Advanced Search

Log In | Not a Member? Contact ADC

Memory Mapping and Block Copying

Memory mapping is one of the more common means of communicating between two applications or between an application and the kernel. While occasionally used by itself, it is usually used in conjunction with one of the other means of boundary crossing.

One way of using memory mapping is known as shared memory. In this form, one or more pages of memory are mapped into the address space of two processes. Either process can then access or modify the data stored in those shared pages. This is useful when moving large quantities of data between processes, as it allows direct communication without multiple user-kernel boundary crossings. Thus, when moving large amounts of data between processes, this is preferable to traditional message passing.

The same holds true with memory mapping between an application and the kernel. The BSD sysctl and syscall interfaces (and to an extent, Mach IPC) were designed to transfer small units of data of known size, such as an array of four integers. In this regard, they are much like a traditional C function call. If you need to pass a large amount of data to a function in C, you should pass a pointer. This is also true when passing data between an application and the kernel, with the addition of memory mapping or copying to allow that pointer to be dereferenced in the kernel.

There are a number of limitations to the way that memory mapping can be used to exchange data between an application and the kernel. For one, memory allocated in the kernel cannot be written to by applications, including those running as root (unless the kernel is running in an insecure mode, such as single user mode). For this reason, if a buffer must be modified by an application, the buffer must be allocated by that program, not by the kernel.

When you use memory mapping for passing data to the kernel, the application allocates a block of memory and fills it with data. It then performs a system call that passes the address to the appropriate function in kernel space. It should be noted, however, that the address being passed is a virtual address, not a physical address, and more importantly, it is relative to the address space of the program, which is not the same as the address space of the kernel.

Since the address is a user-space virtual address, the kernel must call special functions to copy the block of memory into a kernel buffer or to map the block of memory into the kernel’s address space.

In the Mac OS X kernel, data is most easily copied into kernel space with the BSD copyin function, and back out to user space with the copyout function. For large blocks of data, entire pages will be memory mapped using copy-on-write. For this reason, it is generally not useful to do memory mapping by hand.

Getting data from the kernel to an application can be done in a number of ways. The most common method is the reverse of the above, in which the application passes in a buffer pointer, the kernel scribbles on a chunk of data, uses copyout to copy the buffer data into the address space of the application, and returns KERN_SUCCESS. Note that this is really using the buffer allocated in the application, even though the physical memory may have actually been allocated by the kernel. Assuming the kernel frees its reference to the buffer, no memory is wasted.

A special case of memory mapping occurs when doing I/O to a device from user space. Since I/O operations can, in some cases, be performed by DMA hardware that operates based on physical addressing, it is vital that the memory associated with I/O buffers not be paged out while the hardware is copying data to or from the buffer.

For this reason, when a driver or other kernel entity needs a buffer for I/O, it must take steps to mark it as not pageable. This step is referred to as wiring the pages in memory.

Wiring pages into memory can also be helpful where high bandwidth, low latency communication is desired, as it prevents shared buffers from being paged out to disk. In general, however, this sort of workaround should be unnecessary, and is considered to be bad programming practice.

Pages can be wired in two ways. When a memory region is allocated, it may be allocated in a nonpageable fashion. The details of allocating memory for I/O differ, depending on what part of the kernel you are modifying. This is described in more detail in the appropriate sections of this document, or in the case of the I/O Kit, in the API reference documentation (available from the developer section of Apple’s web site). Alternately, individual pages may be wired after allocation.

The recommended way to do this is through a call to vm_wire in BSD parts of the kernel, with mlock from applications (but only by processes running as root), or with IOMemoryDescriptor::prepare in the I/O Kit. Because this can fail for a number of reasons, it is particularly crucial to check return values when wiring memory. The vm_wire call and other virtual memory topics are discussed in more detail in “Memory and Virtual Memory”. The IOMemoryDescriptor class is described in more detail in the I/O Kit API reference available from the developer section of Apple’s web site.

Did this document help you? Yes: Tell us what works for you. It’s good, but: Report typos, inaccuracies, and so forth. It wasn’t helpful: Tell us what would have helped.

	Get information on Apple products. Visit the Apple Store online or at retail locations. 1-800-MY-APPLE Copyright © 2007 Apple Inc. All rights reserved. | Terms of use | Privacy Notice