We have observed slow performance on mmap read on some applications.

The problem is if access pattern is neither sequential nor stride reading, but
still adjacent in a small range and then seek a random position.

So the pattern could be something like this:

[1M data] [hole..] [0.5M data] [hole]......[0.7M data]......[1M data]

So every time application want to access some data, the data is not only 4K but
cosed to let's say 1M range, so if we could predict this kind of behavior and next time we hit page miss, we use last time range for prefetching.