Index: libcfa/src/concurrency/ready_queue.cfa
===================================================================
--- libcfa/src/concurrency/ready_queue.cfa	(revision 24e321c53a4cf2985e9827928927e3b7d42e1791)
+++ libcfa/src/concurrency/ready_queue.cfa	(revision c60e50946cae4e184f97d9820dbf62a9ce0f287f)
@@ -100,6 +100,6 @@
 	#define __kernel_rseq_unregister rseq_unregister_current_thread
 #elif defined(CFA_HAVE_LINUX_RSEQ_H)
-	void __kernel_raw_rseq_register  (void);
-	void __kernel_raw_rseq_unregister(void);
+	static void __kernel_raw_rseq_register  (void);
+	static void __kernel_raw_rseq_unregister(void);
 
 	#define __kernel_rseq_register __kernel_raw_rseq_register
@@ -294,6 +294,7 @@
 
 		processor * const proc = kernelTLS().this_processor;
-		const bool external = !push_local || (!proc) || (cltr != proc->cltr);
-
+		const bool external = (!proc) || (cltr != proc->cltr);
+
+		// Figure out the current cpu and make sure it is valid
 		const int cpu = __kernel_getcpu();
 		/* paranoid */ verify(cpu >= 0);
@@ -301,5 +302,15 @@
 		/* paranoid */ verify(cpu * READYQ_SHARD_FACTOR < lanes.count);
 
-		const cpu_map_entry_t & map = cpu_info.llc_map[cpu];
+		// Figure out where thread was last time and make sure it's
+		/* paranoid */ verify(thrd->preferred >= 0);
+		/* paranoid */ verify(thrd->preferred < cpu_info.hthrd_count);
+		/* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count);
+		const int prf = thrd->preferred * READYQ_SHARD_FACTOR;
+
+		const cpu_map_entry_t & map;
+		choose(hint) {
+			case UNPARK_LOCAL : &map = &cpu_info.llc_map[cpu];
+			case UNPARK_REMOTE: &map = &cpu_info.llc_map[prf];
+		}
 		/* paranoid */ verify(map.start * READYQ_SHARD_FACTOR < lanes.count);
 		/* paranoid */ verify(map.self * READYQ_SHARD_FACTOR < lanes.count);
@@ -312,5 +323,8 @@
 			if(unlikely(external)) { r = __tls_rand(); }
 			else { r = proc->rdq.its++; }
-			i = start + (r % READYQ_SHARD_FACTOR);
+			choose(hint) {
+				case UNPARK_LOCAL : i = start + (r % READYQ_SHARD_FACTOR);
+				case UNPARK_REMOTE: i = prf   + (r % READYQ_SHARD_FACTOR);
+			}
 			// If we can't lock it retry
 		} while( !__atomic_try_acquire( &lanes.data[i].lock ) );
@@ -965,5 +979,5 @@
 	extern void __enable_interrupts_hard();
 
-	void __kernel_raw_rseq_register  (void) {
+	static void __kernel_raw_rseq_register  (void) {
 		/* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
 
@@ -983,5 +997,5 @@
 	}
 
-	void __kernel_raw_rseq_unregister(void) {
+	static void __kernel_raw_rseq_unregister(void) {
 		/* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );