Re: [PATCH] Adding Solaris-style "owner of records" to rwlocks

2006/8/8, Suleiman Souhlal <ssouhlal_at_freebsd.org>:
> Attilio Rao wrote:
> > This is a first implementation of the owner of records concept in rwlocks.
> > It allows to avoid the priority inversion problem in the current
> > rwlocks implementation (for readers).
> >
> > The main idea (that John and I discussed) is to have as owner of
> > records the first rlock'er for a "class contention".
> > The implementation consists in adding two flags (RW_LOCK_OWNED and
> > RW_LOCK_EXEMPTED) which are used in order to not penalyze the easy
> > case, and syncronizing the operation of acquiring and dropping the
> > owner of records with the turnstile spin-lock.
> > The main scheme might work in this way:
> >
> > thread1::rlock() -> sets the owner of records
> > thread2::rlock() -> checks for RW_LOCK_OWNED bit and, if it is set, go
> > in the easy case
> > thread3::rlock() -> checks for RW_LOCK_OWNED...
> > thread4::wlock() -> blocks and land its priority to thread1
> > thread1::runlock() -> disable the owner of records (disowning the
> > associated turnstile) and sets the RW_LOCK_EXEMPTED flag. In this way
> > other threads will treact as an easy case.
> > ...
>
> Aren't you missing the hard part: transferring ownership from one reader
> to another? If you don't, you'll still have priority inversions as soon
> as the initial reader unlocks..

Exactly, but having a complete owner switching would be:
1) too hard to achieve in terms of resource taken
2) will imply too many races and we might get a too hard function

With this implementation, only the first rlock (for every class
contention) will be penalyzed while the other are treacted as the
easy/hard case.
It doesn't completely solve the priority inversion problem, but it's
the better compromise between performances/correctnes.

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein