The default mechanism by which SQLite accesses and updates database disk files is the xRead() and xWrite() methods of the sqlite3_io_methods VFS object. These methods are typically implemented as "read()" and "write()" system calls which cause the operating system to copy disk content between the kernel buffer cache and user space.
There are advantages and disadvantages to using memory-mapped I/O. Advantages include:
Many operations, especially I/O intensive operations, can be much faster since content does need to be copied between kernel space and user space. In some cases, performance can nearly double.
The SQLite library may need less RAM since it shares pages with the operating-system page cache and does not always need its own copy of working pages.
But there are also disadvantages:
A stray pointer or buffer overflow in the application program might change the content of mapped memory, potentially corrupting the database file. Bugs in application program that overwrite memory assigned to the SQLite library could, in theory, also cause corruption in read/write mode, but that would require that the overwrite be followed by a call to xWrite() and in practice the buggy application usually errors out or crashes before that point. With memory-mapped I/O, the database corruption occurs immediately and is thus more of a risk.
An I/O error on a memory-mapped file cannot be caught and dealt with by SQLite. Instead, the I/O error causes a signal which, if not caught by the application, results in a program crash.
The operating system must have a unified buffer cache in order for the memory-mapped I/O extension to work correctly, especially in situations where two processes are accessing the same database file and one process is using memory-mapped I/O while the other is not. Not all operating systems have a unified buffer cache. In some operating systems that claim to have a unified buffer cache, the implementation is buggy and can lead to corrupt databases.
Performance does not always increase with memory-mapped I/O. In fact, it is possible to construct test cases where performance is reduced by the use of memory-mapped I/O, though this is hard to do.
Windows is unable to truncate a memory-mapped file. Hence, on Windows, if an operation such as VACUUM or auto_vacuum tries to reduce the size of a memory-mapped database file, the size reduction attempt will silently fail, leaving unused space at the end of the database file. No data is lost due to this problem, and the unused space will be reused again the next time the database grows. However if a version of SQLite prior to 3.7.0 runs PRAGMA integrity_check on such a database, it will (incorrectly) report database corruption due to the unused space at the end. Or if a version of SQLite prior to 3.7.0 writes to the database while it still has unused space at the end, it may make that unused space inaccessible and unavailable for reuse until after the next VACUUM.
Because of the potential disadvantages, memory-mapped I/O is turned off by default. To activate memory-mapped I/O, use the mmap_size pragma and set the mmap_size to some large number, usually 256MB or larger, depending on how much address space your application can spare. The rest is automatic. The PRAGMA mmap_size statement will be a silent no-op on systems that do not support memory-mapped I/O.
To read a page of database content using the legacy xRead() method, SQLite first allocates a page-size chunk of heap memory then invokes the xRead() method which causes the database page content to be copied into the newly allocated heap memory. This involves (at a minimum) a copy of the entire page.
But if SQLite wants to access a page of the database file and memory mapped I/O is enabled, it first calls the xFetch() method. The xFetch() method asks the operating system to return a pointer to the requested page, if possible. If the requested page has been or can be mapped into the application address space, then xFetch returns a pointer to that page for SQLite to use without having to copy anything. Skipping the copy step is what makes memory mapped I/O faster.
SQLite does not assume that the xFetch() method will work. If a call to xFetch() returns a NULL pointer (indicating that the requested page is not currently mapped into the applications address space) then SQLite silently falls back to using xRead(). An error is only reported if xRead() also fails.
When updating the database file, SQLite always makes a copy of the page content into heap memory before modifying the page. This is necessary since the changes are not supposed to be visible to other processes until after the transaction commits and so the changes must occur in private memory. After all needed changes are completed, xWrite() is used to move the content back into the database file. The current xWrite() implementations for both unix and windows check to see if section of the file being written is mapped into the applications address space, and if it is the write operation is implemented using memcpy() rather than invoking a "write()" system call, but that is just an implementation detail. A memory copy occurs either way. So the use of memory mapped I/O does not significantly change the performance of database changes. Memory mapped I/O is mostly a benefit for queries.
The "mmap_size" is the maximum number of bytes of the database file that SQLite will try to map into the process address space at one time. The mmap_size applies separately to each database file, so the total amount of process address space that could potentially be used is the mmap_size times the number of open database files.
To activate memory-mapped I/O, an application can set the mmap_size to some large value. For example:
To disable memory-mapped I/O, simply set the mmap_size to zero:
If mmap_size is set to N then all current implementations map the first N bytes of the database file and use legacy xRead() calls for any content beyond N bytes. If the database file is smaller than N bytes, then the entire file is mapped. In the future, new OS interfaces could, in theory, map regions of the file other than the first N bytes, but no such implementation currently exists.
The mmap_size is set separately for each database file using the "PRAGMA mmap_size" statement. The usual default mmap_size is zero, meaning that memory mapped I/O is disabled by default. However, the default mmap_size can be increased either at compile-time using the SQLITE_DEFAULT_MMAP_SIZE macro or at start-time using the sqlite3_config(SQLITE_CONFIG_MMAP_SIZE,...) interface.
SQLite also maintains a hard upper bound on the mmap_size. Attempts to increase the mmap_size above this hard upper bound (using PRAGMA mmap_size) will automatically cap the mmap_size at the hard upper bound. If the hard upper bound is zero, then memory mapped I/O is impossible. The hard upper bound can be set at compile-time using the SQLITE_MAX_MMAP_SIZE macro. If SQLITE_MAX_MMAP_SIZE is set to zero, then the code used to implement memory mapped I/O is omitted from the build. The hard upper bound is automatically set to zero on certain platforms (ex: OpenBSD) where memory mapped I/O does not work due to the lack of a unified buffer cache.
If the hard upper bound on mmap_size is non-zero at compilation time, it may still be reduced or zeroed at start-time using the sqlite3_config(SQLITE_CONFIG_MMAP_SIZE,X,Y) interface. The X and Y parameters must both be 64-bit signed integers. The X parameter is the default mmap_size of the process and the Y is the new hard upper bound. The hard upper bound cannot be increased above its compile-time setting using SQLITE_CONFIG_MMAP_SIZE but it can be reduced or zeroed.
*** DRAFT ***