4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 ******************************************************************************
13 ** This file contains the VFS implementation for unix-like operating systems
14 ** include Linux, MacOSX, *BSD, QNX, VxWorks, AIX, HPUX, and others.
16 ** There are actually several different VFS implementations in this file.
17 ** The differences are in the way that file locking is done. The default
18 ** implementation uses Posix Advisory Locks. Alternative implementations
19 ** use flock(), dot-files, various proprietary locking schemas, or simply
20 ** skip locking all together.
22 ** This source file is organized into divisions where the logic for various
23 ** subfunctions is contained within the appropriate division. PLEASE
24 ** KEEP THE STRUCTURE OF THIS FILE INTACT. New code should be placed
25 ** in the correct division and should be clearly labeled.
27 ** The layout of divisions is as follows:
29 ** * General-purpose declarations and utility functions.
30 ** * Unique file ID logic used by VxWorks.
31 ** * Various locking primitive implementations (all except proxy locking):
32 ** + for Posix Advisory Locks
34 ** + for dot-file locks
35 ** + for flock() locking
36 ** + for named semaphore locks (VxWorks only)
37 ** + for AFP filesystem locks (MacOSX only)
38 ** * sqlite3_file methods not associated with locking.
39 ** * Definitions of sqlite3_io_methods objects for all locking
40 ** methods plus "finder" functions for each locking method.
41 ** * sqlite3_vfs method implementations.
42 ** * Locking primitives for the proxy uber-locking-method. (MacOSX only)
43 ** * Definitions of sqlite3_vfs objects for all locking methods
44 ** plus implementations of sqlite3_os_init() and sqlite3_os_end().
46 #include "sqliteInt.h"
47 #if SQLITE_OS_UNIX /* This file is used on unix only */
50 ** There are various methods for file locking used for concurrency
53 ** 1. POSIX locking (the default),
55 ** 3. Dot-file locking,
56 ** 4. flock() locking,
57 ** 5. AFP locking (OSX only),
58 ** 6. Named POSIX semaphores (VXWorks only),
59 ** 7. proxy locking. (OSX only)
61 ** Styles 4, 5, and 7 are only available of SQLITE_ENABLE_LOCKING_STYLE
62 ** is defined to 1. The SQLITE_ENABLE_LOCKING_STYLE also enables automatic
63 ** selection of the appropriate locking style based on the filesystem
64 ** where the database is located.
66 #if !defined(SQLITE_ENABLE_LOCKING_STYLE)
67 # if defined(__APPLE__)
68 # define SQLITE_ENABLE_LOCKING_STYLE 1
70 # define SQLITE_ENABLE_LOCKING_STYLE 0
74 /* Use pread() and pwrite() if they are available */
75 #if defined(__APPLE__)
77 # define HAVE_PWRITE 1
79 #if defined(HAVE_PREAD64) && defined(HAVE_PWRITE64)
81 # define USE_PREAD64 1
82 #elif defined(HAVE_PREAD) && defined(HAVE_PWRITE)
88 ** standard include files.
90 #include <sys/types.h>
93 #include <sys/ioctl.h>
98 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
99 # include <sys/mman.h>
102 #if SQLITE_ENABLE_LOCKING_STYLE
103 # include <sys/ioctl.h>
104 # include <sys/file.h>
105 # include <sys/param.h>
106 #endif /* SQLITE_ENABLE_LOCKING_STYLE */
108 #if defined(__APPLE__) && ((__MAC_OS_X_VERSION_MIN_REQUIRED > 1050) || \
109 (__IPHONE_OS_VERSION_MIN_REQUIRED > 2000))
110 # if (!defined(TARGET_OS_EMBEDDED) || (TARGET_OS_EMBEDDED==0)) \
111 && (!defined(TARGET_IPHONE_SIMULATOR) || (TARGET_IPHONE_SIMULATOR==0))
112 # define HAVE_GETHOSTUUID 1
114 # warning "gethostuuid() is disabled."
120 # include <sys/ioctl.h>
121 # include <semaphore.h>
123 #endif /* OS_VXWORKS */
125 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
126 # include <sys/mount.h>
134 ** Allowed values of unixFile.fsFlags
136 #define SQLITE_FSFLAGS_IS_MSDOS 0x1
139 ** If we are to be thread-safe, include the pthreads header and define
140 ** the SQLITE_UNIX_THREADS macro.
142 #if SQLITE_THREADSAFE
143 # include <pthread.h>
144 # define SQLITE_UNIX_THREADS 1
148 ** Default permissions when creating a new file
150 #ifndef SQLITE_DEFAULT_FILE_PERMISSIONS
151 # define SQLITE_DEFAULT_FILE_PERMISSIONS 0644
155 ** Default permissions when creating auto proxy dir
157 #ifndef SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
158 # define SQLITE_DEFAULT_PROXYDIR_PERMISSIONS 0755
162 ** Maximum supported path-length.
164 #define MAX_PATHNAME 512
167 ** Maximum supported symbolic links
169 #define SQLITE_MAX_SYMLINKS 100
171 /* Always cast the getpid() return type for compatibility with
172 ** kernel modules in VxWorks. */
173 #define osGetpid(X) (pid_t)getpid()
176 ** Only set the lastErrno if the error code is a real error and not
177 ** a normal expected return code of SQLITE_BUSY or SQLITE_OK
179 #define IS_LOCK_ERROR(x) ((x != SQLITE_OK) && (x != SQLITE_BUSY))
181 /* Forward references */
182 typedef struct unixShm unixShm
; /* Connection shared memory */
183 typedef struct unixShmNode unixShmNode
; /* Shared memory instance */
184 typedef struct unixInodeInfo unixInodeInfo
; /* An i-node */
185 typedef struct UnixUnusedFd UnixUnusedFd
; /* An unused file descriptor */
188 ** Sometimes, after a file handle is closed by SQLite, the file descriptor
189 ** cannot be closed immediately. In these cases, instances of the following
190 ** structure are used to store the file descriptor while waiting for an
191 ** opportunity to either close or reuse it.
193 struct UnixUnusedFd
{
194 int fd
; /* File descriptor to close */
195 int flags
; /* Flags this file descriptor was opened with */
196 UnixUnusedFd
*pNext
; /* Next unused file descriptor on same file */
200 ** The unixFile structure is subclass of sqlite3_file specific to the unix
201 ** VFS implementations.
203 typedef struct unixFile unixFile
;
205 sqlite3_io_methods
const *pMethod
; /* Always the first entry */
206 sqlite3_vfs
*pVfs
; /* The VFS that created this unixFile */
207 unixInodeInfo
*pInode
; /* Info about locks on this inode */
208 int h
; /* The file descriptor */
209 unsigned char eFileLock
; /* The type of lock held on this fd */
210 unsigned short int ctrlFlags
; /* Behavioral bits. UNIXFILE_* flags */
211 int lastErrno
; /* The unix errno from last I/O error */
212 void *lockingContext
; /* Locking style specific state */
213 UnixUnusedFd
*pPreallocatedUnused
; /* Pre-allocated UnixUnusedFd */
214 const char *zPath
; /* Name of the file */
215 unixShm
*pShm
; /* Shared memory segment information */
216 int szChunk
; /* Configured by FCNTL_CHUNK_SIZE */
217 #if SQLITE_MAX_MMAP_SIZE>0
218 int nFetchOut
; /* Number of outstanding xFetch refs */
219 sqlite3_int64 mmapSize
; /* Usable size of mapping at pMapRegion */
220 sqlite3_int64 mmapSizeActual
; /* Actual size of mapping at pMapRegion */
221 sqlite3_int64 mmapSizeMax
; /* Configured FCNTL_MMAP_SIZE value */
222 void *pMapRegion
; /* Memory mapped region */
224 int sectorSize
; /* Device sector size */
225 int deviceCharacteristics
; /* Precomputed device characteristics */
226 #if SQLITE_ENABLE_LOCKING_STYLE
227 int openFlags
; /* The flags specified at open() */
229 #if SQLITE_ENABLE_LOCKING_STYLE || defined(__APPLE__)
230 unsigned fsFlags
; /* cached details from statfs() */
233 struct vxworksFileId
*pId
; /* Unique file ID */
236 /* The next group of variables are used to track whether or not the
237 ** transaction counter in bytes 24-27 of database files are updated
238 ** whenever any part of the database changes. An assertion fault will
239 ** occur if a file is updated without also updating the transaction
240 ** counter. This test is made to avoid new problems similar to the
241 ** one described by ticket #3584.
243 unsigned char transCntrChng
; /* True if the transaction counter changed */
244 unsigned char dbUpdate
; /* True if any part of database file changed */
245 unsigned char inNormalWrite
; /* True if in a normal write operation */
250 /* In test mode, increase the size of this structure a bit so that
251 ** it is larger than the struct CrashFile defined in test6.c.
257 /* This variable holds the process id (pid) from when the xRandomness()
258 ** method was called. If xOpen() is called from a different process id,
259 ** indicating that a fork() has occurred, the PRNG will be reset.
261 static pid_t randomnessPid
= 0;
264 ** Allowed values for the unixFile.ctrlFlags bitmask:
266 #define UNIXFILE_EXCL 0x01 /* Connections from one process only */
267 #define UNIXFILE_RDONLY 0x02 /* Connection is read only */
268 #define UNIXFILE_PERSIST_WAL 0x04 /* Persistent WAL mode */
269 #ifndef SQLITE_DISABLE_DIRSYNC
270 # define UNIXFILE_DIRSYNC 0x08 /* Directory sync needed */
272 # define UNIXFILE_DIRSYNC 0x00
274 #define UNIXFILE_PSOW 0x10 /* SQLITE_IOCAP_POWERSAFE_OVERWRITE */
275 #define UNIXFILE_DELETE 0x20 /* Delete on close */
276 #define UNIXFILE_URI 0x40 /* Filename might have query parameters */
277 #define UNIXFILE_NOLOCK 0x80 /* Do no file locking */
280 ** Include code that is common to all os_*.c files
282 #include "os_common.h"
285 ** Define various macros that are missing from some systems.
288 # define O_LARGEFILE 0
290 #ifdef SQLITE_DISABLE_LFS
292 # define O_LARGEFILE 0
295 # define O_NOFOLLOW 0
302 ** The threadid macro resolves to the thread-id or to 0. Used for
303 ** testing and debugging only.
305 #if SQLITE_THREADSAFE
306 #define threadid pthread_self()
312 ** HAVE_MREMAP defaults to true on Linux and false everywhere else.
314 #if !defined(HAVE_MREMAP)
315 # if defined(__linux__) && defined(_GNU_SOURCE)
316 # define HAVE_MREMAP 1
318 # define HAVE_MREMAP 0
323 ** Explicitly call the 64-bit version of lseek() on Android. Otherwise, lseek()
324 ** is the 32-bit version, even if _FILE_OFFSET_BITS=64 is defined.
327 # define lseek lseek64
332 ** Linux-specific IOCTL magic numbers used for controlling F2FS
334 #define F2FS_IOCTL_MAGIC 0xf5
335 #define F2FS_IOC_START_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 1)
336 #define F2FS_IOC_COMMIT_ATOMIC_WRITE _IO(F2FS_IOCTL_MAGIC, 2)
337 #define F2FS_IOC_START_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 3)
338 #define F2FS_IOC_ABORT_VOLATILE_WRITE _IO(F2FS_IOCTL_MAGIC, 5)
339 #define F2FS_IOC_GET_FEATURES _IOR(F2FS_IOCTL_MAGIC, 12, u32)
340 #define F2FS_FEATURE_ATOMIC_WRITE 0x0004
341 #endif /* __linux__ */
345 ** Different Unix systems declare open() in different ways. Same use
346 ** open(const char*,int,mode_t). Others use open(const char*,int,...).
347 ** The difference is important when using a pointer to the function.
349 ** The safest way to deal with the problem is to always use this wrapper
350 ** which always has the same well-defined interface.
352 static int posixOpen(const char *zFile
, int flags
, int mode
){
353 return open(zFile
, flags
, mode
);
356 /* Forward reference */
357 static int openDirectory(const char*, int*);
358 static int unixGetpagesize(void);
361 ** Many system calls are accessed through pointer-to-functions so that
362 ** they may be overridden at runtime to facilitate fault injection during
363 ** testing and sandboxing. The following array holds the names and pointers
364 ** to all overrideable system calls.
366 static struct unix_syscall
{
367 const char *zName
; /* Name of the system call */
368 sqlite3_syscall_ptr pCurrent
; /* Current value of the system call */
369 sqlite3_syscall_ptr pDefault
; /* Default value */
371 { "open", (sqlite3_syscall_ptr
)posixOpen
, 0 },
372 #define osOpen ((int(*)(const char*,int,int))aSyscall[0].pCurrent)
374 { "close", (sqlite3_syscall_ptr
)close
, 0 },
375 #define osClose ((int(*)(int))aSyscall[1].pCurrent)
377 { "access", (sqlite3_syscall_ptr
)access
, 0 },
378 #define osAccess ((int(*)(const char*,int))aSyscall[2].pCurrent)
380 { "getcwd", (sqlite3_syscall_ptr
)getcwd
, 0 },
381 #define osGetcwd ((char*(*)(char*,size_t))aSyscall[3].pCurrent)
383 { "stat", (sqlite3_syscall_ptr
)stat
, 0 },
384 #define osStat ((int(*)(const char*,struct stat*))aSyscall[4].pCurrent)
387 ** The DJGPP compiler environment looks mostly like Unix, but it
388 ** lacks the fcntl() system call. So redefine fcntl() to be something
389 ** that always succeeds. This means that locking does not occur under
390 ** DJGPP. But it is DOS - what did you expect?
394 #define osFstat(a,b,c) 0
396 { "fstat", (sqlite3_syscall_ptr
)fstat
, 0 },
397 #define osFstat ((int(*)(int,struct stat*))aSyscall[5].pCurrent)
400 { "ftruncate", (sqlite3_syscall_ptr
)ftruncate
, 0 },
401 #define osFtruncate ((int(*)(int,off_t))aSyscall[6].pCurrent)
403 { "fcntl", (sqlite3_syscall_ptr
)fcntl
, 0 },
404 #define osFcntl ((int(*)(int,int,...))aSyscall[7].pCurrent)
406 { "read", (sqlite3_syscall_ptr
)read
, 0 },
407 #define osRead ((ssize_t(*)(int,void*,size_t))aSyscall[8].pCurrent)
409 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
410 { "pread", (sqlite3_syscall_ptr
)pread
, 0 },
412 { "pread", (sqlite3_syscall_ptr
)0, 0 },
414 #define osPread ((ssize_t(*)(int,void*,size_t,off_t))aSyscall[9].pCurrent)
416 #if defined(USE_PREAD64)
417 { "pread64", (sqlite3_syscall_ptr
)pread64
, 0 },
419 { "pread64", (sqlite3_syscall_ptr
)0, 0 },
421 #define osPread64 ((ssize_t(*)(int,void*,size_t,off64_t))aSyscall[10].pCurrent)
423 { "write", (sqlite3_syscall_ptr
)write
, 0 },
424 #define osWrite ((ssize_t(*)(int,const void*,size_t))aSyscall[11].pCurrent)
426 #if defined(USE_PREAD) || SQLITE_ENABLE_LOCKING_STYLE
427 { "pwrite", (sqlite3_syscall_ptr
)pwrite
, 0 },
429 { "pwrite", (sqlite3_syscall_ptr
)0, 0 },
431 #define osPwrite ((ssize_t(*)(int,const void*,size_t,off_t))\
432 aSyscall[12].pCurrent)
434 #if defined(USE_PREAD64)
435 { "pwrite64", (sqlite3_syscall_ptr
)pwrite64
, 0 },
437 { "pwrite64", (sqlite3_syscall_ptr
)0, 0 },
439 #define osPwrite64 ((ssize_t(*)(int,const void*,size_t,off64_t))\
440 aSyscall[13].pCurrent)
442 { "fchmod", (sqlite3_syscall_ptr
)fchmod
, 0 },
443 #define osFchmod ((int(*)(int,mode_t))aSyscall[14].pCurrent)
445 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
446 { "fallocate", (sqlite3_syscall_ptr
)posix_fallocate
, 0 },
448 { "fallocate", (sqlite3_syscall_ptr
)0, 0 },
450 #define osFallocate ((int(*)(int,off_t,off_t))aSyscall[15].pCurrent)
452 { "unlink", (sqlite3_syscall_ptr
)unlink
, 0 },
453 #define osUnlink ((int(*)(const char*))aSyscall[16].pCurrent)
455 { "openDirectory", (sqlite3_syscall_ptr
)openDirectory
, 0 },
456 #define osOpenDirectory ((int(*)(const char*,int*))aSyscall[17].pCurrent)
458 { "mkdir", (sqlite3_syscall_ptr
)mkdir
, 0 },
459 #define osMkdir ((int(*)(const char*,mode_t))aSyscall[18].pCurrent)
461 { "rmdir", (sqlite3_syscall_ptr
)rmdir
, 0 },
462 #define osRmdir ((int(*)(const char*))aSyscall[19].pCurrent)
464 #if defined(HAVE_FCHOWN)
465 { "fchown", (sqlite3_syscall_ptr
)fchown
, 0 },
467 { "fchown", (sqlite3_syscall_ptr
)0, 0 },
469 #define osFchown ((int(*)(int,uid_t,gid_t))aSyscall[20].pCurrent)
471 { "geteuid", (sqlite3_syscall_ptr
)geteuid
, 0 },
472 #define osGeteuid ((uid_t(*)(void))aSyscall[21].pCurrent)
474 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
475 { "mmap", (sqlite3_syscall_ptr
)mmap
, 0 },
477 { "mmap", (sqlite3_syscall_ptr
)0, 0 },
479 #define osMmap ((void*(*)(void*,size_t,int,int,int,off_t))aSyscall[22].pCurrent)
481 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
482 { "munmap", (sqlite3_syscall_ptr
)munmap
, 0 },
484 { "munmap", (sqlite3_syscall_ptr
)0, 0 },
486 #define osMunmap ((void*(*)(void*,size_t))aSyscall[23].pCurrent)
488 #if HAVE_MREMAP && (!defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0)
489 { "mremap", (sqlite3_syscall_ptr
)mremap
, 0 },
491 { "mremap", (sqlite3_syscall_ptr
)0, 0 },
493 #define osMremap ((void*(*)(void*,size_t,size_t,int,...))aSyscall[24].pCurrent)
495 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
496 { "getpagesize", (sqlite3_syscall_ptr
)unixGetpagesize
, 0 },
498 { "getpagesize", (sqlite3_syscall_ptr
)0, 0 },
500 #define osGetpagesize ((int(*)(void))aSyscall[25].pCurrent)
502 #if defined(HAVE_READLINK)
503 { "readlink", (sqlite3_syscall_ptr
)readlink
, 0 },
505 { "readlink", (sqlite3_syscall_ptr
)0, 0 },
507 #define osReadlink ((ssize_t(*)(const char*,char*,size_t))aSyscall[26].pCurrent)
509 #if defined(HAVE_LSTAT)
510 { "lstat", (sqlite3_syscall_ptr
)lstat
, 0 },
512 { "lstat", (sqlite3_syscall_ptr
)0, 0 },
514 #define osLstat ((int(*)(const char*,struct stat*))aSyscall[27].pCurrent)
516 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
517 { "ioctl", (sqlite3_syscall_ptr
)ioctl
, 0 },
519 { "ioctl", (sqlite3_syscall_ptr
)0, 0 },
521 #define osIoctl ((int(*)(int,int,...))aSyscall[28].pCurrent)
523 }; /* End of the overrideable system calls */
527 ** On some systems, calls to fchown() will trigger a message in a security
528 ** log if they come from non-root processes. So avoid calling fchown() if
529 ** we are not running as root.
531 static int robustFchown(int fd
, uid_t uid
, gid_t gid
){
532 #if defined(HAVE_FCHOWN)
533 return osGeteuid() ? 0 : osFchown(fd
,uid
,gid
);
540 ** This is the xSetSystemCall() method of sqlite3_vfs for all of the
541 ** "unix" VFSes. Return SQLITE_OK opon successfully updating the
542 ** system call pointer, or SQLITE_NOTFOUND if there is no configurable
543 ** system call named zName.
545 static int unixSetSystemCall(
546 sqlite3_vfs
*pNotUsed
, /* The VFS pointer. Not used */
547 const char *zName
, /* Name of system call to override */
548 sqlite3_syscall_ptr pNewFunc
/* Pointer to new system call value */
551 int rc
= SQLITE_NOTFOUND
;
553 UNUSED_PARAMETER(pNotUsed
);
555 /* If no zName is given, restore all system calls to their default
556 ** settings and return NULL
559 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
560 if( aSyscall
[i
].pDefault
){
561 aSyscall
[i
].pCurrent
= aSyscall
[i
].pDefault
;
565 /* If zName is specified, operate on only the one system call
568 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
569 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ){
570 if( aSyscall
[i
].pDefault
==0 ){
571 aSyscall
[i
].pDefault
= aSyscall
[i
].pCurrent
;
574 if( pNewFunc
==0 ) pNewFunc
= aSyscall
[i
].pDefault
;
575 aSyscall
[i
].pCurrent
= pNewFunc
;
584 ** Return the value of a system call. Return NULL if zName is not a
585 ** recognized system call name. NULL is also returned if the system call
586 ** is currently undefined.
588 static sqlite3_syscall_ptr
unixGetSystemCall(
589 sqlite3_vfs
*pNotUsed
,
594 UNUSED_PARAMETER(pNotUsed
);
595 for(i
=0; i
<sizeof(aSyscall
)/sizeof(aSyscall
[0]); i
++){
596 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) return aSyscall
[i
].pCurrent
;
602 ** Return the name of the first system call after zName. If zName==NULL
603 ** then return the name of the first system call. Return NULL if zName
604 ** is the last system call or if zName is not the name of a valid
607 static const char *unixNextSystemCall(sqlite3_vfs
*p
, const char *zName
){
612 for(i
=0; i
<ArraySize(aSyscall
)-1; i
++){
613 if( strcmp(zName
, aSyscall
[i
].zName
)==0 ) break;
616 for(i
++; i
<ArraySize(aSyscall
); i
++){
617 if( aSyscall
[i
].pCurrent
!=0 ) return aSyscall
[i
].zName
;
623 ** Do not accept any file descriptor less than this value, in order to avoid
624 ** opening database file using file descriptors that are commonly used for
625 ** standard input, output, and error.
627 #ifndef SQLITE_MINIMUM_FILE_DESCRIPTOR
628 # define SQLITE_MINIMUM_FILE_DESCRIPTOR 3
632 ** Invoke open(). Do so multiple times, until it either succeeds or
633 ** fails for some reason other than EINTR.
635 ** If the file creation mode "m" is 0 then set it to the default for
636 ** SQLite. The default is SQLITE_DEFAULT_FILE_PERMISSIONS (normally
637 ** 0644) as modified by the system umask. If m is not 0, then
638 ** make the file creation mode be exactly m ignoring the umask.
640 ** The m parameter will be non-zero only when creating -wal, -journal,
641 ** and -shm files. We want those files to have *exactly* the same
642 ** permissions as their original database, unadulterated by the umask.
643 ** In that way, if a database file is -rw-rw-rw or -rw-rw-r-, and a
644 ** transaction crashes and leaves behind hot journals, then any
645 ** process that is able to write to the database will also be able to
646 ** recover the hot journals.
648 static int robust_open(const char *z
, int f
, mode_t m
){
650 mode_t m2
= m
? m
: SQLITE_DEFAULT_FILE_PERMISSIONS
;
652 #if defined(O_CLOEXEC)
653 fd
= osOpen(z
,f
|O_CLOEXEC
,m2
);
658 if( errno
==EINTR
) continue;
661 if( fd
>=SQLITE_MINIMUM_FILE_DESCRIPTOR
) break;
663 sqlite3_log(SQLITE_WARNING
,
664 "attempt to open \"%s\" as file descriptor %d", z
, fd
);
666 if( osOpen("/dev/null", f
, m
)<0 ) break;
671 if( osFstat(fd
, &statbuf
)==0
672 && statbuf
.st_size
==0
673 && (statbuf
.st_mode
&0777)!=m
678 #if defined(FD_CLOEXEC) && (!defined(O_CLOEXEC) || O_CLOEXEC==0)
679 osFcntl(fd
, F_SETFD
, osFcntl(fd
, F_GETFD
, 0) | FD_CLOEXEC
);
686 ** Helper functions to obtain and relinquish the global mutex. The
687 ** global mutex is used to protect the unixInodeInfo and
688 ** vxworksFileId objects used by this file, all of which may be
689 ** shared by multiple threads.
691 ** Function unixMutexHeld() is used to assert() that the global mutex
692 ** is held when required. This function is only used as part of assert()
696 ** assert( unixMutexHeld() );
699 static void unixEnterMutex(void){
700 sqlite3_mutex_enter(sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_VFS1
));
702 static void unixLeaveMutex(void){
703 sqlite3_mutex_leave(sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_VFS1
));
706 static int unixMutexHeld(void) {
707 return sqlite3_mutex_held(sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_VFS1
));
712 #ifdef SQLITE_HAVE_OS_TRACE
714 ** Helper function for printing out trace information from debugging
715 ** binaries. This returns the string representation of the supplied
716 ** integer lock-type.
718 static const char *azFileLock(int eFileLock
){
720 case NO_LOCK
: return "NONE";
721 case SHARED_LOCK
: return "SHARED";
722 case RESERVED_LOCK
: return "RESERVED";
723 case PENDING_LOCK
: return "PENDING";
724 case EXCLUSIVE_LOCK
: return "EXCLUSIVE";
730 #ifdef SQLITE_LOCK_TRACE
732 ** Print out information about all locking operations.
734 ** This routine is used for troubleshooting locks on multithreaded
735 ** platforms. Enable by compiling with the -DSQLITE_LOCK_TRACE
736 ** command-line option on the compiler. This code is normally
739 static int lockTrace(int fd
, int op
, struct flock
*p
){
740 char *zOpName
, *zType
;
745 }else if( op
==F_SETLK
){
748 s
= osFcntl(fd
, op
, p
);
749 sqlite3DebugPrintf("fcntl unknown %d %d %d\n", fd
, op
, s
);
752 if( p
->l_type
==F_RDLCK
){
754 }else if( p
->l_type
==F_WRLCK
){
756 }else if( p
->l_type
==F_UNLCK
){
761 assert( p
->l_whence
==SEEK_SET
);
762 s
= osFcntl(fd
, op
, p
);
764 sqlite3DebugPrintf("fcntl %d %d %s %s %d %d %d %d\n",
765 threadid
, fd
, zOpName
, zType
, (int)p
->l_start
, (int)p
->l_len
,
767 if( s
==(-1) && op
==F_SETLK
&& (p
->l_type
==F_RDLCK
|| p
->l_type
==F_WRLCK
) ){
770 osFcntl(fd
, F_GETLK
, &l2
);
771 if( l2
.l_type
==F_RDLCK
){
773 }else if( l2
.l_type
==F_WRLCK
){
775 }else if( l2
.l_type
==F_UNLCK
){
780 sqlite3DebugPrintf("fcntl-failure-reason: %s %d %d %d\n",
781 zType
, (int)l2
.l_start
, (int)l2
.l_len
, (int)l2
.l_pid
);
787 #define osFcntl lockTrace
788 #endif /* SQLITE_LOCK_TRACE */
791 ** Retry ftruncate() calls that fail due to EINTR
793 ** All calls to ftruncate() within this file should be made through
794 ** this wrapper. On the Android platform, bypassing the logic below
795 ** could lead to a corrupt database.
797 static int robust_ftruncate(int h
, sqlite3_int64 sz
){
800 /* On Android, ftruncate() always uses 32-bit offsets, even if
801 ** _FILE_OFFSET_BITS=64 is defined. This means it is unsafe to attempt to
802 ** truncate a file to any size larger than 2GiB. Silently ignore any
804 if( sz
>(sqlite3_int64
)0x7FFFFFFF ){
808 do{ rc
= osFtruncate(h
,sz
); }while( rc
<0 && errno
==EINTR
);
813 ** This routine translates a standard POSIX errno code into something
814 ** useful to the clients of the sqlite3 functions. Specifically, it is
815 ** intended to translate a variety of "try again" errors into SQLITE_BUSY
816 ** and a variety of "please close the file descriptor NOW" errors into
819 ** Errors during initialization of locks, or file system support for locks,
820 ** should handle ENOLCK, ENOTSUP, EOPNOTSUPP separately.
822 static int sqliteErrorFromPosixError(int posixError
, int sqliteIOErr
) {
823 assert( (sqliteIOErr
== SQLITE_IOERR_LOCK
) ||
824 (sqliteIOErr
== SQLITE_IOERR_UNLOCK
) ||
825 (sqliteIOErr
== SQLITE_IOERR_RDLOCK
) ||
826 (sqliteIOErr
== SQLITE_IOERR_CHECKRESERVEDLOCK
) );
827 switch (posixError
) {
834 /* random NFS retry error, unless during file system support
835 * introspection, in which it actually means what it says */
847 /******************************************************************************
848 ****************** Begin Unique File ID Utility Used By VxWorks ***************
850 ** On most versions of unix, we can get a unique ID for a file by concatenating
851 ** the device number and the inode number. But this does not work on VxWorks.
852 ** On VxWorks, a unique file id must be based on the canonical filename.
854 ** A pointer to an instance of the following structure can be used as a
855 ** unique file ID in VxWorks. Each instance of this structure contains
856 ** a copy of the canonical filename. There is also a reference count.
857 ** The structure is reclaimed when the number of pointers to it drops to
860 ** There are never very many files open at one time and lookups are not
861 ** a performance-critical path, so it is sufficient to put these
862 ** structures on a linked list.
864 struct vxworksFileId
{
865 struct vxworksFileId
*pNext
; /* Next in a list of them all */
866 int nRef
; /* Number of references to this one */
867 int nName
; /* Length of the zCanonicalName[] string */
868 char *zCanonicalName
; /* Canonical filename */
873 ** All unique filenames are held on a linked list headed by this
876 static struct vxworksFileId
*vxworksFileList
= 0;
879 ** Simplify a filename into its canonical form
880 ** by making the following changes:
882 ** * removing any trailing and duplicate /
883 ** * convert /./ into just /
884 ** * convert /A/../ where A is any simple name into just /
886 ** Changes are made in-place. Return the new name length.
888 ** The original filename is in z[0..n-1]. Return the number of
889 ** characters in the simplified name.
891 static int vxworksSimplifyName(char *z
, int n
){
893 while( n
>1 && z
[n
-1]=='/' ){ n
--; }
894 for(i
=j
=0; i
<n
; i
++){
896 if( z
[i
+1]=='/' ) continue;
897 if( z
[i
+1]=='.' && i
+2<n
&& z
[i
+2]=='/' ){
901 if( z
[i
+1]=='.' && i
+3<n
&& z
[i
+2]=='.' && z
[i
+3]=='/' ){
902 while( j
>0 && z
[j
-1]!='/' ){ j
--; }
915 ** Find a unique file ID for the given absolute pathname. Return
916 ** a pointer to the vxworksFileId object. This pointer is the unique
919 ** The nRef field of the vxworksFileId object is incremented before
920 ** the object is returned. A new vxworksFileId object is created
921 ** and added to the global list if necessary.
923 ** If a memory allocation error occurs, return NULL.
925 static struct vxworksFileId
*vxworksFindFileId(const char *zAbsoluteName
){
926 struct vxworksFileId
*pNew
; /* search key and new file ID */
927 struct vxworksFileId
*pCandidate
; /* For looping over existing file IDs */
928 int n
; /* Length of zAbsoluteName string */
930 assert( zAbsoluteName
[0]=='/' );
931 n
= (int)strlen(zAbsoluteName
);
932 pNew
= sqlite3_malloc64( sizeof(*pNew
) + (n
+1) );
933 if( pNew
==0 ) return 0;
934 pNew
->zCanonicalName
= (char*)&pNew
[1];
935 memcpy(pNew
->zCanonicalName
, zAbsoluteName
, n
+1);
936 n
= vxworksSimplifyName(pNew
->zCanonicalName
, n
);
938 /* Search for an existing entry that matching the canonical name.
939 ** If found, increment the reference count and return a pointer to
940 ** the existing file ID.
943 for(pCandidate
=vxworksFileList
; pCandidate
; pCandidate
=pCandidate
->pNext
){
944 if( pCandidate
->nName
==n
945 && memcmp(pCandidate
->zCanonicalName
, pNew
->zCanonicalName
, n
)==0
954 /* No match was found. We will make a new file ID */
957 pNew
->pNext
= vxworksFileList
;
958 vxworksFileList
= pNew
;
964 ** Decrement the reference count on a vxworksFileId object. Free
965 ** the object when the reference count reaches zero.
967 static void vxworksReleaseFileId(struct vxworksFileId
*pId
){
969 assert( pId
->nRef
>0 );
972 struct vxworksFileId
**pp
;
973 for(pp
=&vxworksFileList
; *pp
&& *pp
!=pId
; pp
= &((*pp
)->pNext
)){}
980 #endif /* OS_VXWORKS */
981 /*************** End of Unique File ID Utility Used By VxWorks ****************
982 ******************************************************************************/
985 /******************************************************************************
986 *************************** Posix Advisory Locking ****************************
988 ** POSIX advisory locks are broken by design. ANSI STD 1003.1 (1996)
989 ** section 6.5.2.2 lines 483 through 490 specify that when a process
990 ** sets or clears a lock, that operation overrides any prior locks set
991 ** by the same process. It does not explicitly say so, but this implies
992 ** that it overrides locks set by the same process using a different
993 ** file descriptor. Consider this test case:
995 ** int fd1 = open("./file1", O_RDWR|O_CREAT, 0644);
996 ** int fd2 = open("./file2", O_RDWR|O_CREAT, 0644);
998 ** Suppose ./file1 and ./file2 are really the same file (because
999 ** one is a hard or symbolic link to the other) then if you set
1000 ** an exclusive lock on fd1, then try to get an exclusive lock
1001 ** on fd2, it works. I would have expected the second lock to
1002 ** fail since there was already a lock on the file due to fd1.
1003 ** But not so. Since both locks came from the same process, the
1004 ** second overrides the first, even though they were on different
1005 ** file descriptors opened on different file names.
1007 ** This means that we cannot use POSIX locks to synchronize file access
1008 ** among competing threads of the same process. POSIX locks will work fine
1009 ** to synchronize access for threads in separate processes, but not
1010 ** threads within the same process.
1012 ** To work around the problem, SQLite has to manage file locks internally
1013 ** on its own. Whenever a new database is opened, we have to find the
1014 ** specific inode of the database file (the inode is determined by the
1015 ** st_dev and st_ino fields of the stat structure that fstat() fills in)
1016 ** and check for locks already existing on that inode. When locks are
1017 ** created or removed, we have to look at our own internal record of the
1018 ** locks to see if another thread has previously set a lock on that same
1021 ** (Aside: The use of inode numbers as unique IDs does not work on VxWorks.
1022 ** For VxWorks, we have to use the alternative unique ID system based on
1023 ** canonical filename and implemented in the previous division.)
1025 ** The sqlite3_file structure for POSIX is no longer just an integer file
1026 ** descriptor. It is now a structure that holds the integer file
1027 ** descriptor and a pointer to a structure that describes the internal
1028 ** locks on the corresponding inode. There is one locking structure
1029 ** per inode, so if the same inode is opened twice, both unixFile structures
1030 ** point to the same locking structure. The locking structure keeps
1031 ** a reference count (so we will know when to delete it) and a "cnt"
1032 ** field that tells us its internal lock status. cnt==0 means the
1033 ** file is unlocked. cnt==-1 means the file has an exclusive lock.
1034 ** cnt>0 means there are cnt shared locks on the file.
1036 ** Any attempt to lock or unlock a file first checks the locking
1037 ** structure. The fcntl() system call is only invoked to set a
1038 ** POSIX lock if the internal lock structure transitions between
1039 ** a locked and an unlocked state.
1041 ** But wait: there are yet more problems with POSIX advisory locks.
1043 ** If you close a file descriptor that points to a file that has locks,
1044 ** all locks on that file that are owned by the current process are
1045 ** released. To work around this problem, each unixInodeInfo object
1046 ** maintains a count of the number of pending locks on tha inode.
1047 ** When an attempt is made to close an unixFile, if there are
1048 ** other unixFile open on the same inode that are holding locks, the call
1049 ** to close() the file descriptor is deferred until all of the locks clear.
1050 ** The unixInodeInfo structure keeps a list of file descriptors that need to
1051 ** be closed and that list is walked (and cleared) when the last lock
1054 ** Yet another problem: LinuxThreads do not play well with posix locks.
1056 ** Many older versions of linux use the LinuxThreads library which is
1057 ** not posix compliant. Under LinuxThreads, a lock created by thread
1058 ** A cannot be modified or overridden by a different thread B.
1059 ** Only thread A can modify the lock. Locking behavior is correct
1060 ** if the appliation uses the newer Native Posix Thread Library (NPTL)
1061 ** on linux - with NPTL a lock created by thread A can override locks
1062 ** in thread B. But there is no way to know at compile-time which
1063 ** threading library is being used. So there is no way to know at
1064 ** compile-time whether or not thread A can override locks on thread B.
1065 ** One has to do a run-time check to discover the behavior of the
1068 ** SQLite used to support LinuxThreads. But support for LinuxThreads
1069 ** was dropped beginning with version 3.7.0. SQLite will still work with
1070 ** LinuxThreads provided that (1) there is no more than one connection
1071 ** per database file in the same process and (2) database connections
1072 ** do not move across threads.
1076 ** An instance of the following structure serves as the key used
1077 ** to locate a particular unixInodeInfo object.
1080 dev_t dev
; /* Device number */
1082 struct vxworksFileId
*pId
; /* Unique file ID for vxworks. */
1084 /* We are told that some versions of Android contain a bug that
1085 ** sizes ino_t at only 32-bits instead of 64-bits. (See
1086 ** https://android-review.googlesource.com/#/c/115351/3/dist/sqlite3.c)
1087 ** To work around this, always allocate 64-bits for the inode number.
1088 ** On small machines that only have 32-bit inodes, this wastes 4 bytes,
1089 ** but that should not be a big deal. */
1090 /* WAS: ino_t ino; */
1091 u64 ino
; /* Inode number */
1096 ** An instance of the following structure is allocated for each open
1097 ** inode. Or, on LinuxThreads, there is one of these structures for
1098 ** each inode opened by each thread.
1100 ** A single inode can have multiple file descriptors, so each unixFile
1101 ** structure contains a pointer to an instance of this object and this
1102 ** object keeps a count of the number of unixFile pointing to it.
1104 struct unixInodeInfo
{
1105 struct unixFileId fileId
; /* The lookup key */
1106 int nShared
; /* Number of SHARED locks held */
1107 unsigned char eFileLock
; /* One of SHARED_LOCK, RESERVED_LOCK etc. */
1108 unsigned char bProcessLock
; /* An exclusive process lock is held */
1109 int nRef
; /* Number of pointers to this structure */
1110 unixShmNode
*pShmNode
; /* Shared memory associated with this inode */
1111 int nLock
; /* Number of outstanding file locks */
1112 UnixUnusedFd
*pUnused
; /* Unused file descriptors to close */
1113 unixInodeInfo
*pNext
; /* List of all unixInodeInfo objects */
1114 unixInodeInfo
*pPrev
; /* .... doubly linked */
1115 #if SQLITE_ENABLE_LOCKING_STYLE
1116 unsigned long long sharedByte
; /* for AFP simulated shared lock */
1119 sem_t
*pSem
; /* Named POSIX semaphore */
1120 char aSemName
[MAX_PATHNAME
+2]; /* Name of that semaphore */
1125 ** A lists of all unixInodeInfo objects.
1127 static unixInodeInfo
*inodeList
= 0; /* All unixInodeInfo objects */
1128 static unsigned int nUnusedFd
= 0; /* Total unused file descriptors */
1132 ** This function - unixLogErrorAtLine(), is only ever called via the macro
1135 ** It is invoked after an error occurs in an OS function and errno has been
1136 ** set. It logs a message using sqlite3_log() containing the current value of
1137 ** errno and, if possible, the human-readable equivalent from strerror() or
1140 ** The first argument passed to the macro should be the error code that
1141 ** will be returned to SQLite (e.g. SQLITE_IOERR_DELETE, SQLITE_CANTOPEN).
1142 ** The two subsequent arguments should be the name of the OS function that
1143 ** failed (e.g. "unlink", "open") and the associated file-system path,
1146 #define unixLogError(a,b,c) unixLogErrorAtLine(a,b,c,__LINE__)
1147 static int unixLogErrorAtLine(
1148 int errcode
, /* SQLite error code */
1149 const char *zFunc
, /* Name of OS function that failed */
1150 const char *zPath
, /* File path associated with error */
1151 int iLine
/* Source line number where error occurred */
1153 char *zErr
; /* Message from strerror() or equivalent */
1154 int iErrno
= errno
; /* Saved syscall error number */
1156 /* If this is not a threadsafe build (SQLITE_THREADSAFE==0), then use
1157 ** the strerror() function to obtain the human-readable error message
1158 ** equivalent to errno. Otherwise, use strerror_r().
1160 #if SQLITE_THREADSAFE && defined(HAVE_STRERROR_R)
1162 memset(aErr
, 0, sizeof(aErr
));
1165 /* If STRERROR_R_CHAR_P (set by autoconf scripts) or __USE_GNU is defined,
1166 ** assume that the system provides the GNU version of strerror_r() that
1167 ** returns a pointer to a buffer containing the error message. That pointer
1168 ** may point to aErr[], or it may point to some static storage somewhere.
1169 ** Otherwise, assume that the system provides the POSIX version of
1170 ** strerror_r(), which always writes an error message into aErr[].
1172 ** If the code incorrectly assumes that it is the POSIX version that is
1173 ** available, the error message will often be an empty string. Not a
1174 ** huge problem. Incorrectly concluding that the GNU version is available
1175 ** could lead to a segfault though.
1177 #if defined(STRERROR_R_CHAR_P) || defined(__USE_GNU)
1180 strerror_r(iErrno
, aErr
, sizeof(aErr
)-1);
1182 #elif SQLITE_THREADSAFE
1183 /* This is a threadsafe build, but strerror_r() is not available. */
1186 /* Non-threadsafe build, use strerror(). */
1187 zErr
= strerror(iErrno
);
1190 if( zPath
==0 ) zPath
= "";
1191 sqlite3_log(errcode
,
1192 "os_unix.c:%d: (%d) %s(%s) - %s",
1193 iLine
, iErrno
, zFunc
, zPath
, zErr
1200 ** Close a file descriptor.
1202 ** We assume that close() almost always works, since it is only in a
1203 ** very sick application or on a very sick platform that it might fail.
1204 ** If it does fail, simply leak the file descriptor, but do log the
1207 ** Note that it is not safe to retry close() after EINTR since the
1208 ** file descriptor might have already been reused by another thread.
1209 ** So we don't even try to recover from an EINTR. Just log the error
1212 static void robust_close(unixFile
*pFile
, int h
, int lineno
){
1214 unixLogErrorAtLine(SQLITE_IOERR_CLOSE
, "close",
1215 pFile
? pFile
->zPath
: 0, lineno
);
1220 ** Set the pFile->lastErrno. Do this in a subroutine as that provides
1221 ** a convenient place to set a breakpoint.
1223 static void storeLastErrno(unixFile
*pFile
, int error
){
1224 pFile
->lastErrno
= error
;
1228 ** Close all file descriptors accumuated in the unixInodeInfo->pUnused list.
1230 static void closePendingFds(unixFile
*pFile
){
1231 unixInodeInfo
*pInode
= pFile
->pInode
;
1233 UnixUnusedFd
*pNext
;
1234 for(p
=pInode
->pUnused
; p
; p
=pNext
){
1236 robust_close(pFile
, p
->fd
, __LINE__
);
1240 pInode
->pUnused
= 0;
1244 ** Release a unixInodeInfo structure previously allocated by findInodeInfo().
1246 ** The mutex entered using the unixEnterMutex() function must be held
1247 ** when this function is called.
1249 static void releaseInodeInfo(unixFile
*pFile
){
1250 unixInodeInfo
*pInode
= pFile
->pInode
;
1251 assert( unixMutexHeld() );
1252 if( ALWAYS(pInode
) ){
1254 if( pInode
->nRef
==0 ){
1255 assert( pInode
->pShmNode
==0 );
1256 closePendingFds(pFile
);
1257 if( pInode
->pPrev
){
1258 assert( pInode
->pPrev
->pNext
==pInode
);
1259 pInode
->pPrev
->pNext
= pInode
->pNext
;
1261 assert( inodeList
==pInode
);
1262 inodeList
= pInode
->pNext
;
1264 if( pInode
->pNext
){
1265 assert( pInode
->pNext
->pPrev
==pInode
);
1266 pInode
->pNext
->pPrev
= pInode
->pPrev
;
1268 sqlite3_free(pInode
);
1271 assert( inodeList
!=0 || nUnusedFd
==0 );
1275 ** Given a file descriptor, locate the unixInodeInfo object that
1276 ** describes that file descriptor. Create a new one if necessary. The
1277 ** return value might be uninitialized if an error occurs.
1279 ** The mutex entered using the unixEnterMutex() function must be held
1280 ** when this function is called.
1282 ** Return an appropriate error code.
1284 static int findInodeInfo(
1285 unixFile
*pFile
, /* Unix file with file desc used in the key */
1286 unixInodeInfo
**ppInode
/* Return the unixInodeInfo object here */
1288 int rc
; /* System call return code */
1289 int fd
; /* The file descriptor for pFile */
1290 struct unixFileId fileId
; /* Lookup key for the unixInodeInfo */
1291 struct stat statbuf
; /* Low-level file information */
1292 unixInodeInfo
*pInode
= 0; /* Candidate unixInodeInfo object */
1294 assert( unixMutexHeld() );
1296 /* Get low-level information about the file that we can used to
1297 ** create a unique name for the file.
1300 rc
= osFstat(fd
, &statbuf
);
1302 storeLastErrno(pFile
, errno
);
1303 #if defined(EOVERFLOW) && defined(SQLITE_DISABLE_LFS)
1304 if( pFile
->lastErrno
==EOVERFLOW
) return SQLITE_NOLFS
;
1306 return SQLITE_IOERR
;
1310 /* On OS X on an msdos filesystem, the inode number is reported
1311 ** incorrectly for zero-size files. See ticket #3260. To work
1312 ** around this problem (we consider it a bug in OS X, not SQLite)
1313 ** we always increase the file size to 1 by writing a single byte
1314 ** prior to accessing the inode number. The one byte written is
1315 ** an ASCII 'S' character which also happens to be the first byte
1316 ** in the header of every SQLite database. In this way, if there
1317 ** is a race condition such that another thread has already populated
1318 ** the first page of the database, no damage is done.
1320 if( statbuf
.st_size
==0 && (pFile
->fsFlags
& SQLITE_FSFLAGS_IS_MSDOS
)!=0 ){
1321 do{ rc
= osWrite(fd
, "S", 1); }while( rc
<0 && errno
==EINTR
);
1323 storeLastErrno(pFile
, errno
);
1324 return SQLITE_IOERR
;
1326 rc
= osFstat(fd
, &statbuf
);
1328 storeLastErrno(pFile
, errno
);
1329 return SQLITE_IOERR
;
1334 memset(&fileId
, 0, sizeof(fileId
));
1335 fileId
.dev
= statbuf
.st_dev
;
1337 fileId
.pId
= pFile
->pId
;
1339 fileId
.ino
= (u64
)statbuf
.st_ino
;
1341 assert( inodeList
!=0 || nUnusedFd
==0 );
1343 while( pInode
&& memcmp(&fileId
, &pInode
->fileId
, sizeof(fileId
)) ){
1344 pInode
= pInode
->pNext
;
1347 pInode
= sqlite3_malloc64( sizeof(*pInode
) );
1349 return SQLITE_NOMEM_BKPT
;
1351 memset(pInode
, 0, sizeof(*pInode
));
1352 memcpy(&pInode
->fileId
, &fileId
, sizeof(fileId
));
1354 pInode
->pNext
= inodeList
;
1356 if( inodeList
) inodeList
->pPrev
= pInode
;
1366 ** Return TRUE if pFile has been renamed or unlinked since it was first opened.
1368 static int fileHasMoved(unixFile
*pFile
){
1370 return pFile
->pInode
!=0 && pFile
->pId
!=pFile
->pInode
->fileId
.pId
;
1373 return pFile
->pInode
!=0 &&
1374 (osStat(pFile
->zPath
, &buf
)!=0
1375 || (u64
)buf
.st_ino
!=pFile
->pInode
->fileId
.ino
);
1381 ** Check a unixFile that is a database. Verify the following:
1383 ** (1) There is exactly one hard link on the file
1384 ** (2) The file is not a symbolic link
1385 ** (3) The file has not been renamed or unlinked
1387 ** Issue sqlite3_log(SQLITE_WARNING,...) messages if anything is not right.
1389 static void verifyDbFile(unixFile
*pFile
){
1393 /* These verifications occurs for the main database only */
1394 if( pFile
->ctrlFlags
& UNIXFILE_NOLOCK
) return;
1396 rc
= osFstat(pFile
->h
, &buf
);
1398 sqlite3_log(SQLITE_WARNING
, "cannot fstat db file %s", pFile
->zPath
);
1401 if( buf
.st_nlink
==0 ){
1402 sqlite3_log(SQLITE_WARNING
, "file unlinked while open: %s", pFile
->zPath
);
1405 if( buf
.st_nlink
>1 ){
1406 sqlite3_log(SQLITE_WARNING
, "multiple links to file: %s", pFile
->zPath
);
1409 if( fileHasMoved(pFile
) ){
1410 sqlite3_log(SQLITE_WARNING
, "file renamed while open: %s", pFile
->zPath
);
1417 ** This routine checks if there is a RESERVED lock held on the specified
1418 ** file by this or any other process. If such a lock is held, set *pResOut
1419 ** to a non-zero value otherwise *pResOut is set to zero. The return value
1420 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
1422 static int unixCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
1425 unixFile
*pFile
= (unixFile
*)id
;
1427 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
1430 assert( pFile
->eFileLock
<=SHARED_LOCK
);
1431 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
1433 /* Check if a thread in this process holds such a lock */
1434 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
1438 /* Otherwise see if some other process holds it.
1441 if( !reserved
&& !pFile
->pInode
->bProcessLock
){
1443 lock
.l_whence
= SEEK_SET
;
1444 lock
.l_start
= RESERVED_BYTE
;
1446 lock
.l_type
= F_WRLCK
;
1447 if( osFcntl(pFile
->h
, F_GETLK
, &lock
) ){
1448 rc
= SQLITE_IOERR_CHECKRESERVEDLOCK
;
1449 storeLastErrno(pFile
, errno
);
1450 } else if( lock
.l_type
!=F_UNLCK
){
1457 OSTRACE(("TEST WR-LOCK %d %d %d (unix)\n", pFile
->h
, rc
, reserved
));
1459 *pResOut
= reserved
;
1464 ** Attempt to set a system-lock on the file pFile. The lock is
1465 ** described by pLock.
1467 ** If the pFile was opened read/write from unix-excl, then the only lock
1468 ** ever obtained is an exclusive lock, and it is obtained exactly once
1469 ** the first time any lock is attempted. All subsequent system locking
1470 ** operations become no-ops. Locking operations still happen internally,
1471 ** in order to coordinate access between separate database connections
1472 ** within this process, but all of that is handled in memory and the
1473 ** operating system does not participate.
1475 ** This function is a pass-through to fcntl(F_SETLK) if pFile is using
1476 ** any VFS other than "unix-excl" or if pFile is opened on "unix-excl"
1477 ** and is read-only.
1479 ** Zero is returned if the call completes successfully, or -1 if a call
1480 ** to fcntl() fails. In this case, errno is set appropriately (by fcntl()).
1482 static int unixFileLock(unixFile
*pFile
, struct flock
*pLock
){
1484 unixInodeInfo
*pInode
= pFile
->pInode
;
1485 assert( unixMutexHeld() );
1486 assert( pInode
!=0 );
1487 if( (pFile
->ctrlFlags
& (UNIXFILE_EXCL
|UNIXFILE_RDONLY
))==UNIXFILE_EXCL
){
1488 if( pInode
->bProcessLock
==0 ){
1490 assert( pInode
->nLock
==0 );
1491 lock
.l_whence
= SEEK_SET
;
1492 lock
.l_start
= SHARED_FIRST
;
1493 lock
.l_len
= SHARED_SIZE
;
1494 lock
.l_type
= F_WRLCK
;
1495 rc
= osFcntl(pFile
->h
, F_SETLK
, &lock
);
1496 if( rc
<0 ) return rc
;
1497 pInode
->bProcessLock
= 1;
1503 rc
= osFcntl(pFile
->h
, F_SETLK
, pLock
);
1509 ** Lock the file with the lock specified by parameter eFileLock - one
1510 ** of the following:
1513 ** (2) RESERVED_LOCK
1515 ** (4) EXCLUSIVE_LOCK
1517 ** Sometimes when requesting one lock state, additional lock states
1518 ** are inserted in between. The locking might fail on one of the later
1519 ** transitions leaving the lock state different from what it started but
1520 ** still short of its goal. The following chart shows the allowed
1521 ** transitions and the inserted intermediate states:
1523 ** UNLOCKED -> SHARED
1524 ** SHARED -> RESERVED
1525 ** SHARED -> (PENDING) -> EXCLUSIVE
1526 ** RESERVED -> (PENDING) -> EXCLUSIVE
1527 ** PENDING -> EXCLUSIVE
1529 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
1530 ** routine to lower a locking level.
1532 static int unixLock(sqlite3_file
*id
, int eFileLock
){
1533 /* The following describes the implementation of the various locks and
1534 ** lock transitions in terms of the POSIX advisory shared and exclusive
1535 ** lock primitives (called read-locks and write-locks below, to avoid
1536 ** confusion with SQLite lock names). The algorithms are complicated
1537 ** slightly in order to be compatible with Windows95 systems simultaneously
1538 ** accessing the same database file, in case that is ever required.
1540 ** Symbols defined in os.h indentify the 'pending byte' and the 'reserved
1541 ** byte', each single bytes at well known offsets, and the 'shared byte
1542 ** range', a range of 510 bytes at a well known offset.
1544 ** To obtain a SHARED lock, a read-lock is obtained on the 'pending
1545 ** byte'. If this is successful, 'shared byte range' is read-locked
1546 ** and the lock on the 'pending byte' released. (Legacy note: When
1547 ** SQLite was first developed, Windows95 systems were still very common,
1548 ** and Widnows95 lacks a shared-lock capability. So on Windows95, a
1549 ** single randomly selected by from the 'shared byte range' is locked.
1550 ** Windows95 is now pretty much extinct, but this work-around for the
1551 ** lack of shared-locks on Windows95 lives on, for backwards
1554 ** A process may only obtain a RESERVED lock after it has a SHARED lock.
1555 ** A RESERVED lock is implemented by grabbing a write-lock on the
1558 ** A process may only obtain a PENDING lock after it has obtained a
1559 ** SHARED lock. A PENDING lock is implemented by obtaining a write-lock
1560 ** on the 'pending byte'. This ensures that no new SHARED locks can be
1561 ** obtained, but existing SHARED locks are allowed to persist. A process
1562 ** does not have to obtain a RESERVED lock on the way to a PENDING lock.
1563 ** This property is used by the algorithm for rolling back a journal file
1566 ** An EXCLUSIVE lock, obtained after a PENDING lock is held, is
1567 ** implemented by obtaining a write-lock on the entire 'shared byte
1568 ** range'. Since all other locks require a read-lock on one of the bytes
1569 ** within this range, this ensures that no other locks are held on the
1573 unixFile
*pFile
= (unixFile
*)id
;
1574 unixInodeInfo
*pInode
;
1579 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (unix)\n", pFile
->h
,
1580 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
1581 azFileLock(pFile
->pInode
->eFileLock
), pFile
->pInode
->nShared
,
1584 /* If there is already a lock of this type or more restrictive on the
1585 ** unixFile, do nothing. Don't use the end_lock: exit path, as
1586 ** unixEnterMutex() hasn't been called yet.
1588 if( pFile
->eFileLock
>=eFileLock
){
1589 OSTRACE(("LOCK %d %s ok (already held) (unix)\n", pFile
->h
,
1590 azFileLock(eFileLock
)));
1594 /* Make sure the locking sequence is correct.
1595 ** (1) We never move from unlocked to anything higher than shared lock.
1596 ** (2) SQLite never explicitly requests a pendig lock.
1597 ** (3) A shared lock is always held when a reserve lock is requested.
1599 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
1600 assert( eFileLock
!=PENDING_LOCK
);
1601 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
1603 /* This mutex is needed because pFile->pInode is shared across threads
1606 pInode
= pFile
->pInode
;
1608 /* If some thread using this PID has a lock via a different unixFile*
1609 ** handle that precludes the requested lock, return BUSY.
1611 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
1612 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
1618 /* If a SHARED lock is requested, and some thread using this PID already
1619 ** has a SHARED or RESERVED lock, then increment reference counts and
1620 ** return SQLITE_OK.
1622 if( eFileLock
==SHARED_LOCK
&&
1623 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
1624 assert( eFileLock
==SHARED_LOCK
);
1625 assert( pFile
->eFileLock
==0 );
1626 assert( pInode
->nShared
>0 );
1627 pFile
->eFileLock
= SHARED_LOCK
;
1634 /* A PENDING lock is needed before acquiring a SHARED lock and before
1635 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
1639 lock
.l_whence
= SEEK_SET
;
1640 if( eFileLock
==SHARED_LOCK
1641 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
1643 lock
.l_type
= (eFileLock
==SHARED_LOCK
?F_RDLCK
:F_WRLCK
);
1644 lock
.l_start
= PENDING_BYTE
;
1645 if( unixFileLock(pFile
, &lock
) ){
1647 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1648 if( rc
!=SQLITE_BUSY
){
1649 storeLastErrno(pFile
, tErrno
);
1656 /* If control gets to this point, then actually go ahead and make
1657 ** operating system calls for the specified lock.
1659 if( eFileLock
==SHARED_LOCK
){
1660 assert( pInode
->nShared
==0 );
1661 assert( pInode
->eFileLock
==0 );
1662 assert( rc
==SQLITE_OK
);
1664 /* Now get the read-lock */
1665 lock
.l_start
= SHARED_FIRST
;
1666 lock
.l_len
= SHARED_SIZE
;
1667 if( unixFileLock(pFile
, &lock
) ){
1669 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1672 /* Drop the temporary PENDING lock */
1673 lock
.l_start
= PENDING_BYTE
;
1675 lock
.l_type
= F_UNLCK
;
1676 if( unixFileLock(pFile
, &lock
) && rc
==SQLITE_OK
){
1677 /* This could happen with a network mount */
1679 rc
= SQLITE_IOERR_UNLOCK
;
1683 if( rc
!=SQLITE_BUSY
){
1684 storeLastErrno(pFile
, tErrno
);
1688 pFile
->eFileLock
= SHARED_LOCK
;
1690 pInode
->nShared
= 1;
1692 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
1693 /* We are trying for an exclusive lock but another thread in this
1694 ** same process is still holding a shared lock. */
1697 /* The request was for a RESERVED or EXCLUSIVE lock. It is
1698 ** assumed that there is a SHARED or greater lock on the file
1701 assert( 0!=pFile
->eFileLock
);
1702 lock
.l_type
= F_WRLCK
;
1704 assert( eFileLock
==RESERVED_LOCK
|| eFileLock
==EXCLUSIVE_LOCK
);
1705 if( eFileLock
==RESERVED_LOCK
){
1706 lock
.l_start
= RESERVED_BYTE
;
1709 lock
.l_start
= SHARED_FIRST
;
1710 lock
.l_len
= SHARED_SIZE
;
1713 if( unixFileLock(pFile
, &lock
) ){
1715 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
1716 if( rc
!=SQLITE_BUSY
){
1717 storeLastErrno(pFile
, tErrno
);
1724 /* Set up the transaction-counter change checking flags when
1725 ** transitioning from a SHARED to a RESERVED lock. The change
1726 ** from SHARED to RESERVED marks the beginning of a normal
1727 ** write operation (not a hot journal rollback).
1730 && pFile
->eFileLock
<=SHARED_LOCK
1731 && eFileLock
==RESERVED_LOCK
1733 pFile
->transCntrChng
= 0;
1734 pFile
->dbUpdate
= 0;
1735 pFile
->inNormalWrite
= 1;
1740 if( rc
==SQLITE_OK
){
1741 pFile
->eFileLock
= eFileLock
;
1742 pInode
->eFileLock
= eFileLock
;
1743 }else if( eFileLock
==EXCLUSIVE_LOCK
){
1744 pFile
->eFileLock
= PENDING_LOCK
;
1745 pInode
->eFileLock
= PENDING_LOCK
;
1750 OSTRACE(("LOCK %d %s %s (unix)\n", pFile
->h
, azFileLock(eFileLock
),
1751 rc
==SQLITE_OK
? "ok" : "failed"));
1756 ** Add the file descriptor used by file handle pFile to the corresponding
1759 static void setPendingFd(unixFile
*pFile
){
1760 unixInodeInfo
*pInode
= pFile
->pInode
;
1761 UnixUnusedFd
*p
= pFile
->pPreallocatedUnused
;
1762 p
->pNext
= pInode
->pUnused
;
1763 pInode
->pUnused
= p
;
1765 pFile
->pPreallocatedUnused
= 0;
1770 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1771 ** must be either NO_LOCK or SHARED_LOCK.
1773 ** If the locking level of the file descriptor is already at or below
1774 ** the requested locking level, this routine is a no-op.
1776 ** If handleNFSUnlock is true, then on downgrading an EXCLUSIVE_LOCK to SHARED
1777 ** the byte range is divided into 2 parts and the first part is unlocked then
1778 ** set to a read lock, then the other part is simply unlocked. This works
1779 ** around a bug in BSD NFS lockd (also seen on MacOSX 10.3+) that fails to
1780 ** remove the write lock on a region when a read lock is set.
1782 static int posixUnlock(sqlite3_file
*id
, int eFileLock
, int handleNFSUnlock
){
1783 unixFile
*pFile
= (unixFile
*)id
;
1784 unixInodeInfo
*pInode
;
1789 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (unix)\n", pFile
->h
, eFileLock
,
1790 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
1793 assert( eFileLock
<=SHARED_LOCK
);
1794 if( pFile
->eFileLock
<=eFileLock
){
1798 pInode
= pFile
->pInode
;
1799 assert( pInode
->nShared
!=0 );
1800 if( pFile
->eFileLock
>SHARED_LOCK
){
1801 assert( pInode
->eFileLock
==pFile
->eFileLock
);
1804 /* When reducing a lock such that other processes can start
1805 ** reading the database file again, make sure that the
1806 ** transaction counter was updated if any part of the database
1807 ** file changed. If the transaction counter is not updated,
1808 ** other connections to the same file might not realize that
1809 ** the file has changed and hence might not know to flush their
1810 ** cache. The use of a stale cache can lead to database corruption.
1812 pFile
->inNormalWrite
= 0;
1815 /* downgrading to a shared lock on NFS involves clearing the write lock
1816 ** before establishing the readlock - to avoid a race condition we downgrade
1817 ** the lock in 2 blocks, so that part of the range will be covered by a
1818 ** write lock until the rest is covered by a read lock:
1824 if( eFileLock
==SHARED_LOCK
){
1825 #if !defined(__APPLE__) || !SQLITE_ENABLE_LOCKING_STYLE
1826 (void)handleNFSUnlock
;
1827 assert( handleNFSUnlock
==0 );
1829 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
1830 if( handleNFSUnlock
){
1831 int tErrno
; /* Error code from system call errors */
1832 off_t divSize
= SHARED_SIZE
- 1;
1834 lock
.l_type
= F_UNLCK
;
1835 lock
.l_whence
= SEEK_SET
;
1836 lock
.l_start
= SHARED_FIRST
;
1837 lock
.l_len
= divSize
;
1838 if( unixFileLock(pFile
, &lock
)==(-1) ){
1840 rc
= SQLITE_IOERR_UNLOCK
;
1841 storeLastErrno(pFile
, tErrno
);
1844 lock
.l_type
= F_RDLCK
;
1845 lock
.l_whence
= SEEK_SET
;
1846 lock
.l_start
= SHARED_FIRST
;
1847 lock
.l_len
= divSize
;
1848 if( unixFileLock(pFile
, &lock
)==(-1) ){
1850 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_RDLOCK
);
1851 if( IS_LOCK_ERROR(rc
) ){
1852 storeLastErrno(pFile
, tErrno
);
1856 lock
.l_type
= F_UNLCK
;
1857 lock
.l_whence
= SEEK_SET
;
1858 lock
.l_start
= SHARED_FIRST
+divSize
;
1859 lock
.l_len
= SHARED_SIZE
-divSize
;
1860 if( unixFileLock(pFile
, &lock
)==(-1) ){
1862 rc
= SQLITE_IOERR_UNLOCK
;
1863 storeLastErrno(pFile
, tErrno
);
1867 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
1869 lock
.l_type
= F_RDLCK
;
1870 lock
.l_whence
= SEEK_SET
;
1871 lock
.l_start
= SHARED_FIRST
;
1872 lock
.l_len
= SHARED_SIZE
;
1873 if( unixFileLock(pFile
, &lock
) ){
1874 /* In theory, the call to unixFileLock() cannot fail because another
1875 ** process is holding an incompatible lock. If it does, this
1876 ** indicates that the other process is not following the locking
1877 ** protocol. If this happens, return SQLITE_IOERR_RDLOCK. Returning
1878 ** SQLITE_BUSY would confuse the upper layer (in practice it causes
1879 ** an assert to fail). */
1880 rc
= SQLITE_IOERR_RDLOCK
;
1881 storeLastErrno(pFile
, errno
);
1886 lock
.l_type
= F_UNLCK
;
1887 lock
.l_whence
= SEEK_SET
;
1888 lock
.l_start
= PENDING_BYTE
;
1889 lock
.l_len
= 2L; assert( PENDING_BYTE
+1==RESERVED_BYTE
);
1890 if( unixFileLock(pFile
, &lock
)==0 ){
1891 pInode
->eFileLock
= SHARED_LOCK
;
1893 rc
= SQLITE_IOERR_UNLOCK
;
1894 storeLastErrno(pFile
, errno
);
1898 if( eFileLock
==NO_LOCK
){
1899 /* Decrement the shared lock counter. Release the lock using an
1900 ** OS call only when all threads in this same process have released
1904 if( pInode
->nShared
==0 ){
1905 lock
.l_type
= F_UNLCK
;
1906 lock
.l_whence
= SEEK_SET
;
1907 lock
.l_start
= lock
.l_len
= 0L;
1908 if( unixFileLock(pFile
, &lock
)==0 ){
1909 pInode
->eFileLock
= NO_LOCK
;
1911 rc
= SQLITE_IOERR_UNLOCK
;
1912 storeLastErrno(pFile
, errno
);
1913 pInode
->eFileLock
= NO_LOCK
;
1914 pFile
->eFileLock
= NO_LOCK
;
1918 /* Decrement the count of locks against this same file. When the
1919 ** count reaches zero, close any other file descriptors whose close
1920 ** was deferred because of outstanding locks.
1923 assert( pInode
->nLock
>=0 );
1924 if( pInode
->nLock
==0 ){
1925 closePendingFds(pFile
);
1931 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
1936 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
1937 ** must be either NO_LOCK or SHARED_LOCK.
1939 ** If the locking level of the file descriptor is already at or below
1940 ** the requested locking level, this routine is a no-op.
1942 static int unixUnlock(sqlite3_file
*id
, int eFileLock
){
1943 #if SQLITE_MAX_MMAP_SIZE>0
1944 assert( eFileLock
==SHARED_LOCK
|| ((unixFile
*)id
)->nFetchOut
==0 );
1946 return posixUnlock(id
, eFileLock
, 0);
1949 #if SQLITE_MAX_MMAP_SIZE>0
1950 static int unixMapfile(unixFile
*pFd
, i64 nByte
);
1951 static void unixUnmapfile(unixFile
*pFd
);
1955 ** This function performs the parts of the "close file" operation
1956 ** common to all locking schemes. It closes the directory and file
1957 ** handles, if they are valid, and sets all fields of the unixFile
1960 ** It is *not* necessary to hold the mutex when this routine is called,
1961 ** even on VxWorks. A mutex will be acquired on VxWorks by the
1962 ** vxworksReleaseFileId() routine.
