[PATCH v2 7/8] locking/rwsem: Use the force

[Waiman Long <longman@redhat.com>]

From: Peter Zijlstra <peterz@infradead.org>

Now that the writer adjustment is done from the wakeup side and
HANDOFF guarantees spinning/stealing is disabled, use the combined
guarantee it ignore spurious READER_BIAS and directly claim the lock.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/locking/rwsem.c | 33 +++++++++++++++++++++++++++++----
 1 file changed, 29 insertions(+), 4 deletions(-)

diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
index ee8861effcc2..7bd26e64827a 100644
--- a/kernel/locking/rwsem.c
+++ b/kernel/locking/rwsem.c
@@ -377,8 +377,8 @@ rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
 /*
  * Remove a waiter from the wait_list and clear flags.
  *
- * Both rwsem_reader_wake() and rwsem_try_write_lock() contain a full 'copy' of
- * this function. Modify with care.
+ * Both rwsem_{reader,writer}_wake() and rwsem_try_write_lock() contain a full
+ * 'copy' of this function. Modify with care.
  *
  * Return: true if wait_list isn't empty and false otherwise
  */
@@ -479,8 +479,33 @@ static void rwsem_writer_wake(struct rw_semaphore *sem,
 			      struct rwsem_waiter *waiter,
 			      struct wake_q_head *wake_q)
 {
-	if (rwsem_try_write_lock(sem, waiter))
-		rwsem_waiter_wake(waiter, wake_q);
+	long count = atomic_long_read(&sem->count);
+
+	/*
+	 * Since rwsem_mark_wake() is only called (with WAKE_ANY) when
+	 * the lock is unlocked, and the HANDOFF bit guarantees that
+	 * all spinning / stealing is disabled, it is posssible to
+	 * unconditionally claim the lock -- any READER_BIAS will be
+	 * temporary.
+	 */
+	if ((count & (RWSEM_FLAG_HANDOFF|RWSEM_WRITER_LOCKED)) == RWSEM_FLAG_HANDOFF) {
+		unsigned long adjustment = RWSEM_WRITER_LOCKED - RWSEM_FLAG_HANDOFF;
+
+		if (list_is_singular(&sem->wait_list))
+			adjustment -= RWSEM_FLAG_WAITERS;
+
+		atomic_long_add(adjustment, &sem->count);
+		/*
+		 * Have rwsem_writer_wake() fully imply rwsem_del_waiter() on
+		 * success.
+		 */
+		list_del(&waiter->list);
+		atomic_long_set(&sem->owner, (long)waiter->task);
+
+	} else if (!rwsem_try_write_lock(sem, waiter))
+		return;
+
+	rwsem_waiter_wake(waiter, wake_q);
 }
 
 static void rwsem_reader_wake(struct rw_semaphore *sem,
-- 
2.31.1