> I recall a discussion some time ago about heavy use of mmap()
> actually _decreasing_ performace of reads from mmap()ped files.

A while ago, NetBSD had a bug that meant that reads and writes via
mmap() and accessing the memory were decoupled from reads and writes
via read() and write().  That is, if you wrote to a mapped memory
section and then later used read(), you might not see the writes as
having been done; I assume something similar could happen the other
way, but I don't recall seeing it tested.  (This is the "unified buffer
cache" issue - the buffer cache used by read()/write() was a rather
ugly hack and it didn't cooperate nicely with the VM subsystem used by
mmap()ed accesses.)

I don't know whether we still have this bug.  I would hope that at the
very least UVM would fix it, but as I haven't tested for its presence
under any kernel less than a few months old I can't say about either
UVM or -current old VM.

					der Mouse

			       mouse@rodents.montreal.qc.ca
		     7D C8 61 52 5D E7 2D 39  4E F1 31 3E E8 B3 27 4B