Linux System Programming, 2nd Edition

Chapter 3. Buffered I/O

Recall from Chapter 1 that the block is an abstraction representing the smallest unit of storage on a filesystem. Inside the kernel, all filesystem operations occur in terms of blocks. Indeed, the block is the lingua franca of I/O. Consequently, no I/O operation may execute on an amount of data less than the block size or that is not an integer multiple of the block size. If you only want to read a byte, too bad: you’ll have to read a whole block. Want to write 4.5 blocks worth of data? You’ll need to write 5 blocks, which implies reading that partial block in its entirety, updating only the half you’ve changed, and then writing back the whole block.

You can see where this is leading: partial block operations are inefficient. The operating system has to “fix up” your I/O by ensuring that everything occurs on block-aligned boundaries and rounding up to the next largest block. Unfortunately, this is not how user-space applications are generally written. Most applications operate in terms of higher-level abstractions, such as fields and strings, whose size varies independently of the block size. At its worst, a user-space application might read and write but a single byte at a time! That’s a lot of waste. Each of those one-byte writes is actually writing a whole block.

The situation is exacerbated by the extraneous system calls required to read, say, a single byte 1,024 times rather than a single 1,024-byte block all at once. The solution to this pathological ...