*** DRAFT ***

Moving From SQLite 3.4.2 to 3.5.0

SQLite version 3.5.0 introduces a new OS interface layer that is incompatible with all prior versions of SQLite. In addition, a few existing interfaces have been generalized to work across all database connections within a process rather than just all connections within a thread. The purpose of this article is to describe the changes to 3.5.0 in detail so that users of prior versions of SQLite can judge what, if any, effort will be required to upgrade to newer versions.

1.0 Overview Of Changes

A quick enumeration of the changes in SQLite version 3.5.0 is provide here. Subsequent sections will describe these changes in more detail.

  1. The OS interface layer has been completely reworked:
    1. The undocumented sqlite3_os_switch() interface has been removed.
    2. The SQLITE_ENABLE_REDEF_IO compile-time flag no longer functions. I/O procedures are now always redefinable.
    3. Three new objects are defined for specifying I/O procedures: sqlite3_vfs, sqlite3_file, and sqlite3_io_methods.
    4. Three new interfaces are used to create alternative OS interfaces: sqlite3_vfs_register(), sqlite3_vfs_unregister(), and sqlite3_vfs_find().
    5. A new interface has been added to provided additional control over the creation of new database connections: sqlite3_open_v2(). The legacy interfaces of sqlite3_open() and sqlite3_open16() continue to be fully supported.
  2. The optional shared cache and memory management features that were introduced in version 3.3.0 can now be used across multiple threads within the same process. Formerly, these extensions only applied to database connections operating within a single thread.
    1. The sqlite3_enable_shared_cache() interface now applies to all threads within a process, not to just the one thread in which it was run.
    2. The sqlite3_soft_heap_limit() interface now applies to all threads within a process, not to just the one thread in which it was run.
    3. The sqlite3_release_memory() interface will now attempt to reduce the memory usages across all database connections in all threads, not just connections in the thread where the interface is called.
    4. The sqlite3_thread_cleanup() interface has become a no-op.
  3. Restrictions on the use of the same database connection by multiple threads have been dropped. It is now safe for multiple threads to use the same database connection at the same time.
  4. There is now a compile-time option that allows an application to define alternative malloc()/free() implementations without having to modify any core SQLite code.
  5. There is now a compile-time option that allows an application to define alternative mutex implementations without having to modify any core SQLite code.

Of these changes, only 1a and 2a through 2c are incompatibilities in any formal sense. But users who have previously made custom modifications to the SQLite source (for example to add a custom OS layer for embedded hardware) might find that these changes have a larger impact. On the other hand, an important goal of these changes is to make it much easier to customize SQLite for use on different operating systems.

2.0 The OS Interface Layer

If your system defines a custom OS interface for SQLite or if you were using the undocumented sqlite3_os_switch() interface, then you will need to make modifications in order to upgrade to SQLite version 3.5.0. This may seem painful at first glance. But as you look more closely, you will probably discover that your changes are made smaller and easier to understand and manage by the new SQLite interface. It is likely that your changes will now also work seamlessly with the SQLite amalgamation. You will no longer need to make any changes to the code SQLite source code. All of your changes can be effected by application code and you can link against a standard, unmodified version of the SQLite amalgamation. Furthermore, the OS interface layer, which was formerly undocumented, is now an officially support interface for SQLite. So you have some assurance that this will be a one-time change and that your new backend will continue to work in future versions of SQLite.

2.1 The Virtual File System Object

The new OS interface for SQLite is built around an object named sqlite3_vfs. The "vfs" stands for "Virtual File System". The sqlite3_vfs object is basically a structure containing pointers to functions that implement the primitive disk I/O operations that SQLite needs to perform in order to read and write databases. In this article, we will often refer a sqlite3_vfs objects as a "VFS".

SQLite is able to use multiple VFSes at the same time. Each individual database connection is associated with just one VFS. But if you have multiple database connections, each connection can be associated with a different VFS.

There is always a default VFS. The legacy interfaces sqlite3_open() and sqlite3_open16() always use the default VFS. The new interface for creating database connections, sqlite3_open_v2(), allows you to specify which VFS you want to use by name.

2.1.1 Registering New VFS Objects

Standard builds of SQLite for Unix or Windows come with a single VFS named "unix" or "win32", as appropriate. This one VFS is also the default. So if you are using the legacy open functions, everything will continue to operate as it has before. The change is that an application now has the flexibility of adding new VFS modules to implement a customized OS layer. The sqlite3_vfs_register() API can be used to tell SQLite about one or more application-defined VFS modules:

int sqlite3_vfs_register(sqlite3_vfs*, int makeDflt);

Applications can call sqlite3_vfs_register at any time, though of course a VFS needs to be registered before it can be used. The first argument is a pointer to a customized VFS object that the application has prepared. The second argument is true to make the new VFS the default VFS so that it will be used by the legacy sqlite3_open() and sqlite3_open16() APIs. If the new VFS is not the default, then you will probably have to use the new sqlite3_open_v2() API to use it. Note, however, that if a new VFS is the only VFS known to SQLite (if SQLite was compiled without its usual default VFS or if the precompiled default VFS was removed using sqlite3_vfs_unregister()) then the new VFS automatic becomes the default VFS regardless of the makeDflt argument to sqlite3_vfs_register().

Standard builds include the default "unix" or "win32" VFSes. But if you use the -DOS_OTHER=1 compile-time option, then SQLite is built without a default VFS. In that case, the application must register at least one VFS prior to calling sqlite3_open(). This is the approach that embedded applications should use. Rather than modifying the SQLite source to insert an alternative OS layer as was done in prior releases of SQLite, instead compile an unmodified SQLite source file (preferably the amalgamation) with the -DOS_OTHER=1 option, then invoke sqlite3_vfs_register() to define the interface to the underlying filesystem prior to creating any database connections.

2.1.2 Additional Control Over VFS Objects

The sqlite3_vfs_unregister() API is used to remove an existing VFS from the system.

int sqlite3_vfs_unregister(sqlite3_vfs*);

The sqlite3_vfs_find() API is used to locate a particular VFS by name. Its prototype is as follows:

sqlite3_vfs *sqlite3_vfs_find(const char *zVfsName);

The argument is the symbolic name for the desired VFS. If the argument is a NULL pointer, then the default VFS is returned. The function returns a pointer to the sqlite3_vfs object that implements the VFS. Or it returns a NULL pointer if no object could be found that matched the search criteria.

2.1.3 Modifications Of Existing VFSes

Once a VFS has been registered, it should never be modified. If a change in behavior is required, a new VFS should be registered. The application could, perhaps, use sqlite3_vfs_find() to locate the old VFS, make a copy of the old VFS into a new sqlite3_vfs object, make the desired modifications to the new VFS, unregister the old VFS, the register the new VFS in its place. Existing database connections would continue to use the old VFS even after it is unregistered, but new database connections would use the new VFS.

2.1.4 The VFS Object

A VFS object is an instance of the following structure:

typedef struct sqlite3_vfs sqlite3_vfs;
struct sqlite3_vfs {
  int iVersion;            /* Structure version number */
  int szOsFile;            /* Size of subclassed sqlite3_file */
  int mxPathname;          /* Maximum file pathname length */
  sqlite3_vfs *pNext;      /* Next registered VFS */
  const char *zName;       /* Name of this virtual file system */
  void *pAppData;          /* Pointer to application-specific data */
  int (*xOpen)(sqlite3_vfs*, const char *zName, sqlite3_file*,
               int flags, int *pOutFlags);
  int (*xDelete)(sqlite3_vfs*, const char *zName, int syncDir);
  int (*xAccess)(sqlite3_vfs*, const char *zName, int flags);
  int (*xGetTempName)(sqlite3_vfs*, char *zOut);
  int (*xFullPathname)(sqlite3_vfs*, const char *zName, char *zOut);
  void *(*xDlOpen)(sqlite3_vfs*, const char *zFilename);
  void (*xDlError)(sqlite3_vfs*, int nByte, char *zErrMsg);
  void *(*xDlSym)(sqlite3_vfs*,void*, const char *zSymbol);
  void (*xDlClose)(sqlite3_vfs*, void*);
  int (*xRandomness)(sqlite3_vfs*, int nByte, char *zOut);
  int (*xSleep)(sqlite3_vfs*, int microseconds);
  int (*xCurrentTime)(sqlite3_vfs*, double*);
  /* New fields may be appended in figure versions.  The iVersion
  ** value will increment whenever this happens. */
};

