Linux also counts files on tmpfs as part of the cache number--it's really annoying that no number in top can serve to show you a decent approximation of how many page-size malloc()s you can do before OOM (with overhead and fragmentation and so on, of course it's infeasible to get an exact number).
It's worse. No line of /proc/meminfo has that value, and since not all of the cache can be instantly discarded (I didn't know that), I don't see any combination of them that would help.
That sentence:
> There are no downsides, except for confusing newbies.
Just isn't true. There is the downside that we can't know how much memory is available.
Yeah, and obviously some (though not all) forms of shared memory are actually implemented using tmpfs behind the scenes which makes life even more interesting. I think the problem I encountered involved SysV shared memory which isn't backed by tmpfs and does show up in /proc/meminfo, whereas other types may not.
