Possible hang in InnoDB rw_lock_lock_word_decr()

In MDEV-10813, the function rw_lock_lock_word_decr() was simplified in an incorrect way:

@@ -275,12 +275,11 @@ rw_lock_lock_word_decr(

 	os_rmb;

 	local_lock_word = lock->lock_word;

 	while (local_lock_word > threshold) {

-		if (os_compare_and_swap_lint(&lock->lock_word,

-					     local_lock_word,

-					     local_lock_word - amount)) {

+		if (my_atomic_caslint(&lock->lock_word,

+				      &local_lock_word,

+				      local_lock_word - amount)) {

 			return(true);

 		}

-		local_lock_word = lock->lock_word;

 	}

 	return(false);

 #else /* INNODB_RW_LOCKS_USE_ATOMICS */

Because we sample the local_lock_word only once at the start of the loop, it is possible that the following happens during rw_lock_s_lock():

1. Another thread is holding an S-latch, so that local_lock_word will be X_LOCK_DECR-1.
2. The other thread will release its S-latch and never come back.
3. An infinite loop will occur, because lock->lock_word will be X_LOCK_DECR from that point onwards.
4. Only a subsequent S-latch acquisition from another thread could resolve the hang.