Re: [PATCH v5] locking/rwsem: Make handoff bit handling more consistent

[Hillf Danton <hdanton@sina.com>]

On Tue, 19 Jul 2022 08:30:02 -0700 Doug Anderson wrote:
> 
> I haven't done any stress testing other than my test case, though, so
> I can't speak to whether there might be any other unintended issues.

The diff below is prepared for any regressions I can imagine in stress
tests by adding changes to both read and write acquirer slow pathes.

On the read side, make lock stealing more aggressive; on the other hand,
write acquirers try to set HANDOFF after a RWSEM_WAIT_TIMEOUT nap to
force the reader acquirers to take the slow path.

Hillf

--- a/kernel/locking/rwsem.c
+++ b/kernel/locking/rwsem.c
@@ -992,13 +992,7 @@ rwsem_down_read_slowpath(struct rw_semap
 	struct rwsem_waiter waiter;
 	DEFINE_WAKE_Q(wake_q);
 
-	/*
-	 * To prevent a constant stream of readers from starving a sleeping
-	 * waiter, don't attempt optimistic lock stealing if the lock is
-	 * currently owned by readers.
-	 */
-	if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
-	    (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
+	if (WARN_ON_ONCE(count & RWSEM_FLAG_READFAIL))
 		goto queue;
 
 	/*
@@ -1169,7 +1163,11 @@ rwsem_down_write_slowpath(struct rw_sema
 				goto trylock_again;
 		}
 
-		schedule();
+		if (RWSEM_FLAG_HANDOFF & atomic_long_read(&sem->count))
+			schedule();
+		else
+			schedule_timeout(1 + RWSEM_WAIT_TIMEOUT);
+
 		lockevent_inc(rwsem_sleep_writer);
 		set_current_state(state);
 trylock_again:
--