Next: , Previous: Scatter-Gather, Up: Low-Level I/O

13.7 Memory-mapped I/O

On modern operating systems, it is possible to mmap (pronounced “em-map”) a file to a region of memory. When this is done, the file can be accessed just like an array in the program.

This is more efficient than read or write, as only the regions of the file that a program actually accesses are loaded. Accesses to not-yet-loaded parts of the mmapped region are handled in the same way as swapped out pages.

Since mmapped pages can be stored back to their file when physical memory is low, it is possible to mmap files orders of magnitude larger than both the physical memory and swap space. The only limit is address space. The theoretical limit is 4GB on a 32-bit machine - however, the actual limit will be smaller since some areas will be reserved for other purposes. If the LFS interface is used the file size on 32-bit systems is not limited to 2GB (offsets are signed which reduces the addressable area of 4GB by half); the full 64-bit are available.

Memory mapping only works on entire pages of memory. Thus, addresses for mapping must be page-aligned, and length values will be rounded up. To determine the size of a page the machine uses one should use 

     size_t page_size = (size_t) sysconf (_SC_PAGESIZE);

These functions are declared in sys/mman.h.

— Function: void * mmap (void *address, size_t length, int protect, int flags, int filedes, off_t offset)

The mmap function creates a new mapping, connected to bytes (offset) to (offset + length - 1) in the file open on filedes. A new reference for the file specified by filedes is created, which is not removed by closing the file.

address gives a preferred starting address for the mapping. NULL expresses no preference. Any previous mapping at that address is automatically removed. The address you give may still be changed, unless you use the MAP_FIXED flag.

protect contains flags that control what kind of access is permitted. They include PROT_READ, PROT_WRITE, and PROT_EXEC, which permit reading, writing, and execution, respectively. Inappropriate access will cause a segfault (see Program Error Signals).

Note that most hardware designs cannot support write permission without read permission, and many do not distinguish read and execute permission. Thus, you may receive wider permissions than you ask for, and mappings of write-only files may be denied even if you do not use PROT_READ.

flags contains flags that control the nature of the map. One of MAP_SHARED or MAP_PRIVATE must be specified.

They include:

MAP_PRIVATE
This specifies that writes to the region should never be written back to the attached file. Instead, a copy is made for the process, and the region will be swapped normally if memory runs low. No other process will see the changes.

Since private mappings effectively revert to ordinary memory when written to, you must have enough virtual memory for a copy of the entire mmapped region if you use this mode with PROT_WRITE.

MAP_SHARED
This specifies that writes to the region will be written back to the file. Changes made will be shared immediately with other processes mmaping the same file.

Note that actual writing may take place at any time. You need to use msync, described below, if it is important that other processes using conventional I/O get a consistent view of the file.

MAP_FIXED
This forces the system to use the exact mapping address specified in address and fail if it can't.
MAP_ANONYMOUS
MAP_ANON
This flag tells the system to create an anonymous mapping, not connected to a file. filedes and off are ignored, and the region is initialized with zeros.

Anonymous maps are used as the basic primitive to extend the heap on some systems. They are also useful to share data between multiple tasks without creating a file.

On some systems using private anonymous mmaps is more efficient than using malloc for large blocks. This is not an issue with the GNU C Library, as the included malloc automatically uses mmap where appropriate.

mmap returns the address of the new mapping, or -1 for an error.

Possible errors include:

EINVAL
Either address was unusable, or inconsistent flags were given.
EACCES
filedes was not open for the type of access specified in protect.
ENOMEM
Either there is not enough memory for the operation, or the process is out of address space.
ENODEV
This file is of a type that doesn't support mapping.
ENOEXEC
The file is on a filesystem that doesn't support mapping.
— Function: void * mmap64 (void *address, size_t length, int protect, int flags, int filedes, off64_t offset)

The mmap64 function is equivalent to the mmap function but the offset parameter is of type off64_t. On 32-bit systems this allows the file associated with the filedes descriptor to be larger than 2GB. filedes must be a descriptor returned from a call to open64 or fopen64 and freopen64 where the descriptor is retrieved with fileno.

