On Wed, Oct 11, 2017 at 02:01:53PM -0400, Waiman Long wrote:

+/*
+ * The definition of the atomic counter in the semaphore:
+ *
+ * Bit  0    - writer locked bit
+ * Bit  1    - waiters present bit
+ * Bits 2-7  - reserved
+ * Bits 8-31 - 24-bit reader count
+ *
+ * atomic_fetch_add() is used to obtain reader lock, whereas atomic_cmpxchg()
+ * will be used to obtain writer lock.
+ */
+#define RWSEM_WRITER_LOCKED	0X00000001
+#define RWSEM_FLAG_WAITERS	0X00000002
+#define RWSEM_READER_BIAS	0x00000100
+#define RWSEM_READER_SHIFT	8
+#define RWSEM_READER_MASK	(~((1U << RWSEM_READER_SHIFT) - 1))
+#define RWSEM_LOCK_MASK 	(RWSEM_WRITER_LOCKED|RWSEM_READER_MASK)
+#define RWSEM_READ_FAILED_MASK	(RWSEM_WRITER_LOCKED|RWSEM_FLAG_WAITERS)
+
+#define RWSEM_COUNT_IS_LOCKED(c)	((c) & RWSEM_LOCK_MASK)
+
+/*
+ * lock for reading
+ */
+static inline void __down_read(struct rw_semaphore *sem)
+{
+	if (unlikely(atomic_fetch_add_acquire(RWSEM_READER_BIAS, &sem->count)
+		     & RWSEM_READ_FAILED_MASK))
+		rwsem_down_read_failed(sem);
+}

So I implemented rwsem-mutex (also qrwlock based) that puts

  (unsigned long)current | RWSEM_WRITER

in the atomic_long_t rw_semaphore::owner field. The down-side is that
you can't do fetch_add based __down_read, because that would clobber the
pointer. The up-side is that we have a stable owner pointer (which is
what I needed for PI like things).

I've yet to do performance tests -- and I've not done a bunch of the
obvious optimisations either.