, writeBuffer
, writeSize
, 0);
7149 full_fsync(conchFile
->h
,0,0);
7150 /* If we created a new conch file (not just updated the contents of a
7151 ** valid conch file), try to match the permissions of the database
7153 if( rc
==SQLITE_OK
&& createConch
){
7155 int err
= osFstat(pFile
->h
, &buf
);
7157 mode_t cmode
= buf
.st_mode
&(S_IRUSR
|S_IWUSR
| S_IRGRP
|S_IWGRP
|
7159 /* try to match the database file R/W permissions, ignore failure */
7160 #ifndef SQLITE_PROXY_DEBUG
7161 osFchmod(conchFile
->h
, cmode
);
7164 rc
= osFchmod(conchFile
->h
, cmode
);
7165 }while( rc
==(-1) && errno
==EINTR
);
7168 fprintf(stderr
, "fchmod %o FAILED with %d %s\n",
7169 cmode
, code
, strerror(code
));
7171 fprintf(stderr
, "fchmod %o SUCCEDED\n",cmode
);
7175 fprintf(stderr
, "STAT FAILED[%d] with %d %s\n",
7176 err
, code
, strerror(code
));
7181 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
7184 OSTRACE(("TRANSPROXY: CLOSE %d\n", pFile
->h
));
7185 if( rc
==SQLITE_OK
&& pFile
->openFlags
){
7188 robust_close(pFile
, pFile
->h
, __LINE__
);
7191 fd
= robust_open(pCtx
->dbPath
, pFile
->openFlags
, 0);
7192 OSTRACE(("TRANSPROXY: OPEN %d\n", fd
));
7196 rc
=SQLITE_CANTOPEN_BKPT
; /* SQLITE_BUSY? proxyTakeConch called
7200 if( rc
==SQLITE_OK
&& !pCtx
->lockProxy
){
7201 char *path
= tempLockPath
? tempLockPath
: pCtx
->lockProxyPath
;
7202 rc
= proxyCreateUnixFile(path
, &pCtx
->lockProxy
, 1);
7203 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_NOMEM
&& tryOldLockPath
){
7204 /* we couldn't create the proxy lock file with the old lock file path
7205 ** so try again via auto-naming
7207 forceNewLockPath
= 1;
7209 continue; /* go back to the do {} while start point, try again */
7212 if( rc
==SQLITE_OK
){
7213 /* Need to make a copy of path if we extracted the value
7214 ** from the conch file or the path was allocated on the stack
7217 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, tempLockPath
);
7218 if( !pCtx
->lockProxyPath
){
7219 rc
= SQLITE_NOMEM_BKPT
;
7223 if( rc
==SQLITE_OK
){
7224 pCtx
->conchHeld
= 1;
7226 if( pCtx
->lockProxy
->pMethod
== &afpIoMethods
){
7227 afpLockingContext
*afpCtx
;
7228 afpCtx
= (afpLockingContext
*)pCtx
->lockProxy
->lockingContext
;
7229 afpCtx
->dbPath
= pCtx
->lockProxyPath
;
7232 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7234 OSTRACE(("TAKECONCH %d %s\n", conchFile
->h
,
7235 rc
==SQLITE_OK
?"ok":"failed"));
7237 } while (1); /* in case we need to retry the :auto: lock file -
7238 ** we should never get here except via the 'continue' call. */
7243 ** If pFile holds a lock on a conch file, then release that lock.
7245 static int proxyReleaseConch(unixFile
*pFile
){
7246 int rc
= SQLITE_OK
; /* Subroutine return code */
7247 proxyLockingContext
*pCtx
; /* The locking context for the proxy lock */
7248 unixFile
*conchFile
; /* Name of the conch file */
7250 pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7251 conchFile
= pCtx
->conchFile
;
7252 OSTRACE(("RELEASECONCH %d for %s pid=%d\n", conchFile
->h
,
7253 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
7255 if( pCtx
->conchHeld
>0 ){
7256 rc
= conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7258 pCtx
->conchHeld
= 0;
7259 OSTRACE(("RELEASECONCH %d %s\n", conchFile
->h
,
7260 (rc
==SQLITE_OK
? "ok" : "failed")));
7265 ** Given the name of a database file, compute the name of its conch file.
7266 ** Store the conch filename in memory obtained from sqlite3_malloc64().
7267 ** Make *pConchPath point to the new name. Return SQLITE_OK on success
7268 ** or SQLITE_NOMEM if unable to obtain memory.
7270 ** The caller is responsible for ensuring that the allocated memory
7271 ** space is eventually freed.
