diff --git a/include/haproxy/task-t.h b/include/haproxy/task-t.h
index 54455c7..1b64b8f 100644
--- a/include/haproxy/task-t.h
+++ b/include/haproxy/task-t.h
@@ -59,6 +59,9 @@
 #define TASK_F_USR1       0x00010000  /* preserved user flag 1, application-specific, def:0 */
 /* unused: 0x20000..0x80000000 */
 
+/* These flags are persistent across scheduler calls */
+#define TASK_PERSISTENT   (TASK_SHARED_WQ | TASK_SELF_WAKING | TASK_KILLED | \
+                           TASK_HEAVY | TASK_F_TASKLET | TASK_F_USR1)
 
 enum {
 	TL_URGENT = 0,   /* urgent tasklets (I/O callbacks) */
diff --git a/src/task.c b/src/task.c
index 8b08675..84873f0 100644
--- a/src/task.c
+++ b/src/task.c
@@ -490,11 +490,10 @@
 		}
 
 		budgets[queue]--;
-		t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
-		state = t->state & (TASK_SHARED_WQ|TASK_SELF_WAKING|TASK_HEAVY|TASK_F_TASKLET|TASK_KILLED|TASK_F_USR1);
-
 		ti->flags &= ~TI_FL_STUCK; // this thread is still running
 		activity[tid].ctxsw++;
+
+		t = (struct task *)LIST_ELEM(tl_queues[queue].n, struct tasklet *, list);
 		ctx = t->context;
 		process = t->process;
 		t->calls++;
@@ -502,7 +501,7 @@
 
 		_HA_ATOMIC_DEC(&sched->rq_total);
 
-		if (state & TASK_F_TASKLET) {
+		if (t->state & TASK_F_TASKLET) {
 			uint64_t before = 0;
 
 			LIST_DEL_INIT(&((struct tasklet *)t)->list);
@@ -519,7 +518,7 @@
 #endif
 			}
 
-			state = _HA_ATOMIC_XCHG(&t->state, state);
+			state = _HA_ATOMIC_FETCH_AND(&t->state, TASK_PERSISTENT);
 			__ha_barrier_atomic_store();
 
 			process(t, ctx, state);
@@ -537,7 +536,11 @@
 
 		LIST_DEL_INIT(&((struct tasklet *)t)->list);
 		__ha_barrier_store();
-		state = _HA_ATOMIC_XCHG(&t->state, state|TASK_RUNNING|TASK_F_USR1);
+
+		state = t->state;
+		while (!_HA_ATOMIC_CAS(&t->state, &state, (state & TASK_PERSISTENT) | TASK_RUNNING))
+			;
+
 		__ha_barrier_atomic_store();
 
 		/* OK then this is a regular task */