1964 static int closeUnixFile(sqlite3_file
*id
){
1965 unixFile
*pFile
= (unixFile
*)id
;
1966 #if SQLITE_MAX_MMAP_SIZE>0
1967 unixUnmapfile(pFile
);
1970 robust_close(pFile
, pFile
->h
, __LINE__
);
1975 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
1976 osUnlink(pFile
->pId
->zCanonicalName
);
1978 vxworksReleaseFileId(pFile
->pId
);
1982 #ifdef SQLITE_UNLINK_AFTER_CLOSE
1983 if( pFile
->ctrlFlags
& UNIXFILE_DELETE
){
1984 osUnlink(pFile
->zPath
);
1985 sqlite3_free(*(char**)&pFile
->zPath
);
1989 OSTRACE(("CLOSE %-3d\n", pFile
->h
));
1991 sqlite3_free(pFile
->pPreallocatedUnused
);
1992 memset(pFile
, 0, sizeof(unixFile
));
1999 static int unixClose(sqlite3_file
*id
){
2001 unixFile
*pFile
= (unixFile
*)id
;
2002 verifyDbFile(pFile
);
2003 unixUnlock(id
, NO_LOCK
);
2006 /* unixFile.pInode is always valid here. Otherwise, a different close
2007 ** routine (e.g. nolockClose()) would be called instead.
2009 assert( pFile
->pInode
->nLock
>0 || pFile
->pInode
->bProcessLock
==0 );
2010 if( ALWAYS(pFile
->pInode
) && pFile
->pInode
->nLock
){
2011 /* If there are outstanding locks, do not actually close the file just
2012 ** yet because that would clear those locks. Instead, add the file
2013 ** descriptor to pInode->pUnused list. It will be automatically closed
2014 ** when the last lock is cleared.
2016 setPendingFd(pFile
);
2018 releaseInodeInfo(pFile
);
2019 rc
= closeUnixFile(id
);
2024 /************** End of the posix advisory lock implementation *****************
2025 ******************************************************************************/
2027 /******************************************************************************
2028 ****************************** No-op Locking **********************************
2030 ** Of the various locking implementations available, this is by far the
2031 ** simplest: locking is ignored. No attempt is made to lock the database
2032 ** file for reading or writing.
2034 ** This locking mode is appropriate for use on read-only databases
2035 ** (ex: databases that are burned into CD-ROM, for example.) It can
2036 ** also be used if the application employs some external mechanism to
2037 ** prevent simultaneous access of the same database by two or more
2038 ** database connections. But there is a serious risk of database
2039 ** corruption if this locking mode is used in situations where multiple
2040 ** database connections are accessing the same database file at the same
2041 ** time and one or more of those connections are writing.
2044 static int nolockCheckReservedLock(sqlite3_file
*NotUsed
, int *pResOut
){
2045 UNUSED_PARAMETER(NotUsed
);
2049 static int nolockLock(sqlite3_file
*NotUsed
, int NotUsed2
){
2050 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2053 static int nolockUnlock(sqlite3_file
*NotUsed
, int NotUsed2
){
2054 UNUSED_PARAMETER2(NotUsed
, NotUsed2
);
2061 static int nolockClose(sqlite3_file
*id
) {
2062 return closeUnixFile(id
);
2065 /******************* End of the no-op lock implementation *********************
2066 ******************************************************************************/
2068 /******************************************************************************
2069 ************************* Begin dot-file Locking ******************************
2071 ** The dotfile locking implementation uses the existence of separate lock
2072 ** files (really a directory) to control access to the database. This works
2073 ** on just about every filesystem imaginable. But there are serious downsides:
2075 ** (1) There is zero concurrency. A single reader blocks all other
2076 ** connections from reading or writing the database.
2078 ** (2) An application crash or power loss can leave stale lock files
2079 ** sitting around that need to be cleared manually.
2081 ** Nevertheless, a dotlock is an appropriate locking mode for use if no
2082 ** other locking strategy is available.
2084 ** Dotfile locking works by creating a subdirectory in the same directory as
2085 ** the database and with the same name but with a ".lock" extension added.
2086 ** The existence of a lock directory implies an EXCLUSIVE lock. All other
2087 ** lock types (SHARED, RESERVED, PENDING) are mapped into EXCLUSIVE.
2091 ** The file suffix added to the data base filename in order to create the
2094 #define DOTLOCK_SUFFIX ".lock"
2097 ** This routine checks if there is a RESERVED lock held on the specified
2098 ** file by this or any other process. If such a lock is held, set *pResOut
2099 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2100 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2102 ** In dotfile locking, either a lock exists or it does not. So in this
2103 ** variation of CheckReservedLock(), *pResOut is set to true if any lock
2104 ** is held on the file and false if the file is unlocked.
2106 static int dotlockCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2109 unixFile
*pFile
= (unixFile
*)id
;
2111 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2114 reserved
= osAccess((const char*)pFile
->lockingContext
, 0)==0;
2115 OSTRACE(("TEST WR-LOCK %d %d %d (dotlock)\n", pFile
->h
, rc
, reserved
));
2116 *pResOut
= reserved
;
2121 ** Lock the file with the lock specified by parameter eFileLock - one
2122 ** of the following:
2125 ** (2) RESERVED_LOCK
2127 ** (4) EXCLUSIVE_LOCK
2129 ** Sometimes when requesting one lock state, additional lock states
2130 ** are inserted in between. The locking might fail on one of the later
2131 ** transitions leaving the lock state different from what it started but
2132 ** still short of its goal. The following chart shows the allowed
2133 ** transitions and the inserted intermediate states:
2135 ** UNLOCKED -> SHARED
2136 ** SHARED -> RESERVED
2137 ** SHARED -> (PENDING) -> EXCLUSIVE
2138 ** RESERVED -> (PENDING) -> EXCLUSIVE
2139 ** PENDING -> EXCLUSIVE
2141 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2142 ** routine to lower a locking level.
2144 ** With dotfile locking, we really only support state (4): EXCLUSIVE.
2145 ** But we track the other locking levels internally.
2147 static int dotlockLock(sqlite3_file
*id
, int eFileLock
) {
2148 unixFile
*pFile
= (unixFile
*)id
;
2149 char *zLockFile
= (char *)pFile
->lockingContext
;
2153 /* If we have any lock, then the lock file already exists. All we have
2154 ** to do is adjust our internal record of the lock level.
2156 if( pFile
->eFileLock
> NO_LOCK
){
2157 pFile
->eFileLock
= eFileLock
;
2158 /* Always update the timestamp on the old file */
2160 utime(zLockFile
, NULL
);
2162 utimes(zLockFile
, NULL
);
2167 /* grab an exclusive lock */
2168 rc
= osMkdir(zLockFile
, 0777);
2170 /* failed to open/create the lock directory */
2172 if( EEXIST
== tErrno
){
2175 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2176 if( rc
!=SQLITE_BUSY
){
2177 storeLastErrno(pFile
, tErrno
);
2183 /* got it, set the type and return ok */
2184 pFile
->eFileLock
= eFileLock
;
2189 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2190 ** must be either NO_LOCK or SHARED_LOCK.
2192 ** If the locking level of the file descriptor is already at or below
2193 ** the requested locking level, this routine is a no-op.
2195 ** When the locking level reaches NO_LOCK, delete the lock file.
2197 static int dotlockUnlock(sqlite3_file
*id
, int eFileLock
) {
2198 unixFile
*pFile
= (unixFile
*)id
;
2199 char *zLockFile
= (char *)pFile
->lockingContext
;
2203 OSTRACE(("UNLOCK %d %d was %d pid=%d (dotlock)\n", pFile
->h
, eFileLock
,
2204 pFile
->eFileLock
, osGetpid(0)));
2205 assert( eFileLock
<=SHARED_LOCK
);
2207 /* no-op if possible */
2208 if( pFile
->eFileLock
==eFileLock
){
2212 /* To downgrade to shared, simply update our internal notion of the
2213 ** lock state. No need to mess with the file on disk.
2215 if( eFileLock
==SHARED_LOCK
){
2216 pFile
->eFileLock
= SHARED_LOCK
;
2220 /* To fully unlock the database, delete the lock file */
2221 assert( eFileLock
==NO_LOCK
);
2222 rc
= osRmdir(zLockFile
);
2225 if( tErrno
==ENOENT
){
2228 rc
= SQLITE_IOERR_UNLOCK
;
2229 storeLastErrno(pFile
, tErrno
);
2233 pFile
->eFileLock
= NO_LOCK
;
2238 ** Close a file. Make sure the lock has been released before closing.
2240 static int dotlockClose(sqlite3_file
*id
) {
2241 unixFile
*pFile
= (unixFile
*)id
;
2243 dotlockUnlock(id
, NO_LOCK
);
2244 sqlite3_free(pFile
->lockingContext
);
2245 return closeUnixFile(id
);
2247 /****************** End of the dot-file lock implementation *******************
2248 ******************************************************************************/
2250 /******************************************************************************
2251 ************************** Begin flock Locking ********************************
2253 ** Use the flock() system call to do file locking.
2255 ** flock() locking is like dot-file locking in that the various
2256 ** fine-grain locking levels supported by SQLite are collapsed into
2257 ** a single exclusive lock. In other words, SHARED, RESERVED, and
2258 ** PENDING locks are the same thing as an EXCLUSIVE lock. SQLite
2259 ** still works when you do this, but concurrency is reduced since
2260 ** only a single process can be reading the database at a time.
2262 ** Omit this section if SQLITE_ENABLE_LOCKING_STYLE is turned off
2264 #if SQLITE_ENABLE_LOCKING_STYLE
2267 ** Retry flock() calls that fail with EINTR
2270 static int robust_flock(int fd
, int op
){
2272 do{ rc
= flock(fd
,op
); }while( rc
<0 && errno
==EINTR
);
2276 # define robust_flock(a,b) flock(a,b)
2281 ** This routine checks if there is a RESERVED lock held on the specified
2282 ** file by this or any other process. If such a lock is held, set *pResOut
2283 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2284 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2286 static int flockCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2289 unixFile
*pFile
= (unixFile
*)id
;
2291 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2295 /* Check if a thread in this process holds such a lock */
2296 if( pFile
->eFileLock
>SHARED_LOCK
){
2300 /* Otherwise see if some other process holds it. */
2302 /* attempt to get the lock */
2303 int lrc
= robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
);
2305 /* got the lock, unlock it */
2306 lrc
= robust_flock(pFile
->h
, LOCK_UN
);
2309 /* unlock failed with an error */
2310 lrc
= SQLITE_IOERR_UNLOCK
;
2311 storeLastErrno(pFile
, tErrno
);
2317 /* someone else might have it reserved */
2318 lrc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2319 if( IS_LOCK_ERROR(lrc
) ){
2320 storeLastErrno(pFile
, tErrno
);
2325 OSTRACE(("TEST WR-LOCK %d %d %d (flock)\n", pFile
->h
, rc
, reserved
));
2327 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2328 if( (rc
& 0xff) == SQLITE_IOERR
){
2332 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2333 *pResOut
= reserved
;
2338 ** Lock the file with the lock specified by parameter eFileLock - one
2339 ** of the following:
2342 ** (2) RESERVED_LOCK
2344 ** (4) EXCLUSIVE_LOCK
2346 ** Sometimes when requesting one lock state, additional lock states
2347 ** are inserted in between. The locking might fail on one of the later
2348 ** transitions leaving the lock state different from what it started but
2349 ** still short of its goal. The following chart shows the allowed
2350 ** transitions and the inserted intermediate states:
2352 ** UNLOCKED -> SHARED
2353 ** SHARED -> RESERVED
2354 ** SHARED -> (PENDING) -> EXCLUSIVE
2355 ** RESERVED -> (PENDING) -> EXCLUSIVE
2356 ** PENDING -> EXCLUSIVE
2358 ** flock() only really support EXCLUSIVE locks. We track intermediate
2359 ** lock states in the sqlite3_file structure, but all locks SHARED or
2360 ** above are really EXCLUSIVE locks and exclude all other processes from
2363 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2364 ** routine to lower a locking level.
2366 static int flockLock(sqlite3_file
*id
, int eFileLock
) {
2368 unixFile
*pFile
= (unixFile
*)id
;
2372 /* if we already have a lock, it is exclusive.
2373 ** Just adjust level and punt on outta here. */
2374 if (pFile
->eFileLock
> NO_LOCK
) {
2375 pFile
->eFileLock
= eFileLock
;
2379 /* grab an exclusive lock */
2381 if (robust_flock(pFile
->h
, LOCK_EX
| LOCK_NB
)) {
2383 /* didn't get, must be busy */
2384 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_LOCK
);
2385 if( IS_LOCK_ERROR(rc
) ){
2386 storeLastErrno(pFile
, tErrno
);
2389 /* got it, set the type and return ok */
2390 pFile
->eFileLock
= eFileLock
;
2392 OSTRACE(("LOCK %d %s %s (flock)\n", pFile
->h
, azFileLock(eFileLock
),
2393 rc
==SQLITE_OK
? "ok" : "failed"));
2394 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2395 if( (rc
& 0xff) == SQLITE_IOERR
){
2398 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2404 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2405 ** must be either NO_LOCK or SHARED_LOCK.
2407 ** If the locking level of the file descriptor is already at or below
2408 ** the requested locking level, this routine is a no-op.
2410 static int flockUnlock(sqlite3_file
*id
, int eFileLock
) {
2411 unixFile
*pFile
= (unixFile
*)id
;
2414 OSTRACE(("UNLOCK %d %d was %d pid=%d (flock)\n", pFile
->h
, eFileLock
,
2415 pFile
->eFileLock
, osGetpid(0)));
2416 assert( eFileLock
<=SHARED_LOCK
);
2418 /* no-op if possible */
2419 if( pFile
->eFileLock
==eFileLock
){
2423 /* shared can just be set because we always have an exclusive */
2424 if (eFileLock
==SHARED_LOCK
) {
2425 pFile
->eFileLock
= eFileLock
;
2429 /* no, really, unlock. */
2430 if( robust_flock(pFile
->h
, LOCK_UN
) ){
2431 #ifdef SQLITE_IGNORE_FLOCK_LOCK_ERRORS
2433 #endif /* SQLITE_IGNORE_FLOCK_LOCK_ERRORS */
2434 return SQLITE_IOERR_UNLOCK
;
2436 pFile
->eFileLock
= NO_LOCK
;
2444 static int flockClose(sqlite3_file
*id
) {
2446 flockUnlock(id
, NO_LOCK
);
2447 return closeUnixFile(id
);
2450 #endif /* SQLITE_ENABLE_LOCKING_STYLE && !OS_VXWORK */
2452 /******************* End of the flock lock implementation *********************
2453 ******************************************************************************/
2455 /******************************************************************************
2456 ************************ Begin Named Semaphore Locking ************************
2458 ** Named semaphore locking is only supported on VxWorks.
2460 ** Semaphore locking is like dot-lock and flock in that it really only
2461 ** supports EXCLUSIVE locking. Only a single process can read or write
2462 ** the database file at a time. This reduces potential concurrency, but
2463 ** makes the lock implementation much easier.
2468 ** This routine checks if there is a RESERVED lock held on the specified
2469 ** file by this or any other process. If such a lock is held, set *pResOut
2470 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2471 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2473 static int semXCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
2476 unixFile
*pFile
= (unixFile
*)id
;
2478 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2482 /* Check if a thread in this process holds such a lock */
2483 if( pFile
->eFileLock
>SHARED_LOCK
){
2487 /* Otherwise see if some other process holds it. */
2489 sem_t
*pSem
= pFile
->pInode
->pSem
;
2491 if( sem_trywait(pSem
)==-1 ){
2493 if( EAGAIN
!= tErrno
){
2494 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_CHECKRESERVEDLOCK
);
2495 storeLastErrno(pFile
, tErrno
);
2497 /* someone else has the lock when we are in NO_LOCK */
2498 reserved
= (pFile
->eFileLock
< SHARED_LOCK
);
2501 /* we could have it if we want it */
2505 OSTRACE(("TEST WR-LOCK %d %d %d (sem)\n", pFile
->h
, rc
, reserved
));
2507 *pResOut
= reserved
;
2512 ** Lock the file with the lock specified by parameter eFileLock - one
2513 ** of the following:
2516 ** (2) RESERVED_LOCK
2518 ** (4) EXCLUSIVE_LOCK
2520 ** Sometimes when requesting one lock state, additional lock states
2521 ** are inserted in between. The locking might fail on one of the later
2522 ** transitions leaving the lock state different from what it started but
2523 ** still short of its goal. The following chart shows the allowed
2524 ** transitions and the inserted intermediate states:
2526 ** UNLOCKED -> SHARED
2527 ** SHARED -> RESERVED
2528 ** SHARED -> (PENDING) -> EXCLUSIVE
2529 ** RESERVED -> (PENDING) -> EXCLUSIVE
2530 ** PENDING -> EXCLUSIVE
2532 ** Semaphore locks only really support EXCLUSIVE locks. We track intermediate
2533 ** lock states in the sqlite3_file structure, but all locks SHARED or
2534 ** above are really EXCLUSIVE locks and exclude all other processes from
2537 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2538 ** routine to lower a locking level.
2540 static int semXLock(sqlite3_file
*id
, int eFileLock
) {
2541 unixFile
*pFile
= (unixFile
*)id
;
2542 sem_t
*pSem
= pFile
->pInode
->pSem
;
2545 /* if we already have a lock, it is exclusive.
2546 ** Just adjust level and punt on outta here. */
2547 if (pFile
->eFileLock
> NO_LOCK
) {
2548 pFile
->eFileLock
= eFileLock
;
2553 /* lock semaphore now but bail out when already locked. */
2554 if( sem_trywait(pSem
)==-1 ){
2559 /* got it, set the type and return ok */
2560 pFile
->eFileLock
= eFileLock
;
2567 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2568 ** must be either NO_LOCK or SHARED_LOCK.
2570 ** If the locking level of the file descriptor is already at or below
2571 ** the requested locking level, this routine is a no-op.
2573 static int semXUnlock(sqlite3_file
*id
, int eFileLock
) {
2574 unixFile
*pFile
= (unixFile
*)id
;
2575 sem_t
*pSem
= pFile
->pInode
->pSem
;
2579 OSTRACE(("UNLOCK %d %d was %d pid=%d (sem)\n", pFile
->h
, eFileLock
,
2580 pFile
->eFileLock
, osGetpid(0)));
2581 assert( eFileLock
<=SHARED_LOCK
);
2583 /* no-op if possible */
2584 if( pFile
->eFileLock
==eFileLock
){
2588 /* shared can just be set because we always have an exclusive */
2589 if (eFileLock
==SHARED_LOCK
) {
2590 pFile
->eFileLock
= eFileLock
;
2594 /* no, really unlock. */
2595 if ( sem_post(pSem
)==-1 ) {
2596 int rc
, tErrno
= errno
;
2597 rc
= sqliteErrorFromPosixError(tErrno
, SQLITE_IOERR_UNLOCK
);
2598 if( IS_LOCK_ERROR(rc
) ){
2599 storeLastErrno(pFile
, tErrno
);
2603 pFile
->eFileLock
= NO_LOCK
;
2610 static int semXClose(sqlite3_file
*id
) {
2612 unixFile
*pFile
= (unixFile
*)id
;
2613 semXUnlock(id
, NO_LOCK
);
2616 releaseInodeInfo(pFile
);
2623 #endif /* OS_VXWORKS */
2625 ** Named semaphore locking is only available on VxWorks.
2627 *************** End of the named semaphore lock implementation ****************
2628 ******************************************************************************/
2631 /******************************************************************************
2632 *************************** Begin AFP Locking *********************************
2634 ** AFP is the Apple Filing Protocol. AFP is a network filesystem found
2635 ** on Apple Macintosh computers - both OS9 and OSX.
2637 ** Third-party implementations of AFP are available. But this code here
2638 ** only works on OSX.
2641 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
2643 ** The afpLockingContext structure contains all afp lock specific state
2645 typedef struct afpLockingContext afpLockingContext
;
2646 struct afpLockingContext
{
2648 const char *dbPath
; /* Name of the open file */
2651 struct ByteRangeLockPB2
2653 unsigned long long offset
; /* offset to first byte to lock */
2654 unsigned long long length
; /* nbr of bytes to lock */
2655 unsigned long long retRangeStart
; /* nbr of 1st byte locked if successful */
2656 unsigned char unLockFlag
; /* 1 = unlock, 0 = lock */
2657 unsigned char startEndFlag
; /* 1=rel to end of fork, 0=rel to start */
2658 int fd
; /* file desc to assoc this lock with */
2661 #define afpfsByteRangeLock2FSCTL _IOWR('z', 23, struct ByteRangeLockPB2)
2664 ** This is a utility for setting or clearing a bit-range lock on an
2667 ** Return SQLITE_OK on success, SQLITE_BUSY on failure.
2669 static int afpSetLock(
2670 const char *path
, /* Name of the file to be locked or unlocked */
2671 unixFile
*pFile
, /* Open file descriptor on path */
2672 unsigned long long offset
, /* First byte to be locked */
2673 unsigned long long length
, /* Number of bytes to lock */
2674 int setLockFlag
/* True to set lock. False to clear lock */
2676 struct ByteRangeLockPB2 pb
;
2679 pb
.unLockFlag
= setLockFlag
? 0 : 1;
2680 pb
.startEndFlag
= 0;
2685 OSTRACE(("AFPSETLOCK [%s] for %d%s in range %llx:%llx\n",
2686 (setLockFlag
?"ON":"OFF"), pFile
->h
, (pb
.fd
==-1?"[testval-1]":""),
2688 err
= fsctl(path
, afpfsByteRangeLock2FSCTL
, &pb
, 0);
2692 OSTRACE(("AFPSETLOCK failed to fsctl() '%s' %d %s\n",
2693 path
, tErrno
, strerror(tErrno
)));
2694 #ifdef SQLITE_IGNORE_AFP_LOCK_ERRORS
2697 rc
= sqliteErrorFromPosixError(tErrno
,
2698 setLockFlag
? SQLITE_IOERR_LOCK
: SQLITE_IOERR_UNLOCK
);
2699 #endif /* SQLITE_IGNORE_AFP_LOCK_ERRORS */
2700 if( IS_LOCK_ERROR(rc
) ){
2701 storeLastErrno(pFile
, tErrno
);
2710 ** This routine checks if there is a RESERVED lock held on the specified
2711 ** file by this or any other process. If such a lock is held, set *pResOut
2712 ** to a non-zero value otherwise *pResOut is set to zero. The return value
2713 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
2715 static int afpCheckReservedLock(sqlite3_file
*id
, int *pResOut
){
2718 unixFile
*pFile
= (unixFile
*)id
;
2719 afpLockingContext
*context
;
2721 SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK
; );
2724 context
= (afpLockingContext
*) pFile
->lockingContext
;
2725 if( context
->reserved
){
2729 unixEnterMutex(); /* Because pFile->pInode is shared across threads */
2731 /* Check if a thread in this process holds such a lock */
2732 if( pFile
->pInode
->eFileLock
>SHARED_LOCK
){
2736 /* Otherwise see if some other process holds it.
2739 /* lock the RESERVED byte */
2740 int lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2741 if( SQLITE_OK
==lrc
){
2742 /* if we succeeded in taking the reserved lock, unlock it to restore
2743 ** the original state */
2744 lrc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
2746 /* if we failed to get the lock then someone else must have it */
2749 if( IS_LOCK_ERROR(lrc
) ){
2755 OSTRACE(("TEST WR-LOCK %d %d %d (afp)\n", pFile
->h
, rc
, reserved
));
2757 *pResOut
= reserved
;
2762 ** Lock the file with the lock specified by parameter eFileLock - one
2763 ** of the following:
2766 ** (2) RESERVED_LOCK
2768 ** (4) EXCLUSIVE_LOCK
2770 ** Sometimes when requesting one lock state, additional lock states
2771 ** are inserted in between. The locking might fail on one of the later
2772 ** transitions leaving the lock state different from what it started but
2773 ** still short of its goal. The following chart shows the allowed
2774 ** transitions and the inserted intermediate states:
2776 ** UNLOCKED -> SHARED
2777 ** SHARED -> RESERVED
2778 ** SHARED -> (PENDING) -> EXCLUSIVE
2779 ** RESERVED -> (PENDING) -> EXCLUSIVE
2780 ** PENDING -> EXCLUSIVE
2782 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
2783 ** routine to lower a locking level.
2785 static int afpLock(sqlite3_file
*id
, int eFileLock
){
2787 unixFile
*pFile
= (unixFile
*)id
;
2788 unixInodeInfo
*pInode
= pFile
->pInode
;
2789 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
2792 OSTRACE(("LOCK %d %s was %s(%s,%d) pid=%d (afp)\n", pFile
->h
,
2793 azFileLock(eFileLock
), azFileLock(pFile
->eFileLock
),
2794 azFileLock(pInode
->eFileLock
), pInode
->nShared
, osGetpid(0)));
2796 /* If there is already a lock of this type or more restrictive on the
2797 ** unixFile, do nothing. Don't use the afp_end_lock: exit path, as
2798 ** unixEnterMutex() hasn't been called yet.
2800 if( pFile
->eFileLock
>=eFileLock
){
2801 OSTRACE(("LOCK %d %s ok (already held) (afp)\n", pFile
->h
,
2802 azFileLock(eFileLock
)));
2806 /* Make sure the locking sequence is correct
2807 ** (1) We never move from unlocked to anything higher than shared lock.
2808 ** (2) SQLite never explicitly requests a pendig lock.
2809 ** (3) A shared lock is always held when a reserve lock is requested.
2811 assert( pFile
->eFileLock
!=NO_LOCK
|| eFileLock
==SHARED_LOCK
);
2812 assert( eFileLock
!=PENDING_LOCK
);
2813 assert( eFileLock
!=RESERVED_LOCK
|| pFile
->eFileLock
==SHARED_LOCK
);
2815 /* This mutex is needed because pFile->pInode is shared across threads
2818 pInode
= pFile
->pInode
;
2820 /* If some thread using this PID has a lock via a different unixFile*
2821 ** handle that precludes the requested lock, return BUSY.
2823 if( (pFile
->eFileLock
!=pInode
->eFileLock
&&
2824 (pInode
->eFileLock
>=PENDING_LOCK
|| eFileLock
>SHARED_LOCK
))
2830 /* If a SHARED lock is requested, and some thread using this PID already
2831 ** has a SHARED or RESERVED lock, then increment reference counts and
2832 ** return SQLITE_OK.
2834 if( eFileLock
==SHARED_LOCK
&&
2835 (pInode
->eFileLock
==SHARED_LOCK
|| pInode
->eFileLock
==RESERVED_LOCK
) ){
2836 assert( eFileLock
==SHARED_LOCK
);
2837 assert( pFile
->eFileLock
==0 );
2838 assert( pInode
->nShared
>0 );
2839 pFile
->eFileLock
= SHARED_LOCK
;
2845 /* A PENDING lock is needed before acquiring a SHARED lock and before
2846 ** acquiring an EXCLUSIVE lock. For the SHARED lock, the PENDING will
2849 if( eFileLock
==SHARED_LOCK
2850 || (eFileLock
==EXCLUSIVE_LOCK
&& pFile
->eFileLock
<PENDING_LOCK
)
2853 failed
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 1);
2860 /* If control gets to this point, then actually go ahead and make
2861 ** operating system calls for the specified lock.
2863 if( eFileLock
==SHARED_LOCK
){
2864 int lrc1
, lrc2
, lrc1Errno
= 0;
2867 assert( pInode
->nShared
==0 );
2868 assert( pInode
->eFileLock
==0 );
2870 mask
= (sizeof(long)==8) ? LARGEST_INT64
: 0x7fffffff;
2871 /* Now get the read-lock SHARED_LOCK */
2872 /* note that the quality of the randomness doesn't matter that much */
2874 pInode
->sharedByte
= (lk
& mask
)%(SHARED_SIZE
- 1);
2875 lrc1
= afpSetLock(context
->dbPath
, pFile
,
2876 SHARED_FIRST
+pInode
->sharedByte
, 1, 1);
2877 if( IS_LOCK_ERROR(lrc1
) ){
2878 lrc1Errno
= pFile
->lastErrno
;
2880 /* Drop the temporary PENDING lock */
2881 lrc2
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
2883 if( IS_LOCK_ERROR(lrc1
) ) {
2884 storeLastErrno(pFile
, lrc1Errno
);
2887 } else if( IS_LOCK_ERROR(lrc2
) ){
2890 } else if( lrc1
!= SQLITE_OK
) {
2893 pFile
->eFileLock
= SHARED_LOCK
;
2895 pInode
->nShared
= 1;
2897 }else if( eFileLock
==EXCLUSIVE_LOCK
&& pInode
->nShared
>1 ){
2898 /* We are trying for an exclusive lock but another thread in this
2899 ** same process is still holding a shared lock. */
2902 /* The request was for a RESERVED or EXCLUSIVE lock. It is
2903 ** assumed that there is a SHARED or greater lock on the file
2907 assert( 0!=pFile
->eFileLock
);
2908 if (eFileLock
>= RESERVED_LOCK
&& pFile
->eFileLock
< RESERVED_LOCK
) {
2909 /* Acquire a RESERVED lock */
2910 failed
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1,1);
2912 context
->reserved
= 1;
2915 if (!failed
&& eFileLock
== EXCLUSIVE_LOCK
) {
2916 /* Acquire an EXCLUSIVE lock */
2918 /* Remove the shared lock before trying the range. we'll need to
2919 ** reestablish the shared lock if we can't get the afpUnlock
2921 if( !(failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
+
2922 pInode
->sharedByte
, 1, 0)) ){
2923 int failed2
= SQLITE_OK
;
2924 /* now attemmpt to get the exclusive lock range */
2925 failed
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
,
2927 if( failed
&& (failed2
= afpSetLock(context
->dbPath
, pFile
,
2928 SHARED_FIRST
+ pInode
->sharedByte
, 1, 1)) ){
2929 /* Can't reestablish the shared lock. Sqlite can't deal, this is
2930 ** a critical I/O error
2932 rc
= ((failed
& 0xff) == SQLITE_IOERR
) ? failed2
:
2945 if( rc
==SQLITE_OK
){
2946 pFile
->eFileLock
= eFileLock
;
2947 pInode
->eFileLock
= eFileLock
;
2948 }else if( eFileLock
==EXCLUSIVE_LOCK
){
2949 pFile
->eFileLock
= PENDING_LOCK
;
2950 pInode
->eFileLock
= PENDING_LOCK
;
2955 OSTRACE(("LOCK %d %s %s (afp)\n", pFile
->h
, azFileLock(eFileLock
),
2956 rc
==SQLITE_OK
? "ok" : "failed"));
2961 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
2962 ** must be either NO_LOCK or SHARED_LOCK.
2964 ** If the locking level of the file descriptor is already at or below
2965 ** the requested locking level, this routine is a no-op.
2967 static int afpUnlock(sqlite3_file
*id
, int eFileLock
) {
2969 unixFile
*pFile
= (unixFile
*)id
;
2970 unixInodeInfo
*pInode
;
2971 afpLockingContext
*context
= (afpLockingContext
*) pFile
->lockingContext
;
2978 OSTRACE(("UNLOCK %d %d was %d(%d,%d) pid=%d (afp)\n", pFile
->h
, eFileLock
,
2979 pFile
->eFileLock
, pFile
->pInode
->eFileLock
, pFile
->pInode
->nShared
,
2982 assert( eFileLock
<=SHARED_LOCK
);
2983 if( pFile
->eFileLock
<=eFileLock
){
2987 pInode
= pFile
->pInode
;
2988 assert( pInode
->nShared
!=0 );
2989 if( pFile
->eFileLock
>SHARED_LOCK
){
2990 assert( pInode
->eFileLock
==pFile
->eFileLock
);
2991 SimulateIOErrorBenign(1);
2992 SimulateIOError( h
=(-1) )
2993 SimulateIOErrorBenign(0);
2996 /* When reducing a lock such that other processes can start
2997 ** reading the database file again, make sure that the
2998 ** transaction counter was updated if any part of the database
2999 ** file changed. If the transaction counter is not updated,
3000 ** other connections to the same file might not realize that
3001 ** the file has changed and hence might not know to flush their
3002 ** cache. The use of a stale cache can lead to database corruption.
3004 assert( pFile
->inNormalWrite
==0
3005 || pFile
->dbUpdate
==0
3006 || pFile
->transCntrChng
==1 );
3007 pFile
->inNormalWrite
= 0;
3010 if( pFile
->eFileLock
==EXCLUSIVE_LOCK
){
3011 rc
= afpSetLock(context
->dbPath
, pFile
, SHARED_FIRST
, SHARED_SIZE
, 0);
3012 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1) ){
3013 /* only re-establish the shared lock if necessary */
3014 int sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3015 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 1);
3020 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=PENDING_LOCK
){
3021 rc
= afpSetLock(context
->dbPath
, pFile
, PENDING_BYTE
, 1, 0);
3023 if( rc
==SQLITE_OK
&& pFile
->eFileLock
>=RESERVED_LOCK
&& context
->reserved
){
3024 rc
= afpSetLock(context
->dbPath
, pFile
, RESERVED_BYTE
, 1, 0);
3026 context
->reserved
= 0;
3029 if( rc
==SQLITE_OK
&& (eFileLock
==SHARED_LOCK
|| pInode
->nShared
>1)){
3030 pInode
->eFileLock
= SHARED_LOCK
;
3033 if( rc
==SQLITE_OK
&& eFileLock
==NO_LOCK
){
3035 /* Decrement the shared lock counter. Release the lock using an
3036 ** OS call only when all threads in this same process have released
3039 unsigned long long sharedLockByte
= SHARED_FIRST
+pInode
->sharedByte
;
3041 if( pInode
->nShared
==0 ){
3042 SimulateIOErrorBenign(1);
3043 SimulateIOError( h
=(-1) )
3044 SimulateIOErrorBenign(0);
3046 rc
= afpSetLock(context
->dbPath
, pFile
, sharedLockByte
, 1, 0);
3049 pInode
->eFileLock
= NO_LOCK
;
3050 pFile
->eFileLock
= NO_LOCK
;
3053 if( rc
==SQLITE_OK
){
3055 assert( pInode
->nLock
>=0 );
3056 if( pInode
->nLock
==0 ){
3057 closePendingFds(pFile
);
3063 if( rc
==SQLITE_OK
) pFile
->eFileLock
= eFileLock
;
3068 ** Close a file & cleanup AFP specific locking context
3070 static int afpClose(sqlite3_file
*id
) {
3072 unixFile
*pFile
= (unixFile
*)id
;
3074 afpUnlock(id
, NO_LOCK
);
3076 if( pFile
->pInode
&& pFile
->pInode
->nLock
){
3077 /* If there are outstanding locks, do not actually close the file just
3078 ** yet because that would clear those locks. Instead, add the file
3079 ** descriptor to pInode->aPending. It will be automatically closed when
3080 ** the last lock is cleared.
3082 setPendingFd(pFile
);
3084 releaseInodeInfo(pFile
);
3085 sqlite3_free(pFile
->lockingContext
);
3086 rc
= closeUnixFile(id
);
3091 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3093 ** The code above is the AFP lock implementation. The code is specific
3094 ** to MacOSX and does not work on other unix platforms. No alternative
3095 ** is available. If you don't compile for a mac, then the "unix-afp"
3096 ** VFS is not available.
3098 ********************* End of the AFP lock implementation **********************
3099 ******************************************************************************/
3101 /******************************************************************************
3102 *************************** Begin NFS Locking ********************************/
3104 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
3106 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
3107 ** must be either NO_LOCK or SHARED_LOCK.
3109 ** If the locking level of the file descriptor is already at or below
3110 ** the requested locking level, this routine is a no-op.
3112 static int nfsUnlock(sqlite3_file
*id
, int eFileLock
){
3113 return posixUnlock(id
, eFileLock
, 1);
3116 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
3118 ** The code above is the NFS lock implementation. The code is specific
3119 ** to MacOSX and does not work on other unix platforms. No alternative
3122 ********************* End of the NFS lock implementation **********************
3123 ******************************************************************************/
3125 /******************************************************************************
3126 **************** Non-locking sqlite3_file methods *****************************
3128 ** The next division contains implementations for all methods of the
3129 ** sqlite3_file object other than the locking methods. The locking
3130 ** methods were defined in divisions above (one locking method per
3131 ** division). Those methods that are common to all locking modes
3132 ** are gather together into this division.
3136 ** Seek to the offset passed as the second argument, then read cnt
3137 ** bytes into pBuf. Return the number of bytes actually read.
3139 ** NB: If you define USE_PREAD or USE_PREAD64, then it might also
3140 ** be necessary to define _XOPEN_SOURCE to be 500. This varies from
3141 ** one system to another. Since SQLite does not define USE_PREAD
3142 ** in any form by default, we will not attempt to define _XOPEN_SOURCE.
3143 ** See tickets #2741 and #2681.
3145 ** To avoid stomping the errno value on a failed read the lastErrno value
3146 ** is set before returning.
3148 static int seekAndRead(unixFile
*id
, sqlite3_int64 offset
, void *pBuf
, int cnt
){
3151 #if (!defined(USE_PREAD) && !defined(USE_PREAD64))
3155 assert( cnt
==(cnt
&0x1ffff) );
3158 #if defined(USE_PREAD)
3159 got
= osPread(id
->h
, pBuf
, cnt
, offset
);
3160 SimulateIOError( got
= -1 );
3161 #elif defined(USE_PREAD64)
3162 got
= osPread64(id
->h
, pBuf
, cnt
, offset
);
3163 SimulateIOError( got
= -1 );
3165 newOffset
= lseek(id
->h
, offset
, SEEK_SET
);
3166 SimulateIOError( newOffset
= -1 );
3168 storeLastErrno((unixFile
*)id
, errno
);
3171 got
= osRead(id
->h
, pBuf
, cnt
);
3173 if( got
==cnt
) break;
3175 if( errno
==EINTR
){ got
= 1; continue; }
3177 storeLastErrno((unixFile
*)id
, errno
);
3183 pBuf
= (void*)(got
+ (char*)pBuf
);
3187 OSTRACE(("READ %-3d %5d %7lld %llu\n",
3188 id
->h
, got
+prior
, offset
-prior
, TIMER_ELAPSED
));
3193 ** Read data from a file into a buffer. Return SQLITE_OK if all
3194 ** bytes were read successfully and SQLITE_IOERR if anything goes
3197 static int unixRead(
3201 sqlite3_int64 offset
3203 unixFile
*pFile
= (unixFile
*)id
;
3206 assert( offset
>=0 );
3209 /* If this is a database file (not a journal, master-journal or temp
3210 ** file), the bytes in the locking range should never be read or written. */
3212 assert( pFile
->pPreallocatedUnused
==0
3213 || offset
>=PENDING_BYTE
+512
3214 || offset
+amt
<=PENDING_BYTE
3218 #if SQLITE_MAX_MMAP_SIZE>0
3219 /* Deal with as much of this read request as possible by transfering
3220 ** data from the memory mapping using memcpy(). */
3221 if( offset
<pFile
->mmapSize
){
3222 if( offset
+amt
<= pFile
->mmapSize
){
3223 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], amt
);
3226 int nCopy
= pFile
->mmapSize
- offset
;
3227 memcpy(pBuf
, &((u8
*)(pFile
->pMapRegion
))[offset
], nCopy
);
3228 pBuf
= &((u8
*)pBuf
)[nCopy
];
3235 got
= seekAndRead(pFile
, offset
, pBuf
, amt
);
3239 /* lastErrno set by seekAndRead */
3240 return SQLITE_IOERR_READ
;
3242 storeLastErrno(pFile
, 0); /* not a system error */
3243 /* Unread parts of the buffer must be zero-filled */
3244 memset(&((char*)pBuf
)[got
], 0, amt
-got
);
3245 return SQLITE_IOERR_SHORT_READ
;
3250 ** Attempt to seek the file-descriptor passed as the first argument to
3251 ** absolute offset iOff, then attempt to write nBuf bytes of data from
3252 ** pBuf to it. If an error occurs, return -1 and set *piErrno. Otherwise,
3253 ** return the actual number of bytes written (which may be less than
3256 static int seekAndWriteFd(
3257 int fd
, /* File descriptor to write to */
3258 i64 iOff
, /* File offset to begin writing at */
3259 const void *pBuf
, /* Copy data from this buffer to the file */
3260 int nBuf
, /* Size of buffer pBuf in bytes */
3261 int *piErrno
/* OUT: Error number if error occurs */
3263 int rc
= 0; /* Value returned by system call */
3265 assert( nBuf
==(nBuf
&0x1ffff) );
3267 assert( piErrno
!=0 );
3271 #if defined(USE_PREAD)
3272 do{ rc
= (int)osPwrite(fd
, pBuf
, nBuf
, iOff
); }while( rc
<0 && errno
==EINTR
);
3273 #elif defined(USE_PREAD64)
3274 do{ rc
= (int)osPwrite64(fd
, pBuf
, nBuf
, iOff
);}while( rc
<0 && errno
==EINTR
);
3277 i64 iSeek
= lseek(fd
, iOff
, SEEK_SET
);
3278 SimulateIOError( iSeek
= -1 );
3283 rc
= osWrite(fd
, pBuf
, nBuf
);
3284 }while( rc
<0 && errno
==EINTR
);
3288 OSTRACE(("WRITE %-3d %5d %7lld %llu\n", fd
, rc
, iOff
, TIMER_ELAPSED
));
3290 if( rc
<0 ) *piErrno
= errno
;
3296 ** Seek to the offset in id->offset then read cnt bytes into pBuf.
3297 ** Return the number of bytes actually read. Update the offset.
3299 ** To avoid stomping the errno value on a failed write the lastErrno value
3300 ** is set before returning.
3302 static int seekAndWrite(unixFile
*id
, i64 offset
, const void *pBuf
, int cnt
){
3303 return seekAndWriteFd(id
->h
, offset
, pBuf
, cnt
, &id
->lastErrno
);
3308 ** Write data from a buffer into a file. Return SQLITE_OK on success
3309 ** or some other error code on failure.
3311 static int unixWrite(
3315 sqlite3_int64 offset
3317 unixFile
*pFile
= (unixFile
*)id
;
3322 /* If this is a database file (not a journal, master-journal or temp
3323 ** file), the bytes in the locking range should never be read or written. */
3325 assert( pFile
->pPreallocatedUnused
==0
3326 || offset
>=PENDING_BYTE
+512
3327 || offset
+amt
<=PENDING_BYTE
3332 /* If we are doing a normal write to a database file (as opposed to
3333 ** doing a hot-journal rollback or a write to some file other than a
3334 ** normal database file) then record the fact that the database
3335 ** has changed. If the transaction counter is modified, record that
3338 if( pFile
->inNormalWrite
){
3339 pFile
->dbUpdate
= 1; /* The database has been modified */
3340 if( offset
<=24 && offset
+amt
>=27 ){
3343 SimulateIOErrorBenign(1);
3344 rc
= seekAndRead(pFile
, 24, oldCntr
, 4);
3345 SimulateIOErrorBenign(0);
3346 if( rc
!=4 || memcmp(oldCntr
, &((char*)pBuf
)[24-offset
], 4)!=0 ){
3347 pFile
->transCntrChng
= 1; /* The transaction counter has changed */
3353 #if defined(SQLITE_MMAP_READWRITE) && SQLITE_MAX_MMAP_SIZE>0
3354 /* Deal with as much of this write request as possible by transfering
3355 ** data from the memory mapping using memcpy(). */
3356 if( offset
<pFile
->mmapSize
){
3357 if( offset
+amt
<= pFile
->mmapSize
){
3358 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, amt
);
3361 int nCopy
= pFile
->mmapSize
- offset
;
3362 memcpy(&((u8
*)(pFile
->pMapRegion
))[offset
], pBuf
, nCopy
);
3363 pBuf
= &((u8
*)pBuf
)[nCopy
];
3370 while( (wrote
= seekAndWrite(pFile
, offset
, pBuf
, amt
))<amt
&& wrote
>0 ){
3373 pBuf
= &((char*)pBuf
)[wrote
];
3375 SimulateIOError(( wrote
=(-1), amt
=1 ));
3376 SimulateDiskfullError(( wrote
=0, amt
=1 ));
3379 if( wrote
<0 && pFile
->lastErrno
!=ENOSPC
){
3380 /* lastErrno set by seekAndWrite */
3381 return SQLITE_IOERR_WRITE
;
3383 storeLastErrno(pFile
, 0); /* not a system error */
3393 ** Count the number of fullsyncs and normal syncs. This is used to test
3394 ** that syncs and fullsyncs are occurring at the right times.
3396 int sqlite3_sync_count
= 0;
3397 int sqlite3_fullsync_count
= 0;
3401 ** We do not trust systems to provide a working fdatasync(). Some do.
3402 ** Others do no. To be safe, we will stick with the (slightly slower)
3403 ** fsync(). If you know that your system does support fdatasync() correctly,
3404 ** then simply compile with -Dfdatasync=fdatasync or -DHAVE_FDATASYNC
3406 #if !defined(fdatasync) && !HAVE_FDATASYNC
3407 # define fdatasync fsync
3411 ** Define HAVE_FULLFSYNC to 0 or 1 depending on whether or not
3412 ** the F_FULLFSYNC macro is defined. F_FULLFSYNC is currently
3413 ** only available on Mac OS X. But that could change.
3416 # define HAVE_FULLFSYNC 1
3418 # define HAVE_FULLFSYNC 0
3423 ** The fsync() system call does not work as advertised on many
3424 ** unix systems. The following procedure is an attempt to make
3427 ** The SQLITE_NO_SYNC macro disables all fsync()s. This is useful
3428 ** for testing when we want to run through the test suite quickly.
3429 ** You are strongly advised *not* to deploy with SQLITE_NO_SYNC
3430 ** enabled, however, since with SQLITE_NO_SYNC enabled, an OS crash
3431 ** or power failure will likely corrupt the database file.
3433 ** SQLite sets the dataOnly flag if the size of the file is unchanged.
3434 ** The idea behind dataOnly is that it should only write the file content
3435 ** to disk, not the inode. We only set dataOnly if the file size is
3436 ** unchanged since the file size is part of the inode. However,
3437 ** Ted Ts'o tells us that fdatasync() will also write the inode if the
3438 ** file size has changed. The only real difference between fdatasync()
3439 ** and fsync(), Ted tells us, is that fdatasync() will not flush the
3440 ** inode if the mtime or owner or other inode attributes have changed.
3441 ** We only care about the file size, not the other file attributes, so
3442 ** as far as SQLite is concerned, an fdatasync() is always adequate.
3443 ** So, we always use fdatasync() if it is available, regardless of
3444 ** the value of the dataOnly flag.
3446 static int full_fsync(int fd
, int fullSync
, int dataOnly
){
3449 /* The following "ifdef/elif/else/" block has the same structure as
3450 ** the one below. It is replicated here solely to avoid cluttering
3451 ** up the real code with the UNUSED_PARAMETER() macros.
3453 #ifdef SQLITE_NO_SYNC
3454 UNUSED_PARAMETER(fd
);
3455 UNUSED_PARAMETER(fullSync
);
3456 UNUSED_PARAMETER(dataOnly
);
3457 #elif HAVE_FULLFSYNC
3458 UNUSED_PARAMETER(dataOnly
);
3460 UNUSED_PARAMETER(fullSync
);
3461 UNUSED_PARAMETER(dataOnly
);
3464 /* Record the number of times that we do a normal fsync() and
3465 ** FULLSYNC. This is used during testing to verify that this procedure
3466 ** gets called with the correct arguments.
3469 if( fullSync
) sqlite3_fullsync_count
++;
3470 sqlite3_sync_count
++;
3473 /* If we compiled with the SQLITE_NO_SYNC flag, then syncing is a
3474 ** no-op. But go ahead and call fstat() to validate the file
3475 ** descriptor as we need a method to provoke a failure during
3476 ** coverate testing.
3478 #ifdef SQLITE_NO_SYNC
3481 rc
= osFstat(fd
, &buf
);
3483 #elif HAVE_FULLFSYNC
3485 rc
= osFcntl(fd
, F_FULLFSYNC
, 0);
3489 /* If the FULLFSYNC failed, fall back to attempting an fsync().
3490 ** It shouldn't be possible for fullfsync to fail on the local
3491 ** file system (on OSX), so failure indicates that FULLFSYNC
3492 ** isn't supported for this file system. So, attempt an fsync
3493 ** and (for now) ignore the overhead of a superfluous fcntl call.
3494 ** It'd be better to detect fullfsync support once and avoid
3495 ** the fcntl call every time sync is called.
3497 if( rc
) rc
= fsync(fd
);
3499 #elif defined(__APPLE__)
3500 /* fdatasync() on HFS+ doesn't yet flush the file size if it changed correctly
3501 ** so currently we default to the macro that redefines fdatasync to fsync
3507 if( rc
==-1 && errno
==ENOTSUP
){
3510 #endif /* OS_VXWORKS */
3511 #endif /* ifdef SQLITE_NO_SYNC elif HAVE_FULLFSYNC */
3513 if( OS_VXWORKS
&& rc
!= -1 ){
3520 ** Open a file descriptor to the directory containing file zFilename.
3521 ** If successful, *pFd is set to the opened file descriptor and
3522 ** SQLITE_OK is returned. If an error occurs, either SQLITE_NOMEM
3523 ** or SQLITE_CANTOPEN is returned and *pFd is set to an undefined
3526 ** The directory file descriptor is used for only one thing - to
3527 ** fsync() a directory to make sure file creation and deletion events
3528 ** are flushed to disk. Such fsyncs are not needed on newer
3529 ** journaling filesystems, but are required on older filesystems.
3531 ** This routine can be overridden using the xSetSysCall interface.
3532 ** The ability to override this routine was added in support of the
3533 ** chromium sandbox. Opening a directory is a security risk (we are
3534 ** told) so making it overrideable allows the chromium sandbox to
3535 ** replace this routine with a harmless no-op. To make this routine
3536 ** a no-op, replace it with a stub that returns SQLITE_OK but leaves
3537 ** *pFd set to a negative number.
3539 ** If SQLITE_OK is returned, the caller is responsible for closing
3540 ** the file descriptor *pFd using close().
3542 static int openDirectory(const char *zFilename
, int *pFd
){
3545 char zDirname
[MAX_PATHNAME
+1];
3547 sqlite3_snprintf(MAX_PATHNAME
, zDirname
, "%s", zFilename
);
3548 for(ii
=(int)strlen(zDirname
); ii
>0 && zDirname
[ii
]!='/'; ii
--);
3550 zDirname
[ii
] = '\0';
3552 if( zDirname
[0]!='/' ) zDirname
[0] = '.';
3555 fd
= robust_open(zDirname
, O_RDONLY
|O_BINARY
, 0);
3557 OSTRACE(("OPENDIR %-3d %s\n", fd
, zDirname
));
3560 if( fd
>=0 ) return SQLITE_OK
;
3561 return unixLogError(SQLITE_CANTOPEN_BKPT
, "openDirectory", zDirname
);
3565 ** Make sure all writes to a particular file are committed to disk.
3567 ** If dataOnly==0 then both the file itself and its metadata (file
3568 ** size, access time, etc) are synced. If dataOnly!=0 then only the
3569 ** file data is synced.
3571 ** Under Unix, also make sure that the directory entry for the file
3572 ** has been created by fsync-ing the directory that contains the file.
3573 ** If we do not do this and we encounter a power failure, the directory
3574 ** entry for the journal might not exist after we reboot. The next
3575 ** SQLite to access the file will not know that the journal exists (because
3576 ** the directory entry for the journal was never created) and the transaction
3577 ** will not roll back - possibly leading to database corruption.
3579 static int unixSync(sqlite3_file
*id
, int flags
){
3581 unixFile
*pFile
= (unixFile
*)id
;
3583 int isDataOnly
= (flags
&SQLITE_SYNC_DATAONLY
);
3584 int isFullsync
= (flags
&0x0F)==SQLITE_SYNC_FULL
;
3586 /* Check that one of SQLITE_SYNC_NORMAL or FULL was passed */
3587 assert((flags
&0x0F)==SQLITE_SYNC_NORMAL
3588 || (flags
&0x0F)==SQLITE_SYNC_FULL
3591 /* Unix cannot, but some systems may return SQLITE_FULL from here. This
3592 ** line is to test that doing so does not cause any problems.
3594 SimulateDiskfullError( return SQLITE_FULL
);
3597 OSTRACE(("SYNC %-3d\n", pFile
->h
));
3598 rc
= full_fsync(pFile
->h
, isFullsync
, isDataOnly
);
3599 SimulateIOError( rc
=1 );
3601 storeLastErrno(pFile
, errno
);
3602 return unixLogError(SQLITE_IOERR_FSYNC
, "full_fsync", pFile
->zPath
);
3605 /* Also fsync the directory containing the file if the DIRSYNC flag
3606 ** is set. This is a one-time occurrence. Many systems (examples: AIX)
3607 ** are unable to fsync a directory, so ignore errors on the fsync.
3609 if( pFile
->ctrlFlags
& UNIXFILE_DIRSYNC
){
3611 OSTRACE(("DIRSYNC %s (have_fullfsync=%d fullsync=%d)\n", pFile
->zPath
,
3612 HAVE_FULLFSYNC
, isFullsync
));
3613 rc
= osOpenDirectory(pFile
->zPath
, &dirfd
);
3614 if( rc
==SQLITE_OK
){
3615 full_fsync(dirfd
, 0, 0);
3616 robust_close(pFile
, dirfd
, __LINE__
);
3618 assert( rc
==SQLITE_CANTOPEN
);
3621 pFile
->ctrlFlags
&= ~UNIXFILE_DIRSYNC
;
3627 ** Truncate an open file to a specified size
3629 static int unixTruncate(sqlite3_file
*id
, i64 nByte
){
3630 unixFile
*pFile
= (unixFile
*)id
;
3633 SimulateIOError( return SQLITE_IOERR_TRUNCATE
);
3635 /* If the user has configured a chunk-size for this file, truncate the
3636 ** file so that it consists of an integer number of chunks (i.e. the
3637 ** actual file size after the operation may be larger than the requested
3640 if( pFile
->szChunk
>0 ){
3641 nByte
= ((nByte
+ pFile
->szChunk
- 1)/pFile
->szChunk
) * pFile
->szChunk
;
3644 rc
= robust_ftruncate(pFile
->h
, nByte
);
3646 storeLastErrno(pFile
, errno
);
3647 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3650 /* If we are doing a normal write to a database file (as opposed to
3651 ** doing a hot-journal rollback or a write to some file other than a
3652 ** normal database file) and we truncate the file to zero length,
3653 ** that effectively updates the change counter. This might happen
3654 ** when restoring a database using the backup API from a zero-length
3657 if( pFile
->inNormalWrite
&& nByte
==0 ){
3658 pFile
->transCntrChng
= 1;
3662 #if SQLITE_MAX_MMAP_SIZE>0
3663 /* If the file was just truncated to a size smaller than the currently
3664 ** mapped region, reduce the effective mapping size as well. SQLite will
3665 ** use read() and write() to access data beyond this point from now on.
3667 if( nByte
<pFile
->mmapSize
){
3668 pFile
->mmapSize
= nByte
;
3677 ** Determine the current size of a file in bytes
3679 static int unixFileSize(sqlite3_file
*id
, i64
*pSize
){
3683 rc
= osFstat(((unixFile
*)id
)->h
, &buf
);
3684 SimulateIOError( rc
=1 );
3686 storeLastErrno((unixFile
*)id
, errno
);
3687 return SQLITE_IOERR_FSTAT
;
3689 *pSize
= buf
.st_size
;
3691 /* When opening a zero-size database, the findInodeInfo() procedure
3692 ** writes a single byte into that file in order to work around a bug
3693 ** in the OS-X msdos filesystem. In order to avoid problems with upper
3694 ** layers, we need to report this file size as zero even though it is
3695 ** really 1. Ticket #3260.
3697 if( *pSize
==1 ) *pSize
= 0;
3703 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3705 ** Handler for proxy-locking file-control verbs. Defined below in the
3706 ** proxying locking division.
3708 static int proxyFileControl(sqlite3_file
*,int,void*);
3712 ** This function is called to handle the SQLITE_FCNTL_SIZE_HINT
3713 ** file-control operation. Enlarge the database to nBytes in size
3714 ** (rounded up to the next chunk-size). If the database is already
3715 ** nBytes or larger, this routine is a no-op.
3717 static int fcntlSizeHint(unixFile
*pFile
, i64 nByte
){
3718 if( pFile
->szChunk
>0 ){
3719 i64 nSize
; /* Required file size */
3720 struct stat buf
; /* Used to hold return values of fstat() */
3722 if( osFstat(pFile
->h
, &buf
) ){
3723 return SQLITE_IOERR_FSTAT
;
3726 nSize
= ((nByte
+pFile
->szChunk
-1) / pFile
->szChunk
) * pFile
->szChunk
;
3727 if( nSize
>(i64
)buf
.st_size
){
3729 #if defined(HAVE_POSIX_FALLOCATE) && HAVE_POSIX_FALLOCATE
3730 /* The code below is handling the return value of osFallocate()
3731 ** correctly. posix_fallocate() is defined to "returns zero on success,
3732 ** or an error number on failure". See the manpage for details. */
3735 err
= osFallocate(pFile
->h
, buf
.st_size
, nSize
-buf
.st_size
);
3736 }while( err
==EINTR
);
3737 if( err
) return SQLITE_IOERR_WRITE
;
3739 /* If the OS does not have posix_fallocate(), fake it. Write a
3740 ** single byte to the last byte in each block that falls entirely
3741 ** within the extended region. Then, if required, a single byte
3742 ** at offset (nSize-1), to set the size of the file correctly.
3743 ** This is a similar technique to that used by glibc on systems
3744 ** that do not have a real fallocate() call.
3746 int nBlk
= buf
.st_blksize
; /* File-system block size */
3747 int nWrite
= 0; /* Number of bytes written by seekAndWrite */
3748 i64 iWrite
; /* Next offset to write to */
3750 iWrite
= (buf
.st_size
/nBlk
)*nBlk
+ nBlk
- 1;
3751 assert( iWrite
>=buf
.st_size
);
3752 assert( ((iWrite
+1)%nBlk
)==0 );
3753 for(/*no-op*/; iWrite
<nSize
+nBlk
-1; iWrite
+=nBlk
){
3754 if( iWrite
>=nSize
) iWrite
= nSize
- 1;
3755 nWrite
= seekAndWrite(pFile
, iWrite
, "", 1);
3756 if( nWrite
!=1 ) return SQLITE_IOERR_WRITE
;
3762 #if SQLITE_MAX_MMAP_SIZE>0
3763 if( pFile
->mmapSizeMax
>0 && nByte
>pFile
->mmapSize
){
3765 if( pFile
->szChunk
<=0 ){
3766 if( robust_ftruncate(pFile
->h
, nByte
) ){
3767 storeLastErrno(pFile
, errno
);
3768 return unixLogError(SQLITE_IOERR_TRUNCATE
, "ftruncate", pFile
->zPath
);
3772 rc
= unixMapfile(pFile
, nByte
);
3781 ** If *pArg is initially negative then this is a query. Set *pArg to
3782 ** 1 or 0 depending on whether or not bit mask of pFile->ctrlFlags is set.
3784 ** If *pArg is 0 or 1, then clear or set the mask bit of pFile->ctrlFlags.
3786 static void unixModeBit(unixFile
*pFile
, unsigned char mask
, int *pArg
){
3788 *pArg
= (pFile
->ctrlFlags
& mask
)!=0;
3789 }else if( (*pArg
)==0 ){
3790 pFile
->ctrlFlags
&= ~mask
;
3792 pFile
->ctrlFlags
|= mask
;
3796 /* Forward declaration */
3797 static int unixGetTempname(int nBuf
, char *zBuf
);
3800 ** Information and control of an open file handle.
3802 static int unixFileControl(sqlite3_file
*id
, int op
, void *pArg
){
3803 unixFile
*pFile
= (unixFile
*)id
;
3805 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3806 case SQLITE_FCNTL_BEGIN_ATOMIC_WRITE
: {
3807 int rc
= osIoctl(pFile
->h
, F2FS_IOC_START_ATOMIC_WRITE
);
3808 return rc
? SQLITE_IOERR_BEGIN_ATOMIC
: SQLITE_OK
;
3810 case SQLITE_FCNTL_COMMIT_ATOMIC_WRITE
: {
3811 int rc
= osIoctl(pFile
->h
, F2FS_IOC_COMMIT_ATOMIC_WRITE
);
3812 return rc
? SQLITE_IOERR_COMMIT_ATOMIC
: SQLITE_OK
;
3814 case SQLITE_FCNTL_ROLLBACK_ATOMIC_WRITE
: {
3815 int rc
= osIoctl(pFile
->h
, F2FS_IOC_ABORT_VOLATILE_WRITE
);
3816 return rc
? SQLITE_IOERR_ROLLBACK_ATOMIC
: SQLITE_OK
;
3818 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3820 case SQLITE_FCNTL_LOCKSTATE
: {
3821 *(int*)pArg
= pFile
->eFileLock
;
3824 case SQLITE_FCNTL_LAST_ERRNO
: {
3825 *(int*)pArg
= pFile
->lastErrno
;
3828 case SQLITE_FCNTL_CHUNK_SIZE
: {
3829 pFile
->szChunk
= *(int *)pArg
;
3832 case SQLITE_FCNTL_SIZE_HINT
: {
3834 SimulateIOErrorBenign(1);
3835 rc
= fcntlSizeHint(pFile
, *(i64
*)pArg
);
3836 SimulateIOErrorBenign(0);
3839 case SQLITE_FCNTL_PERSIST_WAL
: {
3840 unixModeBit(pFile
, UNIXFILE_PERSIST_WAL
, (int*)pArg
);
3843 case SQLITE_FCNTL_POWERSAFE_OVERWRITE
: {
3844 unixModeBit(pFile
, UNIXFILE_PSOW
, (int*)pArg
);
3847 case SQLITE_FCNTL_VFSNAME
: {
3848 *(char**)pArg
= sqlite3_mprintf("%s", pFile
->pVfs
->zName
);
3851 case SQLITE_FCNTL_TEMPFILENAME
: {
3852 char *zTFile
= sqlite3_malloc64( pFile
->pVfs
->mxPathname
);
3854 unixGetTempname(pFile
->pVfs
->mxPathname
, zTFile
);
3855 *(char**)pArg
= zTFile
;
3859 case SQLITE_FCNTL_HAS_MOVED
: {
3860 *(int*)pArg
= fileHasMoved(pFile
);
3863 #if SQLITE_MAX_MMAP_SIZE>0
3864 case SQLITE_FCNTL_MMAP_SIZE
: {
3865 i64 newLimit
= *(i64
*)pArg
;
3867 if( newLimit
>sqlite3GlobalConfig
.mxMmap
){
3868 newLimit
= sqlite3GlobalConfig
.mxMmap
;
3871 /* The value of newLimit may be eventually cast to (size_t) and passed
3872 ** to mmap(). Restrict its value to 2GB if (size_t) is not at least a
3874 if( newLimit
>0 && sizeof(size_t)<8 ){
3875 newLimit
= (newLimit
& 0x7FFFFFFF);
3878 *(i64
*)pArg
= pFile
->mmapSizeMax
;
3879 if( newLimit
>=0 && newLimit
!=pFile
->mmapSizeMax
&& pFile
->nFetchOut
==0 ){
3880 pFile
->mmapSizeMax
= newLimit
;
3881 if( pFile
->mmapSize
>0 ){
3882 unixUnmapfile(pFile
);
3883 rc
= unixMapfile(pFile
, -1);
3890 /* The pager calls this method to signal that it has done
3891 ** a rollback and that the database is therefore unchanged and
3892 ** it hence it is OK for the transaction change counter to be
3895 case SQLITE_FCNTL_DB_UNCHANGED
: {
3896 ((unixFile
*)id
)->dbUpdate
= 0;
3900 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
3901 case SQLITE_FCNTL_SET_LOCKPROXYFILE
:
3902 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
3903 return proxyFileControl(id
,op
,pArg
);
3905 #endif /* SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__) */
3907 return SQLITE_NOTFOUND
;
3911 ** If pFd->sectorSize is non-zero when this function is called, it is a
3912 ** no-op. Otherwise, the values of pFd->sectorSize and
3913 ** pFd->deviceCharacteristics are set according to the file-system
3916 ** There are two versions of this function. One for QNX and one for all
3920 static void setDeviceCharacteristics(unixFile
*pFd
){
3921 assert( pFd
->deviceCharacteristics
==0 || pFd
->sectorSize
!=0 );
3922 if( pFd
->sectorSize
==0 ){
3923 #if defined(__linux__) && defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
3927 /* Check for support for F2FS atomic batch writes. */
3928 res
= osIoctl(pFd
->h
, F2FS_IOC_GET_FEATURES
, &f
);
3929 if( res
==0 && (f
& F2FS_FEATURE_ATOMIC_WRITE
) ){
3930 pFd
->deviceCharacteristics
= SQLITE_IOCAP_BATCH_ATOMIC
;
3932 #endif /* __linux__ && SQLITE_ENABLE_BATCH_ATOMIC_WRITE */
3934 /* Set the POWERSAFE_OVERWRITE flag if requested. */
3935 if( pFd
->ctrlFlags
& UNIXFILE_PSOW
){
3936 pFd
->deviceCharacteristics
|= SQLITE_IOCAP_POWERSAFE_OVERWRITE
;
3939 pFd
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
3943 #include <sys/dcmd_blk.h>
3944 #include <sys/statvfs.h>
3945 static void setDeviceCharacteristics(unixFile
*pFile
){
3946 if( pFile
->sectorSize
== 0 ){
3947 struct statvfs fsInfo
;
3949 /* Set defaults for non-supported filesystems */
3950 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
3951 pFile
->deviceCharacteristics
= 0;
3952 if( fstatvfs(pFile
->h
, &fsInfo
) == -1 ) {
3953 return pFile
->sectorSize
;
3956 if( !strcmp(fsInfo
.f_basetype
, "tmp") ) {
3957 pFile
->sectorSize
= fsInfo
.f_bsize
;
3958 pFile
->deviceCharacteristics
=
3959 SQLITE_IOCAP_ATOMIC4K
| /* All ram filesystem writes are atomic */
3960 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3961 ** the write succeeds */
3962 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3963 ** so it is ordered */
3965 }else if( strstr(fsInfo
.f_basetype
, "etfs") ){
3966 pFile
->sectorSize
= fsInfo
.f_bsize
;
3967 pFile
->deviceCharacteristics
=
3968 /* etfs cluster size writes are atomic */
3969 (pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) |
3970 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3971 ** the write succeeds */
3972 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3973 ** so it is ordered */
3975 }else if( !strcmp(fsInfo
.f_basetype
, "qnx6") ){
3976 pFile
->sectorSize
= fsInfo
.f_bsize
;
3977 pFile
->deviceCharacteristics
=
3978 SQLITE_IOCAP_ATOMIC
| /* All filesystem writes are atomic */
3979 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
3980 ** the write succeeds */
3981 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3982 ** so it is ordered */
3984 }else if( !strcmp(fsInfo
.f_basetype
, "qnx4") ){
3985 pFile
->sectorSize
= fsInfo
.f_bsize
;
3986 pFile
->deviceCharacteristics
=
3987 /* full bitset of atomics from max sector size and smaller */
3988 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
3989 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3990 ** so it is ordered */
3992 }else if( strstr(fsInfo
.f_basetype
, "dos") ){
3993 pFile
->sectorSize
= fsInfo
.f_bsize
;
3994 pFile
->deviceCharacteristics
=
3995 /* full bitset of atomics from max sector size and smaller */
3996 ((pFile
->sectorSize
/ 512 * SQLITE_IOCAP_ATOMIC512
) << 1) - 2 |
3997 SQLITE_IOCAP_SEQUENTIAL
| /* The ram filesystem has no write behind
3998 ** so it is ordered */
4001 pFile
->deviceCharacteristics
=
4002 SQLITE_IOCAP_ATOMIC512
| /* blocks are atomic */
4003 SQLITE_IOCAP_SAFE_APPEND
| /* growing the file does not occur until
4004 ** the write succeeds */
4008 /* Last chance verification. If the sector size isn't a multiple of 512
4009 ** then it isn't valid.*/
4010 if( pFile
->sectorSize
% 512 != 0 ){
4011 pFile
->deviceCharacteristics
= 0;
4012 pFile
->sectorSize
= SQLITE_DEFAULT_SECTOR_SIZE
;
4018 ** Return the sector size in bytes of the underlying block device for
4019 ** the specified file. This is almost always 512 bytes, but may be
4020 ** larger for some devices.
4022 ** SQLite code assumes this function cannot fail. It also assumes that
4023 ** if two files are created in the same file-system directory (i.e.
4024 ** a database and its journal file) that the sector size will be the
4027 static int unixSectorSize(sqlite3_file
*id
){
4028 unixFile
*pFd
= (unixFile
*)id
;
4029 setDeviceCharacteristics(pFd
);
4030 return pFd
->sectorSize
;
4034 ** Return the device characteristics for the file.
4036 ** This VFS is set up to return SQLITE_IOCAP_POWERSAFE_OVERWRITE by default.
4037 ** However, that choice is controversial since technically the underlying
4038 ** file system does not always provide powersafe overwrites. (In other
4039 ** words, after a power-loss event, parts of the file that were never
4040 ** written might end up being altered.) However, non-PSOW behavior is very,
4041 ** very rare. And asserting PSOW makes a large reduction in the amount
4042 ** of required I/O for journaling, since a lot of padding is eliminated.
4043 ** Hence, while POWERSAFE_OVERWRITE is on by default, there is a file-control
4044 ** available to turn it off and URI query parameter available to turn it off.
4046 static int unixDeviceCharacteristics(sqlite3_file
*id
){
4047 unixFile
*pFd
= (unixFile
*)id
;
4048 setDeviceCharacteristics(pFd
);
4049 return pFd
->deviceCharacteristics
;
4052 #if !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0
4055 ** Return the system page size.
4057 ** This function should not be called directly by other code in this file.
4058 ** Instead, it should be called via macro osGetpagesize().
4060 static int unixGetpagesize(void){
4063 #elif defined(_BSD_SOURCE)
4064 return getpagesize();
4066 return (int)sysconf(_SC_PAGESIZE
);
4070 #endif /* !defined(SQLITE_OMIT_WAL) || SQLITE_MAX_MMAP_SIZE>0 */
4072 #ifndef SQLITE_OMIT_WAL
4075 ** Object used to represent an shared memory buffer.
4077 ** When multiple threads all reference the same wal-index, each thread
4078 ** has its own unixShm object, but they all point to a single instance
4079 ** of this unixShmNode object. In other words, each wal-index is opened
4080 ** only once per process.
4082 ** Each unixShmNode object is connected to a single unixInodeInfo object.
4083 ** We could coalesce this object into unixInodeInfo, but that would mean
4084 ** every open file that does not use shared memory (in other words, most
4085 ** open files) would have to carry around this extra information. So
4086 ** the unixInodeInfo object contains a pointer to this unixShmNode object
4087 ** and the unixShmNode object is created only when needed.
4089 ** unixMutexHeld() must be true when creating or destroying
4090 ** this object or while reading or writing the following fields:
4094 ** The following fields are read-only after the object is created:
4099 ** Either unixShmNode.mutex must be held or unixShmNode.nRef==0 and
4100 ** unixMutexHeld() is true when reading or writing any other field
4101 ** in this structure.
4103 struct unixShmNode
{
4104 unixInodeInfo
*pInode
; /* unixInodeInfo that owns this SHM node */
4105 sqlite3_mutex
*mutex
; /* Mutex to access this object */
4106 char *zFilename
; /* Name of the mmapped file */
4107 int h
; /* Open file descriptor */
4108 int szRegion
; /* Size of shared-memory regions */
4109 u16 nRegion
; /* Size of array apRegion */
4110 u8 isReadonly
; /* True if read-only */
4111 u8 isUnlocked
; /* True if no DMS lock held */
4112 char **apRegion
; /* Array of mapped shared-memory regions */
4113 int nRef
; /* Number of unixShm objects pointing to this */
4114 unixShm
*pFirst
; /* All unixShm objects pointing to this */
4116 u8 exclMask
; /* Mask of exclusive locks held */
4117 u8 sharedMask
; /* Mask of shared locks held */
4118 u8 nextShmId
; /* Next available unixShm.id value */
4123 ** Structure used internally by this VFS to record the state of an
4124 ** open shared memory connection.
4126 ** The following fields are initialized when this object is created and
4127 ** are read-only thereafter:
4132 ** All other fields are read/write. The unixShm.pFile->mutex must be held
4133 ** while accessing any read/write fields.
4136 unixShmNode
*pShmNode
; /* The underlying unixShmNode object */
4137 unixShm
*pNext
; /* Next unixShm with the same unixShmNode */
4138 u8 hasMutex
; /* True if holding the unixShmNode mutex */
4139 u8 id
; /* Id of this connection within its unixShmNode */
4140 u16 sharedMask
; /* Mask of shared locks held */
4141 u16 exclMask
; /* Mask of exclusive locks held */
4145 ** Constants used for locking
4147 #define UNIX_SHM_BASE ((22+SQLITE_SHM_NLOCK)*4) /* first lock byte */
4148 #define UNIX_SHM_DMS (UNIX_SHM_BASE+SQLITE_SHM_NLOCK) /* deadman switch */
4151 ** Apply posix advisory locks for all bytes from ofst through ofst+n-1.
4153 ** Locks block if the mask is exactly UNIX_SHM_C and are non-blocking
4156 static int unixShmSystemLock(
4157 unixFile
*pFile
, /* Open connection to the WAL file */
4158 int lockType
, /* F_UNLCK, F_RDLCK, or F_WRLCK */
4159 int ofst
, /* First byte of the locking range */
4160 int n
/* Number of bytes to lock */
4162 unixShmNode
*pShmNode
; /* Apply locks to this open shared-memory segment */
4163 struct flock f
; /* The posix advisory locking structure */
4164 int rc
= SQLITE_OK
; /* Result code form fcntl() */
4166 /* Access to the unixShmNode object is serialized by the caller */
4167 pShmNode
= pFile
->pInode
->pShmNode
;
4168 assert( sqlite3_mutex_held(pShmNode
->mutex
) || pShmNode
->nRef
==0 );
4170 /* Shared locks never span more than one byte */
4171 assert( n
==1 || lockType
!=F_RDLCK
);
4173 /* Locks are within range */
4174 assert( n
>=1 && n
<=SQLITE_SHM_NLOCK
);
4176 if( pShmNode
->h
>=0 ){
4177 /* Initialize the locking parameters */
4178 memset(&f
, 0, sizeof(f
));
4179 f
.l_type
= lockType
;
4180 f
.l_whence
= SEEK_SET
;
4184 rc
= osFcntl(pShmNode
->h
, F_SETLK
, &f
);
4185 rc
= (rc
!=(-1)) ? SQLITE_OK
: SQLITE_BUSY
;
4188 /* Update the global lock state and do debug tracing */
4191 OSTRACE(("SHM-LOCK "));
4192 mask
= ofst
>31 ? 0xffff : (1<<(ofst
+n
)) - (1<<ofst
);
4193 if( rc
==SQLITE_OK
){
4194 if( lockType
==F_UNLCK
){
4195 OSTRACE(("unlock %d ok", ofst
));
4196 pShmNode
->exclMask
&= ~mask
;
4197 pShmNode
->sharedMask
&= ~mask
;
4198 }else if( lockType
==F_RDLCK
){
4199 OSTRACE(("read-lock %d ok", ofst
));
4200 pShmNode
->exclMask
&= ~mask
;
4201 pShmNode
->sharedMask
|= mask
;
4203 assert( lockType
==F_WRLCK
);
4204 OSTRACE(("write-lock %d ok", ofst
));
4205 pShmNode
->exclMask
|= mask
;
4206 pShmNode
->sharedMask
&= ~mask
;
4209 if( lockType
==F_UNLCK
){
4210 OSTRACE(("unlock %d failed", ofst
));
4211 }else if( lockType
==F_RDLCK
){
4212 OSTRACE(("read-lock failed"));
4214 assert( lockType
==F_WRLCK
);
4215 OSTRACE(("write-lock %d failed", ofst
));
4218 OSTRACE((" - afterwards %03x,%03x\n",
4219 pShmNode
->sharedMask
, pShmNode
->exclMask
));
4227 ** Return the minimum number of 32KB shm regions that should be mapped at
4228 ** a time, assuming that each mapping must be an integer multiple of the
4229 ** current system page-size.
4231 ** Usually, this is 1. The exception seems to be systems that are configured
4232 ** to use 64KB pages - in this case each mapping must cover at least two
4235 static int unixShmRegionPerMap(void){
4236 int shmsz
= 32*1024; /* SHM region size */
4237 int pgsz
= osGetpagesize(); /* System page size */
4238 assert( ((pgsz
-1)&pgsz
)==0 ); /* Page size must be a power of 2 */
4239 if( pgsz
<shmsz
) return 1;
4244 ** Purge the unixShmNodeList list of all entries with unixShmNode.nRef==0.
4246 ** This is not a VFS shared-memory method; it is a utility function called
4247 ** by VFS shared-memory methods.
4249 static void unixShmPurge(unixFile
*pFd
){
4250 unixShmNode
*p
= pFd
->pInode
->pShmNode
;
4251 assert( unixMutexHeld() );
4252 if( p
&& ALWAYS(p
->nRef
==0) ){
4253 int nShmPerMap
= unixShmRegionPerMap();
4255 assert( p
->pInode
==pFd
->pInode
);
4256 sqlite3_mutex_free(p
->mutex
);
4257 for(i
=0; i
<p
->nRegion
; i
+=nShmPerMap
){
4259 osMunmap(p
->apRegion
[i
], p
->szRegion
);
4261 sqlite3_free(p
->apRegion
[i
]);
4264 sqlite3_free(p
->apRegion
);
4266 robust_close(pFd
, p
->h
, __LINE__
);
4269 p
->pInode
->pShmNode
= 0;
4275 ** The DMS lock has not yet been taken on shm file pShmNode. Attempt to
4276 ** take it now. Return SQLITE_OK if successful, or an SQLite error
4279 ** If the DMS cannot be locked because this is a readonly_shm=1
4280 ** connection and no other process already holds a lock, return
4281 ** SQLITE_READONLY_CANTINIT and set pShmNode->isUnlocked=1.
4283 static int unixLockSharedMemory(unixFile
*pDbFd
, unixShmNode
*pShmNode
){
4287 /* Use F_GETLK to determine the locks other processes are holding
4288 ** on the DMS byte. If it indicates that another process is holding
4289 ** a SHARED lock, then this process may also take a SHARED lock
4290 ** and proceed with opening the *-shm file.
4292 ** Or, if no other process is holding any lock, then this process
4293 ** is the first to open it. In this case take an EXCLUSIVE lock on the
4294 ** DMS byte and truncate the *-shm file to zero bytes in size. Then
4295 ** downgrade to a SHARED lock on the DMS byte.
4297 ** If another process is holding an EXCLUSIVE lock on the DMS byte,
4298 ** return SQLITE_BUSY to the caller (it will try again). An earlier
4299 ** version of this code attempted the SHARED lock at this point. But
4300 ** this introduced a subtle race condition: if the process holding
4301 ** EXCLUSIVE failed just before truncating the *-shm file, then this
4302 ** process might open and use the *-shm file without truncating it.
4303 ** And if the *-shm file has been corrupted by a power failure or
4304 ** system crash, the database itself may also become corrupt. */
4305 lock
.l_whence
= SEEK_SET
;
4306 lock
.l_start
= UNIX_SHM_DMS
;
4308 lock
.l_type
= F_WRLCK
;
4309 if( osFcntl(pShmNode
->h
, F_GETLK
, &lock
)!=0 ) {
4310 rc
= SQLITE_IOERR_LOCK
;
4311 }else if( lock
.l_type
==F_UNLCK
){
4312 if( pShmNode
->isReadonly
){
4313 pShmNode
->isUnlocked
= 1;
4314 rc
= SQLITE_READONLY_CANTINIT
;
4316 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, UNIX_SHM_DMS
, 1);
4317 if( rc
==SQLITE_OK
&& robust_ftruncate(pShmNode
->h
, 0) ){
4318 rc
= unixLogError(SQLITE_IOERR_SHMOPEN
,"ftruncate",pShmNode
->zFilename
);
4321 }else if( lock
.l_type
==F_WRLCK
){
4325 if( rc
==SQLITE_OK
){
4326 assert( lock
.l_type
==F_UNLCK
|| lock
.l_type
==F_RDLCK
);
4327 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, UNIX_SHM_DMS
, 1);
4333 ** Open a shared-memory area associated with open database file pDbFd.
4334 ** This particular implementation uses mmapped files.
4336 ** The file used to implement shared-memory is in the same directory
4337 ** as the open database file and has the same name as the open database
4338 ** file with the "-shm" suffix added. For example, if the database file
4339 ** is "/home/user1/config.db" then the file that is created and mmapped
4340 ** for shared memory will be called "/home/user1/config.db-shm".
4342 ** Another approach to is to use files in /dev/shm or /dev/tmp or an
4343 ** some other tmpfs mount. But if a file in a different directory
4344 ** from the database file is used, then differing access permissions
4345 ** or a chroot() might cause two different processes on the same
4346 ** database to end up using different files for shared memory -
4347 ** meaning that their memory would not really be shared - resulting
4348 ** in database corruption. Nevertheless, this tmpfs file usage
4349 ** can be enabled at compile-time using -DSQLITE_SHM_DIRECTORY="/dev/shm"
4350 ** or the equivalent. The use of the SQLITE_SHM_DIRECTORY compile-time
4351 ** option results in an incompatible build of SQLite; builds of SQLite
4352 ** that with differing SQLITE_SHM_DIRECTORY settings attempt to use the
4353 ** same database file at the same time, database corruption will likely
4354 ** result. The SQLITE_SHM_DIRECTORY compile-time option is considered
4355 ** "unsupported" and may go away in a future SQLite release.
4357 ** When opening a new shared-memory file, if no other instances of that
4358 ** file are currently open, in this process or in other processes, then
4359 ** the file must be truncated to zero length or have its header cleared.
4361 ** If the original database file (pDbFd) is using the "unix-excl" VFS
4362 ** that means that an exclusive lock is held on the database file and
4363 ** that no other processes are able to read or write the database. In
4364 ** that case, we do not really need shared memory. No shared memory
4365 ** file is created. The shared memory will be simulated with heap memory.
4367 static int unixOpenSharedMemory(unixFile
*pDbFd
){
4368 struct unixShm
*p
= 0; /* The connection to be opened */
4369 struct unixShmNode
*pShmNode
; /* The underlying mmapped file */
4370 int rc
= SQLITE_OK
; /* Result code */
4371 unixInodeInfo
*pInode
; /* The inode of fd */
4372 char *zShm
; /* Name of the file used for SHM */
4373 int nShmFilename
; /* Size of the SHM filename in bytes */
4375 /* Allocate space for the new unixShm object. */
4376 p
= sqlite3_malloc64( sizeof(*p
) );
4377 if( p
==0 ) return SQLITE_NOMEM_BKPT
;
4378 memset(p
, 0, sizeof(*p
));
4379 assert( pDbFd
->pShm
==0 );
4381 /* Check to see if a unixShmNode object already exists. Reuse an existing
4382 ** one if present. Create a new one if necessary.
4385 pInode
= pDbFd
->pInode
;
4386 pShmNode
= pInode
->pShmNode
;
4388 struct stat sStat
; /* fstat() info for database file */
4389 #ifndef SQLITE_SHM_DIRECTORY
4390 const char *zBasePath
= pDbFd
->zPath
;
4393 /* Call fstat() to figure out the permissions on the database file. If
4394 ** a new *-shm file is created, an attempt will be made to create it
4395 ** with the same permissions.
4397 if( osFstat(pDbFd
->h
, &sStat
) ){
4398 rc
= SQLITE_IOERR_FSTAT
;
4402 #ifdef SQLITE_SHM_DIRECTORY
4403 nShmFilename
= sizeof(SQLITE_SHM_DIRECTORY
) + 31;
4405 nShmFilename
= 6 + (int)strlen(zBasePath
);
4407 pShmNode
= sqlite3_malloc64( sizeof(*pShmNode
) + nShmFilename
);
4409 rc
= SQLITE_NOMEM_BKPT
;
4412 memset(pShmNode
, 0, sizeof(*pShmNode
)+nShmFilename
);
4413 zShm
= pShmNode
->zFilename
= (char*)&pShmNode
[1];
4414 #ifdef SQLITE_SHM_DIRECTORY
4415 sqlite3_snprintf(nShmFilename
, zShm
,
4416 SQLITE_SHM_DIRECTORY
"/sqlite-shm-%x-%x",
4417 (u32
)sStat
.st_ino
, (u32
)sStat
.st_dev
);
4419 sqlite3_snprintf(nShmFilename
, zShm
, "%s-shm", zBasePath
);
4420 sqlite3FileSuffix3(pDbFd
->zPath
, zShm
);
4423 pDbFd
->pInode
->pShmNode
= pShmNode
;
4424 pShmNode
->pInode
= pDbFd
->pInode
;
4425 if( sqlite3GlobalConfig
.bCoreMutex
){
4426 pShmNode
->mutex
= sqlite3_mutex_alloc(SQLITE_MUTEX_FAST
);
4427 if( pShmNode
->mutex
==0 ){
4428 rc
= SQLITE_NOMEM_BKPT
;
4433 if( pInode
->bProcessLock
==0 ){
4434 if( 0==sqlite3_uri_boolean(pDbFd
->zPath
, "readonly_shm", 0) ){
4435 pShmNode
->h
= robust_open(zShm
, O_RDWR
|O_CREAT
, (sStat
.st_mode
&0777));
4437 if( pShmNode
->h
<0 ){
4438 pShmNode
->h
= robust_open(zShm
, O_RDONLY
, (sStat
.st_mode
&0777));
4439 if( pShmNode
->h
<0 ){
4440 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zShm
);
4443 pShmNode
->isReadonly
= 1;
4446 /* If this process is running as root, make sure that the SHM file
4447 ** is owned by the same user that owns the original database. Otherwise,
4448 ** the original owner will not be able to connect.
4450 robustFchown(pShmNode
->h
, sStat
.st_uid
, sStat
.st_gid
);
4452 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4453 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_READONLY_CANTINIT
) goto shm_open_err
;
4457 /* Make the new connection a child of the unixShmNode */
4458 p
->pShmNode
= pShmNode
;
4460 p
->id
= pShmNode
->nextShmId
++;
4466 /* The reference count on pShmNode has already been incremented under
4467 ** the cover of the unixEnterMutex() mutex and the pointer from the
4468 ** new (struct unixShm) object to the pShmNode has been set. All that is
4469 ** left to do is to link the new object into the linked list starting
4470 ** at pShmNode->pFirst. This must be done while holding the pShmNode->mutex
4473 sqlite3_mutex_enter(pShmNode
->mutex
);
4474 p
->pNext
= pShmNode
->pFirst
;
4475 pShmNode
->pFirst
= p
;
4476 sqlite3_mutex_leave(pShmNode
->mutex
);
4479 /* Jump here on any error */
4481 unixShmPurge(pDbFd
); /* This call frees pShmNode if required */
4488 ** This function is called to obtain a pointer to region iRegion of the
4489 ** shared-memory associated with the database file fd. Shared-memory regions
4490 ** are numbered starting from zero. Each shared-memory region is szRegion
4493 ** If an error occurs, an error code is returned and *pp is set to NULL.
4495 ** Otherwise, if the bExtend parameter is 0 and the requested shared-memory
4496 ** region has not been allocated (by any client, including one running in a
4497 ** separate process), then *pp is set to NULL and SQLITE_OK returned. If
4498 ** bExtend is non-zero and the requested shared-memory region has not yet
4499 ** been allocated, it is allocated by this function.
4501 ** If the shared-memory region has already been allocated or is allocated by
4502 ** this call as described above, then it is mapped into this processes
4503 ** address space (if it is not already), *pp is set to point to the mapped
4504 ** memory and SQLITE_OK returned.
4506 static int unixShmMap(
4507 sqlite3_file
*fd
, /* Handle open on database file */
4508 int iRegion
, /* Region to retrieve */
4509 int szRegion
, /* Size of regions */
4510 int bExtend
, /* True to extend file if necessary */
4511 void volatile **pp
/* OUT: Mapped memory */
4513 unixFile
*pDbFd
= (unixFile
*)fd
;
4515 unixShmNode
*pShmNode
;
4517 int nShmPerMap
= unixShmRegionPerMap();
4520 /* If the shared-memory file has not yet been opened, open it now. */
4521 if( pDbFd
->pShm
==0 ){
4522 rc
= unixOpenSharedMemory(pDbFd
);
4523 if( rc
!=SQLITE_OK
) return rc
;
4527 pShmNode
= p
->pShmNode
;
4528 sqlite3_mutex_enter(pShmNode
->mutex
);
4529 if( pShmNode
->isUnlocked
){
4530 rc
= unixLockSharedMemory(pDbFd
, pShmNode
);
4531 if( rc
!=SQLITE_OK
) goto shmpage_out
;
4532 pShmNode
->isUnlocked
= 0;
4534 assert( szRegion
==pShmNode
->szRegion
|| pShmNode
->nRegion
==0 );
4535 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4536 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4537 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4539 /* Minimum number of regions required to be mapped. */
4540 nReqRegion
= ((iRegion
+nShmPerMap
) / nShmPerMap
) * nShmPerMap
;
4542 if( pShmNode
->nRegion
<nReqRegion
){
4543 char **apNew
; /* New apRegion[] array */
4544 int nByte
= nReqRegion
*szRegion
; /* Minimum required file size */
4545 struct stat sStat
; /* Used by fstat() */
4547 pShmNode
->szRegion
= szRegion
;
4549 if( pShmNode
->h
>=0 ){
4550 /* The requested region is not mapped into this processes address space.
4551 ** Check to see if it has been allocated (i.e. if the wal-index file is
4552 ** large enough to contain the requested region).
4554 if( osFstat(pShmNode
->h
, &sStat
) ){
4555 rc
= SQLITE_IOERR_SHMSIZE
;
4559 if( sStat
.st_size
<nByte
){
4560 /* The requested memory region does not exist. If bExtend is set to
4561 ** false, exit early. *pp will be set to NULL and SQLITE_OK returned.
4567 /* Alternatively, if bExtend is true, extend the file. Do this by
4568 ** writing a single byte to the end of each (OS) page being
4569 ** allocated or extended. Technically, we need only write to the
4570 ** last page in order to extend the file. But writing to all new
4571 ** pages forces the OS to allocate them immediately, which reduces
4572 ** the chances of SIGBUS while accessing the mapped region later on.
4575 static const int pgsz
= 4096;
4578 /* Write to the last byte of each newly allocated or extended page */
4579 assert( (nByte
% pgsz
)==0 );
4580 for(iPg
=(sStat
.st_size
/pgsz
); iPg
<(nByte
/pgsz
); iPg
++){
4582 if( seekAndWriteFd(pShmNode
->h
, iPg
*pgsz
+ pgsz
-1, "", 1, &x
)!=1 ){
4583 const char *zFile
= pShmNode
->zFilename
;
4584 rc
= unixLogError(SQLITE_IOERR_SHMSIZE
, "write", zFile
);
4592 /* Map the requested memory region into this processes address space. */
4593 apNew
= (char **)sqlite3_realloc(
4594 pShmNode
->apRegion
, nReqRegion
*sizeof(char *)
4597 rc
= SQLITE_IOERR_NOMEM_BKPT
;
4600 pShmNode
->apRegion
= apNew
;
4601 while( pShmNode
->nRegion
<nReqRegion
){
4602 int nMap
= szRegion
*nShmPerMap
;
4605 if( pShmNode
->h
>=0 ){
4606 pMem
= osMmap(0, nMap
,
4607 pShmNode
->isReadonly
? PROT_READ
: PROT_READ
|PROT_WRITE
,
4608 MAP_SHARED
, pShmNode
->h
, szRegion
*(i64
)pShmNode
->nRegion
4610 if( pMem
==MAP_FAILED
){
4611 rc
= unixLogError(SQLITE_IOERR_SHMMAP
, "mmap", pShmNode
->zFilename
);
4615 pMem
= sqlite3_malloc64(szRegion
);
4617 rc
= SQLITE_NOMEM_BKPT
;
4620 memset(pMem
, 0, szRegion
);
4623 for(i
=0; i
<nShmPerMap
; i
++){
4624 pShmNode
->apRegion
[pShmNode
->nRegion
+i
] = &((char*)pMem
)[szRegion
*i
];
4626 pShmNode
->nRegion
+= nShmPerMap
;
4631 if( pShmNode
->nRegion
>iRegion
){
4632 *pp
= pShmNode
->apRegion
[iRegion
];
4636 if( pShmNode
->isReadonly
&& rc
==SQLITE_OK
) rc
= SQLITE_READONLY
;
4637 sqlite3_mutex_leave(pShmNode
->mutex
);
4642 ** Change the lock state for a shared-memory segment.
4644 ** Note that the relationship between SHAREd and EXCLUSIVE locks is a little
4645 ** different here than in posix. In xShmLock(), one can go from unlocked
4646 ** to shared and back or from unlocked to exclusive and back. But one may
4647 ** not go from shared to exclusive or from exclusive to shared.
4649 static int unixShmLock(
4650 sqlite3_file
*fd
, /* Database file holding the shared memory */
4651 int ofst
, /* First lock to acquire or release */
4652 int n
, /* Number of locks to acquire or release */
4653 int flags
/* What to do with the lock */
4655 unixFile
*pDbFd
= (unixFile
*)fd
; /* Connection holding shared memory */
4656 unixShm
*p
= pDbFd
->pShm
; /* The shared memory being locked */
4657 unixShm
*pX
; /* For looping over all siblings */
4658 unixShmNode
*pShmNode
= p
->pShmNode
; /* The underlying file iNode */
4659 int rc
= SQLITE_OK
; /* Result code */
4660 u16 mask
; /* Mask of locks to take or release */
4662 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4663 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4664 assert( ofst
>=0 && ofst
+n
<=SQLITE_SHM_NLOCK
);
4666 assert( flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_SHARED
)
4667 || flags
==(SQLITE_SHM_LOCK
| SQLITE_SHM_EXCLUSIVE
)
4668 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_SHARED
)
4669 || flags
==(SQLITE_SHM_UNLOCK
| SQLITE_SHM_EXCLUSIVE
) );
4670 assert( n
==1 || (flags
& SQLITE_SHM_EXCLUSIVE
)!=0 );
4671 assert( pShmNode
->h
>=0 || pDbFd
->pInode
->bProcessLock
==1 );
4672 assert( pShmNode
->h
<0 || pDbFd
->pInode
->bProcessLock
==0 );
4674 mask
= (1<<(ofst
+n
)) - (1<<ofst
);
4675 assert( n
>1 || mask
==(1<<ofst
) );
4676 sqlite3_mutex_enter(pShmNode
->mutex
);
4677 if( flags
& SQLITE_SHM_UNLOCK
){
4678 u16 allMask
= 0; /* Mask of locks held by siblings */
4680 /* See if any siblings hold this same lock */
4681 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4682 if( pX
==p
) continue;
4683 assert( (pX
->exclMask
& (p
->exclMask
|p
->sharedMask
))==0 );
4684 allMask
|= pX
->sharedMask
;
4687 /* Unlock the system-level locks */
4688 if( (mask
& allMask
)==0 ){
4689 rc
= unixShmSystemLock(pDbFd
, F_UNLCK
, ofst
+UNIX_SHM_BASE
, n
);
4694 /* Undo the local locks */
4695 if( rc
==SQLITE_OK
){
4696 p
->exclMask
&= ~mask
;
4697 p
->sharedMask
&= ~mask
;
4699 }else if( flags
& SQLITE_SHM_SHARED
){
4700 u16 allShared
= 0; /* Union of locks held by connections other than "p" */
4702 /* Find out which shared locks are already held by sibling connections.
4703 ** If any sibling already holds an exclusive lock, go ahead and return
4706 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4707 if( (pX
->exclMask
& mask
)!=0 ){
4711 allShared
|= pX
->sharedMask
;
4714 /* Get shared locks at the system level, if necessary */
4715 if( rc
==SQLITE_OK
){
4716 if( (allShared
& mask
)==0 ){
4717 rc
= unixShmSystemLock(pDbFd
, F_RDLCK
, ofst
+UNIX_SHM_BASE
, n
);
4723 /* Get the local shared locks */
4724 if( rc
==SQLITE_OK
){
4725 p
->sharedMask
|= mask
;
4728 /* Make sure no sibling connections hold locks that will block this
4729 ** lock. If any do, return SQLITE_BUSY right away.
4731 for(pX
=pShmNode
->pFirst
; pX
; pX
=pX
->pNext
){
4732 if( (pX
->exclMask
& mask
)!=0 || (pX
->sharedMask
& mask
)!=0 ){
4738 /* Get the exclusive locks at the system level. Then if successful
4739 ** also mark the local connection as being locked.
4741 if( rc
==SQLITE_OK
){
4742 rc
= unixShmSystemLock(pDbFd
, F_WRLCK
, ofst
+UNIX_SHM_BASE
, n
);
4743 if( rc
==SQLITE_OK
){
4744 assert( (p
->sharedMask
& mask
)==0 );
4745 p
->exclMask
|= mask
;
4749 sqlite3_mutex_leave(pShmNode
->mutex
);
4750 OSTRACE(("SHM-LOCK shmid-%d, pid-%d got %03x,%03x\n",
4751 p
->id
, osGetpid(0), p
->sharedMask
, p
->exclMask
));
4756 ** Implement a memory barrier or memory fence on shared memory.
4758 ** All loads and stores begun before the barrier must complete before
4759 ** any load or store begun after the barrier.
4761 static void unixShmBarrier(
4762 sqlite3_file
*fd
/* Database file holding the shared memory */
4764 UNUSED_PARAMETER(fd
);
4765 sqlite3MemoryBarrier(); /* compiler-defined memory barrier */
4766 unixEnterMutex(); /* Also mutex, for redundancy */
4771 ** Close a connection to shared-memory. Delete the underlying
4772 ** storage if deleteFlag is true.
4774 ** If there is no shared memory associated with the connection then this
4775 ** routine is a harmless no-op.
4777 static int unixShmUnmap(
4778 sqlite3_file
*fd
, /* The underlying database file */
4779 int deleteFlag
/* Delete shared-memory if true */
4781 unixShm
*p
; /* The connection to be closed */
4782 unixShmNode
*pShmNode
; /* The underlying shared-memory file */
4783 unixShm
**pp
; /* For looping over sibling connections */
4784 unixFile
*pDbFd
; /* The underlying database file */
4786 pDbFd
= (unixFile
*)fd
;
4788 if( p
==0 ) return SQLITE_OK
;
4789 pShmNode
= p
->pShmNode
;
4791 assert( pShmNode
==pDbFd
->pInode
->pShmNode
);
4792 assert( pShmNode
->pInode
==pDbFd
->pInode
);
4794 /* Remove connection p from the set of connections associated
4796 sqlite3_mutex_enter(pShmNode
->mutex
);
4797 for(pp
=&pShmNode
->pFirst
; (*pp
)!=p
; pp
= &(*pp
)->pNext
){}
4800 /* Free the connection p */
4803 sqlite3_mutex_leave(pShmNode
->mutex
);
4805 /* If pShmNode->nRef has reached 0, then close the underlying
4806 ** shared-memory file, too */
4808 assert( pShmNode
->nRef
>0 );
4810 if( pShmNode
->nRef
==0 ){
4811 if( deleteFlag
&& pShmNode
->h
>=0 ){
4812 osUnlink(pShmNode
->zFilename
);
4814 unixShmPurge(pDbFd
);
4823 # define unixShmMap 0
4824 # define unixShmLock 0
4825 # define unixShmBarrier 0
4826 # define unixShmUnmap 0
4827 #endif /* #ifndef SQLITE_OMIT_WAL */
4829 #if SQLITE_MAX_MMAP_SIZE>0
4831 ** If it is currently memory mapped, unmap file pFd.
4833 static void unixUnmapfile(unixFile
*pFd
){
4834 assert( pFd
->nFetchOut
==0 );
4835 if( pFd
->pMapRegion
){
4836 osMunmap(pFd
->pMapRegion
, pFd
->mmapSizeActual
);
4837 pFd
->pMapRegion
= 0;
4839 pFd
->mmapSizeActual
= 0;
4844 ** Attempt to set the size of the memory mapping maintained by file
4845 ** descriptor pFd to nNew bytes. Any existing mapping is discarded.
4847 ** If successful, this function sets the following variables:
4849 ** unixFile.pMapRegion
4850 ** unixFile.mmapSize
4851 ** unixFile.mmapSizeActual
4853 ** If unsuccessful, an error message is logged via sqlite3_log() and
4854 ** the three variables above are zeroed. In this case SQLite should
4855 ** continue accessing the database using the xRead() and xWrite()
4858 static void unixRemapfile(
4859 unixFile
*pFd
, /* File descriptor object */
4860 i64 nNew
/* Required mapping size */
4862 const char *zErr
= "mmap";
4863 int h
= pFd
->h
; /* File descriptor open on db file */
4864 u8
*pOrig
= (u8
*)pFd
->pMapRegion
; /* Pointer to current file mapping */
4865 i64 nOrig
= pFd
->mmapSizeActual
; /* Size of pOrig region in bytes */
4866 u8
*pNew
= 0; /* Location of new mapping */
4867 int flags
= PROT_READ
; /* Flags to pass to mmap() */
4869 assert( pFd
->nFetchOut
==0 );
4870 assert( nNew
>pFd
->mmapSize
);
4871 assert( nNew
<=pFd
->mmapSizeMax
);
4873 assert( pFd
->mmapSizeActual
>=pFd
->mmapSize
);
4874 assert( MAP_FAILED
!=0 );
4876 #ifdef SQLITE_MMAP_READWRITE
4877 if( (pFd
->ctrlFlags
& UNIXFILE_RDONLY
)==0 ) flags
|= PROT_WRITE
;
4882 i64 nReuse
= pFd
->mmapSize
;
4884 const int szSyspage
= osGetpagesize();
4885 i64 nReuse
= (pFd
->mmapSize
& ~(szSyspage
-1));
4887 u8
*pReq
= &pOrig
[nReuse
];
4889 /* Unmap any pages of the existing mapping that cannot be reused. */
4890 if( nReuse
!=nOrig
){
4891 osMunmap(pReq
, nOrig
-nReuse
);
4895 pNew
= osMremap(pOrig
, nReuse
, nNew
, MREMAP_MAYMOVE
);
4898 pNew
= osMmap(pReq
, nNew
-nReuse
, flags
, MAP_SHARED
, h
, nReuse
);
4899 if( pNew
!=MAP_FAILED
){
4901 osMunmap(pNew
, nNew
- nReuse
);
4909 /* The attempt to extend the existing mapping failed. Free it. */
4910 if( pNew
==MAP_FAILED
|| pNew
==0 ){
4911 osMunmap(pOrig
, nReuse
);
4915 /* If pNew is still NULL, try to create an entirely new mapping. */
4917 pNew
= osMmap(0, nNew
, flags
, MAP_SHARED
, h
, 0);
4920 if( pNew
==MAP_FAILED
){
4923 unixLogError(SQLITE_OK
, zErr
, pFd
->zPath
);
4925 /* If the mmap() above failed, assume that all subsequent mmap() calls
4926 ** will probably fail too. Fall back to using xRead/xWrite exclusively
4928 pFd
->mmapSizeMax
= 0;
4930 pFd
->pMapRegion
= (void *)pNew
;
4931 pFd
->mmapSize
= pFd
->mmapSizeActual
= nNew
;
4935 ** Memory map or remap the file opened by file-descriptor pFd (if the file
4936 ** is already mapped, the existing mapping is replaced by the new). Or, if
4937 ** there already exists a mapping for this file, and there are still
4938 ** outstanding xFetch() references to it, this function is a no-op.
4940 ** If parameter nByte is non-negative, then it is the requested size of
4941 ** the mapping to create. Otherwise, if nByte is less than zero, then the
4942 ** requested size is the size of the file on disk. The actual size of the
4943 ** created mapping is either the requested size or the value configured
4944 ** using SQLITE_FCNTL_MMAP_LIMIT, whichever is smaller.
4946 ** SQLITE_OK is returned if no error occurs (even if the mapping is not
4947 ** recreated as a result of outstanding references) or an SQLite error
4950 static int unixMapfile(unixFile
*pFd
, i64 nMap
){
4951 assert( nMap
>=0 || pFd
->nFetchOut
==0 );
4952 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
4953 if( pFd
->nFetchOut
>0 ) return SQLITE_OK
;
4956 struct stat statbuf
; /* Low-level file information */
4957 if( osFstat(pFd
->h
, &statbuf
) ){
4958 return SQLITE_IOERR_FSTAT
;
4960 nMap
= statbuf
.st_size
;
4962 if( nMap
>pFd
->mmapSizeMax
){
4963 nMap
= pFd
->mmapSizeMax
;
4966 assert( nMap
>0 || (pFd
->mmapSize
==0 && pFd
->pMapRegion
==0) );
4967 if( nMap
!=pFd
->mmapSize
){
4968 unixRemapfile(pFd
, nMap
);
4973 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
4976 ** If possible, return a pointer to a mapping of file fd starting at offset
4977 ** iOff. The mapping must be valid for at least nAmt bytes.
4979 ** If such a pointer can be obtained, store it in *pp and return SQLITE_OK.
4980 ** Or, if one cannot but no error occurs, set *pp to 0 and return SQLITE_OK.
4981 ** Finally, if an error does occur, return an SQLite error code. The final
4982 ** value of *pp is undefined in this case.
4984 ** If this function does return a pointer, the caller must eventually
4985 ** release the reference by calling unixUnfetch().
4987 static int unixFetch(sqlite3_file
*fd
, i64 iOff
, int nAmt
, void **pp
){
4988 #if SQLITE_MAX_MMAP_SIZE>0
4989 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
4993 #if SQLITE_MAX_MMAP_SIZE>0
4994 if( pFd
->mmapSizeMax
>0 ){
4995 if( pFd
->pMapRegion
==0 ){
4996 int rc
= unixMapfile(pFd
, -1);
4997 if( rc
!=SQLITE_OK
) return rc
;
4999 if( pFd
->mmapSize
>= iOff
+nAmt
){
5000 *pp
= &((u8
*)pFd
->pMapRegion
)[iOff
];
5009 ** If the third argument is non-NULL, then this function releases a
5010 ** reference obtained by an earlier call to unixFetch(). The second
5011 ** argument passed to this function must be the same as the corresponding
5012 ** argument that was passed to the unixFetch() invocation.
5014 ** Or, if the third argument is NULL, then this function is being called
5015 ** to inform the VFS layer that, according to POSIX, any existing mapping
5016 ** may now be invalid and should be unmapped.
5018 static int unixUnfetch(sqlite3_file
*fd
, i64 iOff
, void *p
){
5019 #if SQLITE_MAX_MMAP_SIZE>0
5020 unixFile
*pFd
= (unixFile
*)fd
; /* The underlying database file */
5021 UNUSED_PARAMETER(iOff
);
5023 /* If p==0 (unmap the entire file) then there must be no outstanding
5024 ** xFetch references. Or, if p!=0 (meaning it is an xFetch reference),
5025 ** then there must be at least one outstanding. */
5026 assert( (p
==0)==(pFd
->nFetchOut
==0) );
5028 /* If p!=0, it must match the iOff value. */
5029 assert( p
==0 || p
==&((u8
*)pFd
->pMapRegion
)[iOff
] );
5037 assert( pFd
->nFetchOut
>=0 );
5039 UNUSED_PARAMETER(fd
);
5040 UNUSED_PARAMETER(p
);
5041 UNUSED_PARAMETER(iOff
);
5047 ** Here ends the implementation of all sqlite3_file methods.
5049 ********************** End sqlite3_file Methods *******************************
5050 ******************************************************************************/
5053 ** This division contains definitions of sqlite3_io_methods objects that
5054 ** implement various file locking strategies. It also contains definitions
5055 ** of "finder" functions. A finder-function is used to locate the appropriate
5056 ** sqlite3_io_methods object for a particular database file. The pAppData
5057 ** field of the sqlite3_vfs VFS objects are initialized to be pointers to
5058 ** the correct finder-function for that VFS.
5060 ** Most finder functions return a pointer to a fixed sqlite3_io_methods
5061 ** object. The only interesting finder-function is autolockIoFinder, which
5062 ** looks at the filesystem type and tries to guess the best locking
5063 ** strategy from that.
5065 ** For finder-function F, two objects are created:
5067 ** (1) The real finder-function named "FImpt()".
5069 ** (2) A constant pointer to this function named just "F".
5072 ** A pointer to the F pointer is used as the pAppData value for VFS
5073 ** objects. We have to do this instead of letting pAppData point
5074 ** directly at the finder-function since C90 rules prevent a void*
5075 ** from be cast into a function pointer.
5078 ** Each instance of this macro generates two objects:
5080 ** * A constant sqlite3_io_methods object call METHOD that has locking
5081 ** methods CLOSE, LOCK, UNLOCK, CKRESLOCK.
5083 ** * An I/O method finder function called FINDER that returns a pointer
5084 ** to the METHOD object in the previous bullet.
5086 #define IOMETHODS(FINDER,METHOD,VERSION,CLOSE,LOCK,UNLOCK,CKLOCK,SHMMAP) \
5087 static const sqlite3_io_methods METHOD = { \
5088 VERSION, /* iVersion */ \
5089 CLOSE, /* xClose */ \
5090 unixRead, /* xRead */ \
5091 unixWrite, /* xWrite */ \
5092 unixTruncate, /* xTruncate */ \
5093 unixSync, /* xSync */ \
5094 unixFileSize, /* xFileSize */ \
5096 UNLOCK, /* xUnlock */ \
5097 CKLOCK, /* xCheckReservedLock */ \
5098 unixFileControl, /* xFileControl */ \
5099 unixSectorSize, /* xSectorSize */ \
5100 unixDeviceCharacteristics, /* xDeviceCapabilities */ \
5101 SHMMAP, /* xShmMap */ \
5102 unixShmLock, /* xShmLock */ \
5103 unixShmBarrier, /* xShmBarrier */ \
5104 unixShmUnmap, /* xShmUnmap */ \
5105 unixFetch, /* xFetch */ \
5106 unixUnfetch, /* xUnfetch */ \
5108 static const sqlite3_io_methods *FINDER##Impl(const char *z, unixFile *p){ \
5109 UNUSED_PARAMETER(z); UNUSED_PARAMETER(p); \
5112 static const sqlite3_io_methods *(*const FINDER)(const char*,unixFile *p) \
5116 ** Here are all of the sqlite3_io_methods objects for each of the
5117 ** locking strategies. Functions that return pointers to these methods
5118 ** are also created.
5121 posixIoFinder
, /* Finder function name */
5122 posixIoMethods
, /* sqlite3_io_methods object name */
5123 3, /* shared memory and mmap are enabled */
5124 unixClose
, /* xClose method */
5125 unixLock
, /* xLock method */
5126 unixUnlock
, /* xUnlock method */
5127 unixCheckReservedLock
, /* xCheckReservedLock method */
5128 unixShmMap
/* xShmMap method */
5131 nolockIoFinder
, /* Finder function name */
5132 nolockIoMethods
, /* sqlite3_io_methods object name */
5133 3, /* shared memory is disabled */
5134 nolockClose
, /* xClose method */
5135 nolockLock
, /* xLock method */
5136 nolockUnlock
, /* xUnlock method */
5137 nolockCheckReservedLock
, /* xCheckReservedLock method */
5138 0 /* xShmMap method */
5141 dotlockIoFinder
, /* Finder function name */
5142 dotlockIoMethods
, /* sqlite3_io_methods object name */
5143 1, /* shared memory is disabled */
5144 dotlockClose
, /* xClose method */
5145 dotlockLock
, /* xLock method */
5146 dotlockUnlock
, /* xUnlock method */
5147 dotlockCheckReservedLock
, /* xCheckReservedLock method */
5148 0 /* xShmMap method */
5151 #if SQLITE_ENABLE_LOCKING_STYLE
5153 flockIoFinder
, /* Finder function name */
5154 flockIoMethods
, /* sqlite3_io_methods object name */
5155 1, /* shared memory is disabled */
5156 flockClose
, /* xClose method */
5157 flockLock
, /* xLock method */
5158 flockUnlock
, /* xUnlock method */
5159 flockCheckReservedLock
, /* xCheckReservedLock method */
5160 0 /* xShmMap method */
5166 semIoFinder
, /* Finder function name */
5167 semIoMethods
, /* sqlite3_io_methods object name */
5168 1, /* shared memory is disabled */
5169 semXClose
, /* xClose method */
5170 semXLock
, /* xLock method */
5171 semXUnlock
, /* xUnlock method */
5172 semXCheckReservedLock
, /* xCheckReservedLock method */
5173 0 /* xShmMap method */
5177 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5179 afpIoFinder
, /* Finder function name */
5180 afpIoMethods
, /* sqlite3_io_methods object name */
5181 1, /* shared memory is disabled */
5182 afpClose
, /* xClose method */
5183 afpLock
, /* xLock method */
5184 afpUnlock
, /* xUnlock method */
5185 afpCheckReservedLock
, /* xCheckReservedLock method */
5186 0 /* xShmMap method */
5191 ** The proxy locking method is a "super-method" in the sense that it
5192 ** opens secondary file descriptors for the conch and lock files and
5193 ** it uses proxy, dot-file, AFP, and flock() locking methods on those
5194 ** secondary files. For this reason, the division that implements
5195 ** proxy locking is located much further down in the file. But we need
5196 ** to go ahead and define the sqlite3_io_methods and finder function
5197 ** for proxy locking here. So we forward declare the I/O methods.
5199 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5200 static int proxyClose(sqlite3_file
*);
5201 static int proxyLock(sqlite3_file
*, int);
5202 static int proxyUnlock(sqlite3_file
*, int);
5203 static int proxyCheckReservedLock(sqlite3_file
*, int*);
5205 proxyIoFinder
, /* Finder function name */
5206 proxyIoMethods
, /* sqlite3_io_methods object name */
5207 1, /* shared memory is disabled */
5208 proxyClose
, /* xClose method */
5209 proxyLock
, /* xLock method */
5210 proxyUnlock
, /* xUnlock method */
5211 proxyCheckReservedLock
, /* xCheckReservedLock method */
5212 0 /* xShmMap method */
5216 /* nfs lockd on OSX 10.3+ doesn't clear write locks when a read lock is set */
5217 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5219 nfsIoFinder
, /* Finder function name */
5220 nfsIoMethods
, /* sqlite3_io_methods object name */
5221 1, /* shared memory is disabled */
5222 unixClose
, /* xClose method */
5223 unixLock
, /* xLock method */
5224 nfsUnlock
, /* xUnlock method */
5225 unixCheckReservedLock
, /* xCheckReservedLock method */
5226 0 /* xShmMap method */
5230 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5232 ** This "finder" function attempts to determine the best locking strategy
5233 ** for the database file "filePath". It then returns the sqlite3_io_methods
5234 ** object that implements that strategy.
5236 ** This is for MacOSX only.
5238 static const sqlite3_io_methods
*autolockIoFinderImpl(
5239 const char *filePath
, /* name of the database file */
5240 unixFile
*pNew
/* open file object for the database file */
5242 static const struct Mapping
{
5243 const char *zFilesystem
; /* Filesystem type name */
5244 const sqlite3_io_methods
*pMethods
; /* Appropriate locking method */
5246 { "hfs", &posixIoMethods
},
5247 { "ufs", &posixIoMethods
},
5248 { "afpfs", &afpIoMethods
},
5249 { "smbfs", &afpIoMethods
},
5250 { "webdav", &nolockIoMethods
},
5254 struct statfs fsInfo
;
5255 struct flock lockInfo
;
5258 /* If filePath==NULL that means we are dealing with a transient file
5259 ** that does not need to be locked. */
5260 return &nolockIoMethods
;
5262 if( statfs(filePath
, &fsInfo
) != -1 ){
5263 if( fsInfo
.f_flags
& MNT_RDONLY
){
5264 return &nolockIoMethods
;
5266 for(i
=0; aMap
[i
].zFilesystem
; i
++){
5267 if( strcmp(fsInfo
.f_fstypename
, aMap
[i
].zFilesystem
)==0 ){
5268 return aMap
[i
].pMethods
;
5273 /* Default case. Handles, amongst others, "nfs".
5274 ** Test byte-range lock using fcntl(). If the call succeeds,
5275 ** assume that the file-system supports POSIX style locks.
5278 lockInfo
.l_start
= 0;
5279 lockInfo
.l_whence
= SEEK_SET
;
5280 lockInfo
.l_type
= F_RDLCK
;
5281 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5282 if( strcmp(fsInfo
.f_fstypename
, "nfs")==0 ){
5283 return &nfsIoMethods
;
5285 return &posixIoMethods
;
5288 return &dotlockIoMethods
;
5291 static const sqlite3_io_methods
5292 *(*const autolockIoFinder
)(const char*,unixFile
*) = autolockIoFinderImpl
;
5294 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
5298 ** This "finder" function for VxWorks checks to see if posix advisory
5299 ** locking works. If it does, then that is what is used. If it does not
5300 ** work, then fallback to named semaphore locking.
5302 static const sqlite3_io_methods
*vxworksIoFinderImpl(
5303 const char *filePath
, /* name of the database file */
5304 unixFile
*pNew
/* the open file object */
5306 struct flock lockInfo
;
5309 /* If filePath==NULL that means we are dealing with a transient file
5310 ** that does not need to be locked. */
5311 return &nolockIoMethods
;
5314 /* Test if fcntl() is supported and use POSIX style locks.
5315 ** Otherwise fall back to the named semaphore method.
5318 lockInfo
.l_start
= 0;
5319 lockInfo
.l_whence
= SEEK_SET
;
5320 lockInfo
.l_type
= F_RDLCK
;
5321 if( osFcntl(pNew
->h
, F_GETLK
, &lockInfo
)!=-1 ) {
5322 return &posixIoMethods
;
5324 return &semIoMethods
;
5327 static const sqlite3_io_methods
5328 *(*const vxworksIoFinder
)(const char*,unixFile
*) = vxworksIoFinderImpl
;
5330 #endif /* OS_VXWORKS */
5333 ** An abstract type for a pointer to an IO method finder function:
5335 typedef const sqlite3_io_methods
*(*finder_type
)(const char*,unixFile
*);
5338 /****************************************************************************
5339 **************************** sqlite3_vfs methods ****************************
5341 ** This division contains the implementation of methods on the
5342 ** sqlite3_vfs object.
5346 ** Initialize the contents of the unixFile structure pointed to by pId.
5348 static int fillInUnixFile(
5349 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
5350 int h
, /* Open file descriptor of file being opened */
5351 sqlite3_file
*pId
, /* Write to the unixFile structure here */
5352 const char *zFilename
, /* Name of the file being opened */
5353 int ctrlFlags
/* Zero or more UNIXFILE_* values */
5355 const sqlite3_io_methods
*pLockingStyle
;
5356 unixFile
*pNew
= (unixFile
*)pId
;
5359 assert( pNew
->pInode
==NULL
);
5361 /* No locking occurs in temporary files */
5362 assert( zFilename
!=0 || (ctrlFlags
& UNIXFILE_NOLOCK
)!=0 );
5364 OSTRACE(("OPEN %-3d %s\n", h
, zFilename
));
5367 pNew
->zPath
= zFilename
;
5368 pNew
->ctrlFlags
= (u8
)ctrlFlags
;
5369 #if SQLITE_MAX_MMAP_SIZE>0
5370 pNew
->mmapSizeMax
= sqlite3GlobalConfig
.szMmap
;
5372 if( sqlite3_uri_boolean(((ctrlFlags
& UNIXFILE_URI
) ? zFilename
: 0),
5373 "psow", SQLITE_POWERSAFE_OVERWRITE
) ){
5374 pNew
->ctrlFlags
|= UNIXFILE_PSOW
;
5376 if( strcmp(pVfs
->zName
,"unix-excl")==0 ){
5377 pNew
->ctrlFlags
|= UNIXFILE_EXCL
;
5381 pNew
->pId
= vxworksFindFileId(zFilename
);
5383 ctrlFlags
|= UNIXFILE_NOLOCK
;
5384 rc
= SQLITE_NOMEM_BKPT
;
5388 if( ctrlFlags
& UNIXFILE_NOLOCK
){
5389 pLockingStyle
= &nolockIoMethods
;
5391 pLockingStyle
= (**(finder_type
*)pVfs
->pAppData
)(zFilename
, pNew
);
5392 #if SQLITE_ENABLE_LOCKING_STYLE
5393 /* Cache zFilename in the locking context (AFP and dotlock override) for
5394 ** proxyLock activation is possible (remote proxy is based on db name)
5395 ** zFilename remains valid until file is closed, to support */
5396 pNew
->lockingContext
= (void*)zFilename
;
5400 if( pLockingStyle
== &posixIoMethods
5401 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
5402 || pLockingStyle
== &nfsIoMethods
5406 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5407 if( rc
!=SQLITE_OK
){
5408 /* If an error occurred in findInodeInfo(), close the file descriptor
5409 ** immediately, before releasing the mutex. findInodeInfo() may fail
5410 ** in two scenarios:
5412 ** (a) A call to fstat() failed.
5413 ** (b) A malloc failed.
5415 ** Scenario (b) may only occur if the process is holding no other
5416 ** file descriptors open on the same file. If there were other file
5417 ** descriptors on this file, then no malloc would be required by
5418 ** findInodeInfo(). If this is the case, it is quite safe to close
5419 ** handle h - as it is guaranteed that no posix locks will be released
5422 ** If scenario (a) caused the error then things are not so safe. The
5423 ** implicit assumption here is that if fstat() fails, things are in
5424 ** such bad shape that dropping a lock or two doesn't matter much.
5426 robust_close(pNew
, h
, __LINE__
);
5432 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5433 else if( pLockingStyle
== &afpIoMethods
){
5434 /* AFP locking uses the file path so it needs to be included in
5435 ** the afpLockingContext.
5437 afpLockingContext
*pCtx
;
5438 pNew
->lockingContext
= pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
5440 rc
= SQLITE_NOMEM_BKPT
;
5442 /* NB: zFilename exists and remains valid until the file is closed
5443 ** according to requirement F11141. So we do not need to make a
5444 ** copy of the filename. */
5445 pCtx
->dbPath
= zFilename
;
5449 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5450 if( rc
!=SQLITE_OK
){
5451 sqlite3_free(pNew
->lockingContext
);
5452 robust_close(pNew
, h
, __LINE__
);
5460 else if( pLockingStyle
== &dotlockIoMethods
){
5461 /* Dotfile locking uses the file path so it needs to be included in
5462 ** the dotlockLockingContext
5466 assert( zFilename
!=0 );
5467 nFilename
= (int)strlen(zFilename
) + 6;
5468 zLockFile
= (char *)sqlite3_malloc64(nFilename
);
5470 rc
= SQLITE_NOMEM_BKPT
;
5472 sqlite3_snprintf(nFilename
, zLockFile
, "%s" DOTLOCK_SUFFIX
, zFilename
);
5474 pNew
->lockingContext
= zLockFile
;
5478 else if( pLockingStyle
== &semIoMethods
){
5479 /* Named semaphore locking uses the file path so it needs to be
5480 ** included in the semLockingContext
5483 rc
= findInodeInfo(pNew
, &pNew
->pInode
);
5484 if( (rc
==SQLITE_OK
) && (pNew
->pInode
->pSem
==NULL
) ){
5485 char *zSemName
= pNew
->pInode
->aSemName
;
5487 sqlite3_snprintf(MAX_PATHNAME
, zSemName
, "/%s.sem",
5488 pNew
->pId
->zCanonicalName
);
5489 for( n
=1; zSemName
[n
]; n
++ )
5490 if( zSemName
[n
]=='/' ) zSemName
[n
] = '_';
5491 pNew
->pInode
->pSem
= sem_open(zSemName
, O_CREAT
, 0666, 1);
5492 if( pNew
->pInode
->pSem
== SEM_FAILED
){
5493 rc
= SQLITE_NOMEM_BKPT
;
5494 pNew
->pInode
->aSemName
[0] = '\0';
5501 storeLastErrno(pNew
, 0);
5503 if( rc
!=SQLITE_OK
){
5504 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5506 osUnlink(zFilename
);
5507 pNew
->ctrlFlags
|= UNIXFILE_DELETE
;
5510 if( rc
!=SQLITE_OK
){
5511 if( h
>=0 ) robust_close(pNew
, h
, __LINE__
);
5513 pNew
->pMethod
= pLockingStyle
;
5521 ** Return the name of a directory in which to put temporary files.
5522 ** If no suitable temporary file directory can be found, return NULL.
5524 static const char *unixTempFileDir(void){
5525 static const char *azDirs
[] = {
5535 const char *zDir
= sqlite3_temp_directory
;
5537 if( !azDirs
[0] ) azDirs
[0] = getenv("SQLITE_TMPDIR");
5538 if( !azDirs
[1] ) azDirs
[1] = getenv("TMPDIR");
5541 && osStat(zDir
, &buf
)==0
5542 && S_ISDIR(buf
.st_mode
)
5543 && osAccess(zDir
, 03)==0
5547 if( i
>=sizeof(azDirs
)/sizeof(azDirs
[0]) ) break;
5554 ** Create a temporary file name in zBuf. zBuf must be allocated
5555 ** by the calling process and must be big enough to hold at least
5556 ** pVfs->mxPathname bytes.
5558 static int unixGetTempname(int nBuf
, char *zBuf
){
5562 /* It's odd to simulate an io-error here, but really this is just
5563 ** using the io-error infrastructure to test that SQLite handles this
5564 ** function failing.
5567 SimulateIOError( return SQLITE_IOERR
);
5569 zDir
= unixTempFileDir();
5570 if( zDir
==0 ) return SQLITE_IOERR_GETTEMPPATH
;
5573 sqlite3_randomness(sizeof(r
), &r
);
5576 sqlite3_snprintf(nBuf
, zBuf
, "%s/"SQLITE_TEMP_FILE_PREFIX
"%llx%c",
5578 if( zBuf
[nBuf
-2]!=0 || (iLimit
++)>10 ) return SQLITE_ERROR
;
5579 }while( osAccess(zBuf
,0)==0 );
5583 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
5585 ** Routine to transform a unixFile into a proxy-locking unixFile.
5586 ** Implementation in the proxy-lock division, but used by unixOpen()
5587 ** if SQLITE_PREFER_PROXY_LOCKING is defined.
5589 static int proxyTransformUnixFile(unixFile
*, const char*);
5593 ** Search for an unused file descriptor that was opened on the database
5594 ** file (not a journal or master-journal file) identified by pathname
5595 ** zPath with SQLITE_OPEN_XXX flags matching those passed as the second
5596 ** argument to this function.
5598 ** Such a file descriptor may exist if a database connection was closed
5599 ** but the associated file descriptor could not be closed because some
5600 ** other file descriptor open on the same file is holding a file-lock.
5601 ** Refer to comments in the unixClose() function and the lengthy comment
5602 ** describing "Posix Advisory Locking" at the start of this file for
5603 ** further details. Also, ticket #4018.
5605 ** If a suitable file descriptor is found, then it is returned. If no
5606 ** such file descriptor is located, -1 is returned.
5608 static UnixUnusedFd
*findReusableFd(const char *zPath
, int flags
){
5609 UnixUnusedFd
*pUnused
= 0;
5611 /* Do not search for an unused file descriptor on vxworks. Not because
5612 ** vxworks would not benefit from the change (it might, we're not sure),
5613 ** but because no way to test it is currently available. It is better
5614 ** not to risk breaking vxworks support for the sake of such an obscure
5617 struct stat sStat
; /* Results of stat() call */
5621 /* A stat() call may fail for various reasons. If this happens, it is
5622 ** almost certain that an open() call on the same path will also fail.
5623 ** For this reason, if an error occurs in the stat() call here, it is
5624 ** ignored and -1 is returned. The caller will try to open a new file
5625 ** descriptor on the same path, fail, and return an error to SQLite.
5627 ** Even if a subsequent open() call does succeed, the consequences of
5628 ** not searching for a reusable file descriptor are not dire. */
5629 if( nUnusedFd
>0 && 0==osStat(zPath
, &sStat
) ){
5630 unixInodeInfo
*pInode
;
5633 while( pInode
&& (pInode
->fileId
.dev
!=sStat
.st_dev
5634 || pInode
->fileId
.ino
!=(u64
)sStat
.st_ino
) ){
5635 pInode
= pInode
->pNext
;
5639 for(pp
=&pInode
->pUnused
; *pp
&& (*pp
)->flags
!=flags
; pp
=&((*pp
)->pNext
));
5643 *pp
= pUnused
->pNext
;
5648 #endif /* if !OS_VXWORKS */
5653 ** Find the mode, uid and gid of file zFile.
5655 static int getFileMode(
5656 const char *zFile
, /* File name */
5657 mode_t
*pMode
, /* OUT: Permissions of zFile */
5658 uid_t
*pUid
, /* OUT: uid of zFile. */
5659 gid_t
*pGid
/* OUT: gid of zFile. */
5661 struct stat sStat
; /* Output of stat() on database file */
5663 if( 0==osStat(zFile
, &sStat
) ){
5664 *pMode
= sStat
.st_mode
& 0777;
5665 *pUid
= sStat
.st_uid
;
5666 *pGid
= sStat
.st_gid
;
5668 rc
= SQLITE_IOERR_FSTAT
;
5674 ** This function is called by unixOpen() to determine the unix permissions
5675 ** to create new files with. If no error occurs, then SQLITE_OK is returned
5676 ** and a value suitable for passing as the third argument to open(2) is
5677 ** written to *pMode. If an IO error occurs, an SQLite error code is
5678 ** returned and the value of *pMode is not modified.
5680 ** In most cases, this routine sets *pMode to 0, which will become
5681 ** an indication to robust_open() to create the file using
5682 ** SQLITE_DEFAULT_FILE_PERMISSIONS adjusted by the umask.
5683 ** But if the file being opened is a WAL or regular journal file, then
5684 ** this function queries the file-system for the permissions on the
5685 ** corresponding database file and sets *pMode to this value. Whenever
5686 ** possible, WAL and journal files are created using the same permissions
5687 ** as the associated database file.
5689 ** If the SQLITE_ENABLE_8_3_NAMES option is enabled, then the
5690 ** original filename is unavailable. But 8_3_NAMES is only used for
5691 ** FAT filesystems and permissions do not matter there, so just use
5692 ** the default permissions.
5694 static int findCreateFileMode(
5695 const char *zPath
, /* Path of file (possibly) being created */
5696 int flags
, /* Flags passed as 4th argument to xOpen() */
5697 mode_t
*pMode
, /* OUT: Permissions to open file with */
5698 uid_t
*pUid
, /* OUT: uid to set on the file */
5699 gid_t
*pGid
/* OUT: gid to set on the file */
5701 int rc
= SQLITE_OK
; /* Return Code */
5705 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5706 char zDb
[MAX_PATHNAME
+1]; /* Database file path */
5707 int nDb
; /* Number of valid bytes in zDb */
5709 /* zPath is a path to a WAL or journal file. The following block derives
5710 ** the path to the associated database file from zPath. This block handles
5711 ** the following naming conventions:
5713 ** "<path to db>-journal"
5714 ** "<path to db>-wal"
5715 ** "<path to db>-journalNN"
5716 ** "<path to db>-walNN"
5718 ** where NN is a decimal number. The NN naming schemes are
5719 ** used by the test_multiplex.c module.
5721 nDb
= sqlite3Strlen30(zPath
) - 1;
5722 while( zPath
[nDb
]!='-' ){
5723 /* In normal operation, the journal file name will always contain
5724 ** a '-' character. However in 8+3 filename mode, or if a corrupt
5725 ** rollback journal specifies a master journal with a goofy name, then
5726 ** the '-' might be missing. */
5727 if( nDb
==0 || zPath
[nDb
]=='.' ) return SQLITE_OK
;
5730 memcpy(zDb
, zPath
, nDb
);
5733 rc
= getFileMode(zDb
, pMode
, pUid
, pGid
);
5734 }else if( flags
& SQLITE_OPEN_DELETEONCLOSE
){
5736 }else if( flags
& SQLITE_OPEN_URI
){
5737 /* If this is a main database file and the file was opened using a URI
5738 ** filename, check for the "modeof" parameter. If present, interpret
5739 ** its value as a filename and try to copy the mode, uid and gid from
5741 const char *z
= sqlite3_uri_parameter(zPath
, "modeof");
5743 rc
= getFileMode(z
, pMode
, pUid
, pGid
);
5750 ** Open the file zPath.
5752 ** Previously, the SQLite OS layer used three functions in place of this
5755 ** sqlite3OsOpenReadWrite();
5756 ** sqlite3OsOpenReadOnly();
5757 ** sqlite3OsOpenExclusive();
5759 ** These calls correspond to the following combinations of flags:
5761 ** ReadWrite() -> (READWRITE | CREATE)
5762 ** ReadOnly() -> (READONLY)
5763 ** OpenExclusive() -> (READWRITE | CREATE | EXCLUSIVE)
5765 ** The old OpenExclusive() accepted a boolean argument - "delFlag". If
5766 ** true, the file was configured to be automatically deleted when the
5767 ** file handle closed. To achieve the same effect using this new
5768 ** interface, add the DELETEONCLOSE flag to those specified above for
5771 static int unixOpen(
5772 sqlite3_vfs
*pVfs
, /* The VFS for which this is the xOpen method */
5773 const char *zPath
, /* Pathname of file to be opened */
5774 sqlite3_file
*pFile
, /* The file descriptor to be filled in */
5775 int flags
, /* Input flags to control the opening */
5776 int *pOutFlags
/* Output flags returned to SQLite core */
5778 unixFile
*p
= (unixFile
*)pFile
;
5779 int fd
= -1; /* File descriptor returned by open() */
5780 int openFlags
= 0; /* Flags to pass to open() */
5781 int eType
= flags
&0xFFFFFF00; /* Type of file to open */
5782 int noLock
; /* True to omit locking primitives */
5783 int rc
= SQLITE_OK
; /* Function Return Code */
5784 int ctrlFlags
= 0; /* UNIXFILE_* flags */
5786 int isExclusive
= (flags
& SQLITE_OPEN_EXCLUSIVE
);
5787 int isDelete
= (flags
& SQLITE_OPEN_DELETEONCLOSE
);
5788 int isCreate
= (flags
& SQLITE_OPEN_CREATE
);
5789 int isReadonly
= (flags
& SQLITE_OPEN_READONLY
);
5790 int isReadWrite
= (flags
& SQLITE_OPEN_READWRITE
);
5791 #if SQLITE_ENABLE_LOCKING_STYLE
5792 int isAutoProxy
= (flags
& SQLITE_OPEN_AUTOPROXY
);
5794 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
5795 struct statfs fsInfo
;
5798 /* If creating a master or main-file journal, this function will open
5799 ** a file-descriptor on the directory too. The first time unixSync()
5800 ** is called the directory file descriptor will be fsync()ed and close()d.
5802 int syncDir
= (isCreate
&& (
5803 eType
==SQLITE_OPEN_MASTER_JOURNAL
5804 || eType
==SQLITE_OPEN_MAIN_JOURNAL
5805 || eType
==SQLITE_OPEN_WAL
5808 /* If argument zPath is a NULL pointer, this function is required to open
5809 ** a temporary file. Use this buffer to store the file name in.
5811 char zTmpname
[MAX_PATHNAME
+2];
5812 const char *zName
= zPath
;
5814 /* Check the following statements are true:
5816 ** (a) Exactly one of the READWRITE and READONLY flags must be set, and
5817 ** (b) if CREATE is set, then READWRITE must also be set, and
5818 ** (c) if EXCLUSIVE is set, then CREATE must also be set.
5819 ** (d) if DELETEONCLOSE is set, then CREATE must also be set.