To create a new VFS, an application fills in an instance of this structure with appropriate values and then calls sqlite3_vfs_register().

The iVersion field of sqlite3_vfs should be 1 for SQLite version 3.5.0. This number may increase in future versions of SQLite if we have to modify the VFS object in some way. We hope that this never happens, but the provision is made in case it does.

The szOsFile field is the size in bytes of the structure that defines an open file: the sqlite3_file object. This object will be described more fully below. The point here is that each VFS implementation can define its own sqlite3_file object containing whatever information the VFS implementation needs to store about an open file. SQLite needs to know how big this object is, however, in order to preallocate enough space to hold it.

The mxPathname field is the maximum length of a file pathname that this VFS can use. SQLite sometimes has to preallocate buffers of this size, so it should be as small as reasonably possible. Some filesystems permit huge pathnames, but in practice pathnames rarely extend beyond 100 bytes or so. You do not have to put the longest pathname that the underlying filesystem can handle here. You only have to put the longest pathname that you want SQLite to be able to handle. A few hundred is a good value in most cases.

The pNext field is used internally by SQLite. Specifically, SQLite uses this field to form a linked list of registered VFSes.

The zName field is the symbolic name of the VFS. This is the name that the sqlite3_vfs_find() compares against when it is looking for a VFS.

The pAppData pointer is unused by the SQLite core. The pointer is available to store auxiliary information that a VFS information might want to carry around.

The remaining fields of the sqlite3_vfs object all store pointer to functions that implement primitive operations. We call these "methods". The first methods, xOpen, is used to open files on the underlying storage media. The result is an sqlite3_file object. There are additional methods, defined by the sqlite3_file object itself that are used to read and write and close the file. The additional methods are detailed below. The filename is in UTF-8. SQLite will guarantee that the zFilename string passed to xOpen() is a full pathname as generated by xFullPathname() and that the string will be valid and unchanged until xClose() is called. So the sqlite3_file can store a pointer to the filename if it needs to remember the filename for some reason. The flags argument to xOpen() is a copy of the flags argument to sqlite3_open_v2(). If sqlite3_open() or sqlite3_open16() is used, then flags is SQLITE_OPEN_READWRITE | SQLITE_OPEN_CREATE. If xOpen() opens a file read-only then it sets *pOutFlags to include SQLITE_OPEN_READONLY. Other bits in *pOutFlags may be set. SQLite will also add one of the following flags to the xOpen() call, depending on the object being opened:

The file I/O implementation can use the object type flags to changes the way it deals with files. For example, an application that does not care about crash recovery or rollback, might make the open of a journal file a no-op. Writes to this journal are also a no-op. Any attempt to read the journal returns SQLITE_IOERR. Or the implementation might recognize the a database file will be doing page-aligned sector reads and writes in a random order and set up its I/O subsystem accordingly. SQLite might also add one of the following flags to the xOpen method: The SQLITE_OPEN_DELETEONCLOSE flag means the file should be deleted when it is closed. This will always be set for TEMP databases and journals and for subjournals. The SQLITE_OPEN_EXCLUSIVE flag means the file should be opened for exclusive access. This flag is set for all files except for the main database file. The sqlite3_file structure passed as the third argument to xOpen is allocated by the caller. xOpen just fills it in. The caller allocates a minimum of szOsFile bytes for the sqlite3_file structure.