When the sources are translated with _FILE_OFFSET_BITS == 64 this function is actually available under the name mmap. I.e., the new, extended API using 64 bit file sizes and offsets transparently replaces the old API.

— Function: int munmap (void *addr, size_t length)

munmap removes any memory maps from (addr) to (addr + length). length should be the length of the mapping.

It is safe to unmap multiple mappings in one command, or include unmapped space in the range. It is also possible to unmap only part of an existing mapping. However, only entire pages can be removed. If length is not an even number of pages, it will be rounded up.

It returns 0 for success and -1 for an error.

One error is possible:

EINVAL
The memory range given was outside the user mmap range or wasn't page aligned.
— Function: int msync (void *address, size_t length, int flags)

When using shared mappings, the kernel can write the file at any time before the mapping is removed. To be certain data has actually been written to the file and will be accessible to non-memory-mapped I/O, it is necessary to use this function.

It operates on the region address to (address + length). It may be used on part of a mapping or multiple mappings, however the region given should not contain any unmapped space.

flags can contain some options:

MS_SYNC
This flag makes sure the data is actually written to disk. Normally msync only makes sure that accesses to a file with conventional I/O reflect the recent changes.
MS_ASYNC
This tells msync to begin the synchronization, but not to wait for it to complete.

msync returns 0 for success and -1 for error. Errors include:

EINVAL
An invalid region was given, or the flags were invalid.
EFAULT
There is no existing mapping in at least part of the given region.
— Function: void * mremap (void *address, size_t length, size_t new_length, int flag)

This function can be used to change the size of an existing memory area. address and length must cover a region entirely mapped in the same mmap statement. A new mapping with the same characteristics will be returned with the length new_length.

One option is possible, MREMAP_MAYMOVE. If it is given in flags, the system may remove the existing mapping and create a new one of the desired length in another location.

The address of the resulting mapping is returned, or -1. Possible error codes include:

EFAULT
There is no existing mapping in at least part of the original region, or the region covers two or more distinct mappings.
EINVAL
The address given is misaligned or inappropriate.
EAGAIN
The region has pages locked, and if extended it would exceed the process's resource limit for locked pages. See Limits on Resources.
ENOMEM
The region is private writable, and insufficient virtual memory is available to extend it. Also, this error will occur if MREMAP_MAYMOVE is not given and the extension would collide with another mapped region.

This function is only available on a few systems. Except for performing optional optimizations one should not rely on this function.

Not all file descriptors may be mapped. Sockets, pipes, and most devices only allow sequential access and do not fit into the mapping abstraction. In addition, some regular files may not be mmapable, and older kernels may not support mapping at all. Thus, programs using mmap should have a fallback method to use should it fail. See Mmap.

— Function: int madvise (void *addr, size_t length, int advice)

This function can be used to provide the system with advice about the intended usage patterns of the memory region starting at addr and extending length bytes.

The valid BSD values for advice are:

MADV_NORMAL
The region should receive no further special treatment.
MADV_RANDOM
The region will be accessed via random page references. The kernel should page-in the minimal number of pages for each page fault.
MADV_SEQUENTIAL
The region will be accessed via sequential page references. This may cause the kernel to aggressively read-ahead, expecting further sequential references after any page fault within this region.
MADV_WILLNEED
The region will be needed. The pages within this region may be pre-faulted in by the kernel.
MADV_DONTNEED
The region is no longer needed. The kernel may free these pages, causing any changes to the pages to be lost, as well as swapped out pages to be discarded.

The POSIX names are slightly different, but with the same meanings:

POSIX_MADV_NORMAL
This corresponds with BSD's MADV_NORMAL.
POSIX_MADV_RANDOM
This corresponds with BSD's MADV_RANDOM.
POSIX_MADV_SEQUENTIAL
This corresponds with BSD's MADV_SEQUENTIAL.
POSIX_MADV_WILLNEED
This corresponds with BSD's MADV_WILLNEED.
POSIX_MADV_DONTNEED
This corresponds with BSD's MADV_DONTNEED.

madvise returns 0 for success and -1 for error. Errors include:

EINVAL
An invalid region was given, or the advice was invalid.
EFAULT
There is no existing mapping in at least part of the given region.