5821 assert((isReadonly
==0 || isReadWrite
==0) && (isReadWrite
|| isReadonly
));
5822 assert(isCreate
==0 || isReadWrite
);
5823 assert(isExclusive
==0 || isCreate
);
5824 assert(isDelete
==0 || isCreate
);
5826 /* The main DB, main journal, WAL file and master journal are never
5827 ** automatically deleted. Nor are they ever temporary files. */
5828 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_DB
);
5829 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MAIN_JOURNAL
);
5830 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_MASTER_JOURNAL
);
5831 assert( (!isDelete
&& zName
) || eType
!=SQLITE_OPEN_WAL
);
5833 /* Assert that the upper layer has set one of the "file-type" flags. */
5834 assert( eType
==SQLITE_OPEN_MAIN_DB
|| eType
==SQLITE_OPEN_TEMP_DB
5835 || eType
==SQLITE_OPEN_MAIN_JOURNAL
|| eType
==SQLITE_OPEN_TEMP_JOURNAL
5836 || eType
==SQLITE_OPEN_SUBJOURNAL
|| eType
==SQLITE_OPEN_MASTER_JOURNAL
5837 || eType
==SQLITE_OPEN_TRANSIENT_DB
|| eType
==SQLITE_OPEN_WAL
5840 /* Detect a pid change and reset the PRNG. There is a race condition
5841 ** here such that two or more threads all trying to open databases at
5842 ** the same instant might all reset the PRNG. But multiple resets
5845 if( randomnessPid
!=osGetpid(0) ){
5846 randomnessPid
= osGetpid(0);
5847 sqlite3_randomness(0,0);
5850 memset(p
, 0, sizeof(unixFile
));
5852 if( eType
==SQLITE_OPEN_MAIN_DB
){
5853 UnixUnusedFd
*pUnused
;
5854 pUnused
= findReusableFd(zName
, flags
);
5858 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
5860 return SQLITE_NOMEM_BKPT
;
5863 p
->pPreallocatedUnused
= pUnused
;
5865 /* Database filenames are double-zero terminated if they are not
5866 ** URIs with parameters. Hence, they can always be passed into
5867 ** sqlite3_uri_parameter(). */
5868 assert( (flags
& SQLITE_OPEN_URI
) || zName
[strlen(zName
)+1]==0 );
5871 /* If zName is NULL, the upper layer is requesting a temp file. */
5872 assert(isDelete
&& !syncDir
);
5873 rc
= unixGetTempname(pVfs
->mxPathname
, zTmpname
);
5874 if( rc
!=SQLITE_OK
){
5879 /* Generated temporary filenames are always double-zero terminated
5880 ** for use by sqlite3_uri_parameter(). */
5881 assert( zName
[strlen(zName
)+1]==0 );
5884 /* Determine the value of the flags parameter passed to POSIX function
5885 ** open(). These must be calculated even if open() is not called, as
5886 ** they may be stored as part of the file handle and used by the
5887 ** 'conch file' locking functions later on. */
5888 if( isReadonly
) openFlags
|= O_RDONLY
;
5889 if( isReadWrite
) openFlags
|= O_RDWR
;
5890 if( isCreate
) openFlags
|= O_CREAT
;
5891 if( isExclusive
) openFlags
|= (O_EXCL
|O_NOFOLLOW
);
5892 openFlags
|= (O_LARGEFILE
|O_BINARY
);
5895 mode_t openMode
; /* Permissions to create file with */
5896 uid_t uid
; /* Userid for the file */
5897 gid_t gid
; /* Groupid for the file */
5898 rc
= findCreateFileMode(zName
, flags
, &openMode
, &uid
, &gid
);
5899 if( rc
!=SQLITE_OK
){
5900 assert( !p
->pPreallocatedUnused
);
5901 assert( eType
==SQLITE_OPEN_WAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
);
5904 fd
= robust_open(zName
, openFlags
, openMode
);
5905 OSTRACE(("OPENX %-3d %s 0%o\n", fd
, zName
, openFlags
));
5906 assert( !isExclusive
|| (openFlags
& O_CREAT
)!=0 );
5907 if( fd
<0 && errno
!=EISDIR
&& isReadWrite
){
5908 /* Failed to open the file for read/write access. Try read-only. */
5909 flags
&= ~(SQLITE_OPEN_READWRITE
|SQLITE_OPEN_CREATE
);
5910 openFlags
&= ~(O_RDWR
|O_CREAT
);
5911 flags
|= SQLITE_OPEN_READONLY
;
5912 openFlags
|= O_RDONLY
;
5914 fd
= robust_open(zName
, openFlags
, openMode
);
5917 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "open", zName
);
5921 /* If this process is running as root and if creating a new rollback
5922 ** journal or WAL file, set the ownership of the journal or WAL to be
5923 ** the same as the original database.
5925 if( flags
& (SQLITE_OPEN_WAL
|SQLITE_OPEN_MAIN_JOURNAL
) ){
5926 robustFchown(fd
, uid
, gid
);
5934 if( p
->pPreallocatedUnused
){
5935 p
->pPreallocatedUnused
->fd
= fd
;
5936 p
->pPreallocatedUnused
->flags
= flags
;
5942 #elif defined(SQLITE_UNLINK_AFTER_CLOSE)
5943 zPath
= sqlite3_mprintf("%s", zName
);
5945 robust_close(p
, fd
, __LINE__
);
5946 return SQLITE_NOMEM_BKPT
;
5952 #if SQLITE_ENABLE_LOCKING_STYLE
5954 p
->openFlags
= openFlags
;
5958 #if defined(__APPLE__) || SQLITE_ENABLE_LOCKING_STYLE
5959 if( fstatfs(fd
, &fsInfo
) == -1 ){
5960 storeLastErrno(p
, errno
);
5961 robust_close(p
, fd
, __LINE__
);
5962 return SQLITE_IOERR_ACCESS
;
5964 if (0 == strncmp("msdos", fsInfo
.f_fstypename
, 5)) {
5965 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
5967 if (0 == strncmp("exfat", fsInfo
.f_fstypename
, 5)) {
5968 ((unixFile
*)pFile
)->fsFlags
|= SQLITE_FSFLAGS_IS_MSDOS
;
5972 /* Set up appropriate ctrlFlags */
5973 if( isDelete
) ctrlFlags
|= UNIXFILE_DELETE
;
5974 if( isReadonly
) ctrlFlags
|= UNIXFILE_RDONLY
;
5975 noLock
= eType
!=SQLITE_OPEN_MAIN_DB
;
5976 if( noLock
) ctrlFlags
|= UNIXFILE_NOLOCK
;
5977 if( syncDir
) ctrlFlags
|= UNIXFILE_DIRSYNC
;
5978 if( flags
& SQLITE_OPEN_URI
) ctrlFlags
|= UNIXFILE_URI
;
5980 #if SQLITE_ENABLE_LOCKING_STYLE
5981 #if SQLITE_PREFER_PROXY_LOCKING
5984 if( isAutoProxy
&& (zPath
!=NULL
) && (!noLock
) && pVfs
->xOpen
){
5985 char *envforce
= getenv("SQLITE_FORCE_PROXY_LOCKING");
5988 /* SQLITE_FORCE_PROXY_LOCKING==1 means force always use proxy, 0 means
5989 ** never use proxy, NULL means use proxy for non-local files only. */
5990 if( envforce
!=NULL
){
5991 useProxy
= atoi(envforce
)>0;
5993 useProxy
= !(fsInfo
.f_flags
&MNT_LOCAL
);
5996 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
5997 if( rc
==SQLITE_OK
){
5998 rc
= proxyTransformUnixFile((unixFile
*)pFile
, ":auto:");
5999 if( rc
!=SQLITE_OK
){
6000 /* Use unixClose to clean up the resources added in fillInUnixFile
6001 ** and clear all the structure's references. Specifically,
6002 ** pFile->pMethods will be NULL so sqlite3OsClose will be a no-op
6013 assert( zPath
==0 || zPath
[0]=='/'
6014 || eType
==SQLITE_OPEN_MASTER_JOURNAL
|| eType
==SQLITE_OPEN_MAIN_JOURNAL
6016 rc
= fillInUnixFile(pVfs
, fd
, pFile
, zPath
, ctrlFlags
);
6019 if( rc
!=SQLITE_OK
){
6020 sqlite3_free(p
->pPreallocatedUnused
);
6027 ** Delete the file at zPath. If the dirSync argument is true, fsync()
6028 ** the directory after deleting the file.
6030 static int unixDelete(
6031 sqlite3_vfs
*NotUsed
, /* VFS containing this as the xDelete method */
6032 const char *zPath
, /* Name of file to be deleted */
6033 int dirSync
/* If true, fsync() directory after deleting file */
6036 UNUSED_PARAMETER(NotUsed
);
6037 SimulateIOError(return SQLITE_IOERR_DELETE
);
6038 if( osUnlink(zPath
)==(-1) ){
6041 || osAccess(zPath
,0)!=0
6044 rc
= SQLITE_IOERR_DELETE_NOENT
;
6046 rc
= unixLogError(SQLITE_IOERR_DELETE
, "unlink", zPath
);
6050 #ifndef SQLITE_DISABLE_DIRSYNC
6051 if( (dirSync
& 1)!=0 ){
6053 rc
= osOpenDirectory(zPath
, &fd
);
6054 if( rc
==SQLITE_OK
){
6055 if( full_fsync(fd
,0,0) ){
6056 rc
= unixLogError(SQLITE_IOERR_DIR_FSYNC
, "fsync", zPath
);
6058 robust_close(0, fd
, __LINE__
);
6060 assert( rc
==SQLITE_CANTOPEN
);
6069 ** Test the existence of or access permissions of file zPath. The
6070 ** test performed depends on the value of flags:
6072 ** SQLITE_ACCESS_EXISTS: Return 1 if the file exists
6073 ** SQLITE_ACCESS_READWRITE: Return 1 if the file is read and writable.
6074 ** SQLITE_ACCESS_READONLY: Return 1 if the file is readable.
6076 ** Otherwise return 0.
6078 static int unixAccess(
6079 sqlite3_vfs
*NotUsed
, /* The VFS containing this xAccess method */
6080 const char *zPath
, /* Path of the file to examine */
6081 int flags
, /* What do we want to learn about the zPath file? */
6082 int *pResOut
/* Write result boolean here */
6084 UNUSED_PARAMETER(NotUsed
);
6085 SimulateIOError( return SQLITE_IOERR_ACCESS
; );
6086 assert( pResOut
!=0 );
6088 /* The spec says there are three possible values for flags. But only
6089 ** two of them are actually used */
6090 assert( flags
==SQLITE_ACCESS_EXISTS
|| flags
==SQLITE_ACCESS_READWRITE
);
6092 if( flags
==SQLITE_ACCESS_EXISTS
){
6094 *pResOut
= (0==osStat(zPath
, &buf
) && buf
.st_size
>0);
6096 *pResOut
= osAccess(zPath
, W_OK
|R_OK
)==0;
6104 static int mkFullPathname(
6105 const char *zPath
, /* Input path */
6106 char *zOut
, /* Output buffer */
6107 int nOut
/* Allocated size of buffer zOut */
6109 int nPath
= sqlite3Strlen30(zPath
);
6111 if( zPath
[0]!='/' ){
6112 if( osGetcwd(zOut
, nOut
-2)==0 ){
6113 return unixLogError(SQLITE_CANTOPEN_BKPT
, "getcwd", zPath
);
6115 iOff
= sqlite3Strlen30(zOut
);
6118 if( (iOff
+nPath
+1)>nOut
){
6119 /* SQLite assumes that xFullPathname() nul-terminates the output buffer
6120 ** even if it returns an error. */
6122 return SQLITE_CANTOPEN_BKPT
;
6124 sqlite3_snprintf(nOut
-iOff
, &zOut
[iOff
], "%s", zPath
);
6129 ** Turn a relative pathname into a full pathname. The relative path
6130 ** is stored as a nul-terminated string in the buffer pointed to by
6133 ** zOut points to a buffer of at least sqlite3_vfs.mxPathname bytes
6134 ** (in this case, MAX_PATHNAME bytes). The full-path is written to
6135 ** this buffer before returning.
6137 static int unixFullPathname(
6138 sqlite3_vfs
*pVfs
, /* Pointer to vfs object */
6139 const char *zPath
, /* Possibly relative input path */
6140 int nOut
, /* Size of output buffer in bytes */
6141 char *zOut
/* Output buffer */
6143 #if !defined(HAVE_READLINK) || !defined(HAVE_LSTAT)
6144 return mkFullPathname(zPath
, zOut
, nOut
);
6148 int nLink
= 1; /* Number of symbolic links followed so far */
6149 const char *zIn
= zPath
; /* Input path for each iteration of loop */
6152 assert( pVfs
->mxPathname
==MAX_PATHNAME
);
6153 UNUSED_PARAMETER(pVfs
);
6155 /* It's odd to simulate an io-error here, but really this is just
6156 ** using the io-error infrastructure to test that SQLite handles this
6157 ** function failing. This function could fail if, for example, the
6158 ** current working directory has been unlinked.
6160 SimulateIOError( return SQLITE_ERROR
);
6164 /* Call stat() on path zIn. Set bLink to true if the path is a symbolic
6165 ** link, or false otherwise. */
6168 if( osLstat(zIn
, &buf
)!=0 ){
6169 if( errno
!=ENOENT
){
6170 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "lstat", zIn
);
6173 bLink
= S_ISLNK(buf
.st_mode
);
6178 zDel
= sqlite3_malloc(nOut
);
6179 if( zDel
==0 ) rc
= SQLITE_NOMEM_BKPT
;
6180 }else if( ++nLink
>SQLITE_MAX_SYMLINKS
){
6181 rc
= SQLITE_CANTOPEN_BKPT
;
6184 if( rc
==SQLITE_OK
){
6185 nByte
= osReadlink(zIn
, zDel
, nOut
-1);
6187 rc
= unixLogError(SQLITE_CANTOPEN_BKPT
, "readlink", zIn
);
6191 for(n
= sqlite3Strlen30(zIn
); n
>0 && zIn
[n
-1]!='/'; n
--);
6192 if( nByte
+n
+1>nOut
){
6193 rc
= SQLITE_CANTOPEN_BKPT
;
6195 memmove(&zDel
[n
], zDel
, nByte
+1);
6196 memcpy(zDel
, zIn
, n
);
6207 assert( rc
!=SQLITE_OK
|| zIn
!=zOut
|| zIn
[0]=='/' );
6208 if( rc
==SQLITE_OK
&& zIn
!=zOut
){
6209 rc
= mkFullPathname(zIn
, zOut
, nOut
);
6211 if( bLink
==0 ) break;
6213 }while( rc
==SQLITE_OK
);
6217 #endif /* HAVE_READLINK && HAVE_LSTAT */
6221 #ifndef SQLITE_OMIT_LOAD_EXTENSION
6223 ** Interfaces for opening a shared library, finding entry points
6224 ** within the shared library, and closing the shared library.
6227 static void *unixDlOpen(sqlite3_vfs
*NotUsed
, const char *zFilename
){
6228 UNUSED_PARAMETER(NotUsed
);
6229 return dlopen(zFilename
, RTLD_NOW
| RTLD_GLOBAL
);
6233 ** SQLite calls this function immediately after a call to unixDlSym() or
6234 ** unixDlOpen() fails (returns a null pointer). If a more detailed error
6235 ** message is available, it is written to zBufOut. If no error message
6236 ** is available, zBufOut is left unmodified and SQLite uses a default
6239 static void unixDlError(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBufOut
){
6241 UNUSED_PARAMETER(NotUsed
);
6245 sqlite3_snprintf(nBuf
, zBufOut
, "%s", zErr
);
6249 static void (*unixDlSym(sqlite3_vfs
*NotUsed
, void *p
, const char*zSym
))(void){
6251 ** GCC with -pedantic-errors says that C90 does not allow a void* to be
6252 ** cast into a pointer to a function. And yet the library dlsym() routine
6253 ** returns a void* which is really a pointer to a function. So how do we
6254 ** use dlsym() with -pedantic-errors?
6256 ** Variable x below is defined to be a pointer to a function taking
6257 ** parameters void* and const char* and returning a pointer to a function.
6258 ** We initialize x by assigning it a pointer to the dlsym() function.
6259 ** (That assignment requires a cast.) Then we call the function that
6262 ** This work-around is unlikely to work correctly on any system where
6263 ** you really cannot cast a function pointer into void*. But then, on the
6264 ** other hand, dlsym() will not work on such a system either, so we have
6265 ** not really lost anything.
6267 void (*(*x
)(void*,const char*))(void);
6268 UNUSED_PARAMETER(NotUsed
);
6269 x
= (void(*(*)(void*,const char*))(void))dlsym
;
6270 return (*x
)(p
, zSym
);
6272 static void unixDlClose(sqlite3_vfs
*NotUsed
, void *pHandle
){
6273 UNUSED_PARAMETER(NotUsed
);
6276 #else /* if SQLITE_OMIT_LOAD_EXTENSION is defined: */
6277 #define unixDlOpen 0
6278 #define unixDlError 0
6280 #define unixDlClose 0
6284 ** Write nBuf bytes of random data to the supplied buffer zBuf.
6286 static int unixRandomness(sqlite3_vfs
*NotUsed
, int nBuf
, char *zBuf
){
6287 UNUSED_PARAMETER(NotUsed
);
6288 assert((size_t)nBuf
>=(sizeof(time_t)+sizeof(int)));
6290 /* We have to initialize zBuf to prevent valgrind from reporting
6291 ** errors. The reports issued by valgrind are incorrect - we would
6292 ** prefer that the randomness be increased by making use of the
6293 ** uninitialized space in zBuf - but valgrind errors tend to worry
6294 ** some users. Rather than argue, it seems easier just to initialize
6295 ** the whole array and silence valgrind, even if that means less randomness
6296 ** in the random seed.
6298 ** When testing, initializing zBuf[] to zero is all we do. That means
6299 ** that we always use the same random number sequence. This makes the
6300 ** tests repeatable.
6302 memset(zBuf
, 0, nBuf
);
6303 randomnessPid
= osGetpid(0);
6304 #if !defined(SQLITE_TEST) && !defined(SQLITE_OMIT_RANDOMNESS)
6307 fd
= robust_open("/dev/urandom", O_RDONLY
, 0);
6311 memcpy(zBuf
, &t
, sizeof(t
));
6312 memcpy(&zBuf
[sizeof(t
)], &randomnessPid
, sizeof(randomnessPid
));
6313 assert( sizeof(t
)+sizeof(randomnessPid
)<=(size_t)nBuf
);
6314 nBuf
= sizeof(t
) + sizeof(randomnessPid
);
6316 do{ got
= osRead(fd
, zBuf
, nBuf
); }while( got
<0 && errno
==EINTR
);
6317 robust_close(0, fd
, __LINE__
);
6326 ** Sleep for a little while. Return the amount of time slept.
6327 ** The argument is the number of microseconds we want to sleep.
6328 ** The return value is the number of microseconds of sleep actually
6329 ** requested from the underlying operating system, a number which
6330 ** might be greater than or equal to the argument, but not less
6331 ** than the argument.
6333 static int unixSleep(sqlite3_vfs
*NotUsed
, int microseconds
){
6337 sp
.tv_sec
= microseconds
/ 1000000;
6338 sp
.tv_nsec
= (microseconds
% 1000000) * 1000;
6339 nanosleep(&sp
, NULL
);
6340 UNUSED_PARAMETER(NotUsed
);
6341 return microseconds
;
6342 #elif defined(HAVE_USLEEP) && HAVE_USLEEP
6343 usleep(microseconds
);
6344 UNUSED_PARAMETER(NotUsed
);
6345 return microseconds
;
6347 int seconds
= (microseconds
+999999)/1000000;
6349 UNUSED_PARAMETER(NotUsed
);
6350 return seconds
*1000000;
6355 ** The following variable, if set to a non-zero value, is interpreted as
6356 ** the number of seconds since 1970 and is used to set the result of
6357 ** sqlite3OsCurrentTime() during testing.
6360 int sqlite3_current_time
= 0; /* Fake system time in seconds since 1970. */
6364 ** Find the current time (in Universal Coordinated Time). Write into *piNow
6365 ** the current time and date as a Julian Day number times 86_400_000. In
6366 ** other words, write into *piNow the number of milliseconds since the Julian
6367 ** epoch of noon in Greenwich on November 24, 4714 B.C according to the
6368 ** proleptic Gregorian calendar.
6370 ** On success, return SQLITE_OK. Return SQLITE_ERROR if the time and date
6373 static int unixCurrentTimeInt64(sqlite3_vfs
*NotUsed
, sqlite3_int64
*piNow
){
6374 static const sqlite3_int64 unixEpoch
= 24405875*(sqlite3_int64
)8640000;
6376 #if defined(NO_GETTOD)
6379 *piNow
= ((sqlite3_int64
)t
)*1000 + unixEpoch
;
6381 struct timespec sNow
;
6382 clock_gettime(CLOCK_REALTIME
, &sNow
);
6383 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_nsec
/1000000;
6385 struct timeval sNow
;
6386 (void)gettimeofday(&sNow
, 0); /* Cannot fail given valid arguments */
6387 *piNow
= unixEpoch
+ 1000*(sqlite3_int64
)sNow
.tv_sec
+ sNow
.tv_usec
/1000;
6391 if( sqlite3_current_time
){
6392 *piNow
= 1000*(sqlite3_int64
)sqlite3_current_time
+ unixEpoch
;
6395 UNUSED_PARAMETER(NotUsed
);
6399 #ifndef SQLITE_OMIT_DEPRECATED
6401 ** Find the current time (in Universal Coordinated Time). Write the
6402 ** current time and date as a Julian Day number into *prNow and
6403 ** return 0. Return 1 if the time and date cannot be found.
6405 static int unixCurrentTime(sqlite3_vfs
*NotUsed
, double *prNow
){
6406 sqlite3_int64 i
= 0;
6408 UNUSED_PARAMETER(NotUsed
);
6409 rc
= unixCurrentTimeInt64(0, &i
);
6410 *prNow
= i
/86400000.0;
6414 # define unixCurrentTime 0
6418 ** The xGetLastError() method is designed to return a better
6419 ** low-level error message when operating-system problems come up
6420 ** during SQLite operation. Only the integer return code is currently
6423 static int unixGetLastError(sqlite3_vfs
*NotUsed
, int NotUsed2
, char *NotUsed3
){
6424 UNUSED_PARAMETER(NotUsed
);
6425 UNUSED_PARAMETER(NotUsed2
);
6426 UNUSED_PARAMETER(NotUsed3
);
6432 ************************ End of sqlite3_vfs methods ***************************
6433 ******************************************************************************/
6435 /******************************************************************************
6436 ************************** Begin Proxy Locking ********************************
6438 ** Proxy locking is a "uber-locking-method" in this sense: It uses the
6439 ** other locking methods on secondary lock files. Proxy locking is a
6440 ** meta-layer over top of the primitive locking implemented above. For
6441 ** this reason, the division that implements of proxy locking is deferred
6442 ** until late in the file (here) after all of the other I/O methods have
6443 ** been defined - so that the primitive locking methods are available
6444 ** as services to help with the implementation of proxy locking.
6448 ** The default locking schemes in SQLite use byte-range locks on the
6449 ** database file to coordinate safe, concurrent access by multiple readers
6450 ** and writers [http://sqlite.org/lockingv3.html]. The five file locking
6451 ** states (UNLOCKED, PENDING, SHARED, RESERVED, EXCLUSIVE) are implemented
6452 ** as POSIX read & write locks over fixed set of locations (via fsctl),
6453 ** on AFP and SMB only exclusive byte-range locks are available via fsctl
6454 ** with _IOWR('z', 23, struct ByteRangeLockPB2) to track the same 5 states.
6455 ** To simulate a F_RDLCK on the shared range, on AFP a randomly selected
6456 ** address in the shared range is taken for a SHARED lock, the entire
6457 ** shared range is taken for an EXCLUSIVE lock):
6459 ** PENDING_BYTE 0x40000000
6460 ** RESERVED_BYTE 0x40000001
6461 ** SHARED_RANGE 0x40000002 -> 0x40000200
6463 ** This works well on the local file system, but shows a nearly 100x
6464 ** slowdown in read performance on AFP because the AFP client disables
6465 ** the read cache when byte-range locks are present. Enabling the read
6466 ** cache exposes a cache coherency problem that is present on all OS X
6467 ** supported network file systems. NFS and AFP both observe the
6468 ** close-to-open semantics for ensuring cache coherency
6469 ** [http://nfs.sourceforge.net/#faq_a8], which does not effectively
6470 ** address the requirements for concurrent database access by multiple
6471 ** readers and writers
6472 ** [http://www.nabble.com/SQLite-on-NFS-cache-coherency-td15655701.html].
6474 ** To address the performance and cache coherency issues, proxy file locking
6475 ** changes the way database access is controlled by limiting access to a
6476 ** single host at a time and moving file locks off of the database file
6477 ** and onto a proxy file on the local file system.
6480 ** Using proxy locks
6481 ** -----------------
6485 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_SET_LOCKPROXYFILE,
6486 ** <proxy_path> | ":auto:");
6487 ** sqlite3_file_control(db, dbname, SQLITE_FCNTL_GET_LOCKPROXYFILE,
6493 ** PRAGMA [database.]lock_proxy_file=<proxy_path> | :auto:
6494 ** PRAGMA [database.]lock_proxy_file
6496 ** Specifying ":auto:" means that if there is a conch file with a matching
6497 ** host ID in it, the proxy path in the conch file will be used, otherwise
6498 ** a proxy path based on the user's temp dir
6499 ** (via confstr(_CS_DARWIN_USER_TEMP_DIR,...)) will be used and the
6500 ** actual proxy file name is generated from the name and path of the
6501 ** database file. For example:
6503 ** For database path "/Users/me/foo.db"
6504 ** The lock path will be "<tmpdir>/sqliteplocks/_Users_me_foo.db:auto:")
6506 ** Once a lock proxy is configured for a database connection, it can not
6507 ** be removed, however it may be switched to a different proxy path via
6508 ** the above APIs (assuming the conch file is not being held by another
6509 ** connection or process).
6512 ** How proxy locking works
6513 ** -----------------------
6515 ** Proxy file locking relies primarily on two new supporting files:
6517 ** * conch file to limit access to the database file to a single host
6520 ** * proxy file to act as a proxy for the advisory locks normally
6521 ** taken on the database
6523 ** The conch file - to use a proxy file, sqlite must first "hold the conch"
6524 ** by taking an sqlite-style shared lock on the conch file, reading the
6525 ** contents and comparing the host's unique host ID (see below) and lock
6526 ** proxy path against the values stored in the conch. The conch file is
6527 ** stored in the same directory as the database file and the file name
6528 ** is patterned after the database file name as ".<databasename>-conch".
6529 ** If the conch file does not exist, or its contents do not match the
6530 ** host ID and/or proxy path, then the lock is escalated to an exclusive
6531 ** lock and the conch file contents is updated with the host ID and proxy
6532 ** path and the lock is downgraded to a shared lock again. If the conch
6533 ** is held by another process (with a shared lock), the exclusive lock
6534 ** will fail and SQLITE_BUSY is returned.
6536 ** The proxy file - a single-byte file used for all advisory file locks
6537 ** normally taken on the database file. This allows for safe sharing
6538 ** of the database file for multiple readers and writers on the same
6539 ** host (the conch ensures that they all use the same local lock file).
6541 ** Requesting the lock proxy does not immediately take the conch, it is
6542 ** only taken when the first request to lock database file is made.
6543 ** This matches the semantics of the traditional locking behavior, where
6544 ** opening a connection to a database file does not take a lock on it.
6545 ** The shared lock and an open file descriptor are maintained until
6546 ** the connection to the database is closed.
6548 ** The proxy file and the lock file are never deleted so they only need
6549 ** to be created the first time they are used.
6551 ** Configuration options
6552 ** ---------------------
6554 ** SQLITE_PREFER_PROXY_LOCKING
6556 ** Database files accessed on non-local file systems are
6557 ** automatically configured for proxy locking, lock files are
6558 ** named automatically using the same logic as
6559 ** PRAGMA lock_proxy_file=":auto:"
6561 ** SQLITE_PROXY_DEBUG
6563 ** Enables the logging of error messages during host id file
6564 ** retrieval and creation
6568 ** Overrides the default directory used for lock proxy files that
6569 ** are named automatically via the ":auto:" setting
6571 ** SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
6573 ** Permissions to use when creating a directory for storing the
6574 ** lock proxy files, only used when LOCKPROXYDIR is not set.
6577 ** As mentioned above, when compiled with SQLITE_PREFER_PROXY_LOCKING,
6578 ** setting the environment variable SQLITE_FORCE_PROXY_LOCKING to 1 will
6579 ** force proxy locking to be used for every database file opened, and 0
6580 ** will force automatic proxy locking to be disabled for all database
6581 ** files (explicitly calling the SQLITE_FCNTL_SET_LOCKPROXYFILE pragma or
6582 ** sqlite_file_control API is not affected by SQLITE_FORCE_PROXY_LOCKING).
6586 ** Proxy locking is only available on MacOSX
6588 #if defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE
6591 ** The proxyLockingContext has the path and file structures for the remote
6592 ** and local proxy files in it
6594 typedef struct proxyLockingContext proxyLockingContext
;
6595 struct proxyLockingContext
{
6596 unixFile
*conchFile
; /* Open conch file */
6597 char *conchFilePath
; /* Name of the conch file */
6598 unixFile
*lockProxy
; /* Open proxy lock file */
6599 char *lockProxyPath
; /* Name of the proxy lock file */
6600 char *dbPath
; /* Name of the open file */
6601 int conchHeld
; /* 1 if the conch is held, -1 if lockless */
6602 int nFails
; /* Number of conch taking failures */
6603 void *oldLockingContext
; /* Original lockingcontext to restore on close */
6604 sqlite3_io_methods
const *pOldMethod
; /* Original I/O methods for close */
6608 ** The proxy lock file path for the database at dbPath is written into lPath,
6609 ** which must point to valid, writable memory large enough for a maxLen length
6612 static int proxyGetLockPath(const char *dbPath
, char *lPath
, size_t maxLen
){
6618 len
= strlcpy(lPath
, LOCKPROXYDIR
, maxLen
);
6620 # ifdef _CS_DARWIN_USER_TEMP_DIR
6622 if( !confstr(_CS_DARWIN_USER_TEMP_DIR
, lPath
, maxLen
) ){
6623 OSTRACE(("GETLOCKPATH failed %s errno=%d pid=%d\n",
6624 lPath
, errno
, osGetpid(0)));
6625 return SQLITE_IOERR_LOCK
;
6627 len
= strlcat(lPath
, "sqliteplocks", maxLen
);
6630 len
= strlcpy(lPath
, "/tmp/", maxLen
);
6634 if( lPath
[len
-1]!='/' ){
6635 len
= strlcat(lPath
, "/", maxLen
);
6638 /* transform the db path to a unique cache name */
6639 dbLen
= (int)strlen(dbPath
);
6640 for( i
=0; i
<dbLen
&& (i
+len
+7)<(int)maxLen
; i
++){
6642 lPath
[i
+len
] = (c
=='/')?'_':c
;
6645 strlcat(lPath
, ":auto:", maxLen
);
6646 OSTRACE(("GETLOCKPATH proxy lock path=%s pid=%d\n", lPath
, osGetpid(0)));
6651 ** Creates the lock file and any missing directories in lockPath
6653 static int proxyCreateLockPath(const char *lockPath
){
6655 char buf
[MAXPATHLEN
];
6658 assert(lockPath
!=NULL
);
6659 /* try to create all the intermediate directories */
6660 len
= (int)strlen(lockPath
);
6661 buf
[0] = lockPath
[0];
6662 for( i
=1; i
<len
; i
++ ){
6663 if( lockPath
[i
] == '/' && (i
- start
> 0) ){
6664 /* only mkdir if leaf dir != "." or "/" or ".." */
6665 if( i
-start
>2 || (i
-start
==1 && buf
[start
] != '.' && buf
[start
] != '/')
6666 || (i
-start
==2 && buf
[start
] != '.' && buf
[start
+1] != '.') ){
6668 if( osMkdir(buf
, SQLITE_DEFAULT_PROXYDIR_PERMISSIONS
) ){
6671 OSTRACE(("CREATELOCKPATH FAILED creating %s, "
6672 "'%s' proxy lock path=%s pid=%d\n",
6673 buf
, strerror(err
), lockPath
, osGetpid(0)));
6680 buf
[i
] = lockPath
[i
];
6682 OSTRACE(("CREATELOCKPATH proxy lock path=%s pid=%d\n",lockPath
,osGetpid(0)));
6687 ** Create a new VFS file descriptor (stored in memory obtained from
6688 ** sqlite3_malloc) and open the file named "path" in the file descriptor.
6690 ** The caller is responsible not only for closing the file descriptor
6691 ** but also for freeing the memory associated with the file descriptor.
6693 static int proxyCreateUnixFile(
6694 const char *path
, /* path for the new unixFile */
6695 unixFile
**ppFile
, /* unixFile created and returned by ref */
6696 int islockfile
/* if non zero missing dirs will be created */
6701 int openFlags
= O_RDWR
| O_CREAT
;
6702 sqlite3_vfs dummyVfs
;
6704 UnixUnusedFd
*pUnused
= NULL
;
6706 /* 1. first try to open/create the file
6707 ** 2. if that fails, and this is a lock file (not-conch), try creating
6708 ** the parent directories and then try again.
6709 ** 3. if that fails, try to open the file read-only
6710 ** otherwise return BUSY (if lock file) or CANTOPEN for the conch file
6712 pUnused
= findReusableFd(path
, openFlags
);
6716 pUnused
= sqlite3_malloc64(sizeof(*pUnused
));
6718 return SQLITE_NOMEM_BKPT
;
6722 fd
= robust_open(path
, openFlags
, 0);
6724 if( fd
<0 && errno
==ENOENT
&& islockfile
){
6725 if( proxyCreateLockPath(path
) == SQLITE_OK
){
6726 fd
= robust_open(path
, openFlags
, 0);
6731 openFlags
= O_RDONLY
;
6732 fd
= robust_open(path
, openFlags
, 0);
6743 return SQLITE_IOERR_LOCK
; /* even though it is the conch */
6745 return SQLITE_CANTOPEN_BKPT
;
6749 pNew
= (unixFile
*)sqlite3_malloc64(sizeof(*pNew
));
6751 rc
= SQLITE_NOMEM_BKPT
;
6752 goto end_create_proxy
;
6754 memset(pNew
, 0, sizeof(unixFile
));
6755 pNew
->openFlags
= openFlags
;
6756 memset(&dummyVfs
, 0, sizeof(dummyVfs
));
6757 dummyVfs
.pAppData
= (void*)&autolockIoFinder
;
6758 dummyVfs
.zName
= "dummy";
6760 pUnused
->flags
= openFlags
;
6761 pNew
->pPreallocatedUnused
= pUnused
;
6763 rc
= fillInUnixFile(&dummyVfs
, fd
, (sqlite3_file
*)pNew
, path
, 0);
6764 if( rc
==SQLITE_OK
){
6769 robust_close(pNew
, fd
, __LINE__
);
6771 sqlite3_free(pUnused
);
6776 /* simulate multiple hosts by creating unique hostid file paths */
6777 int sqlite3_hostid_num
= 0;
6780 #define PROXY_HOSTIDLEN 16 /* conch file host id length */
6782 #ifdef HAVE_GETHOSTUUID
6783 /* Not always defined in the headers as it ought to be */
6784 extern int gethostuuid(uuid_t id
, const struct timespec
*wait
);
6787 /* get the host ID via gethostuuid(), pHostID must point to PROXY_HOSTIDLEN
6788 ** bytes of writable memory.