The differences between an SQLITE_OPEN_TEMP_DB database and an SQLITE_OPEN_TRANSIENT_DB database is this: The SQLITE_OPEN_TEMP_DB is used for explicitly declared and named TEMP tables (using the CREATE TEMP TABLE syntax) or for named tables in a temporary database that is created by opening a database with a filename that is an empty string. An SQLITE_OPEN_TRANSIENT_DB holds a database table that SQLite creates automatically in order to evaluate a subquery or ORDER BY or GROUP BY clause. Both TEMP_DB and TRANSIENT_DB databases are private and are deleted automatically. TEMP_DB databases last for the duration of the database connection. TRANSIENT_DB databases last only for the duration of a single SQL statement.

The xDelete method is used delete a file. The name of the file is given in the second parameter. The filename will be in UTF-8. The VFS must convert the filename into whatever character representation the underlying operating system expects. If the syncDir parameter is true, then the xDelete method should not return until the change to the directory contents for the directory containing the deleted file have been synced to disk in order to ensure that the file does not "reappear" if a power failure occurs soon after.

The xAccess method is used to check for access permissions on a file. The filename will be UTF-8 encoded. The flags argument will be SQLITE_ACCESS_EXISTS to check for the existence of the file, SQLITE_ACCESS_READWRITE to check to see if the file is both readable and writable, or SQLITE_ACCESS_READ to check to see if the file is at least readable. The "file" named by the second parameter might be a directory or folder name.

The xGetTempName method computes the name of a temporary file that SQLite can use. The name should be written into the buffer given by the second parameter. SQLite will size that buffer to hold at least mxPathname bytes. The generated filename should be in UTF-8. To avoid security problems, the generated temporary filename should contain enough randomness to prevent an attacker from guessing the temporary filename in advance.

The xFullPathname method is used to convert a relative pathname into a full pathname. The resulting full pathname is written into the buffer provided by the third parameter. SQLite will size the output buffer to at least mxPathname bytes. Both the input and output names should be in UTF-8.

The xDlOpen, xDlError, xDlSym, and xDlClose methods are all used for accessing shared libraries at run-time. These methods may be omitted (and their pointers set to zero) if the library is compiled with SQLITE_OMIT_LOAD_EXTENSION or if the sqlite3_enable_load_extension() interface is never used to enable dynamic extension loading. The xDlOpen method opens a shared library or DLL and returns a pointer to a handle. NULL is returned if the open fails. If the open fails, the xDlError method can be used to obtain a text error message. The message is written into the zErrMsg buffer of the third parameter which is at least nByte bytes in length. The xDlSym returns a pointer to a symbol in the shared library. The name of the symbol is given by the second parameter. UTF-8 encoding is assumed. If the symbol is not found a NULL pointer is returned. The xDlClose routine closes the shared library.

The xRandomness method is used exactly once to initialize the pseudo-random number generator (PRNG) inside of SQLite. Only the xRandomness method on the default VFS is used. The xRandomness methods on other VFSes are never accessed by SQLite. The xRandomness routine requests that nByte bytes of randomness be written into zOut. The routine returns the actual number of bytes of randomness obtained. The quality of the randomness so obtained will determine the quality of the randomness generated by built-in SQLite functions such as random() and randomblob(). SQLite also uses its PRNG to generate temporary file names.. On some platforms (ex: Windows) SQLite assumes that temporary file names are unique without actually testing for collisions, so it is important to have good-quality randomness even if the random() and randomblob() functions are never used.

The xSleep method is used to suspend the calling thread for at least the number of microseconds given. This method is used to implement the sqlite3_sleep() and sqlite3_busy_timeout() APIs. In the case of sqlite3_sleep() the xSleep method of the default VFS is always used. If the underlying system does not have a microsecond resolution sleep capability, then the sleep time should be rounded up. xSleep returns this rounded-up value.

The xCurrentTime method finds the current time and date and writes the result as double-precision floating point value into pointer provided by the second parameter. The time and date is in coordinated universal time (UTC) and is a fractional Julian day number.

