This patch (made against linux-2.4.4-pre6) makes a number of changes to the
rwsem implementation:

 (1) Fixes a subtle contention bug between up_write and the down_* functions.

 (2) Optimises the i386 fastpath implementation and changed the slowpath
     implementation to aid it.
     - The arch/i386/lib/rwsem.c is now gone.
     - Better inline asm constraints have been applied.

 (3) Changed the sparc64 fastpath implementation to use revised slowpath
     interface.
     [Dave Miller: can you check this please]

 (4) Makes the generic spinlock implementation non-inline.
     - lib/rwsem.c has been duplicated to lib/rwsem-spinlock.c and a
       slightly different algorithm has been created. This one is simpler
       since it does not have to use atomic operations on the counters as
       all accesses to them are governed by a blanket spinlock.

David