7273 ** *pConchPath is set to NULL if a memory allocation error occurs.
7275 static int proxyCreateConchPathname(char *dbPath
, char **pConchPath
){
7276 int i
; /* Loop counter */
7277 int len
= (int)strlen(dbPath
); /* Length of database filename - dbPath */
7278 char *conchPath
; /* buffer in which to construct conch name */
7280 /* Allocate space for the conch filename and initialize the name to
7281 ** the name of the original database file. */
7282 *pConchPath
= conchPath
= (char *)sqlite3_malloc64(len
+ 8);
7284 return SQLITE_NOMEM_BKPT
;
7286 memcpy(conchPath
, dbPath
, len
+1);
7288 /* now insert a "." before the last / character */
7289 for( i
=(len
-1); i
>=0; i
-- ){
7290 if( conchPath
[i
]=='/' ){
7297 conchPath
[i
+1]=dbPath
[i
];
7301 /* append the "-conch" suffix to the file */
7302 memcpy(&conchPath
[i
+1], "-conch", 7);
7303 assert( (int)strlen(conchPath
) == len
+7 );
7309 /* Takes a fully configured proxy locking-style unix file and switches
7310 ** the local lock file path
7312 static int switchLockProxyPath(unixFile
*pFile
, const char *path
) {
7313 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7314 char *oldPath
= pCtx
->lockProxyPath
;
7317 if( pFile
->eFileLock
!=NO_LOCK
){
7321 /* nothing to do if the path is NULL, :auto: or matches the existing path */
7322 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ||
7323 (oldPath
&& !strncmp(oldPath
, path
, MAXPATHLEN
)) ){
7326 unixFile
*lockProxy
= pCtx
->lockProxy
;
7327 pCtx
->lockProxy
=NULL
;
7328 pCtx
->conchHeld
= 0;
7329 if( lockProxy
!=NULL
){
7330 rc
=lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7332 sqlite3_free(lockProxy
);
7334 sqlite3_free(oldPath
);
7335 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, path
);
7342 ** pFile is a file that has been opened by a prior xOpen call. dbPath
7343 ** is a string buffer at least MAXPATHLEN+1 characters in size.
7345 ** This routine find the filename associated with pFile and writes it
7348 static int proxyGetDbPathForUnixFile(unixFile
*pFile
, char *dbPath
){
7349 #if defined(__APPLE__)
7350 if( pFile
->pMethod
== &afpIoMethods
){
7351 /* afp style keeps a reference to the db path in the filePath field
7353 assert( (int)strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7354 strlcpy(dbPath
, ((afpLockingContext
*)pFile
->lockingContext
)->dbPath
,
7358 if( pFile
->pMethod
== &dotlockIoMethods
){
7359 /* dot lock style uses the locking context to store the dot lock
7361 int len
= strlen((char *)pFile
->lockingContext
) - strlen(DOTLOCK_SUFFIX
);
7362 memcpy(dbPath
, (char *)pFile
->lockingContext
, len
+ 1);
7364 /* all other styles use the locking context to store the db file path */
7365 assert( strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7366 strlcpy(dbPath
, (char *)pFile
->lockingContext
, MAXPATHLEN
);
7372 ** Takes an already filled in unix file and alters it so all file locking
7373 ** will be performed on the local proxy lock file. The following fields
7374 ** are preserved in the locking context so that they can be restored and
7375 ** the unix structure properly cleaned up at close time:
7379 static int proxyTransformUnixFile(unixFile
*pFile
, const char *path
) {
7380 proxyLockingContext
*pCtx
;
7381 char dbPath
[MAXPATHLEN
+1]; /* Name of the database file */
7382 char *lockPath
=NULL
;
7385 if( pFile
->eFileLock
!=NO_LOCK
){
7388 proxyGetDbPathForUnixFile(pFile
, dbPath
);
7389 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ){
7392 lockPath
=(char *)path
;
7395 OSTRACE(("TRANSPROXY %d for %s pid=%d\n", pFile
->h
,
7396 (lockPath
? lockPath
: ":auto:"), osGetpid(0)));
7398 pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
7400 return SQLITE_NOMEM_BKPT
;
7402 memset(pCtx
, 0, sizeof(*pCtx
));
7404 rc
= proxyCreateConchPathname(dbPath
, &pCtx
->conchFilePath
);
7405 if( rc
==SQLITE_OK
){
7406 rc
= proxyCreateUnixFile(pCtx
->conchFilePath
, &pCtx
->conchFile
, 0);
7407 if( rc
==SQLITE_CANTOPEN
&& ((pFile
->openFlags
&O_RDWR
) == 0) ){
7408 /* if (a) the open flags are not O_RDWR, (b) the conch isn't there, and
7409 ** (c) the file system is read-only, then enable no-locking access.