2.1.5 The Open File Object

The result of opening a file is an instance of an sqlite3_file object. The sqlite3_file object is an abstract base class defined as follows:

typedef struct sqlite3_file sqlite3_file;
struct sqlite3_file {
  const struct sqlite3_io_methods *pMethods;
};

Each VFS implementation will subclass the sqlite3_file by adding additional fields at the end to hold whatever information the VFS needs to know about an open file. It does not matter what information is stored as long as the total size of the structure does not exceed the szOsFile value recorded in the sqlite3_vfs object.

The sqlite3_io_methods object is a structure that contains pointers to methods for reading, writing, and otherwise dealing with files. This object is defined as follows:

typedef struct sqlite3_io_methods sqlite3_io_methods;
struct sqlite3_io_methods {
  int iVersion;
  int (*xClose)(sqlite3_file*);
  int (*xRead)(sqlite3_file*, void*, int iAmt, sqlite3_int64 iOfst);
  int (*xWrite)(sqlite3_file*, const void*, int iAmt, sqlite3_int64 iOfst);
  int (*xTruncate)(sqlite3_file*, sqlite3_int64 size);
  int (*xSync)(sqlite3_file*, int flags);
  int (*xFileSize)(sqlite3_file*, sqlite3_int64 *pSize);
  int (*xLock)(sqlite3_file*, int);
  int (*xUnlock)(sqlite3_file*, int);
  int (*xCheckReservedLock)(sqlite3_file*);
  int (*xFileControl)(sqlite3_file*, int op, void *pArg);
  int (*xSectorSize)(sqlite3_file*);
  int (*xDeviceCharacteristics)(sqlite3_file*);
  /* Additional methods may be added in future releases */
};

The iVersion field of sqlite3_io_methods is provided as insurance against future enhancements. The iVersion value should always be 1 for SQLite version 3.5.

The xClose method closes the file. The space for the sqlite3_file structure is deallocated by the caller. But if the sqlite3_file contains pointers to other allocated memory or resources, those allocations should be released by the xClose method.

The xRead method reads iAmt bytes from the file beginning at a byte offset to iOfst. The data read is stored in the pointer of the second parameter. xRead returns the SQLITE_OK on success, SQLITE_IOERR_SHORT_READ if it was not able to read the full number of bytes because it reached end-of-file, or SQLITE_IOERR_READ for any other error.

The xWrite method writes iAmt bytes of data from the second parameter into the file beginning at an offset of iOfst bytes. If the size of the file is less than iOfst bytes prior to the write, then xWrite should ensure that the file is extended with zeros up to iOfst bytes prior to beginning its write. xWrite continues to extends the file as necessary so that the size of the file is at least iAmt+iOfst bytes at the conclusion of the xWrite call. The xWrite method returns SQLITE_OK on success. If the write cannot complete because the underlying storage medium is full, then SQLITE_FULL is returned. SQLITE_IOERR_WRITE should be returned for any other error.

The xTruncate method truncates a file to be nByte bytes in length. If the file is already nByte bytes or less in length then this method is a no-op. The xTruncate method returns SQLITE_OK on success and SQLITE_IOERR_TRUNCATE if anything goes wrong.

The xSync method is used to force previously written data out of operating system cache and into non-volatile memory. The second parameter is usually SQLITE_SYNC_NORMAL. If the second parameter is SQLITE_SYNC_FULL then the xSync method should make sure that data has also been flushed through the disk controllers cache. The SQLITE_SYNC_FULL parameter is the equivalent of the F_FULLSYNC ioctl() on Mac OS X. The xSync method returns SQLITE_OK on success and SQLITE_IOERR_FSYNC if anything goes wrong.

The xFileSize() method determines the current size of the file in bytes and writes that value into *pSize. It returns SQLITE_OK on success and SQLITE_IOERR_FSTAT if something goes wrong.

