Re: rwsem: down_read_unfair() proposal

[Michel Lespinasse]

On Wed, May 05, 2010 at 11:06:44AM +0100, David Howells wrote:
> Michel Lespinasse <walken@xxxxxxxxxx> wrote:
> 
> > and looks like it's doable with the x86 rwsem implementation as well in a
> > way that would only involve changes to the rwsem spinlock-protected slow
> > paths in lib/rwsem.c .
> 
> It's not as easy as it seems.  Once an XADD-based rwsem is contended, you
> cannot necessarily tell without looking at the queue whether the rwsem is
> currently write-locked or read-locked.

I only said it was doable :) Not done with the implementation yet, but I can
describe the general idea if that helps. The high part of the rwsem is
decremented by two for each thread holding or trying to acquire a write lock;
additionally the high part of the rwsem is decremented by one for the first
thread getting queued. Since queuing is done under a spinlock, it is easy
to decrement only for the first blocked thread there. In down_read_unfair(),
the rwsem value is compared with RWSEM_WAITING_BIAS (== -1 << 16 or 32);
if it's smaller then the rwsem might be write owned and we have to block;
otherwise it only has waiters which we can decide to ignore. This is the
idea in a nutshell.

-- 
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/