The Linux Programming Interface

1026 Chapter 49

Figure 49-2: Two processes with a shared mapping of the same region of a file

Memory-mapped I/O

Since the contents of the shared file mapping are initialized from the file, and any

modifications to the contents of the mapping are automatically carried through to

the file, we can perform file I/O simply by accessing bytes of memory, relying on the

kernel to ensure that the changes to memory are propagated to the mapped file.

(Typically, a program would define a structured data type that corresponds to the

contents of the disk file, and then use that data type to cast the contents of the

mapping.) This technique is referred to as memory-mapped I/O, and is an alternative

to using read() and write() to access the contents of a file.

Memory-mapped I/O has two potential advantages:

z By replacing read() and write() system calls with memory accesses, it can simplify

the logic of some applications.

z It can, in some circumstances, provide better performance than file I/O carried

out using the conventional I/O system calls.

The reasons that memory-mapped I/O can provide performance benefits are as

follows:

z A normal read() or write() involves two transfers: one between the file and the

kernel buffer cache, and the other between the buffer cache and a user-space

buffer. Using mmap() eliminates the second of these transfers. For input, the

data is available to the user process as soon as the kernel has mapped the

Process A

page table

PT entries

for mapped

region

Process B

page table

PT entries

for mapped

region

Mapped

pages

Physical

memory

Mapped

region of

file

Open file

I/O managed

by kernel