Re: pserialized reader/writer locks

[Taylor R Campbell <riastradh%NetBSD.org@localhost>]

   Date: Sat, 26 Jul 2014 19:16:20 +0100
   From: Mindaugas Rasiukevicius <rmind%netbsd.org@localhost>

   Taylor R Campbell <riastradh%NetBSD.org@localhost> wrote:
   >    Date: Sat, 26 Jul 2014 16:48:27 +0100
   >    From: Mindaugas Rasiukevicius <rmind%netbsd.org@localhost>
   > 
   >    http://www.netbsd.org/~rmind/kern_rdlock.c
   > 
   > That makes a single reader cheap; doesn't make multiple readers
   > scalable.

   Huh?

Sorry, I misread some of the comments at the top while skimming too
quickly.  On closer look, I think that basically does the same thing
my non-recursive prwlock does, but maybe with less overhead due to
pserialize(9) (and three times the amount of code so it's a little
harder to see what's going on).  In particular:

- Read enter waits for writer if any, does some per-CPU bookkeeping.
- Read exit notifies writer if there is one waiting.
- Writer enter announce itself, waits until no readers on any CPU.
- Writer exit wakes waiting readers.

If there's no writer, readers need only write to per-CPU structures.
Writers must trigger activity on all CPUs and traverse a list of all
readers.

Many details in how ad's rdlock and my prwlock work are different, but
they both follow that same idea.  It looks like ad's rdlock is not
recursive, so it wouldn't be fit for fstrans(9).  I'm not sure how
hard it would be to adapt it for that purpose.