7410 ** Ugh, since O_RDONLY==0x0000 we test for !O_RDWR since unixOpen asserts
7411 ** that openFlags will have only one of O_RDONLY or O_RDWR.
7413 struct statfs fsInfo
;
7414 struct stat conchInfo
;
7417 if( osStat(pCtx
->conchFilePath
, &conchInfo
) == -1 ) {
7419 if( (err
==ENOENT
) && (statfs(dbPath
, &fsInfo
) != -1) ){
7420 goLockless
= (fsInfo
.f_flags
&MNT_RDONLY
) == MNT_RDONLY
;
7424 pCtx
->conchHeld
= -1; /* read only FS/ lockless */
7429 if( rc
==SQLITE_OK
&& lockPath
){
7430 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, lockPath
);
7433 if( rc
==SQLITE_OK
){
7434 pCtx
->dbPath
= sqlite3DbStrDup(0, dbPath
);
7435 if( pCtx
->dbPath
==NULL
){
7436 rc
= SQLITE_NOMEM_BKPT
;
7439 if( rc
==SQLITE_OK
){
7440 /* all memory is allocated, proxys are created and assigned,
7441 ** switch the locking context and pMethod then return.
7443 pCtx
->oldLockingContext
= pFile
->lockingContext
;
7444 pFile
->lockingContext
= pCtx
;
7445 pCtx
->pOldMethod
= pFile
->pMethod
;
7446 pFile
->pMethod
= &proxyIoMethods
;
7448 if( pCtx
->conchFile
){
7449 pCtx
->conchFile
->pMethod
->xClose((sqlite3_file
*)pCtx
->conchFile
);
7450 sqlite3_free(pCtx
->conchFile
);
7452 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7453 sqlite3_free(pCtx
->conchFilePath
);
7456 OSTRACE(("TRANSPROXY %d %s\n", pFile
->h
,
7457 (rc
==SQLITE_OK
? "ok" : "failed")));
7463 ** This routine handles sqlite3_file_control() calls that are specific
7464 ** to proxy locking.
7466 static int proxyFileControl(sqlite3_file
*id
, int op
, void *pArg
){
7468 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
7469 unixFile
*pFile
= (unixFile
*)id
;
7470 if( pFile
->pMethod
== &proxyIoMethods
){
7471 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7472 proxyTakeConch(pFile
);
7473 if( pCtx
->lockProxyPath
){
7474 *(const char **)pArg
= pCtx
->lockProxyPath
;
7476 *(const char **)pArg
= ":auto: (not held)";
7479 *(const char **)pArg
= NULL
;
7483 case SQLITE_FCNTL_SET_LOCKPROXYFILE
: {
7484 unixFile
*pFile
= (unixFile
*)id
;
7486 int isProxyStyle
= (pFile
->pMethod
== &proxyIoMethods
);
7487 if( pArg
==NULL
|| (const char *)pArg
==0 ){
7489 /* turn off proxy locking - not supported. If support is added for
7490 ** switching proxy locking mode off then it will need to fail if
7491 ** the journal mode is WAL mode.
7493 rc
= SQLITE_ERROR
/*SQLITE_PROTOCOL? SQLITE_MISUSE?*/;
7495 /* turn off proxy locking - already off - NOOP */
7499 const char *proxyPath
= (const char *)pArg
;
7501 proxyLockingContext
*pCtx
=
7502 (proxyLockingContext
*)pFile
->lockingContext
;
7503 if( !strcmp(pArg
, ":auto:")
7504 || (pCtx
->lockProxyPath
&&
7505 !strncmp(pCtx
->lockProxyPath
, proxyPath
, MAXPATHLEN
))
7509 rc
= switchLockProxyPath(pFile
, proxyPath
);
7512 /* turn on proxy file locking */
7513 rc
= proxyTransformUnixFile(pFile
, proxyPath
);
7519 assert( 0 ); /* The call assures that only valid opcodes are sent */
7523 return SQLITE_ERROR
;
7527 ** Within this division (the proxying locking implementation) the procedures
7528 ** above this point are all utilities. The lock-related methods of the
7529 ** proxy-locking sqlite3_io_method object follow.
7534 ** This routine checks if there is a RESERVED lock held on the specified
7535 ** file by this or any other process. If such a lock is held, set *pResOut
7536 ** to a non-zero value otherwise *pResOut is set to zero. The return value
7537 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
7539 static int proxyCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
7540 unixFile
*pFile
= (unixFile
*)id
;
7541 int rc
= proxyTakeConch(pFile
);
7542 if( rc
==SQLITE_OK
){
7543 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7544 if( pCtx
->conchHeld
>0 ){
7545 unixFile
*proxy
= pCtx
->lockProxy
;
7546 return proxy
->pMethod
->xCheckReservedLock((sqlite3_file
*)proxy
, pResOut
);
7547 }else{ /* conchHeld < 0 is lockless */
7555 ** Lock the file with the lock specified by parameter eFileLock - one
7556 ** of the following:
7559 ** (2) RESERVED_LOCK
7561 ** (4) EXCLUSIVE_LOCK
7563 ** Sometimes when requesting one lock state, additional lock states
7564 ** are inserted in between. The locking might fail on one of the later
7565 ** transitions leaving the lock state different from what it started but
7566 ** still short of its goal. The following chart shows the allowed
7567 ** transitions and the inserted intermediate states:
7569 ** UNLOCKED -> SHARED
7570 ** SHARED -> RESERVED
7571 ** SHARED -> (PENDING) -> EXCLUSIVE
7572 ** RESERVED -> (PENDING) -> EXCLUSIVE
7573 ** PENDING -> EXCLUSIVE
7575 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
7576 ** routine to lower a locking level.
7578 static int proxyLock(sqlite3_file
*id
, int eFileLock
) {
7579 unixFile
*pFile
= (unixFile
*)id
;
7580 int rc
= proxyTakeConch(pFile
);
7581 if( rc
==SQLITE_OK
){
7582 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7583 if( pCtx
->conchHeld
>0 ){
7584 unixFile
*proxy
= pCtx
->lockProxy
;
7585 rc
= proxy
->pMethod
->xLock((sqlite3_file
*)proxy
, eFileLock
);
7586 pFile
->eFileLock
= proxy
->eFileLock
;
7588 /* conchHeld < 0 is lockless */
7596 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
7597 ** must be either NO_LOCK or SHARED_LOCK.
7599 ** If the locking level of the file descriptor is already at or below
7600 ** the requested locking level, this routine is a no-op.
7602 static int proxyUnlock(sqlite3_file
*id
, int eFileLock
) {
7603 unixFile
*pFile
= (unixFile
*)id
;
7604 int rc
= proxyTakeConch(pFile
);
7605 if( rc
==SQLITE_OK
){
7606 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7607 if( pCtx
->conchHeld
>0 ){
7608 unixFile
*proxy
= pCtx
->lockProxy
;
7609 rc
= proxy
->pMethod
->xUnlock((sqlite3_file
*)proxy
, eFileLock
);
7610 pFile
->eFileLock
= proxy
->eFileLock
;
7612 /* conchHeld < 0 is lockless */
7619 ** Close a file that uses proxy locks.
7621 static int proxyClose(sqlite3_file
*id
) {
7623 unixFile
*pFile
= (unixFile
*)id
;
7624 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7625 unixFile
*lockProxy
= pCtx
->lockProxy
;
7626 unixFile
*conchFile
= pCtx
->conchFile
;
7630 rc
= lockProxy
->pMethod
->xUnlock((sqlite3_file
*)lockProxy
, NO_LOCK
);
7632 rc
= lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7634 sqlite3_free(lockProxy
);
7635 pCtx
->lockProxy
= 0;
7638 if( pCtx
->conchHeld
){
7639 rc
= proxyReleaseConch(pFile
);
7642 rc
= conchFile
->pMethod
->xClose((sqlite3_file
*)conchFile
);
7644 sqlite3_free(conchFile
);
7646 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7647 sqlite3_free(pCtx
->conchFilePath
);
7648 sqlite3DbFree(0, pCtx
->dbPath
);
7649 /* restore the original locking context and pMethod then close it */
7650 pFile
->lockingContext
= pCtx
->oldLockingContext
;
7651 pFile
->pMethod
= pCtx
->pOldMethod
;
7653 return pFile
->pMethod
->xClose(id
);
7660 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
7662 ** The proxy locking style is intended for use with AFP filesystems.
7663 ** And since AFP is only supported on MacOSX, the proxy locking is also
7664 ** restricted to MacOSX.
7667 ******************* End of the proxy lock implementation **********************
7668 ******************************************************************************/
7671 ** Initialize the operating system interface.
7673 ** This routine registers all VFS implementations for unix-like operating
7674 ** systems. This routine, and the sqlite3_os_end() routine that follows,
7675 ** should be the only routines in this file that are visible from other
7678 ** This routine is called once during SQLite initialization and by a
7679 ** single thread. The memory allocation and mutex subsystems have not
7680 ** necessarily been initialized when this routine is called, and so they
7681 ** should not be used.
