Re: rwlock changes

[yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi)]

> Hi,
> 
> http://www.netbsd.org/~ad/rwlock.diff
> 
> This diff does the following:
> 
> - Use atomic ops directly, since rwlocks work the same way on all platforms.
> - Try to make it a bit more cache efficient, and use branch hints.
> - Fix a bug in rw_downgrade() where the turnstile lock was not released.
> - Remove a couple of redundant assertions.
> - Use atomic_swap instead of atomic_cas in one place where it's safe.
> - After acquiring turnstile lock in rw_vector_enter, check if owner is
>   running again and spin if so.
> 
> It also changes the handoff/release algorithm to work like this:
> 
>          * If we are releasing a write lock, then prefer to wake all
>          * outstanding readers.  Otherwise, wake one writer if there
>          * are outstanding readers, or all writers if there are no
>          * pending readers.  If waking one specific writer, the writer
>          * is handed the lock here.  If waking multiple writers, we
>          * set WRITE_WANTED to block out new readers, and let them
>          * do the work of acquring the lock in rw_vector_enter().
> 
> So, direct handoff is only done when there are readers involved (wakeing
> readers, or readers pending on the lock). Otherwise threads being awoken
> must acquire the lock themselves, making it work like an adaptive mutex in
> that case. It weakens the defence against starvation slightly but I'm
> fairly confident that it shouldn't be a problem.
> 
> The sum of these changes is a reduction of 4% in build time on an 8-core
> machine. There is less idle time during the build because the windows where
> a rwlock can be held but the owner is not running is reduced.
> 
> Comments?
> 
> Thanks,
> Andrew

- please keep the rw_owner bitfield comment.
- rw_swap needs RW_INHERITDEBUG.

YAMAMOTO Takashi