The xLock and xUnlock methods are used to set and clear file locks. SQLite supports five levels of file locks, in order:

The underlying implementation can support some subset of these locking levels as long as it meets the other requirements of this paragraph. The locking level is specified as the second argument to both xLock and xUnlock. The xLock method increases the locking level to the specified locking level or higher. The xUnlock method decreases the locking level to no lower than the level specified. SQLITE_LOCK_NONE means that the file is unlocked. SQLITE_LOCK_SHARED gives permission to read the file. Multiple database connections can hold SQLITE_LOCK_SHARED at the same time. SQLITE_LOCK_RESERVED is like SQLITE_LOCK_SHARED in that its is permission to read the file. But only a single connection can hold a reserved lock at any point in time. The SQLITE_LOCK_PENDING is also permission to read the file. Other connections can continue to read the file as well, but no other connection is allowed to escalate a lock from none to shared. SQLITE_LOCK_EXCLUSIVE is permission to write on the file. Only a single connection can hold an exclusive lock and no other connection can hold any lock (other than "none") while one connection is hold an exclusive lock. The xLock returns SQLITE_OK on success, SQLITE_BUSY if it is unable to obtain the lock, or SQLITE_IOERR_RDLOCK if something else goes wrong. The xUnlock method returns SQLITE_OK on success and SQLITE_IOERR_UNLOCK for problems.

The xCheckReservedLock method checks to see if another connection or another process is currently holding a reserved, pending, or exclusive lock on the file. It returns true or false.

The xFileControl() method is a generic interface that allows custom VFS implementations to directly control an open file using the (new and experimental) sqlite3_file_control() interface. The second "op" argument is an integer opcode. The third argument is a generic pointer which is intended to be a pointer to a structure that may contain arguments or space in which to write return values. Potential uses for xFileControl() might be functions to enable blocking locks with timeouts, to change the locking strategy (for example to use dot-file locks), to inquire about the status of a lock, or to break stale locks. The SQLite core reserves opcodes less than 100 for its own use. A list of opcodes less than 100 is available. Applications that define a custom xFileControl method should use opcodes greater than 100 to avoid conflicts.

The xSectorSize returns the "sector size" of the underlying non-volatile media. A "sector" is defined as the smallest unit of storage that can be written without disturbing adjacent storage. On a disk drive the "sector size" has until recently been 512 bytes, though there is a push to increase this value to 4KiB. SQLite needs to know the sector size so that it can write a full sector at a time, and thus avoid corrupting adjacent storage space if a power lose occurs in the middle of a write.

The xDeviceCharacteristics method returns an integer bit vector that defines any special properties that the underlying storage medium might have that SQLite can use to increase performance. The allowed return is the bit-wise OR of the following values:

The SQLITE_IOCAP_ATOMIC bit means that all writes to this device are atomic in the sense that either the entire write occurs or none of it occurs. The other SQLITE_IOCAP_ATOMICnnn values indicate that writes of aligned blocks of the indicated size are atomic. SQLITE_IOCAP_SAFE_APPEND means that when extending a file with new data, the new data is written first and then the file size is updated. So if a power failure occurs, there is no chance that the file might have been extended with randomness. The SQLITE_IOCAP_SEQUENTIAL bit means that all writes occur in the order that they are issued and are not reordered by the underlying file system.

2.1.6 Checklist For Constructing A New VFS

The preceding paragraphs contain a lot of information. To ease the task of constructing a new VFS for SQLite we offer the following implementation checklist:

  1. Define an appropriate subclass of the sqlite3_file object.
  2. Implement the methods required by the sqlite3_io_methods object.
  3. Create a static and constant sqlite3_io_methods object containing pointers to the methods from the previous step.
  4. Implement the xOpen method that opens a file and populates an sqlite3_file object, including setting pMethods to point to the sqlite3_io_methods object from the previous step.
  5. Implement the other methods required by sqlite3_vfs.
  6. Define a static (but not constant) sqlite3_vfs structure that contains pointers to the xOpen method and the other methods and which contains the appropriate values for iVersion, szOsFile, mxPathname, zName, and pAppData.
  7. Implement a procedure that calls sqlite3_vfs_register() and passes it a pointer to the sqlite3_vfs structure from the previous step. This procedure is probably the only exported symbol in the source file that implements your VFS.

Within your application, call the procedure implemented in the last step above as part of your initialization process before any database connections are opened.

