Dynamic reads without locking.

[Tim Kientzle]

Bruce M Simpson wrote:
> On Fri, Oct 10, 2003 at 10:31:21PM -0700, Tim Kientzle wrote:
> 
>>On further inspection, I'm pretty sure that sys/kern/subr_devstat.c
>>is not correct.
> 
> OK. What about the shared page interface? Specifically the comment
> above devstat_end_transaction().

I'm not entirely happy about it, but maybe I just need to study it more.

The atomic* operations here are implementing a partial lock.  What's
not entirely clear to me is whether it's sufficient or not.  The fact
that the transaction interface is using a different locking mechanism
than the rest of the code is suspicious.

 > The generation count is used by
> the old sysctl interface. The shared page interface has a liberal
> sprinkling of atomic*() instructions.

Back when I was following the Java threading mailing list, I saw a lot
of "what about this" examples that I studied and lots of other people 
studied, and it often took a long time for someone to figure out why it
was wrong.  And, for as long as I was following the list, it
was _always_ wrong.  Every single trick for eliminating locks
that someone came up with didn't work.  Usually, the problem was
some compiler optimization, processor optimization, or cache
behavior that would break the code.

Using generation counts in sysctl code makes perfect sense as a tool
for notifying user space whether something has changed since they
last checked.  Using it as a tool to eliminate locking is another
thing entirely.

Tim