I did benchmarks on OS X (10.5) between asynchronous IO, mmap, stdio, and normal unix open/read (disabling caching) — it all came down to about the same after I had fiddled with page size, prefetch size, etc. to optimize each.
This of course depends on your OS (and version). My advice would be to pick API based on “semantics” (or need for portability). If you need to read through a file, use the API intended for that because mmap will not be optimized for that access pattern, but stdio will.
Mostly because mmap is not portable. If you're writing a Linux specific version of your code then mmap is fine. If you foresee running your code on other platforms then you shouldn't use it.
Basically, there's a reason why it's called "standard" IO. Generally you shouldn't use system calls unless you are absolutely, 100% sure that your code will only ever need to work on one system.
I've worked on a project that successfully uses memory mapping on all the major OSes.
mmap is (allegedly) portable across Unixes:
http://opengroup.org/onlinepubs/007908775/xsh/mmap.html
and you can do the same thing (with a few more hoops) on Windows.
Many people may not care, but mmap() often doesn't exist on hardware without a MMU.
I usually just use standard IO, unless I have a large file that might be randomly accessed by multiple processes. In that case I'll use a mmap. But, I don't code in C very much anymore, so I was wondering how much of a difference there really is.