Here's a patch that fixes RW semaphores on the i386 architecture. It is very
simple in the way it works.

The lock counter is dealt with as two semi-independent words: the LSW is the
number of active (granted) locks, and the MSW, if negated, is the number of
active writers (0 or 1) plus the number of waiting lockers of any type.

The fast paths are either two or three instructions long.

This algorithm should also be totally fair! Contentious lockers get put on the
back of the wait queue, and a waker function wakes them starting at the front,
but only wakes either one writer or the first consecutive bundle of readers.

The disadvantage is that the maximum number of active locks is 65535, and the
maximum number of waiting locks is 32766 (this can be extended to 65534 by not
relying on the S flag).

I've included a revised testing module (rwsem.c) that allows read locks to be
obtained as well as write locks and a revised driver program (driver.c) that
can use rwsem.c. Try the following tests:

	driver -200 & driver 200 # fork 200 writers and then 200 readers
	driver 200 & driver -200 # fork 200 readers and then 200 writers

David Howells