6790 static int proxyGetHostID(unsigned char *pHostID
, int *pError
){
6791 assert(PROXY_HOSTIDLEN
== sizeof(uuid_t
));
6792 memset(pHostID
, 0, PROXY_HOSTIDLEN
);
6793 #ifdef HAVE_GETHOSTUUID
6795 struct timespec timeout
= {1, 0}; /* 1 sec timeout */
6796 if( gethostuuid(pHostID
, &timeout
) ){
6801 return SQLITE_IOERR
;
6805 UNUSED_PARAMETER(pError
);
6808 /* simulate multiple hosts by creating unique hostid file paths */
6809 if( sqlite3_hostid_num
!= 0){
6810 pHostID
[0] = (char)(pHostID
[0] + (char)(sqlite3_hostid_num
& 0xFF));
6817 /* The conch file contains the header, host id and lock file path
6819 #define PROXY_CONCHVERSION 2 /* 1-byte header, 16-byte host id, path */
6820 #define PROXY_HEADERLEN 1 /* conch file header length */
6821 #define PROXY_PATHINDEX (PROXY_HEADERLEN+PROXY_HOSTIDLEN)
6822 #define PROXY_MAXCONCHLEN (PROXY_HEADERLEN+PROXY_HOSTIDLEN+MAXPATHLEN)
6825 ** Takes an open conch file, copies the contents to a new path and then moves
6826 ** it back. The newly created file's file descriptor is assigned to the
6827 ** conch file structure and finally the original conch file descriptor is
6828 ** closed. Returns zero if successful.
6830 static int proxyBreakConchLock(unixFile
*pFile
, uuid_t myHostID
){
6831 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6832 unixFile
*conchFile
= pCtx
->conchFile
;
6833 char tPath
[MAXPATHLEN
];
6834 char buf
[PROXY_MAXCONCHLEN
];
6835 char *cPath
= pCtx
->conchFilePath
;
6838 char errmsg
[64] = "";
6841 UNUSED_PARAMETER(myHostID
);
6843 /* create a new path by replace the trailing '-conch' with '-break' */
6844 pathLen
= strlcpy(tPath
, cPath
, MAXPATHLEN
);
6845 if( pathLen
>MAXPATHLEN
|| pathLen
<6 ||
6846 (strlcpy(&tPath
[pathLen
-5], "break", 6) != 5) ){
6847 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"path error (len %d)",(int)pathLen
);
6850 /* read the conch content */
6851 readLen
= osPread(conchFile
->h
, buf
, PROXY_MAXCONCHLEN
, 0);
6852 if( readLen
<PROXY_PATHINDEX
){
6853 sqlite3_snprintf(sizeof(errmsg
),errmsg
,"read error (len %d)",(int)readLen
);
6856 /* write it out to the temporary break file */
6857 fd
= robust_open(tPath
, (O_RDWR
|O_CREAT
|O_EXCL
), 0);
6859 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "create failed (%d)", errno
);
6862 if( osPwrite(fd
, buf
, readLen
, 0) != (ssize_t
)readLen
){
6863 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "write failed (%d)", errno
);
6866 if( rename(tPath
, cPath
) ){
6867 sqlite3_snprintf(sizeof(errmsg
), errmsg
, "rename failed (%d)", errno
);
6871 fprintf(stderr
, "broke stale lock on %s\n", cPath
);
6872 robust_close(pFile
, conchFile
->h
, __LINE__
);
6874 conchFile
->openFlags
= O_RDWR
| O_CREAT
;
6880 robust_close(pFile
, fd
, __LINE__
);
6882 fprintf(stderr
, "failed to break stale lock on %s, %s\n", cPath
, errmsg
);
6887 /* Take the requested lock on the conch file and break a stale lock if the
6890 static int proxyConchLock(unixFile
*pFile
, uuid_t myHostID
, int lockType
){
6891 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6892 unixFile
*conchFile
= pCtx
->conchFile
;
6895 struct timespec conchModTime
;
6897 memset(&conchModTime
, 0, sizeof(conchModTime
));
6899 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
6901 if( rc
==SQLITE_BUSY
){
6902 /* If the lock failed (busy):
6903 * 1st try: get the mod time of the conch, wait 0.5s and try again.
6904 * 2nd try: fail if the mod time changed or host id is different, wait
6905 * 10 sec and try again
6906 * 3rd try: break the lock unless the mod time has changed.
6909 if( osFstat(conchFile
->h
, &buf
) ){
6910 storeLastErrno(pFile
, errno
);
6911 return SQLITE_IOERR_LOCK
;
6915 conchModTime
= buf
.st_mtimespec
;
6916 usleep(500000); /* wait 0.5 sec and try the lock again*/
6921 if( conchModTime
.tv_sec
!= buf
.st_mtimespec
.tv_sec
||
6922 conchModTime
.tv_nsec
!= buf
.st_mtimespec
.tv_nsec
){
6927 char tBuf
[PROXY_MAXCONCHLEN
];
6928 int len
= osPread(conchFile
->h
, tBuf
, PROXY_MAXCONCHLEN
, 0);
6930 storeLastErrno(pFile
, errno
);
6931 return SQLITE_IOERR_LOCK
;
6933 if( len
>PROXY_PATHINDEX
&& tBuf
[0]==(char)PROXY_CONCHVERSION
){
6934 /* don't break the lock if the host id doesn't match */
6935 if( 0!=memcmp(&tBuf
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
) ){
6939 /* don't break the lock on short read or a version mismatch */
6942 usleep(10000000); /* wait 10 sec and try the lock again */
6946 assert( nTries
==3 );
6947 if( 0==proxyBreakConchLock(pFile
, myHostID
) ){
6949 if( lockType
==EXCLUSIVE_LOCK
){
6950 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
6953 rc
= conchFile
->pMethod
->xLock((sqlite3_file
*)conchFile
, lockType
);
6957 } while( rc
==SQLITE_BUSY
&& nTries
<3 );
6962 /* Takes the conch by taking a shared lock and read the contents conch, if
6963 ** lockPath is non-NULL, the host ID and lock file path must match. A NULL
6964 ** lockPath means that the lockPath in the conch file will be used if the
6965 ** host IDs match, or a new lock path will be generated automatically
6966 ** and written to the conch file.
6968 static int proxyTakeConch(unixFile
*pFile
){
6969 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
6971 if( pCtx
->conchHeld
!=0 ){
6974 unixFile
*conchFile
= pCtx
->conchFile
;
6977 char readBuf
[PROXY_MAXCONCHLEN
];
6978 char lockPath
[MAXPATHLEN
];
6979 char *tempLockPath
= NULL
;
6981 int createConch
= 0;
6982 int hostIdMatch
= 0;
6984 int tryOldLockPath
= 0;
6985 int forceNewLockPath
= 0;
6987 OSTRACE(("TAKECONCH %d for %s pid=%d\n", conchFile
->h
,
6988 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
6991 rc
= proxyGetHostID(myHostID
, &pError
);
6992 if( (rc
&0xff)==SQLITE_IOERR
){
6993 storeLastErrno(pFile
, pError
);
6996 rc
= proxyConchLock(pFile
, myHostID
, SHARED_LOCK
);
6997 if( rc
!=SQLITE_OK
){
7000 /* read the existing conch file */
7001 readLen
= seekAndRead((unixFile
*)conchFile
, 0, readBuf
, PROXY_MAXCONCHLEN
);
7003 /* I/O error: lastErrno set by seekAndRead */
7004 storeLastErrno(pFile
, conchFile
->lastErrno
);
7005 rc
= SQLITE_IOERR_READ
;
7007 }else if( readLen
<=(PROXY_HEADERLEN
+PROXY_HOSTIDLEN
) ||
7008 readBuf
[0]!=(char)PROXY_CONCHVERSION
){
7009 /* a short read or version format mismatch means we need to create a new
7014 /* if the host id matches and the lock path already exists in the conch
7015 ** we'll try to use the path there, if we can't open that path, we'll
7016 ** retry with a new auto-generated path
7018 do { /* in case we need to try again for an :auto: named lock file */
7020 if( !createConch
&& !forceNewLockPath
){
7021 hostIdMatch
= !memcmp(&readBuf
[PROXY_HEADERLEN
], myHostID
,
7023 /* if the conch has data compare the contents */
7024 if( !pCtx
->lockProxyPath
){
7025 /* for auto-named local lock file, just check the host ID and we'll
7026 ** use the local lock file path that's already in there
7029 size_t pathLen
= (readLen
- PROXY_PATHINDEX
);
7031 if( pathLen
>=MAXPATHLEN
){
7032 pathLen
=MAXPATHLEN
-1;
7034 memcpy(lockPath
, &readBuf
[PROXY_PATHINDEX
], pathLen
);
7035 lockPath
[pathLen
] = 0;
7036 tempLockPath
= lockPath
;
7038 /* create a copy of the lock path if the conch is taken */
7041 }else if( hostIdMatch
7042 && !strncmp(pCtx
->lockProxyPath
, &readBuf
[PROXY_PATHINDEX
],
7043 readLen
-PROXY_PATHINDEX
)
7045 /* conch host and lock path match */
7050 /* if the conch isn't writable and doesn't match, we can't take it */
7051 if( (conchFile
->openFlags
&O_RDWR
) == 0 ){
7056 /* either the conch didn't match or we need to create a new one */
7057 if( !pCtx
->lockProxyPath
){
7058 proxyGetLockPath(pCtx
->dbPath
, lockPath
, MAXPATHLEN
);
7059 tempLockPath
= lockPath
;
7060 /* create a copy of the lock path _only_ if the conch is taken */
7063 /* update conch with host and path (this will fail if other process
7064 ** has a shared lock already), if the host id matches, use the big
7067 futimes(conchFile
->h
, NULL
);
7068 if( hostIdMatch
&& !createConch
){
7069 if( conchFile
->pInode
&& conchFile
->pInode
->nShared
>1 ){
7070 /* We are trying for an exclusive lock but another thread in this
7071 ** same process is still holding a shared lock. */
7074 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7077 rc
= proxyConchLock(pFile
, myHostID
, EXCLUSIVE_LOCK
);
7079 if( rc
==SQLITE_OK
){
7080 char writeBuffer
[PROXY_MAXCONCHLEN
];
7083 writeBuffer
[0] = (char)PROXY_CONCHVERSION
;
7084 memcpy(&writeBuffer
[PROXY_HEADERLEN
], myHostID
, PROXY_HOSTIDLEN
);
7085 if( pCtx
->lockProxyPath
!=NULL
){
7086 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], pCtx
->lockProxyPath
,
7089 strlcpy(&writeBuffer
[PROXY_PATHINDEX
], tempLockPath
, MAXPATHLEN
);
7091 writeSize
= PROXY_PATHINDEX
+ strlen(&writeBuffer
[PROXY_PATHINDEX
]);
7092 robust_ftruncate(conchFile
->h
, writeSize
);
7093 rc
= unixWrite((sqlite3_file
*)conchFile
, writeBuffer
, writeSize
, 0);
7094 full_fsync(conchFile
->h
,0,0);
7095 /* If we created a new conch file (not just updated the contents of a
7096 ** valid conch file), try to match the permissions of the database
7098 if( rc
==SQLITE_OK
&& createConch
){
7100 int err
= osFstat(pFile
->h
, &buf
);
7102 mode_t cmode
= buf
.st_mode
&(S_IRUSR
|S_IWUSR
| S_IRGRP
|S_IWGRP
|
7104 /* try to match the database file R/W permissions, ignore failure */
7105 #ifndef SQLITE_PROXY_DEBUG
7106 osFchmod(conchFile
->h
, cmode
);
7109 rc
= osFchmod(conchFile
->h
, cmode
);
7110 }while( rc
==(-1) && errno
==EINTR
);
7113 fprintf(stderr
, "fchmod %o FAILED with %d %s\n",
7114 cmode
, code
, strerror(code
));
7116 fprintf(stderr
, "fchmod %o SUCCEDED\n",cmode
);
7120 fprintf(stderr
, "STAT FAILED[%d] with %d %s\n",
7121 err
, code
, strerror(code
));
7126 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, SHARED_LOCK
);
7129 OSTRACE(("TRANSPROXY: CLOSE %d\n", pFile
->h
));
7130 if( rc
==SQLITE_OK
&& pFile
->openFlags
){
7133 robust_close(pFile
, pFile
->h
, __LINE__
);
7136 fd
= robust_open(pCtx
->dbPath
, pFile
->openFlags
, 0);
7137 OSTRACE(("TRANSPROXY: OPEN %d\n", fd
));
7141 rc
=SQLITE_CANTOPEN_BKPT
; /* SQLITE_BUSY? proxyTakeConch called
7145 if( rc
==SQLITE_OK
&& !pCtx
->lockProxy
){
7146 char *path
= tempLockPath
? tempLockPath
: pCtx
->lockProxyPath
;
7147 rc
= proxyCreateUnixFile(path
, &pCtx
->lockProxy
, 1);
7148 if( rc
!=SQLITE_OK
&& rc
!=SQLITE_NOMEM
&& tryOldLockPath
){
7149 /* we couldn't create the proxy lock file with the old lock file path
7150 ** so try again via auto-naming
7152 forceNewLockPath
= 1;
7154 continue; /* go back to the do {} while start point, try again */
7157 if( rc
==SQLITE_OK
){
7158 /* Need to make a copy of path if we extracted the value
7159 ** from the conch file or the path was allocated on the stack
7162 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, tempLockPath
);
7163 if( !pCtx
->lockProxyPath
){
7164 rc
= SQLITE_NOMEM_BKPT
;
7168 if( rc
==SQLITE_OK
){
7169 pCtx
->conchHeld
= 1;
7171 if( pCtx
->lockProxy
->pMethod
== &afpIoMethods
){
7172 afpLockingContext
*afpCtx
;
7173 afpCtx
= (afpLockingContext
*)pCtx
->lockProxy
->lockingContext
;
7174 afpCtx
->dbPath
= pCtx
->lockProxyPath
;
7177 conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7179 OSTRACE(("TAKECONCH %d %s\n", conchFile
->h
,
7180 rc
==SQLITE_OK
?"ok":"failed"));
7182 } while (1); /* in case we need to retry the :auto: lock file -
7183 ** we should never get here except via the 'continue' call. */
7188 ** If pFile holds a lock on a conch file, then release that lock.
7190 static int proxyReleaseConch(unixFile
*pFile
){
7191 int rc
= SQLITE_OK
; /* Subroutine return code */
7192 proxyLockingContext
*pCtx
; /* The locking context for the proxy lock */
7193 unixFile
*conchFile
; /* Name of the conch file */
7195 pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7196 conchFile
= pCtx
->conchFile
;
7197 OSTRACE(("RELEASECONCH %d for %s pid=%d\n", conchFile
->h
,
7198 (pCtx
->lockProxyPath
? pCtx
->lockProxyPath
: ":auto:"),
7200 if( pCtx
->conchHeld
>0 ){
7201 rc
= conchFile
->pMethod
->xUnlock((sqlite3_file
*)conchFile
, NO_LOCK
);
7203 pCtx
->conchHeld
= 0;
7204 OSTRACE(("RELEASECONCH %d %s\n", conchFile
->h
,
7205 (rc
==SQLITE_OK
? "ok" : "failed")));
7210 ** Given the name of a database file, compute the name of its conch file.
7211 ** Store the conch filename in memory obtained from sqlite3_malloc64().
7212 ** Make *pConchPath point to the new name. Return SQLITE_OK on success
7213 ** or SQLITE_NOMEM if unable to obtain memory.
7215 ** The caller is responsible for ensuring that the allocated memory
7216 ** space is eventually freed.
7218 ** *pConchPath is set to NULL if a memory allocation error occurs.
7220 static int proxyCreateConchPathname(char *dbPath
, char **pConchPath
){
7221 int i
; /* Loop counter */
7222 int len
= (int)strlen(dbPath
); /* Length of database filename - dbPath */
7223 char *conchPath
; /* buffer in which to construct conch name */
7225 /* Allocate space for the conch filename and initialize the name to
7226 ** the name of the original database file. */
7227 *pConchPath
= conchPath
= (char *)sqlite3_malloc64(len
+ 8);
7229 return SQLITE_NOMEM_BKPT
;
7231 memcpy(conchPath
, dbPath
, len
+1);
7233 /* now insert a "." before the last / character */
7234 for( i
=(len
-1); i
>=0; i
-- ){
7235 if( conchPath
[i
]=='/' ){
7242 conchPath
[i
+1]=dbPath
[i
];
7246 /* append the "-conch" suffix to the file */
7247 memcpy(&conchPath
[i
+1], "-conch", 7);
7248 assert( (int)strlen(conchPath
) == len
+7 );
7254 /* Takes a fully configured proxy locking-style unix file and switches
7255 ** the local lock file path
7257 static int switchLockProxyPath(unixFile
*pFile
, const char *path
) {
7258 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7259 char *oldPath
= pCtx
->lockProxyPath
;
7262 if( pFile
->eFileLock
!=NO_LOCK
){
7266 /* nothing to do if the path is NULL, :auto: or matches the existing path */
7267 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ||
7268 (oldPath
&& !strncmp(oldPath
, path
, MAXPATHLEN
)) ){
7271 unixFile
*lockProxy
= pCtx
->lockProxy
;
7272 pCtx
->lockProxy
=NULL
;
7273 pCtx
->conchHeld
= 0;
7274 if( lockProxy
!=NULL
){
7275 rc
=lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7277 sqlite3_free(lockProxy
);
7279 sqlite3_free(oldPath
);
7280 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, path
);
7287 ** pFile is a file that has been opened by a prior xOpen call. dbPath
7288 ** is a string buffer at least MAXPATHLEN+1 characters in size.
7290 ** This routine find the filename associated with pFile and writes it
7293 static int proxyGetDbPathForUnixFile(unixFile
*pFile
, char *dbPath
){
7294 #if defined(__APPLE__)
7295 if( pFile
->pMethod
== &afpIoMethods
){
7296 /* afp style keeps a reference to the db path in the filePath field
7298 assert( (int)strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7299 strlcpy(dbPath
, ((afpLockingContext
*)pFile
->lockingContext
)->dbPath
,
7303 if( pFile
->pMethod
== &dotlockIoMethods
){
7304 /* dot lock style uses the locking context to store the dot lock
7306 int len
= strlen((char *)pFile
->lockingContext
) - strlen(DOTLOCK_SUFFIX
);
7307 memcpy(dbPath
, (char *)pFile
->lockingContext
, len
+ 1);
7309 /* all other styles use the locking context to store the db file path */
7310 assert( strlen((char*)pFile
->lockingContext
)<=MAXPATHLEN
);
7311 strlcpy(dbPath
, (char *)pFile
->lockingContext
, MAXPATHLEN
);
7317 ** Takes an already filled in unix file and alters it so all file locking
7318 ** will be performed on the local proxy lock file. The following fields
7319 ** are preserved in the locking context so that they can be restored and
7320 ** the unix structure properly cleaned up at close time:
7324 static int proxyTransformUnixFile(unixFile
*pFile
, const char *path
) {
7325 proxyLockingContext
*pCtx
;
7326 char dbPath
[MAXPATHLEN
+1]; /* Name of the database file */
7327 char *lockPath
=NULL
;
7330 if( pFile
->eFileLock
!=NO_LOCK
){
7333 proxyGetDbPathForUnixFile(pFile
, dbPath
);
7334 if( !path
|| path
[0]=='\0' || !strcmp(path
, ":auto:") ){
7337 lockPath
=(char *)path
;
7340 OSTRACE(("TRANSPROXY %d for %s pid=%d\n", pFile
->h
,
7341 (lockPath
? lockPath
: ":auto:"), osGetpid(0)));
7343 pCtx
= sqlite3_malloc64( sizeof(*pCtx
) );
7345 return SQLITE_NOMEM_BKPT
;
7347 memset(pCtx
, 0, sizeof(*pCtx
));
7349 rc
= proxyCreateConchPathname(dbPath
, &pCtx
->conchFilePath
);
7350 if( rc
==SQLITE_OK
){
7351 rc
= proxyCreateUnixFile(pCtx
->conchFilePath
, &pCtx
->conchFile
, 0);
7352 if( rc
==SQLITE_CANTOPEN
&& ((pFile
->openFlags
&O_RDWR
) == 0) ){
7353 /* if (a) the open flags are not O_RDWR, (b) the conch isn't there, and
7354 ** (c) the file system is read-only, then enable no-locking access.
7355 ** Ugh, since O_RDONLY==0x0000 we test for !O_RDWR since unixOpen asserts
7356 ** that openFlags will have only one of O_RDONLY or O_RDWR.
7358 struct statfs fsInfo
;
7359 struct stat conchInfo
;
7362 if( osStat(pCtx
->conchFilePath
, &conchInfo
) == -1 ) {
7364 if( (err
==ENOENT
) && (statfs(dbPath
, &fsInfo
) != -1) ){
7365 goLockless
= (fsInfo
.f_flags
&MNT_RDONLY
) == MNT_RDONLY
;
7369 pCtx
->conchHeld
= -1; /* read only FS/ lockless */
7374 if( rc
==SQLITE_OK
&& lockPath
){
7375 pCtx
->lockProxyPath
= sqlite3DbStrDup(0, lockPath
);
7378 if( rc
==SQLITE_OK
){
7379 pCtx
->dbPath
= sqlite3DbStrDup(0, dbPath
);
7380 if( pCtx
->dbPath
==NULL
){
7381 rc
= SQLITE_NOMEM_BKPT
;
7384 if( rc
==SQLITE_OK
){
7385 /* all memory is allocated, proxys are created and assigned,
7386 ** switch the locking context and pMethod then return.
7388 pCtx
->oldLockingContext
= pFile
->lockingContext
;
7389 pFile
->lockingContext
= pCtx
;
7390 pCtx
->pOldMethod
= pFile
->pMethod
;
7391 pFile
->pMethod
= &proxyIoMethods
;
7393 if( pCtx
->conchFile
){
7394 pCtx
->conchFile
->pMethod
->xClose((sqlite3_file
*)pCtx
->conchFile
);
7395 sqlite3_free(pCtx
->conchFile
);
7397 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7398 sqlite3_free(pCtx
->conchFilePath
);
7401 OSTRACE(("TRANSPROXY %d %s\n", pFile
->h
,
7402 (rc
==SQLITE_OK
? "ok" : "failed")));
7408 ** This routine handles sqlite3_file_control() calls that are specific
7409 ** to proxy locking.
7411 static int proxyFileControl(sqlite3_file
*id
, int op
, void *pArg
){
7413 case SQLITE_FCNTL_GET_LOCKPROXYFILE
: {
7414 unixFile
*pFile
= (unixFile
*)id
;
7415 if( pFile
->pMethod
== &proxyIoMethods
){
7416 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7417 proxyTakeConch(pFile
);
7418 if( pCtx
->lockProxyPath
){
7419 *(const char **)pArg
= pCtx
->lockProxyPath
;
7421 *(const char **)pArg
= ":auto: (not held)";
7424 *(const char **)pArg
= NULL
;
7428 case SQLITE_FCNTL_SET_LOCKPROXYFILE
: {
7429 unixFile
*pFile
= (unixFile
*)id
;
7431 int isProxyStyle
= (pFile
->pMethod
== &proxyIoMethods
);
7432 if( pArg
==NULL
|| (const char *)pArg
==0 ){
7434 /* turn off proxy locking - not supported. If support is added for
7435 ** switching proxy locking mode off then it will need to fail if
7436 ** the journal mode is WAL mode.
7438 rc
= SQLITE_ERROR
/*SQLITE_PROTOCOL? SQLITE_MISUSE?*/;
7440 /* turn off proxy locking - already off - NOOP */
7444 const char *proxyPath
= (const char *)pArg
;
7446 proxyLockingContext
*pCtx
=
7447 (proxyLockingContext
*)pFile
->lockingContext
;
7448 if( !strcmp(pArg
, ":auto:")
7449 || (pCtx
->lockProxyPath
&&
7450 !strncmp(pCtx
->lockProxyPath
, proxyPath
, MAXPATHLEN
))
7454 rc
= switchLockProxyPath(pFile
, proxyPath
);
7457 /* turn on proxy file locking */
7458 rc
= proxyTransformUnixFile(pFile
, proxyPath
);
7464 assert( 0 ); /* The call assures that only valid opcodes are sent */
7468 return SQLITE_ERROR
;
7472 ** Within this division (the proxying locking implementation) the procedures
7473 ** above this point are all utilities. The lock-related methods of the
7474 ** proxy-locking sqlite3_io_method object follow.
7479 ** This routine checks if there is a RESERVED lock held on the specified
7480 ** file by this or any other process. If such a lock is held, set *pResOut
7481 ** to a non-zero value otherwise *pResOut is set to zero. The return value
7482 ** is set to SQLITE_OK unless an I/O error occurs during lock checking.
7484 static int proxyCheckReservedLock(sqlite3_file
*id
, int *pResOut
) {
7485 unixFile
*pFile
= (unixFile
*)id
;
7486 int rc
= proxyTakeConch(pFile
);
7487 if( rc
==SQLITE_OK
){
7488 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7489 if( pCtx
->conchHeld
>0 ){
7490 unixFile
*proxy
= pCtx
->lockProxy
;
7491 return proxy
->pMethod
->xCheckReservedLock((sqlite3_file
*)proxy
, pResOut
);
7492 }else{ /* conchHeld < 0 is lockless */
7500 ** Lock the file with the lock specified by parameter eFileLock - one
7501 ** of the following:
7504 ** (2) RESERVED_LOCK
7506 ** (4) EXCLUSIVE_LOCK
7508 ** Sometimes when requesting one lock state, additional lock states
7509 ** are inserted in between. The locking might fail on one of the later
7510 ** transitions leaving the lock state different from what it started but
7511 ** still short of its goal. The following chart shows the allowed
7512 ** transitions and the inserted intermediate states:
7514 ** UNLOCKED -> SHARED
7515 ** SHARED -> RESERVED
7516 ** SHARED -> (PENDING) -> EXCLUSIVE
7517 ** RESERVED -> (PENDING) -> EXCLUSIVE
7518 ** PENDING -> EXCLUSIVE
7520 ** This routine will only increase a lock. Use the sqlite3OsUnlock()
7521 ** routine to lower a locking level.
7523 static int proxyLock(sqlite3_file
*id
, int eFileLock
) {
7524 unixFile
*pFile
= (unixFile
*)id
;
7525 int rc
= proxyTakeConch(pFile
);
7526 if( rc
==SQLITE_OK
){
7527 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7528 if( pCtx
->conchHeld
>0 ){
7529 unixFile
*proxy
= pCtx
->lockProxy
;
7530 rc
= proxy
->pMethod
->xLock((sqlite3_file
*)proxy
, eFileLock
);
7531 pFile
->eFileLock
= proxy
->eFileLock
;
7533 /* conchHeld < 0 is lockless */
7541 ** Lower the locking level on file descriptor pFile to eFileLock. eFileLock
7542 ** must be either NO_LOCK or SHARED_LOCK.
7544 ** If the locking level of the file descriptor is already at or below
7545 ** the requested locking level, this routine is a no-op.
7547 static int proxyUnlock(sqlite3_file
*id
, int eFileLock
) {
7548 unixFile
*pFile
= (unixFile
*)id
;
7549 int rc
= proxyTakeConch(pFile
);
7550 if( rc
==SQLITE_OK
){
7551 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7552 if( pCtx
->conchHeld
>0 ){
7553 unixFile
*proxy
= pCtx
->lockProxy
;
7554 rc
= proxy
->pMethod
->xUnlock((sqlite3_file
*)proxy
, eFileLock
);
7555 pFile
->eFileLock
= proxy
->eFileLock
;
7557 /* conchHeld < 0 is lockless */
7564 ** Close a file that uses proxy locks.
7566 static int proxyClose(sqlite3_file
*id
) {
7568 unixFile
*pFile
= (unixFile
*)id
;
7569 proxyLockingContext
*pCtx
= (proxyLockingContext
*)pFile
->lockingContext
;
7570 unixFile
*lockProxy
= pCtx
->lockProxy
;
7571 unixFile
*conchFile
= pCtx
->conchFile
;
7575 rc
= lockProxy
->pMethod
->xUnlock((sqlite3_file
*)lockProxy
, NO_LOCK
);
7577 rc
= lockProxy
->pMethod
->xClose((sqlite3_file
*)lockProxy
);
7579 sqlite3_free(lockProxy
);
7580 pCtx
->lockProxy
= 0;
7583 if( pCtx
->conchHeld
){
7584 rc
= proxyReleaseConch(pFile
);
7587 rc
= conchFile
->pMethod
->xClose((sqlite3_file
*)conchFile
);
7589 sqlite3_free(conchFile
);
7591 sqlite3DbFree(0, pCtx
->lockProxyPath
);
7592 sqlite3_free(pCtx
->conchFilePath
);
7593 sqlite3DbFree(0, pCtx
->dbPath
);
7594 /* restore the original locking context and pMethod then close it */
7595 pFile
->lockingContext
= pCtx
->oldLockingContext
;
7596 pFile
->pMethod
= pCtx
->pOldMethod
;
7598 return pFile
->pMethod
->xClose(id
);
7605 #endif /* defined(__APPLE__) && SQLITE_ENABLE_LOCKING_STYLE */
7607 ** The proxy locking style is intended for use with AFP filesystems.
7608 ** And since AFP is only supported on MacOSX, the proxy locking is also
7609 ** restricted to MacOSX.
7612 ******************* End of the proxy lock implementation **********************
7613 ******************************************************************************/
7616 ** Initialize the operating system interface.
7618 ** This routine registers all VFS implementations for unix-like operating
7619 ** systems. This routine, and the sqlite3_os_end() routine that follows,
7620 ** should be the only routines in this file that are visible from other
7623 ** This routine is called once during SQLite initialization and by a
7624 ** single thread. The memory allocation and mutex subsystems have not
7625 ** necessarily been initialized when this routine is called, and so they
7626 ** should not be used.
7628 int sqlite3_os_init(void){
7630 ** The following macro defines an initializer for an sqlite3_vfs object.
7631 ** The name of the VFS is NAME. The pAppData is a pointer to a pointer
7632 ** to the "finder" function. (pAppData is a pointer to a pointer because
7633 ** silly C90 rules prohibit a void* from being cast to a function pointer
7634 ** and so we have to go through the intermediate pointer to avoid problems
7635 ** when compiling with -pedantic-errors on GCC.)
7637 ** The FINDER parameter to this macro is the name of the pointer to the
7638 ** finder-function. The finder-function returns a pointer to the
7639 ** sqlite_io_methods object that implements the desired locking
7640 ** behaviors. See the division above that contains the IOMETHODS
7641 ** macro for addition information on finder-functions.
7643 ** Most finders simply return a pointer to a fixed sqlite3_io_methods
7644 ** object. But the "autolockIoFinder" available on MacOSX does a little
7645 ** more than that; it looks at the filesystem type that hosts the
7646 ** database file and tries to choose an locking method appropriate for
7647 ** that filesystem time.
7649 #define UNIXVFS(VFSNAME, FINDER) { \
7651 sizeof(unixFile), /* szOsFile */ \
7652 MAX_PATHNAME, /* mxPathname */ \
7654 VFSNAME, /* zName */ \
7655 (void*)&FINDER, /* pAppData */ \
7656 unixOpen, /* xOpen */ \
7657 unixDelete, /* xDelete */ \
7658 unixAccess, /* xAccess */ \
7659 unixFullPathname, /* xFullPathname */ \
7660 unixDlOpen, /* xDlOpen */ \
7661 unixDlError, /* xDlError */ \
7662 unixDlSym, /* xDlSym */ \
7663 unixDlClose, /* xDlClose */ \
7664 unixRandomness, /* xRandomness */ \
7665 unixSleep, /* xSleep */ \
7666 unixCurrentTime, /* xCurrentTime */ \
7667 unixGetLastError, /* xGetLastError */ \
7668 unixCurrentTimeInt64, /* xCurrentTimeInt64 */ \
7669 unixSetSystemCall, /* xSetSystemCall */ \
7670 unixGetSystemCall, /* xGetSystemCall */ \
7671 unixNextSystemCall, /* xNextSystemCall */ \
7675 ** All default VFSes for unix are contained in the following array.
7677 ** Note that the sqlite3_vfs.pNext field of the VFS object is modified
7678 ** by the SQLite core when the VFS is registered. So the following
7679 ** array cannot be const.
7681 static sqlite3_vfs aVfs
[] = {
7682 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7683 UNIXVFS("unix", autolockIoFinder
),
7685 UNIXVFS("unix", vxworksIoFinder
),
7687 UNIXVFS("unix", posixIoFinder
),
7689 UNIXVFS("unix-none", nolockIoFinder
),
7690 UNIXVFS("unix-dotfile", dotlockIoFinder
),
7691 UNIXVFS("unix-excl", posixIoFinder
),
7693 UNIXVFS("unix-namedsem", semIoFinder
),
7695 #if SQLITE_ENABLE_LOCKING_STYLE || OS_VXWORKS
7696 UNIXVFS("unix-posix", posixIoFinder
),
7698 #if SQLITE_ENABLE_LOCKING_STYLE
7699 UNIXVFS("unix-flock", flockIoFinder
),
7701 #if SQLITE_ENABLE_LOCKING_STYLE && defined(__APPLE__)
7702 UNIXVFS("unix-afp", afpIoFinder
),
7703 UNIXVFS("unix-nfs", nfsIoFinder
),
7704 UNIXVFS("unix-proxy", proxyIoFinder
),
7707 unsigned int i
; /* Loop counter */
7709 /* Double-check that the aSyscall[] array has been constructed
7710 ** correctly. See ticket [bb3a86e890c8e96ab] */
7711 assert( ArraySize(aSyscall
)==29 );
7713 /* Register all VFSes defined in the aVfs[] array */
7714 for(i
=0; i
<(sizeof(aVfs
)/sizeof(sqlite3_vfs
)); i
++){
7715 sqlite3_vfs_register(&aVfs
[i
], i
==0);
7721 ** Shutdown the operating system interface.
7723 ** Some operating systems might need to do some cleanup in this routine,
7724 ** to release dynamically allocated objects. But not on unix.
7725 ** This routine is a no-op for unix.
7727 int sqlite3_os_end(void){
7731 #endif /* SQLITE_OS_UNIX */