3.0 The Memory Allocation Subsystem

Beginning with version 3.5, SQLite obtains all of the heap memory it needs using the routines sqlite3_malloc(), sqlite3_free(), and sqlite3_realloc(). These routines have existed in prior versions of SQLite, but SQLite has previously bypassed these routines and used its own memory allocator. This all changes in version 3.5.0.

The SQLite source tree actually contains multiple versions of the memory allocator. The default high-speed version found in the "mem1.c" source file is used for most builds. But if the SQLITE_MEMDEBUG flag is enabled, a separate memory allocator the "mem2.c" source file is used instead. The mem2.c allocator implements lots of hooks to do error checking and to simulate memory allocation failures for testing purposes. Both of these allocators use the malloc()/free() implementation in the standard C library.

Applications are not required to use either of these standard memory allocators. If SQLite is compiled with SQLITE_OMIT_MEMORY_ALLOCATION then no implementation for the sqlite3_malloc(), sqlite3_realloc(), and sqlite3_free() functions is provided. Instead, the application that links against SQLite must provide its own implementation of these functions. The application provided memory allocator is not required to use the malloc()/free() implementation in the standard C library. An embedded application might provide an alternative memory allocator that uses memory for a fixed memory pool set aside for the exclusive use of SQLite, for example.

Applications that implement their own memory allocator must provide implementation for the usual three allocation functions sqlite3_malloc(), sqlite3_realloc(), and sqlite3_free(). And they must also implement a fourth function:

int sqlite3_memory_alarm(
  void(*xCallback)(void *pArg, sqlite3_int64 used, int N),
  void *pArg,
  sqlite3_int64 iThreshold
);

The sqlite3_memory_alarm routine is used to register a callback on memory allocation events. This routine registers or clears a callbacks that fires when the amount of memory allocated exceeds iThreshold. Only a single callback can be registered at a time. Each call to sqlite3_memory_alarm() overwrites the previous callback. The callback is disabled by setting xCallback to a NULL pointer.

The parameters to the callback are the pArg value, the amount of memory currently in use, and the size of the allocation that provoked the callback. The callback will presumably invoke sqlite3_free() to free up memory space. The callback may invoke sqlite3_malloc() or sqlite3_realloc() but if it does, no additional callbacks will be invoked by the recursive calls.

The sqlite3_soft_heap_limit() interface works by registering a memory alarm at the soft heap limit and invoking sqlite3_release_memory() in the alarm callback. Application programs should not attempt to use the sqlite3_memory_alarm() interface because doing so will interfere with the sqlite3_soft_heap_limit() module. This interface is exposed only so that applications can provide their own alternative implementation when the SQLite core is compiled with SQLITE_OMIT_MEMORY_ALLOCATION.

The built-in memory allocators in SQLite also provide the following additional interfaces:

sqlite3_int64 sqlite3_memory_used(void);
sqlite3_int64 sqlite3_memory_highwater(int resetFlag);

These interfaces can be used by an application to monitor how much memory SQLite is using. The sqlite3_memory_used() routine returns the number of bytes of memory currently in use and the sqlite3_memory_highwater() returns the maximum instantaneous memory usage. Neither routine includes the overhead associated with the memory allocator. These routines are provided for use by the application. SQLite never invokes them itself. So if the application is providing its own memory allocation subsystem, it can omit these interfaces if desired.

4.0 The Mutex Subsystem

SQLite has always been threadsafe in the sense that it is safe to use different SQLite database connections in different threads at the same time. The constraint was that the same database connection could not be used in two separate threads at once. SQLite version 3.5.0 relaxes this constraint.

In order to allow multiple threads to use the same database connection at the same time, SQLite must make extensive use of mutexes. And for this reason a new mutex subsystem as been added. The mutex subsystem as the following interface:

sqlite3_mutex *sqlite3_mutex_alloc(int);
void sqlite3_mutex_free(sqlite3_mutex*);
void sqlite3_mutex_enter(sqlite3_mutex*);
int sqlite3_mutex_try(sqlite3_mutex*);
void sqlite3_mutex_leave(sqlite3_mutex*);

Though these routines exist for the use of the SQLite core, application code is free to use these routines as well, if desired. A mutex is an sqlite3_mutex object. The sqlite3_mutex_alloc() routine allocates a new mutex object and returns a pointer to it. The argument to sqlite3_mutex_alloc() should be SQLITE_MUTEX_FAST or SQLITE_MUTEX_RECURSIVE for non-recursive and recursive mutexes, respectively. If the underlying system does not provide non-recursive mutexes, then a recursive mutex can be substituted in that case. The argument to sqlite3_mutex_alloc() can also be a constant designating one of several static mutexes:

These static mutexes are reserved for use internally by SQLite and should not be used by the application. The static mutexes are all non-recursive.

The sqlite3_mutex_free() routine should be used to deallocate a non-static mutex. If a static mutex is passed to this routine then the behavior is undefined.

The sqlite3_mutex_enter() attempts to enter the mutex and blocks if another threads is already there. sqlite3_mutex_try() attempts to enter and returns SQLITE_OK on success or SQLITE_BUSY if another thread is already there. sqlite3_mutex_leave() exits a mutex. The mutex is held until the number of exits matches the number of entrances. If sqlite3_mutex_leave() is called on a mutex that the thread is not currently holding, then the behavior is undefined. If any routine is called for a deallocated mutex, then the behavior is undefined.

The SQLite source code provides multiple implementations of these APIs, suitable for varying environments. If SQLite is compiled with the SQLITE_THREADSAFE=0 flag then a no-op mutex implementation that is fast but does no real mutual exclusion is provided. That implementation is suitable for use in single-threaded applications or applications that only use SQLite in a single thread. Other real mutex implementations are provided based on the underlying operating system.

Embedded applications may wish to provide their own mutex implementation. If SQLite is compiled with the -DSQLITE_MUTEX_APPDEF=1 compile-time flag then the SQLite core provides no mutex subsystem and a mutex subsystem that matches the interface described above must be provided by the application that links against SQLite.

5.0 Other Interface Changes

Version 3.5.0 of SQLite changes the behavior of a few APIs in ways that are technically incompatible. However, these APIs are seldom used and even when they are used it is difficult to imagine a scenario where the change might break something. The changes actually makes these interface much more useful and powerful.

Prior to version 3.5.0, the sqlite3_enable_shared_cache() API would enable and disable the shared cache feature for all connections within a single thread - the same thread from which the sqlite3_enable_shared_cache() routine was called. Database connections that used the shared cache were restricted to running in the same thread in which they were opened. Beginning with version 3.5.0, the sqlite3_enable_shared_cache() applies to all database connections in all threads within the process. Now database connections running in separate threads can share a cache. And database connections that use shared cache can migrate from one thread to another.

Prior to version 3.5.0 the sqlite3_soft_heap_limit() set an upper bound on heap memory usage for all database connections within a single thread. Each thread could have its own heap limit. Beginning in version 3.5.0, there is a single heap limit for the entire process. This seems more restrictive (one limit as opposed to many) but in practice it is what most users want.

Prior to version 3.5.0 the sqlite3_release_memory() function would try to reclaim memory from all database connections in the same thread as the sqlite3_release_memory() call. Beginning with version 3.5.0, the sqlite3_release_memory() function will attempt to reclaim memory from all database connections in all threads.

6.0 Summary

The transition from SQLite version 3.4.2 to 3.5.0 is a major change. Every source code file in the SQLite core had to be modified, some extensively. And the change introduced some minor incompatibilities in the C interface. But we feel that the benefits of the transition from 3.4.2 to 3.5.0 far outweigh the pain of porting. The new VFS layer is now well-defined and stable and should simplify future customizations. The VFS layer, and the separable memory allocator and mutex subsystems allow a standard SQLite source code amalgamation to be used in an embedded project without change, greatly simplifying configuration management. And the resulting system is much more tolerant of highly threaded designs.

*** DRAFT ***