7683 int sqlite3_os_init(void){
7685 ** The following macro defines an initializer for an sqlite3_vfs object.
7686 ** The name of the VFS is NAME. The pAppData is a pointer to a pointer
7687 ** to the "finder" function. (pAppData is a pointer to a pointer because
7688 ** silly C90 rules prohibit a void* from being cast to a function pointer
7689 ** and so we have to go through the intermediate pointer to avoid problems
7690 ** when compiling with -pedantic-errors on GCC.)
7692 ** The FINDER parameter to this macro is the name of the pointer to the
7693 ** finder-function. The finder-function returns a pointer to the
7694 ** sqlite_io_methods object that implements the desired locking
7695 ** behaviors. See the division above that contains the IOMETHODS
7696 ** macro for addition information on finder-functions.
7698 ** Most finders simply return a pointer to a fixed sqlite3_io_methods
7699 ** object. But the "autolockIoFinder" available on MacOSX does a little
7700 ** more than that; it looks at the filesystem type that hosts the
7701 ** database file and tries to choose an locking method appropriate for
7702 ** that filesystem time.
7704 #define UNIXVFS(VFSNAME, FINDER) { \
7706 sizeof(unixFile), /* szOsFile */ \
7707 MAX_PATHNAME, /* mxPathname */ \
7709 VFSNAME, /* zName */ \
7710 (void*)&FINDER, /* pAppData */ \
7711 unixOpen, /* xOpen */ \
7712 unixDelete, /* xDelete */ \
7713 unixAccess, /* xAccess */ \
7714 unixFullPathname, /* xFullPathname */ \
7715 unixDlOpen, /* xDlOpen */ \
7716 unixDlError, /* xDlError */ \
7717 unixDlSym, /* xDlSym */ \
7718 unixDlClose, /* xDlClose */ \
7719 unixRandomness, /* xRandomness */ \
7720 unixSleep, /* xSleep */ \
7721 unixCurrentTime, /* xCurrentTime */ \
7722 unixGetLastError, /* xGetLastError */ \
7723 unixCurrentTimeInt64, /* xCurrentTimeInt64 */ \
7724 unixSetSystemCall, /* xSetSystemCall */ \
7725 unixGetSystemCall, /* xGetSystemCall */ \
7726 unixNextSystemCall, /* xNextSystemCall */ \
7730 ** All default VFSes for unix are contained in the following array.
7732 ** Note that the sqlite3_vfs.pNext field of the VFS object is modified
7733 ** by the SQLite core when the VFS is registered. So the following
7734 ** array cannot be const.
7736 static sqlite3_vfs aVfs
[] = {
7737 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7738 UNIXVFS("unix", autolockIoFinder
),
7740 UNIXVFS("unix", vxworksIoFinder
),
7742 UNIXVFS("unix", posixIoFinder
),
7744 UNIXVFS("unix-none", nolockIoFinder
),
7745 UNIXVFS("unix-dotfile", dotlockIoFinder
),
7746 UNIXVFS("unix-excl", posixIoFinder
),
7748 UNIXVFS("unix-namedsem", semIoFinder
),
7750 #if SQLITE_ENABLE_LOCKING_STYLE || OS_VXWORKS
7751 UNIXVFS("unix-posix", posixIoFinder
),
7753 #if SQLITE_ENABLE_LOCKING_STYLE
7754 UNIXVFS("unix-flock", flockIoFinder
),
7756 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7757 UNIXVFS("unix-afp", afpIoFinder
),
7758 UNIXVFS("unix-nfs", nfsIoFinder
),
7759 UNIXVFS("unix-proxy", proxyIoFinder
),
7762 unsigned int i
; /* Loop counter */
7764 /* Double-check that the aSyscall[] array has been constructed
7765 ** correctly. See ticket [bb3a86e890c8e96ab] */
7766 assert( ArraySize(aSyscall
)==29 );
7768 /* Register all VFSes defined in the aVfs[] array */
7769 for(i
=0; i
<(sizeof(aVfs
)/sizeof(sqlite3_vfs
)); i
++){
7770 sqlite3_vfs_register(&aVfs
[i
], i
==0);
7772 unixBigLock
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_VFS1
);
7777 ** Shutdown the operating system interface.
7779 ** Some operating systems might need to do some cleanup in this routine,
7780 ** to release dynamically allocated objects. But not on unix.
7781 ** This routine is a no-op for unix.
7783 int sqlite3_os_end(void){
7788 #endif /* SQLITE_OS_UNIX */